Research On Underwater Target Recognition And Classification Method

The current marine environment is becoming more and more complex,and the traditional underwater target recognition technology is facing unprecedented challenges.On the one hand,with the continuous development of underwater moving target vibration reduction and noise reduction technology,the noise level of the underwater vehicle has been reduced to or close to the marine background noise,and the low signal-to-noise ratio signals make the task of underwater target recognition more difficult.On the other hand,China has a long coastline with wide latitude coverage and complex underwater conditions.Due to its limited ability of feature expression,the traditional target recognition technology can only identify the target within a limited range of hydrological conditions and cannot be carried out under complex hydrological conditions.In addition,with the development of the coastal economy,the noisy background noise caused by the high-intensity activities of marine vessels further increases the task difficulty of target recognition.Therefore,in the face of the current complex underwater environment,the underwater target recognition task needs a recognition model with stronger feature representation ability and higher recognition accuracy.In this paper,the deep learning method is used to study the radiated noise signals of underwater targets,and the end-to-end processing of an underwater target signal is realized based on the convolutional neural network structure.The whole process covers four stages:vector representation,feature extraction,model reduction and target recognition of underwater target noise signals.In the stage of vector representation of underwater target noise signal,an improved anti-noise method of power-law normalized cepstrum coefficient is proposed in this paper,and in the stage of feature extraction,the feature weighting layer structure is proposed.In the stage of model design,a Fast reducted dimension convolution model with attention is proposed.In the stage of recognition and classification,an incremental integration method based on clustering is realized for underwater target incremental classification.Overall,the research in this paper mainly includes the following four aspects:Firstly,in view of the serious reverberation of radiated noise of underwater targets in the complex underwater acoustic environment,an improved anti-noise Power-Normalized Cepstral Coefficients(ia-PNCC)is proposed to improve the anti-noise capability on the basis of comparing a lot of noise feature description schemes.ia-PNCC uses multi-orthogonal windows and normalized Gammatone filter banks to enhance the anti-noise ability of PNCC inthe representation of underwater target noise features.The experimental results show that the acoustic features expressed by ia-PNCC have strong anti-noise ability and are more suitable for underwater target recognition model based on convolutional neural structure.Compared with other single acoustic features and convolution neural network,the ia-PNCC features greatly improve the accuracy of underwater target recognition.Secondly,aiming at the problem that the underwater target noise data set,which is used to train deep convolution neural networks,is generally small in size and insufficient to support the training of deep convolution neural network models.From the data expansion and model simplification two aspects to begin to solve.In the aspect of data expansion,the symmetric learning data extension model(Symmetric Learning Data Augmentation Model,SLDAM)is proposed to effectively constrain the extended data samples from the two angles of generating sample loss and generating sample feature loss.So as to ensure the quality of the generated data and the reliability of the generated data in the classifier.In the aspect of model simplification,a Faster Reduced Dimensional Convolution Model with Attention(FRD-CMA)is proposed to solve the problem of over-fitting of the training model on the sample data set with the small amount of data.The FRD-CMA model takes the direction of convergence as the direction of pooling operation and improves the problem of rough dimensionality reduction when pooling operations are separated from task in convolution neural networks.Experiments show that the FRD-CMA model can effectively reduce the risk of over-fitting in small data sets when the data is small,and the dimension reduction of the model based on the direction of feature focus can effectively reduce the risk of over-fitting in small data sets.Thirdly,in view of the difficulty of feature extraction and the low utilization ratio of underwater target recognition and classification task,by analyzing the characteristics of convolution neural network structure,the convolution operation process is used to preserve the position information of convolutional kernel.The Feature-Weighted Layer(FWL)is introduced in front of the fully connected layer.The FWL structure weights the target features extracted from the convolution neural network from the two dimensions of the plane position and the spatial channel.While preserving the feature position information,the use efficiency of the feature is improved.The experimental results show that the weighted process of FWL does not introduce additional parameters to reduce the burden of network training.At the same time,the convolution neural network is used to automatically extract the features of samples from the actual data set.The feature can be fully optimized by the efficient iterative process trained by the model.Finally,in view of the existence of underwater reverberation in a complex environment,the difference of radiated noise becomes smaller,and the underwater target classifier needs task-oriented dynamic adjustment of the performance of the classifier.An Underwater Targets Incremental Classification based on Clustering(UTICC)is proposed.UTICC adjusts the sensitivity of underwater targets by integrating different underwater target classifiers.By using the classification error probability of the base classifier as the difference between the classifiers,UTICC can directly participate in the training process of the deep learning model and realize the support of the deep learning model to the incremental data.The experimental results show that UTICC method can effectively emphasize the difference between the base classifiers and enhance the generalization performance of the model on the dataset.