Underwater target radiation noise identification method based on deep learning

Abstract

The invention discloses an underwater target radiation noise identification method based on deep learning. The method comprises the following steps: 1, acquiring a logarithmic Mel frequency band energy matrix of a target radiation noise signal; 2, obtaining a processed logarithm Mel band energy matrix through distortion mask of the logarithm Mel band energy matrix; 3, constructing an underwater target radiation noise identification model; 4, feature extraction of the processed logarithmic Mel frequency band energy matrix; 5, training an underwater target radiation noise identification model based on the target category probability vector; and 6, recognizing subsequent target radiation noise signals by using the trained underwater target radiation noise recognition model. The method is reasonable in design, the logarithmic Mel frequency band energy matrix for obtaining the target radiation noise signal is processed through the random mask matrix, the processing capacity of the model for the distortion condition is improved, the robustness of the model is improved, richer features of the full text are captured through multi-scale feature extraction in the model, and the robustness of the model is improved. And a sliding window is used to pay attention to extraction of local detail features, so that the recognition accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of underwater target radiation noise identification, and specifically relates to an underwater target radiation noise identification method based on deep learning. Background Art

[0002] Existing methods for identifying underwater target radiated noise primarily rely on machine learning and deep learning. Machine learning methods use feature-extracted underwater target radiated noise as input and train the noise radiated by different targets. The trained model can predict unknown noise signals and identify the corresponding underwater target type. However, these methods are highly data-dependent; if the dataset contains noise, model performance degrades, and model generalization is limited. Deep learning methods typically use automatic feature learning to achieve end-to-end recognition, improving model performance. However, these models suffer from low robustness and poor handling of distortion.

[0003] Therefore, there is currently a lack of a well-designed method for underwater target radiation noise recognition based on deep learning. The logarithmic Mel-band energy matrix of the target radiation noise signal is obtained and processed through a random mask matrix, which increases the model's ability to handle distortion and improves the robustness of the model. Multi-scale feature extraction is used in the underwater target radiation noise recognition model to capture richer features of the full text, and sliding window attention is used to focus on the extraction of local detail features, thereby improving the recognition accuracy of the model. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to address the deficiencies in the above-mentioned prior art and provide a method for identifying underwater target radiation noise based on deep learning. The method has simple steps and a reasonable design. The logarithmic Mel-band energy matrix of the target radiation noise signal is obtained and processed by a random mask matrix, thereby increasing the model's ability to handle distortion and improving the robustness of the model. Multi-scale feature extraction is used in the underwater target radiation noise recognition model to capture richer features of the full text, and a sliding window is used to focus on the extraction of local detail features, thereby improving the recognition accuracy of the model.

[0005] To solve the above technical problems, the present invention adopts a technical solution: a method for identifying underwater target radiated noise based on deep learning, the method comprising the following steps:

[0006] Step 1: Obtain the logarithmic Mel-band energy matrix of the target radiation noise signal; where the mth logarithmic Mel-band energy matrix is denoted as H m ;

[0007] Step 2: The distortion mask of the logarithmic Mel-band energy matrix is obtained to obtain the processed logarithmic Mel-band energy matrix:

[0008] Step 201: Generate a random mask matrix A using a computer; wherein the element value in the random mask matrix A is 0 or 1;

[0009] Step 202: The mth logarithmic Mel-band energy matrix H m Multiply element-by-element with the random mask matrix A to obtain the mth processed logarithmic Mel-band energy matrix H′ m ;

[0010] Step 3: Construction of underwater target radiation noise recognition model:

[0011] The underwater target radiation noise recognition model includes a first convolution + BN layer, a first maximum pooling layer, a multi-scale feature module, a second maximum pooling layer, a sliding window attention mechanism, a flattening layer, a first fully connected layer, a second fully connected layer and a Softmax layer;

[0012] Step 4: Feature extraction of the processed logarithmic Mel-band energy matrix:

[0013] The mth processed logarithmic Mel-band energy matrix is passed through the underwater target radiation noise recognition model for feature extraction, and the target category probability vector is output;

[0014] Step 5: Training the underwater target radiation noise recognition model based on the target category probability vector;

[0015] Step 6: Use the trained underwater target radiation noise recognition model to identify subsequent target radiation noise signals.

