CN-122020321-A - Ship noise identification method and system based on deep learning

CN122020321ACN 122020321 ACN122020321 ACN 122020321ACN-122020321-A

Abstract

The invention relates to the technical field of noise identification, in particular to a ship noise identification method and a ship noise identification system based on deep learning, wherein the method comprises the following steps of collecting ship noise signals and generating a two-dimensional time-frequency spectrogram of the ship noise signals through self-adaptive time-frequency conversion; generating a time sequence feature vector sequence through a convolution neural network of global time sequence perception according to a two-dimensional time spectrum chart, inputting the time sequence feature vector sequence into a multi-scale time sequence mode extraction network to obtain multi-scale time sequence fusion features of ship noise signals, and inputting the multi-scale time sequence fusion features into a noise classification network to realize noise identification. According to the invention, a more adaptive time-frequency spectrogram can be generated through self-adaptive time-frequency conversion, the capturing capacity of the noise periodic mode and the sensitivity of anomaly detection are improved through the convolution neural network of global time sequence perception and the multi-scale time sequence mode extraction network, and the accuracy and the intelligent degree of noise identification in a complex marine environment are improved.

Inventors

XUE RUICHAO
LI JIANLONG
FANG RONG

Assignees

自然资源部第三海洋研究所
浙江大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (9)

1. The ship noise identification method based on deep learning is characterized by comprising the following steps of: acquiring a ship noise signal, and generating a two-dimensional time-frequency spectrogram of the ship noise signal through self-adaptive time-frequency conversion; Generating a time sequence feature vector sequence of the ship noise signal through a global time sequence perceived convolution neural network according to the two-dimensional time-frequency spectrogram; Inputting the time sequence feature vector sequence into a multi-scale time sequence mode extraction network to obtain multi-scale time sequence fusion features of the ship noise signals; inputting the multi-scale time sequence fusion characteristics into a noise classification network to realize noise identification; the method for acquiring the ship noise signal and generating the two-dimensional time-frequency spectrogram of the ship noise signal through self-adaptive time-frequency conversion comprises the following steps: Collecting a ship noise signal; Constructing a differentiable short-time Fourier transform network structure for realizing self-adaptive time-frequency transformation of the ship noise signal so as to obtain complex tensors; And generating a two-dimensional time-frequency spectrogram of the ship noise signal according to the complex tensor.
2. The method for identifying ship noise based on deep learning according to claim 1, wherein the constructing a differentiable short-time fourier transform network structure is used for realizing adaptive time-frequency transform of the ship noise signal, and further obtaining complex tensors, and comprises the following steps: Constructing a differentiable short-time Fourier transform network structure; inputting the ship noise signal into the differentiable short-time Fourier transform network structure to obtain a real part tensor and an imaginary part tensor of the ship noise signal; And splicing the real part tensor and the imaginary part tensor to obtain the corresponding complex tensor.
3. The deep learning-based ship noise identification method as claimed in claim 2, wherein: The differentiable short-time Fourier transform network structure comprises two parallel 1D convolution layers with the same parameter setting, wherein one 1D convolution layer outputs the real part tensor, the other 1D convolution layer outputs the imaginary part tensor, and the output shapes of the two 1D convolution layers are the same.
4. The deep learning-based ship noise identification method as claimed in claim 1, wherein: the global time sequence aware convolutional neural network comprises an input module, a convolutional module and an attention module, wherein the convolutional module at least comprises three time sequence expansion convolutional layers which are stacked and the time expansion rate increases exponentially.
5. The method for identifying ship noise based on deep learning according to claim 4, wherein the generating the sequence of time sequence feature vectors of the ship noise signal through a global time sequence aware convolutional neural network according to the two-dimensional time spectrum diagram comprises the following steps: inputting the two-dimensional time-frequency spectrogram into the convolution neural network of the global time sequence perception, and extracting spectrogram characteristics through the convolution module; And according to the spectrogram characteristics, compressing the information of the frequency dimension into a characteristic channel through the attention module, and outputting a time sequence characteristic vector sequence of the ship noise signal.
6. The deep learning-based ship noise identification method as claimed in claim 1, wherein: The multi-scale time sequence mode extraction network comprises a component feature extraction module and a feature fusion module, wherein the component feature extraction module comprises a plurality of parallel cyclic neural network branches, and each cyclic neural network branch adopts different time steps to process the time sequence feature vector sequence.
7. The method for identifying ship noise based on deep learning according to claim 6, wherein the step of inputting the sequence of time sequence feature vectors into a multi-scale time sequence pattern extraction network to obtain multi-scale time sequence fusion features of the ship noise signal comprises the steps of: Inputting the time sequence feature vector sequence into the multi-scale time sequence mode extraction network, and acquiring multi-scale time sequence feature components through the component feature extraction module; and carrying out weighted splicing on each time sequence feature component through the feature fusion module to obtain a multi-scale time sequence fusion feature.
8. The deep learning-based ship noise identification method as claimed in claim 1, wherein: And forming a ship noise recognition model by using the differentiable short-time Fourier transform network structure, the convolution neural network of global time sequence perception, the multi-scale time sequence mode extraction network and the noise classification network, and performing end-to-end training on the ship noise recognition model.
9. A deep learning based marine noise identification system, comprising a data input device, a data output device, a processor, and a storage, the storage comprising a computer readable storage medium having a computer program stored therein, the computer program comprising program instructions which, when executed by the processor, cause the processor to implement a deep learning based marine noise identification method as claimed in any one of claims 1-8.

