Passive sonar target identity correlation method based on space-time frequency feature fusion
By using a deep neural network model that fuses spatial, temporal, and frequency features, the problems of low feature utilization and poor environmental adaptability in traditional passive sonar target identification are solved, and more efficient target identity association is achieved.
Patent Information
- Application Number
- CN202511816343.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-04
- Publication Date
- 2026-02-27
AI Technical Summary
Traditional passive sonar target recognition methods rely on artificial physical feature construction, resulting in low feature utilization, difficulty in accurate extraction, poor environmental adaptability, and difficulty in achieving efficient target identity association.
By employing a deep neural network model that fuses spatial, temporal, and frequency features, and constructing standardized samples and an end-to-end deep neural network, we can achieve in-depth mining and dynamic fusion of the spatial, temporal, and frequency features of the target, thereby improving the accuracy and reliability of identity association.
It significantly improves the confidence and generalization ability of passive sonar target identity association, overcomes the problems of single and fragmented features in traditional methods, and achieves more efficient feature mining and fusion.
Smart Images
Figure CN121578282A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of underwater target recognition, and mainly relates to a passive sonar target identity correlation method based on space-time-frequency feature fusion. BACKGROUND
[0002] Passive sonar target recognition is a core technology of modern underwater detection systems. Identity correlation is an important means to realize reliable target recognition by comparing and identifying the passive target data obtained in real time on site with the target features in the library. The traditional method mainly relies on manual physical feature construction extraction and template matching method, and has the disadvantages of low feature utilization rate, great difficulty in accurate extraction, poor environmental adaptability, etc.
[0003] At present, the deep learning technology develops very rapidly and has been researched and applied in the fields of face verification and speaker recognition. In the field of passive sonar target recognition, deep learning has also been widely concerned, but most of them are aimed at processing the noise frequency domain features of the tracked target. Passive sonar targets have different mechanical structures, working modes and navigation situations, and all show corresponding characteristics in space, time and frequency. Therefore, by mining and utilizing the space-time-frequency features of the target, the target identification capability can be effectively improved.
[0004] Therefore, there is an urgent need for a passive sonar target identity correlation method which can comprehensively utilize the space-time-frequency multi-dimensional features and realize deeper feature mining and fusion, so as to improve the accuracy and reliability of the correlation. SUMMARY
[0005] The purpose of the present application is to overcome the deficiencies in the prior art and provide a passive sonar target identity correlation method based on space-time-frequency feature fusion. The method realizes deep mining, dynamic fusion and efficient correlation of the space-time-frequency features of the target by constructing a standardized multi-domain sample and an end-to-end deep neural network model, thereby significantly improving the confidence and generalization ability of the identity correlation.
[0006] The purpose of the present application is achieved by the following technical scheme. A passive sonar target identity correlation method based on space-time-frequency feature fusion comprises the following steps: (1) Data preprocessing and sample generation: target detection and tracking are performed on the acoustic signals received by the passive sonar, and space domain samples, time domain samples and frequency domain samples are constructed, and the three are combined to generate standardized samples, and further construct a training sample set and a specific target sample set; (2) Constructing an air-time-frequency multi-dimensional feature-mining-cross-domain fusion-identity correlation network model: including parallel space domain feature extraction module, time domain feature extraction module and frequency domain feature extraction module, and serial cross-domain dynamic attention fusion module and feature secondary fusion and identity correlation module; (3) Model training: design loss function and optimizer, generate standardized comparison samples based on training sample set, train the model until convergence; (4) Identity correlation decision: generate standardized samples for unknown targets, compare with samples in specific target sample set, make identity correlation decision based on model output confidence.
[0007] As preferably, the data preprocessing and sample generation in step (1) comprises: (1.1) Target detection and tracking: adopt uniform linear array to passively receive acoustic signals, calculate spatial spectrum through conventional beamforming, track targets based on Kalman filtering, output target waveform sequence, bearing and signal-to-noise ratio every 0.5s, start sample construction after accumulating 60 batches; (1.2) Spatial sample construction: normalize time-bearing history and time-signal-to-noise ratio history and splice into 1×60×2 spatial samples; (1.3) Temporal sample construction: extract 512-point sequence from target waveform sequence and normalize to obtain temporal feature 1; extract envelope through Hilbert transform and normalize to obtain temporal feature 2; splice the two into 1×30720×2 temporal samples; (1.4) Frequency sample construction: perform LOFAR spectrum and DEMON spectrum analysis on time-amplitude sequence to obtain frequency feature 1 and frequency feature 2 respectively, splice into 60×1024×2 frequency samples; (1.5) Combine spatial, temporal and frequency samples into standardized samples; (1.6) Generate training sample set and specific target sample set based on labeled data.