[0016] The above-mentioned method for identifying underwater target radiation noise based on deep learning is characterized by: Step 1, the specific process is as follows:

[0017] Step 101: Select a target radiation noise signal from the underwater ship radiation noise dataset ShipsEar, and perform frame processing on each target radiation noise signal according to a set frame length to obtain a framed signal; then sort the framed signals to obtain the mth framed signal; where m is a positive integer and 1≤m≤M, M is the total number of samples, and the number of time domain sampling points of the mth framed signal is 216;

[0018] Step 102: Using a computer, zero padding is performed on the mth framed signal to obtain the mth transformed signal; wherein the number of FFT points of the mth transformed signal is 2048;

[0019] Step 103: Process the mth transformed signal using the Mel-frequency cepstral coefficient method using a computer to obtain the energy S of the mth transformed signal output after passing through the i-th triangular filter. m(i); where i is a positive integer and 1≤i≤I; I is the total number of triangular Mel filters and I is 128;

[0020] Step 104: Using a computer to calculate the m (i) = lnS m (i) The logarithmic energy E of the m-th transformed signal output by the i-th triangular filter is obtained. m (i);

[0021] Step 105: According to the method of steps 103 and 104, the logarithmic energy vector of the m-th transformed signal output through I triangular filters is obtained and recorded as the m-th logarithmic energy vector;

[0022] Step 106: Combine the 216 mth logarithmic energy vectors to obtain the logarithmic Mel-band energy matrix corresponding to the mth framed signal, and record it as the mth logarithmic Mel-band energy matrix H m , H m The size of is E×I, and E=216.

[0023] The above-mentioned method for identifying underwater target radiation noise based on deep learning is characterized in that: in step 2, the size of the random mask matrix A is E×I, and the element value A in the e-th row and i-th column of the random mask matrix A is ei is 0 or 1, and the number of 0 elements in the random mask matrix A is [35% MI] to [40% MI]; where [·] represents a rounding operation;

[0024] The mth processed log-Mel band energy matrix H′ m The value of the element in row e and column i is E′ ei , and E′ ei =E ei ×A ei ;E ei Represents the mth log-Mel band energy matrix H m The value of the element in the e-th row and i-th column in ; e is a positive integer, and 1≤e≤E.

[0025] The above-mentioned method for identifying underwater target radiation noise based on deep learning is characterized by: Step 3, the specific process is as follows:

[0026] Step 301: Use a computer to convert the mth processed logarithmic Mel-band energy matrix H′ m Recorded as a two-dimensional tensor, and the two-dimensional tensor is input into the first convolution + BN layer for feature extraction to obtain the first feature map;

[0027] Step 302: Input the first feature map into the first maximum pooling layer for feature extraction to obtain a second feature map;

[0028] Step 303: Input the second feature map into the multi-scale feature module and pass it through the convolution layer 1, convolution layer 2, and convolution layer 3 respectively to obtain the first intermediate feature map, the second intermediate feature map, and the third intermediate feature map;

[0029] Step 304: Concatenate the first intermediate feature map, the second intermediate feature map, and the third intermediate feature map through concat to obtain a first concatenated feature map;

[0030] Step 305: Input the first concatenated feature map into the second maximum pooling layer for feature extraction to obtain a third feature map;

[0031] Step 306: Process the third feature map through a sliding window attention mechanism to obtain a fourth feature map;

[0032] Step 307: After the fourth feature map is flattened, it is converted into a one-dimensional vector;

[0033] Step 308: Pass the one-dimensional vector through the first fully connected layer to output an intermediate one-dimensional vector;

[0034] Step 309: Pass the intermediate one-dimensional vector through the second fully connected layer to output the converted one-dimensional vector;

[0035] Step 3010: Pass the converted one-dimensional vector through the Softmax layer to output the target category probability vector.

