Embark volume 1 by College of Engineering

EMBARK NORTHEASTERN UNDERGRADUATE ENGINEERING REVIEW

Volume I November 2016

Papers Digitally Tunable Lowpass-Notch Filter Design for Analog Front-Ends in Brain Signal Measurement Applications Kaidi Du, Student, Marvin Onabajo, Professor………………………………………………1 Design of Liquid Nitrogen Capsules for Forest Fire Suppression Craig W. Martland, Student, David P. Marchessault, Student, Andrew McGarey, Student, Diego Rivas, Student, Kevin W. Stanley, Student, and Yiannis Levendis, Professor………12 Maximum Likelihood Image Reconstruction using Data Fusion between X-Ray and Microwave Radar Matthew T. Tivnan, Student, Carey M. Rappaport, Professor………………………………22 Programming Acoustic Modems for Underwater Networking Andrew Tu, Student Member, IEEE, Brian Wilcox, Student Member, IEEE, Mark German, Yashar M. Aval, Member, IEEE, and Stefano Basagni, Senior Member, IEEE……………...28 Microwave ignition for nanostructured reactive composites Gianmarco Vella, Student at Advanced Materials Processing Lab (AMPL)………………...34

In This Volume Digitally Tunable Lowpass-Notch Filter Design for Analog Front-Ends in Brain Signal Measurement Applications Kaidi Du, Student, Marvin Onabajo, Professor A digitally tunable Transconductance-Capacitor Low-pass Notch Filter (LPNF) for Electroencephalography (EEG) application is presented in this research report. Since EEG signals fall into four basic frequency bands, δ (1-4Hz), θ (4-8Hz), α (8-13Hz), and β (13-40Hz), but the power line interference at 60Hz, created by electrode cable and circuitry, has much higher power than the brain signals, the power line interference negatively affects the accuracy of the EEG system. Therefore, a combination of a notch filter and a high-order low-pass filter is employed in this work. With the development of microcontrollers, digital control methods are becoming more frequent in integrated circuit (IC) implementations. Hence, a digital tuning method for this LPNF is in high demand. Due to the digital tuning approach, an automatic calibration of this Gm-C LPNF through a microcontroller can be realized in the future.

Design of Liquid Nitrogen Capsules for Forest Fire Suppression Craig W. Martland, Student, David P. Marchessault, Student, Andrew McGarey, Student, Diego Rivas, Student, Kevin W. Stanley, Student, and Yiannis Levendis, Professor In recent years forest fires have become increasingly frequent, increasingly large and, hence, increasingly catastrophic. As these fires burn unchecked, firefighters strive to extinguish them by dropping water onto affected areas with aerial delivery methods, such as planes and helicopters. Past research at Northeastern University, showed that direct application of liquid nitrogen is very effective at extinguishing fuel pool fires and, thus, research was initiated to explore the application of liquid nitrogen to forest fires. It is hypothesized that liquid nitrogen would be effective at suppressing forest fires, most likely as a twopart approach. Initial application of liquid nitrogen can suppress the flames and subsequent application of water can extinguish deep-seated fires in the pores of the wood. Herein, as an initial step to realize this approach, a capsule was designed to deliver liquid nitrogen to forest fires. This capsule is designed to insulate the liquid nitrogen and minimize in-transit vaporization, whereas incorporation of exterior fins is expected to impart a controlled spin as the capsule falls from the helicopter. This spin will eject liquid nitrogen, which can create a sprinkling effect as it reaches a crown fire whereas any liquid nitrogen remaining in the capsule will be ejected upon impact and will affect the bottom fire. The capsule is made of a single injection molded piece to be cost-effective. Initial tests proved the insulating, spinning and spilling capabilities of the capsule. No fire tests have been conducted yet.

Maximum Likelihood Image Reconstruction using Data Fusion between X-Ray and Microwave Radar Matthew T. Tivnan, Student, Carey M. Rappaport, Professor Data fusion is the process by which measurements collected by two or more sensors are combined

iii

to produce a better result than could have been produced by any of the sensors acting individually. X-ray transmission and Microwave Tomography (MWT) are good candidates for data fusion because of their complementary strengths. For example, X-Ray is known for high spatial resolution structural imaging and MWT provides higher contrast in the physical properties for certain applications. In this work, a simple image reconstruction algorithm is presented which utilizes data fusion between X-Ray and MWT measurements. One possible application in neuroimaging is then simulated in a numerical experiment. The final results show that data fusion has significant advantages over conventional approaches.

Programming Acoustic Modems for Underwater Networking Andrew Tu, Student Member, IEEE, Brian Wilcox, Student Member, IEEE, Mark German, Yashar M. Aval, Member, IEEE, and Stefano Basagni, Senior Member, IEEE Underwater acoustic communication and networks have attracted significant attention in recent years, with applications ranging from ocean monitoring to off-shore sensor control, and port surveillance. Experimental data are required to test and develop effective underwater networking protocols before underwater networks can be successfully deployed for real world applications. Unfortunately, there are very few permanent underwater acoustic testbeds currently in operation, making it difficult for full scale tests to be conducted. To meet the demands for experimental data, we are working to deploy a permanent underwater acoustic network at the Northeastern University Marine Science Center in Nahant, MA. At the final stage, the network will consist of at least five SM 975 Teledyne Benthos acoustic smart modems, with one wirelessly connected to the shore through a smart buoy of our design. This paper describes the interface for programming these modems and how we used it to implement a fundamental protocol to be used as performance benchmark for more advanced underwater solutions.

Microwave ignition for nanostructured reactive composites Gianmarco Vella, Student at Advanced Materials Processing Lab (AMPL) The need for heating at nanoscale has pushed researchers in the study of reactive, nanostructured composites known as nanoheaters. Major topics of interest are the best conditions for consolidation, composition, and ignition of these innovative heat sources. This work presents a new method of ignition for nanoheaters, known as non contact microwave ignition, distinguishing itself from previously developed direct heat application methods. Al-Ni nanoheaters were fabricated through ultrasonic powder consolidation (UPC) with embedded aluminum and copper wires. The conductive properties of the embedded wires, acting as susceptors when exposed to electromagnetic radiation in the microwave range, were found to induce enough heat to Al-Ni nanoheaters to facilitate ignition. This nullifies the requirement of direct heat application to the fabricated nanoheaters to produce ignition. In addition to testing in gaseous environment, this new method of ignition for nanostructured, reactive composites was also tested in vacuum, verifying its effectiveness in a non-gaseous environment.

Digitally Tunable Lowpass-Notch Filter Design for Analog Front-Ends in Brain Signal Measurement Applications Kaidi Du, Student, Marvin Onabajo, Professor

Abstract—A digitally tunable Transconductance-Capacitor Low-pass Notch Filter (LPNF) for Electroencephalography (EEG) application is presented in this research report. Since EEG signals fall into four basic frequency bands, δ (1-4Hz), θ (4-8Hz), α (8-13Hz), and β (13-40Hz), but the power line interference at 60Hz, created by electrode cable and circuitry, has much higher power than the brain signals, the power line interference negatively affects the accuracy of the EEG system. Therefore, a combination of a notch filter and a high-order low-pass filter is employed in this work. With the development of microcontrollers, digital control methods are becoming more frequent in integrated circuit (IC) implementations. Hence, a digital tuning method for this LPNF is in high demand. Due to the digital tuning approach, an automatic calibration of this Gm-C LPNF through a microcontroller can be realized in the future. Keywords—Digital Tuning Method; Binary Way; Lowpass-Notch Filter; Operational Transconductance Amplifier.

I. INTRODUCTION Electroencephalography systems monitor activities in the brain by recording electrical signals from cerebral nerve cells along the scalp. As the only non-invasive method for measuring brain activities from human scalps, it plays an important role in effectively diagnosing patients with severe neuron muscular disorders, and therefore, it is widely used in research of brain nerves and clinical applications [1]. EEG signals can be categorized into four basic bands, δ (1-4Hz), θ (4-8Hz), α (8-13Hz), and β (13-40Hz) [2]. Since brain signals are very weak, ranging from 2 to 200µV, compared to the Signal to Noise Ratio (SNR) from the power line, the strongest noise source, it is beneficial to include a LPNF in the analog front-end (AFE) of EEG measurement devices [3]. However, fabrications and environmental variations may cause the actual notch frequency to deviate from the nominal value, so there must be a reliable tuning method for the LPNF architecture in order that the notch frequency can be easily calibrated. As a microcontroller has advantages of small size, programmable input and output peripherals, all-in-one design (containing the processor, RAM, and I/O), and low cost, it can

be used to test preliminary designs prior to integrating analog signal processing and digital calibration circuits on the same chip, for which the work described in this report is preparing. The LPNF used in this work is mainly referenced by a single-ended Gm-C filter structure first proposed by Qian et al. and a fully-differential CMOS LPNF designed by Kainan Wang from the Analog and Mixed-Signal Integrated Circuit Research Laboratory at Northeastern University [2], [3]. In Section II, a digitally tunable notch filter is elaborated upon. Then, a comprehensive simulation assessment with Cadence software tools and actual measurements is presented in Section III and Section IV. The focus in Section V is on the future improvements of this digitally tunable LPNF. Conclusions are given in Section VI.

II. TUNABLE FILTER DESIGN A. Fifth Order Elliptic Low-pass Notch Filter The Transconductance-C (OTA-C) filter in Fig. 1 realizes a fifth-order elliptic filter with a low pass transfer function that contains a notch in Fig. 2. The null frequencies play an important role in determining the notch frequency and stopband ripple in the low-pass filter. Because the brain signals of interest only fall into low frequencies ranging up to 40Hz, the signals out of this range will be treated as noise. In order to remove these noises, it is necessary to have a low-pass filter with extremely narrow transition band. Not only does this kind of elliptic filter act as a low-pass filter, but also it is used as a notch filter. Since the strong power line interference of the EEG system is around 60Hz, it is required to have sufficient attenuation at the notch frequency. In this fifth-order elliptic filter, the null frequencies are very close to the cut-off frequency of the transfer function, which is why the notch is just at the outside of the passband range. Hence, this fifth-order elliptic filter can combine the advantages of both notch filters and low-pass filters. The schematic of the active LPNF is shown in Fig. 1. Since this LPNF is mainly designed for on-chip circuits, the size (chip area) of the filter is one of the main concerns. As OTAs in integrated circuits are much smaller than that of the on-chip inductors, the OTA-C filter is a better design choice for implementation on chips. Additionally, it allows designing the passband gain of the active filters by using Gm-C structures. In this design, all six OTAs have the same transconductance Gm, so the pass band gain is 0dB. Moreover, in order to have deeper

2 notch, two null frequencies of the OTA-C filters are combined to a single notch in this fifth-order elliptic filter. According to the schematic in Fig. 1, these two null frequencies can be expressed as: đ?&#x2018;&#x201C;" =

1 đ??ş()*+ Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â (1) 2đ?&#x153;&#x2039; đ??ś-" đ??ś.

đ?&#x2018;&#x201C;. =

1 đ??ş()*+ Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â (2) 2đ?&#x153;&#x2039; đ??ś-. đ??ś2 đ?&#x2018;&#x201C;" = đ?&#x2018;&#x201C;. Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â (3)

Because these two notches overlap each other, the final notch in this design is deeper compared to the standard elliptic filter design. As tuning for GmOTA may affect the ripples in the pass-band specification, it is better to fix the value of Gm during the tuning process. Therefore, the only way to tune the notch frequency is changing the values of CL1 and CL2 or C2 and C4 by using variable capacitors or by adding extra capacitors with switches. Usually, the switch is easier to operate when one of its two terminals is grounded. Fig. 1 shows that one terminal of CL1 and CL2 is grounded but C2 and C4 are floating between two nodes, so adding extra capacitors with switches to CL1 and CL2 is a better way to tune the notch frequency than using C2 and C4.

Fig. 1. Schematic of the fifth-order single-ended low-pass notch filter [2].

B. Â Digital Notch Frequency Tuning Since microcontrollers can read, implement, and generate digital signals at I/O interfaces, which are chip-operable, inexpensive, and programmable devices with low power consumptions and supported programming environments across multiple standard platforms; they are widely used in the electronics field [5]. The LPNF in this paper is intended for an on-chip application, and therefore, a tuning method involving digital controlling by microcontrollers or on-chip digital circuits will help to compensate for manufacturing process variations. According to the equation (1) and (2), tuning the grounded capacitors, which are CL1 and CL2 shown in Fig. 1, will lead to changes in the notch frequency. Hence, if the value of the grounded capacitors can be controlled by digital signals, digitally tuning the notch frequency will be put into practice. In other words, the notch frequency of the LPNF can be digitally

Fig. 2. Typical fifth-order elliptic filter with a low pass transfer function [3]

controlled by connecting digital switches in series with extra capacitors that are parallel with the fixed-value grounded capacitors (CL1 and CL2). The detailed schematic is presented in Fig A-1. In the circuit, all switches are implemented with discrete N-type metal-oxide-semiconductor (NMOS) transistors. Each NMOS switch controls the connection of one extra-grounded capacitor. The drain of the NMOS connects to one lead of an extra-grounded capacitor and its source terminal connects to the ground. In the future, the gate voltage of the NMOS can be controlled by a programmable microcontroller. Currently, the on and off state of the NMOS switch is manually controlled by connecting the gate either to the low voltage power line (3-4V) or to ground, which imitated the digital output signal of the microcontroller. If the voltage at the gate is higher than the threshold voltage of the NMOS, which is equal to voltage value of the low voltage power line, the extra capacitor will be connected in parallel with the fixed value grounded capacitors so that the total value of the grounded capacitors increases and the notch frequency decreases. Digital tuning of the notch frequency is possible by controlling the on and off states of the NMOS switches.

C. Â Binary Digital Tuning Approach For the purpose of tuning the notch frequency of the LPNF with the largest tuning range and the smallest increment, it is effective and economic to use a binary-weighted method for changing the total values of grounded capacitors. There are 8 bits available in the LPNF design of this work. The value of capacitor at each bit is calculated based on the binary arithmetic rules: đ??ś4 â&#x2030;&#x2026; đ??ś6 Ă&#x2014;24 Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â (4) Where đ??ś4 Â is the capacitor value at đ?&#x2018;&#x203A;:; bit, and đ??ś6 is the capacitor

value at LSB.The calculated capacitor value for each code is shown in Table 1.

In order to actually build this filter, the component selections also depend on the availability of standard values offered by

3 TABLE I CALCULATED EXTRA-G ROUNDED CAPACITOR VALUES AT EACH CODE

component distributors. To make calculations simple, all GmOTA values of OTAs are set to be 9.6mS; C1, C3, and C5 are equal to 30ÂľF; C2 and C4 are equal to 20 ÂľF (Fig. A-1). Cancelling out the power interference at 60Hz is the main goal of this LPNF, so the notch frequency is set to 60Hz. If all values are substituted into equations (1)-(3), the desired notch frequency would be achieved: AB

đ?&#x2018;&#x201C;< = 60 =

1 9.6Ă&#x2014;10 Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â 5 2đ?&#x153;&#x2039; đ??śCDEF4G_:E:IJ_< Ă&#x2014;20Ă&#x2014;10AK

Where Â đ?&#x2018;&#x201C;< , đ?&#x2018;&#x201C;" , đ?&#x2018;&#x17D;đ?&#x2018;&#x203A;đ?&#x2018;&#x2018;đ?&#x2018;&#x201C;. are all equal to the notch frequency and đ??śCDEF4GPQPRS = đ??ś-< + T

đ??śđ?&#x2018;&#x2122;đ?&#x2018;&#x2013;X4 Â Â Â Â Â Â Â Â Â Â Â Â Â (7) 4

To determine the acceptable range of CLi, the worst case must be considered. According to the data sheets [6]-[9], all capacitors have the same 20% tolerance and the range of the GmOTA is from 6.7Ă&#x2014;10AB đ?&#x2018;&#x2020; to 13Ă&#x2014;10AB đ?&#x2018;&#x2020;. Therefore, there are two extreme cases:

a. Â If GmOTA is at the lower bound (i.e., GmOTA=6.7Ă&#x2014;10AB đ?&#x2018;&#x2020;)

and all capacitors are at the upper bound (i.e., đ??śĂ&#x2014;(1 + 20%), when only đ??ś-< is used), the notch frequency, which is at the upper bound of the notch frequency range, must be greater than 60Hz. The mathematical inequality expression is K.bĂ&#x2014;"6cd

efT_gQhTgRS Ă&#x2014;("i.6%)Ă&#x2014;.6â&#x2C6;&#x2014;("i.6%)Ă&#x2014;"6ck

Â Â Â Â Â Â Â Â 8

which implies đ??ś-<_4E(<4IJ < 10.96ÂľÎźđ??š Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â (9)

nth Binary Code 0 (LSB) 1 2 3 4 5 6 7 (MSB) The sum of all extra capacitors

"B.6Ă&#x2014;"6cd

" .a

(efT ieopqh )Ă&#x2014;("A.6%)Ă&#x2014;.6â&#x2C6;&#x2014;("A.6%)Ă&#x2014;"6ck

Â Â Â Â Â Â Â Â Â Â Â Â 10

where the sum of the nominal value of all extra capacitors is CX_Sum, implying that:

đ??śđ?&#x2018;&#x2122;đ?&#x2018;&#x2013;X4 Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â (6)

đ??śCDEF4GPQPRS â&#x2030;&#x2026; 32.42 Â ÂľÎźđ??š = đ??ś-< +

60 >

đ??ś-< + đ??śr_sF( > 92.90ÂľÎźđ??š Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â (11)

Where đ?&#x2018;&#x2013; = {1,2}, đ?&#x2018;&#x203A; = {0,1,2,3,4,5,6,7} Â . Rearranging equation (4) results in

