A passive target line spectrum feature extraction method based on a sliding window deep convolutional neural network and reinforcement learning

Abstract

The application discloses a passive target line spectrum feature extraction method based on a sliding window deep convolutional neural network and reinforcement learning, and the method comprises the following steps: S1, constructing a LOFAR spectrum graph training / verification sample set; S2, constructing a sliding window deep convolutional neural network model based on a Pytorch framework; S3, constructing a model training loss function; S4, designing a model training optimizer and a parameter setting strategy; and S5, repeatedly iteratively training the sliding window deep convolutional neural network model constructed in S2 by using the LOFAR spectrum graph training / verification sample set constructed in S1, combining the loss function designed in S3 and the model training optimizer and the parameter setting strategy designed in S4. The application has a strong nonlinear relationship mapping capability, and the model training under the line spectrum reinforcement learning architecture improves the weak feature capturing capability of the model, and the processing result has high tolerance.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of sonar, and particularly relates to a passive target line spectrum feature extraction method based on a sliding window deep convolutional neural network and reinforcement learning. BACKGROUND

[0002] Passive sonar target recognition is a technology for distinguishing target types or ship types by using passive target radiation noise received by a sonar and other sensor information and providing target feature information, and is one of core key technologies in the field of sonar technology and an important means for providing target information for decision-making of commanders.

[0003] The core research content of passive sonar target recognition is extraction and modeling of underwater target noise features. Traditionally, a feature with clear physical meaning and separability and statistical invariance is extracted for identification. As one of the most representative passive target physical features, a line spectrum has been widely used in actual scenarios. However, for a long time, due to factors such as complex target sound generation mechanism, serious interference of platform and environmental noise, and serious time and space changes of the ocean channel, the traditional method depending on prior rules and signal processing means has limited ability to extract weak line spectrum features of targets, and it is difficult to meet the actual use requirements in the case of multi-target strong interference and complex environment.

[0004] In recent years, artificial intelligence technologies represented by deep learning have developed rapidly, and have been widely applied in computer vision, natural language processing and other fields, greatly promoting the innovation of information processing technology and having a profound impact on all aspects of human society. At present, many scholars have conducted research on deep learning methods and applications for passive target physical feature extraction and recognition. However, they usually directly borrow the successful application experience of mainstream deep learning methods in other fields, and the model construction, training and optimization also use existing mature frameworks and schemes, without combining with actual target feature knowledge, so the extraction and generalization ability of the model for high complexity and low signal-to-noise ratio line spectrum still needs to be improved. SUMMARY

[0005] The technical problem to be solved by the application is to provide a passive target line spectrum feature extraction method based on a sliding window deep convolutional neural network and reinforcement learning. The sliding window deep convolutional neural network architecture and line spectrum reinforcement learning method suitable for the characteristics of passive target data are constructed, and the deep learning network is used to realize autonomous extraction of line spectrum features. The deep learning network model structure proposed by the method is adapted to the LOFAR spectrum feature, has strong non-linear relationship mapping ability, and the model training under the line spectrum reinforcement learning architecture improves the weak feature capturing ability of the model, and the processing result has high tolerance.

[0006] The technical solution of this invention is to provide a passive target line spectrum feature extraction method based on a sliding window deep convolutional neural network and reinforcement learning. This method includes...

[0007] S1, Construct the LOFAR spectrum training / validation sample set;

[0008] S2, based on the PyTorch framework, constructs a sliding window deep convolutional neural network model, specifically including:

[0009] S2.1, construct 3 basic modules, namely basic module 1, basic module 2 and basic module 3;

[0010] S2.2, construct 4 basic units, namely basic unit 1, basic unit 2, basic unit 3, and basic unit 4;

[0011] S2.3, construct the entire sliding window deep convolutional neural network model. The specific construction method is as follows: add 4 parallel branches. Branch 1 includes a low-frequency data slicing layer, basic unit 1, and a Dropout layer in sequence. The data slicing layer is used to segment low-frequency segments x from the entire input sample. Patch,1 , that is, x Patch,1 =x(f start,1 :f end,1 ), where x is the complete sample, f start f1 is the starting frequency point, and f1 is the ending frequency point; branch 2 includes a low-frequency data slicing layer, basic unit 2, and a dropout layer in sequence, where the data slicing layer realizes the segmentation of low-frequency segments x from the entire input sample. Patch,2 , that is, x Patch,2 =x(f start,2 :f end,2 Branch 3 includes a mid-to-high frequency data slicing layer, basic unit 3, and a Dropout layer, where the data slicing layer segments the entire input sample to obtain mid-to-high frequency fragments x. Patch,3 , that is, x Patch,3 =x(f start,3 :f end,3 Branch 4 consists of a high-frequency data slicing layer, basic unit 4, and a Dropout layer. The data slicing layer segments the entire input sample to obtain high-frequency segments x. Patch,4 , that is, x Patch,4 =x(f start,4 :f end,4 Add a network integration layer to concatenate the outputs of the four branches; a fully connected layer reduces the dimensionality of the input features to match the label dimensions, and outputs the result.