[0036] The above-mentioned method for identifying underwater target radiation noise based on deep learning is characterized in that: the first convolution + BN layer includes a Conv1 convolution layer and a batch normalization layer, the number of convolution kernels in the Conv1 convolution layer is 64, the size of the convolution kernel is 3×3, the sliding step is 1, and the padding is 1;

[0037] The size of the pooling window in the first maximum pooling layer is 2×2, the window sliding step is 2, and the padding is 0;

[0038] The multi-scale feature module includes convolution layer 1, convolution layer 2, and convolution layer 3. The number of convolution kernels in convolution layer 1 is 64, the size of the convolution kernel is 3×3, the sliding step is 1, and the padding is 1.

[0039] The number of convolution kernels in the second convolution layer is 64, the size of the convolution kernel is 5×5, the sliding step is 1, and the padding is 2;

[0040] The number of convolution kernels in convolution layer 3 is 64, the size of the convolution kernel is 7×7, the sliding step is 1, and the padding is 3;

[0041] The size of the pooling window in the second maximum pooling layer is 2×2, the window sliding step is 2, and the padding is 0;

[0042] The number of the first fully connected layers is 1024, and the number of the second fully connected layers is 5.

[0043] The above-mentioned method for identifying underwater target radiation noise based on deep learning is characterized by: Step 5, the specific process is as follows:

[0044] Step 501: Obtain the j-th category probability value corresponding to the m-th framed signal from the target category probability vector. Where j is a positive integer, 1≤j≤C, and C represents the total number of underwater target categories;

[0045] Step 502: Using a computer Get the cross entropy loss function L; where, Indicates the true label corresponding to the signal after the mth frame, the value is 0 or 1; m ranges from 1 to P, where P is the batch size;

[0046] Step 503: The computer uses the Adma optimizer to iteratively train the underwater target radiation noise recognition model using the cross entropy loss function L until the entire training set is trained, completing one iterative training.

[0047] Step 504: Repeat steps 502 and 503, and perform iterative training until a preset number of iterative training times is met, thereby obtaining a trained underwater target radiation noise recognition model.

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] 1. The method of the present invention has simple steps and reasonable design, and solves the problem of identifying the radiation noise of underwater targets.

[0050] 2. The present invention obtains the logarithmic Mel-band energy matrix of the target radiation noise signal, and distorts it through a random mask matrix to obtain the processed logarithmic Mel-band energy matrix. The random mask matrix is generated for masking to achieve the effect of data distortion, so that subsequent processing can be carried out under distortion, thereby improving the subsequent model's ability to handle distortion and improving the robustness of the model.

[0051] 3. A multi-scale feature module and a sliding window attention mechanism are set in the underwater target radiation noise recognition model of the present invention. Using the multi-scale feature module, the model can simultaneously extract features of multiple scales through convolution kernels of different sizes. Smaller convolution kernels can capture detailed features, while larger convolution kernels can capture broader contextual information, which increases the flexibility of the model and can cope with targets of different sizes and shapes; the sliding window attention mechanism is used to extract in each local area, paying more attention to local features, which can greatly reduce the amount of calculation and improve the efficiency of the network model; secondly, the multi-scale feature extraction is used to capture richer features of the full text, and the sliding window attention mechanism is used to focus on the extraction of local detail features, which cooperate with each other to improve the recognition accuracy of the model.

[0052] In summary, the method of the present invention has simple steps and reasonable design. The logarithmic Mel-band energy matrix of the target radiation noise signal is obtained and processed by a random mask matrix, which increases the model's ability to handle distortion and improves the robustness of the model. Multi-scale feature extraction is used in the underwater target radiation noise recognition model to capture richer features of the full text, and the sliding window attention is used to focus on the extraction of local detail features, thereby improving the recognition accuracy of the model.

[0053] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flowchart of the method of the present invention.