60 <

b. Â If GmOTA is at the lower bound (i.e. GmOTA=13.0Ă&#x2014;10AB đ?&#x2018;&#x2020;), and all capacitors are at the upper bound (i.e. đ??śĂ&#x2014;(1 â&#x2C6;&#x2019; 20%) , when all extra capacitors are used), the notch frequency, which is at the lower bound of the notch frequency range, must be less than 60Hz. The mathematical inequality expression is

Capacitor Value at Each Code 0.33 Â ÂľÎźđ??š 0.66 Â ÂľÎźđ??š 1.32 Â ÂľÎźđ??š 2.64 Â ÂľÎźđ??š 5.28 Â ÂľÎźđ??š 10.56 Â ÂľÎźđ??š 21.12 Â ÂľÎźđ??š 42.24 Â ÂľÎźđ??š 84.15 Â ÂľÎźđ??š

The larger đ??ś-< is, the smaller đ??śr_sF( is. Because it is better to design with a small increment, đ??śr_sF( Â should be small. In this TABLE II AVAILABLE EXTRA-G ROUNDED CAPACITOR V ALUES AT EACH CODE

prototype, the largest đ??ś-< value is 10 Â ÂľÎźđ??š, requiring that đ??śr_sF( must be greater than 82.92 Â ÂľÎźđ??š. Hence, the capacitor value at the LSB must be đ??ś-st >

82.92ÂľÎźđ??š â&#x2030;&#x2026; 0.325ÂľÎźđ??š Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â (12) 2u â&#x2C6;&#x2019; 1

The closest available capacitor value of 0.33 Â ÂľÎźđ??š is selected. The closest available capacitors at other bits are summarized in Table 2. The expected results for the typical and two worst cases are listed in Table 3.

nth Binary Code

Typical Value

Anticipated minimum value with variation

0 (LSB) 1 2 3 4 5 6 7 (MSB)

0.33 Â ÂľÎźđ??š 0.68 Â ÂľÎźđ??š 1.2 Â ÂľÎźđ??š 2.7 Â ÂľÎźđ??š 5.6 Â ÂľÎźđ??š 12 Â ÂľÎźđ??š 22 Â ÂľÎźđ??š 47 Â ÂľÎźđ??š

0.26 Â ÂľÎźđ??š 0.54 Â ÂľÎźđ??š 0.96 Â ÂľÎźđ??š 2.16 Â ÂľÎźđ??š 4.48 Â ÂľÎźđ??š 9.60 Â ÂľÎźđ??š 17.60 Â ÂľÎźđ??š 37.60 Â ÂľÎźđ??š

Anticipated maximum value with variation 0.40 Â ÂľÎźđ??š 0.82 Â ÂľÎźđ??š 1.44 Â ÂľÎźđ??š 3.24 Â ÂľÎźđ??š 6.72 Â ÂľÎźđ??š 14.40 Â ÂľÎźđ??š 26.40 Â ÂľÎźđ??š 56.40 Â ÂľÎźđ??š

The sum of all extra capacitors (CX_Sum)

91.51 Â ÂľÎźđ??š

73.21 Â ÂľÎźđ??š

109.81 Â ÂľÎźđ??š

4 TABLE IV PARAMETERS FOR SIMULATIONS IN CADENCE Â

TABLE III EXPECTED NOTCH FREQUENCY RESULTS IN DIFFERENT SIMULATIONS

Case Typical Case Worst Case 1 (G_min, C_max) Worst Case 2 (G_max, C_min)

gm 9600 Â ÂľÎźđ?&#x2018;&#x2020;

6700 Â ÂľÎźđ?&#x2018;&#x2020;

13000 Â ÂľÎźđ?&#x2018;&#x2020;

fmin 33.90954426 Hz 19.72169675 Hz

57.39896815 Hz

fmax 108.0379579 Hz 62.83457623 Hz

182.8767517 Hz

Parameters in Cadence Simulation Typical

GminCmax

GmaxCmin

Code when Notch @60Hz

01000110

00000110

11101101

Voltage to Open the Switch

< 999.9mV

> 1V

Voltage to Close the Switch

> 1V

R_OpenSwitch

192Kâ&#x201E;Ś

R_ClosedSwitch

365.2mâ&#x201E;Ś

CL_1

10ÂľF

12ÂľF

8ÂľF

CL_2

10ÂľF

12ÂľF

8ÂľF

cl2_D7

47ÂľF

56.4ÂľF

37.6ÂľF

cl2_D6

22ÂľF

26.4ÂľF

17.6ÂľF

cl2_D5

12ÂľF

14.4ÂľF

9.6ÂľF

cl2_D4

5.6ÂľF

6.72ÂľF

4.48ÂľF

cl2_D3

2.7ÂľF

3.24ÂľF

2.16ÂľF

cl2_D2

1.2ÂľF

1.44ÂľF

0.96ÂľF

cl2_D1

0.68ÂľF

0.82ÂľF

0.54ÂľF

cl2_D0

0.33ÂľF

0.4ÂľF

0.26ÂľF

D27

D26

D25

D24

D23

D22

D21

D20

cl1_D7

47ÂľF

56.4ÂľF

37.6ÂľF

cl1_D6

22ÂľF

26.4ÂľF

17.6ÂľF

cl1_D5

12ÂľF

14.4ÂľF

9.6ÂľF

cl1_D4

5.6ÂľF

6.72ÂľF

4.48ÂľF

cl1_D3

2.7ÂľF

3.24ÂľF

2.16ÂľF

cl1_D2

1.2ÂľF

1.44ÂľF

0.96ÂľF

cl1_D1

0.68ÂľF

0.82ÂľF

0.54ÂľF

According to the simulation results in Table 5, it is clear that if Gm is fixed, the notch frequency is inversely correlated to the value of the grounded capacitors (CL1 and CL2). On the other hand, if the value of the grounded capacitors is fixed, the notch frequency increases with the value of Gm. This simulation results agree with the calculated results based on the equations (1) and (2).

cl1_D0

0.33ÂľF

0.4ÂľF

0.26ÂľF

D17

D16

D15

D14

D13

D12

III. Â SIMULATION RESULTS All circuits were designed and simulated with ideal components in Cadence. The schematics in Fig. A-2 and Fig. A-3 show the macro model of the LPNF in Cadence. The OTA is modeled as shown in Fig. 3. Table 4 summarizes the simulated specifications of the standalone LPNF. The results are indicated in Table 5. The outputs of the Cadence simulations are shown in Fig. 4, Fig. 5, and Fig. 6. Fig. 4 represents the simulation result when all capacitors and OTAs are at the nominal values. When the ideal switch codes are 11111111 (all NMOS switches are on), the simulated notch frequency is at 34.51Hz; when ideal switches codes are 01000110, the notch frequency is at 60.20Hz; when ideal switches codes are 00000000 (all NMOS switches are off), the notch frequency is at 111.68Hz. Table 5 shows the simulation results in other two extreme cases (the largest gm with smallest capacitors and the smallest gm with largest capacitors, which corresponds to Fig. 5 and Fig. 6).

Fig. 3. OTA macro model.

5 TABLE IV (CONT.) PARAMETERS FOR SIMULATIONS IN CADENCE

Typical

GminCmax

GmaxCmin

D11

D10

30µF

36µF

24µF

20µF

24µF

16µF

30µF

36µF

24µF

20µF

24µF

16µF

30µF

36µF

24µF

9600µS

6700µS

13000µS

Rin

26KΩ

Rout

1MΩ

Vsin

1mV

Fig. 5. Simulated results with minimum Gm and maximum capacitor values.

TABLE V SUMMARY OF SIMULATION RESULTS

G_typ and C_typ

G_min and C_max

G_max and C_min

The closest frequency to 60Hz

59.98 Hz

59.70 Hz

60.25 Hz

fmin

34.51 Hz

20.14 Hz

58.34 Hz

fmax

111.17 Hz

65.16 Hz

187.93 Hz

Gain @60 Hz

-61.64 dB

-58.78 dB

-66.79 dB

Gain @fmin(11111111)

-64.77 dB

-62.64 dB

-66.83 dB

Gain @fmax(00000000)

-60.83 dB

-59.28 dB

-62.55 dB

Switch Code to place the notch at 60HZ (MSB→LSB)

01000110

00000110

11101101

Fig. 4. Simulated results with typical values of Gm and for capacitors.

Fig. 6. Simulated results with maximum Gm and minimum capacitor values.

IV. MEASUREMENTS The proposed filter was assembled on a breadboard with components ordered from Mouser Electronics. Table 6 lists the values of key components parameters shown in Fig. A-1. The photo of the circuit is displayed in Fig. 7. All OTAs are biased identically and operate with the same supply voltages of ±18V. There are some differences between Table 6 and Table 4. All ideal switches in Fig. A-1 are replaced by NMOS switches with the same threshold voltage. The values of CL_1 and CL_2 in measurements are larger than those in simulations. The reason is that manufacturing variations were higher than the expected. After adding extra capacitors to CL_1 and CL_2, the total value of the fixed grounded capacitors was 40µF to place the notch frequency at 60Hz. More specifically, 30µF capacitors were added to the original CL_1 (10µF) and CL_1 (10µF), which are marked as four 15µF capacitors in Fig. 7. As a result, the smallest frequency which the notch can reach is about 58Hz.

6 cl2_D7

47µF

cl2_D6

22µF

cl2_D5

12µF

cl2_D4

5.6µF

cl2_D3

2.7µF

cl2_D2

1.2µF

cl2_D1

0.68µF

cl2_D0

0.33µF

cl1_D7

47µF

cl1_D6

22µF

cl1_D5

12µF

cl1_D4

5.6µF

cl1_D3

2.7µF

cl1_D2

1.2µF

cl1_D1

0.68µF

cl1_D0

0.33µF

TABLE VI PARAMETERS DURING MEASUREMENT BASED ON SCHEMATIC IN FIG. A-1

30µF

20µF

Measurement

30µF

20µF

30µF

vsin

1mV

Fig. 7. Photograph of the LPNF.

Notch Frequency (Hz) Code

60 Hz (Typical)

58 Hz

120 Hz

11111100

11111111

00000000

D27

D26

D25

D24

D23

D22

D21

D20

D17

D16

D15

D14

D13

D12

D11

D10 NMOS Threshold VDC

3.75V

R_buffer1

10KΩ

R_buffer2

62KΩ

C_buffer

0.001µF

CL_1

40µF

CL_2

40µF

Fig. 8 shows the measured transfer function of the LPNF in Fig. 7. This transfer function is based on the collected data from measuring output voltage and input voltage at different frequencies with a Tektronix DPO2024B oscilloscope. It shows that the range of notch frequency in the measurements is from 58Hz to 120Hz depending on the digital code setting. The captured signals at points 1, 2 and 3 in Fig. 8 are shown in Figurers 9-14 (point 1, 2, and 3 are all for the transfer function with the code for a notch at 60Hz). In Figures 9-14, the unit of the vertical axis is dBV(RMS) and the unit of the horizontal axis is Hz. Fig. 9 and Fig. 10 show that when the input signal is at point 1 (5Hz) in Fig. 8, the output signal has

Fig. 8. Transfer function of the LPNF circuit.

7 approximately the same amplitude as the input signal. In other words, the gain at 5Hz is almost 0dB as expected. Fig. 11 and Fig. 12 show that the gain at point 2 (55 Hz) in Fig. 8, where stop-band attenuation occurs, is about -70.3-(-30.3)=-40dB. Fig. 13 and Fig. 14 show that when the input signal is at point 3 (60Hz) in Fig. 8, where the notch frequency is located, the gain is -98.1-(-28.8)=-69.7dB. The flat line in Fig. 14 indicates that the output signal is below the noise level, which agrees with expectations. Fig. 13. Input signal waveform and FFT at point 3.

Fig. 9. Input signal waveform and FFT at point 1.

Fig. 14. Output signal waveform and FFT at point 3.

V. FUTURE WORK — AUTOMATIC DIGITALLY-ASSISTED

CALIBRATION

Fig. 10. Output signal waveform and FFT at point 1.

Fig. 11. Input signal waveform and FFT at point 2.

Fig. 12. Output signal waveform and FFT at point 2.

In the future, microcontrollers can be added to control this LPNF design. The microcontrollers will not only control on and off states of NMOS switches, but also automatically determine which switches are on. In other words, microcontrollers will realize the automatic calibration of the LPNF. In order to achieve this goal, there must be a feedback circuit returning the information at the output of the LPNF to the microcontroller. In this case, the Beagle Bone Black microcontroller board will be used as an example. The LPNF design in this paper needs sixteen digital output pins for controlling on and off states of the NMOS switches and two input pins for connecting the feedback circuit. The Beagle Bone Black microcontroller board has sixty five available digital I/Os [10]. The output “high” voltage of each digital output is 3.3V, which is higher than the threshold voltage (1-2V) of the NMOS used in this measurement. The major part of the feedback circuit design is the Amplitude Detector. In each cycle of the calibration, the output of the LPNF will be monitored by the Amplitude Detector and then send the result to the microcontroller. The logic structure of the envisioned amplitude detection scheme is shown in Fig. 15. There are two outputs in this Amplitude Detector—the Amplitude Detector Output and the On Detector Output. These two outputs will connect with the input pins of the microcontroller. When the calibration of the LPNF starts, a 60Hz test signal will enter the filter, and then the output signal of the LPNF will be compared with the reference voltages (VREF_TYP and VREF_ON) in the Amplitude Detector. VREF_TYP is the threshold voltage for detecting that the signal at 60Hz is cancelled out. If the amplitude of the output signal from LPNF is smaller than VREF_TYP but larger than VREF_ON, it means that the signal at 60Hz is filtered.

8 Similarly, VREF_ON is the threshold voltage for detecting that the LPNF is on. When the amplitude of the output signal from LPNF is smaller than VREF_ON, the LPNF is off. For instance, in Fig. 15, there are three sample output waveforms from the LPNF: Waveform 1 occurs when the LPNF is powered off, so there will be the signal which is smaller than VREF_ON coming into the Amplitude Detector; Waveform 2 happens when the notch frequency of the LPNF is at 60Hz, the amplitude of the output signal from the LPNF will be smaller than the unfiltered typical signal amplitude VREF_TYP, which means that most of the output signals of LPNF at 60Hz are canceled out; Waveform 3 happens when the notch frequency of the LPNF is not at 60Hz so the amplitude of the output signal from the LPNF is larger than VREF_TYP. To sum up, the Amplitude Detector’s output results of each waveform are listed in Table 7.

Fig. 15. Amplitude detection scheme in the feedback circuit for the LPNF [12].

The microcontroller will start calibrations with switch codes from 00000000 to 11111111, which means that the notch frequency begins at the maximum value and gradually decreases to the minimum value. To realize this function, the microcontroller will increase one LSB of the switches code each time; i.e., sequentially closing switches to increase the total grounded capacitors value with an increment of 0.33µF. When the feedback signal from the amplitude detector is “11,” the notch frequency of the LPNF is at 60Hz and the microcontroller will stop increasing the switches code; i.e., closing NMOS switches. Consequently, the automatically digitally-assisted calibration through microcontrollers will be realized with the feedback circuit in the future. TABLE VII OUTPUT CODE FOR DIFFERENT WAVEFORMS

Wave form

Amplitude Detector Output

On Detector Output

Meaning

LPNF is Power Off

Notch Frequency Calibrated

Notch Frequency is not at 60Hz

VI. CONCLUSION A digitally tunable fifth order Transconductance-Capacitor (Gm-C) Low-pass Notch Filter for EEG applications has been presented. The main goal of this design is to realize digitally tuning of the notch frequency to be at 60Hz so that it can filter out the power line interference signal of the EEG front-end system. In the simulations, this LPNF achieves a 61.64dB deep notch at 60 Hz with typical values for each component (using macro models). The notch frequency can be tuned from 34.51 Hz to 111.17Hz. Comparatively, in the real measurement, the LPNF achieves a 69.7dB deep notch at 60Hz. The notch frequency of the real LPNF can be tuned from 58Hz to 120Hz through changing the total value of the grounded capacitors. Because the transconductance of the OTA may vary due to manufacturing process variations or changes of its temperature, a difference between simulations and actual measurements is normal. In the actual LPNF, the total value of the grounded capacitors can be controlled by the on and off state of the NMOS switches, which connect with extra grounded capacitors (cl1_D1~7 and cl2_D1~7 shown in Table 6) in series. By changing the gate voltage of each NMOS switch, the connection of the extra grounded capacitors can be easily added to the fixed grounded capacitors (CL_1 and CL_2). As the values of the extra grounded capacitors are distributed in a binary way, changing the switches code; i.e., changing the gate voltage, will lead to the movements of the notch frequency. The gate voltage of each NMOS switch can be controlled by a microcontroller, such as the Beagle Bone Black microcontroller board, and the feedback circuit will help the microcontroller to calibrate the notch frequency to be at 60Hz. In the future, a microcontroller and a feedback circuit can be added to this LPNF design.

APPENDIX

Fig. A-1. Actual LPNF circuit connection.

Fig. A-2. Macro Model layout design of LPNF in Cadence.

Fig. A-3. Zoomed-in macro model for OTA.

