| At present,activities in the sea are becoming more and more frequent.The exchange of information between unidentified targets by acoustic signals,such as marine organisms,ships,communication equipment,detection equipment,etc.,forms a turbulent storm under waves.As a result,the security situation of the national maritime border is even more complicated and unexpected.Therefore,developing analysis and recognition methods and a platform for underwater acoustic(UWA)signals and detecting signals of unknown targets in time are of great significance for ensuring the safety of sea areas and protecting marine life.TF transform algorithms play important roles in UWA signal recognition since they can extract time-frequency(TF)features to distinguish acoustic signals from diffferent underwater targets.When UWA signals are contaminated by strong background noise from the marine environment and their frequency modulated(FM)components have characteristics of fast time-varying,close,and even overlapping frequency points,traditional TF transform methods cannot effectively extract the high-resolution TF spectrum of such signals.The low discrimination of TF spectrum of different targets results in low accuracy of UWA signal recognition.Besides,the signal recognition model using deep learning cannot currently meet the requirement for detecting multi-target UWA signals and their elements.To address the above problems,this thesis aims to research and develop a UWA signal analysis and recognition system,explores the new theory of TF feature extraction from the perspective of enhancing the TF resolution of nonlinear FM(NLFM)components,and its applications in UWA signal recognition based on deep learning.The main research works of this thesis include:(1)The synchro-compensating chirplet transform is proposed to enhance the TF resolution of NLFM components in high noise.By introducing the self-tuning demodulated operator and an instantaneous rotating operator,a series of instantaneous chirp rate coefficients are acquired to match the local curvature of the ridges of NLFM components.The TF energy is synchronously concentrated on the ridges along the instantaneous frequency.The frequency reassignment is then applied to obtain a high-resolution TF spectrum.Experimental results show that the proposed algorithm enhances the TF resolution of NLFM components that are close or even overlapping for high-noise environments.(2)An anisotropic chirplet transform(ACT)is proposed to enhance the TF resolution of multimodal components.Such transform applies a TF-varying two-dimensional window allowing the adjustment between time and frequency resolution of multimodal components and noise suppression during the TF energy concentration.Directional chirplet ridges combine with frequency reassignment to enhance TF resolution of pulsed components and tonal components.A structure-split-merge algorithm is developed to decompose and extract tonal and pulsed components with close and even overlapping in the TF spectrum.Relying on these components,the pulse-to-tone strength ratio is defined to measure the variation of the two-mode components of UWA signals.(3)A TF feature network is proposed based on TF feature and deep learning to detect the sound elements in UWA signals.The ACT uses the TF reassignment to enhance the TF resolution.To improve the efficiency of the TF feature extraction,the two-dimensional TF-varying window in the ACT is decomposed into two onedimensional windows such that a linear convolution calculates TF reassignment.In the proposed efficient feature pyramid structure,the dilated convolution is applied to enlarge the receptive field of the feature maps at different dilation rates.These feature maps are cascaded to enhance the features of TF components in the deep layers.To remove the gridding artefacts,each branch of the feature pyramid utilizes a forward feature fusion.Experimental results on UWA communication and whale signals show that the proposed TF feature network has good performance.(4)Since the classification model based on deep learning and TF feature has poor stability for UWA signals and low accuracy for ship signal classification,a classification model based on deep learning and multi-feature fusion is designed by combining TF features with time-domain and frequency-domain features.Depending on the analysis of a large number of UWA signals,three time-domain features,six frequencydomain features,and one TF feature are selected,which are respectively embedded into the long-short-term memory(LSTM)network,Bidirectional LSTM,and the convolutional neural network.The network structure with multi-feature spatial fusion is introduced to map features of three branch networks into features with the same scale for fused parameters learning.The Softmax layer of the network calculates a probability distribution over predicted output classes.Experimental results on UWA signals show that the proposed multi-feature fusion network has great stability and can improve the accuracy of recognition in UWA signals from multiple targets.(5)For the management,feature extraction and recognition of multi-target UWA signals,the system of UWA signal TF analysis and recognition is designed by using a hierarchical architecture.In the system,the proposed TF transforms implements the TF feature extraction of NLFM components,and the proposed deep learning models implement recognition of multi-target UWA signals.The function of the system is tested and the practicability of the proposed techniques is verified by four kinds of signals,namely,marine biological signals,sonar signals,UWA communication signals and ship signals.Experimental results show that the system has convenient functions for signal preprocessing,feature extraction and recognition,and can realize the TF feature extraction of NLFM components and the recognition of multi-target signals and sound events. |
