Implementation of an Intelligent Target Classifier with Bicoherence Feature Set

Transcription

1 ISSN: International Journal o Innovative Research in Science, (An ISO 397: 007 Certiied Organization Vol. 3, Issue, November 04 Implementation o an Intelligent Target Classiier with Bicoherence Feature Set Mohankumar K., Supriya M. H., P. R. Saseendran Pillai Department o Electronics, Cochin University o Science and Technology, Cochin, India ABSTRACT: This paper examines the easibility o bispectral analysing o acoustic signals emanated rom underwater targets, or the purpose o classiication. Higher order analysis, especially bispectral analysis has been widely used to analyse signals when non-gaussianity and non-linearity are involved. Bicoherence, which is a normalized orm o bispectrum, has been used to extract source speciic eatures, which is inally ed to a neural network classiier. Vector quantization has been used to reduce the dimensionality o the eature set, thereby reducing computational costs. Simulations were carried out with linear, tan and log-sigmoid transer unctions and also with dierent code book sizes. It is ound that the bicoherence eature set can provide acceptable levels o classiication accuracy with a properly trained neural network classiier. KEYWORDS: Bispectrum, Bicoherence, Neural Networks, Target Classiication. I. INTRODUCTION The development o intelligent systems or classiying marine noise sources, based on their acoustic eatures, has gained considerable attention due to their wide applicability in military as well as commercial ields. Traditionally, power spectral analysis, which shows the distribution o the periodic components in a signal, and its variants have been used as the eature extraction technique or such systems. However, being a linear method, and most complex signals being nonlinear, the use o power spectral analysis may turn out to be inappropriate in certain cases. Nonlinear methods must be used in such cases, in order to gain a more complete understanding o signal dynamics. The bispectrum, which is based on the third order cumulant sequence o a signal, can play a key role in characterizing non-linearities o the underlying signal generating mechanisms, especially those containing quadratic non-linearities[]. Bispectral analysis is also capable o providing inormation pertaining to deviations rom Gaussianity o a stochastic process and thus, is ound to be capable o providing much more classiication clues than the conventional power spectral methods. Higher order spectral analysis, especially the bispectrum, has been used in many signal processing applications including transient signal reconstruction [], speaker identiication[3], biomedical signal analysis [4], [5], radar target identiication [6] etc. This paper investigates the easibility o implementing an artiicial neural network based classiier or ocean noises, making use o the higher order spectral eatures. II. RELATED WORKS Underwater target recognition and classiication has been a hot area o research or decades. Most o the methods rely on eature extraction using classical power spectral analysis. An overview o underwater acoustic signal recognition methods, ocussing on Mel-Frequency Cepstral Coeicients (MFCC and Linear Predictive Coding derived Cepstral Coeicients (LPCC has been given in [7]. Feature extraction o biological noises using wavelet decomposition is described [8]. In [9], the classiication ramework include eature extractor using wavelet packets in conjunction with linear predictive coding (LPCand a backpropagation neural-network classiier. Classiication using chaotic eatures and ractal based eatures has been explored in[0] and [] respectively. Empirical Mode Decomposition (EMD has been used in [] or classiying ships.ship noise classiication using Probabilistic Neural Network and AR Model Coeicients has been explored in [3]. An adaptive ramework or underwater noise classiication using Neural Network, with Wavelet eature extraction is described in [4]. However, the classical methods based on power spectral analysis cannot provide inormation regarding non-linearity and couplings arising due to non-linear interactions. Such limitations can be surpassed by the use o higher order spectral analysis techniques like bispectrum. DOI: /IJIRSET Copyright to IJIRSET 755

