Hardware Implementation of K-means Clustering Based Spectrum Sensing Using USRP in a Cognitive Radio System

Transcription

1 Hardware Implementation of K-means Clustering Based Spectrum Sensing Using USRP in a Cognitive Radio System Anirudh Agarwal 1, Himanshu Jain 1, Ranjan Gangopadhyay 2 and Soumitra Debnath 2 Department of ECE The LNM Institute of Information Technology Jaipur, India 1 {a1anirudh, himanshujain.lnmiit}@gmail.com, 2 {ranjan_iitkgp, soumitra_deb@yahoo.com Abstract An accurate detection of spectrum opportunities is a key factor in governing the efficient spectrum usage in a cognitive radio (CR) system. Energy detection based spectrum sensing has been widely used due to its ease of implementation with lower computational complexity; however, its robustness and performance are highly affected by the noise uncertainty. In the present work, a real time hardware implementable spectrum sensor has been realized and tested for an unsupervised learning based K-means clustering approach, to detect the white spaces in the spectrum. A CR network with one primary transmitter and two secondary nodes has been considered for which the data is collected over an FM band using a software defined radio peripheral, i.e. USRP B210. The whole system has been implemented with the help of MATLAB Simulink & Xilinx System Generator. The decision accuracy of the proposed algorithm is verified at different values of the signal-to-noise ratios (SNRs) and found that the classification based sensing is quite accurate even at low SNR region. Keywords- machine learning; K-means clustering; real time hardware implementation; USRP; dynamic spectrum access; spectrum sensing. I. INTRODUCTION In the recent past, CR technology has attracted significant attention towards detecting the vacant spaces in the spectrum and intelligently utilize them without creating harmful interference to the primary (licensed) incumbents. Spectrum sensing (SS) is, therefore, the most important function of CR for accurate detection of white spectrum. So far, several SS techniques have been discussed in the literature, among which Energy detection is the most widely used, due to its simplicity in implementation and probably the lowest computational complexity [1]. But, unfortunately, in low SNR regimes, its performance is drastically degraded. In order to combat the noise uncertainty, eigen value based sensing method has been proved to be robust; however, it is mathematically more complex to realize and implement [2]. Apart from these two, there are other cumbersome spectrum sensing techniques which have been investigated but almost of them are not blind and require knowledge about the primary user (PU) activity [3-5]. Moreover, the performance of every SS technique is severely affected by the inaccurate choice of the decision threshold. A decision threshold cannot be fixed, unlike the case in most of the available methods of SS, and needs to be dynamically updated depending upon the nature of the PU traffic and channel conditions. This calls for an online learning method which would adaptively set the threshold in accordance with the data received at the secondary node. Machine learning (ML) algorithms have plethora of applications ranging from problems like classification and clustering to problems like modeling and prediction. Consequently, it is recognized as a powerful tool for a CR system as well opening up new possibilities of versatile dynamic spectrum allocation (DSA) applications viz. spectrum sensing, spectrum occupancy modeling, spectrum prediction etc [6]. In supervised ML, input and output datasets are provided to train the learning model while there is no provision of the information of output in the unsupervised ML and hence the model has to take its decision exclusively based on the input. Spectrum sensing can be easily thought of as a binary classification problem. In case of unsupervised learning based SS, once the model is trained, the decision threshold is found, which is further used for testing. The primary advantage of ML over other existing SS techniques is that it is an adaptive process of training a system with the efficacy and simplicity of implementation. Moreover, it doesn't require prior knowledge the PU signal, so it is a blind SS technique. There has not been much work done in the context of ML based SS in the existing literature. In [7], authors have used Support Vector Machines (SVM) and weighted K-nearest neighbor (KNN) techniques for SS in a cooperative scenario. Various ML techniques have been proposed and implemented in [8] i.e. K-mean clustering, Gaussian Mixture Model, SVM and KNN but all are the results are shown based on simulations. K-mean clustering has been used for cooperative SS in generalized к-μ fading channel in [9]. In this paper, we have realized the hardware implementation of K-mean clustering based unsupervised learning technique for SS in a CR system. To the best of authors' knowledge, the present work is the first real-time hardware implementation of an SS technique which is based on ML. We have utilized two Universal Software Radio Peripheral (USRP) devices for receiving the

