Underwater acoustic spectrum segmentation communication method based on deep learning

Abstract

The present invention relates to the field of information processing, and in particular to an underwater acoustic spectrum segmentation communication method based on deep learning. A deep learning-based underwater acoustic spectrum segmentation communication method comprises: acquiring an underwater acoustic signal; generating an underwater acoustic spectrum graph through Mel-frequency cepstrum coefficients of the underwater acoustic signal; sending the underwater acoustic spectrum graph into an underwater acoustic spectrum segmentation model for processing to obtain an underwater acoustic spectrum annotation graph; mapping the available spectrum area in the underwater acoustic spectrum annotation graph to the underwater acoustic signal frequency band; and allocating all the underwater acoustic signal frequency bands to different network channels for transmission. The present invention converts the underwater acoustic signal into an underwater acoustic spectrum graph, and sends the underwater acoustic spectrum graph into an underwater acoustic spectrum segmentation model for processing. The underwater acoustic spectrum segmentation model adopts deep learning technology, and can segment and mark the available spectrum in the underwater acoustic spectrum graph based on the feature information in the underwater acoustic spectrum graph, thereby realizing the extraction and segmentation of the available frequency bands in the underwater acoustic signal.

Description

Technical Field

[0001] The present invention relates to the field of information processing, and in particular to an underwater acoustic spectrum segmentation communication method based on deep learning. Background Art

[0002] With the continuous development of marine resources and the growing demand for underwater communications, the application of underwater acoustic communication technology in military, scientific research, marine engineering and other fields has become increasingly important. However, due to the complexity and variability of the underwater environment, underwater acoustic communication faces many challenges such as signal attenuation, noise interference and multipath effects. Traditional underwater acoustic communication methods usually rely on specific frequency bands and preset channels, have poor adaptability to the environment, and are difficult to meet the needs of efficient and stable communication. Summary of the invention

[0003] The present invention converts the hydroacoustic signal into a hydroacoustic spectrum graph and feeds the hydroacoustic spectrum graph into a hydroacoustic spectrum segmentation model for processing. The hydroacoustic spectrum segmentation model adopts deep learning technology and can segment and mark the available spectrum in the hydroacoustic spectrum graph based on the feature information in the hydroacoustic spectrum graph, thereby realizing the extraction and segmentation of the available frequency bands in the hydroacoustic signal.

[0004] A deep learning-based underwater acoustic spectrum segmentation communication method, comprising:

[0005] Acquire hydroacoustic signals;

[0006] Generate a water acoustic spectrum diagram through the Mel frequency cepstrum coefficient of the water acoustic signal;

[0007] The underwater acoustic spectrum graph is sent to an underwater acoustic spectrum segmentation model for processing to obtain an underwater acoustic spectrum annotation graph, wherein the underwater acoustic spectrum annotation graph includes an available spectrum region and a noise spectrum region;

[0008] The available spectrum area in the hydroacoustic spectrum annotation diagram is mapped to the hydroacoustic signal frequency band;

[0009] All underwater acoustic signal frequency bands are allocated to different network channels for transmission to achieve underwater acoustic spectrum segmentation communication.

[0010] As a preferred aspect, the underwater acoustic spectrum segmentation model includes a preprocessing layer, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer and a graph segmentation layer, wherein the preprocessing layer is used to perform convolution operations on the underwater acoustic spectrum graph using convolution kernels of different sizes to construct a fused feature graph; the first feature is used to perform convolution, pooling and full connection operations on the fused feature graph to construct an underwater acoustic spectrum feature vector; the second feature extraction layer is used to perform temporal feature extraction on the underwater acoustic spectrum feature vector to construct an underwater acoustic spectrum temporal feature graph; the third feature extraction layer is used to perform feature extraction on the underwater acoustic spectrum temporal feature graph using a self-attention mechanism to construct an enhanced feature graph; the graph segmentation layer is used to perform deconvolution operations on the enhanced feature graph to construct an underwater acoustic spectrum annotation graph.

[0011] As a preferred aspect, the underwater acoustic spectrum graph is sent to an underwater acoustic spectrum segmentation model for processing to obtain an underwater acoustic spectrum annotation graph, which specifically includes the following steps:

[0012] In the preprocessing layer, the underwater acoustic spectrum is convolved by N convolution kernels of different sizes, and the N outputs of the underwater acoustic spectrum after the convolution operation are spliced ​​according to the channel to construct a fusion feature map, where N is 10, and the size of the N convolution kernels is calculated by simulation using the swarm optimization algorithm;

[0013] In the first feature extraction layer, convolution, pooling and full connection operations are performed on the fused feature map to construct an underwater acoustic spectrum feature vector;