[0055] Figure 2 This is a model block diagram of the present invention. DETAILED DESCRIPTION

[0056] like Figure 1 and Figure 2 As shown, the method for identifying underwater target radiation noise based on deep learning of the present invention includes the following steps:

[0057] Step 1: Obtain the logarithmic Mel-band energy matrix of the target radiation noise signal; where the mth logarithmic Mel-band energy matrix is denoted as H m ;

[0058] Step 2: The distortion mask of the logarithmic Mel-band energy matrix is obtained to obtain the processed logarithmic Mel-band energy matrix:

[0059] Step 201: Generate a random mask matrix A using a computer; wherein the element value in the random mask matrix A is 0 or 1;

[0060] Step 202: The mth logarithmic Mel-band energy matrix H mMultiply element-by-element with the random mask matrix A to obtain the mth processed logarithmic Mel-band energy matrix H′ m ;

[0061] Step 3: Construction of underwater target radiation noise recognition model:

[0062] The underwater target radiation noise recognition model includes a first convolution + BN layer, a first maximum pooling layer, a multi-scale feature module, a second maximum pooling layer, a sliding window attention mechanism, a flattening layer, a first fully connected layer, a second fully connected layer and a Softmax layer;

[0063] Step 4: Feature extraction of the processed logarithmic Mel-band energy matrix:

[0064] The mth processed logarithmic Mel-band energy matrix is passed through the underwater target radiation noise recognition model for feature extraction, and the target category probability vector is output;

[0065] Step 5: Training the underwater target radiation noise recognition model based on the target category probability vector;

[0066] Step 6: Use the trained underwater target radiation noise recognition model to identify subsequent target radiation noise signals.

[0067] In this embodiment, the specific process of step one is as follows:

[0068] Step 101: Select a target radiation noise signal from the underwater ship radiation noise dataset ShipsEar, and perform frame processing on each target radiation noise signal according to a set frame length to obtain a framed signal; then sort the framed signals to obtain the mth framed signal; where m is a positive integer and 1≤m≤M, M is the total number of samples, and the number of time domain sampling points of the mth framed signal is 216;

[0069] Step 102: Using a computer, zero padding is performed on the mth framed signal to obtain the mth transformed signal; wherein the number of FFT points of the mth transformed signal is 2048;

[0070] Step 103: Process the mth transformed signal using the Mel-frequency cepstral coefficient method using a computer to obtain the energy S of the mth transformed signal output after passing through the i-th triangular filter. m (i); where i is a positive integer and 1≤i≤I; I is the total number of triangular Mel filters and I is 128;

[0071] Step 104: Using a computer to calculate the m (i) = lnS m (i) The logarithmic energy E of the m-th transformed signal output by the i-th triangular filter is obtained. m(i);

[0072] Step 105: According to the method of steps 103 and 104, the logarithmic energy vector of the m-th transformed signal output through I triangular filters is obtained and recorded as the m-th logarithmic energy vector;

[0073] Step 106: Combine the 216 mth logarithmic energy vectors to obtain the logarithmic Mel-band energy matrix corresponding to the mth framed signal, and record it as the mth logarithmic Mel-band energy matrix H m , H m The size of is E×I, and E=216.

[0074] In this embodiment, the size of the random mask matrix A in step 2 is E×I, and the element value A in the e-th row and i-th column of the random mask matrix A is ei is 0 or 1, and the number of 0 elements in the random mask matrix A is [35% MI] to [40% MI]; where [·] represents a rounding operation;

[0075] The mth processed log-Mel band energy matrix H′ m The value of the element in row e and column i is E′ ei , and E′ ei =E ei ×A ei ;E ei Represents the mth log-Mel band energy matrix H m The value of the element in the e-th row and i-th column in ; e is a positive integer, and 1≤e≤E.

