518

Page 1

Full Paper Proc. of Int. Conf. on Advances in Design and Construction of Structures 2012

Application of Neural Networks for Piezoelectric Sensor based Strength Monitoring of Concrete Cube Gaurav Duggal1, Venu Gopal Annamdas2,Saurab Verma3 1

Laboratory of Monitoring Science, Birla Institute of Technology and Science (BITS), Pilani, Hyderabad Campus, Hyderabad, India - 500078, 2 School of Mechanical and Aerospace Engineering Nanyang Technological University, Singapore – 639718 3 Laboratory of Monitoring Science, Birla Institute of Technology and Science (BITS), Pilani, Hyderabad Campus, Hyderabad, India - 500078, Email: 1gauravduggal@valonia.co.in,2anammdas@ntu.edu.sg,3saurabverma@gmail.com Abstract - Artificial Neural Networks (ANN) is an algorithm used in Artificial Intelligence and is used for modeling complex data. They have applications in various fields like Computer Vision, Self Driving Vehicles, amongst various other fields. In this paper we discuss its uses in Civil Engineering applications. Piezoelectric sensors which are used for Structural Health Monitoring (SHM) are being used to produce data required for ANN. The frequency output response (Impedance) of the piezoelectric sensors is obtained using an Impedance Analyzer. We get a characteristic impedance signal which is a function of frequency. This signal comprising of several peaks and valleys varies with change in strength of the concrete cube. The peaks/valleys can shift in amplitude or move left / right in the graph. This forms a preferable environment for use of neural networks since the variation in data (signal) is a complex function of the inputs. An ANN was constructed and trained on some training data. After this, the trained ANN was used to predict the strength of the cube based on the new signal data obtained from Analyzer, which is unseen data for ANN. This paper is expected to be useful for application of ANN in new civil engineering field with a good accuracy and efficiency. Keywords - Conductance, Susceptance, One vs. all Classification, Impedance, Piezoelectric sensor, Training data

1. The way different layers are connected to each other. 2. The training process for updating the weights (theta values) corresponding to each layer. 3. The ‘activation function’ that is the basic processing done by a node to convert an input to an output. Each consecutive layer has weights associated with them. Theses weights corresponding to each layer control ‘how the inputs produce a desired output’. These weights need to be ‘learned’ by using some training data. This training data should be a random collection of the various inputs (signals) that needs to be processed. The training data should be categorized manually based on desired output and both input signal and output values should be provided to ANN during training process so that it learns and is able to adopt itself for unseen but similar signal. This training changes the weights corresponding to each layer and hence the neural network is trained. This is basically a classification algorithm. The ANN tries to identify patterns among the data and then classifies them according to it. There is a cost (www.ml-class.org ) associated with a collection of training data and the ANN itself. This cost is minimized by changing the weights corresponding to each layer.

I. INTRODUCTION TO ARTIFICIAL NEURAL NETWORKS (ANN) ANN is an algorithm widely used in machine learning. It has been used in diverse fields like Autonomous Navigation Systems ([1] Pomerleau, 1991), Radar Systems ([2] Cheikh et al), Power Electronics ([3] Bose et al, 2001), Flight Control Systems ([4] Nakanishi et al, 2003), Economics ([5] Li et al, 2010), etc. It can be used in where there is a need to model complex functions of some inputs to specific outputs. It tries to work the way biological neurons work in our brains. They are used for purposes of classification of data or to find complex patterns in data. The three main features (arrangement and processing of nodes and layers) of a neural network can be identified in the pictorial representation of an ANN as shown in Fig. 1. ([6] Wikipedia 2011) as: 47 © 2012 ACEE DOI: 02.ADCS.2012.1. 518

Fig. 1. This shows a 3 layer Artificial Neural Network. The input layer consists of 6 nodes, there is 1 hidden layer consisting of 4 nodes and the output layer consists of 2 nodes


Full Paper Proc. of Int. Conf. on Advances in Design and Construction of Structures 2012 The arrows represent connections between the nodes. Each node in the input layer has a connection to each node in the hidden layer which is connected to the output layer. This type of ANN is also called a Feed Forward Artificial Neural Network.This is because it processes the data as it is fed forward from one layer to the other and eventually when it reaches the output layer it should be completely processed. As an example the input value of the topmost node in the hidden layer (Fig1) is the weighted sum of the input layer nodes as indicated by the arrows. The weights are assigned values depending on the data and have to be ‘learned’ from the training data. This is elaborated later in this paper. Each node (except in the input layer) consists of the activation function. This does the processing on the data and gives the ANN its non linear characteristics. The activation function, which we have used is the sigmoid function P(t). The input to the activation function is the weighted sum using the corresponding layers weights and over the input layer values.