[0012] S3, Construct the model training loss function;

[0013] S4, design model training optimizer and parameter setting strategy;

[0014] S5, use the LOFAR spectrum constructed in S1 to train / validate the sample set, combine the loss function designed in S3 and the model training optimizer and parameter setting strategy designed in S4, and repeatedly iterate the sliding window deep convolutional neural network model constructed in S2 to train;

[0015] S6, line spectrum extraction is performed on the unknown underwater acoustic target LOFAR spectrum, the sliding window deep convolutional neural network model is directly processed based on the trained sliding window deep convolutional neural network model, and the line spectrum extraction result is obtained.

[0016] As preferred, in S1, a large-scale target time-domain waveform with randomly distributed line spectrum quantity, amplitude, frequency, duration, bending / offset / undulation is generated based on the underwater acoustic target radiation noise model; LOFAR analysis processing is performed on the target time-domain waveform to generate a large-scale normalized LOFAR spectrum sample with a frequency resolution of 0.1 Hz, a frequency point number of 1024, and a cumulative batch number of 32, and a corresponding line spectrum sequence label; and the sample is divided into a training sample set and a verification sample set independently according to a ratio of 100:1.

[0017] As preferred, the construction method of S2.1 is as follows,

[0018] S2.1.1, a basic module 1 is constructed, the output channel number is a settable parameter x, and three parallel branches are added, wherein branch 1 includes a convolution layer (7x7, x, 1), a LayerNorm layer, a convolution layer (1x1, 4x, 1), a GELU activation function, a convolution layer (1x1, x, 1), and a coefficient weighting layer in sequence, wherein for (7x7, x, 1), 7x7 is the convolution kernel size, x is the convolution channel number, and 1 is the step length; for the coefficient weighting layer, it realizes point multiplication of the convolution feature in the channel dimension with a weighting vector, and the length of the weighting vector is the same as the convolution feature channel number; branch 2 includes a convolution layer (5x5, x, 1), a LayerNorm layer, a convolution layer (1x1, 4x, 1), a GELU activation function, a convolution layer (1x1, x, 1), and a coefficient weighting layer in sequence; branch 3 is a direct connection layer; the convolution features output by the three branches are added to obtain the final output result of the basic module 1;

[0019] S2.1.2, a basic module 2 is constructed, the output channel number is a settable parameter x, and two parallel branches are added, wherein branch 1 includes a convolution layer (7x7, x, 1), a LayerNorm layer, a convolution layer (1x1, 4x, 1), a GELU activation function, a convolution layer (1x1, x, 1), and a coefficient weighting layer in sequence; branch 2 is a direct connection layer; the convolution features output by the two branches are added to obtain the final output result of the basic module 2;

[0020] S2.1.3, construct a basic module 3, the number of output channels is a settable parameter x, add 2 parallel branches, wherein branch 1 comprises, in sequence, a convolutional layer (5x5, x, 1), a LayerNorm layer, a convolutional layer (1x1, 4x, 1), a GELU activation function, a convolutional layer (1x1, x, 1), and a coefficient weighting layer; branch 2 is a straight connection layer; the convolutional features output by the two branches are added to obtain the final output result of the basic module 3.

[0021] As a preference, S2.2 is constructed as follows,

[0022] S2.2.1, construct a basic unit 1, add a convolutional layer (1x4, 128, 2); add a LayerNorm layer; add 3 basic modules 1, each of which has an output channel number x of 128; add a LayerNorm layer and a convolutional layer (2x2, 256, 2); add 3 basic modules 1, each of which has an output channel number x of 256; add a LayerNorm layer and a convolutional layer (2x2, 512, 2); add 15 basic modules 1, each of which has an output channel number x of 512; add a LayerNorm layer and a convolutional layer (2x2, 1024, 2); add 3 basic modules 1, each of which has an output channel number x of 1024; add a global average pooling layer; and add a fully connected layer (1024, 256), with the data in the parentheses being the input and output data dimensions, respectively;

[0023] S2.2.2, construct a basic unit 2, add a convolutional layer (1x4, 128, 2); add a LayerNorm layer; add 3 basic modules 2, each of which has an output channel number x of 128; add a LayerNorm layer and a convolutional layer (2x2, 256, 2); add 3 basic modules 2, each of which has an output channel number x of 256; add a LayerNorm layer and a convolutional layer (2x2, 512, 2); add 15 basic modules 2, each of which has an output channel number x of 512; add a LayerNorm layer and a convolutional layer (2x2, 1024, 2); add 3 basic modules 2, each of which has an output channel number x of 1024; add a global average pooling layer; and add a fully connected layer (1024, 256);

