Online modulation recognition of analog communication signals using neural network

Transcription

1 Expert Systems with Applications Expert Systems with Applications 33 (7) Online modulation recognition of analog communication signals using neural network H. Guldemir *, A. Sengur Firat University, Technical Education Faculty, 39 Elazig, Turkey Abstract In this paper, a neural network based online analog modulation recognition of communication signals is presented. The proposed system can discriminate between amplitude modulation (), frequency modulation (), double sideband (), upper sideband (), lower sideband () and continuous wave (CW) modulations. A matlab graphical user interface (GUI) is designed to see the intercepted signal, its power spectral density, frequency and modulation type on the screen of the personnel computer. To achieve correct classification, extensive simulations have been done for training the neural network. Theoretical simulations and experimental results indicate good performance even at signal-to-noise ratios as low as 5 db. Ó 6 Elsevier Ltd. All rights reserved. Keywords: Modulation; Recognition; Neural network; Classification; Key features. Introduction * Corresponding author. Tel.: x488. addresses: hguldemir@firat.edu.tr (H. Guldemir), ksengur@ firat.edu.tr (A. Sengur). The modulation type is one of the most important variables of the communication signals. Automatic modulation recognition is the detection of random modulated signals without a priori information to determine the type of modulation used. The automatic modulation recognition of communication signal is a valuable tool in both military and civilian purposes, especially in electronic surveillance systems. In the past, radar systems and communication systems have relied on operator interpretation of measured parameters to provide classification and identification of emitters (Hsue & Soliman, 99). These non-automatic recognition procedures are subjected to operator particular conditions and they are to slow in a hostile environment (Dominguez, Borello, Garcia, & Mezaua, 99). For fast response, automatic processing techniques are required for the recognition systems. The problem of replacing operator interpretations by microcomputer based recognition has recently received attention and new recognition procedures based on time averaged behaviour of instantaneous envelope, frequency and zero-phase have been applied to the received signals (Aisbett, 997). Several modulation recognition procedures have been developed for various applications (Azzous & Nandi, 996a; Chan & Gadboia, 989; Dubuc, Boudreau, Patenaude, & Inkol, 999; Guldemir & Sengur, 6; Jondral, 985; Mammone, Rothaker, & Podilchuk, 987). Most of these techniques are used to classify,, SSB, and signals. Automatic modulation recognition can be accomplished in several different ways such as to use intelligent decision algorithms and to use spectral analysis techniques together with some form of statistical comparisons. In the former, a probabilistic and hypothesis testing arguments are employed to formulate the recognition problem. In the latter, the recognition system is composed of two subsystems. The first is the feature extraction subsystem, which extracts predefined features from the incoming signal. The second one is the pattern recognition subsystem, which finds the modulation type of the received signal. In an effective surveillance system the classification and modulation type of the communication signal have to be determined quickly and precisely in real time in order to /$ - see front matter Ó 6 Elsevier Ltd. All rights reserved. doi:.6/j.eswa.6.4.5