The weights or theta values are learned using the backpropagation algorithm. This algorithm generates the errors between a desired output and the ANN output and tries to minimize them by controlling theta values. II. PIEZOELECTRIC SENSORS Piezoelectric sensors (PZT’s) are being used in SHM applications based on Electromechanical Impedance (EMI) technology ([8] Annamdas et al, 2010). It consists of a piezoelectric element which is bonded to the host structure on its surface or it may even be embedded inside the structure. The main characteristics of PZT’s are high elastic modulus, low tensile strength, and brittleness. In the EMI technique ([9] Yang et al, 2008), the PZT is fed with a sinusoidal voltage signal of varying frequency. The admittance (inverse of impedance) versus frequency is plotted to produce a structural response which has many peaks and valleys. This is known as the ‘electromechanical (EM) admittance signal’. The EM admittance signal depends on both the structure as well as the PZT material. We intend to use its dependence on the structural properties to predict the strength gain of concrete cube. The EM admittance signal depends upon the stiffness, mass, and damping of the host structure, the length, width, thickness and orientation of the PZT ([9] Yang et al, 2008). The admittance is expressed as a complex number, with the real part representing the conductance and the imaginary part representing the susceptance of the PZT as shown in Figs 3-4 (representative signals obtained for strength gain of cube).

Thus, this activation function was chosen because the minimum value it has is 0 and the maximum value is 1. Between just 1 order (substituting -10 to 10 in‘t’) it changes from close to 0 to almost 1. Now we have a decision box acting over this sigmoid function. If P (t) is close to zero (below the threshold value of t) we say the output of the node should be zero else the output should be 1. Now we can control the threshold, when the node should output 0 or 1 by controlling the weights. Now our problem is a multiclass configuration problem. This means the output layer corresponding to our ANN has more than one node (one for each class). The way the ANN tells us that the input data is in a specific class is by a one versus all representation ([7] www.ml-class.org) This means an output 1 will appear in the output node corresponding to the category which the input data belongs to, while the other output nodes will have 0’s. So for a particular set of learned theta (weights) values and for a particular input data sample we get 1 in a single node and zeros in all the other nodes (Fig. 2).This is how the ANN classifies data into categories (denoted by the output layer nodes).

Fig. 3. Frequency vs Susceptance

Fig. 2. One Vs All Classification. The ‘1’ in the second node in the output layer means the layer has classified the current input data sample in the second category. Fig. 4. Frequency vs Conductance

© 2012 ACEE DOI: 02.ADCS.2012.1. 518

48


Full Paper Proc. of Int. Conf. on Advances in Design and Construction of Structures 2012 The impedance values were measured using an impedance analyzer. Impedance Analyzers such as 4294A Precision Impedance Analyzer made by Agilent Technologies were used. The circuit as shown in Fig 5 is used for measuring impedance. The impedance analyzer applies a voltage V0 through a lead wire (or circuit) and measures the current passing through the circuit. On dividing the voltage by current, we get the magnitude of the impedance. The phase values of the impedance are measured by calculating the phase difference between the current and voltage. The impedance can therefore be represented by a complex number since we know both the magnitude and the phase angle. On multiplying the cosine of the phase angle with the magnitude of impedance we get the magnitude of conductance whereas multiplying the sine of the phase angle with the magnitude of impedance we get the susceptance values. All this is done within the Impedance analyzer itself.

As concrete cures (as time passes) its strength increases. A signal reading was taken approximately 3 times a day. The signal was taken over a frequency range of 0-500kHz in intervals of 0.5 kHz, so there were a total of 1000 conductance values per signal. An ANN consisting of 3 layers (including 1 hidden layer) was constructed in MATLAB. We decided to classify the data into 3 categories depending upon the amount of strength gain. So the ANN would predict the strength of a given conductance signal by classifying it in one of the 3 categories with 1 being weakest and 3 the strongest. This would mean the ANN would predict a given signal of unknown strength and place it one of these 3 categories. This sets the size of the output layer to 3 nodes (as per one versus all representation). The data was first labeled with each signal corresponding to a number between 1 and 3 corresponding to its strength. Then it was arranged in order of increasing strength since we know the strength of concrete increases from the time it is set. Then the first 29 signals (of 86 signals) were placed in category I, the next 29 signals in category II and the remaining 28 signals in category III. To generate the test dataset 5 random signals from each category were selected from the previous step (Fig. 6). The remaining set is the training data.

Fig. 5. Circuit associated with the PZT sensor.