[0024] S2.2.3, construct basic unit 3, add convolutional layer (1×4, 96, 2); add LayerNorm layer; add 3 basic modules 2, each with output channel number x set to 96; add LayerNorm layer and convolutional layer (2×2, 192, 2); add 3 basic modules 2, each with output channel number x set to 192; add LayerNorm layer and convolutional layer (2×2, 192, 2); add 9 basic modules 2, each with output channel number x set to 384; add LayerNorm layer and convolutional layer (2×2, 768, 2); add 3 basic modules 2, each with output channel number x set to 768; add global average pooling layer; add fully connected layer (768, 256);

[0025] S2.2.4, construct basic unit 4, add convolutional layer (1×4, 96, 2); add LayerNorm layer; add 3 basic modules 3, each with output channel number x set to 96; add LayerNorm layer and convolutional layer (2×2, 192, 2); add 3 basic modules 3, each with output channel number x set to 192; add LayerNorm layer and convolutional layer (2×2, 192, 2); add 9 basic modules 3, each with output channel number x set to 384; add LayerNorm layer and convolutional layer (2×2, 768, 2); add 3 basic modules 3, each with output channel number x set to 768; add global average pooling layer; add fully connected layer (768, 256).

[0026] As a preferred option, the process of constructing the model training loss function in S3 is as follows, setting the model output as R. Model The sample label is y Label , will y Label Divide into n subsequences according to amplitude, that is

[0027]

[0028] Where L0 represents the result of the radio spectrum in the tag, and L1~L n This indicates the result of dividing the line spectrum in the label according to amplitude. length(y Label ) represents the length of the label sequence; based on y Label The partitioning result, i.e., L0~L n The corresponding positions i0~i n , for R Model The division is as shown in the following formula.

[0029]

[0030] Based on the above amplitude division results, a loss function is constructed as shown in the following formula.

[0031]

[0032] Where, N B The batch size during training, γ0~γ n These are weighting coefficients, which can be set according to the actual situation.

[0033] Furthermore, in S4, the process of designing the model training optimizer and parameter setting strategy is as follows:

[0034] S4.1 divides the training process into several stages according to the number of training rounds. The first stage of model training uses the SDG optimizer, and the remaining stages of model training use the Adam optimizer.

[0035] S4.2, during the training process of the model at each stage, several line spectrum extraction accuracy evaluations are performed based on the validation dataset. The model loss function value is recorded simultaneously during the evaluation, forming a set of validation accuracy and a set of loss function values ​​for that stage. The set of validation accuracy is then divided into two subsets A in chronological order. Present,1 and A Present,2 Simultaneously, the set of loss function values ​​is divided into two subsets L in chronological order. Present,1 and L Present,2 The learning rate adjustment method for the next stage of model training is shown in the following formula.

[0036]

[0037] Where, N Present Let Lr be the number of validations during the current training phase, α and β be the decision coefficients, and Lr be the number of validations. Next and Lr Present These are the learning rates for the next stage and the current stage of training, respectively.

[0038] Compared with the prior art, the present invention has the following advantages after adopting the above solution:

[0039] This invention utilizes a sliding window deep convolutional neural network architecture and a line spectrum reinforcement learning method adapted to the characteristics of passive target data to achieve autonomous extraction of line spectrum features. The deep learning network model structure proposed in this invention is adapted to LOFAR spectral features, exhibiting strong nonlinear relationship mapping capabilities. Simultaneously, model training under the line spectrum reinforcement learning architecture enhances the model's ability to capture weak features, resulting in highly tolerant processing results. Attached image description:

[0040] Figure 1 The following is a flowchart of signal processing;

[0041] Figure 2 The image shows the sliding window deep convolutional neural network architecture proposed in this invention;

[0042] Figure 3 The diagram shows three basic modules used in constructing the sliding window deep convolutional neural network of this invention.

[0043] Figure 4 The figure shown is for the present invention. Figure 2 A schematic diagram of the basic structure for modeling using basic units;

[0044] Figure 5 The diagram shown is a schematic of the reinforcement learning framework of a sliding window deep convolutional neural network model.

