A method for intelligent extraction and classification of physical features of multidimensional underwater acoustic targets

By constructing a multi-dimensional physical feature intelligent extraction and classification network model, the problem of accurate extraction and recognition of traditional underwater acoustic targets under low signal-to-noise ratio was solved, and more efficient feature fusion and target classification were achieved.

CN117668659BActive Publication Date: 2026-11-17THE 715TH RES INST OF CHINA SHIPBUILDING IND CORP +1
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202311665695.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-12-06
Publication Date
2026-11-17
Estimated Expiration
2043-12-06

AI Technical Summary

Technical Problem

Traditional underwater acoustic target recognition methods struggle to achieve accurate physical feature extraction and target recognition under low signal-to-noise ratio conditions, and are significantly affected by ocean channels and background noise.

Method used

A multi-dimensional physical feature intelligent extraction and classification network model is constructed, including line spectrum feature and propeller harmonic feature extraction network units. Feature fusion and target classification are achieved through multi-stage training and loss function weighting.

Benefits of technology

It improves the accuracy of target physical feature extraction and recognition in low signal-to-noise ratio environments, and enhances the comprehensive utilization capability of underwater acoustic target classification.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117668659B_ABST
    Figure CN117668659B_ABST
Patent Text Reader

Abstract

This invention discloses a method for intelligent extraction and classification of multidimensional underwater acoustic target physical features, comprising: constructing a multidimensional physical feature intelligent extraction and classification network model, including a pre-set parallel line spectrum feature extraction network unit and a propeller harmonic feature extraction network unit, and a post-set multidimensional physical feature fusion and classification unit; constructing a loss function for the model, which is a weighted sum of the line spectrum feature extraction loss, the propeller harmonic feature extraction loss, and the multidimensional physical feature fusion and classification loss; using labeled underwater target radiated noise data, training the model in multiple stages, with each stage using weighted loss function values ​​and model training parameters to achieve model convergence; using the trained model to identify unknown underwater acoustic target radiated noise data and outputting classification results. This invention improves the recognition effect by extracting two typical physical features, line spectrum features and propeller harmonic features, and achieving target classification and recognition through feature fusion.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of underwater target recognition and artificial intelligence technology, specifically relating to a method for intelligent extraction and classification of multidimensional underwater acoustic target physical features. Background Technology

[0002] Underwater target radiated noise identification, a crucial function of passive sonar, provides vital information for comprehensive decision-making and is one of the main research directions in underwater acoustic signal processing. Traditional target classification and identification typically starts from the target's acoustic mechanism, combining signal processing methods to construct and extract separable physical features, and then using template matching, classifiers, and other methods to achieve target identification. Line spectra and propeller parameters, as typical physical features, have been applied in many underwater target identification scenarios. However, due to factors such as ocean channels, background noise, and multi-target interference, the underwater acoustic data received by sonar is highly complex, and target noise often has a very low signal-to-noise ratio. Traditional signal processing methods are limited by their performance bottlenecks, making it difficult to achieve accurate estimation and extraction of weak features, thus hindering the improvement of target identification performance.

[0003] Currently, deep learning is a hot research area in machine learning and has been widely applied in fields such as computer vision and natural language processing. Deep learning, by employing a multi-hidden-layer structure with complex connections, can achieve complex computational and fitting capabilities. Therefore, applying deep learning to extract physical features of underwater acoustic targets holds promise for achieving corresponding feature estimation and extraction by learning the nonlinear relationship between high-dimensional raw data and feature values. Summary of the Invention

[0004] The purpose of this invention is to provide a method for intelligent extraction and classification of multidimensional underwater acoustic target physical features to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for intelligent extraction and classification of multidimensional underwater acoustic target physical features includes the following steps:

[0007] Construct a multidimensional physical feature intelligent extraction and classification network model, including,

[0008] Two feature extraction network units are constructed in a pre-positioned and parallel configuration, used for line spectrum feature extraction and propeller harmonic feature extraction, respectively.

[0009] A post-constructed multi-dimensional physical feature fusion and classification recognition unit is constructed for feature fusion and target classification recognition based on fused features;

[0010] The loss function of the intelligent extraction and classification network model of multidimensional physical features is constructed as a weighted sum of the line spectrum feature extraction loss, the propeller harmonic feature extraction loss, and the multidimensional physical feature fusion and classification recognition loss.