[0008] As preferably, the space-time-frequency multi-dimensional feature-mining-cross-domain fusion-identity correlation network model in step (2) comprises: (2.1) Build basic convolution operator: including basic module 1 and basic module 2, each module contains parallel branches, respectively performing convolution, LayerNorm, GELU activation, and adding output with straight-through branch; (2.2) Spatial feature extraction module: sequentially includes convolution layer, basic module 1, bidirectional LSTM layer and global average pooling, outputs 512-dimensional feature vector; (2.3) Frequency feature extraction module: sequentially includes convolution layer, basic module 2 and global average pooling, outputs 512-dimensional feature vector; (2.4) Temporal feature extraction module: includes three parallel convolution branches, after splicing, sequentially passes through convolution layer, basic module 1, bidirectional LSTM layer and global average pooling, outputs 512-dimensional feature vector; (2.5) Cross-domain dynamic attention fusion module: project the spatial, frequency and time domain features into a common space, input the concatenated features into a multi-head attention sub-module, and output the fused features; (2.6) Secondary feature fusion and identity association module: including nonlinear transformation and residual connection, feature flattening, feature dimension reduction and classification sub-module, output identity association result and confidence.
[0009] As a preferred, the cross-domain dynamic attention fusion module comprises: (2.5.1) Multi-domain feature projection sub-module: project each domain feature into query, key and value vectors through a learnable linear transformation; (2.5.2) Attention input sub-module: concatenate the query, key and value vectors of each domain into a matrix; (2.5.3) Multi-head attention calculation sub-module: including 8 attention heads, each head calculates the subspace projection and outputs the attention weighting value, and finally concatenates the fused features.
[0010] As a preferred, the secondary feature fusion and identity association module comprises: (2.6.1) Nonlinear transformation and residual connection sub-module: feature transformation through fully connected layer and residual connection; (2.6.2) Feature flattening sub-module: flatten the feature into a 1x768 dimensional vector; (2.6.3) Feature dimension reduction and classification sub-module: output the confidence of whether it is the same target through fully connected layer, Dropout and Softmax function.
[0011] As a preferred, the model training in step (3) comprises: (3.1) Design the loss function and use AdamW optimizer, configure the cosine annealing learning rate scheduling strategy; (3.2) Randomly select two standardized samples from the training samples, concatenate their spatial, temporal and frequency domain samples in the channel dimension to generate standardized comparison samples, and the label is whether it is the same target; (3.3) Repeat the selection of comparison samples for training until the model converges.
[0012] As a preferred, the identity association decision in step (4) comprises: (4.1) Generate a standardized sample for the unknown target; (4.2) Combine each sample in the specific target sample set into a comparison sample one by one, input the trained model to get the confidence; (4.3) Group the confidence by target, calculate the average value of each group, and if it exceeds the set threshold, it is determined as the same target, and the result is output according to the confidence.
[0013] Preferably, the confidence threshold is 0.7.
[0014] The beneficial effects of this invention are as follows: This invention proposes a passive sonar target identity association method based on spatiotemporal-frequency feature fusion. It designs a spatiotemporal-frequency standardized sample construction method, combines convolution and attention mechanisms, and considers sample characteristics to construct a spatiotemporal-frequency multi-dimensional feature-mining-cross-domain fusion-identity association network model. This achieves end-to-end integrated multi-domain feature extraction, deep fusion, and identity association. Compared with existing passive sonar target identity association methods, the method proposed in this invention offers more comprehensive feature mining, stronger fusion and utilization capabilities, and can effectively improve the stability and generalization of identity association. Specifically, it is manifested in: 1. Comprehensive utilization of multi-dimensional features: For the first time, the spatial (azimuth, signal-to-noise ratio), temporal (waveform, envelope), and frequency (LOFAR spectrum, DEMON spectrum) features of the target are jointly modeled, overcoming the problem of single features in traditional methods and making more comprehensive use of information.