[0045] Figure 6 The image shows the results of line spectral feature extraction from a simulated LOFAR spectrum based on a trained sliding window deep convolutional neural network model. Figure 6 (a) shows the original simulated LOFAR spectrum. Figure 6 (b) shows the actual line spectrum sequence in the simulated LOFAR spectrum. Figure 6 (c) shows the line spectrum extraction results;

[0046] Figure 7 The image shows the result of line spectral feature extraction from the LOFAR spectrum of an actual sea trial target based on this sliding window deep convolutional neural network model. Figure 7 (a) shows the original simulated LOFAR spectrum. Figure 7 (b) shows the line spectrum extraction results. Detailed implementation method:

[0047] The present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0048] This invention first constructs a sliding window deep convolutional neural network based on the characteristics of LOFAR spectra and knowledge of passive target line spectra. Simultaneously, to enhance the extraction capability of weak line spectra features, a model reinforcement learning method is designed, considering loss function, optimizer, and learning rate setting strategies. Secondly, a large-scale simulation training / validation sample set is constructed to train the line spectra feature extraction model. Finally, based on the trained line spectra feature extraction model, features are extracted from unknown LOFAR spectra, and the line spectra extraction results are output.

[0049] Its flowchart is as follows Figure 1 As shown, a passive target line spectrum feature extraction method based on sliding window deep convolutional neural network and reinforcement learning is presented. This method includes...

[0050] S1, Construct the LOFAR spectral training / validation sample set, the specific implementation is as follows.

[0051] S1.1, Generate the target time-domain waveform with random distribution of line spectrum number, amplitude, frequency, duration, bending / offset / fluctuation based on the underwater acoustic target radiation noise model;

[0052] S1.2, The target time-domain waveform is processed using the traditional LOFAR processing method to generate LOFAR spectrum samples with a frequency resolution of 0.1Hz, a number of frequency points of 1024, and a cumulative batch number of 32. The corresponding label is a one-dimensional sequence with a length of 1024, which is the same as the number of frequency points in the LOFAR spectrum. The sequence is assigned values ​​according to the corresponding frequency points and intensities based on the theoretical line spectrum to generate sample labels.

[0053] S1.3 Normalize the LOFAR spectral samples and their labels.

[0054] S1.4, repeat S1.1 to S1.3, generate more labeled LOFAR spectral samples to form a large-scale sample set, and divide it into a 100:1 ratio of training sample set and validation sample set.

[0055] This completes the construction of the training sample set.

[0056] S2, based on the PyTorch framework, constructs a sliding window deep convolutional neural network model, specifically including:

[0057] S2.1, construct 3 basic modules, namely basic module 1, basic module 2 and basic module 3;

[0058] S2.2, construct 4 basic units, namely basic unit 1, basic unit 2, basic unit 3, and basic unit 4;

[0059] S2.3, Construct the entire sliding window deep convolutional neural network model.

[0060] Specifically, such as Figure 2 , 3 As shown in Figure 4, Figure 2The diagram shows the sliding window deep convolutional neural network architecture proposed in this invention. Addressing the high frequency resolution and significantly greater frequency dimension points compared to the time dimension characteristics of LOFAR spectra, it constructs low-frequency, mid-low-frequency, mid-high-frequency, and high-frequency data slicing layers. A sliding window approach is used to truncate the original LOFAR spectra with a certain overlap, forming multi-scale segments. Based on this, corresponding basic units are constructed for each frequency band to achieve deep mining of line spectrum features. This process must fully consider the richness of feature information in each frequency band and combine the line spectrum distribution patterns of different platforms to construct model units with a structure size adapted to the data processing complexity. Then, a Dropout processing layer is placed after each basic unit. Finally, the line spectrum features extracted from each frequency band are combined based on the network integration layer and the fully connected layer to output a line spectrum sequence. Figure 3 The diagram shows the three basic modules used in constructing the sliding window deep convolutional neural network of this invention. Unlike general convolutional neural networks, the basic modules in this patent all adopt the LN layer normalization method and the GELU activation function. At the same time, the basic modules enhance feature extraction capabilities by combining multiple convolutional layers in series and parallel. In addition, by using direct connection branches, each module is suitable for constructing large deep neural networks.

[0061] The specific construction method of S2.1 is as follows:

[0062] S2.1.1, construct basic module 1, with the number of output channels settable by parameter x. Add 3 parallel branches. Branch 1 includes a convolutional layer (7×7,x,1), a LayerNorm layer, a convolutional layer (1×1,4x,1), a GELU activation function, a convolutional layer (1×1,x,1), and a coefficient-weighted layer. For (7×7,x,1), 7×7 is the kernel size, x is the number of convolutional channels, and 1 is the stride. For the coefficient-weighted layer, it performs a dot-multiplication of the convolutional features in the channel dimension to form a weighted vector. The length of this weighted vector is the same as the number of convolutional feature channels. Branch 2 includes a convolutional layer (5×5,x,1), a LayerNorm layer, a convolutional layer (1×1,4x,1), a GELU activation function, a convolutional layer (1×1,x,1), and a coefficient-weighted layer. Branch 3 is a direct connection layer. The convolutional features output from the three branches are summed to obtain the final output of basic module 1.