ACKNOWLEDGMENT This work was supported by Analog & Mixed-Signal Integrated Circuit (AMSIC) Research Laboratory at Northeastern University. The authors thank to Kainan Wang, Li Xu, and Alireza Zahrai for valuable discussions. Kaidi Du received a B.S. degree (summa cum laude) in Electrical Engineering from Northeastern University in 2016. She conducted this independent study project and got the college honors distinction in Electrical Engineering at Northeastern University. She will continue her graduate study in Electrical Engineering at UC Berkeley in Fall 2016. Marvin Onabajo is an Assistant Professor in the Electrical and Computer Engineering Department at Northeastern University. He received a B.S. degree (summa cum laude) in Electrical Engineering from The University of Texas at Arlington in 2003 as well as the M.S. and Ph.D. degrees in Electrical Engineering from Texas A&M University in 2007 and 2011, respectively. During his final year at UT-Arlington he worked in the Analog and Mixed-Signal IC group in affiliation with the National Science Foundation’s Research Experiences for Undergraduates program. From 2004 to 2005, he was Electrical Test/Product Engineer at Intel Corp. in Hillsboro, Oregon. He joined the Analog and Mixed-Signal Center at Texas A&M University in 2005, where he was engaged in research projects involving analog built-in testing, data converters, and on-chip temperature sensors for thermal monitoring. In the Spring 2011 semester, he worked as a Design Engineering Intern in the Broadband RF/Tuner Development group at Broadcom Corp. in Irvine, California. Marvin Onabajo has been at Northeastern University since the Fall 2011 semester. His current research areas are analog/RF integrated circuit design, on-chip built-in testing and calibration, mixed-signal integrated circuits for medical applications, data converters, and on-chip sensors for thermal monitoring. He received the 2015 CAREER Award from the National Science Foundation.

REFERENCES [1]

[2]

[3]

[4] [5]

X. Li, et al., “A High Precision EEG Acquisition System Based on the CompactPCI Platform,” in Proc. IEEE Intl. Conf. on BioMedical Engineering and Informatic (BMEI), 2014, pp. 511-516. doi: 10.1109/BMEI.2014.7002828 X. Qian, Y. P. Xu, and X. Li, “A CMOS continuous-time low-pass notch filter for EEG system,” Analog Integrated Circuits Signal Processing, vol. 44, no. 3, pp. 231-238, Sep. 2005. W. Kainan, C. Chun-hsiang, M. Onabajo, “A Fully-Differential CMOS Low-Pass Notch Filter for Brain signals Measurement Devices with High Interference Rejection,” in Proc. IEEE Intl. Conf. on Circuits and Systems (MWSCAS), July 2014, pp. 1041-1044. doi: 10.1109/MWSCAS.2014.6908596 J. David, M. Ken, “Continuous-time Filters,” Analog Integrated Circuit Design, New York: John Wiley & Sons, 1997, pp. 574-642. H. Kenneth, T. Daniel, “Computer, Microprocessor, and Microcontroller Architectures,” Microcontrollers :

[6]

[7]

[8]

[9]

[10]

[11]

architecture, implementation and programming, New York: McGraw-Hill, 1992, pp. 1-40. NJR Corporation. “DUAL OPERATIONAL TRANSCONDUCTANCE AMPLIFIER,” NJM13600/13700 datasheet [Online]. Available: http://www.mouser.com/ds/2/294/NJM13600_NJM137 00_E-346722.pdf April 2006. [Aug. 2015]. Panasonic. “Aluminum Electrolytic Capacitors,” KS datasheet [Online]. Available: http://www.mouser.com/ds/2/315/KK-197348.pdf [Aug. 2015]. Panasonic. “Aluminum Electrolytic Capacitors,” GA datasheet [Online]. Available: http://www.mouser.com/ds/2/315/ABA0000C1041-947 478.pdf [Aug. 2015]. nichicon. “Aluminum Electrolytic Capacitors,” ULD datasheet [Online]. Available: http://www.mouser.com/ds/2/293/e-uld-884079.pdf [Aug. 2015]. BeagleBone, “65 possible digital I/Os,” Beagleboard.org. Available: http://beagleboard.org/support/bone101 Mar. 2015. [Aug. 2015]. L. Xu, F. Junpeng, N. Yuchi, M. Onabajo, “Test Signal Generation for the Claibration of Analog Front-End Circuits in Biopotential Measurement Applications,” in Proc. IEEE Intl. Conf. on Circuits and Systems (MWSCAS, Aug. 2014, pp. 949-952. doi: 10.1109/MWSCAS.2014.6908573

Design of Liquid Nitrogen Capsules for Forest Fire Suppression Craig W. Martland, Student, David P. Marchessault, Student, Andrew McGarey, Student, Diego Rivas, Student, Kevin W. Stanley, Student, and Yiannis Levendis, Professor Department of Mechanical and Industrial Engineering Northeastern University Abstractâ&#x20AC;&#x201D; In recent years forest fires have become increasingly frequent, increasingly large and, hence, increasingly catastrophic. As these fires burn unchecked, firefighters strive to extinguish them by dropping water onto affected areas with aerial delivery methods, such as planes and helicopters. Past research at Northeastern University [1], [2] showed that direct application of liquid nitrogen is very effective at extinguishing fuel pool fires and, thus, research was initiated to explore the application of liquid nitrogen to forest fires. It is hypothesized that liquid nitrogen would be effective at suppressing forest fires, most likely as a twopart approach. Initial application of liquid nitrogen can suppress the flames and subsequent application of water can extinguish deep-seated fires in the pores of the wood [1]. Herein, as an initial step to realize this approach, a capsule was designed to deliver liquid nitrogen to forest fires. This capsule is designed to insulate the liquid nitrogen and minimize in-transit vaporization, whereas incorporation of exterior fins is expected to impart a controlled spin as the capsule falls from the helicopter. This spin will eject liquid nitrogen, which can create a sprinkling effect as it reaches a crown fire whereas any liquid nitrogen remaining in the capsule will be ejected upon impact and will affect the bottom fire. The capsule is made of a single injection molded piece to be cost-effective. Initial tests proved the insulating, spinning and spilling capabilities of the capsule. No fire tests have been conducted yet.

I. Â INTRODUCTION Forest fires are becoming increasingly common around the globe as increasing population, water usage, droughts, and global climate change induce favorable conditions. From 2007 to 2011 an average of over 75,000 wildfires occurred each year costing a total of over $8 billion [3], [4]. Currently, ground crews use a combination of water, fire, and trenches to contain a fire long enough that it can burn out [5]. Once a fire has spread to an area, it cannot be permanently extinguished, and the area cannot be saved. Airplanes and helicopters cannot carry enough water to effectively suppress forest fires, and much of the water evaporates before reaching the fire [5]. More acres of woodland were damaged from 2005-2009 than in the entire 1990â&#x20AC;&#x2122;s [4]. Current fire containment methods are failing to keep pace with the number and intensity of forest fires that are seen today, suggesting that additional research is required into more effective fire suppression methods.

Previous research at Northeastern University demonstrated the effectiveness of direct application of liquid nitrogen (LN2) in extinguishing fuel pool fires [1], [2]. It was found that bringing the cryogen into contact with a pyrolysing/burning surface causes an abrupt phase change followed by a very large thermal expansion. LN2 absorbs more energy to vaporize and heat up in a fire than water (H2O), as it is colder and has a thermal expansion ratio of 1:694, and approximately 1:1,000 when exposed to flame temperatures [6]. The vaporizing cryogen forms a cloud over the pyrolysing/burning surface, thus cooling the surface and thereby reducing its pyrolysis rate. Therefore, the pyrolysis gases become inert and the fire is starved of air. These phenomena lead to expedient fire extinction. The pyrolysing surface is then blanketed for a considerable period of time with nitrogen gas and re-ignition is impeded. Forest fires can be of considerable magnitude, and may require tremendous resources to suppress and, eventually, extinguish. Therefore, it is expected that the application of LN2 to forest fires will be particularly challenging, due to (a) the very large size of such fires and (b) to the fact that fire enters the pores of wood as it burns. Hence, extinction of the burning pyrolyzates of wood can still leave smoldering embers burning behind. Regarding the size of the fire, it would be prudent to confine this technique in addressing only limited areas of critical need (nascent fires, important infrastructure, houses, etc.). Regarding the extinction of smoldering wood embers, a two-step approach is envisioned: application of LN2 to suppress the flames followed by sequential application of water to quench the smoldering wood embers. Overcoming the energy transfer as the cryogen approaches the fire is critical. The temperature of a forest fire is considerably high, rendering both convection and radiation as significant modes of energy transfer to the cryogen. Forest fires are, by nature, huge and difficult to extinguish. This means that any potential solution has to be implemented on a large scale and, more realistically, to a targeted area. This suggests the solution should be economical enough to be implemented on a large scale, scalable from a production and manufacturing perspective, and environmentally benign. In order for the proposed method to be acceptable, it would have to outperform or at least complement conventional fire suppression methods in all of the major categories stated above. Extensive research did not yield a single solution, either using cryogen or with conventional methods, which could target all three phases of the forest fire. Furthermore, no cryogen

2 solutions found were economically feasible, as delivery costs were significantly higher than the proposed solution, with a significantly lower coverage. The capsule solution proposed in this paper was developed by using an analysis-based approach validated by testing. Initially, the problem was approached from a thermal perspective, and a mathematical model was created of the thermal effects of a forest fire. This was based heavily on previous research [7], [8], [10]. This model was then expanded to include a droplet of liquid nitrogen descending to the fire below. Further expansion lead to a model that was based on the capsule approach presented herein. Testing was simultaneously conducted to validate this thermal model with steady state nitrogen vaporization. Additional testing was conducted to determine coefficients of drag and lift. Fluid ejection testing was conducted to validate the dispersion mechanism and the calculations for fluid coverage. Finally, the capsules were dropped to see all aspects of the design at work. Throughout this process, capsule parameters were changed and the capsule evolved from a purely theoretical design to a complex part with moldable geometry fit for mass production.

due to gravity in kg/m3, and P is the height for which the temperature is being calculated in meters. This correlation in Eq. 1 approaches infinity as P approaches zero; therefore the maximum temperature of the air was set to equal the temperature of the fire. The temperature distribution has been plotted in Fig. 1. It was hypothesized that these elevated temperatures, even high above the fire, would impact vaporization during transport. Testing later validated this hypothesis. In response, a conceptual design of an insulated and automated capsule release device was developed.

II. Â THERMAL CONSIDERATIONS The first part of this problem was ensuring that LN2 would reach the forest fire without vaporizing. While gaseous nitrogen can extinguish a fire, delivering nitrogen in a liquid state extinguishes more fire per unit of nitrogen. In theory, vaporized nitrogen would continue to descend due to its lower temperature and density, but the intense updrafts known to occur in forest fires, as well as potential wind conditions, make delivery of gaseous nitrogen undesirable. The nitrogen must also be transported to the drop location, and during transportation vaporization will occur.

A. Â Anticipated Forest Fire Temperatures The minimum safe distance from a forest fire was used as the minimum drop height. This was based on a review paper of several models used to calculate the safe distance based on radiation. A distance of four times the flame height was considered safe by all models in all conditions [7]. Eq. 1 below was used to calculate the temperature profile above a fire. This was developed in an analysis of the temperature distribution of the air above a burning house, and slightly modified for this application [8]. đ?&#x2018;&#x2021;"#$ = đ?&#x2018;&#x2021;& +

7 ( , Ă&#x2014;., Ă&#x2014;01 )*+ / , 23*4) Ă&#x2014;5, Ă&#x2014;6

Ă&#x2014;

8.::; =

(1)

Where Af is the approximate area of the forest fire in m2, Tair is the temperature of the air at that height in Â°K, Tâ&#x2C6;&#x17E; is the temperature of the air far away in Â°K, qrad is the energy flux due to radiation in W/m2, Cpair is the specific heat of air in J/kg Â°K, đ?&#x153;&#x152; is the density of the air in kg/m3, g is the acceleration

Fig. 1. Calculated temperature distribution at a given distance above a forest fire. Notice that the temperature at the capsule release point, 200 m above the fire, is still approximately 322 Â°K (120 Â°F).

B. Â Anticipated Forest Fire Radiative Heating Using a previous study on the radiative heat flux emitted by various types of fires as a starting point, the variation of the heat flux over a distance was found [9]. The magnitude of the radiation was determined based on research into radiative heating due to large-scale forest fires [10]. It has been reported that the maximum radiative energy flux given off by a forest fire was 300 MW/m2 [10]. This energy flux was measured between the forest fire and a measuring device initially at room temperature [10]. Since radiative energy flux is known to vary according to the difference in temperature to the fourth power, the difference in temperature between the measurement probe used in the study and the LN2 used in that application was accounted for [9]. The resulting expression for radiative heat flux at a given distance from the fire is given in Eq. 3. Notice that the variation of heat flux with respect to distance is approximately halfway between the inverse r2 law and the temperature variation found by S. Yokoi in Eq. 1, which reinforces the plausibility of the result obtained herein [8], [9]. To calculate the energy transferred by radiation the view factor for an infinite plane on a spherical object (1/2) was used [9]. đ?&#x2018;&#x17E;$"@ =

ABBĂ&#x2014;:B7 <C.DE

Ă&#x2014;

0/F GHHF 0/F GABBF

(3)

3 Where qrad is the energy flux due to radiation in W/m2, Tf is the temperature of the forest fire in Â°K, and P is the height for which the energy flux due to radiation is being calculated in meters.

C. Â Calculation of Liquid Nitrogen Droplet Vaporization This first simulation was run to determine how a droplet of LN2 would behave as it descends into a forest fire. This information is important in the determination of an effective delivery method. This simulation uses energy balance to calculate the mass of LN2, which is vaporized as a droplet falls towards a forest fire. The basic energy balance is shown in Eq. 4. (đ?&#x2018;&#x17E;$"@ + đ?&#x2018;&#x17E;KLMN )Ă&#x2014;â&#x2C6;&#x2020;đ?&#x2018;Ą = â&#x2C6;&#x2020;đ?&#x2018;&#x161;Ă&#x2014;đ??ťT6

(4)

Where qrad is the energy flux due to radiation in W/m2, qconv is the energy transferred due to convection in W, A is the surface area of the liquid droplet in m2, â&#x2C6;&#x2020;đ?&#x2018;Ą is the time step duration in s, â&#x2C6;&#x2020;đ?&#x2018;&#x161; is the vaporized mass in kg, and Hfg is the heat of vaporization for N2 in Joules. Due to the complex nonlinear relationship between these variables, this equation was not solved. Instead, a time based numerical solution was used to find the amount of nitrogen vaporized during each time step. The position of a free-falling droplet is changing as a function of time due to gravity. The position is calculated from the velocity at the beginning of each time step, which is in turn calculated using its acceleration. The acceleration of the droplet at any instant can be calculated through Newtonâ&#x20AC;&#x2122;s second law, with gravity and drag considered. The drag force is calculated by using the drag coefficient correlations for a spherical object. As can be seen from Fig. 2, a droplet would need to have an unrealistically large diameter to reach the fire with a reasonable efficiency. When a bucket of water is poured out, the water separates into smaller droplets during its descent. This was not taken into consideration in this model, due to its complexity, but is not significant because even a large initial volume would be inefficient. It was determined that convection was the dominant form of heat transfer.

Fig. 2. Calculated thermal energy removed from the fire with different droplet starting diameters. Note that individual droplets would need to have a minimum diameter on the order of 2 cm, with a diameter closer to 10 cm being ideal. This is highly unrealistic, and prompted the development of a capsule based solution. The jump observed around 0.06 m diameter corresponds to a discontinuity in the drag correlations for a sphere near the Reynolds numbers observed.

D. Â Steady State Liquid Nitrogen Vaporization Testing Steady state LN2 vaporization testing was conducted to validate the model created to simulate the heat transfer effects through a capsule of various geometries and materials. This involved a series of experiments that included filling various cylindrical vessels and capsules of different volumes and materials with LN2. Materials, geometry, and volumes were selected so as to get multiple data points in all tested rows and columns of the testing matrix. Statistical analysis was used to extrapolate the effect of the relevant parameters. These parameters included thermal conductivity, wall thickness, surface area, volume, and relevant free convection coefficient. Rate of vaporization and overall time of vaporization was measured and compared to the predicted times. This allowed for verification of the heat transfer equations used in the model, and provided insight into desired material and geometric properties. The containers were tested after the temperature distribution within the container wall had reached steady state.

4 While this finite element solution is different than the lumped capacitance method, the assumption for the lumped capacitance method is that the temperature within the wall is essentially constant. For sufficiently small elements this assumption is still valid. The initial temperatures for the elements in the capsule are calculated using steady state equations. It is assumed that the capsule is filled and then, after an amount of time, the capsule is topped off to replace the nitrogen, which is vaporized cooling down the capsule.

Fig. 3. A representative time series for cryogen vaporization at steady state conditions. This particular test corresponded to an HDPE cylinder. All model results were conservative like this figure shows. The major source of error in this model is the boiling equations, which had known errors of +/- 100% [11].

A representative cryogen vaporization time series is shown in Fig. 3. This is the predicted versus experimental data for a HDPE Cylinder filled with LN2. It can be seen that the model was found to be conservative. This was the case for every test run.

E. Â Falling Capsule Model

The final heat transfer region is from the capsule walls into the LN2 cavity. This is governed by boiling correlations based mostly on convection. The amount of heat transferred to the nitrogen via boiling then results in the mass of nitrogen lost. For each time step this mass is calculated, and the simulation ends when the capsule either runs out of nitrogen, or reaches the ground. This model was run for a series of spherical capsules made of PTFE. It was found that the capsule solution is much more efficient. Fig. 4 shows the percent of the LN2, which reaches the forest fire for given dimensions of a capsule. The tradeoff of wall thickness and amount of remaining cryogen in the capsule undergoing a 200-meter drop can be noticed.

