Research On Underwater Target Classification Technology Based On Deep Neural Network

Underwater target classification can be classified into two categories,one is to process echo signals from active sonar,and the other is to process noise radiated by targets.Since the radiated noise of ships contains large quantities of ship features that can be used for classification,it’s chosen to accomplish the task of underwater targets classifying and identifying.To identify targets by passive sonar,features of the received ship radiated noise should be extracted in the first step,a suitable classifier should be designed next.Firstly,this paper studies and simulates the radiated noise of ships,including continuous spectrum,line spectrum and modulation spectrum,features of which are extracted next.The frequency doubling method and the greatest common divisor method are used to identify the fundamental frequency of the modulation spectrum,and then the blade number identification expert system is used to identify the number of blades of the target.Due to the different continuum structures of different targets,this paper applies Deep Neural Networks(DNN)to classify ship radiated noise signals.Confusion matrix,receiver operating characteristic curve(ROC)curve and precision-recall(Precision Recall,PR)curve are used to evaluate the performance of the classifier.Stochastic Gradient Descent(SGD),gradient descent with momentum(momentum),Root Mean Square Prop(RMSprop)and Adaptive Moment Estimation(Adam)optimization algorithms are compared,because the Adam optimization algorithm has the best comprehensive performance,this paper uses the Adam optimization algorithm as the optimization algorithm of the DNN network.Influence of the signal-to-noise ratio(SNR)on the classification performance of DNN network is studied,and then the influence of the different SNR of the training set sample and the test set sample SNR on the classification and performance of the DNN network is studied.The simulation results show that when the SNR of the training set is higher than the SNR of the test set,the classification performance of the DNN network will be seriously degraded.However,when the test set SNR is higher than the training set SNR,the performance change is not obvious.The influence of the training set sample size on the classification performance of DNN network is studied.The simulation results show that when the sample size is small,the performance of the classifier is more sensitive to the change of the sample size.Since the DNN network will have a serious performance degradation in the case of small samples,in order to solve this problem,Generative Adversarial Networks(GAN)are used to expand the samples.In order to measure the quality of the generated samples,the Maximum Mean Discrepancy(MMD)algorithm is used to measure the difference between the generated sample distribution and the real sample distribution.Simulation results show that the performance of augmenting each class of samples individually with GAN network is better than augmenting all class samples together.When expanding each type of samples individually,labeled samples are required,and there are always more unlabeled samples of actual underwater acoustic targets.Therefore,this paper uses Semi-Supervised Generative Adversarial Network(Semi-Supervised GAN,SGAN)to expand the real samples.At the same time,the SGAN network can also complete the task of classification of the target.Compared with the GAN network’s performance of augmenting each class of samples individually,the SGAN network’s performance of augmenting all samples at the same time is not much different.The DNN network is trained with the samples expanded by the GAN network,and the classification performance of the DNN network is slightly improved,while the classifier of the SGAN network is used to classify and recognize the target,and the performance is significantly improved.Finally,the Ships Ear dataset is used to construct a dataset of five types of targets: dredgers,motorboats,passenger ships,roller ships and ocean liners.On the problem of sample expansion,the GAN network has little difference between the sample expansion performance of each category of samples and the sample expansion performance of the SGAN network,while the performance of using GAN to expand all categories of samples together is poor.In the problem of small sample classification,compared with directly using the DNN network for classification,first using the GAN network for sample expansion,and then using the DNN network for classification,the performance has been significantly improved.Using the SGAN network for classification,the performance is greatly improved.