[0063] S2.1.2, construct basic module 2, with the number of output channels settable by parameter x. Add two parallel branches. Branch 1 consists of a convolutional layer (7×7,x,1), a LayerNorm layer, a convolutional layer (1×1,4x,1), a GELU activation function, a convolutional layer (1×1,x,1), and a coefficient-weighted layer; Branch 2 is a direct-connected layer; sum the convolutional features output from the three branches to obtain the final output of basic module 2.

[0064] In step S2.1.3, construct basic module 3 with the number of output channels settable by parameter x. Add two parallel branches. Branch 1 consists of a convolutional layer (5×5, x, 1), a LayerNorm layer, a convolutional layer (1×1, 4x, 1), a GELU activation function, a convolutional layer (1×1, x, 1), and a coefficient-weighted layer. Branch 2 is a direct-connected layer. The convolutional features output from the three branches are summed to obtain the final output of basic module 3.

[0065] Figure 4 The figure shown is for the present invention. Figure 2 The basic structure for modeling using basic units mainly includes multiple convolutional layers for feature size compression and several basic modules for feature depth extraction. Each unit will select and combine basic modules according to the computational power requirements for line spectrum feature extraction, as detailed in the specific construction method in S2.2.

[0066] S2.2.1, Construct basic unit 1. Add a convolutional layer (1×4, 128, 2); add a LayerNorm layer; add 3 basic modules 1, each with an output channel count x of 128; add a LayerNorm layer and a convolutional layer (2×2, 256, 2); add 3 basic modules 1, each with an output channel count x of 256; add a LayerNorm layer and a convolutional layer (2×2, 512, 2); add 15 basic modules 1, each with an output channel count x of 512; add a LayerNorm layer and a convolutional layer (2×2, 1024, 2); add 3 basic modules 1, each with an output channel count x of 1024; add a global average pooling layer; add a fully connected layer (1024, 256). The data in parentheses represent the input and output data dimensions, respectively, and the same applies below.

[0067] S2.2.2, Construct basic unit 2. Add a convolutional layer (1×4, 128, 2); add a LayerNorm layer; add 3 basic modules 2, each with an output channel count x of 128; add a LayerNorm layer and a convolutional layer (2×2, 256, 2); add 3 basic modules 2, each with an output channel count x of 256; add a LayerNorm layer and a convolutional layer (2×2, 512, 2); add 15 basic modules 2, each with an output channel count x of 512; add a LayerNorm layer and a convolutional layer (2×2, 1024, 2); add 3 basic modules 2, each with an output channel count x of 1024; add a global average pooling layer; add a fully connected layer (1024, 256).

[0068] S2.2.3, Construct basic unit 3. Add a convolutional layer (1×4, 96, 2); add a LayerNorm layer; add 3 basic modules 2, each with an output channel count x of 96; add a LayerNorm layer and a convolutional layer (2×2, 192, 2); add 3 basic modules 2, each with an output channel count x of 192; add a LayerNorm layer and a convolutional layer (2×2, 192, 2); add 9 basic modules 2, each with an output channel count x of 384; add a LayerNorm layer and a convolutional layer (2×2, 768, 2); add 3 basic modules 2, each with an output channel count x of 768; add a global average pooling layer; add a fully connected layer (768, 256).

[0069] S2.2.4, Construct basic unit 4. Add a convolutional layer (1×4, 96, 2); add a LayerNorm layer; add 3 basic modules 3, each with an output channel count x of 96; add a LayerNorm layer and a convolutional layer (2×2, 192, 2); add 3 basic modules 3, each with an output channel count x of 192; add a LayerNorm layer and a convolutional layer (2×2, 192, 2); add 9 basic modules 3, each with an output channel count x of 384; add a LayerNorm layer and a convolutional layer (2×2, 768, 2); add 3 basic modules 3, each with an output channel count x of 768; add a global average pooling layer; add a fully connected layer (768, 256).

[0070] In S2.3, the specific construction method of the entire sliding window deep convolutional neural network model is as follows: four parallel branches are added. Branch 1 includes a low-frequency data slicing layer, basic unit 1, and a Dropout layer in sequence. The data slicing layer is used to segment low-frequency segments x from the entire input sample. Patch,1 , that is, x Patch,1 =x(f start,1 :f end,1 ), where x is the complete sample, f start f1 is the starting frequency point, and f1 is the ending frequency point; branch 2 includes a low-frequency data slicing layer, basic unit 2, and a dropout layer in sequence, where the data slicing layer realizes the segmentation of low-frequency segments x from the entire input sample. Patch,2 , that is, x Patch,2 =x(f start,2 :f end,2 Branch 3 includes a mid-to-high frequency data slicing layer, basic unit 3, and a Dropout layer, where the data slicing layer segments the entire input sample to obtain mid-to-high frequency fragments x. Patch,3 , that is, x Patch,3 =x(fstart,3 :f end,3 Branch 4 consists of a high-frequency data slicing layer, basic unit 4, and a Dropout layer. The data slicing layer segments the entire input sample to obtain high-frequency segments x. Patch,4 , that is, x Patch,4 =x(f start,4 :f end,4 Add a network integration layer to concatenate the outputs of the four branches; a fully connected layer reduces the dimensionality of the input features to match the label dimension, and then outputs the results.