Description

Ship noise identification method and system based on deep learning Technical Field The invention relates to the technical field of noise identification, in particular to a ship noise identification method and system based on deep learning. Background With the increasing frequency and complexity of marine activities, higher requirements are placed on the accuracy, real-time and robustness of marine noise identification. The existing ship noise identification method is mainly based on the traditional signal processing technology and some basic machine learning algorithms. In the aspect of traditional signal processing technology, common methods comprise short-time Fourier transform, wavelet transform and the like, and the methods generally adopt fixed parameters and a transformation mode to perform time-frequency analysis on ship noise signals, so that when complex and changeable ship noise is processed, the time-frequency resolution is not ideal enough, key characteristic information in the noise signals is difficult to accurately capture, and the follow-up recognition accuracy is affected. In the ship noise identification method based on machine learning, part of the method only uses a common convolutional neural network to perform feature extraction on a time-frequency spectrogram of the ship noise, so that the long-distance dependence of the ship noise on the time dimension is ignored, key information in a time sequence mode cannot be effectively captured, or a single-scale feature extraction mode is adopted, the features of the ship noise cannot be comprehensively described from multiple angles, so that the generalization capability of a model is poor when the ship noise under different types and different working conditions is faced, the identification performance is unstable, and the high-precision and strong robustness requirements of the ship noise identification in practical application are difficult to meet. Disclosure of Invention Aiming at the defects in the prior art, the invention provides a ship noise identification method and a ship noise identification system based on deep learning. In order to achieve the purpose, in a first aspect, the invention provides a ship noise identification method based on deep learning, which comprises the following steps of collecting ship noise signals, generating a two-dimensional time-frequency spectrogram of the ship noise signals through self-adaptive time-frequency conversion, generating a time sequence feature vector sequence of the ship noise signals through a global time sequence perceived convolution neural network according to the two-dimensional time-frequency spectrogram, inputting the time sequence feature vector sequence into a multi-scale time sequence mode extraction network to obtain multi-scale time sequence fusion features of the ship noise signals, and inputting the multi-scale time sequence fusion features into a noise classification network to realize noise identification. According to the invention, a more adaptive time-frequency spectrogram can be generated through self-adaptive time-frequency conversion, the capturing capacity of the noise periodic mode and the sensitivity of anomaly detection are improved through the convolution neural network of global time sequence perception and the multi-scale time sequence mode extraction network, and the accuracy and the intelligent degree of noise identification in a complex marine environment are improved. Optionally, the acquiring the ship noise signal and generating the two-dimensional time-frequency spectrogram of the ship noise signal through adaptive time-frequency transformation includes the following steps: Collecting a ship noise signal; Constructing a differentiable short-time Fourier transform network structure for realizing self-adaptive time-frequency transformation of the ship noise signal so as to obtain complex tensors; And generating a two-dimensional time-frequency spectrogram of the ship noise signal according to the complex tensor. Optionally, the constructing a differentiable short-time fourier transform network structure is configured to implement adaptive time-frequency transform on the ship noise signal, thereby obtaining a complex tensor, and includes the following steps: Constructing a differentiable short-time Fourier transform network structure; inputting the ship noise signal into the differentiable short-time Fourier transform network structure to obtain a real part tensor and an imaginary part tensor of the ship noise signal; And splicing the real part tensor and the imaginary part tensor to obtain the corresponding complex tensor. Optionally, the differentiable short-time fourier transform network structure includes two parallel 1D convolution layers with the same parameter setting, one of the 1D convolution layers outputs the real part tensor, the other 1D convolution layer outputs the imaginary part tensor, and the output shapes of the two 1D convolution layers a