[0015] 2. Deep Feature Mining and Adaptive Fusion: By using parallel deep convolutional networks and LSTM networks to mine deep features in each domain, and innovatively introducing a cross-domain dynamic attention mechanism, it can adaptively learn and fuse the importance of features from different domains, thereby improving the representational ability of the fused features.
[0016] 3. End-to-end integrated association: An end-to-end model is constructed from the original multi-domain data to the identity association result, which avoids the problem of feature extraction and classification decision being separated in traditional methods, and improves the overall performance and efficiency.
[0017] 4. Significantly improved association performance: Validated by sea trial data, the identity association accuracy of the method of this invention is significantly higher than that of traditional line spectrum matching methods and deep learning methods that only use frequency domain features, with higher confidence and better environmental adaptability. Attached Figure Description
[0018] Figure 1 This is a flowchart of the overall signal processing method of the present invention.
[0019] Figure 2 This is a schematic diagram illustrating the working principle of the algorithm proposed in this invention.
[0020] Figure 3 This is a diagram of the basic convolutional residual operator structure used to construct the feature extraction module in this invention, where (a) is basic module 1 and (b) is basic module 2.
[0021] Figure 4 This is a top-level feature scatter plot obtained by processing sea trial data using the method of this invention. Detailed Implementation
[0022] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort belong to the protection scope of the present application.
[0023] As shown in the figure, the implementation of the method of the present application mainly includes four stages: data preprocessing and sample set construction, space-time-frequency multi-dimensional feature-mining-cross-domain fusion-identity correlation network model construction, space-time-frequency multi-dimensional feature-mining-cross-domain fusion-identity correlation network model training, and identity correlation decision. Figure 1 The mechanism of the present application is as follows: first, data preprocessing is carried out, and on the basis of target detection and tracking, space domain samples are generated through target time-azimuth history and target time-signal-to-noise ratio history analysis, time domain samples are generated through target time-amplitude waveform and envelope analysis, and frequency domain samples are generated through target LOFAR spectrum and DEMON spectrum analysis. Standardized samples are obtained by integrating the above multi-domain samples, and further, training sample sets and specific target sample sets are generated based on the above processing mode; second, a space-time-frequency multi-dimensional feature-mining-cross-domain fusion-identity correlation network model is constructed, and based on convolution and residual operators and LSTM operators, parallel space domain feature extraction modules, time domain feature extraction modules and frequency domain feature extraction modules are constructed. On this basis, based on attention mechanism operators and convolution, full connection and other operators, a serial cross-domain dynamic attention fusion module and a feature secondary fusion and identity correlation module are constructed; third, a model training loss function is designed and an optimizer is configured, standardized comparison samples are generated based on the training sample set, and model training is completed; finally, unknown passive target noise is correlated with the specific target sample set, and the identity correlation of unknown targets and specific targets is completed based on confidence statistics and threshold determination. Compared with the existing passive sonar target identity correlation method, the method proposed in the present application has more in-depth mining and utilization of target useful feature information, and has higher identity correlation confidence.
[0024] As shown in the figure, the working principle of the algorithm proposed in the present application is as follows: by using space-time-frequency multi-domain features, target samples are constructed, and standardized samples are generated by combining two samples to be compared. On this basis, an end-to-end space-time-frequency multi-dimensional feature-mining-cross-domain fusion-identity correlation network model is designed based on various deep learning operators, realizing integrated processing from original samples to target identity correlation.
[0025] Figure 2 The specific steps are as follows:
[0026] The specific steps are as follows: (1) Data preprocessing and sample generation, the main process is as follows: (1.1) Target detection and tracking: uniform linear array is used to passively receive acoustic signals, and spatial spectrum is calculated by conventional beam forming (CBF), azimuth range 0~180°, resolution 1°; select a certain azimuth target from the spatial spectrum and track based on Kalman filtering algorithm, output target waveform sequence (length 0.5 , s), the azimuth and the signal-to-noise ratio relative to the background, denoted as 1 batch of characteristic quantities, and the above output characteristic quantities are accumulated in time sequence to obtain time-azimuth, time-signal-to-noise ratio and time-amplitude (waveform) sequence; when the target tracking time reaches 30s, i.e. 60 batches are accumulated, sample construction begins; (1.2) Spatial sample construction: time-azimuth history and time-signal-to-noise ratio history are normalized respectively to obtain spatial feature 1 and spatial feature 2, both of which are 1×60 length sequences; spatial feature 1 and spatial feature 2 are spliced to obtain a spatial sample of 1×60×2.