Upon extensive deliberations, it was decided that the most viable method for delivering the cryogen to the forest fire was a capsule. This resulted in a new model being developed, which calculated the heat transferred in three different regions: heat transferred to the capsule, heat transfer within the capsule, and heat transfer between the capsule and the LN2. Heat transfer in the first region was dominated by both convection and radiation, in the second region conduction was the only mode, and the in third region phase change (boiling) of the LN2 was dominant. As the capsule design had an open or vented top, no pressure buildup was considered. The first region, i.e., heat transfer into the capsule, was similar to the droplet model. The main differences being that the capsule outer diameter is constant, some heat transferred into the capsule is used to heat the capsule itself, thus, the temperature of the capsule was not assumed constant. The second region, heat transfer within the capsule, cannot be analyzed using traditional transient conduction. The difficulty is that this would have yielded a nonlinear result for temperature versus time [9]. Since the heat transfer equations used for boiling are highly temperature dependent, and nonlinear, this equation becomes unsolvable. The solution to this was to write a computer program, which divides the capsule into a series of elements, and calculates the heat transfer in each element using energy balance. This finite element method is more accurate as the number of elements increases and the time step decreases. However, due to the duration of the fall and the number of elements, this solution increases the computation time exponentially. In order to balance this, the elements are sized to be the largest possible elements for which the lumped capacitance method is valid.

Fig. 4. Percent of original dropped volume remaining. The percent remaining reaches a peak value due to the combination of two factors. The first is that as the wall thickness decreases, any nitrogen vaporized corresponds to a smaller percentage of the nitrogen. The second is that as the walls become too thin they are no longer capable of insulating the nitrogen.

III. Â CAPSULE SPIN At the desired forest fire location, the capsules will be released and will begin descending towards the fire. As a capsule starts falling, air will move across the fins on the outside of the capsule, which will impart a spinning force on the capsule. The inner wall of the capsule will exert a centripetal force on the nitrogen inside the capsule. This contact force, like all

5 elastic contact forces, is perpendicular to the surface of the capsule. Since the capsule geometry is semi-ellipsoidal, this force has a vertical component at all points on the inner surface. At low rotational velocities this force is balanced out by the gravitational force on the cryogen, but as the capsule starts spinning faster, the cryogen will start moving upward and will exit the capsule.

transfer in the form of convection into the nitrogen within the capsule, and mass of nitrogen removed. The equations of motion of the capsule are the following:

Since a certain rotational velocity is required for the nitrogen to leave the capsule, and the rotational velocity is a function of the lift force on the outside of the capsule, the capsule can be tuned to release the nitrogen at specific heights. This allows the capsule to target all three forest fire zones (crown, surface, and ground), while still insulating the nitrogen for the descent.

Where J is the mass moment of inertia of the capsule in kg m2, D is the effective distance in m, đ??šY#TZ is the lift on the capsule fins from the vertical velocity in N, đ??š\$"6, Â Â Â b#M is the drag on the fins due to the vertical velocity is N, đ??š\$"6, Â Â Â 2"_^eYf is the drag on the capsule due to the vertical velocity in N, đ??š\$"6, Â Â Â ^_#M is the drag on the fins due to the rotational velocity in N, m is mass in kg, g is the gravitational acceleration in m/s, and đ??šY#TZ, Â Â Â ^_#M is the lift on the capsule fins due to the rotational velocity in N.

A. Â Capsule Descent Modeling To perform an analysis of the drag force versus the lift of the body, the fins were approximated to be inclined at an angle tangent to the midpoint of the curved fin. The free body diagram in Fig. 5 shows the geometry for a falling capsule including the forces acting on it and the direction of those forces. The CD and CL values used were for a flat plate at a specific aspect ratio and angle of attack [12]. The velocity used to find the spin lift and drag components was approximated as the tangential velocity at the furthest distance from the fin tip to the central axis, in order to prevent integration, while ensuring a conservative estimate.

đ??˝đ?&#x153;&#x192; = đ??ˇ đ??šY#TZ â&#x2C6;&#x2019; đ??š\$"6, Â Â Â ^_#M

(5)

đ?&#x2018;&#x17D; = đ?&#x2018;&#x201D; â&#x2C6;&#x2019; đ??š\$"6, Â Â Â b#M + đ??šc#TZ, Â Â Â d_#M â&#x2C6;&#x2019; đ??š\$"6, Â Â Â 2"_^eYf

(6)

B. Â Wind Tunnel Testing The paper used for lift and drag correlations had a wide range of values [12]. Since this was the governing parameter in nitrogen release, the team conducted wind tunnel testing to determine actual values. Airspeed measurements were taken on a Pitot tube, and a Go-Pro was used to record the position of the capsule. After testing, the video recordings from the Go-Pro were used to calculate the angular speed for the three different wind speeds that were used. The data can be seen in the graph in Fig. 6 where the predicted values were close to the experimental data for capsules S-1 and S-2, but capsule S-3 was shifted about 2 m/s. This is acceptable because the wind speeds that the capsule would see would not be constant and a variation of this magnitude is small enough where it would not significantly affect the rotation.

Fig. 5. Free body diagram of the fins on an early capsule design. The fins are the protrusions on the outside of the capsule. The capsule is hollow, with the liquid nitrogen residing within. From this diagram it can be seen that the drag on the fins reduces the capsuleâ&#x20AC;&#x2122;s velocity towards the ground, the lift on the fins causes the rotation of the capsule, the drag due to spinning opposes the rotation of the capsule, and the lift due to spinning propels the capsule downwards.

For the purpose of torque calculations, the distance where the spin drag and lift forces were applied at was two thirds of the way up the capsule, similar to the center of area of a triangle. The results from this model were later confirmed in wind tunnel testing. These calculations were then integrated into a thermal finite element code. This resulted in a final model which calculated a list of variables including; position, linear and angular acceleration, linear and angular velocity, heat transfer in the form of convection and conduction into the capsule, conduction between elements within the capsule wall, heat

Rotations Per Second

Wind Tunnel Testing 20.0

15.0

10.0

5.0

Predicted S1

0.0 0

5 10 15 Wind Speed (m/s)

Predicted S2 Predicted S3

Fig. 6. Wind tunnel testing data for three experimental capsules. These capsules were an early design that was scaled down for testing purposes.

Based on these results final values for the lift and drag coefficient were chosen. Values of 0.9 for CD and 1.5 for CL were used. These are within the ranges provided in the paper.

C. Fluid Ejection Testing Fluid ejection testing was performed to determine the minimum angular velocity that the LN2 would be ejected from the fin channels and the dispersion radius of the capsule. Water was used for this testing because it is easier to handle and has a similar density and viscosity to LN2. Video recording was used to ensure accuracy. The minimum release velocity was found to be 2.65 revolutions per second or 159 revolutions per minute. Fig. 7 shows the actual dispersion of the water from the capsule. The dispersion radius was found to be approximately 40 inches for this design.

Fig. 8. Plot showing how capsule scale up affects the fluid release height. Plots of this nature were generated for multiple different capsule designs to determine which would perform best, and to inform design decisions. Variables like fin size can be varied to shift these curves.

IV. CAPSULE DROP TESTING

Fig. 7. Spin testing with an early capsule design. This design had small channels for fluid dispersion, which were later modified to allow the more dramatic release seen in subsequent drop testing of the capsule.

D. Fluid Ejection Modeling Due to the complex nature of the fluid dynamics inside the capsule, it was much easier to perform the testing previously outlined than to create a CFD model for every capsule geometry required. As such, the critical angular velocity measured in testing was used as the angular release velocity, and then the area covered by the capsule as it fell was calculated using ballistic motion. The angle of the helical flow channels will be the initial release angle of the fluid, due to contact forces. The tangential velocity can be calculated using the rotational velocity and the distance at the tube exit, and then using the height and gravity, the radius of dispersion can be calculated. This was integrated into the model so that when a set of parameters were run, the fluid dispersion area could be calculated. This was particularly important for ensuring the capsules could be scaled up to reasonable sizes and quantities. The results of this model can be seen in Fig. 8.

The capsule drop testing that the team conducted was done to determine how well the capsule spins and disperses LN2 in a real life situation. After determining that the capsule would spin from the wind test and that the fluid would eject from the stationary spin test, the drop test was the last step needed to conclude that the design was promising. The drop testing was conducted from the rooftop of a local parking garage. This gave a drop height of approximately 17 meters, much lower than the 200 meters the capsules were designed for. However, it was the highest point available at the time of testing. Testing was performed at the presence of a university police detail.

A. Procedure The setup for the drop test consisted of a support structure to extend the capsule over the edge of the parking garage and control the drop of the capsule. In order to compensate for the reduced fall time, the fixture had the capability to spin the capsules prior to releasing them. This better simulated the angular speed of the capsule if it had been dropped from the proper height. Several capsules were dropped into large cardboard boxes filled with packaging to prevent them from breaking. The capsules tested can be seen in Fig. 9.

V. COST MODELING One of the final steps to validate this method of firefighting was to perform a cost analysis in conjunction with the heat transfer modeling. The results of this model showed that utilizing a two-phase process of dropping the LN2 capsules to suppress the fire followed by a load of water to extinguish the fire would cost approximately $9,869 and could extinguish 2 acres in about thirty minutes. This is compared to the standard method of dropping water from a helicopter, which would cost approximately $7,910 and take four hours and thirty minutes to extinguish the same area. The material used for this analysis was SMMA copolymer due to the favorable mechanical and thermal characteristics, as well as very low cost.

A. Material Cost

Fig. 9. Capsuled drop tested. On the left is an old version of the capsule design, which had internal fluid channels and could only be 3-D printed. On the right were two capsules of a newer, moldable design. One panel of the capsules was painted red in order to better visualize the rotation during the drop testing.

B. Results Without being spun on the test stand prior to release, the capsules started spinning slightly before impact, with minimal water dispersion. This was expected based on the low drop height. The maximum radius for the capsule dropped with prespin was approximately eight feet. Overall, the drop testing proved that the capsules were able to spin and disperse water as they were falling through the air. This test was able to prove the effectiveness of the design of the capsule for both spin capacity and fluid dispersion. An image of the capsule dispersing water during the pre-spin period and during the drop of the capsule can be seen in Fig. 10.

Fig. 10. Capsule releasing died water. On the left the capsule can be shown early in the descent, just as the liquid begins to release the water. The dispersion is more visible on the right.

This test was a culmination of all the testing and design that had been done on the capsule and proved that the design works and can be effective.

The maximum weight that could be carried by a representative helicopter was found through research. Using material costs for the capsule and the nitrogen, and a rough calculation of the injection molding manufacturing cost, the team was able to determine the material cost of a helicopter load. Then, the thermal analysis was implemented to determine the efficiency of the LN2 during transport and drop to assess how much LN2 was delivered to the fire. An impact analysis of the falling capsule was performed in the software package MATLAB, which showed that the gas would rise to a height that is about one third of its diameter as it expands. From this, the area of fire suppressed from a single capsule was determined. When combined with the cost of the liquid nitrogen and capsule, this calculation was used to determine the cost of materials per acre extinguished: $4,614. This value proved to be reasonable enough to make the LN2 suppression method an economically viable option to replace helicopter loads of water.

B. Time and Operational Cost In order to find the costs of fighting with water, research found from the Camping and RVing British Columbia Coalition dictated the quantity of water required to extinguish a fixed fire size [13]. After LN2 proved to cover more area than water, the authors contacted CalFire Aviation and communicated with the Chief of Aviation, Bill Payne. Chief Payne is an expert in managing aerial vehicles for the purpose of fighting large forest fires. He offered valuable information regarding the operational costs of these pieces of equipment. The scenario of dropping a single load of LN2 filled capsules, followed with a load of water was compared to current methods. It was determined that it would take more than 26 helicopter loads of water to cover that same area as one load of LN2 followed by a load of water. The results of these calculations showed that a two-phase LN2 implementation plan would cost an extra $1,960, but save four hours. This translates to a significant savings in avoided damages making it a preferred firefighting method when a fire is moving at a fast rate and may create thousands of dollars of damages in a small amount of time.

VI. CONCLUSIONS The problem of suppressing forest fires with LN2 is extremely unique and challenging. This topic required significant research into the properties of both forest fires and LN2. Based on this research, it is hypothesized that a two-stage process, using water as a second agent, is needed to fully extinguish the fire beyond re-ignition. The proposed solution is to carry LN2 in a high capacity capsule. It needs to be effectively insulated to maintain nitrogen in its fully liquid state, and to incorporate a series of spiraled fins with internal groves for fluid rise flow. The spiraled fins can induce a spin on the capsule as the airspeed of the capsule increases, causing the LN2 in the capsule to move outwards and disperse to the fire. In order to understand the vaporization process, mathematical models were developed for the nitrogen’s descent to the fire below. A model was then developed for the proposed capsule solution. From this model it was concluded that less nitrogen would vaporize during the descent of a properly designed capsule than in the lengthier transportation to the fire. A steady state vaporization model was developed to evaluate this conclusion, and testing was done to validate the model. This model allowed optimization of the geometry for insulation under a certain temperature given a capsule of a certain material. An analytical model was constructed to predict the expected angular speeds of the capsule for different drop heights. This was verified with a wind tunnel test that simulated expected wind speeds encountered at different heights. This was incorporated into the thermal model to account for second order effects on the capsules position, velocity, and convection values as the capsules fell.

Further research still needs to be conducted to develop and design the release mechanism required to optimally release the capsules on critical areas of the fire. This would require sufficient cooling or insulation to protect the nitrogen on its journey, while keeping the pilots concentration of the operation of the vehicle as opposed to the payload. This work can serve as a foundation for the development of this solution.

ACKNOWLEDGMENT Financial support for this project was provided by Northeastern University’s Department of Mechanical Engineering Capstone Program. An acknowledgement also goes to Professor Bridget Smyser, Kevin McCue and Jack Price for their assistance in setting up experiments safely, and providing recommendations on the best way to run the experimentation and take data. Professor Taslim and Professor Jalili are acknowledged for their assistance in aerodynamics and sealing mechanisms respectively. Finally, Cal Fire Aviation Chief Bill Payne is acknowledged for sharing specific data regarding costs and methods of extinguishing forest fires. Craig W. Martland is a senior mechanical engineering student at Northeastern University pursuing a dual Bachelor’s and Master’s degree with a concentration in mechanics and graduating in May 2016. He is interested in continuing to design and develop innovative solutions to problems. cmartland@comcast.net David P. Marchessault is a senior mechanical engineering student at Northeastern University pursuing a Bachelor’s degree and graduating in May 2016. He is interested in product design and alternative energy technologies and hopes to pursue a career that enables these interests. Dpm121314@yahoo.com

Then, the area covered by the nitrogen as it left the capsule had to be determined. This was done experimentally, by spinning a 1.9-liter capsule and measuring the spread of water. The area coverage was determined to have increased from 5 m2 merely from impact, to a conservative 8.7 m2 from preliminary fluid ejection calculations.

Andrew McGarey is a senior mechanical engineering student at Northeastern University graduating in May 2016. He is interested in design and manufacturing and will be working at Sikorsky Aircraft Corporation doing tooling and manufacturing design after graduation. AndrewMcGarey@gmail.com

The capsule was then dropped from the rooftop of a university building to evaluate the capsule in action. The capsule rotated and the coverage area from the drop testing was found to be approximately 23 m2.

Diego Rivas is a senior mechanical engineering student at Northeastern University. He will be pursuing his Master’s degree at UC Berkeley in product design, with an interest in developing innovative technologies that solve problems with a significant social impact. diegorivasco@gmail.com

Finally, an economic analysis was conducted to determine if this solution was cost effective. Through research and collaboration with Cal Fire Aviation, the costs of current methods and the proposed solution were estimated. This showed that a helicopter filled with 400 LN2 capsules (25 liters each) could extinguish two acres of fire in approximately thirty minutes, while existing methods using water would require four hours and thirty minutes to extinguish the same area. Using LN2 would cost $9,870 where a method of fighting with just water would cost $7,910.