2 ISSN: International Journal o Innovative Research in Science, (An ISO 397: 007 Certiied Organization Vol. 3, Issue, November 04 III. BISPECTRUM AND BICOHERENCE For a time series. {x(n}, n = 0,,N, with zero mean, the third-order cumulant[5] is deined as * ( k, l E{ x( n x( n k x ( n l} C xx Since the third-order cumulant o a Gaussian process is always equal to zero, this makes it useul or the analysis o non-gaussian signals. The bispectrum B (,, which is the second member o the polyspectrum amily, is deined as the Fourier transorm o the third order cumulant. B(, Cxx( k, lexp( j k *exp( j l k l * E{ X ( X ( X ( } ( (, (. i.e., However, in the case o bispectrum, it is ound that, at the birequency, the complex variance is proportional to the product o the power o the signals[6] at the requencies, and var[ B(, ] Thus, in order to make the bispectrum independent o the energy content at the birequencies, another parameter, reerred to as the bicoherence can be used. Bicoherence, which is a normalized orm o the bispectrum, can be deined as B(, bic(, ( Since the bicoherence is independent o the energy or amplitude o the signal, it can be used as a convenient test statistic or the detection o non-gaussian, non-linear and coupled processes. IV. NEURAL NETWORKS An artiicial neural network (ANN is an inormation processing system[7], composed o simple elements called neurons, operating in parallel. Each neuron is connected to the other neurons by means o directed links, each with an associated weight. The neural network can be trained to perorm a particular unction or task by adjusting the weights between the elements. The basic operation o a neuron involves summing its weighted input signals and applying an output, or activation, unction. The capability o learning rom examples, the ability o reproducing arbitrary nonlinear unctions o input, and the highly parallel and regular structure o ANN make them especially suitable or pattern classiication[8]. ANN could also be used in situations, where the statistical properties o the processes are diicult to predict, as in the case o marine noise signals. In order to train the neural network, or classiication purpose, the eature sets o the objects to be classiied are applied as input to the network. At the training stage, the network adjusts its variable parameters (synaptic weights or capturing the eatures o the object. A. Framing and Bicoherence V. METHODOLOGY The preprocessed noise data waveorms are segmented into records o 0K samples each. The bicoherence o each record is computed using eqn. (, with 56 point DFT. The contour plot o the bicoherence o two targets T and T are presented in Figs. (a and (b. The Bicoherence matrix thus obtained, with a size o 56 x 56, is ound to cause memory limitations or urther computations. In order to reduce the size o the data set, only the values o the irst quadrant were selected or urther DOI: /IJIRSET Copyright to IJIRSET 755

3 ISSN: International Journal o Innovative Research in Science, (An ISO 397: 007 Certiied Organization Vol. 3, Issue, November 04 processing. This could be well justiied due to the six-old symmetry o the bispectrum plot []. Thus, the inal matrix to be processed will have a dimension o n n, where n = 8, which is our times smaller than the original matrix. The reduced n n bicoherence matrix is transormed into a N column vector, where N is equal to n. All the Bicoherence values rom the M records are used to generate a N M matrix, which is vector quantized to get a lower dimension matrix. Fig. (a Bicoherence plot o target T Fig. (b Bicoherence plot o target T B. Vector Quantization Vector Quantization is a lossy data compression method based on the principle o block coding, which codes the values rom a multidimensional vector space into values in a discrete subspace o lower dimension. The Bicoherence data set were vector quantized to reduce the dimensionality. The analysis was carried out with various code book sizes o 3, 64, and 8. C. Training and choice o network parameters The dimensionally reduced bispectral eature set is used to train an artiicial neural network. A eed orward network with back propagation algorithm was used or the analysis. The whole process o eature extraction and training is illustrated in Fig..As the ANN architecture has many parameters which can aect its working, it is important to ind out a proper set o network parameters or the optimum perormance. The various parameters that were analyzed include codebook sizes, the number o hidden layers, and the transer unctions or each layer. Simulation studies were carried out in Matlab environment by varying the relevant parameters to obtain an optimal combination, or making the classiication process more eicient. Fig.. Feature extraction and training From the studies carried out by varying the code book sizes, a code book size o 8 is ound yield satisactory perormance. Fig. 3 shows the variation o average detection rate with respect to various codebook sizes. Simulations DOI: /IJIRSET Copyright to IJIRSET

4 Detection Rate % Detection Rate ISSN: International Journal o Innovative Research in Science, (An ISO 397: 007 Certiied Organization Vol. 3, Issue, November 04 were also carried out by varying the numbers o hidden layers and a network with two hidden layers is ound to give better perormance. As the number o hidden layers is increased, the learning process is ound to become ineicient, resulting in unacceptable generalization o the network. The number o neurons in the hidden layers was also varied to get the optimal perormance. Simulations were carried out with linear, tan and log-sigmoid transer unctions. Better perormance was observed, when the log-sigmoid transer unction is used or all the layers in the network. Fig. 4 illustrates the variation o the detection rate or various transer unctions, or 3 dierent simulations carried out. It is observed that, in all the three simulations, the network with log-sigmoid transer unction exhibited better perormances. VI. RESULTS AND DISCUSSIONS A neural network with two hidden layers was implemented with log-sigmoid as transer unction, or both the layers, and the output o the vector quantizer was given as the input to the network, so that the quantized code books will be used to train the neural network. Variable Learning Rate Backpropagation algorithm was chosen to be the training algorithm. The network was trained with a total o sixteen targets and the proposed method could classiy targets successully with a success rate o 8%. It is also observed that the waveorms used should not be too small, as the perormance o the system declined i the number o samples used or the bicoherence computation is well below 0K samples. 00 Comparison o Codebook sizes 90 Comparison o transer unctions Sim Sim Sim Codebook size 0 linear tan log Transer unctions Fig. 3. Variation o detection rate with respect to codebook size Fig. 4. Comparison o the perormance o various transer unctions VII. CONCLUSIONS Bicoherence, a normalized orm o bispectrum whose variance is independent o the energy content o the signal can play a key role in the analysis o acoustic noise sources. A properly trained neural network with careully chosen parameters, in combination with dimensionality reduction technique like vector quantization, can be eectively made use o, or implementing intelligent target classiiers. Eorts are being made to urther improve the detection rates by ine tuning the network architecture, incorporating more classiication clues. ACKNOWLEDGEMENTS The authors grateully acknowledge the Department o Electronics, Cochin University o Science and Technology, or extending all the acilities or carrying out this work. DOI: /IJIRSET Copyright to IJIRSET