[0071] The above process describes the construction of a sliding window deep convolutional neural network model.

[0072] Next, we move on to S3, where we construct the model training loss function.

[0073] like Figure 5 The sliding window deep convolutional neural network model reinforcement learning framework shown is designed to enhance the extraction capability of weak line spectrum features. It statistically analyzes and hierarchically categorizes the strength of target line spectrum features based on amplitude, energy, and other factors. When constructing the loss function, it weights line spectrum features at different levels. Furthermore, the model training process is divided into several stages. Based on the loss function value and the changing trend of line spectrum extraction performance on the validation set, the optimizer, learning rate, and weighting coefficients of line spectrum features at the same level are adaptively set for each stage, enhancing the model's ability to learn weak features during training.

[0074] The specific implementation method is as follows, setting the model output to R. Model The sample label is y Label , will y Label Divide into n subsequences according to amplitude, that is

[0075]

[0076] Where L0 represents the result of the radio spectrum in the tag, and L1~L n This indicates the result of dividing the line spectrum in the label according to amplitude. length(y Label ) represents the length of the label sequence. Based on y Label The partitioning result, i.e., L0~L n The corresponding positions i0~i n , for R Model The division is as shown in the following formula.

[0077]

[0078] Based on the above amplitude division results, a loss function is constructed as shown in the following formula.

[0079]

[0080] Where, N B The batch size during training, γ0~γ n These are weighting coefficients, which can be set according to the actual situation.

[0081] S4, Design the model training optimizer and parameter setting strategy.

[0082] The process is implemented as follows.

[0083] S4.1 divides the training process into several stages according to the number of training rounds. The first stage of model training uses the SDG optimizer, while the remaining stages use the Adam optimizer.

[0084] S4.2, during the training process of the model at each stage, several line spectrum extraction accuracy evaluations are performed based on the validation dataset. The model loss function value is recorded simultaneously during the evaluation, forming a set of validation accuracy and a set of loss function values ​​for that stage. The set of validation accuracy is then divided into two subsets A in chronological order. Present,1 and A Present,2 Simultaneously, the set of loss function values ​​is divided into two subsets L in chronological order. Present,1 and L Present,2 The learning rate adjustment method for the next stage of model training is shown in the following formula.

[0085]

[0086] Where, N Present Let Lr be the number of validations during the current training phase, α and β be the decision coefficients, and Lr be the number of validations. Next and Lr Present These are the learning rates for the next stage and the current stage of training, respectively.

[0087] S5 utilizes the LOFAR spectral training / validation sample set constructed in S1, combined with the loss function designed in S3 and the model training optimizer and parameter setting strategy designed in S4, to iteratively train the sliding window deep convolutional neural network model constructed in S2.

[0088] S6. Line spectrum extraction is performed on the LOFAR spectrum of the unknown underwater acoustic target. The LOFAR spectrum is directly processed based on the already trained sliding window deep convolutional neural network model to obtain the line spectrum extraction result.

[0089] like Figure 6 The image shows the result of line spectral feature extraction from a simulated LOFAR spectrum based on a trained sliding-window deep convolutional neural network model. This sliding-window deep convolutional neural network model has been trained on a large amount of labeled simulated LOFAR spectrum data. Figure 6 (a)Figure 6 (b) and Figure 6 (c) As can be seen, the extracted line spectrum sequence is very close to the true value, and the model can also effectively extract line spectra under conditions such as low signal-to-noise ratio and bending shift.

[0090] Figure 7 The image shows the result of line spectral feature extraction from the LOFAR spectrum of an actual sea trial target based on this sliding window deep convolutional neural network model. Figure 7 (a) shows the original simulated LOFAR spectrum. Figure 7 (b) shows the line spectrum extraction results. As can be seen, the model has a good ability to extract actual line spectra.

[0091] This invention utilizes deep learning networks to achieve autonomous extraction of line spectrum features. Compared with traditional line spectrum extraction methods and existing intelligent line spectrum extraction methods, the deep learning network model structure proposed in this invention is more compatible with LOFAR spectral features, has a stronger ability to map nonlinear relationships, and the model training under the line spectrum enhancement learning architecture improves the model's ability to capture weak features, resulting in more tolerant processing results.