Kevin W. Stanley is a senior mechanical engineering student at Northeastern University graduating in May 2016. He hopes to follow his interests in product development and entrepreneurship throughout his career. KevinWilliamStanley@gmail.com Yiannis A. Levendis is a Distinguished Professor of Mechanical and Environmental Engineering. He specializes in Energy Harvesting, Combustion and Air Pollution. He

9 developed the technique of pool fire extinction with direct application of liquid nitrogen. Y.Levendis@northeastern.edu

REFERENCES [1] Y. Levendis and M. Delichatsios, “Cryogenic Supression of Liquid Pool Fires and Wooden Crib Fires”, in National Fire Protection Association Suppression and Detection, Orlando, FL, 2011. [2] K. Miyazaki, S. Inoue, and H. Horiike, 'Small-scale Experiments of Nitrogen Injection Effects on Sodium Fire Extinguishment', Journal of the Atomic Energy Society of Japan / Atomic Energy Society of Japan, vol. 41, no. 10, pp. 1084-1091, 1999. [3] U.S. Administration, 'U.S. fire statistics', Usfa.fema.gov, 2015. [Online]. Available: http://www.usfa.fema.gov/data/statistics/#tab-4. [Accessed: 16- Aug- 2015] [4] Nifc.gov, 'National Interagency Fire Center', 2015. [Online]. Available: http://www.nifc.gov/fireInfo/fireInfo_statistics.html. [Accessed: 16- Aug- 2015] [5] Coloradofirecamp.com, 'Colorado Firecamp, Wildland Fire Suppression Tactics Reference Guide', 2015. [Online]. Available: http://www.coloradofirecamp.com/suppression-tactics. [Accessed: 16- Aug- 2015] [6] Y. Levendis, A. Ergut and M. Delichatsios, 'Cryogenic extinguishment of liquid pool fires', Process Safety Progress, p. NA-NA, 2009. [7] B. Butler, Wildland Firefighter Saftey Zones: A Review of Past Science and Summary of Future Needs, 1st ed. Missoula, MT: US Forest Service, 2015 [Online]. Available: https://www.firescience.gov/projects/07-2-120/project/07-2-120_http_wwwpublishcsiroau_actview_filefile_idWF1302 1.pdf. [Accessed: 16- Aug- 2015] [8] W. Bert, International Symposium on the use of Models in Fire Research. Washington DC: National Academy of Sciences - National Research Council, 1961, pp. 198-204. [9] Ortiz, Xavier, David Rival, and David Wood. "Forces and Moments on Flat Plates of Small Aspect Ratio with Application to PV Wind Loads and Small Wind Turbine Blades." Energies (2015): 2438-453. [10] B. Gabbert, 'At what temperature does a forest fire burn?', Wildfiretoday.com, 2011. [Online]. Available: http://wildfiretoday.com/2011/02/26/at-what-temperaturedoes-a-forest-fire-burn/. [Accessed: 16- Aug- 2015] [11] C. Bergman, Intro to Heat Transfer. Wiley C, 2014. [12] K. Taira, W. B. Dickson, T. Colonius, M. H. Dickinson, and C. W. Rowley, Unsteadiness in Flow over a Flat Plate at Angle-of-Attack at Low Reynolds Numbers', 45th AIAA Aerospace Sciences Meeting, 2007. [13] "The Proven World Standard in Aerial Firefighting." Bambi Bucket – Precision Helicopter Fire Fighting Tool. SEI Industries, 2015. Web. 29 Nov. 2015. [14] Borealforest.org, 'Science and Innovation - Forest Fires', 2015. [Online]. Available: http://www.borealforest.org/world/innova/forest_fire.htm. [Accessed: 16- Aug- 2015]

[15] Airproducts.com, 'Safetygrams', 2015. [Online]. Available: http://www.airproducts.com/Company/Sustainability/envi ronment-health-and-safety/product-safetysafetygrams.aspx. [Accessed: 16- Aug- 2015] [16] R. Ghosh, Cryogenic Nitrogen Gas Cooling for Thermal Spray Coatings, 1st ed. Air Products and Chemicals, Inc, 2015 [Online]. Available: https://www.airproducts.com/~/media/Files/PDF/industrie s/metals-cryogenic-nitrogen-gas-cooling-thermal-spraycoatings.pdf. [Accessed: 16- Aug- 2015] [17] Liquid Nitrogen Cryosurgery, 1st ed. Hadsund, Denmark: Cortex Technology, 2015 [Online]. Available: http://www.cortex.dk/files/manager/documents/infoflyer.pdf?viewer=true. [Accessed: 16- Aug- 2015] [18] Needhamag.com, 'Stream Bars for Uniform Liquid Fertilizer Application - Needham Ag Technologies, LLC', 2015. [Online]. Available: http://www.needhamag.com/innovative_product_sales/str eam_bars_for_uniform_liquid_fertilizer_application.php. [Accessed: 16- Aug- 2015] [19] Ehs.research.uiowa.edu, 'Liquid Nitrogen Handling | Environmental Health and Safety', 2015. [Online]. Available: http://ehs.research.uiowa.edu/liquid-nitrogenhandling. [Accessed: 16- Aug- 2015] [20] Ehs.neu.edu, 'Cryogenic Liquids', 2015. [Online]. Available: http://www.ehs.neu.edu/laboratory_safety/fact_sheets/cry ogenic_liquids/. [Accessed: 16- Aug- 2015] [21] D. An, P. Sunderland and D. Lathrop, 'Suppression of sodium fires with liquid nitrogen', Fire Safety Journal, vol. 58, pp. 204-207, 2013. [22] Z. Fu-bao, S. Bo-bo, C. Jian-wei and M. Ling-jun, 'A New Approach to Control a Serious Mine Fire with Using Liquid Nitrogen as Extinguishing Media', Fire Technol, vol. 51, no. 2, pp. 325-334, 2013. [23] M. Paczkowski and A. Kukuczka, 'Use of Liquid Nitrogen for Extinguishing Fires in Mines', Przegl Gorn, pp. 375-380, 1978. [24] H. McBay, “Fire Extinguishing Capsule”, U.S. Patent no. 5590717, January, 1997. [25] P. Edwards and G. Ruebusch, “Fire Retardant Delivery System”, U.S. Patent no.7083000, August, 2006. [26] B. Ivarez and L. Maria, “Method for Acting on Forest Fires, Pests or Atmospheric Phenomena from the Air”, U.S. Patent no.7690438, April, 2010. [27] F. Cicanese, “Oil Well Fire Suppression Device”, U.S. Patent no. 666278, December, 2003. [28] W. Reed, “Heat-Absorbing Gel material”, U.S. Patent no. 6776920, August 2004. [29] V. Sridharan and R. Vairavan, “Fire Extinguishing by Explosive pulverisation of Projectile Based Frozen Gases and Compacted Solid Extinguishing Agents”, U.S. Patent no. 7478680, Jan. 2009. [30] A. Ozment, “Method for Fighting Fire in Confined Areas Using Nitrogen Expanded Foam”, U.S. Patent no. 7104336, September, 2006. [31] F. White, Fluid mechanics. New York, N.Y.: McGraw Hill, 2011.

10 [32] Sadraey, M., and Dr. Müller. "Drag Force and Drag Coefficients." Web. 28 Sept. 2015. [33] R. Barron, Cryogenic Heat Transfer. CRC Press, May 1, 1999. [34] "The Original Bambi Bucket." SEI Industries 22 Aug. 2013. Web. 29 Nov. 2015. [35] Arney, Donald Brian, Peter Leighton Brooke, and Norman Carter Wagner. "Patent US5829809 - Multidump Fire Fighting Bucket." Google Books. USPTO, 3 Nov. 1998. Web. 29 Nov. 2015. [36] "Bambi Bucket Quick Repair Guide." SEI Industries, June 2014. Web. 29 Nov. 2015. [37] "Campfires in British Columbia | How To Camp." The Camping and RVing British Columbia Coalition. Camping and RVing BC Coalition, 2015. Web. 27 Nov. 2015. [38] Cost Estimator. Computer software. Http://www.custompartnet.com/estimate/injectionmolding/. CustomPartNet, 2009. Web. 11 Nov. 2015. [39] "Aerial Firefighting." Wikipedia. Wikimedia Foundation, 21 Nov. 2015. Web. 27 Nov. 2015. [40] Fan, Karen. "Price of Liquid Nitrogen." Price of Liquid Nitrogen. The Physics Factbook, 2007. Web. 27 Nov. 2015.

Maximum Likelihood Image Reconstruction using Data Fusion between X-Ray and Microwave Radar Matthew T. Tivnan, Carey M. Rappaport

Abstract— Data fusion is the process by which measurements collected by two or more sensors are combined to produce a better result than could have been produced by any of the sensors acting individually. X-ray transmission and Microwave Tomography (MWT) are good candidates for data fusion because of their complementary strengths. For example, X-Ray is known for high spatial resolution structural imaging and MWT provides higher contrast in the physical properties for certain applications. In this work, a simple image reconstruction algorithm is presented which utilizes data fusion between X-Ray and MWT measurements. One possible application in neuroimaging is then simulated in a numerical experiment. The final results show that data fusion has significant advantages over conventional approaches.

I. INTRODUCTION Imaging and reconstruction from indirect measurements is an important issue in science and engineering research with applications in non-destructive testing of materials, airport security, and medical imaging among others. Great technological strides have been made for individual imaging modalities; however, improvements are still possible through data fusion. The goal of the fused approach is to reconstruct an object using data gathered through two or more different physical processes. Data fusion is considered successful if the fused image provides more information than the images generated by the contributing modalities acting individually. Moving forward, there would be no requirement that novel imaging modalities are superior to existing methods; through data fusion, any new information can contribute to an improved image. Small changes in the quality of an image can often contain dramatically important new information. In this work, data fusion using X-ray Computed Tomography (CT) and Microwave Tomography (MWT) measurements for image reconstruction will be considered. X-ray is one of the earliest forms of imaging and remains to this day as the socalled “gold-standard”, especially when it comes to interfaces between two different types of tissue. The absorption of Xrays as they pass through biological tissue can be accurately modeled as a linear process, which makes image processing simple. CT involves the rotation of sensors around the object under test and the collection of X-ray projections at each of the angles, known as a view, to form a complete dataset known as a sinogram. The structure of the image is related to

the sinogram through a mathematical process known as the Radon transform. It can therefore be extracted using the associated inverse process which is known as filtered backprojection. However, X-ray imaging has certain drawbacks, including the relatively low radiological contrast between different types of tissues. Two completely different types of tissues which happen to have the same density can often be indistinguishable. In addition, CT scans deposit harmful radiation. Exposure to radiation can be reduced by decreasing the number of CT views or by reducing the intensity per view – both of which decrease the quality of the final image. A high-quality image may be possible using low-radiation X-ray measurements through data fusion with another modality. Microwave Tomography is an emerging medical imaging modality in which the tissue is probed by a microwave electromagnetic wave. The electromagnetic field which is scattered by the tissue is then measured and used to image the hidden structure. In certain applications, the contrast in the dielectric properties of different materials is higher than the radiological contrast of X-ray, and the radiation is generally considered benign because it is non-ionizing. However, MWT has poor spatial resolution; objects that are small relative to the wavelength of the interrogating electromagnetic wave are often difficult or impossible to resolve. Additionally, the inverse problem, which is to reconstruct the object given the measured scattered electromagnetic field, is known to be illposed. To address this issue, there is a science and an art to the optimal arrangement of antennas and the constraint of possible solutions; however, imaging and reconstruction using MWT is usually very difficult. Data fusion between these two modalities aims to combine the high resolution of X-ray and the high contrast of MWT in a single reconstruction algorithm. In certain applications, fusion can lead to a higher tolerance to noise. In other cases, it can actually extract information which was previously unavailable to either of the sensors acting individually. For these reasons, there has recently been much attention paid to hybrid CTMWT imaging [1-2]. In the opening sections of this work, the physical and mathematical justifications are derived for a maximum likelihood data fusion algorithm. The algorithm is then described in full for a general case such that it can be applied to many situations and configurations. In the final section, a specific example is explored in depth in a numerical experiment. The chosen application is Neuroimaging for detection of carcinoma.

II. Â SENSOR DESCRIPTIONS

B. Â Microwave Tomography

A. Â X-Ray Computed Tomography In order to simulate the measurements collected by a CT scanner, a computational model is designed based on an understanding of X-ray physics. A parallel beam configuration, as shown in Fig. 1 will be considered.

In MWT, electromagnetic waves cannot be accurately modeled as rays passing through the object under test. Instead, one must consider an interdependent network of scattered and re-scattered electromagnetic fields. The amplitude and phase of the total field is measured by the receiving antennas. A typical configuration for MWT is shown in Fig. 2.

Fig. 1. Configuration for parallel beam X-ray CT. X-ray sources and detectors are shown in violet. View angle and ray position are annotated.

Fig. 2. Configuration for MWT. Transmitter shown in gold, receivers shown in blue. Amplitude and phase are annotated in the red break-out box.

As X-rays propagate, their energy is absorbed based on the characteristic properties of the object under test. This attenuation of intensity is described by

A 2D case with the magnetic field oriented in the transverse direction will be considered. In this configuration, the electric field integral equation can be written as

đ??ź đ?&#x153;&#x152;, đ?&#x153;&#x192; = Â đ??ź' đ?&#x2018;&#x2019; )*(,,-)

(1)

đ??¸ đ?&#x2018;Ľ, đ?&#x2018;Ś = đ??¸ ' đ?&#x2018;Ľ, đ?&#x2018;Ś +

đ??ş đ?&#x2018;Ľ, đ?&#x2018;Ś, đ?&#x2018;Ľ B , đ?&#x2018;Ś B đ??˝ đ?&#x2018;Ľ B , đ?&#x2018;Ś B đ?&#x2018;&#x2018;đ?&#x2018;Ľâ&#x20AC;˛đ?&#x2018;&#x2018;đ?&#x2018;Śâ&#x20AC;˛

where đ??ź đ?&#x153;&#x152;, đ?&#x153;&#x192; is the intensity of an X-ray after it passes through the object under test, đ??ź' is the initial intensity, and đ?&#x2018;&#x201D;(đ?&#x153;&#x152;, đ?&#x153;&#x192;) is the distribution of accumulated attenuation which is sometimes referred to as a sinogram, and is given by

where đ??¸ is the total electric field, đ??¸ ' is the incident electric field which depends on the transmitting antennas, and đ??ş is the Greenâ&#x20AC;&#x2122;s function which describes the response for all positions to a small change in the current density đ??˝, which is defined as

đ?&#x2018;&#x201D; đ?&#x153;&#x152;, đ?&#x153;&#x192; = Â

đ?&#x153;&#x2021; đ?&#x2018;Ľ, đ?&#x2018;Ś â&#x2C6;&#x2122; đ?&#x2018;&#x2018;đ?&#x2018;&#x2122; 67,8

(2)

where đ??ż,,- is the linear path followed by the X-ray, and đ?&#x153;&#x2021; is the distribution of mass attenuation coefficients for the object under test. Equation (1) requires that đ?&#x2018;&#x201D; be a unitless quantity, so đ?&#x153;&#x2021; is given in units of cm-1. This transformation from đ?&#x2018;Ľ, đ?&#x2018;Ś space to đ?&#x153;&#x152;, đ?&#x153;&#x192; is called the Radon transform. A discrete version of this integral transformation is easily found by replacing the integral with a summation, and estimating đ?&#x153;&#x2021; đ?&#x2018;Ľ, đ?&#x2018;Ś by interpolation from sampled distribution The measured quantity in X-ray imaging is the intensity, đ??ź. Measurement noise is often normally distributed and referred to as additive Gaussian white noise (AWGN). It is typically identically distributed with respect to intensity. However, the initial transmitted intensity đ??ź' can be adjusted so that pseudomeasurements of đ?&#x2018;&#x201D; have identically distributed AWGN, which is much more convenient for most imaging algorithms. The effective measurement array is therefore given by

đ??˝ đ?&#x2018;Ľ, đ?&#x2018;Ś = đ?&#x153;&#x2019; đ?&#x2018;Ľ, đ?&#x2018;Ś đ??¸(đ?&#x2018;Ľ, đ?&#x2018;Ś)

where đ?&#x153;&#x2019; is the dielectric contrast. The ultimate goal of MWT is to reconstruct đ?&#x153;&#x2019; by measuring đ??¸. Notice that equation (4) defines đ??¸ in terms of đ??˝ and equation (5) defines đ??˝ in terms of đ??¸. This structure is characteristic of a non-linear inverse problem â&#x20AC;&#x201C; that is, đ??¸ cannot be written as a linear transformation of đ?&#x153;&#x2019;, and there is therefore no inverse linear system which would allow for the direct reconstruction of đ?&#x153;&#x2019; from measurements in đ??¸ collected by receiving antennas. However, computational models such as Finite Difference Frequency Domain (FDFD) are able to find an accurate solution for đ??¸ given đ?&#x153;&#x2019; and đ??¸' . The receiving antennas which collect the electric field introduce AWGN. The measurement array can therefore be written as by đ?&#x2018;ŚFG; = đ??¸ + Â đ?&#x2018;§FG;

đ?&#x2018;Ś:; = đ?&#x2018;&#x201D; + Â đ?&#x2018;§:;

(5)

(6)

(3)

where đ?&#x2018;Ś:; contains the CT measurements, and đ?&#x2018;§:; contains independent identically distributed zero-mean AWGN with ? known variance, đ?&#x153;&#x17D;:; . The arrays are equal in length and have one element for each X-ray detector.

where đ?&#x2018;ŚFG; contains the MWT measurements, and đ?&#x2018;§FG; contains independent identically distributed zero-mean ? AWGN with known variance, đ?&#x153;&#x17D;FG; . The arrays are equal in length and have one element for each receiving antenna.

(4)

3 where đ??ś' is simply a normalizing constant so that the sum of the probabilities of all possibilities is equal to 1.