[0011] Using labeled underwater target radiation noise data, a multi-dimensional physical feature intelligent extraction and classification network model is trained in multiple stages. At each stage, the model converges by setting the weighted loss function and model training parameters.

[0012] The trained multidimensional physical feature intelligent extraction and classification network model is used to identify the radiated noise data of unknown underwater acoustic targets and output the classification results.

[0013] Preferably, the construction of the feature extraction network unit includes the following steps:

[0014] By using convolution operators of different scales, a variety of residual convolution blocks for multi-scale and multi-channel feature extraction are constructed, and the number of input channels of each residual convolution block is set to a configurable parameter c. Each residual convolution block is equipped with the GELU activation function.

[0015] By combining residual convolutional blocks, a line spectrum feature extraction network unit and a propeller harmonic feature extraction network unit are constructed.

[0016] Preferably, the line spectrum feature extraction network unit and the propeller harmonic feature extraction network unit each include a convolutional layer, a pooling layer, a linear convolutional network composed of residual convolutional blocks, a global average pooling layer, and a fully connected layer arranged sequentially; the linear convolutional network includes several concatenated convolutional chains, any convolutional chain is composed of the same residual convolutional blocks connected in series, and the number of input channels c of different residual convolutional blocks in the linear convolutional network decreases sequentially.

[0017] Preferably, in the linear convolutional network, any convolutional chain is connected in series with the next convolutional chain through a convolutional layer with a stride of s×s, where s≠1.

[0018] Preferably, the residual convolution module includes three parallel branches, two of which include several concatenated convolutional layers and a GELU activation function; the third branch is a direct-connected layer; the convolutional features output from the three branches are added together to obtain the final output of the module.

[0019] Preferably, the construction of the multidimensional physical feature fusion and classification unit includes: adding a feature splicing function layer to splice the extracted line spectrum features and propeller harmonic features; and sequentially adding a fully connected layer and a Softmax function classifier to classify and identify the spliced ​​features.

[0020] Preferably, the loss function for constructing the multidimensional physical feature intelligent extraction and classification network model includes:

[0021] The loss function J of the line spectrum feature extraction network unit is constructed based on the minimum mean square error method. Line (P Line y Line ), y Line P represents the line spectral feature labels corresponding to the training samples. Line The output of the line spectrum feature extraction network unit;

[0022] The loss function J of the propeller harmonic feature extraction network unit is constructed based on the minimum mean square error method. Har (P Har y Har ), y Har P represents the propeller harmonic feature labels corresponding to the training samples. Har The output of the propeller harmonic feature extraction network unit;

[0023] The loss function J for multi-dimensional physical feature fusion and classification units is constructed based on the cross-entropy method. Class (P Class y T ), P Class y represents the target recognition result output by the multidimensional physical feature fusion and classification unit. T Indicates the target type label of the training sample;

[0024] The loss function J = αJ for constructing a multidimensional physical feature intelligent extraction and classification network model Line +βJ Har +γJ Class , where α, β, and γ are weighting coefficients.

[0025] Preferably, the multi-stage training of the intelligent extraction and classification network model for multi-dimensional physical features includes:

[0026] Radiated noise data x for each tagged underwater target θ (t), θ=1,2,...,n, are preprocessed based on the LOFAR analysis method and the DEMON analysis method respectively to obtain the LOFAR spectrum training sample set and the DEMON spectrum sample set; the line spectrum frequencies and amplitudes of the LOFAR spectrum are manually labeled to obtain the corresponding line spectrum sequence labels; the harmonic frequencies of the DEMON spectrum are manually labeled to obtain the harmonic sequence labels.

[0027] Set α = β = 1 and γ = 0.01, set the optimizer to SGD and the learning rate to 0.001 during iterative training, and conduct the first stage training of the multidimensional physical feature intelligent extraction and classification network model;

[0028] With α = β = 0.1 and γ = 1, and the optimizer set to SGD and the learning rate set to 0.001 during iterative training, the second stage of training of the multidimensional physical feature intelligent extraction and classification network model is carried out.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] This invention comprehensively utilizes two typical physical features: line spectrum and propeller parameters. Simultaneously, using LOFAR and DEMON spectra as inputs, it constructs corresponding line spectrum and propeller parameter feature extraction network units, respectively. Based on this, it achieves target classification and recognition through the fusion of multi-dimensional physical features. Compared with existing underwater acoustic target classification and recognition methods, the method proposed in this invention is more accurate in extracting target physical features under conditions such as low signal-to-noise ratio, has a higher feature comprehensive utilization capability, and can improve recognition performance. Attached Figure Description

[0031] Figure 1 This is a flowchart of the present invention.