[0076] In this embodiment, the specific process of step three is as follows:

[0077] Step 301: Use a computer to convert the mth processed logarithmic Mel-band energy matrix H′ m Recorded as a two-dimensional tensor, and the two-dimensional tensor is input into the first convolution + BN layer for feature extraction to obtain the first feature map;

[0078] Step 302: Input the first feature map into the first maximum pooling layer for feature extraction to obtain a second feature map;

[0079] Step 303: Input the second feature map into the multi-scale feature module and pass it through the convolution layer 1, convolution layer 2, and convolution layer 3 respectively to obtain the first intermediate feature map, the second intermediate feature map, and the third intermediate feature map;

[0080] Step 304: Concatenate the first intermediate feature map, the second intermediate feature map, and the third intermediate feature map through concat to obtain a first concatenated feature map;

[0081] Step 305: Input the first concatenated feature map into the second maximum pooling layer for feature extraction to obtain a third feature map;

[0082] Step 306: Process the third feature map through a sliding window attention mechanism to obtain a fourth feature map;

[0083] Step 307: After the fourth feature map is flattened, it is converted into a one-dimensional vector;

[0084] Step 308: Pass the one-dimensional vector through the first fully connected layer to output an intermediate one-dimensional vector;

[0085] Step 309: Pass the intermediate one-dimensional vector through the second fully connected layer to output the converted one-dimensional vector;

[0086] Step 3010: Pass the converted one-dimensional vector through the Softmax layer to output the target category probability vector.

[0087] In this embodiment, the first convolution+BN layer includes a Conv1 convolution layer and a batch normalization layer. The number of convolution kernels in the Conv1 convolution layer is 64, the size of the convolution kernel is 3×3, the sliding step is 1, and the padding is 1.

[0088] The size of the pooling window in the first maximum pooling layer is 2×2, the window sliding step is 2, and the padding is 0;

[0089] The multi-scale feature module includes convolution layer 1, convolution layer 2, and convolution layer 3. The number of convolution kernels in convolution layer 1 is 64, the size of the convolution kernel is 3×3, the sliding step is 1, and the padding is 1.

[0090] The number of convolution kernels in the second convolution layer is 64, the size of the convolution kernel is 5×5, the sliding step is 1, and the padding is 2;

[0091] The number of convolution kernels in convolution layer 3 is 64, the size of the convolution kernel is 7×7, the sliding step is 1, and the padding is 3;

[0092] The size of the pooling window in the second maximum pooling layer is 2×2, the window sliding step is 2, and the padding is 0;

[0093] The number of the first fully connected layers is 1024, and the number of the second fully connected layers is 5.

[0094] In this embodiment, the specific process of step five is as follows:

[0095] Step 501: Obtain the j-th category probability value corresponding to the m-th framed signal from the target category probability vector. Where j is a positive integer, 1≤j≤C, and C represents the total number of underwater target categories;

[0096] Step 502: Using a computer Get the cross entropy loss function L; where, Indicates the true label corresponding to the signal after the mth frame, the value is 0 or 1; m ranges from 1 to P, where P is the batch size;

[0097] Step 503: The computer uses the Adma optimizer to iteratively train the underwater target radiation noise recognition model using the cross entropy loss function L until the entire training set is trained, completing one iterative training.

[0098] Step 504: Repeat steps 502 and 503, and perform iterative training until a preset number of iterative training times is met, thereby obtaining a trained underwater target radiation noise recognition model.

[0099] In this embodiment, in actual use, the window size set in the sliding window attention mechanism is 2, and the sliding step size is 2.

[0100] In this embodiment, in actual use, the frame length is set to 5 seconds, and the length of the signal after the mth frame is 5 seconds.

[0101] In this embodiment, in actual use, the total number of samples M is 2223.

[0102] In this embodiment, in actual use, the M framed signals correspond to five categories, so the value of C is 5. The first category includes fishing vessels, trawlers, mussel harvesters, tugboats, and dredgers; the second category includes motorboats, pilot boats, and sailboats; the third category includes passenger ferries; the fourth category includes ocean liners and roll-on / roll-off ships; and the fifth category includes marine ambient noise.

[0103] In this embodiment, the total number of samples M is divided into multiple batches for training. The number of batches P in 502 is 50. If the remaining samples are less than 50, the remaining samples are taken as a batch for training.

[0104] In this embodiment, the preset number of iterative training times in 504 is 90.