III. Â ALGORITHM DESCRIPTION A. Â Maximum Likelihood Detection in AWGN In general, measurements collected in the presence of AWGN can be described by the following formula: đ?&#x2018;Ś = đ?&#x2018;Ľ + Â đ?&#x2018;§

The next few steps in the derivation require the utilization of special properties of the maximization operator. The relevant identities are listed below:

(7)

max đ?&#x2018;ĽP = max Â đ?&#x203A;źđ?&#x2018;ĽP Â P

where đ?&#x2018;Ś is an array which contains the measurements, đ?&#x2018;Ľ is the signal of interest, and z contains independent identically distributed AWGN with known variance of Ď&#x192;?J . In Maximum Likelihood detection, there must be some finite number đ?&#x2018;&#x20AC; of possible signals. After the measurements are collected, the goal is to simply find the signal xM which is most likely to produce the measured data. The optimization problem can be formally written as max Â đ?&#x2018;&#x192; đ?&#x2018;Ľ = đ?&#x2018;ĽP Â | Â đ?&#x2018;Ś = Â max Â đ?&#x2018;&#x192; (đ?&#x2018;§ = đ?&#x2018;Ś â&#x2C6;&#x2019; đ?&#x2018;ĽP Â ) P

(8)

Here, notation is borrowed from probability theory. For example, đ?&#x2018;&#x192;(đ??´|đ??ľ) is the probability that event đ??´ will occur given that event đ??ľ has occurred, and max Â CM evaluates to the M

max đ?&#x2018;ĽP = max Â ln đ?&#x2018;ĽP P

max đ?&#x2018;ĽP = min â&#x2C6;&#x2019;đ?&#x2018;ĽP

The first identity shows that maximization of a quantity is equivalent to the maximization of a quantity times any constant đ?&#x203A;ź. The second shows that the maximization of a quantity is the same as the maximization of a logarithm of the same quantity, since the logarithm is a monotonically increasing function. Lastly, finding a maximum of a quantity is the same as finding the minimum of its opposite. Using the identity in (13), the constant outside the repeated product in (11) can be eliminated. [

đ?&#x2018;&#x2019; ) de)fg,e

max P

(9)

max P

đ?&#x2018;&#x192; đ?&#x2018;§Z = đ?&#x2018;ŚZ â&#x2C6;&#x2019; đ?&#x2018;ĽP,Z

(10)

Z\]

Here, capital pi notation is used to represent a repeated product. The random variable đ?&#x2018;§Z has a Gaussian distribution with zero-mean and variance Ď&#x192;?J . Substituting the general form for a Gaussian distribution leads to [

max P

đ??ś' đ?&#x2018;&#x2019;

)

c ] de )fg,e ?ac b

(11)

Z\]

max đ??ś' P

Z\] [

ln đ?&#x2018;&#x2019; ) de)fg,e

max Â P

[ ] c ?h i đ?&#x2018;&#x2019;

[

đ?&#x2018;&#x2019; ) de)fg,e Z\]

(17)

(18)

Z\] [

max P

[

đ?&#x2018;&#x2019; ) de)fg,e Â

max Â ln

Z\]

Here, capital â&#x20AC;&#x153;Uâ&#x20AC;? notation is used to indicate a repeated union of events (event D] AND event D? AND â&#x20AC;Ś event D_ ). The noise for any given sensor is independent of the noise for any other sensor, so by the definition of independence, the optimization problem can be rewritten as

(16)

Z\]

[

đ?&#x2018;§Z = đ?&#x2018;ŚZ â&#x2C6;&#x2019; đ?&#x2018;ĽP,Z

Taking the logarithm of the maximization operand as described in (14) and simplifying leads to

[

max đ?&#x2018;&#x192;

(14) (15)

index đ?&#x2018;&#x161; for which CM of the finite set C is maximum. If y, x, and z each have K elements, then the joint event inside the probability operator in equation (8) can be rewritten in terms of a union between events for each individual sensor as follows:

(13)

â&#x2C6;&#x2019; đ?&#x2018;ŚZ â&#x2C6;&#x2019; đ?&#x2018;ĽP,Z Z\]

(19)

Finally, using the identity in (15) gives [

min P

đ?&#x2018;ŚZ â&#x2C6;&#x2019; đ?&#x2018;ĽP,Z

(20)

Z\]

This is the end of the derivation and the final form of the optimization problem. The solution is chosen as the signal đ?&#x2018;ĽP which leads to the lowest sum of squared residuals with respect to the measurements. This process is commonly referred to as minimizing the least-squared cost function.

(12)

[

đ?&#x2019;ĽP =

đ?&#x2018;ŚZ â&#x2C6;&#x2019; đ?&#x2018;ĽP,Z Z\]

(21)

B. Â Data Fusion Algorithm

IV. Â APPLICATION: NEUROIMAGING

Applying Maximum Likelihood detection to CT and MWT is straight-forward. A finite set of đ?&#x2018;&#x20AC; possible solutions is generated, each corresponding to certain distributions of mass attenuation coefficients đ?&#x153;&#x2021;P (đ?&#x2018;Ľ, đ?&#x2018;Ś) and dielectric contrasts đ?&#x153;&#x2019;P (đ?&#x2018;Ľ, đ?&#x2018;Ś). Next, the signal đ?&#x2018;&#x201D;P (đ?&#x153;&#x152;, đ?&#x153;&#x192;) is generated via the Radon transform, and the signal đ??¸P (đ?&#x2018;Ľ, đ?&#x2018;Ś) is generated via FDFD. These hypothesized signals are compared to the measurements đ?&#x2018;Ś:; and đ?&#x2018;ŚFG; using a pair of least-squared cost functions. [op

đ?&#x2019;Ľ:;,P =

đ?&#x2018;Ś:;,Z â&#x2C6;&#x2019; đ?&#x2018;&#x201D;P,Z

[qrp

đ?&#x2018;ŚFG;,Z â&#x2C6;&#x2019; đ??¸P,Z

A simplified two-dimensional lossless configuration is considered based on an axial scan where the region of interest is composed of homogeneous sub-regions made up of either bone, carcinoma, cerebral spinal fluid (CSF), white matter, gray matter, or air. A single view at 0 degrees is used for CT and a single operating frequency of 1GHz is used for MWT.

(22)

Z\]

đ?&#x2019;ĽFG;,P =

In this numerical experiment, a neuroimaging configuration will be considered. A tumor with a known center but unknown radius and material

(23)

Z\]

In this framework, data fusion is simple â&#x20AC;&#x201C; the fused cost function for the algorithm is defined as đ?&#x2019;ĽP = đ?&#x2019;Ľ:;,P + đ?&#x203A;˝đ?&#x2019;ĽFG;,P

(24)

where đ?&#x203A;˝ is a heuristically chosen constant which weights the relative impacts of CT and MWT. The optimal solution is chosen as the one which results in the lowest fused cost. A flowchart representation of the full algorithm is shown below.

A six-month check-up on a previously detected presence of brain carcinoma will be considered. For a certain fictitious clinical subject, a healthy baseline scan was collected 3 years ago, as well as a scan from six months ago which shows a circular brain carcinoma with a radius of 50mm. In the past six months, treatment has led to a decrease in the size of the tumor and it is now 25mm in radius. Measurements are simulated using the Radon transform and FDFD forward models and adding AWGN such that the signal to noise ratio is 30dB for each sensor. It is the task of the data fusion algorithm to reconstruct this tissue type and radius of a circular region which has the same center as the carcinoma in the six month old scan using the CT and MWT measurements.

Fig. 4.Final results from previous scans. These are treated as prior knowledge.

Each material has characteristic mass attenuation coefficients and dielectric constants. Data from [3] are used for mass attenuation coefficients and data from [4] and [5] are used to define dielectric properties. The material properties of each material are shown in the figure below.

Fig. 3. Flow chart for complete maximum likelihood data fusion algorithm

Fig. 5. Relevant mass attenuation coefficients and dielectric constants

The finite set of đ?&#x2018;&#x20AC; possible solutions is generated by inserting a circular region at a given center point into the healthy image from the three year old scan. Radii between 0 and 50 mm,

5 mass attenuation coefficients between 0.07 and 0.12 cm-1 , and dielectric constants between 50 and 80 are considered. This solution set corresponds to a 3 dimensional parameter space. The CT-based cost function will depend on the radius and the mass attenuation coefficient of the circular region, the MWTbased cost function will depend on the radius and dielectric constant, and the fused cost function will depend on all three parameters. For each possible case, the signals đ?&#x2018;&#x201D;P and đ??¸P are computed using the Radon transform and FDFD, respectively. The resulting CT-based and MWT-based cost functions are shown in the contour plot below. Fig. 8. Reconstructed image showing the carcinoma with a radius of 25mm

V. Â CONCLUSION

Fig. 6. CT-based and MWT-based cost function contour plots. The region inside the innermost blue contour corresponds to the most likely cases.

The CT-based cost function is lowest for brain carcinoma and white matter, and the MWT-based cost function is lowest for brain carcinoma and CSF. It is already clear that brain carcinoma is the only overlap between possible solutions yielded by the two sensors. The fused cost function is then evaluated for all possible solutions. The weighting parameter đ?&#x203A;˝ is chosen such that the each sensor type has the same impact on the solution â&#x20AC;&#x201C; that is, the lowest occurrence in each cost function is normalized to the same value. The final results are shown in the figures below

Figure 6 shows that neither modality is able to completely isolate the correct solution acting individually. However, the two methods contain complimentary information which is exploited by the data fusion algorithm. In the end, the correct image is reconstructed as shown in Figures 7 and 8. The numerical experiment serves as an example of the advantages offered by data fusion. In this case, information was extracted through data fusion which would not have been available using either modality individually.

VI. Â PERSPECTIVE APPLICATION: AIRPORT SECURITY The same process may also have applications in airport security screening if microwave radar and x-ray CT measurements are both collected. Currently airport security CT scanners rely on image processing algorithms applied to the final imaging result to detect threats. If however, MWT sensors could be incorporated, threats could be detected automatically based on their material properties.

ACKNOWLEDGMENT Financial support for this work was provided by the Department of Homeland Security Science and Engineering Workforce Development Program.

REFERENCES

Fig. 7. CT-based, MWT-based and Fused cost functions plotted vs. radius.

Figure 7 shows that the CT-based and MWT-based cost functions are insufficient on their own. However, the fused cost function has a clear minimum for brain carcinoma with a radius of 25mm â&#x20AC;&#x201C; the correct solution.

[1] Â M. Tivnan, C. Rappaport, J. Martinez-Lorenzo, A. Morgenthaler. (2014, April) FDFD Microwave Modeling of Realistic, Inhomogeneous Breast Tissue Based on Digital Breast Tomosynthesis Priors for Cancer Detection. Presented at Northeast Bioengineering Conference. Print. [2] Â M. Tivnan, A. Morgenthaler, J. Martinez-Lorenzo, R. Moore, C. Rappaport. (2015, March). Fusion of Digital Breast Tomosynthesis and Microwave Radar Imaging for a High Contrast Breast Cancer Imaging Algorithm. Presented at URSI Journees Scientifiques. Print. [3] Â ICRU â&#x20AC;&#x153;Tissue Substitutes in Radiation Dosimetry and Measurementâ&#x20AC;?, Report 44 of the International Commission on Radiation Units and Measurements Bethesda, MD, 1989 [4] Â C. Gabriel, â&#x20AC;&#x153;Compliation of The Dielectric Properties of Body Tissues at RF and Microwave Frequencies,â&#x20AC;? Occ. and Env. Health Directorate, Brooks Air Force Base, TX, Rep., 1996. [5] Â Y. Done-Sik, K. Bong-Seok, C. Hyung-Do, L. Ae-Kyoung, P. Jeongâ&#x20AC;&#x201C;Ki. October 2004. Dielectric Properties of Carcinomas. Bioelectromagnetics. Print. Volume 25 (issue 7), pages 492â&#x20AC;&#x201C;497, Available: site/path/file

STUDENT BIO: MATT TIVNAN Matt Tivnan is a Senior at Northeastern University pursuing a combined degree in Electrical Engineering and Physics. He has conducted advanced imaging research for four years with a focus on Microwave Radar Imaging in biological tissues. His co-ops have included the École Supérieure d'Electricité in Paris, France where he developed extensions to his undergraduate research and PhotoDiagnostic Systems in Boxboro, MA where he helped design and build CT and PET scanners. Matt is a mentor in the Gordon Scholars program, a COE mentor in the first-year program, a Community Outreach Officer for the IEEE student chapter, and a member of the ALERT student leadership counsel. He is also the first recipient of the undergraduate fellowship sponsored by the Department of Homeland Security Science and Engineering Workforce Development Program. Matt plans to pursue a graduate education and a career in the field of advanced medical imaging technology.

Programming Acoustic Modems for Underwater Networking Andrew Tu, Student Member, IEEE, Brian Wilcox Student Member, IEEE,, Mark German, Yashar M. Aval, Member, IEEE, and Stefano Basagni, Senior Member, IEEE

Abstract—Underwater acoustic communication and networks have attracted significant attention in recent years, with applications ranging from ocean monitoring to off-shore sensor control, and port surveillance. Experimental data are required to test and develop effective underwater networking protocols before underwater networks can be successfully deployed for real world applications. Unfortunately, there are very few permanent underwater acoustic testbeds currently in operation, making it difficult for full scale tests to be conducted. To meet the demands for experimental data, we are working to deploy a permanent underwater acoustic network at the Northeastern University Marine Science Center in Nahant, MA. At the final stage, the network will consist of at least five SM 975 Teledyne Benthos acoustic smart modems, with one wirelessly connected to the shore through a smart buoy of our design. This paper describes the interface for programming these modems and how we used it to implement a fundamental protocol to be used as performance benchmark for more advanced underwater solutions.

I. I NTRODUCTION With over 70% of Earth surface covered in water, underwater communications and networking are critical technological developments with countless applications in the research and commercial sectors. Wireless underwater networks will enable the remote monitoring and control of sensors and equipment that can be used to collect marine data and analyze their patterns, thus fostering new applications for the sustainable exploitation of this still unknown realm of our planet. Prevailing forms of terrestrial wireless communications, especially radio, are ill adept for long distance underwater transmissions. As such, acoustic communication provides a crucial component towards implementing large scale underwater networks. Currently, there are very few permanent testbeds for underwater acoustic networks (UANs) in the world, severely limiting experimental results on the design of effective networking protocols. Our end goal is to build a permanent testbed that will enable easier, more efficient testing of protocols for UANs at all layers of the protocol stack. Northeastern University is set to design, develop and deploy a permanent testbed at its Marine Science Center campus in Nahant, MA. The network, called NU MONET for Northeastern University Marine Observatory NETwork, will consist of at least five SM 975 Teledyne Benthos acoustic smart modems with one modem connected to the shore via a radio link to program and control the network and to relay data to the final user. A sketch of the planned NU MONET is depicted in Fig. 1. A. Tu, B. Wilcox, M. German, Y. M. Aval and S. Basagni are with the Department of Electrical and Computer Engineering, Northeastern University, Boston, MA, USA. Corresponding authors e-mail: tu.a@husky.neu.edu.

Fig. 1: The NU MONET to be deployed in Nahant, MA. The connections between the modems are underwater acoustic links (shown in white). One modem acts as the network gateway to the terrestrial Internet via a smart buoy (of our design). The connection between the buoy and the shore station is radio (ZigBee1 -based; black line in the picture).

This paper concerns programming the Teledyne Benthos modems as a fundamental step towards building the NU MONET. Modem programming largely depends on the API provided with the devices. Teledyne Benthos modems, for instance, come with BenthoNet as the interface for implementing basic networking functionalities (see Section II). Other approaches to underwater network programming go well beyond the modem’s relatively limited interface, and concern the use of suites for integrated simulation, emulation, and testbed trials of protocols at all levels of the networking stack. Examples of this approach are provided by tools such as SUNSET [1] and DESERT [2]. The SUNSET (Sapienza University Networking framework for underwater Simulation, Emulation and reallife Testing) framework was developed by a team at the University of Rome “La Sapienza.” The package is designed to allow developers to easily conduct simulations, emulations and field tests of underwater communication protocols [1]. DESERT (DEsign, Simulate, Emulate and Realize Test-beds for Underwater network protocols) is a set of C++ libraries extending the NS-MIRACLE simulation framework to support the development and implementation of UANs [2]. The key idea of both packages is that of using the same code for emulation and simulations of underwater protocol and also, once the physical layer is provided by actual modems, to

run that same code for experimentation in the water. These software packages are designed to be largely platform independent, in that given the appropriate driver, they can be used irrespective of the specific modem. The approach to modem programming described in this paper is of the first kind, i.e., built on the specific interface of the Teledyne Benthos modems. In particular, we develop Matlab-based usecase cases for controlling the modem from an outside device running Matlab (i.e., a laptop or other small-factor computer that can be deployed underwater with the modem itself), and with which we can implement different communication protocols. Our cases range from low-level, driver-like interfaces between Matlab and the hexadecimal-based commands of the modems, to more “autonomous” cases taking care of basic networking primitives, such as sending and receiving data, topology setup, etc. As a case study, we show how we use these use cases to implement one of the simplest medium access control (MAC) protocols, namely, ALOHA, for which we show results obtained in water that are based on our implementation. The rest of the paper is organized as follows. Section II describes the SM 975 Teledyne Benthos acoustic smart modem. In Section III we show how to control the modem via Matlab-based programming. In the following Section IV we demonstrate our programming primitives for implementing ALOHA and show some experimental results obtained through our implementation. Finally, Section V concludes the paper. II. T HE SM 975 T ELEDYNE B ENTHOS ACOUSTIC M ODEM The SM 975 Teledyne Benthos Acoustic Smart Modem is one of the fundamental components of our project. The modem electronics are housed within a vacuum sealed glass sphere, nestled within a two piece polyethylene hardhat. The vacuum seal allows the modem to descend to depths up to 6,700 meters underwater. An omnidirectional transducer extends from the top glass hemisphere and through a cutout in the hardhat. The bottom of the hardhat has an electrolytic dissolving “burn wire” which allows the remote release of the modem from the sea floor.2 Users can interact and control the modem via a proprietary serial port/power connection near the modem base. The modem contains a 28 Ah battery allowing the modem to run on battery life for up to one year between charges depending on usage. A sketch of the SM 975 is shown in Fig. 2. The low frequency acoustic modem can emit sounds between 9-14 kHz and has two modes of operation. The first mode uses coherent detection and employs phase shift keying (PSK) modulation to deliver baud rates of up to 15,360 bits/sec, while the second mode is based on incoherent detection and employs frequency shift keying (FSK) modulation, delivering baud rates up to 2,400 bits/sec. Note that while FSK modulation cannot deliver data rates as high as PSK modulation, it has improved reliability as compared to PSK modulation, which makes it a practical solution for more challenging channels. 2 The modem is positively buoyant and will ascend to the surface when released.

