Transient signal identification method based on multi-scale multi-resolution feature fusion
Through the multi-scale multi-resolved feature fusion method, deep learning technology is used to jointly learn time-domain and time-frequency domains of underwater transient signals, effectively identifying transient signals and improving the accuracy of underwater target detection.
Patent Information
- Application Number
- CN202510377805.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-28
- Publication Date
- 2025-07-18
AI Technical Summary
The prior art is difficult to effectively identify underwater transient signals, especially impact signals and pulsating signals, which makes it difficult to detect underwater targets.
The multi-scale multi-resolution feature fusion method is adopted to achieve effective identification of transient signals through time domain feature extraction, multi-scale multi-resolution time spectrum generation, cross-attention mechanism feature fusion and classifier classification.
The time-domain energy mutation and time-frequency domain feature information of transient signals are fully utilized to improve the recognition performance of transient signals and solve the problem of difficult features due to short duration.
Smart Images
Figure CN120337001A_ABST
Abstract
Description
Technical Field:
[0001] The present invention belongs to the technical field of underwater acoustic engineering, and particularly relates to a transient signal identification method based on multi-scale and multi-resolution feature fusion. Background Art:
[0002] The noise signals radiated by underwater targets can be divided into two categories. One is continuously and repeatedly generated, and the other is transiently changing. There are many reasons for underwater targets to generate transient signals, such as changes in the operating states of machinery and equipment, changes in working conditions, sudden signals during ship navigation, weapon launches, etc. In the early stage, underwater transient signals were regarded as interference signals. However, as ships become quieter, it has become increasingly difficult to detect underwater targets using steady-state radiated noise. Since these occasional transient signals are difficult to control and will inevitably appear during the navigation of underwater targets, using transient signals to detect and identify underwater targets has become a new method for passive sonar target detection.
[0003] Underwater transient signal types can be divided into impact signals, pulsating signals, and state mutation signals, etc. according to their generation principles. Impact signals have a short duration in the time domain and large envelope changes, and have the characteristics of a wide frequency band and a large dynamic range in the frequency domain; pulsating signals have a relatively longer duration in the time domain compared to impact signals, and the envelope changes in segments, and their frequency domain characteristics are low frequency and band-limited. Currently, most studies use the energy mutation characteristics of transient signals in the time domain or time-frequency domain to detect transient signals, and there is less research on the identification of different types of transient signals, and the difficulty of transient signal identification is relatively large. However, although the duration of transient signals is short, their frequency bands are relatively wide, and there are still exploitable characteristic information in the entire time-frequency space. If the time-domain energy mutation of transient signals and the characteristic information in the time-frequency domain can be fully utilized, effective identification of underwater transient signals will be achieved. Summary of the Invention:
[0004] The technical problem to be solved by the present invention is to provide a transient signal identification method based on multi-scale and multi-resolution feature fusion, which can effectively identify underwater transient signals.
[0005] The technical solution of the present invention is to provide a transient signal identification method based on multi-scale and multi-resolution feature fusion. First, extract the time-domain features of the transient signal; then, expand the transient time-domain signal through upsampling or deconvolution to generate a multi-resolution time-frequency spectrum, and then use a feature pyramid to achieve multi-scale and multi-resolution feature spectrum fusion to obtain multi-scale and multi-resolution features; next, perform feature fusion based on the cross-attention mechanism on the time-domain features and the multi-scale and multi-resolution time-frequency features to achieve joint learning in the time domain and the time-frequency domain; finally, use a classifier to identify the types of transient signals.
[0006] The specific steps are as follows:
[0007] Step 1: Extract time-domain features of the transient signal. Input the transient time-domain signal into the time-domain feature extraction network to extract the energy distribution characteristics of the transient signal in the time domain, and obtain the output feature X t . The feature extraction network can be 1D-ViT, TCN, 1D-ResNet, etc.
[0008] Step 2: Extract multi-scale and multi-resolution time-frequency features of the transient signal. Specifically as follows:
[0009] Step 2.1, to obtain multi-scale and multi-resolution time-frequency features from the transient signal with a short duration, use one of the following two methods to expand the original transient time-domain signal to N1, N2, L, N times the original length, where M is the number of selected multi-scales. M times, where M is the number of selected multi-scales.
[0010] Method 1: Perform deconvolution on the original transient signal using multiple one-dimensional deconvolution filters with different parameters;
[0011] Method 2: Upsample the original transient signal using methods such as linear interpolation and spline interpolation.
[0012] Step 2.2, preprocess the original transient signal and the expanded transient signal respectively, transform them into time-frequency spectra, and obtain M+1 multi-resolution time-frequency spectra S1, S2, L, S M+1 . The preprocessing method can be high-resolution time-frequency analysis methods such as wavelet transform, WVD transform, and Hilbert-Huang transform, or a combination of multiple time-frequency analysis methods.
