Research On Underwater Target Recognition Technology Based On Bionic Dolphins

With the deepening of marine construction projects,the detection and identification of underwater unknown targets is the basis and key for improving the transparency of underwater information in the future.The sonar system of dolphin-like creatures is able to adapt to complex environments and has excellent performance under non-stationary conditions.This paper therefore exploits the physical mechanism and recognition basis of dolphin-like organisms in the recognition process,and delves into the implementation of a classification and recognition method for underwater targets from the perspective of bionic dolphins.Based on this,this paper investigates the design of bionic emission signals and the acquisition of bionic underwater target echoes,and achieves the bionic feature extraction of target echoes from both postprocessing and pre-processing directions in the dolphin target recognition process,and finally completes the classification and recognition of targets by combining with neural network models.At first,based on the real collected dolphin-emitted Click signals in the time-frequency domain,the time-frequency characteristics of the dolphin-emitted signals were qualitatively analyzed and the different components of the dolphin signal were separated in the fractionalorder Fourier domain,according to which a negative linear frequency modulation signal delay model was derived.Secondly,the performance of synthetic bionic signal is also analyzed by Q-function,and making use of bionic signal as acoustic excitation,corresponding bionic underwater target echoes under numerical and finite element simulation conditions are obtained respectively.Secondly,based on the statistical features in the time-frequency domain in the post-processing of dolphin classification and recognition process,synchronous extraction transformation based on time-frequency energy rearrangement is introduced as a method for time-frequency feature extraction of the echoes,and the best time-frequency detail feature portrayal capability of the SET method is demonstrated through Rayleigh entropy.A feature extraction method based on the biomimetic dolphin auditory system model is simultaneously investigated from the echo pre-processing process based on the dolphin’s biological perception mechanism,and the feature representation mechanism of the auditory spectrogram of the model is explained in two dimensions,namely time domain and frequency domain.Then,for the feature matrices under the two kinds of feature extraction methods,the singular value decomposition is introduced to dimensionally compress the feature matrices,and a specific parametric analysis of the amount of matrix information contained in the singular values under different orders,demonstrating the sparsity property of the matrix singular values.On this basis,kernel typical correlation analysis is used to fuse the different feature vectors at the feature level,and the effectiveness of this feature fusion method is illustrated in terms of inter-class and intra-class distances in simulation.Finally,radial basis function neural network was introduced as a classifier for bionic dolphin target classification recognition by combining the intelligent processing mechanism of echo information in the head of dolphin,and the performance of the classification recognition effect of synchronous extraction transformation features and biomimetic dolphin auditory system model auditory spectrum features in different feature dimensions and different signalto-noise ratios was analyzed on the basis of pool experimental echo data.The results show that performance of two methods is comparable under high SNR conditions,while the SET features are more effective under low SNR conditions,but its feature dimension is an order of magnitude larger than that of biomimetic dolphin auditory system model,while the feature vector after fusing by kernel typical correlation analysis is more effective than two single features in terms of recognition rate,and significantly reduces the feature dimension,which has stronger feature characterization and classification recognition ability.