[0027] (1.3) Time domain sample construction: each batch of target waveform sequence in the time-amplitude sequence is processed, and 512 point sequences are taken from the middle, and all 60 batches of sequences are further arranged in time and normalized to form a time domain feature 1 with a length of 1×30720; the time-amplitude sequence envelope is obtained by Hilbert transform, and the envelope is further linearly decimated and normalized to obtain a time domain feature 2 with a length of 1×30720; time domain feature 1 and time domain feature 2 are spliced to obtain a time domain sample of 1×30720×2.
[0028] (1.4) Frequency domain sample construction: LOFAR spectrum analysis is performed on the time-amplitude sequence to generate a LOFAR spectrum, with a frequency resolution of not less than 0.1Hz, and further linearly decimated and normalized to obtain a frequency domain feature 1 of 60×1024 dimensions; DEMON spectrum analysis is performed on the time-amplitude sequence to generate a DEMON spectrum, with a frequency resolution of not less than 0.01Hz, and further linearly decimated and normalized to obtain a frequency domain feature 2 of 60×1024 dimensions; frequency domain feature 1 and frequency domain feature 2 are spliced to obtain a time domain sample of 60×1024×2.
[0029] (1.5) Combine the spatial domain sample, the frequency domain sample and the time domain sample to generate a standardized sample.
[0030] (1.6) Sample set construction: according to the method described in steps (1.1)~(1.5), process the labeled actual data to generate a training sample set, and select specific target samples of interest to generate a specific target sample set.
[0031] (2) Spatiotemporal multi-dimensional feature-mining-cross-domain fusion-identity correlation network model construction, the main process is as follows: (2.1) Basic convolution operator construction, Figure 3 The basic module used when constructing the spatial feature extraction module, the temporal feature extraction module and the frequency domain feature extraction module in the patent is shown, and by configuring different size convolution kernels, residual convolution operators suitable for target feature mining are formed, and the specific process is as follows.
[0032] (2.1.1) Construct basic module 1, the output channel number is a settable parameter x. Add 2 parallel branches. Branch 1 includes convolution layer(1x7,4x,[1,1]) in turn, LayerNorm layer, convolution layer(1x3,x,[1,1]), GELU activation function, wherein the convolution layer parameter(1x7,x,[1,1]) represents the convolution kernel size is 1x7, the convolution channel number is x, the convolution kernel horizontal and vertical step length is 1, the same below; Branch 2 is a straight connection layer; the convolution features output by the two branches are added to obtain the final output result of the module.
[0033] (2.1.2) Construct basic module 2, the output channel number is a settable parameter x. Add 2 parallel branches. Branch 1 includes convolution layer(3x1,2x,[1,1]) in turn, LayerNorm layer, convolution layer(1x3,x,[1,1]), GELU activation function; Branch 2 is a straight connection layer; the convolution features output by the two branches are added to obtain the final output result of the module.
[0034] (2.2) Spatial feature extraction module construction: add convolution layer(1x21,64,[1,2]), convolution layer(1x14,128,[1,2]), 3 basic module 1 (output channel number x is set to 256), bidirectional LSTM layer (contains 256 hidden units), global average pooling layer in turn. Finally, the module outputs a 512-dimensional feature vector.
[0035] (2.3) Frequency domain feature extraction module construction: add convolution layer(3x7,32,[1,2]), convolution layer(3x5,128,[1,2]), convolution layer(3x3,256,[2,2]), 3 basic module 2 (output channel number x is set to 256), global average pooling layer in turn. Finally, the module outputs a 512-dimensional feature vector.
