Few-Shot Underwater Acoustic Recognition Based On Knowledge Embedding

The recognition of underwater acoustic targets is the premise of carrying out many works on the seabed.And it is a key research direction of current ocean exploration.The audio signal emitted by the underwater acoustic source is propagated in the complex marine environment,resulting in the weak signal strength collected by the signal receiving end.So that people can’t get a large amount of high-quality data of underwater acoustic signal.Therefore,how to classify and detect efficiently when the number of training samples is insufficient,is the core problem of underwater acoustic target recognition.Existing underwater acoustic target recognition methods can be divided into two parts.One is based on traditional machine learning and another is based on deep learning.In the process of feature extraction,methods based on traditional machine learning is hard to improve the accuracy greatly,because of the limit made by the one-sided representation of shallow features.The deep learning-based methods are able to extract deep features through many nonlinear transformation layers,but they still need a large number of training samples to fit the network for improving the classification accuracy.Aiming at the problem of underwater acoustic classification and recognition under few samples,this thesis proposes a stereo and semantic underwater acoustic recognition method based on knowledge embedding.The research contents of this thesis are as follows:1.According to the characteristics of the underwater acoustic signal,the method extracts a total of eleven underwater acoustic primitive features in five categories.These features include not only the physical features of the signal level,but also the deep learning features of the data level,which characterize the data characteristics in terms of pitch,timbre,loudness,regularity,etc.,so as to reduce the information lost in the process of feature extraction as much as possible;2.In order to avoid feature redundancy caused by directly splicing features,according to the actual physical meaning represented by these audio features and the prior knowledge of experts,an initial knowledge graph that can represent the correlation between audio features is constructed.Node connection relationship,which called semantic-level knowledge,is embedded through knowledge graph.And the knowledge is embedded into the network through a stereo and semantic network,which is designed based on graph convolution;3.Aiming at the problem of incomplete knowledge representation in the designed initial knowledge graph,and the defect that the stereo and semantic network cannot effectively embed time series information,the network which can discriminate node correlation and Graph LSTM network are designed.These two networks solve the update and optimization problem of knowledge graph,and the problem of how data time series information feature embedded respectively.These two networks achieve what people call “update the knowledge”.The proposed method solves the problems that when met few-shot audio data and high intensity noise in the marine environment,people could only get few effective samples collected,then cause low classification and recognition accuracy.Through experiments on the simulated underwater acoustic dataset,the effectiveness of the stereo and semantic network proposed in this thesis is verified when facing the problem of underwater acoustic classification and recognition in the case of few samples.And the accuracy rate is improved significantly compared with the traditional methods.In addition,not only in the case of specific underwater acoustic target recognition,but also in the classification and recognition of natural environment audios in the case of few samples,the proposed stereo and semantic network shows excellent performance compared with other deep learning algorithms.Moreover,it is still robust to high-intensity noise and the number of training samples continues to decrease.