[0032] Figure 2 This is a schematic diagram of the structure of two types of residual convolutional blocks.

[0033] Figure 3 A multi-stage training framework diagram for a network model for intelligent extraction and classification of multi-dimensional physical features.

[0034] Figure 4 This is a line spectrum feature extraction result diagram of the intermediate output when using a multi-dimensional physical feature intelligent extraction and classification network model to process simulated underwater acoustic target data.

[0035] Figure 5 The image shows the propeller harmonic feature extraction result in the intermediate output when processing simulated underwater acoustic target data using a multidimensional physical feature intelligent extraction and classification network model.

[0036] Figure 6 This is a visualization projection of the fused numerical features of the intermediate output when processing two types of actual underwater acoustic target data using a multidimensional physical feature intelligent extraction and classification network model.

[0037] Figure 7 This is a statistical chart showing the accuracy of different identification methods used to identify a batch of sea trial target data. Detailed Implementation

[0038] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0039] Reference Figure 1 As shown, a multi-dimensional underwater acoustic target physical feature intelligent extraction and classification method is proposed. The overall idea of ​​this method is as follows: First, the underwater acoustic target radiated noise data is preprocessed, and standardized LOFAR and DEMON spectra are generated through LOFAR and DEMON spectrum analysis. Second, combined with deep learning methods, a line spectrum feature extraction network unit and a propeller harmonic feature extraction network unit are constructed to extract features from the LOFAR and DEMON spectra respectively, and output the corresponding physical features, including line spectrum features and propeller harmonic features. Finally, a multi-dimensional physical feature fusion and classification unit is constructed to comprehensively process the physical features of the above two dimensions and output the target classification and recognition results.

[0040] The intelligent extraction and classification method for the physical features of multidimensional underwater acoustic targets based on this overall idea consists of the following four steps.

[0041] Step 1: Construct a multi-dimensional physical feature intelligent extraction and classification network model based on the PyTorch framework, which includes two sub-steps.

[0042] Step 1.1: Construct two feature extraction network units that are set up in parallel and pre-positioned, respectively for extracting line spectrum features from the LOFAR spectrum and extracting propeller harmonic features from the DEMON spectrum.

[0043] Step 1.1.1: Construct various residual convolutional blocks for multi-scale, multi-channel feature extraction using convolution operators of different scales. Each residual convolutional block includes three parallel branches. Two of these branches consist of several concatenated convolutional layers and a GELU activation function, while the third branch is a directly connected layer. The convolutional features output from the three branches are connected through a feature concatenation layer. Set the number of input channels for each residual convolutional block to a configurable parameter c.

[0044] Here, different types of residual convolutional blocks have different network structures, and the same type of residual convolutional blocks are considered to be the same residual convolutional block only when the settable parameter values ​​are consistent.

[0045] Step 1.1.2: Construct the line spectrum feature extraction network unit and the propeller harmonic feature extraction network unit. Both include a convolutional layer, a pooling layer, a linear convolutional network composed of residual convolutional blocks, a global average pooling layer, and a fully connected layer arranged sequentially. The linear convolutional network includes several concatenated convolutional chains. Any convolutional chain is connected to the next convolutional chain through a convolutional layer with a stride of s×s, where s≠1. Any convolutional chain is composed of the same residual convolutional blocks connected in series. The number of input channels c of different residual convolutional blocks in the linear convolutional network decreases sequentially.

[0046] Step 1.2: Construct a post-constructed multi-dimensional physical feature fusion and classification unit for feature fusion of line spectrum features and propeller harmonic features, and target classification and recognition of the fused features. This includes: adding a feature splicing function layer to splice the extracted line spectrum features and propeller harmonic features; and sequentially adding a fully connected layer and a Soffmax function classifier to classify and recognize the spliced ​​features.