[0036] (2.4) Temporal feature extraction module construction: add 3 parallel branches, branch 1 is a convolutional layer (1x128, 64, [1, 10]), branch 2 is a convolutional layer (1x256, 64, [1, 10]), branch 3 is a convolutional layer (1x512, 64, [1, 10]); add a branch feature concatenation layer to concatenate the convolutional features output by the three branches in the channel dimension; add a convolutional layer (1x64, 64, [1, 4]), a convolutional layer (1x16, 128, [1, 4]), 3 basic module 1s (the output channel number x is set to 256), a bidirectional LSTM layer (containing 256 hidden units), and a global average pooling layer in turn. Finally, the module outputs a 512-dimensional feature vector.
[0037] (2.5) Cross-domain dynamic attention fusion module construction, the main process is as follows: (2.5.1) Add a multi-domain feature projection sub-module: project the input feature vectors of the three domains into a common feature space through a learnable linear transformation to obtain , , . Among them, , , . Wherein, , represents the domain index, , , represent the spatial domain, the frequency domain and the time domain respectively; is the input feature vector of the i-th domain, with a dimension of 512; is the projection weight matrix of the query vector, is the projection weight matrix of the key vector, is the projection weight matrix of the value vector, , , , all have a dimension of 256x512; is the bias term of the query vector, is the bias term of the key vector, is the bias term of the value vector, , , all have a dimension of 1x256; is the projected query vector, is the projected key vector, is the projected value vector, , , all have a dimension of 1x256.
[0038] (2.5.2) Add an attention input sub-module: project the input feature vectors of the three domains into a common feature space through a learnable linear transformation to obtain , , (2.5.2) Concatenate in the preset order on dimension 1 to obtain the query matrix Dimensions are both 3x256. represents the concatenation operation.
[0039] (2.5.3) Add a multi-head attention calculation submodule: a total of 8 attention heads. For each attention head , calculate the subspace projection matrix , where are the query projection matrix, key projection matrix and value projection matrix of the attention head respectively, and the dimensions are both 256x32; calculate the attention score matrix , where represents the transpose of the matrix ; add a Sigmoid activation function to output ; calculate the output of the attention head ; on this basis, concatenate the outputs of each attention head on the 2nd dimension, i.e. Dimensions are both 3x256.
[0040] (2.6) Feature secondary fusion and identity association module construction, the main process is as follows: (2.6.1) Add a nonlinear transformation and residual connection submodule: take as the input, add a fully connected layer (256, 512) and a fully connected layer (512, 256), and the output is , where fully connected layer (256, 512) represents that the input node number of the layer is 256 and the output node number is 512, and the same below; add a residual connection layer to realize the output , where are weighting coefficients, which are set to 1, 0.25 and 0.25 respectively here; sequentially add a LayerNorm normalization layer, a convolution layer (1x1, 512, [1, 1]), a convolution layer (1x1, 1024, [1, 1]), a convolution layer (1x1, 256, [1, 1]) to obtain the output ; add a residual connection layer to realize the output ; add a LayerNorm normalization layer.
[0041] (2.6.2) Add a feature flattening submodule to obtain wherein represents a flattening operation, the dimension is 1x768; (2.6.3) Add a feature dimension reduction and classification submodule: sequentially add a fully connected layer (768, 512), Dropout (0.5), a fully connected layer (512, 256), Dropout (0.5), a fully connected layer (256, 2), and a Softmax function to obtain the final identity association result and directly calculate the confidence, including the same target and not the same target, so the number of output nodes is set to 2.
[0042] (3) Temporal-spatial-frequency feature mining and fusion recognition network model training, the main process is as follows.
[0043] (3.1) Design a loss function wherein is the BatchSize size, is the sample true label, is the sample predicted label, and are one-hot encoded, is a label smoothing function, which is set to here; set the optimizer to Adam.
[0044] (3.2) Train the temporal-spatial-frequency feature mining and fusion recognition network model based on the training sample set, as follows: (3.2.1) Randomly select two standardized samples from the training samples, and concatenate the spatial domain samples, time domain samples, and frequency domain samples of the two standardized samples in dimension 3 to generate a standardized comparison sample composed of spatial domain comparison samples, time domain comparison samples, and frequency domain comparison samples, with sizes of 1x60x4, 1x30720x4, and 60x1024x4, respectively; if the two standardized samples are the same target, the label is 1, otherwise it is 0; train the model based on the above standardized comparison sample.