[0092] The above description only illustrates preferred embodiments of the present invention and should not be construed as limiting the scope of the claims. Any equivalent structural or procedural modifications made using this specification are included within the patent protection scope of the present invention.

Claims

1. A passive target line spectral feature extraction method based on sliding window deep convolutional neural network and reinforcement learning, characterized in that: The method comprises, S1, constructing a LOFAR spectrum training / verification sample set; S2, constructing a sliding window deep convolutional neural network model based on a Pytorch framework, specifically comprising: S2.1, constructing three basic modules, namely a basic module 1, a basic module 2 and a basic module 3; S2.2, constructing four basic units, namely a basic unit 1, a basic unit 2, a basic unit 3 and a basic unit 4; S2.3, build the entire sliding window deep convolutional neural network model, the specific construction method is as follows, add 4 parallel branches, branch 1 includes low frequency data slice layer, basic unit 1 and Dropout layer in turn, wherein the data slice layer realizes the segmentation of low frequency segment x from the entire input sample Patch,1 , that is, x Patch,1 =x(f start,1 :f end,1 ), x is the complete sample, f start is the starting frequency point, f1 is the cut-off frequency point; branch 2 includes low-middle frequency data slice layer, basic unit 2 and Dropout layer in turn, wherein the data slice layer realizes the segmentation of low-middle frequency segment x Patch,2 , that is, x Patch,2 =x(f start,2 :f end,2 ); branch 3 includes middle-high frequency data slice layer, basic unit 3 and Dropout layer in turn, wherein the data slice layer realizes the segmentation of middle-high frequency segment x Patch,3 , that is, x Patch,3 =x(f start,3 :f end,3 ); branch 4 includes high frequency data slice layer, basic unit 4 and Dropout layer in turn, wherein the data slice layer realizes the segmentation of high frequency segment x Patch,4 , that is, x Patch,4 =x(f start,4 :f end,4 ); add network integration layer to splice the outputs of the 4 branches; full connection layer, reduce the dimension of the input features, so that the dimension is consistent with the label dimension, output the result; S3, constructing a model training loss function; S4, designing a model training optimizer and a parameter setting strategy; S5, using the LOFAR spectrum training / verification sample set constructed in S1, combining the loss function designed in S3 and the model training optimizer and the parameter setting strategy designed in S4, repeatedly iteratively training the sliding window deep convolutional neural network model constructed in S2; S6, performing line spectrum extraction on an unknown underwater target LOFAR spectrum, directly processing the LOFAR spectrum based on the trained sliding window deep convolutional neural network model to obtain a line spectrum extraction result; The construction manner of S2.1 is as follows, S2.1.1, constructing the basic module 1, the output channel number being a settable parameter x, adding three parallel branches, wherein branch 1 comprises, in sequence, a convolutional layer (7x7, x, 1), a LayerNorm layer, a convolutional layer (1x1, 4x, 1), a GELU activation function, a convolutional layer (1x1, x, 1) and a coefficient weighting layer, wherein for (7x7, x, 1), 7x7 is the convolution kernel size, x is the convolution channel number and 1 is the step length, and for the coefficient weighting layer, it realizes point multiplication of the convolution feature on the channel dimension with a weighting vector, and the weighting vector has the same length as the convolution feature channel number; branch 2 comprises, in sequence, a convolutional layer (5x5, x, 1), a LayerNorm layer, a convolutional layer (1x1, 4x, 1), a GELU activation function, a convolutional layer (1x1, x, 1) and a coefficient weighting layer; branch 3 is a direct connection layer; the convolution features output by the three branches are added to obtain the final output result of the basic module 1; S2.1.2, constructing the basic module 2, the output channel number being a settable parameter x, adding two parallel branches, wherein branch 1 comprises, in sequence, a convolutional layer (7x7, x, 1), a LayerNorm layer, a convolutional layer (1x1, 4x, 1), a GELU activation function, a convolutional layer (1x1, x, 1) and a coefficient weighting layer; branch 2 is a direct connection layer; the convolution features output by the two branches are added to obtain the final output result of the basic module 3; S2.1.3, constructing the basic module 3, the output channel number being a settable parameter x, adding two parallel branches, wherein branch 1 comprises, in sequence, a convolutional layer (5x5, x, 1), a LayerNorm layer, a convolutional layer (1x1, 4x, 1), a GELU activation function, a convolutional layer (1x1, x, 1) and a coefficient weighting layer; branch 2 is a direct connection layer; the convolution features output by the two branches are added to obtain the final output result of the basic module 3; The construction manner of S2.2 is as follows, S2.2.1, build basic unit 1, add convolution layer (1x4, 128, 2); add LayerNorm layer; add 3 basic modules 1, each of which is set to 128 in output channel number x; add LayerNorm layer and convolution layer (2x2, 256, 2); add 3 basic modules 1, each of which is set to 256 in output channel number x; add LayerNorm layer and convolution layer (2x2, 512, 2); add 15 basic modules 1, each of which is set to 512 in output channel number x; add LayerNorm layer and convolution layer (2x2, 1024, 2); add 3 basic modules 1, each of which is set to 1024 in output channel number x; add global average pooling layer; add fully connected layer (1024, 256), the data in the parentheses are input and output data dimensions in turn; S2.2.2, build basic unit 2, add convolution layer (1x4, 128, 2); add LayerNorm layer; add 3 basic modules 2, each of which is set to 128 in output channel number x; add LayerNorm layer and convolution layer (2x2, 256, 2); add 3 basic modules 2, each of which is set to 256 in output channel number x; add LayerNorm layer and convolution layer (2x2, 512, 2); add 15 basic modules 2, each of which is set to 512 in output channel number x; add LayerNorm layer and convolution layer (2x2, 1024, 2); add 3 basic modules 2, each of which is set to 1024 in output channel number x; add global average pooling layer; add fully connected layer (1024, 256); S2.2.3, build basic unit 3, add convolution layer (1x4, 96, 2); add LayerNorm layer; add 3 basic modules 2, each of which is set to 96 in output channel number x; add LayerNorm layer and convolution layer (2x2, 192, 2); add 3 basic modules 2, each of which is set to 192 in output channel number x; add LayerNorm layer and convolution layer (2x2, 192, 2); add 9 basic modules 2, each of which is set to 384 in output channel number x; add LayerNorm layer and convolution layer (2x2, 768, 2); add 3 basic modules 2, each of which is set to 768 in output channel number x; add global average pooling layer; add fully connected layer (768, 256); S2.2.4, build the basic unit 4, add a convolution layer (1x4, 96, 2); add a LayerNorm layer; add 3 basic modules 3, and the output channel number x of each module is set to 96; add a LayerNorm layer and a convolution layer (2x2, 192, 2); add 3 basic modules 3, and the output channel number x of each module is set to 192; add a LayerNorm layer and a convolution layer (2x2, 192, 2); add 9 basic modules 3, and the output channel number x of each module is set to 384; add a LayerNorm layer and a convolution layer (2x2, 768, 2); add 3 basic modules 3, and the output channel number x of each module is set to 768; add a global average pooling layer; and add a full connection layer (768, 256). 2.The passive target line spectral feature extraction method based on sliding window deep convolutional neural network and reinforcement learning according to claim 1, characterized in that: In S1, a large-scale target time-domain waveform with random distribution of line spectrum quantity, amplitude, frequency, duration, bending / offset / undulation is generated based on a model of underwater target radiated noise; The target time-domain waveform is subjected to LOFAR analysis processing to generate a large-scale normalized LOFAR spectrum sample with a frequency resolution of 0.1 Hz, a frequency point quantity of 1024 and a cumulative batch quantity of 32, and a corresponding line spectrum sequence label, and the sample is divided into a training sample set and a verification sample set which are independent of each other at a ratio of 100:

1. 3.The passive target line spectral feature extraction method based on sliding window deep convolutional neural network and reinforcement learning according to claim 1, wherein: The process of constructing the model training loss function in S3 is as follows: set the model output as R Model , the sample label as y Label , divide y Label into n subsequences according to the amplitude, that is wherein, L0 represents the result of the line spectrum in the tag, L1 ~ L n represents the result of dividing the line spectrum in the tag by amplitude, length(y Label ) represents the length of the tag sequence; according to y Label the division result, i.e. L0 ~ L n corresponding positions i0 ~ i n , R Model is divided, as shown in the following formula A loss function is constructed according to the above amplitude division result, as shown in the following formula where N B is the Batch size at training, γ0~γ n are the weighting coefficients. 4.The passive target line spectral feature extraction method based on sliding window deep convolutional neural network and reinforcement learning according to claim 1, wherein: In S4, the process of designing a model training optimizer and parameter setting strategy is as follows: In S4.1, the training process is divided into several stages according to the training round number, wherein the model training in the first stage adopts an SDG optimizer, and the model training in the remaining stages adopts an Adam optimizer; S4.2, during the training process of each stage of the model, based on the validation data set, several times of line spectrum extraction accuracy evaluation are carried out, and the model loss function value during evaluation is recorded simultaneously, forming the validation accuracy set and the loss function value set of the model at this stage, and the validation accuracy set is evenly divided into two subsets A Present,1 and A Present,2 , and the loss function value set is evenly divided into two subsets L Present,1 and L Present,2 , then the learning rate adjustment method of the next stage model training is as follows Wherein, N Present is the number of verification in the current stage training process, and α and β are decision coefficients, Lr Next and Lr Present are the learning rates of the next stage and the current stage training, respectively.

Images