2 real time data from an FM transmitter station, thereby investigating a 1x2 SIMO system with one primary transmitter and two secondary receivers. Further processing is done using MATLAB/Simulink software and for FPGA implementation, Xilinx System Generator is used. The remainder of this work is organized as follows: Section II discusses about the SS system model, a brief mathematical description of K-mean clustering algorithm, and its experimental setup. Data collection and hardware implementation are described in Section III and IV respectively. The performance analysis and the results are presented in Section V. Finally, Section 6 concludes the paper. II. SYSTEM MODEL In our model, we have considered a CR network with 1x2 SIMO system, consisting of one PU transmitter and two secondary user (SU) receivers. Let A be the state of the PU transmitter. Correspondingly, A = 0 and A = 1 denote the OFF state and ON state. This in turn suggests that the channel availability, C is given by: 1, = 1; = 1, = 0 where C = -1 indicates that the channel is not available and C = 1 indicates that the channel is available. For multiple receiving nodes, we use a cooperative spectrum sensing based OR fusion rule [4]. So, in case of K-mean clustering, the classifier determines the energy of the samples received at both the receivers and depending upon the clusters, decides whether the channel belongs to available class or unavailable class. The normalized energy received at the k th node, E k is given by: (1) = ( ) (2) where N is the total number of samples in a particular window of observation, S k(j) is the j th received sample at the k th SU node, which is defined as: ( ) = h ( ) ( ) + ( ) (3) where h k(j) is j th channel coefficient, R(j) denotes the j th transmitted signal and N k(j) is the j th sample of an additive white Gaussian noise (AWGN). A. System Level Hardware Description: Real time signals are received at two SU nodes, in the form two USRP B210 devices, from FM transmitter station located in the city. The receiver end system architecture is shown in the Fig. 1. B. K-means clustering based classification For the purpose of classification, the received energy vectors are sent to the K-means clustering classifier, where they are used as the feature vectors (term used in machine learning) for training and get partitioned into K disjoint clusters with different centroids. The procedure for constructing the classifier is as follows: 1. Several energy vectors, E m are used for training the model where m = 1,2,3,..., M; M is a sufficiently large integer, where E = [E 1, E 2, E 3,..., E N]. 2.Once the training is done by K-means clustering algorithm, the location of the centroids is estimated for different clusters. 3.The classifier next predicts the channel availability using the test energy vectors. Let G r denote the centroid of the r th cluster where r=1, 2, 3,..., K with an assumption that G 1 is the average of E m R given that A = 0 due to the fact that the actual cluster will contain the noise only samples. However, the other clusters G r are defined as the average of all the feature vectors used for training in the set µ r where µ r = { E m E m Cluster r m}: = 1 (4) The K-means clustering algorithm finds the clusters based on the following optimization problem:,,, (5) i.e. minimization of the summation of the within cluster squared distance from all the points from their corresponding clusters summed over all the clusters. The whole algorithm is described in Algorithm 1. ALGORITHM 1 1. G 1 = Mean of E, conditioned on A = 0 2. G r : random initialization with r = 2,3,4,...,K 3. while G r is changed in the previous iteration do 4. Assign the closest corresponding cluster to every data point such that Eq. (5) is satisfied 5. = 2,3,4,, end while Once the training phase is over, i.e. all the clusters have been assigned the appropriate centroids (g i, i = 1,2,3,...,K), the prediction based on the test energy vectors (e) is carried out with the help of the ratio (λ) of the distance between the test energy vector & g 1 and the minimum distance between that test energy vector with all other centroids, i.e. λ,,,, (6)