[0045] (3.2.2) Repeat the selection of standardized comparison samples and the training of the model according to the method described in (3.2.1) until convergence.
[0046] (4) Identity association decision, the main process is as follows: (4.1) Preprocess the unknown target passive sonar space-time-frequency data according to the method in steps (1.1)-(1.5) to obtain the unknown target standardized sample.
[0047] (4.2) Select the specific target standardized sample from the specific target sample set in order.
[0048] (4.3) Combine the unknown target standardized sample and the specific target standardized sample into a standardized sample according to step (3.3.1), process the standardized sample based on the trained space-time-frequency multi-dimensional feature-mining-cross-domain fusion-identity correlation network model to obtain an identity correlation direct calculation confidence, if the two samples come from the same target, the confidence directly uses the direct calculation confidence, otherwise the confidence is 1-direct calculation confidence.
[0049] (4.4) Continue to traverse and select the specific target standardized sample of the specific target sample set in order, and complete the identity correlation according to steps (4.1)-(4.3), so that the identity correlation confidence of the unknown target standardized sample and all feature target samples can be obtained.
[0050] (4.5) Group the identity correlation confidence according to the source of the specific target standardized sample, that is, if the specific target standardized sample is generated by the same target, the identity correlation confidence thereof is grouped into a group; calculate the average value of the identity correlation confidence of each group, if the average value exceeds the decision threshold (set to 0.7 here), it is considered that there is an identity correlation possibility between the unknown target and the specific target, the higher the confidence, the greater the identity correlation possibility. All specific targets with identity correlation possibility are pushed as identity correlation results, and are arranged from high to low according to the confidence.
[0051] Experimental results Table 1 shows the results obtained by processing a batch of sea trial data containing multiple targets using the method based on the patent, compared with the traditional line spectrum matching method, the frequency domain feature identity correlation method based on LOFAR spectrum and deep convolutional neural network, it can be seen that the identity correlation accuracy of the method of the patent is obviously higher.
[0052] Table 1 Method Correctness of identity association / % Traditional line spectrum matching method 68.85 Frequency domain feature identity association method based on LOFAR spectrum and deep convolutional neural network 77.87 Patent model 83.61 Figure 4 The top-level result scatter plot when processing a batch of sea trial data containing multiple targets based on the method of the invention is shown, and each color scatter point in the figure represents a target, it can be seen that the method can integrate multi-domain input, effectively mine target individual characteristics, and provide key feature information for identity correlation, and the features extracted by the invention method can better distinguish different targets.
[0053] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of the changes or replacements within the scope disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.
Claims
1. A passive sonar target identity association method based on spatiotemporal-frequency feature fusion, characterized in that, Includes the following steps: (1) Data preprocessing and sample generation: Target detection and tracking are performed on the acoustic signals received by the passive sonar. Spatial domain samples, time domain samples and frequency domain samples are constructed respectively, and the three are combined to generate standardized samples. Training sample set and specific target sample set are then constructed. (2) Construct a multidimensional feature-mining-cross-domain fusion-identity association network model based on spatial, temporal and frequency domains: including parallel spatial feature extraction modules, temporal feature extraction modules and frequency domain feature extraction modules, as well as serial cross-domain dynamic attention fusion modules and feature secondary fusion and identity association modules; (3) Model training: Design loss function and optimizer, generate standardized comparison samples based on training sample set, and train the model until convergence; (4) Identity association judgment: Generate standardized samples for unknown targets, compare them with samples in a specific target sample set, and make identity association judgment based on the confidence level output by the model.