2 H. Guldemir, A. Sengur / Expert Systems with Applications 33 (7) find the location and the source of the signal. This is achieved with modern software and hardware technology. This paper presents such an approach, with emphasis on the problem of modulation recognition. The approach is based on pattern recognition technique and use neural network (NN) classification. The key issue is to identify a set of meaningful features such as amplitude variation, spectrum symmetry and instantaneous phase. The proposed system is able to classify the incoming modulation type and can be used to choose an appropriate demodulator to be used. Alternative uses include spectrum management and surveillance, military threat evaluation, source and interference identification, etc. The use of neural network in this paper provides a quick and robust solution to the modulation recognition. A multi-layer neural network is being used. Algorithms for feature extraction are developed, so that the most efficient and descriptive feature set is chosen. The effectiveness of the proposed scheme is verified by using both theoretically produced and real analog modulated signals. The results indicate about 99% accuracy. The results, showing the effectiveness of the proposed algorithm are presented.. Representation of analog modulated signals A modulated signal s(t) can be expressed by a function of the form sðtþ ¼a c aðtþ cosðpf c t þ uðtþþh Þ where a(t) is the signal envelope, f c is the carrier frequency, u(t) is the phase, h is the initial phase and a c controls the carrier power. Particular modulation types are obtained by encoding the base band message into a(t) and u(t). In the following the mathematical expressions for different types of analog modulated signals (,, SSB, ) are given together with their amplitude and phase relationships. These relationships will later help to classify the distinct type of the modulations... Amplitude modulation () The expression describing an amplitude modulated signal is given by sðtþ ¼½ þ mxðtþ cosðpf c tþš where m is the modulation index, x(t) is the modulating signal and f c is the carrier frequency. The instantaneous amplitude a(t) and phase for this signal is aðtþ ¼j þ mxðtþj and the instantaneous phase u(t) is /ðtþ ¼pf c t ðþ ðþ ð3þ ð4þ.. Double side-band modulation () Double side-band modulated signals can be expressed as sðtþ ¼xðtÞ cosðpf c tþ ð5þ The instantaneous amplitude and phase are given by aðtþ ¼jxðtÞj ð6þ /ðtþ ¼ pf ct if xðtþ > ð7þ pf c t þ p if xðtþ <.3. Single side-band modulation (SSB) Single side-band modulated signals can be expressed as sðtþ ¼xðtÞ cosðpf c tþyðtþsinðpf c tþ ð8þ where x(t) is the modulating signal, y(t) is the Hilbert transform. The negative sign is used for upper side-band () signal generation and the positive sign is used for lower side-band () signal generation. Using the harmonic analysis the SSB signal can be expanded as given by (Azzous & Nandi, 996b) sðtþ ¼ XN x i cosðpðf c f i Þt þ w i Þ ð9þ i¼ and the instantaneous amplitude and phase is given by vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ux N aðtþ ¼ x i þ XN X t N x i x j cos½pðf i f j ÞtŠ; ðþ i¼ i¼ j¼ /ðtþ ¼tan P N i¼ x i sin½pðf c þ f i Þt þ w i Š P N i¼ x i cos½pðf c þ f i Þt þ w i Š.4. Frequency modulation () ðþ In frequency modulation the instantaneous frequency is varied linearly with the modulating signal x(t). The frequency modulated signal is written as Z t sðtþ ¼cos pf c t þ K f xðsþds ðþ where K f is the frequency deviation coefficient. 3. Characteristics of analog modulated signals Examination of instantaneous amplitude and phase expressions for the amplitude modulated signals given in the previous section provides information for the discrimination of the modulation types. Since the envelope and instantaneous frequency variations from its mean are independent of the carrier frequency, parameters obtained from envelope and instantaneous frequency may lead to separation of the modulation types. The first fundamental characteristic is the presence of significant amplitude variation. This characteristic allows an initial discrimination between and signals. The following criteria can be deduced for the modulation types.

3 8 H. Guldemir, A. Sengur / Expert Systems with Applications 33 (7) 6 4 and CW signals have constant envelops, whereas and SSB () signals will have variable envelops. A CW signal has a constant phase while signals have variations in their instantaneous phase. The distinction between CW and signals can be obtained by comparing the value of the variance of the phase to a phase threshold, for the samples having amplitude above their mean. and CW signals have constant instantaneous frequencies, while and SSB signals will have variations in their instantaneous frequencies. Fig. shows the instantaneous amplitude, phase and spectrum of,,, and SSB signals. Instant. Spectrum Instant. Phase Instant. Amp. (a) Frequency(Hz) Instant. Spectrum Instant. Phase Instant. Amp. (b) Frequency(Hz) Instant. Spectrum Instant. Phase Instant. Amp. (c) Frequency(Hz) Instant. Spectrum Instant. Phase Instant. Amp. (d) Frequency(Hz) Fig.. Instantaneous amplitude, phase and spectrum of an analog modulated communication signal. (a) ; (b) ; (c) ; (d) SSB.

4 H. Guldemir, A. Sengur / Expert Systems with Applications 33 (7) Table Variations in the analog modulated signals Amplitude variation Frequency variation Yes No No Yes SSB Yes Yes No No In the following table all the modulation types to be determined are listed to show variations either in their instantaneous amplitude, frequency or both (Table ). The and signals can be distinguished by examination of the instantaneous frequency variations from the mean value and in particular the statistics of the positive and negative peaks. It can also be shown that an instantaneous frequency will have a greater proportion of negative spikes than positive spikes. Conversely, an instantaneous frequency should have more positive spikes than negative (Al-Jalili, 995). 4. Feature extraction In order to reduce the amount of information to the input of the neural network, a process called feature extraction, which computes a small number of salient key features from the raw data is used. These key features are sensitive to differences in modulation schemes and insensitive to signal-noise ratios. The modulation recognition algorithm depends on extracting samples of the instantaneous amplitude and frequency of the received unknown signal. These features are characteristics of the different modulated signal, and hence examination of the instantaneous amplitude and frequency lead to a successful recognition. In the modulation recognition the following key features which are same as those used in (Kremer & Shiels, 997) were used. The feature used to determine the envelope variations is the maximum of the squared Fourier transform of the normalized signal amplitude, defined as c max ¼ max jfftðaþj N where a is the normalised centred instantaneous amplitude of the intercepted signal and N is the number of the sample in the range. This feature is a measure of envelope and allows the reliable separation of non-constant envelope signals from constant envelope signals (CW, ). max dp SNR (db) SNR (db) P SNR (db) Fig.. Variation of c max, r dp and P values for analog modulated signals.

