CN-122020237-A - Underwater target classification method, device, equipment and storage medium

CN122020237ACN 122020237 ACN122020237 ACN 122020237ACN-122020237-A

Abstract

The application provides an underwater target classification method, device, equipment and storage medium, and relates to the technical field of ocean engineering. The method comprises the steps of generating a time-frequency characteristic data set through wavelet transformation of actually measured noise signals, generating an countermeasure network by utilizing an auxiliary classifier to strengthen the data set so as to overcome sample deficiency, extracting multi-scale characteristics through parallel multi-scale expansion convolution, carrying out self-adaptive dynamic weighting on the characteristics by combining a channel time-frequency attention mechanism, highlighting key information, and finally extracting a bidirectional time sequence dependency relationship by adopting a bidirectional long-short period memory network to realize accurate classification. Through effective fusion time-frequency analysis and deep learning, the accuracy and the robustness of underwater target classification are obviously improved under the condition of a small number of samples.

Inventors

XIONG YIWEN
Pang Chunyang
CHEN MING
Dong Yangze
CHEN XIAHUA
Ling wenchang
CHEN YIWEN
GUAN MING

Assignees

南方海洋科学与工程广东省实验室(湛江)

Dates

Publication Date: 20260512
Application Date: 20251226

Claims (10)

1. An underwater object classification method, characterized in that the steps of the underwater object classification method include: wavelet transformation is carried out on the underwater target noise signals obtained through actual measurement, and a corresponding initial time-frequency characteristic data set is formed; generating an countermeasure network model based on an auxiliary classifier, expanding the initial time-frequency characteristic data set, and obtaining a corresponding target time-frequency characteristic data set; based on a parallel multi-scale expansion convolution model, multi-scale feature extraction is carried out on the target time-frequency feature data set, and corresponding target time-frequency features are obtained; Based on a channel time-frequency attention model, carrying out attention weight distribution on channel dimension and time-frequency dimension on the target time-frequency characteristics to obtain corresponding weighted target time-frequency characteristics; And extracting a bidirectional time sequence feature mapping relation corresponding to the weighted target time-frequency feature based on a bidirectional long-short-term memory network model, and outputting an underwater target classification result based on the bidirectional time sequence feature mapping relation.
2. The method of classifying an underwater target as claimed in claim 1, wherein the step of wavelet transforming the actually obtained underwater target noise signal to form a corresponding initial time-frequency characteristic data set includes: Performing continuous wavelet transformation on the underwater target noise signal to obtain a time-frequency distribution map formed by the underwater target noise signal after telescopic translation operation; and acquiring the initial time-frequency characteristic data set based on the time-frequency distribution map.
3. The method for classifying an underwater target as claimed in claim 1, wherein the step of generating an countermeasure network model based on the auxiliary classifier, expanding the initial time-frequency characteristic data set, and obtaining a corresponding target time-frequency characteristic data set, further comprises: Constructing an initial generation countermeasure network model based on the generator and the discriminator; inputting the random noise signals and the class labels into a generator to obtain corresponding synthesized time-frequency characteristic data; the synthesized time-frequency characteristic data and the historical true time-frequency characteristic data are judged and the category is predicted by the discriminator, so that a judging result and a category predicting result are obtained; And carrying out iterative optimization on the generation parameters of the generator and the discrimination parameters of the discriminator in turn alternately based on the discrimination results obtained each time and the category prediction results to obtain the auxiliary classifier to generate an countermeasure network model.
4. The method for classifying an underwater target according to claim 1, wherein the step of extracting the multi-scale features from the target time-frequency feature data set based on the parallel multi-scale expansion convolution model to obtain the corresponding target time-frequency features further comprises: Constructing a plurality of expansion convolution branches with different expansion rate convolution kernels; and arranging the expansion convolution branches in parallel to obtain the parallel multi-scale expansion convolution model.
5. The method for classifying an underwater target according to claim 1, wherein the channel time-frequency attention model includes a channel attention sub-module and a time-frequency attention sub-module, the step of obtaining the corresponding weighted target time-frequency feature by performing attention weight distribution of channel dimension and time-frequency dimension on the target time-frequency feature based on the channel time-frequency attention model includes: acquiring the channel attention weight of the target time-frequency characteristic in the channel dimension through the channel attention submodule; acquiring time-frequency attention weights of the target time-frequency characteristics in time-frequency dimensions through the time-frequency attention submodule; and acquiring the weighted target time-frequency characteristic based on the time-frequency attention weight, the channel attention weight and the target time-frequency characteristic.
6. The method of classifying an underwater target as in claim 5, wherein the step of acquiring the time-frequency attention weight of the target time-frequency feature in the time-frequency dimension by the time-frequency attention sub-module comprises: carrying out cavity convolution on the target time-frequency characteristics, and carrying out global average pooling along a time dimension and a frequency dimension respectively to obtain a time dimension characteristic vector and a frequency dimension characteristic vector; and acquiring the time-frequency attention weight based on the time dimension feature vector and the frequency dimension feature vector.
7. The method for classifying an underwater target according to claim 1, wherein the step of extracting a bidirectional time sequence feature mapping relation corresponding to the weighted target time-frequency feature based on the bidirectional long-short term memory network model and outputting an underwater target classification result based on the bidirectional time sequence feature mapping relation comprises: Based on the two-way long-short-term memory network model, respectively acquiring a forward hiding state sequence and a backward hiding state sequence of the weighted target time-frequency characteristic; Forming the bidirectional time sequence feature mapping relation according to time sequence based on the forward hidden state sequence and the backward hidden state sequence; And outputting an underwater target classification result based on the bidirectional time sequence feature mapping relation.
8. An underwater object classification device, characterized in that the underwater object classification device comprises: the data preprocessing module is used for carrying out wavelet transformation on the underwater target noise signals obtained through actual measurement to form a corresponding initial time-frequency characteristic data set; The data set expansion module is used for generating an countermeasure network model based on the auxiliary classifier, expanding the initial time-frequency characteristic data set and obtaining a corresponding target time-frequency characteristic data set; the feature extraction module is used for carrying out multi-scale feature extraction on the target time-frequency feature data set based on a parallel multi-scale expansion convolution model to obtain corresponding target time-frequency features; The weight distribution module is used for distributing the attention weight of the channel dimension and the time-frequency dimension to the target time-frequency feature based on the channel time-frequency attention model, and obtaining the corresponding weighted target time-frequency feature; and the target classification module is used for extracting a two-way time sequence feature mapping relation corresponding to the weighted target time-frequency feature based on the two-way long-short-term memory network model and outputting an underwater target classification result based on the two-way time sequence feature mapping relation.
9. An underwater object classification apparatus comprising a memory, a processor and an underwater object classification program stored on the memory and executable on the processor, the underwater object classification program being configured to implement the steps of the underwater object classification method as claimed in any of claims 1 to 7.
10. A storage medium, characterized in that the storage medium is a computer-readable storage medium having stored thereon an underwater object classification program which, when executed by a processor, implements the steps of the underwater object classification method according to any of claims 1 to 7.