[0013] Step 2.3, construct a time-frequency feature extraction network. The network contains M+1 branches, each branch contains several basic network units, and each basic network unit consists of several convolutional layers, batch normalization layers, and pooling layers. The network input is M+1 time-frequency spectra, and the output is feature X f .
[0014] Each branch network processes the features of one of the time-frequency spectra. In addition, a feature pyramid structure is used for cross-branch multi-scale feature fusion between different branches.
[0015] Denote the j-th basic network unit of the i-th network branch as F i,j (g), the input feature spectrum of F i,j (g) is f(i,j). Upsample or deconvolve f(i,j) to the size of the next-layer branch feature spectrum f(i+1,j) to obtain g(i+1,j), and superimpose g(i+1,j) and f(i+1,j) as the basic network unit F i+1,j(g) input, through this cross-layer fusion, can achieve the fusion of multi-scale and multi-resolution feature spectra.
[0016] Step 3: For the time-domain feature X t and the multi-scale and multi-resolution time-frequency feature X f perform feature fusion based on cross-attention to achieve joint learning of the transient signal in the time domain and time-frequency domain.
[0017] Step 4: Classify the fused features using a Softmax classifier, and the loss function uses the cross-entropy loss function.
[0018]
[0019] where N is the total number of samples, y i is the true label of the transient signal, is the model-estimated label.
[0020] Compared with the prior art, the present invention has the following advantages:
[0021] (1) The present invention expands the transient signal using operations such as deconvolution or upsampling, and on this basis extracts multi-scale and multi-resolution time-frequency features, and uses the feature pyramid network structure to fuse multi-scale and multi-resolution time-frequency features such as low-resolution, high-resolution, small-scale, and large-scale, making full use of the information of the transient signal within the limited time bandwidth, and solving the problem that it is difficult to extract features due to the short duration of the transient signal.
[0022] (2) The present invention uses deep learning technology to extract time-domain features and multi-scale and multi-resolution time-frequency features from the transient signal, and performs feature fusion based on the attention mechanism, realizing the joint feature extraction and utilization of the transient signal in the time domain and time-frequency domain, and effectively improving the identification performance of the transient signal. Description of the Drawings:
[0023] Figure 1 is the flow chart of the present invention.
[0024] Figure 2 is the schematic diagram of the time-domain waveform and time-frequency spectrum of different transient signals.
[0025] Figure 3 is the schematic diagram of the transient signal identification network structure.
[0026] Figure 4 is the schematic diagram of the network structure of the basic feature extraction module. Specific Embodiments:
[0027] The present invention will be further described below in conjunction with the drawings in terms of specific embodiments:
[0028] Figure 1It is the flowchart of the transient signal identification method based on multi-scale and multi-resolution feature fusion of the present invention. Figure 2 The time-domain waveforms and time-frequency spectrum schematic diagrams of different types of transient signals are given. Among them, Figure 2 (a) of gives the time-domain waveform 1 of the transient signal. Figure 2 (b) of gives the time-domain waveform 2 of the transient signal. Correspondingly, Figure 2 (c) of gives the time-frequency spectrum 1 of the transient signal, while Figure 2 (d) of gives the time-frequency spectrum 2 of the transient signal. It can be seen that different transient signals have different characteristics such as different durations and energy decay rates in the time domain; in the time-frequency domain, they have different time-frequency distribution characteristics. Therefore, comprehensively mining the time-domain and time-frequency domain characteristics of transient signals will provide characteristic support for the identification of different types of transient signals. Figure 3 The schematic diagram of the network structure of the present invention is given, which mainly includes four parts: a time-domain feature extraction module, a multi-scale and multi-resolution time-frequency feature extraction module, a feature fusion module, and a classification module.
[0029] The specific operations are as follows:
[0030] Step 1: Extract the time-domain features of the transient signal. Input the transient time-domain signal into the time-domain feature extraction network to extract the energy distribution characteristics of the transient signal in the time domain, and obtain the output feature X t . The feature extraction network can be 1D-ViT, TCN, 1D-ResNet, etc. In this embodiment, TCN is selected as the feature extraction network.
[0031] Step 2: Extract the multi-scale and multi-resolution time-frequency features of the transient signal. Specifically as follows:
[0032] 2.1 To obtain multi-scale and multi-resolution time-frequency features from transient signals with shorter durations, the method of deconvolution or upsampling is used to expand the original transient time-domain signal. In this embodiment, the original transient signal is upsampled to 4 times and 8 times respectively.
[0033] 2.2 Perform wavelet transforms on the original transient signal and the expanded transient signal respectively to obtain three wavelet transform time-frequency spectra with different sizes and resolutions.
[0034] 2.3 Construct a time-frequency feature extraction network. The network contains 3 branches, each branch contains several basic network units, and each basic network unit is stacked by basic modules. The schematic diagram of the network structure of the basic module is as Figure 4 shown.
[0035] Each branch network processes the features of one of the time-frequency spectra. In addition, a feature pyramid structure is used for cross-branch multi-scale feature fusion between different branches.