[0105] In this embodiment, in actual use, the size of the feature map is expressed as channel × length × width. The size of the first feature map is 64×216×128, the size of the second feature map is 64×108×64, the size of the first intermediate feature map is 64×108×64, the size of the second intermediate feature map is 64×108×64, the size of the third intermediate feature map is 64×108×64, the size of the first spliced feature map is 192×108×64, the size of the third feature map is 192×54×32, the size of the fourth feature map is 192×54×32, the size of the one-dimensional vector is 331776×1, the size of the intermediate one-dimensional vector is 1024×1, and the size of the converted one-dimensional vector is 5×1.

[0106] In summary, the method of the present invention has simple steps and reasonable design. The logarithmic Mel-band energy matrix of the target radiation noise signal is obtained and processed by a random mask matrix, which increases the model's ability to handle distortion and improves the robustness of the model. Multi-scale feature extraction is used in the underwater target radiation noise recognition model to capture richer features of the full text, and the sliding window attention is used to focus on the extraction of local detail features, thereby improving the recognition accuracy of the model.

[0107] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent structural change made to the above embodiment based on the technical essence of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for identifying underwater target radiation noise based on deep learning, characterized in that: The method comprises the following steps: Step 1: Obtain the logarithmic Mel-band energy matrix of the target radiation noise signal; where the mth logarithmic Mel-band energy matrix is denoted as H m ; Step 2: The distortion mask of the logarithmic Mel-band energy matrix is obtained to obtain the processed logarithmic Mel-band energy matrix: Step 201: Generate a random mask matrix A using a computer; wherein the element value in the random mask matrix A is 0 or 1; Step 202: The mth logarithmic Mel-band energy matrix H m Multiply element-by-element with the random mask matrix A to obtain the mth processed logarithmic Mel-band energy matrix H′ m ; Step 3: Construction of underwater target radiation noise recognition model: The underwater target radiation noise recognition model includes a first convolution + BN layer, a first maximum pooling layer, a multi-scale feature module, a second maximum pooling layer, a sliding window attention mechanism, a flattening layer, a first fully connected layer, a second fully connected layer and a Softmax layer; Step 4: Feature extraction of the processed logarithmic Mel-band energy matrix: The mth processed logarithmic Mel-band energy matrix is passed through the underwater target radiation noise recognition model for feature extraction, and the target category probability vector is output; Step 5: Training the underwater target radiation noise recognition model based on the target category probability vector; Step 6: Use the trained underwater target radiation noise recognition model to identify subsequent target radiation noise signals.

2. The method for identifying underwater target radiated noise based on deep learning according to claim 1, characterized in that: Step 1: The specific process is as follows: Step 101: Select a target radiation noise signal from the underwater ship radiation noise dataset ShipsEar, and perform frame processing on each target radiation noise signal according to a set frame length to obtain a framed signal; then sort the framed signals to obtain the mth framed signal; where m is a positive integer and 1≤m≤M, M is the total number of samples, and the number of time domain sampling points of the mth framed signal is 216; Step 102: Using a computer, zero padding is performed on the mth framed signal to obtain the mth transformed signal; wherein the number of FFT points of the mth transformed signal is 2048; Step 103: Process the mth transformed signal using the Mel-frequency cepstral coefficient method using a computer to obtain the energy S of the mth transformed signal output after passing through the i-th triangular filter. m (i); where i is a positive integer and 1≤i≤I; I is the total number of triangular Mel filters and I is 128; Step 104: Using a computer to calculate the m (i) = lnS m (i) The logarithmic energy E of the m-th transformed signal output by the i-th triangular filter is obtained. m (i); Step 105: According to the method of steps 103 and 104, the logarithmic energy vector of the m-th transformed signal output through I triangular filters is obtained and recorded as the m-th logarithmic energy vector; Step 106: Combine the 216 mth logarithmic energy vectors to obtain the logarithmic Mel-band energy matrix corresponding to the mth framed signal, and record it as the mth logarithmic Mel-band energy matrix H m , H m The size of is E×I, and E=216.