5 H. Guldemir, A. Sengur / Expert Systems with Applications 33 (7) 6 4 The second feature is used to determine spectrum symmetry around the carrier frequency. This is based on the spectral powers for the lower and upper sidebands of the intercepted signal and it is defined by P ¼ P L P L P L þ P L where P L ¼ Xfcn i¼ P L ¼ Xfcn i¼ jx c ðiþj jx c ði þ f cn þ Þj where X c (i) is the Fourier transform of the intercepted signal, (f cn + ) is the sample number corresponding to the carrier frequency f c and f cn is defined as f cn ¼ f cn f s The third feature is obtained by extracting information from the instantaneous phase of the intercepted signal. This information is obtained from the standard deviation of the direct value of the centred non-linear component of the instantaneous phase, given by vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi!! X r dp ¼ / NL C ðiþ u X t / C NL ðiþ a nðiþ>a t a nðiþ>a t where / NL (i) is the value of the centred non-linear component of the instantaneous phase at time instants t = i/ f s, C is the number of the samples in / NL (i), and a t is the threshold. The estimation of instantaneous phase is very sensitive to noise below this threshold. The key features c max, r dp and P are plotted in Fig. for all types of the analog modulated signals to be classified in this study against the SNR. The key feature c max represents the spectral density maximum and is zero for and large for all types of modulated signals. r dp is a measure of standard deviation of the centred non-linear component of direct instantaneous phase and can be used to discriminate and signals. The key feature P represents the spectrum symmetry and is used to separate the and signals from the other analog modulated signals even at very low SNR. A decision theoretic algorithm using the three key features can be made for modulation classification using the flowchart given in Fig. 3. In this paper, due to its powerful classification capability a neural network classifier is used. 5. Neural network modulation recognition Modulation recognition of an intercepted signal is realized by a classification algorithm. A feed forward multilayer neural network classifier is used in this study. The parameters of the neural network structure are given in Analog Modulated Signal y(i) i=,...,n s Instantaneous Amplitude max FFT ( y( i) / ma ) γ max = Ns Yes γ max < t( γ max ) No Spectrum Symmetry PL PU P = PL + PU Yes P <t(p) No P< Yes No Instantaneous Phase σ dp = ( ) ( ) φnl i φnl i C C An ( i) > ta An ( i) > ta S S B γ max σ dp P CW Fig. 4. ANN architecture of automatic modulation recognition system. Performance 9.795e-3, Target e- 5-5 Training σ dp < t( σ dp ) No - -5 Target Yes Number of iteration Fig. 3. Flowchart for classification of analog modulated signals. Fig. 5. Performance of the NN training.