[0047] This is a specific embodiment of the present invention. Figure 2 Two network structures with residual convolutional blocks are shown. Figure 2 (a) The first residual convolutional block has an input channel number that can be set by parameter c1. This residual convolutional block includes three parallel branches. The first branch includes a convolutional layer (5×5, c1, [1,1]), a convolutional layer (3×3, 2c1, [1,1]), a convolutional layer (1×1, c1, [1,1]), and a GELU activation function, arranged sequentially. The second branch includes a convolutional layer (1×5, c1, [1,1]), a convolutional layer (5×1, c1, [1,1]), a convolutional layer (1×3, c1, [1,1]), a convolutional layer (3×1, c1, [1,1]), and a GELU activation function, arranged sequentially. The third branch is a direct connection layer. The convolutional features output from the three branches are added together to obtain the final output of the module. Figure 2 (b) The second residual convolutional block has an input channel number that can be set by parameter c2. This residual convolutional block also includes three parallel branches. The first branch includes a convolutional layer (3×3, c2, [1,1]), a convolutional layer (1×1, c2, [1,1]), and a GELU activation function arranged in sequence. The second branch includes a convolutional layer (1×3, c2, [1,1]), a convolutional layer (3×1, 2c2, [1,1]), a convolutional layer (1×1, c2, [1,1]), and a GELU activation function arranged in sequence. The third branch is a direct connection layer. The convolutional features output from the three branches are added together to obtain the final output of the module.

[0048] based on Figure 2 Two residual convolutional blocks are used to construct two numerical feature extraction network units. The specific process is as follows:

[0049] (a) Construct a line spectrum feature extraction network unit with LOFAR spectra as input, and add convolutional layers (7×7, 32, [1, 1]) and pooling layers (3×3, [2, 2]) sequentially; add 5 first residual convolutional blocks, with the number of input channels of each first residual convolutional block being parameter c1 = 128; add convolutional layers (3×3, 512, [2, 2]); add 5 first residual convolutional blocks, where the number of input channels of each first residual convolutional block is parameter c1 = 128. Number c1 = 512; Add convolutional layer (3×3, 1024, [2, 2]); Add 5 second residual convolutional blocks, where the number of input channels of each second residual convolutional block is parameter c2 = 1024; Add convolutional layer (3×3, 2048, [2, 2]); Add 5 second residual convolutional blocks, where the number of input channels of each second residual convolutional block is parameter c2 = 2048; Add global average pooling layer and fully connected layer (2048, f) in sequence. LOFAR_Num ), here f LOFAR_Num This represents the number of frequency points in the LOFAR spectrum, and the fully connected layer parameters (2048, f). LOFAR_Num ) indicates that the number of input and output nodes are 2048 and f, respectively. LOFAR_Num .

[0050] (b) Construct a propeller harmonic feature extraction network unit with DEMON spectrum as input, and add convolutional layers (7×7, 32, [1,1]) and pooling layers (3×3, [2,2]) in sequence; add 5 first residual convolutional blocks, where the number of input channels of each first residual convolutional block is parameter c1 = 256; add convolutional layers (3×3, 512, [2,2]); add 5 first residual convolutional blocks, where the number of input channels of each first residual convolutional block is parameter c1 = 512; add convolutional layers (3×3, 1024, [2,2]); add 8 second residual convolutional blocks, where the number of input channels of each second residual convolutional block is parameter c2 = 1024; add convolutional layers (3×3, 2048, [2,2]); add global average pooling layers; add fully connected layers with 2048 input nodes and the number of output nodes equal to the number of frequency points f in the DEMON spectrum. DEMON_Num , denoted as (2048, f DEMON_Num ).

[0051] Based on the two numerical feature extraction network units mentioned above, a multidimensional physical feature fusion and classification unit is constructed as follows: A feature splicing function layer is added to splice the physical features output by the line spectrum feature extraction network unit and the propeller harmonic feature extraction network unit. The total feature length after splicing is then f. LOFAR_Num +f DEMON_Num Add a fully connected layer with f input nodes. LOFAR_Num +f DEMON_Num The number of nodes output is equal to the number of target categories, N.Class Add a Softmax function classifier.

[0052] This yields a complete multidimensional physical feature intelligent extraction and classification network model, including parallel line spectrum feature extraction network units and propeller harmonic feature extraction network units. The multidimensional physical feature fusion and classification unit takes the features output by the line spectrum feature extraction network units and the propeller harmonic feature extraction network units as input and outputs the target category.