3 USRP B210 Rx K-means classifier USRP Rx front end Figure 1: Receiver end System Level Architecture If eq. (6) is satisfied, then "e" is classified as the channel unavailable class, otherwise it belongs to the available class.the significance of λ (> 0) lies in tuning the values of the error probabilities, i.e. probability of false alarm and probability of misdetection. III. REAL TIME DATA COLLECTION In the present implementation, a real FM band signal with a carrier frequency of 94.3 MHz is received from the FM transmiter station located at Jaipur city, Rajsthan, India. The receiver consists of a bus series Universal Software Radio Peripheral (USRP B210), capable of receiving RF signals, which is interfaced with a host computer employing Software simulations through MATLAB/Simulink and Hardware cosimulation though Xilinx System Generator for DSP which is a part of Xilinx Vivado System Edition Design suite. The receiver end system level architecture has been shown in the Fig. 1. Ettus Research product USRP B210 hardware covers a frequency range of 70 MHz to 6 GHz Xilinx Spartan 6XC6SLX150 FPGA, enabling wide range of experimentation in TV/FM/Cellular/Wi-Fi bands. This transceiver has 2x2 MIMO capability which can stream up to 56 MHz of instantaneous bandwidth in 1x1 and that up to MHz in 2x2. It has an in-built daughterboard where the received signal is digitally processed. However, in Fig. 1, the daughterboard has been exclusively shown. Moreover, USRP B210 can easily be interfaced with various software platforms like GNU Radio and MATLAB/Simulink. The detailed steps for collecting the data and further processing are enumerated as follows: Step1: Firstly, an FM Transmitter model is developed in Simulink for the two USRPs. Step 2: Simultaneously, through a MATLAB script, the complex data received from USRPs is stored in two column vectors which are further converted into corresponding energy vectors. Step 3: Now, half of the data is passed to the K-means clustering classifier for training the model. Step 4: After training, the remaining data is used for testing; thereby verifying if the model is able to correctly classify the input data between channel available class and channel unavailable class. IV. HARDWARE IMPLEMENTATION USING XILINX SYSTEM GENERATOR The whole K-means clustering based spectrum sensing architecture implemented using MATLAB Simulink, with the help of Xilinx System Generator and Vivado High Level Synthesis (HLS), has been presented in Fig. 2. The motivation Figure 2: K-means clustering based Spectrum Sensing Architecture using Simulink and Xilinx System Generator for Hardware Implementation

4 behind using these tools is mainly the complexity of the algorithm under consideration. They, together, provide high flexibility and at the same time, simplicity with robustness in designing such systems, which are to be ultimately implemented over the FPGA (present on USRP B210, in this work) through hardware co-simulation [8]. Once the USRP is interfaced with MATLAB/Simulink, the RF signal complex data is recorded for further processing to get the energy vectors. This data is now fed to Xilinx System Generator block sets. A. Hardware Architecture/Design Description The blocks, colored in blue viz. Accumulator antenna 1, Accumulator antenna 2, K-means clustering algorithm for training and that for testing, which are used in Xilinx System Generator have been manually designed due to unavailability of the required blocks. For this purpose, Vivado HLS tool is utilized, where the required architectural blocks are coded in C++ language and then converted into RTL thereby exporting it to Xilinx System Generator as a System Generator input (IP) block. Those IPs are then connected and configured to give the final architecture. However, the blocks, colored in yellow, are the Xilinx Gateway In and Gateway Out blocks in between which the actual processing in System Generator is carried out. RF complex data is accumulated through two receiver antennas and converted into energy vectors in the Accumulator Antenna 1 & 2 blocks respectively. The energy vectors are further forwarded to the Algorithm 1 block. Some control signals have been used to indicate the valid energy samples and the intermediate delays are used for time synchronization. Moreover, there are provisional buffers in the algorithm block, which store the incoming energy samples until the number reaches a specified limit. The algorithm