Research On The Technologies Of Enhancement And Classification Of Tursiops Aduncus Whistles

Whales are located at the top of the food chain in marine ecosystems and play a vital role in maintaining the stability of marine ecosystems.However,due to human activities such as overfishing,by-catch,pollution,and habitat destruction,whale and dolphin resources worldwide are severely damaged.Many species are highly endangered and on the verge of extinction.Passive acoustic monitoring(PAM)has gradually become an essential technology for measuring and studying cetacean population distribution,habitat behavior,and interaction due to its high reliability,long monitoring time,convenient operation,and low cost.The complex and highly developed communication system of cetaceans reflects the complexity of their social relations,and different types of whistles are highly correlated with the behavior of cetaceans.Therefore,it is significant to research PAM-based whistling signal processing technology for maintaining marine ecosystems.In this paper,several key technologies of underwater PAM are studied :(1)The research on whistle signal enhancement under low SNR broadband stationary noise is carried out.To extract better the envelope characteristics of the whistle signal,a robust unsupervised whistle enhancement scheme based on improved local mean decomposition and adaptive noise estimation and logarithmic spectrum amplitude is proposed in this paper.The enhancement of the whistle signal is divided into three main parts: Firstly,to further alleviate the influence of modal aliasing in decomposition and achieve better modal spectrum separation,a complete ensemble empirical optimal envelope local mean decomposition algorithm with adaptive noise is proposed.The noisy whistle signal is adaptively decomposed into a product function with amplitude and frequency modulation characteristics;Secondly,each noisy product function after decomposition is enhanced by adaptive noise estimation and logarithmic spectral amplitude method.Then,the Hurst index is used to reconstruct the sub-mode with the slightest noise interference to complete the reconstruction of the whistle signal.Finally,in the collected Tursiops aduncus whistle data set,the superiority of the proposed whistle signal enhancement scheme based on broadband stationary noise interference is proved.(2)The research on whistle signal enhancement under low SNR underwater impulse noise interference is carried out.Due to the grouping and clustering characteristics of the time-frequency spectrum of the whistle signal collected by underwater PAM,this paper proposes a whistle enhancement algorithm based on non-convex time-frequency regularization overlapping group shrinkage.By combining the predefined time-frequency structure prior and the non-convex penalty term of the noisy whistle signal,the pure sparse whistle signal is obtained iteratively.Secondly,an iterative minimization algorithm based on the principle of optimization minimization(MM)is used to solve the convex optimization problem quickly.Compared with the conventional convex regularization,the non-convex term can further promote the sparsity of the reconstructed whistle signal.Finally,the proposed scheme is applied to the typical southern bottlenose dolphin whistle signal under the interference of underwater impulse noise(gun shrimp noise)modeled by SαS distribution.The whistle enhancement results under different signal-to-noise ratios and underwater impulse noise based on different parameters are verified,respectively.(3)The whistle event detection of cetaceans under a large number of pulse click sequence interference,and dynamic continuous background noise is studied.In order to filter out the interference of the collected whale whistles in the time-frequency spectrum,this paper proposes an unbiased Gammatone multi-channel Savizky-Golay whistler time-frequency representation enhancement algorithm.Secondly,to accurately find the front and back endpoints of the whistle contour from the enhanced spectrogram,a set of maximum moving filter banks is proposed to extract multi-scale and multi-directional features of whistle signals with different whistle duration and contour diversity in the enhanced spectrogram.The statistical histogram of the feature spectrogram is used to obtain the frame-level binary adaptive decision of the whistle event.Finally,the superiority of the proposed whistle signal event detection scheme is proved in the collected actual Tursiops aduncus whistle data set.(4)The intra-class signal classification of single-population cetaceans was studied under unbalanced whistler signal samples.Firstly,aiming at the problem of an unbalanced sample number under each whistle category,a random tandem amplification algorithm with five audio distortion algorithms is proposed.By setting the corresponding opening probability for each algorithm and randomly selecting each amplification factor in a specific range,the generalization ability of the trained model is improved.Secondly,a lightweight convolutional neural network model for cetacean whistle classification is proposed to aim at the problem of high time complexity in training a large amount of data obtained after passive acoustic monitoring and data amplification.Finally,the superiority of the proposed sample class imbalanced whistle signal classification scheme is proved in the collected actual Tursiops aduncus whistle data set.