[0053] It is important to understand that the parameters of a convolutional layer are, in order, the kernel size, the number of convolutional channels, and the horizontal and vertical strides of the kernel. For example, the parameters of a convolutional layer (5×5, c1, [1, 1]) indicate that the kernel size is 5×5, the number of convolutional channels is c1, and the horizontal and vertical strides of the kernel are both 1. The parameters of a pooling layer are, in order, the pooling kernel size, and the horizontal and vertical strides of the pooling kernel. For example, the parameters of a pooling layer (3×3, [2, 2]) indicate that the pooling kernel size is 3×3 and the horizontal and vertical strides are both 2.

[0054] The first and second residual convolutional blocks in this invention, by configuring multiple network branches and residual structures with convolution operators of different scales, can achieve complex nonlinear relationship fitting and have multi-scale and multi-resolution feature mining capabilities; at the same time, the feature size extracted by the convolution chain is reduced by using convolutional layers with a stride of 2×2.

[0055] Figure 4 and Figure 5 The results of line spectrum feature extraction and propeller harmonic feature extraction are shown respectively when the multidimensional physical feature intelligent extraction and classification network model is used to process simulated underwater acoustic target data. It can be seen that the model can effectively extract line spectrum features and propeller harmonic features from the LOFAR and DEMON spectra.

[0056] Figure 6 In the diagram, the two types of target points are represented by two different colors, respectively. Figure 6 It can be seen that the above features are separable to a certain extent, indicating that the model can effectively extract and integrate multi-dimensional physical features, laying a feature foundation for target recognition.

[0057] Step 2: Construct the loss function J of the multi-dimensional physical feature intelligent extraction and classification network model, which is the line spectrum feature extraction loss J. Line Propeller harmonic feature extraction loss J Har And target classification and recognition loss J Class The weighted sum is calculated as follows.

[0058] Step 2.1: Construct the loss function J of the line spectrum feature extraction network unit based on the minimum mean square error method. Line As shown in equation (1),

[0059]

[0060] Where, N B y represents the number of training samples input in a single batch. Line P represents the line spectral feature labels corresponding to the training samples. LLin The output of the line spectrum feature extraction network unit is the line spectrum sequence.

[0061] Step 2.2: Construct the loss function J of the propeller harmonic feature extraction network unit based on the minimum mean square error method. Har As shown in equation (2),

[0062]

[0063] Where, N B y represents the number of training samples input in a single batch. Har P represents the propeller harmonic feature labels corresponding to the training samples. Har The output of the propeller harmonic feature extraction network unit is the harmonic sequence.

[0064] Step 2.3: Construct the loss function J for multi-dimensional physical feature fusion and classification units based on the cross-entropy method. Class As shown in equation (3),

[0065]

[0066] Among them, P Class y represents the target recognition result output by the multidimensional physical feature fusion and classification unit. T This represents the target type label of the training sample.

[0067] Step 2.4: Construct the loss function J = αJ for the multi-dimensional physical feature intelligent extraction and classification network model. Line +β JHar +γJ Class , where α, β, and γ are weighting coefficients that can be adjusted at different training stages of the model.

[0068] Step 3: Using labeled underwater target radiation noise data, the multi-dimensional physical feature intelligent extraction and classification network model is trained in multiple stages. At each stage, the model converges by setting the weighted loss function and model training parameters. (Refer to...) Figure 3 As shown, the specific process for this step is as follows:

[0069] Step 3.1, analyze the radiated noise data x of each tagged underwater target. θ(t), θ=1,2,...,n, are preprocessed based on the LOFAR analysis method and the DEMON analysis method respectively to obtain the LOFAR spectrum training sample set and the DEMON spectrum sample set; the line spectrum frequencies and amplitudes of the LOFAR spectrum are manually labeled to obtain the corresponding line spectrum sequence labels; the harmonic frequencies of the DEMON spectrum are manually labeled to obtain the harmonic sequence labels.