starts the training process over the received data and generates two centroids for the two clusters. The control signal done indicates the completion of the process. As soon as the training is completed, the final block Testing starts Figure 3: Unclustered training data with 500 sample points of received energy vectors at both antennas classifying the test data based on the computed centroid information through training. The results of the testing are stored and exported as MATLAB workspace variable for further calculation of the performance metrics. B. Real Time Model Verification (Testing phase): After training, for model verification, the USRP is tuned to a busy/idle FM band thereby checking the final result of channel availability in the form of binary decision. For this purpose, initially the energy of the input data samples is calculated. Subsequently, the distance of each energy vector is determined from each centroid and the centroid which is closer, is selected as the correct one assuming that one of the centroids is exclusively the average of noise only samples, i.e. the idle channel samples. V. PERFORMANCE EVALUATION, RESULTS AND DISCUSSION In this study, after the hardware implementation, the results have been obtained after collecting output information in the MATLAB software workspace. The total number of energy vectors that have been used in the present work is 1000, among which half is utilized for training while the remaining half is kept for testing the trained model. The size of each energy vector, i.e. the number of samples received at one SU node, N is taken as The data has been tested for three different SNR values, i.e db, - 4 db and - 8 db. The typical values of SNR are changed by varying the distance between the transmitter and the two receivers, then taking the corresponding average SNR. For this, we initially observe the received power when receiver antennas are tuned over transmitting frequency (case of signal + noise), then observe the same over the frequency where there is no transmission (noise only case). The receivers are kept at sufficiently apart so as to avoid any inter-channel interference. However, the effect of inter-receiver distance over the accuracy of the implemented sensing algorithm has not been covered in this study in order to limit the length of the paper. In Fig. 3, the normalized energy vectors at receiver antenna 1 and receiver antenna 2 are shown respectively at X- axis and Y-axis for SNR = - 4 db. The clustered version of the same dataset after training through K-means clustering classifier is depicted in Fig. 4. The data points colored in blue denote the channel unavailable class while the red colored data points belong to the channel available class. It is clearly visible that even in the real time scenario, the clustering is well performed in that both the clusters for noise only samples and combined signal with noise samples are quite accurately classified. Moreover, the performance of any spectrum sensing algorithm is best evaluated based on the receiver operating characteristics (ROC) curve. Fig. 5 shows the variation of probability of detection (P d) against probability of false-alarm (P fa) at the considered three values of SNR. P fa is a non-linear function of decision-threshold in any spectrum sensing algorithm [9], which in this case is mainly dependent over λ. The value of λ is varied for getting different values of P fa.

5 Normalized energy at Receiver Antenna 2 P d Clustering at SNR=-4 db Noise only samples Signal+Noise samples Centroid (channel available) Centroid (channel unavailable) Normalized energy at Receiver Antenna 1 Figure 4: Clustered data after training with two centroids SNR=0.4 db SNR=-4 db SNR=-8 db P fa Figure 5: Comparison of ROC for K-means clustering based Spectrum Sensing for 3 different SNRs Further, in a real-time situation, it is cumbersome to calculate P d through a closed-form mathematical expression. Hence, for a fixed value of P fa and SNR, P d is determined as the ratio of the number of instances when the test statistic is above threshold and the total number of instances. In Fig. 5, in accordance with the theoretical results of spectrum sensing [9], P d increases with increase in P fa at fixed SNR. In fact, the performance of the algorithm is fair enough at the SNR as lowas -8 db and that too in the presence of fading and various multipath effects when the whole real time system is implemented on hardware. However, some of the curves are not as smooth as simulation based results but that is due to the fluctuations in SNR values, fading