Description

Underwater target classification method, device, equipment and storage medium Technical Field The application relates to the technical field of ocean engineering, in particular to an underwater target classification method, an underwater target classification device, an underwater target classification equipment and a storage medium. Background The underwater target identification has important significance in the fields of ocean monitoring, national defense safety and the like. The traditional method mainly depends on manual listening or conventional signal processing technology, and has the problems of low efficiency, strong subjectivity, poor generalization capability and the like. In recent years, a method based on machine learning is gradually applied to the field, but challenges such as difficulty in acquiring underwater samples, insufficient data volume, complex noise environment and the like are still faced. In the prior art, an original signal is often directly used for modeling, and when a sample is scarce, the model is easy to be over-fitted, so that the recognition accuracy is limited. Therefore, how to realize stable and high-precision automatic classification of underwater targets under limited samples is still a technical problem to be solved currently. Disclosure of Invention The application mainly aims to provide an underwater target classification method, device, equipment and storage medium, which aim to solve the technical problem of how to effectively improve the accuracy and stability of underwater target classification under the condition of rare samples. In order to achieve the above object, the present application provides an underwater object classification method, comprising the steps of: wavelet transformation is carried out on the underwater target noise signals obtained through actual measurement, and a corresponding initial time-frequency characteristic data set is formed; generating an countermeasure network model based on an auxiliary classifier, expanding the initial time-frequency characteristic data set, and obtaining a corresponding target time-frequency characteristic data set; based on a parallel multi-scale expansion convolution model, multi-scale feature extraction is carried out on the target time-frequency feature data set, and corresponding target time-frequency features are obtained; Based on a channel time-frequency attention model, carrying out attention weight distribution on channel dimension and time-frequency dimension on the target time-frequency characteristics to obtain corresponding weighted target time-frequency characteristics; And extracting a bidirectional time sequence feature mapping relation corresponding to the weighted target time-frequency feature based on a bidirectional long-short-term memory network model, and outputting an underwater target classification result based on the bidirectional time sequence feature mapping relation. In one embodiment, the step of performing wavelet transformation on the underwater target noise signal obtained through actual measurement to form a corresponding initial time-frequency characteristic data set includes: Performing continuous wavelet transformation on the underwater target noise signal to obtain a time-frequency distribution map formed by the underwater target noise signal after telescopic translation operation; and acquiring the initial time-frequency characteristic data set based on the time-frequency distribution map. In an embodiment, before the step of generating the countermeasure network model based on the auxiliary classifier and expanding the initial time-frequency characteristic data set to obtain the corresponding target time-frequency characteristic data set, the method further includes: Constructing an initial generation countermeasure network model based on the generator and the discriminator; inputting the random noise signals and the class labels into a generator to obtain corresponding synthesized time-frequency characteristic data; the synthesized time-frequency characteristic data and the historical true time-frequency characteristic data are judged and the category is predicted by the discriminator, so that a judging result and a category predicting result are obtained; And carrying out iterative optimization on the generation parameters of the generator and the discrimination parameters of the discriminator in turn alternately based on the discrimination results obtained each time and the category prediction results to obtain the auxiliary classifier to generate an countermeasure network model. In an embodiment, before the step of extracting the multi-scale features from the target time-frequency feature dataset to obtain the corresponding target time-frequency features based on the parallel multi-scale expansion convolution model, the method further includes: Constructing a plurality of expansion convolution branches with different expansion rate convolution kernels; and arrangi