[0070] In step 3.1 of this invention, the underwater target radiated noise signal library used for training includes multiple labeled underwater target radiated noise data, denoted as x = {x1(t), x2(t), ..., x...} n (t), (n∈N) * For any one of the labeled underwater target radiated noise data x θ For each (t), the corresponding LOFAR and DEMON spectra are obtained through LOFAR and DEMON analysis methods, respectively. Therefore, the multiple LOFAR and DEMON spectra obtained from multiple labeled underwater target radiated noise data serve as training samples, forming the LOFAR spectra training sample set x. LOFAR and DEMON spectral sample set x DEMON The number of samples in the two sample sets is equal and they correspond one-to-one. At the same time, combined with prior knowledge of target characteristics, the human-computer interaction method is used to annotate the line spectrum frequency and amplitude of the LOFAR spectrum to obtain line spectrum sequence labels, and to annotate the harmonic frequencies of the DEMON spectrum to obtain harmonic sequence labels.

[0071] Step 3.2 involves the first stage of training the multidimensional physical feature intelligent extraction and classification network model. Specifically, α = β = 1 and γ = 0.01 are set, along with training parameters such as the optimizer, learning rate, and data block size during iterative training. The optimizer is set to SGD, and the learning rate is set to 0.001. The training sample set x is then used... LOFAR and x DEMON The same number of LOFAR and DEMON spectral samples with one-to-one correspondence are randomly selected with replacement (i.e., the corresponding LOFAR and DEMON spectral samples are generated from the same labeled underwater target radiation noise data); forward inference calculation is performed based on the multidimensional physical feature intelligent extraction and classification network model, and the loss value is calculated by combining the model's weighted loss function. The structural parameters of the multidimensional physical feature intelligent extraction and classification network model are optimized using the set optimizer; the model is trained through multiple rounds until the loss function converges.

[0072] Step 3.3: Perform the second stage training of the multidimensional physical feature intelligent extraction and classification network model. Specifically, set α = β = 0.1 and γ = 1. The specific optimizer, learning rate, data block size and other model training parameters and model training methods are the same as in step 3.2. The model is trained through multiple rounds until the loss function converges.

[0073] In this invention, the structural parameters of the line spectrum feature extraction network unit and the propeller harmonic feature extraction network unit are primarily optimized using the weights of the model loss function set in step 3.2, while also considering the structural parameter optimization of the multi-dimensional physical feature fusion and classification units. After primarily optimizing the structural parameters of the line spectrum feature extraction network unit and the propeller harmonic feature extraction network unit using the weights of the model loss function set in step 3.3, the structural parameters of the multi-dimensional physical feature fusion and classification units are then primarily optimized, while the structural parameters of the line spectrum feature extraction network unit and the propeller harmonic feature extraction network unit are fine-tuned to improve model convergence. Based on the construction of a multi-module weighted loss function, this invention divides model training into two stages, setting appropriate weighting coefficients for the loss function and training parameters such as the optimizer in each stage, achieving successive optimization of the model from local to global levels.

[0074] Step 4: Use the trained multidimensional physical feature intelligent extraction and classification network model to identify the radiated noise data of unknown underwater acoustic targets and output the classification and identification results.

[0075] In step 4 of this invention, the radiated noise data of the unknown underwater acoustic target needs to be preprocessed using the LOFAR analysis method and the DEMON analysis method to obtain the corresponding LOFAR map and DEMON map. Then, these two maps are used as inputs to the trained multidimensional physical feature intelligent extraction and classification network model to perform underwater acoustic target classification and recognition to obtain the classification and recognition results.

[0076] Figure 7 The present invention demonstrates the statistical results of the accuracy of identifying a batch of sea trial target data using the method of the present invention and a single line spectrum feature and propeller parameter feature identification method. It can be seen that the accuracy of the present invention is higher than that of the single feature identification model. This shows that by combining the use of multi-dimensional physical features to achieve comprehensive identification, the separability information of the target can be mined from different angles, thereby improving the identification effect.