6 H. Guldemir, A. Sengur / Expert Systems with Applications 33 (7) 6 4 Analog Signal Data Acquisition and Preprocessing Feature Extraction Neural Network Classification Modulation Type Fig. 6. Subsystems of the modulation recognition. Fig. 7. Screenshots of the designed Matlab GUI. Appendix. The neural network is trained by Levenberg and Marquardt training algorithm and the tangent sigmoid activation function is used. The network is organized by an input layer, a hidden layer and an output layer. Fig. 4 shows the neural network structure used in this study. The number of the input layer neurons is determined by the dimension of the feature vector. Three key features c max, r dp and P given in previous section, are used as the input to the neural network classifier. The number of output layer neurons is determined by the number of modulation types to be determined. In this study, the six outputs correspond to the modulation scheme categories:,,,, and CW. The input layer neurons have linear activation functions with unity slope, but there is a scale factor with each input to convert them to per unit (normalisation) signals. The hidden layer functions to associate the input and output layers. The output signals are converted from per unit signals to actual signals by denormalisation. The normalisation of the datasets is used to reduce the training and learning time. The number of the hidden layer neurons is determined by experiments. A number of tests has been done for choosing an optimum number of hidden layer neurons which will reduce the training time and give higher performance and less sum square error (Sengur & Guldemir, 3). Fig. 5 shows the performance of the neural network training. matching occurs, that is, the error between the desired pattern and actual pattern becomes acceptably small. In order to obtain the correct type of the modulation, the system must be able to classify not only a variety of modulation schemes but also to be able to identify these modulation schemes with different parameters such as: various signal-to-noise ratios, carrier frequencies, data rates, transmitted signals and various modulation parameters (Nandi & Azzous, 997). Because of these reasons, the training continued with a large number of input/output example data patterns. Fifty example data patterns for each modulation type and for each key feature with different modulation type and different modulation indices are used for the training. Totally = 9 test pattern were used. Where 3 is the number of key feature, 6 is the 5.. Training neural network The training of the neural network is done off-line. The training is achieved as follows: For a given input pattern, the actual output pattern is computed and compared with the desired output pattern and weights are adjusted by Levenberg and Marquardt training algorithm until pattern Picture. Online analog modulation recognition system.

7 H. Guldemir, A. Sengur / Expert Systems with Applications 33 (7) 6 4 number of modulation type to be determined and 5 is the number of test pattern with different parameters. 6. Online analog modulation recognition system The online analog modulation recognition system is composed of three main subsystems which are acquisition and pre-processing of the intercepted signals, feature extraction and classification as shown in Fig. 6. In order to process the analog signals in computers, the information carried by the analog signal should be represented in a digital form. The pre-processing block includes the analog/ digital (A/D) conversion which converts analog signals into digital form by a sequence of its instantaneous values measured at discrete time instants. The feature extraction block which gives the feature vectors which are different for different modulation types is the most important block of the overall system. Modulation recognition of an intercepted signal is finally realized by the classification subsystem which classifies the type of the modulation using neural network classification algorithm. Fig. 8. (a) Modulating signal, (b), (c), (d), (e), and (f).

8 H. Guldemir, A. Sengur / Expert Systems with Applications 33 (7) A graphical user interface (GUI) is designed using Matlab and Data Acquisition Toolbox. The developed program receives the analog modulated signals via sound card. After pre-processing, feature extraction and classification, the type of the modulated signal is shown on the GUI together with the original signal waveform, the spectral density of the received signal and the carrier frequency. When the modulation type of the incoming signal or its carrier frequency is changed, the type of the modulation or carrier frequency changes immediately on the GUI showing the correct modulation type or correct carrier frequency online. The Fig. 7 shows the screenshots of the designed Matlab GUI. The system of the automatic modulation recognition of analog communication signals is given in Picture. 7. Theoretical and experimental results Table Performance of neural network based modulation recognition using simulated signals at 5 db SNR Simulated type Deduced modulation type CW Other 98.%.9% 98% % 95.% 4.8% 97.5%.5% 98.5%.5% CW % 99% Table 3 Performance of neural network based modulation recognition using simulated signals at 3 db SNR Simulated type Deduced modulation type CW Other % % 98.%.8% % 99.8%.% CW.4% 99.6% Table 4 Performance of neural network based modulation recognition using real signals at 5 db SNR Real signal Deduced modulation type CW Other 98% % 95% 5% CW 5% 95% Table 5 Performance of neural network based modulation recognition using real signals at 3 db SNR Real signal Deduced modulation type CW Other % 99.5%.5% CW 99.4%.6% A number of simulations have been done with theoretically produced different modulated signals with various SNR, carrier frequencies and various modulation parameters. The modulation types were restricted to the types commonly used in analog communication. The modulation was carried out by using Matlab functions in Communication Toolbox. An additive white Gaussian noise with SNR of between and 4 db is used in the modelling of theoretically produced analog modulated signals. In the simulations, a first degree autoregressive 3 khz band limited speech signal with sampling rate of khz and resampled with 3 khz and modulated by a 5 khz sinusoidal carrier is used. The modulating signal and the,,, and modulated signals with 3 db SNR is shown in Fig. 8. For the experimental verification a B + K Precision Model 34 universal signal generator is used. The signal generator provides outstanding phase-noise performance and analog modulation features for all general purpose test needs. It has a comprehensive analog modulation capabilities including, and CW. A PCL-8PG Enhanced Multi-Lab Card which is a high performance, high speed multi-function data acquisition card for personnel computers, is used to acquire the analog modulated RF signals from the signal generator. The signal generator can apply attenuation of up to 8 db which is necessary when the amplitude of the signal exceeds the data acquisition card limits. The modulation duty control, carrier frequency control and amplitude control can also be done in this signal generator. The length of the screened cable between the computer and signal generator is kept minimum in order to avoid the noise at high frequencies and it is ended by a5x resistor. Using the signals received from the signal generator with different modulation types and different modulation parameters, the neural network classifier is tested. A very good performance has been achieved. The results of the performance evaluation of the neural network based online analog modulation recognition are summarised in Tables and 3 for theoretically simulated signals at 5 db and 3 db, respectively and in Tables 4 and 5 for experimental analog modulated signals at 5 db and 3 db, respectively. It is clear in tables that all types of analog modulations have been correctly recognized with more than 95% success rate at 5 db SNR. If SNR value goes beyond db, the overall success rate of the recognition approaches %. 8. Conclusion An online recognition of the modulation type of analog modulated communication signals using neural network has been presented in this paper. Six analog modulation schemes have been considered and three key features