and other distortion effects in real time case. TABLE I. Resource Utilization Table Name of the block BRAM_18K DSP48E FF LUT Accumulator Algorithm Testing Available A. Resource Utilization Table: The hardware resources utilized by each of the three blocks, i.e. Accumulator, Algorithm 1 and Testing are respectively listed in Table 1. It can be seen that only the algorithm block uses memory units (BRAM_18K) as it holds the incoming stream of data in the form of buffers for processing.clearly, the overall hardware resources utilized by the system whether in terms of memory units, signal processing units (DSP48E), flipflops (FF) or the look-up tables (LUT) are very less as compared to their availability and the system can still be realized quite easily and at a lower cost. The resources utilized depend on the size of training/test data for the algorithm. It increases with increase in data size. In our case, the total size of the data is , which is fairly sufficient for most applications. Similarly, the time complexity of processing of the blocks used in the system architecture is characterized by latency and the time gap one has to maintain between the two consecutive inputs. Based on these two parameters, the time taken by the algorithm to train the model is found to be microseconds. It is to note that the proposed clustering algorithm has been implemented utilizing FM band but it can be extended to any other band of RF spectrum with even more involvement of PU traffic dynamics. VI. CONCLUSION In this paper, we have dealt with the real time hardware realization of a machine learning based K-means clustering spectrum sensing technique which is much more robust and efficient in terms of noise uncertainty as well as computational complexity as compared to the conventional spectrum sensing algorithms. It has been found that the considered the algorithm performs significantly well even in the low SNR scenario. Moreover, the algorithm has an appreciable practical advantage in that its time and space (memory) complexity is remarkably low along with the ease of implementation on hardware with very less utilization of resources. In the future work, we aim to enhance the accuracy of the algorithm with more number of primary transmitters and receiving nodes in a real time case. In this work, we have assumed that the SU receiver sensitivity is optimum and would not affect the spectrum sensing decision. We would try to do away with this assumption and also extend the present hardware implementation to Linux based platforms like GNU

6 radio which is exclusively designed for the experimentation of USRP based applications. ACKNOWLEDGEMENT The authors acknowledge the financial support received from Information Technology Research Academy (ITRA), Department of Electronics and Information Technology (DeitY), Ministry of Communications & IT, Govt. of India. REFERENCES [1] Cordeiro, C., and K. Challapali. "Spectrum agile radios: utilization and sensing architectures." First IEEE International Symposium on New Frontiers in Dynamic Spectrum Access Networks, DySPAN IEEE, " [2] Zeng, Yonghong, and Ying-Chang Liang. "Maximum-minimum eigenvalue detection for cognitive radio." 2007 IEEE 18th International Symposium on Personal, Indoor and Mobile Radio Communications. IEEE, [3] Yucek, Tevfik, and Huseyin Arslan. "A survey of spectrum sensing algorithms for cognitive radio applications." IEEE communications surveys & tutorials 11.1 (2009): [4] Azmat, F., Yunfei Chen, and N. G. Stocks. "Analysis of spectrum occupancy using machine learning algorithms." IEEE Transactions on Vehicular Technology (2015). [5] Thilina, Karaputugala Madushan, et al. "Pattern classification techniques for cooperative spectrum sensing in cognitive radio networks: SVM and W-KNN approaches." Global Communications Conference (GLOBECOM), 2012 IEEE. IEEE, [6] Thilina, Karaputugala Madushan, et al. "Machine learning techniques for cooperative spectrum sensing in cognitive radio networks." IEEE Journal on selected areas in communications (2013): [7] Kumar, Vaibhav, et al. "K-mean clustering based cooperative spectrum sensing in generalized к-μ fading channels." Communication (NCC), 2016 Twenty Second National Conference on. IEEE, [8] Chaitanya, G. V., P. Rajalakshmi, and U. B. Desai. "Real time hardware implementable spectrum sensor for Cognitive Radio applications." 2012 International Conference on Signal Processing and Communications (SPCOM). IEEE, [9] Atapattu, Saman, Chintha Tellambura, and Hai Jiang. Energy detection for spectrum sensing in cognitive radio. Berlin: Springer, 2014.