Claims

1. A method for intelligent extraction and classification of multidimensional underwater acoustic target physical features, characterized in that, Includes the following steps: A multi-dimensional physical feature intelligent extraction and classification network model is constructed, including: constructing two feature extraction network units set up in parallel at the front end, which are used for line spectrum feature extraction and propeller harmonic feature extraction respectively; and constructing a multi-dimensional physical feature fusion and classification recognition unit at the back end, which is used for feature fusion and target classification recognition of fused features. The loss function of the intelligent extraction and classification network model of multidimensional physical features is constructed as a weighted sum of the line spectrum feature extraction loss, the propeller harmonic feature extraction loss, and the multidimensional physical feature fusion and classification recognition loss. Using labeled underwater target radiation noise data, a multi-dimensional physical feature intelligent extraction and classification network model is trained in multiple stages. At each stage, the model converges by setting the weighted loss function and model training parameters. The trained multidimensional physical feature intelligent extraction and classification network model is used to identify the radiated noise data of unknown underwater acoustic targets and output the classification and identification results. The two feature extraction network units, which are constructed in parallel and pre-configured, include: By utilizing convolution operators of different scales, various residual convolutional blocks for multi-scale, multi-channel feature extraction are constructed. The number of input channels for each residual convolutional block is set to a configurable parameter c, and each residual convolutional block is equipped with a GELU activation function. The residual convolutional module includes three parallel branches, two of which include several concatenated convolutional layers and GELU activation functions; the third branch is a direct-connected layer. The convolutional features output from the three branches are summed to obtain the final output of the module. By combining residual convolutional blocks, a line spectrum feature extraction network unit and a propeller harmonic feature extraction network unit are constructed. Each of these units includes, in sequence, a convolutional layer, a pooling layer, a linear convolutional network composed of residual convolutional blocks, a global average pooling layer, and a fully connected layer. The linear convolutional network comprises several concatenated convolutional chains, each consisting of the same residual convolutional block. The number of input channels *c* in different residual convolutional blocks within the linear convolutional network increases sequentially. In the linear convolutional network, each convolutional chain is connected to the next convolutional chain via a convolutional layer with a stride of *s*×*s, where *s*≠1. The constructed post-construction multi-dimensional physical feature fusion and classification recognition unit includes: A feature stitching layer is added to stitch together the extracted line spectrum features and propeller harmonic features. A fully connected layer and a Softmax function classifier are added sequentially to classify and recognize the spliced ​​features.

2. The method for intelligent extraction and classification of multidimensional underwater acoustic target physical features as described in claim 1, characterized in that, The two feature extraction network units, which are constructed in parallel and pre-configured, include: Multiple residual convolutional blocks and second residual convolutional blocks are constructed for multi-scale, multi-channel feature extraction. The first residual convolutional block has a settable number of input channels, c1. This residual convolutional block includes three parallel branches: the first branch includes a convolutional layer (5×5, c1, [1,1]), a convolutional layer (3×3, 2c1, [1,1]), a convolutional layer (1×1, c1, [1,1]), and a GELU activation function, arranged sequentially; the second branch includes a convolutional layer (1×5, c1, [1,1]), a convolutional layer (5×1, c1, [1,1]), a convolutional layer (1×3, c1, [1,1]), a convolutional layer (3×1, c1, [1,1]), and a GELU activation function, arranged sequentially; the third branch is a direct connection layer; the convolutional features output from the three branches are added together to obtain the final output of the module. The second residual convolutional block has an input channel number that can be set by parameter c2. This residual convolutional block also includes three parallel branches, wherein: the first branch includes a convolutional layer (3×3, c2, [1,1]), a convolutional layer (1×1, c2, [1,1]), and a GELU activation function arranged in sequence; the second branch includes a convolutional layer (1×3, c2, [1,1]), a convolutional layer (3×1, 2c2, [1,1]), a convolutional layer (1×1, c2, [1,1]), and a GELU activation function arranged in sequence; the third branch is a direct connection layer; the convolutional features output from the three branches are added together to obtain the final output result of the module; Two feature extraction network units are constructed based on the first residual convolutional block and the second residual convolutional block. The specific process is as follows: (a) Construct a line spectrum feature extraction network unit with LOFAR spectra as input, and add convolutional layers (7×7, 32, [1, 1]) and pooling layers (3×3, [2, 2]) in sequence; add 5 first residual convolutional blocks, with the number of input channels of each first residual convolutional block being parameter c1 = 128; add convolutional layers (3×3, 512, [2, 2]); add 5 first residual convolutional blocks, where the number of input channels of each first residual convolutional block is parameter c1 = 128. Number c1 = 512; Add convolutional layer (3×3, 1024, [2, 2]); Add 5 second residual convolutional blocks, where the number of input channels of each second residual convolutional block is parameter c2 = 1024; Add convolutional layer (3×3, 2048, [2, 2]); Add 5 second residual convolutional blocks, where the number of input channels of each second residual convolutional block is parameter c2 = 2048; Add global average pooling layer and fully connected layer (2048, f) in sequence. LOFAR_Num Here, fLOFAR_Num represents the number of frequency points in the LOFAR spectrum, and the fully connected layer parameters (2048, f LOFAR_Num) This indicates that the number of input and output nodes are 2048 and f, respectively. LOFAR_Num ; (b) Construct a propeller harmonic feature extraction network unit with DEMON spectrum as input, and add convolutional layers (7×7, 32, [1, 1]) and pooling layers (3×3, [2, 2]) in sequence; add 5 first residual convolutional blocks, where the number of input channels of each first residual convolutional block is parameter c1 = 256; add convolutional layers (3×3, 512, [2, 2]); add 5 first residual convolutional blocks, where each first residual convolutional block The number of input channels is parameter c1 = 512; a convolutional layer (3×3, 1024, [2, 2]) is added; 8 second residual convolutional blocks are added, where the number of input channels for each second residual convolutional block is parameter c2 = 1024; a convolutional layer (3×3, 2048, [2, 2]) is added; a global average pooling layer is added; a fully connected layer is added, with 2048 input nodes and the number of output nodes being the number of frequency points f in the DEMON spectrum. DEMON_Num , denoted as (2048, f DEMON_Num ); The constructed post-construction multi-dimensional physical feature fusion and classification recognition unit includes: By adding a feature splicing layer to splice the physical features output by the line spectrum feature extraction network unit and the propeller harmonic feature extraction network unit, the total feature length after splicing is f. LOFAR_Num +f DEMON_Num ; Add a fully connected layer with f input nodes. LOFAR_Num +f DEMON_Num The number of nodes output is the number of target categories, N. Class Add a Softmax function classifier.