Fig. 2: A cut-away view of the SM 975 Teledyne Benthos Smart Modem. The Teledyne modems come pre-loaded with BenthoNet, a software API that enables basic primitives for packet transmission and reception as well as simple networking functionalities, including automatic acknowledgment replies and automatic re-transmission of data packets. We have disabled most of these functionalities as they are not dynamically controllable. This has allowed us to implement channel access protocol with greater control over key parameters. III. P ROGRAMMING THE SM 975 The SM 975 modem is interfaced through a serial port to an external computer, i.e., a laptop or a small form-factor computer such as the BeagleBone Black [3]. Through this interface, we control and program the modem via a series of Matlab functions and scripts. We organize the code for the SM 975 in a series of layers as shown in Fig. 3. Teledyne Benthos modems interact with the user via messages of four different types: “get,” “set,” “execute,” and “notify.” Get is used to read parameters from the modem (e.g., its local address, which is a positive integer in {0, 1, 2, . . . , 63}); set is used for changing a parameter (e.g., setting the local address to a different number), and execute asks the modem to perform a certain task (e.g., send a ping). These three commands are always generated by the user. The notify message, on the other hand, is generated by the modem and carries the modem response to the computer (e.g., the response to a get request), or is used to indicate that an event has occurred (e.g., notify the user about the reception of a packet from another modem). These four message types form the entirety of communication modes to and from the modem. Commands are sent to the modem in a string of hexadecimal bytes through the serial port connection. The available com-

Tdt_ Tda_ Td_ HEX Code SM 975 Acous5c Smart Modem

Fig. 3: The Matlab-based NU MONET code architecture for the SM 975. The users work with the “tdt ” layer that uses functions in lower layers to accomplish tasks. mands are laid out in the proprietary manuals accompanying the SM 975 modems. In order to interface more easily with the modems, we implemented a Matlab base level code to send and receive bytes to/from the modem. This code takes in a command (e.g. “td get(s,“localAddr”, 1)”) and converts it to the appropriate hexadecimal byte string. With the base layer established, we build up a new layer that utilize the “td ”, or “Teledyne”, functions to perform more automated tasks. Functions in this layer of communication are prefixed by “tda ,” where the appended “a” stands for autonomous. For instance, these functions may combine “td get,” “td notify,” and some additional logic for extracting information like the local address from the modem into a usable variable in a protocol code. Desired information returned in the modem response is extracted by the function and stored in memory. On top of the “td ” and “tda ” layers we define a new layer, termed “tdt ” where the extra “t” is short for “timer.” This layer is based on Matlab timers set to periodically trigger the execution of specific functions. Functions defined at the “tdt ” layer are used to implement several use cases for protocol implementation and testing. We describe here five fundamental basic cases. They concern sending a packet, receiving a packet, building the list of a node neighbors,3 namely, a node “address book,” functions for traffic management, and a GUI interface that we built in Matlab for run time control of relevant modem status parameters (e.g., the length of its data queue). Every case utilizes queues to which packets are pushed and pulled. The queues, and a set of global variables acting as “bridges” between case, allow the functions to pass information back and forth. For example, a new data packet generated by the traffic management case is pushed into the global transmit queue, from which the sending case will pop and transmit packets. In the rest of this section we describe these cases and structures in details. The data production and transmission case utilizes the 3 Two nodes are said to be neighbors if they can communicate directly to each other, i.e., when there is a physical channel between them that they can use to exchange information.

transmission queue to pass and store packets. The transmission queue is a series of parallel queues with each queue specific to a modem in the address book. Having several parallel queues specific to a modems affords more versatility than one large queue for all data packets. By separating the packets by modem, the sending function can randomly select a nonempty queue to pull packets from and send. In the event that connection to a modem is lost (thus resulting in delays and dropped messages), the sending function can continue to operate normally via another queue without getting stuck waiting on the retransmission of the lost packet. Furthermore, if a modem is dropped from the network, any packets generated for that modem can be deleted more easily, without having to sift among many packets with different destinations. We have functions for generating data packets (for instance, for testing purposes) which also include pertinent time stamps and diagnostic information. A data generating function is associated with its own timer. This function can be customized to implement different traffic generation patterns, such as Constant Bit Rate traffic, or random traffic generated according to a distribution relevant to a specific application. When a data packet is produced for transmission and its destination node is chosen, it is inserted into the transmit queue for that selected destination. When the tx timer function is called, the lengths of all the transmission queues are evaluated. If the sum of these lengths is greater than 0, (i.e., there are messages ready to be sent), the function generates random numbers between 1 and the current size of the address book to be used as indexes. The queue lengths of the randomly produced indexes are checked until an index with queue length greater than 0 is found. The function extracts the oldest message to be sent from the queue at the given index (i.e., the first message in that queue) and sends it, moving the message after transmission to a buffer queue and shifting all messages in the transmission queue forward. When an acknowledgment (ACK) is received from the recipient, the packet is deleted from the buffer. If an ACK is not received within a designated timeout, the modem will attempt to retransmit the message from the buffer a defined number of times, which is protocol dependent. All notifications that a modem receives or generates sit in the serial buffer pending processing by the notification timer function “tdt notify.” When “tdt notify” is called, the function immediately checks the length of the serial buffer to see if a new packet has been generated. If the length is greater than 0, (i.e., bytes for a notification packet have been generated by the modem and are sitting in the buffer), tdt notify reads the available bytes into a packet. A signature is generated and assigned to the packet as it is processed. The signature, comprised of key elements of the packet, is used to efficiently discriminate and act upon packets depending on the contained data. The information received is interpreted by “tdt notify” upon which it is either pushed to the appropriate queue or passed along to another function for further processing. The data stored in queues will be pulled by the corresponding timer function. This process can also work in reverse mode, in which “tdt notify” reads from the queues of the other functions, and forwards the information to the modem.

IV. C ASE S TUDY: ALOHA We demonstrate the use of our primitives for programming the SM 975 by implementing a basic MAC protocol for channel access, namely, ALOHA. The ALOHA protocol is one of the earliest and most basic random access protocols proposed for wireless broadcast channels. In the original version of this protocol, packets are sent in their entirety immediately upon generation, indiscriminately and without limiting factors [4]. ACKs are sent back to the sending nodes to confirm the successful delivery of a packet. In the event of packet collision, the sending nodes will immediately attempt to retransmit their packets with probability

p, opting to remain silent with probability of 1 − p. The randomness in packet retransmission decreases the probability of repeated collisions. Nodes will attempt to retransmit a total of retransmission attempts times before discarding the packet. Fig. 4 shows the set of actions performed within each of the SM 975 modems for generating a data packet and transmitting it according to our Matlab implementation of ALOHA. Data produced by data produc.on .mer Tx queues

Each modem has an address book containing the addresses, distances, and other pertinent information of every modem within transmission range. New modems can be added to an address book in one of three ways. The first is through a “ping” message broadcast to all modems within transmission range. Any modems who hear the ping respond with their own address and distance at a random time (up to time tping ). After the timeout period, the ping originator produces a notification containing the addresses and distances of all the modems who responded. The function “Tdt notify” identifies this notification based on its signature, and passes the packet into another function, “tdt AddToAddrBook,” with a flag indicating a ping response. The flags “Ping,” “justAddr,” or “Range” specify how the information is being presented to the function so that it can be properly extracted. In particular, “Ping” may contain multiple addresses and distances, “justAddr” only contains a single integer pertaining to the extracted address, and “Range” contains both address and distance updates. The other ways addresses can be added to the address book involve overhearing modems talking to one another. When a message is transmitted from a modem, the modem also sends several system messages with information such as timestamps, range updates, and acoustic information. Any modem that overhears these messages generates notifications indicating their receipt, but ignores the actual data packets unless they are addressed specifically to them. Through signature identification, “tdt notify” can pass information from certain packets to “tdt AddToAddrBook” tagged with either “justAddr” or “Range,” depending on the nature of the receipt. When packet information is passed to “tdt AddToAddrBook,” the address is first checked against the current entries in the address book. If an address already exists in the address book, the function simply updates the modems distance (if applicable). If the address is new, all pertinent information (address, distance, time last heard from, etc.) will be added to the address book. We developed a GUI in Matlab to display the contents of the address books of a node, the lengths of its queues and other status information. The GUI is updated every .5 seconds through a timer. Two curves are depicted that graph the length of the tx queue with respect to time and the end-to-end delay for each received data packet. Additionally, the GUI offers several buttons and sliders to input and change modem settings mid run. This includes a slider to set how often packets are generated, toggles for enabling data generation and sending pings, and buttons to clear the TX queue and send packets.

Modem 1 0 packets

Modem 2 0 packets

Modem 3 1 packets

…

Modem n 0 packets

Tx .mer Td_exec – remote send data Transmission ACK

Buﬀer (awai.ng ACK) No

Delete packet

No ACK Retransmissions >= 3

Yes

Fig. 4: The path of a data packet from production to deletion after transmission in our implementation of ALOHA. The data production timer generates a packet and places it in one of the transmission queues (tx queue) of a selected destination, which is any one of the other n−1 modems. In our experimentation below data packets are generated according to a Poisson distribution, while the data packet destination is any of the other modems, selected randomly and uniformly. All transmissions and retransmissions are coordinated through the tx timer. When the tx timer is called, the generated packet is transmitted by passing it into a “td exec” remote send data command, which converts it into the appropriate hex string for the modem. After transmission, the packet is moved from the tx queue into a buffer queue until an ACK is received from the recipient. If an ACK is not received within a designated timeout, it is assumed that the packet did not reach its intended destination and the modem will attempt to retransmit the packet. The modem will then attempt to resend the packet with probability p. The probability of a successful transmission is derived in [5] as psuccess = p(1 − p)2(n−1) , where n is the number of nodes in the address book. The probability of a successful transmission is maximized by setting the 1 (re)sending probability to p = 2n−1 . The packet is then transferred to the end of the buffer queue while it awaits an ACK. With probability 1 − p, the modem will not attempt to retransmit and will remain silent for the duration of a packet before attempting to retransmit again. This process repeats until the message is transmitted successfully and the sender receives an ACK from the recipient, at which point the packet is deleted from the sender memory. The packet can also be dropped and deleted from memory if it fails to be successfully transmitted within a predefined number of attempts (set to three in our experimental evaluation and in Fig. 4).

We use our implementation to conduct underwater experiments with three SM 975 acoustic smart modems running ALOHA. We deployed the modem in an indoor rectangular tank filled with fresh water positioned in a line (Fig. 5).

of light will instead lead to certain collision. Unfortunately, we did not benefit from this effect as our current testbed setting restricts us to very similar distances between modems. However, future experiments that run in open water (i.e. during final deployment in the ocean) may benefit from this effect due to the more variable distances between modems. V. C ONCLUSIONS AND F UTURE W ORKS

Fig. 5: Three TB SM 975 acoustic modems. In this lab setting the devices are controlled by laptops outside the tank.

0.8

0.6

0.4

0.2

average delay [s]

packet delivery ratio

We investigate the metrics of packet delivery ratio, i.e., the percentage of packet successfully delivered to their intended destination, as well as the end-to-end delay incurred by packets from the moment they are created to when they are delivered. The results of the performance evaluation of ALOHA are shown in Fig. 6.

packet deliver ratio average delay

0 0.05

0.06

0.07

0.08

0.09

0.1

0.11

0.12

0.13

0.14

0 0.15

packet generation rate [packets/s]

Fig. 6: Packet delivery ratio and end-to-end delay for the ALOHA MAC protocol in the setting of Fig. 5. We observe that as the rate of packet generation increases, the packet delivery ratio decreases while the end-to-end delay greatly increases. As for its terrestrial counterpart, the simplistic nature of ALOHA used under water leads to very low channel utilization.4 This is the consequence of not having an effective collision detection/avoidance mechanisms, like those used by CSMA/CD and /CA-based protocols. Curiously, the higher propagation delay of the underwater channel helps improving channel utilization. In fact, two senders at different distances from the same receiver and transmitting a packet at the same time are less likely to incur collision at the receiver than if the networks was terrestrial (radio), where the speed 4 It is notoriously known that the maximum channel utilization of un-slotted ALOHA in radio networks with saturated traffic is about 18% [6].

We used Matlab-based programming to develop a series of use cases for controlling and programing Teledyne Benthos SM 975 Acoustic Smart Modems. We demonstrate how to use these cases by implementing a basic channel-access networking protocol, ALOHA, and running underwater experiments. Eventually, we plan to case these basic use cases for developing more complex modern protocols designed specifically for use in UANs. We have started implementing one such protocol (TARS) and hope to begin testing soon. The TARS (Traffic Adaptive Receiver Synchronized) protocol is a more complex protocol designed specifically for underwater networking that will be implemented via the cases outlined in this paper. Once we have established a stable base of protocols, we hope to expand our small scale underwater testing to the ocean, and collect more realistic results for further analysis. In addition to programming the modems, we are also working on other portions of the NU MONET in preparation for a full deployment of five modems at the Northeastern Marine Science Center campus in Nahant, MA, over the Summer of 2016. This includes the design and construction of a “smart buoy” that will serve as the wireless link to the shore and provide power to the modems. ACKNOWLEDGMENTS This research was supported in part by the NSF MRI grant CNS 1428567. Tu, Wilcox and German were supported by REU grants associated to this MRI award. Andrew Tu and Stefano Basagni were supported by NSF through a GENI SAVI travel grant. The work was also performed under the sponsorship of the EU FP 7 ICT project SUNRISE “Sensing, monitoring and actuating on the UNderwater world through a federated Research InfraStructure Extending the Future Internet.” B IOGRAPHIES Andrew Tu is a rising second year undergraduate student at Northeastern University majoring in computer engineering and computer science. He is a student member of the IEEE and a member of Toastmasters International. Tu has been an NSF REU supported student on the NU MONET project since early October 2015. During the Summer of 2016, Tu was sponsored through a GENI SAVI travel grant to work on the NU MONET project in Rome, Italy. Brian Wilcox is a senior in Electrical and Computer engineering co-oping at MIT Lincoln Laboratory. He is also an IEEE student member and the Vice-President of Northeastern’s Eta Kappa Nu chapter. Mark German is entering his third year in the Computer Engineering program at Northeastern, and currently in a six month co-op at Schlumberger. He received an award for Outstanding Student Research in Engineering and Technology at Northeastern’s RISE 2016 event.

Yashar M. Aval graduated from the University of Tehran, Tehran, Iran, in 2002 and received the M.S. degree in electrical engineering from Sharif University of Technology, Tehran, Iran, in 2005, and his Ph.D. degree in electrical engineering at Northeastern University, Boston, MA, USA, in 2015. Currently, he is working as Associate research engineer at Northeastern university. His research interests include digital communications, OFDM, underwater acoustic communications, and underwater acoustic networks. Stefano Basagni is associate professor at the Department of Electrical and Computer Engineering at Northeastern University, in Boston, MA. He holds a Ph.D. in electrical engineering from the University of Texas at Dallas (December 2001) and a Ph.D. in computer science from the University of Milano, Italy (May 1998). Dr. Basagni’s research interests concern research and implementation aspects of mobile networks and wireless communications systems, radio and acoustic sensor networking, definition and performance evaluation of network protocols and theoretical and practical aspects of distributed algorithms. Dr. Basagni has published over six dozens of refereed technical papers and book chapters that are highly cited (his h-index is currently 32, with over 8500 citations to his works). He is also co-editor of three books. Dr. Basagni serves as a member of the editorial board and of the technical program committee of ACM and IEEE journals and international conferences. He is a senior member of the ACM (including the ACM SIGMOBILE), a senior member of the IEEE (Computer and Communication societies), a member of ASEE (American Society for Engineering Education) and of CUR (Council on Undergraduate Research).

R EFERENCES

[1] C. Petrioli, R. Petroccia, and D. Spaccini, “SUNSET version 2.0: Enhanced framework for simulation, emulation and real-life testing of underwater wireless sensor networks,” in Proceedings of ACM WUWNet 2013, Kaohsiung, Taiwan, November 11–13 2013, pp. 1–8. [2] R. Masiero, S. Azad, F. Favaro, M. Petrani, G. Toso, F. Guerra, P. Casari, and M. Zorzi, “DESERT underwater: an NS-miracle-based framework to DEsign, simulate, emulate and realize test-beds for underwater network protocols,” in Proceedings of IEEE OCEANS 2012, Yeosu, Korea, May, 21–24 2012, pp. 1–10. [3] What is beaglebone black? [Online]. Available: https://beagleboard.org/black [4] N. Abramson, “The ALOHA System—another alternative for computer communications,” in Proceedings of AFIPS 1970, 1970, pp. 281–285. [5] J. F. Kurose and K. W. Ross, Computer Networking—A Top-Down Approach Featuring the Internet, 6th ed. Addison Wesley, 2013. [6] L. G. Roberts, “ALOHA packet system with and without slots and capture,” Stanford Research Institute, Advanced Research Projects Agency, Network Information Center, Stanford, CA, Tech. Rep. ASS note 8, June 1972.

Microwave ignition for nanostructured reactive composites Gianmarco Vella, Student at Advanced Materials Processing Lab (AMPL)

Abstract— The need for heating at nanoscale has pushed researchers in the study of reactive, nanostructured composites known as nanoheaters. Major topics of interest are the best conditions for consolidation, composition, and ignition of these innovative heat sources. This work presents a new method of ignition for nanoheaters, known as non contact microwave ignition, distinguishing itself from previously developed direct heat application methods. Al-Ni nanoheaters were fabricated through ultrasonic powder consolidation (UPC) with embedded aluminum and copper wires. The conductive properties of the embedded wires, acting as susceptors when exposed to electromagnetic radiation in the microwave range, were found to induce enough heat to Al-Ni nanoheaters to facilitate ignition. This nullifies the requirement of direct heat application to the fabricated nanoheaters to produce ignition. In addition to testing in gaseous environment, this new method of ignition for nanostructured, reactive composites was also tested in vacuum, verifying its effectiveness in a non-gaseous environment.