3. The method for identifying underwater target radiated noise based on deep learning according to claim 2, characterized in that: The size of the random mask matrix A in step 2 is E×I, and the element value A in the e-th row and i-th column of the random mask matrix A is ei is 0 or 1, and the number of 0 elements in the random mask matrix A is [35% MI] to [40% MI]; where [·] represents a rounding operation; The mth processed log-Mel band energy matrix H′ m The value of the element in row e and column i is E′ ei , and E′ ei =E ei ×A ei ;E ei Represents the mth log-Mel band energy matrix H m The value of the element in the e-th row and i-th column in ; e is a positive integer, and 1≤e≤E.

4. The method for identifying underwater target radiated noise based on deep learning according to claim 3, characterized in that: Step 3: The specific process is as follows: Step 301: Use a computer to convert the mth processed logarithmic Mel-band energy matrix H′ m Recorded as a two-dimensional tensor, and the two-dimensional tensor is input into the first convolution + BN layer for feature extraction to obtain the first feature map; Step 302: Input the first feature map into the first maximum pooling layer for feature extraction to obtain a second feature map; Step 303: Input the second feature map into the multi-scale feature module and pass it through the convolution layer 1, convolution layer 2, and convolution layer 3 respectively to obtain the first intermediate feature map, the second intermediate feature map, and the third intermediate feature map; Step 304: Concatenate the first intermediate feature map, the second intermediate feature map, and the third intermediate feature map through concat to obtain a first concatenated feature map; Step 305: Input the first concatenated feature map into the second maximum pooling layer for feature extraction to obtain a third feature map; Step 306: Process the third feature map through a sliding window attention mechanism to obtain a fourth feature map; Step 307: After the fourth feature map is flattened, it is converted into a one-dimensional vector; Step 308: Pass the one-dimensional vector through the first fully connected layer to output an intermediate one-dimensional vector; Step 309: Pass the intermediate one-dimensional vector through the second fully connected layer to output the converted one-dimensional vector; Step 3010: Pass the converted one-dimensional vector through the Softmax layer to output the target category probability vector.

5. The method for identifying underwater target radiated noise based on deep learning according to claim 1, characterized in that: The first convolution + BN layer includes a Conv1 convolution layer and a batch normalization layer. The number of convolution kernels in the Conv1 convolution layer is 64, the size of the convolution kernel is 3×3, the sliding step is 1, and the padding is 1. The size of the pooling window in the first maximum pooling layer is 2×2, the window sliding step is 2, and the padding is 0; The multi-scale feature module includes convolution layer 1, convolution layer 2, and convolution layer 3. The number of convolution kernels in convolution layer 1 is 64, the size of the convolution kernel is 3×3, the sliding step is 1, and the padding is 1. The number of convolution kernels in the second convolution layer is 64, the size of the convolution kernel is 5×5, the sliding step is 1, and the padding is 2; The number of convolution kernels in convolution layer 3 is 64, the size of the convolution kernel is 7×7, the sliding step is 1, and the padding is 3; The size of the pooling window in the second maximum pooling layer is 2×2, the window sliding step is 2, and the padding is 0; The number of the first fully connected layers is 1024, and the number of the second fully connected layers is 5.

6. The method for identifying underwater target radiated noise based on deep learning according to claim 4, characterized in that: Step 5: The specific process is as follows: Step 501: Obtain the j-th category probability value corresponding to the m-th framed signal from the target category probability vector. Where j is a positive integer, 1≤j≤C, and C represents the total number of underwater target categories; Step 502: Using a computer Get the cross entropy loss function L; where, Indicates the true label corresponding to the signal after the mth frame, the value is 0 or 1; m ranges from 1 to P, where P is the batch size; Step 503: The computer uses the Adma optimizer to iteratively train the underwater target radiation noise recognition model using the cross entropy loss function L until the entire training set is trained, completing one iterative training. Step 504: Repeat steps 502 and 503, and perform iterative training until a preset number of iterative training times is met, thereby obtaining a trained underwater target radiation noise recognition model.

Images