[0036] Denote the j-th basic network unit of the i-th network branch as F i,j (g), and the input feature spectrum of F i,j (g) is f(i, j). Upsample or deconvolve f(i, j) to the size of the next-layer branch feature spectrum f(i + 1, j) to obtain g(i + 1, j), and superimpose g(i + 1, j) and f(i + 1, j) as the input of the basic network unit F i+1,j (g). In this embodiment, the feature spectrum of the first-layer network branch is upsampled by 4 times and then superimposed with the feature spectrum of the second-layer network branch. The feature spectrum of the second-layer network branch is upsampled by 2 times and then superimposed with the feature spectrum of the third-layer network branch. The original input feature spectrum resolutions of different network branches are different. Through cross-layer fusion, multi-scale and multi-resolution feature spectra such as low resolution, high resolution, small scale, and large scale can be fused.
[0037] Step 3: Perform feature fusion based on cross-attention on the time-domain feature X t and the multi-scale and multi-resolution time-frequency feature X f to achieve joint learning of the transient signal in the time domain and time-frequency domain.
[0038] Step 4: Classify the fused features using a Softmax classifier, and the loss function uses the cross-entropy loss function, where N is the total number of samples, y i is the true label of the transient signal, is the model-estimated label.
[0039] The method of the present invention makes full use of the information within the limited time bandwidth of the transient signal, solves the problem that it is difficult to extract features due to the short duration of the transient signal, and at the same time uses deep learning technology to achieve multi-scale and multi-resolution feature fusion, and feature fusion in the time domain and time-frequency domain, improves the feature representation performance of the transient signal, and realizes transient signal identification.
[0040] The above is only an illustration of the preferred embodiments of the present invention, and it should not be construed as a limitation to the claims. All equivalent process transformations made using the specification of the present invention are included in the patent protection scope of the present invention.
Claims
1. A transient signal identification method based on multi-scale and multi-resolution feature fusion, characterized in that: including the following steps, Step 1: Extract time-domain features from the transient signal. Input the transient time-domain signal into the time-domain feature extraction network to extract the energy distribution characteristics of the transient signal in the time domain, and obtain the output feature X t ; Step 2, perform multi-scale and multi-resolution time-frequency feature extraction on the transient signal; Step 3, perform feature fusion based on cross-attention on the time-domain feature X t and the multi-scale and multi-resolution time-frequency feature X f to achieve joint learning of the transient signal in the time domain and the time-frequency domain; Step 4, classify the fused features using a Softmax classifier, and the loss function uses the cross-entropy loss function, where N is the total number of samples, and y i is the true label of the transient signal, and is the estimated label of the model.
2. The transient signal identification method based on multi-scale and multi-resolution feature fusion according to claim 1, wherein: The feature extraction network can be any one of 1D-ViT, TCN, and 1D-ResNet.
3. The transient signal identification method based on multi-scale and multi-resolution feature fusion according to claim 1, wherein: The specific operation of Step 2 is as follows, Step 2.1, perform deconvolution on the original transient signal using multiple one-dimensional deconvolution filters with different parameters to expand the original transient time-domain signal to N1, N2, L, N times its original length, where M is the number of selected multi-scales; M times, where M is the number of selected multi-scales; Step 2.2, perform preprocessing on the original transient signal and the extended transient signal respectively, transform them into time-frequency spectra, and obtain M + 1 multi-resolution time-frequency spectra S1, S2, …, SM+1 M+1 ; Step 2.3, construct a time-frequency feature extraction network. The network contains M+1 branches, each branch contains several basic network units, and each basic network unit consists of several convolutional layers, batch normalization layers, and pooling layers. The network input is M+1 time-frequency spectra, and the output is the time-frequency feature X f .
4. The transient signal identification method based on multi-scale and multi-resolution feature fusion according to claim 3, characterized in that: In Step 2.1, when expanding the original transient time-domain signal, it can be replaced with the following method, and the original transient signal is upsampled using methods such as linear interpolation and spline interpolation.
5. The transient signal identification method based on multi-scale and multi-resolution feature fusion according to claim 3, characterized in that: The preprocessing method can be one or a combination of wavelet transform, WVD transform, and Hilbert-Huang transform.
6. The transient signal identification method based on multi-scale and multi-resolution feature fusion according to claim 3, characterized in that: In step 2.3, each branch network processes the features of one of the time-frequency spectra, and a feature pyramid structure is adopted between different branches for cross-branch multi-scale feature fusion. Among them, the j-th basic network unit of the i-th network branch is denoted as F i,j (g), the input feature spectrum of F i,j (g) is f(i, j). Upsample or deconvolve f(i, j) to the size of its next-layer branch feature spectrum f(i + 1, j) to obtain g(i + 1, j), and stack g(i + 1, j) and f(i + 1, j) as the input of the basic network unit F i+1,j (g). Through this cross-layer fusion, the fusion of multi-scale and multi-resolution feature spectra is achieved.