There are various factors due to which the impedance values can be corrupted. [10] Park et al (2000) found out temperature affects the signal and can shift the peaks uniformly left/right or up/down, or appropriate combinations of both. [11] Raju(1998) studied the effects of PZT’s lead (wires) lengths and excitation voltage levels and found out that leads less than 30m in length have no adverse effects on measuring impedances. [8] Annamdas (2010) states the frequency range greater than 500 kHz is unfavorable so we have limited frequencies to less than 500 kHz. [12] Sun et al (1995) indicates that conductance signals are good indicators of damage/strength which contradicts our observation that since the variance of the Susceptance was much higher than the conductance signal so the Susceptance signal might give better results than the Conductance signal. Hence we concentrated our efforts on predicting the strength/damage of a concrete cube using only conductance signal.

Fig. 6. The Raw labeled data is split into Test and Training Data

The corresponding strength vector which specifies the strength of each of the training signal was labeled manually as mentioned before. The ‘test’ data which should be unseen by the ANN was made by randomly selecting 5 signals from each class so our training data size reduces to 71 signals (24+24+23).The input features based on which the ANN will classify the data were taken as the individual conductance values at various frequencies, so there were 1000 input layer nodes (as there are 1000 conductance values per signal). The hidden layer size was taken to be made up of 250 nodes. With 1000 iterations of the back propagation algorithm we were able to train the ANN such that it predicted 12 out of the 15 test data correctly. The test data has not been shown to the ANN while it is being trained so it has no knowledge of that. Once the ANN is trained using the training data it learns to recognize what kind of features in a certain signal corresponds to a certain strength category, hence the ANN is able to classify the unknown test dataset into one of the 3 strength categories.

III. APPROACH The test specimen was a 150 x 150 x 150 mm3 concrete cube. A single PZT sensor was bonded on the surface of the cube. We had collected a data of 86 signals for the same cube during a period of 28 days of concrete curing 3 times a day (on 2 days 4 readings were taken). These 86 signals would form our ‘training dataset’ (used to train the ANN) and the ‘test dataset’ (This is for checking unknown signal strengths). © 2012 ACEE DOI: 02.ADCS.2012.1.518

49


Full Paper Proc. of Int. Conf. on Advances in Design and Construction of Structures 2012 CONCLUSION

AutonomousFlight Control System for Unmanned Helicopter by use of Neural Networks.” (2003) [6] http://en.wikipedia.org/wiki/Artificial_neural_network Accessed on October, (2011). [7] Lectures 4 and 5 on ANN in the Online Machine Learning course taken by Dr. Andrew N.G. (Sept-Dec 2011) www.mlclass.org [8] Annamdas V.G.M, Yang Y. and Soh C.K., “Impedance based Concrete Monitoring using Embedded PZT sensors.” (2010) [9] Yang Y., Hui L., Annamdas V.G.M., “Monitoring Damage Propagation using PZT impedance transducers”, (2008). [10] Park, S. Lee, J.J, Yun, C.B. and Inman, D.J. “Electro-Mechanical Impedance-based Wireless Structural Health Monitoring using PCA-Data Compression and k-means Clustering Algorithms; Journal Intelligent Material Systems and Structures, 19:509520. (2008) [11] Raju, V. “Implementing Impedance-based Health Monitoring Technique.” Thesis, Virginia Polytechnic Institute and State University, Blacksburg, VA, USA. (1998) [12] Sun, F.P., Chaudry, Z., Rogers, C.A. Majmundar, M. And Liang, C. “Automated Real-time Structure Health Monitoring Via Signal Pattern Recognition.” In: Proceedings of SPIE, February 27 to March 1, San Diego, California, USA, Vol. 2443, pp, 236-247. (1995).

This paper discussed the application of ANN for Structural Strength Measurement. The basics of ANN were discussed and then applied for a strength gain of concrete cube during period of curing (28 days). A MATLAB program on ANN was written and it successfully predicted the strength of the signal by classifying it in the appropriate class (with an efficiency of 80.28% on the training data and 80% on the test data. REFERENCES [1] D.A. Pomerleau, “Efficient Training of Artificial Neural Networks for Autonomous Navigation” (1991) [2] K. Cheikh and F.Soltani, “Application of Neural Networks to radar signal detection in K-distributed Clutter” IEEE (2006) [3] Bimal K. Bose, “Artificial Neural Network Application in Power Electronics”, in 27th Annual Conference of the IEEE Industrial Electronics Society (2001). [4] Yuhong Li, Weihua Ma, “Application of Artificial Neural Networks in Financial Economics” (2010). [5] Hiroaki Nakanishi and Koichi Inoue, “Development of

© 2012 ACEE DOI: 02.ADCS.2012.1. 518

50


Turn static files into dynamic content formats.

Create a flipbook
Issuu converts static files into: digital portfolios, online yearbooks, online catalogs, digital photo albums and more. Sign up and create your flipbook.