9 4 H. Guldemir, A. Sengur / Expert Systems with Applications 33 (7) 6 4 extracted from the instantaneous amplitude, the instantaneous phase and the instantaneous frequency of the intercepted signal have been used as inputs to the neural network to perform the classification. The intercepted signal, its power spectral density, carrier frequency and the modulation type have been shown on the designed Matlab GUI which can show instantaneous changes in the intercepted signal. Extensive computer simulations of six analog modulations have been carried out at different signal-tonoise ratios. Experimental work of three analog modulations have also been carried out for the verification of the simulation results. Correct classification has been obtained both from theoretically simulated and real analog modulated signals even at low SNR values. The results presented show the effectiveness of the neural network based classification. Appendix See Table A.. Table A. Neural network structure and training parameters Structure of the neural network Neural network model Multilevel feed forward Number of level 3 Number of neurons in the layers Input: 3; hidden: 7; output: 6 Initial weights Random Activation function Tangent hyperbolic Training parameters Training Levenberg Marquardt algorithm Sum square error References Aisbett, J. (997). Automatic modulation recognition using the time domain parameters. Signal Processing, 3, Al-Jalili, Y. O. (995). Identification algorithm of upper sideband and lower sideband SSB signals. Signal Processing, 4, 7 3. Azzous, E. E., & Nandi, A. K. (996a). Procedure for automatic recognition of analog and digital modulations. IEEE Proceedings Communications, 43, Azzous, E. E., & Nandi, A. K. (996b). Automatic modulation recognition of communication signals. London: Kluwer Academic Publishers. Chan, Y. T., & Gadboia, L. G. (989). Identification of the modulation type of a signal. Signal Processing, 6, Dominguez, L. V., Borello, J. M. P., Garcia, J. P., & Mezaua, B. P. (99). A general approach to the automatic classification of radio communication signals. Signal Processing,, Dubuc, C., Boudreau, D., Patenaude, F., & Inkol, R. (999). An automatic modulation recognition algorithm for spectrum monitoring applications. IEEE international conference on communications (ICC 99). Vancouver, Canada. Guldemir, H., & Sengur, A. (6). Comparison of clustering algorithms for analog modulation classification. Expert Systems with Applications, 3, Hsue, S. Z., & Soliman, S. S. (99). Automatic modulation classification using zero crossing. IEE Proceedings, 37, Jondral, F. (985). Automatic classification of high frequency signals. Signal Processing, 9, Kremer, S. C., & Shiels, J. (997). A test bed for automatic modulation recognition using neural networks. Proceedings of the IEEE Canadian Conference on Electrical and Computer Engineering, Mammone, R. J., Rothaker, R. J., & Podilchuk, C. I. (987). Estimation of carrier frequency modulation type and bit rate of an unknown modulated signal. IEEE international conference on communications (ICC 87) (pp. 6 ). Seattle, WA. Nandi, A. K., & Azzous, E. E. (997). Modulation recognition using artificial neural networks. Signal Processing, 56, Sengur, A., & Guldemir, H. (3). Performance comparison of automatic analog modulation classifiers. Third international advanced technologies symposium. Ankara, Turkey.