5 ISSN: International Journal o Innovative Research in Science, (An ISO 397: 007 Certiied Organization Vol. 3, Issue, November 04 REFERENCES [] C. Nikias and M. Raghuveer, Bispectrum estimation: A digital signal processing ramework, Proc. IEEE, vol. 75, no. 7, pp , 987. [] C. K. Papadopoulos and C. L. Nikias, Bispectrum estimation o transient signals, in ICASSP-88., International Conerence on Acoustics, Speech, and Signal Processing, pp , 988,. [3] S. Wenndt and S. Shamsunder, Bispectrum eatures or robust speaker identiication, in 997 IEEE International Conerence on Acoustics, Speech, and Signal Processing, vol., no., pp ,997. [4] J. Jakubowski, K. Kwiatos, A. Chwaleba, and S. Osowski, Higher order statistics and neural network or tremor recognition., IEEE Trans. Biomed. Eng., vol. 49, no., pp. 5 9, Feb. 00. [5] S. Taplidou and L. Hadjileontiadis, Analysis o wheezes using wavelet higher order spectral eatures, Biomed. Eng. IEEE, vol. 57, no. 7, pp , 00. [6] P. O. Molchanov, J. T. Astola, K. O. Egiazarian, and A. V Totsky, Classiication o ground moving targets using bicepstrum-based eatures extracted rom Micro-Doppler radar signatures, EURASIP J. Adv. Signal Process., vol. 03, no., p. 6, 03. [7] M. Kucukbayrak, O. Gunes, and N. Arica, Underwater Acoustic Signal Recognition Methods, J. Nav. Sci. Eng., vol. 5, no. 3, pp , 009. [8] Q. Q. Huynh, L. N. Cooper, N. Intrator, and H. Shouval, Classiication o underwater mammals using eature extraction based on timerequency analysis and BCM theory, IEEE Trans. Signal Process., vol. 46, no. 5, pp. 0 07, May 998. [9] M. R. Azimi-Sadjadi, D. Yao, Q. Huang, and G. J. Dobeck, Underwater target classiication using wavelet packets and neural networks., IEEE Trans. Neural Netw., vol., no. 3, pp , Jan [0] S. Yang and Z. Li, Classiication o ship-radiated signals via chaotic eatures, Electron. Lett., vol. 39, no. 4, p. 395, 003. [] S. Yang, Z. Li, and X. Wang, Ship recognition via its radiated sound: The ractal based approaches, J. Acoust. Soc. Am., vol., no., p. 7, 00. [] F. Bao, C. Li, X. Wang, Q. Wang, and S. Du, Ship classiication using nonlinear eatures o radiated sound: an approach based on empirical mode decomposition., J. Acoust. Soc. Am., vol. 8, no., pp. 06 4, Jul. 00. [3] M. Farrokhrooz and M. Karimi, Ship noise classiication using Probabilistic Neural Network and AR model coeicients, in Europe Oceans 005, vol., pp. 07 0, 005. [4] M. R. Azimi-Sadjadi, D. Yao, a a Jamshidi, and G. J. Dobeck, Underwater target classiication in changing environments using an adaptive eature mapping., IEEE Trans. Neural Netw., vol. 3, no. 5, pp. 099, Jan. 00. [5] M. J. Hinich, Bispectrum o ship-radiated noise, J. Acoust. Soc. Am., vol. 85, no. 4, p. 5, 989. [6] P. L. Brockett and G. R. W. Patrick.L.Brockett, Melvin Hinich, Nonlinear and non-gaussian ocean noise, J. Acoust. Soc. Am., vol. 8, no. 4, pp , 987. [7] A. Jain, J. Mao, and K. Mohiuddin, Artiicial neural networks: A tutorial, IEEE Comput., 996. [8] S. Akhtar, M. Elshaei-Abmed, and M. S. Ahmed, Detection o helicopters using neural nets, IEEE Trans. Instrum. Meas., vol. 50, no. 3, pp , Jun. 00. DOI: /IJIRSET Copyright to IJIRSET