2. The passive sonar target identity association method based on spatiotemporal-frequency feature fusion according to claim 1, characterized in that, The data preprocessing and sample generation in step (1) include: (1.1) Target detection and tracking: A uniform linear array is used to passively receive acoustic signals. The spatial spectrum is calculated through conventional beamforming. The target is tracked based on Kalman filtering. The target waveform sequence, azimuth and signal-to-noise ratio are output every 0.5s. Sample construction begins after accumulating 60 batches. (1.2) Spatial sample construction: normalize the time-azimuth history and time-signal-noise ratio history and splice them into a 1×60×2 spatial sample; (1.3) Construction of time-domain samples: Extract 512 points from the target waveform sequence and normalize them to obtain time-domain feature 1; extract the envelope through Hilbert transform and normalize it to obtain time-domain feature 2; and concatenate the two to form a 1×30720×2 time-domain sample. (1.4) Frequency domain sample construction: LOFAR and DEMON spectrum analyses were performed on the time-amplitude sequence to obtain frequency domain feature 1 and frequency domain feature 2, which were then spliced together to form a 60×1024×2 frequency domain sample; (1.5) Combine spatial, temporal, and frequency domain samples into standardized samples; (1.6) Generate training sample sets and target sample sets based on labeled data.
3. The passive sonar target identity association method based on spatiotemporal-frequency feature fusion according to claim 2, characterized in that, The spatiotemporal-frequency multidimensional feature-mining-cross-domain fusion-identity association network model mentioned in step (2) includes: (2.1) Constructing basic convolution operators: including basic module 1 and basic module 2, each module contains parallel branches, which perform convolution, LayerNorm, and GELU activation respectively, and are added to the directly connected branches for output; (2.2) Spatial feature extraction module: It consists of a convolutional layer, a basic module 1, a bidirectional LSTM layer and global average pooling, and outputs a 512-dimensional feature vector. (2.3) Frequency domain feature extraction module: It consists of convolutional layers, basic module 2 and global average pooling, and outputs a 512-dimensional feature vector. (2.4) Temporal feature extraction module: includes three parallel convolutional branches, which are concatenated and then passed through a convolutional layer, basic module 1, bidirectional LSTM layer and global average pooling in sequence to output a 512-dimensional feature vector; (2.5) Cross-domain dynamic attention fusion module: Projects spatial, frequency, and temporal features onto a common space, stitches them together, inputs them into the multi-head attention submodule, and outputs fused features; (2.6) Feature secondary fusion and identity association module: including nonlinear transformation and residual connection, feature flattening, feature dimensionality reduction and classification sub-modules, outputting identity association results and confidence.
4. The passive sonar target identity association method based on spatiotemporal-frequency feature fusion according to claim 3, characterized in that, The cross-domain dynamic attention fusion module includes: (2.5.1) Multi-domain feature projection submodule: Projects the features of each domain into query, key, and value vectors through learnable linear transformations; (2.5.2) Attention Input Submodule: Concatenates the query, key, and value vectors of each domain into a matrix; (2.5.3) Multi-head attention calculation submodule: includes 8 attention heads, each of which calculates the subspace projection and outputs the attention weight value, and finally splices them into a fused feature.
5. The passive sonar target identity association method based on spatiotemporal-frequency feature fusion according to claim 3, characterized in that, The feature secondary fusion and identity association module includes: (2.6.1) Nonlinear Transformation and Residual Connection Submodule: Feature transformation is performed through fully connected layers and residual connections; (2.6.2) Feature Flattening Submodule: Flattens the features into a 1×768-dimensional vector; (2.6.3) Feature dimensionality reduction and classification submodule: Through fully connected layers, Dropout and Softmax functions, output the confidence score of whether they are the same target.
6. The passive sonar target identity association method based on spatiotemporal feature fusion according to claim 1, characterized in that, The model training described in step (3) includes: (3.1) Design the loss function and use the AdamW optimizer, and configure the cosine annealing learning rate scheduling strategy; (3.2) Randomly select two standardized samples from the training samples, and splice their spatial, temporal, and frequency domain samples in the channel dimension to generate standardized comparison samples, labeled as whether they are the same target; (3.3) Repeatedly select comparison samples for training until the model converges.
7. The passive sonar target identity association method based on spatiotemporal-frequency feature fusion according to claim 1, characterized in that, The identity association decision mentioned in step (4) includes: (4.1) Generate standardized samples for unknown targets; (4.2) Combine each sample with a specific target sample set to form a comparison sample, and input it into the trained model to obtain the confidence score; (4.3) Group the confidence scores by target, calculate the average value of each group, and if the average value exceeds the set threshold, the group is determined to be the same target. Sort the results by confidence score and output them.
8. The passive sonar target identity association method based on spatiotemporal-frequency feature fusion according to claim 7, characterized in that, The confidence threshold is 0.7.