I. INTRODUCTION Nanoheaters are nanostructured reactive composites that allow localized and controlled heat generation when ignited. The advantages of providing controlled heating to a designated volume [1] through the use of this new form of reactive materials, have brought major advancements to a vast number of fields and processes such as microscale joining, microelectronics soldering [2], reshaping of parts for environmental degradation and ablation of biomaterials of cancer cell walls [3]. Current ignition methods of nanoheaters include continuous heating through a heater plate [2], direct heating through laser pulses or plasma torch [3] and direct current passing through the nanoheater [4]. The common aspect among the above ignition methods is the need of direct supply of applied heat to the nanoheater for ignition to occur, which restricts a lack of application to manufacturing processes where direct contact between the nanoheater and the heat source is not feasible. Today’s industrial applications of nanoheaters highlight the need for a non-contact, indirect heating method for the ignition of these nanosized reactive materials. In this paper, a newly developed non-contact microwave ignition method is presented. Leveraging the conductive properties of metals, the fabricated nanoheater with an embedded conductive wire will spark when exposed to electromagnetic radiation. The mechanism behind the

conductive wires acting as susceptors for the nanoheater was investigated by subjecting aluminum and copper wires to electromagnetic radiation and observing their breakdown and sparking patterns. The consolidated nanoheaters with embedded aluminum and copper wires were tested in air and vacuum environments. By eliminating any form of direct application of heat to ignite nanoheaters, the non-contact microwave methodology enables the application of nanoheaters to manufacturing processes where controlled temperature and complex geometries are not to be seen as limitations [4].

II. EXPERIMENTAL PROCEDURE The fabrication of the Al-Ni nanoheaters with conductive embedded wires was divided into three steps.

A. Dimensioning of wires Aluminum and copper wires to be embedded in the nanoheaters were cut to size to investigate their properties as susceptors when exposed to microwave radiation. Two aluminum and copper spools with a wire diameter of 0.405 mm were used in the experiment. The wire spools were purchased from Malin Co. and Arcor Inc., respectively. A Mitutoyo Electronic Caliper was used to accurately measure the sections of conductive wire. A standard box cutter was used to cut the wires. Table I below shows the lengths of wires used in the experiments: TABLE I LENGHTS OF ALUMINUM AND COPPER WIRE SECTIONS

Length #1

Length #2

Length #3

Length #4

Copper

20 mm

25 mm

30 mm

35 mm

Aluminum

24 mm

30 mm

35 mm

Due to the cross section of the wires and the tools used to cut them to size, the radius of the tip of each cut section could not be controlled. Overall, the goal was to cut the wires such that the tip radius resulted to be less than or equal to half of the wire diameter.

B. Mixing process Aluminum and nickel nano-flakes with sub-micron thickness (100-300 nm) were supplied by Fukuda Metal Foil & Powder Co., Ltd. The aluminum and nickel nano-flakes were mixed with a 1:1 molar ratio. The mixing process for the consolidation of the nanoheaters consisted of three steps: dry mixing in air,

2 mixing in ethanol and sonication in ethanol [1]. Table II below summarizes the corresponding times for each step of mixing: TABLE II DURATIONS FOR THE THREE STEPS OF NANO-FLAKE MIXING

Dry mixing in Air

Mixing in Ethanol

Sonication in Ethanol

60 min

120 min

A standard rotary powder mixer and a FS20D Sonicator by Fisher Scientific were used for the mixing process. The blended powders, once sonication was complete, were then left to dry until total evaporation of ethanol and dry mixed for 20 minutes before being used in the consolidation of the nanoheaters. The above three steps of mixing were followed to achieve better deagglomeration and reduction of aluminum and nickel nanoflakes clusters [1].

nanoheater with embedded conductive wire, Fig. 3, was rapidly taken out of the punch and the heater turned off. This procedure was used for the consolidation of all nanoheaters used in the experiments of this paper.

Fig. 2. Schematic of the punch and die for the consolidation of Al-Ni nanoheaters.

C. Consolidation of nanoheaters The metastable state of nanoheaters requires a fabrication methodology that does not employ the use of high temperatures [1]. Ultrasonic powder consolidation (UPC) was employed in this experiment, being a new manufacturing technique that allows powders to metallurgically consolidate when subjected for a few seconds to ultrasonic vibration and low or moderate temperatures. The ultrasonic welder STAPLA Condor® used is a 3 kW setup with a fixed frequency of 20 kHz and an amplitude of vibration at the sonatrode of 9 µm [2]. Additionally, the UPC setup employed a heater plate to perform the consolidation of the specimens at elevated temperatures. A set of a die and a punch was used to consolidate the nanoheater, Fig. 1. The die had dimensions of 12.4 mm x 12.4 mm x 4 mm with a through hole 6.30 mm in diameter, and the punch was a disk of matching diameter of 6.3 mm. Both the die and punch were made of mild steel.

Fig. 3. Schematic of the punch and die for the consolidation of Al-Ni nanoheaters.

D. Encapsulation of nanoheaters To observe the ignition of the consolidated Al-Ni nanoheaters with embedded conductive wires in a vacuum environment, the nanoheaters were encapsulated in such an environment using Pyrex tubes, Fig. 4. A single nanoheater was inserted in the Pyrex tube closed at one end and attached to a Platinum JB pump on the open end. The pump evacuated the Pyrex tube at a rate of 1.5x10-3 m3/s, to a pressure of 4x10-4 mmHg [1] for 20 minutes. Once evacuation was complete, the open end of the Pyrex tube attached to the vacuum pump was sealed by heating with a propane torch.

Fig. 4. The encapsulated nanoheater in a vacuum environment.

Fig. 1. Schematic of the STAPLA Condot UPC setup

The die with an aluminum or copper wire was put in place on the UPC setup and the consolidation temperature was set to 573 K. As the heaters reached the final temperature, approximately 0.1-0.2 grams of previously mixed aluminum and nickel nanoflake mixture was put in the die and a uniaxial pressure of 100 MPa was applied to it through the punch. Once the final temperature of 573 K was reached, 1 second of in-plane vibration was applied, Fig. 2. The consolidated Al-Ni

III. Â RESULTS AND DISCUSSION A. Â Length of wire and time of first spark correlation 14 12

Time Â (s)

10 8 6 4 2 0 18

Length Â (mm)

(a)

sharp electrode. The instance of excessive ionization of a gas physically means that current is conducted through an ionized gas. In the experiment, as the microwave irradiation increases the potential of the wire, the molecules of the gas are also ionized, leading then to a release of the accumulated static charge. The release of static charge is the evidence of a corona discharge, seen by naked eye in the form of luminous and audible sparks. The value of ionization at which the gas starts conducting current is defined as â&#x20AC;&#x153;electric breakdown potentialâ&#x20AC;? of the gas [1]. The electric breakdown potential of air has a value of is 3x106 V/m. The equation for the electric breakdown potential of a gas is shown below: # đ??¸= (1) $%&' ( where Q is the value of free charge to be calculated as a function of wire volume, Îľ0 is the permittivity of free space constant with value of 8.852x10-12 F/m and r is the radius at the two ends of the wire. When the electric breakdown of air occurs, the conductive wires producing sparks melt and break up into smaller sections that are below the minimum sparking length. Fig. 6 below shows copper and aluminum wires after complete breakdown:

Time Â (s)

8 6

a 4 2

0 22

Length Â (mm)

(b)

Fig. 5. Time of first spark leading to electrical breakdown of wire when exposed to microwave radiation (a) copper wire (b) aluminum wire.

Fig. 5 shows the correlation between the time of first spark and length for both copper and aluminum wires subjected to microwaves in air. In this experiment, the time of first spark was defined as the instance at which the first spark is seen on the conductive wires exposed to microwave radiation, leading to their complete melting and breakdown. The time of first spark was found to decrease with increasing wires length. It was noticed that for copper and aluminum wires, below l <20 mm and l <24 mm respectively, no sparks occurred on the wires and therefore no breakdown of the wire, when exposed to microwave radiation. All wires were exposed to the electromagnetic radiation generated by a 1000-Watt commercial microwave with a 2.45 GHz chamber. Each piece of wire was placed on ceramic substrate and then placed in the microwave chamber. A plausible explanation for the sparks observed on the wires, leading then to their melting, can be given in terms of the corona discharge phenomenon. A corona discharge is a release of static charge caused by excessive ionization of a gas surrounding a

d Fig. 6. Wires after being exposed to microwave radiation until complete electrical breakdown (a) 30 mm copper (b) 35 mm copper (c) 30 mm aluminum (d) 35 mm aluminum.

The melting and separation observed in the mechanical breakdown patterns of the conductive wires indicates that heat is generated in the wires. The evidence presented above, supports the hypothesis that the conductive wires, acting as susceptors, absorb the electromagnetic energy. As more electromagnetic energy is accumulated, the potential of the wire increases, causing a current flow. Any form of current flow through a conductor having a resistance, will cause heat generations.

4 This form of heat generation, presumed to be the driving force of this non-contact ignition methodology, is a form of Joule Heating. The equation of Joule heating is shown below: *

đ??ť = đ?&#x2018;&#x2013; - đ?&#x2018;&#x2026;đ?&#x2018;Ą

B. Â Nanoheaters ignition in air Nanoheaters Â with Â embedded Â copper Â wire Copper Â wires Â only

(2)

Where j is a constant, known as Jouleâ&#x20AC;&#x2122;s mechanical equivalent of heat with a value of 4.1860 Jâ&#x20AC;˘cal-1, i is the current flowing through the conductive wire, R is the resistance of the wire and t is the time of current flow. In order to have current flow along the wire, a potential difference must exist between the two wire ends. Therefore, it is presumed that as the sparks appear on the wires, a high voltage difference is created between the two tips, subsequently causing current flow. We can calculate the theoretical voltage in the wire assuming that current flow starts at the instant of electrical breakdown of the ionizing gas: đ?&#x2018;&#x2030; =đ??¸â&#x2C6;&#x2122;đ?&#x2018;&#x; (3)

10 8

Time Â (s)

0 18

= 607 Â đ?&#x2018;&#x2030;

(4)

đ?&#x2018;&#x2026;-A88 Â BC Â DE(F = 0.00260 Â ÎŠ

(5)

đ?&#x2018;&#x2026;-$88 Â HI Â DE(F = 0.00525 Â ÎŠ

(6)

Finally, the theoretical amount of energy (in Joule) generated in the conductive wires when exposed to microwave radiation and therefore leading to the ignition of the nanoheaters, can be calculated as shown below:

đ??˝HIC8EUC8 Â ME(F =

O 7N O

Length Â (mm)

(a)

Nanoheater Â with Â embedded Â aluminum Â wire Aluminum Â wires Â only Â 10

(6AQ)N A.AA-6A Â (6AQ)N

A.AAV-V

= 1.417 â&#x2C6;&#x2122; 10T J/s

(7)

= 7.018 â&#x2C6;&#x2122; 10Q J/s

(8)

Based on previous experiments and related calculation conducted by Gheybi [1], the minimum amount of energy required for the ignition of bi-metallic, 2Al-Fe2O3-3(Al-Ni) nanoheaters was calculated. The same non-contact ignition methodology used in Gheybiâ&#x20AC;&#x2122;s experiments demonstrated that the embedded copper and aluminum conductive wires, produced enough energy for the ignition of the nanoheaters. The section below explores the efficiency of this nanoheaters ignition methodology, proving its effectiveness for the ignition of Al-Ni nanoheaters both in gaseous and non-gaseous environments.

Time Â (s)

â&#x2C6;&#x2122; 0.2025 â&#x2C6;&#x2122; 10

Further on, the theoretical values of resistance for the smallest wires exposed to microwave radiation were calculated as:

đ??˝BKLLF( Â ME(F =

7 106 8

4 2

where E is the electrical breakdown potential of the gas surrounding the conductive wire and r is the radius of cut at the two tips of the wire. For the conductive wires in Air and with a radius of cut of 0.2025 mm we calculate a voltage of: đ?&#x2018;&#x2030; = 3â&#x2C6;&#x2122;

6 4 2 0 18

Length Â (mm)

(b)

Fig. 7. Nanoheaters ignition time with embedded wire in Air (a) copper wire (b) aluminum wire.

Fig. 7 shows the effectiveness of the non-contact nanoheater ignition methodology. The fabricated nanoheaters were exposed to microwave radiation in a gaseous environment, air, all igniting on the first try. From the graphs, it is noticeable that the times of ignition of the nanoheaters, correspond to the times of first spark on the conductive wires tested in section A above. The results support and prove the hypothesis that there is current flow in the wires the moment sparks appear, resulting in enough heat generation to produce the ignition of the nanoheater.

5 C. Nanoheaters ignition in vacuum

for sparks to occur on the wire surface. The major difference in the tested non-gaseous environment is the way electrical breakdown is achieved. The early stages of vacuum breakdown can be justified by the cold electron emission phenomenon. This phenomenon causes electron emission by exposing the impurities and imperfections on the surface of the wires to a high electric field. Electron collisions, the essence of current flow inside the wires, generate heat due to Joule heating, leading to the melting of impurities and release of hot electrons. This chain of electron emissions is the base of the electron avalanche concept, also known as Townsend discharge [6], [7]. The constant heat generation finally leads to the breakdown of the wire and ignition of the nanoheaters.

Time (s)

8 6 4 2 0 18

Length (mm)

In conclusion, reactive Al-Ni composites were ultrasonically consolidated from Al and Ni nano-flakes with embedded Aluminum and Copper conductive wire. The consolidation of the nanoheaters occurred at a temperature of 573 K under 100 MPa of pressure. It was observed that the fabricated nanoheaters ignited within seconds without the need of direct heat application when exposed to microwave irradiation of a conventional 1000 Watt, 2.45 GHz microwave oven.

(a)

Time (s)

8 6 4 2 0 18

IV. CONCLUSION

Length (mm)

(b)

Fig. 8. Ignition times of Al-Ni nanoheaters with embedded conductive wire when exposed to microwave radiation in vacuum (a) copper wires (b) aluminum wires.

To further prove the effectiveness of the non-contact microwave ignition method in non-gaseous medium, the fabricated Al-Ni nanoheaters were ignited in vacuum. Fig. 8 describes the ignition times of Al-Ni nanoheaters with embedded wires of different lengths when exposed to microwave radiation in a vacuum environment. In vacuum, the ignition characteristics and mechanism of the nanoheaters remain unchanged. Current flowing through the conductive embedded wires, which generates enough heat for the ignition of the nanoheaters, is still the supported hypothesis in the tested non-gaseous environment. In Gheybi’s work [1], the ignition of nanoheaters in vacuum is hypothesized to occur at Schwinger’s Limit. In Quantum Electrodynamics (QED), the value of Schwinger limit is of 1.3x1018 V/m, representing the point at which an electromagnetic field becomes non linear [5]. Under such conditions, the dielectric of a material responds non linearly to an applied electric field. In the above condition, the lack of a gas surrounding the conductive wires, automatically excludes the possibility of ionization and electrical breakdown of air as the leading cause

The optimal conditions for the ignition of nanoheaters with embedded conductive wires was observed in vacuum. The AlNi nanoheater with 35 mm embedded copper conductive wire ignited in 2.16 seconds, when exposed to microwave irradiation. Additionally, the fastest ignition in vacuum was produced by nanoheaters with embedded aluminum wires. From the above experiments, it is hypothesized that in a gaseous environment, as dielectric breakdown of the gas is achieved and that the release of static charge, by means of corona discharge phenomenon, induces current flow in the conductive wires, generating enough heat for the ignition of the nanoheaters. In vacuum, it is hypothesized that due to Townsend electron avalanche, the chain of electron emission caused by the exposure of the impurities in the material to a high electric field, induces current flow in the conductive wires, resulting in enough Joule heat generation for the ignition of the nanoheaters. Further work is required for a deeper understanding of ignition characteristics of the nanoheaters with conductive embedded wires when exposed to microwave irradiation. A broader understanding of the nanoheater’s ignition mechanism with conductive embedded wires could enable us to fabricate more compact, efficient and advanced nanoheaters. Possible applications of more advanced nanoheaters and ignition methodologies could have lasting impacts on electronics soldering, ablation of cancer cells and similar fields where localized heat generation by means of indirect heat application to the heat source is required.

6 ACKNOWLEDGEMENT

The author would like to thank Professor Teiichi Ando for the constant help, mentorship and guidance, Dr. Somayeh Gheybi for the valuable advice and support and finally, all the members of AMPL for their friendship and encouragement. Gianmarco Vella is a 4th year Mechanical Engineering student from Milan, Italy. He started working at the Advanced Materials Processing Lab (AMPL) during the summer of his 3rd year. Email: vella.g@husky.neu.edu REFERENCES

[1] S. Gheybi Hashemabad, “Hybrid Bimetallic-Thermite Reactive Composites: Ultrasonic Powder Consolidation, Ignition Characterization and Application to Soldering,” Ph.D. dissertation, Dept. of Mech. and Ind. Eng., Northeastern Univ., Boston, MA, 2015. [2] S. Gheybi Hashemabad and T. Ando, "Ignition characteristics of hybrid Al–Ni–Fe2O3 and Al–Ni–CuO reactive composites fabricated by ultrasonic powder consolidation," Combustion and Flame Elsevier Journal, vol. 162, no. 4, pp. 1144-1152, 2015. [3] Z. Gu, Q. Cui, J. Chen, J. Buckley, T. Ando, D. Erdeniz, P. Wong, A. Hadjiafxenti, P. Epaminonda, I. Gunduz, C. Rebholz and C. Doumanidis, "Fabrication, characterization and applications of novel nanoheater structures," Surface and Coatings Technology, vol. 215, pp. 493-502, 2013. [4] C. Doumanidis, T. Ando, J. Chen and C. Rebholz, "Nanoheater elements, systems and methods of use thereof", US20090235915 A1, 2009. [5] Schwinger, "On Gauge Invariance and Vacuum Polarization," Phys. Rev., vol. 82, no. 5, pp. 664-679, 1951. [6] E. Lauer, "Electron avalanche model of dielectric-vacuum surface breakdown," J. Appl. Phys., vol. 102, no. 11, p. 113306, 2007. [7] R. Marić, K. Stanković, M. Vujisić and P. Osmokrović, "Electrical breakdown mechanisms in vacuum diodes," Vacuum, vol. 84, no. 11, pp. 1291-1295, 2010.