Activity Pattern Aware Spectrum Sensing A CNN-Based Deep Learning Approach
Abstract: In cognitive radio (CR), most spectrum sensing algorithms are model-based and their detection performance relies heavily on the accuracy of the assumed statistical model. In this paper, we propose a convolutional neural network (CNN)based deep learning algorithm for spectrum sensing. Compared with model-based spectrum sensing algorithms, our proposed deep learning approach is data-driven and requires neither signal noise probability model nor primary user (PU) activity pattern model. The proposed algorithm simultaneously takes in the present sensing data and historical sensing data, with which the inherent PU activity pattern can be learned to benefit the detection of PU activity. With extensive numerical simulations, results show that the proposed algorithm outperforms the estimator correlator (E-C) detector and the hidden Markov model (HMM)- based detector in term of correct detection probability. Existing system: The APASS algorithm is data-driven and requires neither signal-noise probability model nor PU activity pattern model. Although DNN’s applications on spectrum
sensing have been studied in existing works, few of them focus on learning PU activity pattern out of sensing data. To make the APASS algorithm capable of exploiting historical PU activity data to improve detection accuracy, we assign two inputs to the deep learning framework: a) the covariance matrix (CM) generated by the sensing data at the present frame for state classification; b) a big matrix constructed by the CMs from the past frames to learn PU activity pattern. As these inputs can be considered as images, for the deep learning structure, we adopt the convolutional neural network (CNN) which excels at image classification. The APASS algorithm consists of two phases: an offline training phase and an online detection phase. Proposed system: A big matrix constructed by the CMs from the past frames to learn PU activity pattern. As these inputs can be considered as images, for the deep learning structure, we adopt the convolutional neural network (CNN) which excels at image classification. The APASS algorithm consists of two phases: an offline training phase and an online detection phase. In the offline training phase, the CNN takes in the present CM, historical CMs and labeled PU state data to train its parameters and in the online phase, the trained CNN can make real-time detection according to the present and historical sensing data. The advantages of the proposed APASS algorithm are three-fold: 1) it can learn the PU activity pattern with no assumption of the PU state dwell time distribution; 2) it assumes no noise or signal model; 3) it has robust performance in low SNR scenarios. Extensive simulations have been done with uncorrelated and correlated signal models and in both cases; our proposed APASS detector outperforms the existing E-C detector and the HMMbased detector with obvious margin. Advantages: The advantages of the proposed APASS algorithm are three-fold: it can learn the PU activity pattern with no assumption of the PU state dwell time distribution; it assumes no noise or signal model; it has robust performance in low SNR scenarios. Extensive simulations have been done with uncorrelated and correlated signal models and in both cases; our proposed APASS detector outperforms the existing E-C detector and the HMM-based detector with obvious margin.
Disadvantages: The APASS algorithm is data-driven and requires neither signal-noise probability model nor PU activity pattern model. Although DNN’s applications on spectrum sensing have been studied in existing works, few of them focus on learning PU activity pattern out of sensing data. To make the APASS algorithm capable of exploiting historical PU activity data to improve detection accuracy, we assign two inputs to the deep learning framework: the covariance matrix (CM) generated by the sensing data at the present frame for state classification. Modules: Blindly combined energy detector: SPECTRUM sensing, which detects the transmission status of the primary users (PUs), is of utmost importance to constrain secondary users’ (SUs) interference on PUs in cognitive radio (CR). In the past decades, many spectrum sensing algorithms have been proposed, including, e.g., the optimal estimator-correlator (E-C) detector, the semi blind energy detector (ED), and the blindly combined energy detector (BCED), et al. One weakness of these detectors is that they do not exploit the historical information of the PU state. In CR, it has been proposed in that the temporal relation between PU states can be depicted by the Markov process or semi-Markov process (SMP) and the hidden Markov model (HMM) has been applied in spectrum sensing to harness such temporal relation to improve the spectrum sensing performance. Nevertheless, HMM-based spectrum sensing methods heavily depend on the availability and the accuracy of statistical models, which may be a strong assumption in the real environment. Convolutional neural network: The APASS algorithm is data-driven and requires neither signal-noise probability model nor PU activity pattern model. Although DNN’s applications on spectrum sensing have been studied in existing works, few of them focus on learning PU activity pattern out of sensing data. To make the APASS algorithm capable of exploiting historical PU activity data to improve detection accuracy, we assign two inputs to the deep learning framework: a) the covariance matrix (CM) generated by
the sensing data at the present frame for state classification; b) a big matrix constructed by the CMs from the past frames to learn PU activity pattern. As these inputs can be considered as images, for the deep learning structure, we adopt the convolutional neural network (CNN) which excels at image classification. The APASS algorithm consists of two phases: an offline training phase and an online detection phase. Training data model: Most existing PU detection algorithms are model-driven. For example, the E-C detector assumes both signal and noise are Gaussian and assumes known mean and covariance matrix for both signal and noise. For model-based algorithms, their detection performance relies on the accuracy of the assumed model and once the model mismatches with the real incoming data, serious performance decline can be incurred. In the era of big data, data-driven PU detectors have been proposed. For data-driven detectors, there are two phases, i.e. the off-line training phase and the on-line detection phase. The advantage of data-driven methods is that the detector is learned from a training process extracting inherent data pattern and, in this way, the model mismatch problem can be alleviated. In this paper, we consider using D sets of training sensing data and each data set consists of the sensing signal as well as the ground truth PU activity labels in U consecutive frames.