3. The intelligent extraction and classification method for multidimensional underwater acoustic target physical features as described in claim 1, characterized in that, The loss function for constructing the intelligent extraction and classification network model of multidimensional physical features includes: The loss function J of the line spectrum feature extraction network unit is constructed based on the minimum mean square error method. Line The formula is as follows: , Where, N B y represents the number of training samples input in a single batch. Line P represents the line spectral feature labels corresponding to the training samples. Line The output of the line spectrum feature extraction network unit; The loss function J of the propeller harmonic feature extraction network unit is constructed based on the minimum mean square error method. Har The formula is as follows: , Where, N B y represents the number of training samples input in a single batch. Har P represents the propeller harmonic feature labels corresponding to the training samples. Har The output of the propeller harmonic feature extraction network unit; The loss function J for multi-dimensional physical feature fusion and classification units is constructed based on the cross-entropy method. Class The formula is as follows: , Among them, P Class y represents the target recognition result output by the multidimensional physical feature fusion and classification unit. T Indicates the target type label of the training sample; The loss function J of the multidimensional physical feature intelligent extraction and classification network model is constructed as follows: J = αJ Line +βJ Har +γJ Class , where α, β, and γ are weighting coefficients.

4. The intelligent extraction and classification method for multidimensional underwater acoustic target physical features as described in claim 1, characterized in that, The multi-stage training of the intelligent extraction and classification network model for multi-dimensional physical features includes: Radiated noise data x for each tagged underwater target θ (t), θ=1,2,…,n, are preprocessed based on the LOFAR analysis method and the DEMON analysis method respectively to obtain the LOFAR spectrum training sample set and the DEMON spectrum sample set; the line spectrum frequencies and amplitudes of the LOFAR spectrum are manually labeled to obtain the corresponding line spectrum sequence labels; the harmonic frequencies of the DEMON spectrum are manually labeled to obtain the harmonic sequence labels. Set α = β = 1 and γ = 0.01, set the optimizer to SGD and the learning rate to 0.001 during iterative training, and conduct the first stage training of the multidimensional physical feature intelligent extraction and classification network model; With α = β = 0.1 and γ = 1, and the optimizer set to SGD and the learning rate set to 0.001 during iterative training, the second stage of training of the multidimensional physical feature intelligent extraction and classification network model is carried out.

Citation Information

Patent Citations

  • Ship radiation noise characteristic recombination method based on statistical distribution

    CN111553207A

  • Small sample underwater target identification method based on multi-platform auditory perception feature deep transfer learning

    CN114202056A

  • Passive target line spectrum feature extraction method based on sliding window deep convolutional neural network and reinforcement learning

    CN115587302A

  • Visible light-infrared sequence target tracking method based on residual learning

    CN115690164A