[0014] The second feature extraction layer includes an LSTM unit. In the second feature extraction layer, a sliding window of size M is used to select the feature vector of the underwater acoustic spectrum, and the step size is H, H is less than M, and a number of local feature vectors are generated; the local feature vectors are sent to the LSTM unit in the order of selection for processing, and the outputs of all LSTM units are spliced ​​in order from top to bottom to construct a time series feature map of the underwater acoustic spectrum;

[0015] In the third feature extraction layer, the self-attention mechanism operation is performed G times for the underwater acoustic spectrum time series feature map, and all the outputs of the self-attention mechanism operation are spliced ​​according to the channel and nonlinearly changed to obtain an enhanced feature map;

[0016] In the graph segmentation layer, the enhanced feature map is deconvolved to construct an underwater acoustic spectrum annotation map.

[0017] As a preferred aspect, training the underwater acoustic spectrum segmentation model specifically includes the following steps:

[0018] A number of underwater acoustic spectrum segmentation training samples are obtained, where the underwater acoustic spectrum segmentation training samples are spectrum diagrams converted from the acquired underwater acoustic signals. The underwater acoustic spectrum segmentation training samples are segmented and labeled, and all the segmented and labeled underwater acoustic spectrum segmentation training samples are combined into an underwater acoustic spectrum segmentation training set. The underwater acoustic spectrum segmentation model is then trained with the underwater acoustic spectrum segmentation training set, and the accuracy of the underwater acoustic spectrum segmentation model is calculated. If the accuracy of the underwater acoustic spectrum segmentation model is higher than a preset value, the trained underwater acoustic spectrum segmentation model is output; otherwise, the underwater acoustic spectrum segmentation model is trained with the underwater acoustic spectrum segmentation training set, and the parameter values ​​in the underwater acoustic spectrum segmentation model are adjusted.

[0019] As a preferred aspect, the sizes of N convolution kernels, the sliding window size M and the step size H are simulated and calculated by a swarm optimization algorithm, specifically including the following steps:

[0020] Construct several parameter simulation individuals, and combine all constructed parameter simulation individuals into a parameter simulation population set, set the maximum number of iterations, and store the values ​​corresponding to the sizes of N convolution kernels, the sliding window size M, and the step size H in the parameter simulation individuals;

[0021] The fitness corresponding to the parameter simulation individual is calculated for each parameter simulation individual in the parameter simulation population set. The fitness corresponding to the parameter simulation individual is calculated as follows: the underwater acoustic spectrum segmentation model is set with the numerical value in the parameter simulation individual, and then the underwater acoustic spectrum segmentation model is trained with the underwater acoustic spectrum segmentation training set to obtain the accuracy A of the underwater acoustic spectrum segmentation model, and the fitness corresponding to the parameter simulation individual is calculated δ=αA+β / B, where B is the floating-point operation number of the set underwater acoustic spectrum segmentation model, and α and β are the accuracy weight coefficient and the floating-point operation weight coefficient respectively;

[0022] With the help of the fitness corresponding to all parameter simulation individuals, the parameter simulation population set is iteratively updated through the group optimization algorithm;

[0023] When the number of iterations reaches the maximum number of iterations, the parameter simulation individuals whose fitness is greater than the fitness threshold are output as the sizes of N convolution kernels, the sliding window size M and the step size H.

[0024] As a preferred aspect, the population optimization algorithm adopts a genetic algorithm, through which the parameter simulation population set is iteratively updated, during which selection, recombination and mutation operations are performed. The selection operation is to select a batch of parameter simulation individuals with the highest fitness in the parameter simulation population set as the parent set; the recombination operation is to perform multi-point crossover transformation on any two parameter simulation individuals in the parent set; the mutation operation is to randomly select a part of the parameter simulation individuals, and make numerical changes to a random position in the selected parameter simulation individuals.

[0025] As a preferred aspect, it also includes real-time training for the underwater acoustic spectrum segmentation model, and the specific steps are as follows:

[0026] The water acoustic spectrum segmentation real-time training data is constructed by using the water acoustic spectrum graph and the corresponding water acoustic spectrum annotation graph each time the water acoustic spectrum segmentation model is used, and a random part of the water acoustic spectrum segmentation real-time training samples are corrected, and the corrected water acoustic spectrum segmentation real-time training data is used as the water acoustic spectrum segmentation real-time training samples. When the water acoustic spectrum segmentation real-time training samples reach a preset upper limit, all the water acoustic spectrum segmentation real-time training samples are combined into a water acoustic spectrum segmentation real-time training set, and then the water acoustic spectrum segmentation model is trained in real time using the water acoustic spectrum segmentation real-time training set.

[0027] As a preferred aspect, the first feature is established based on a convolutional neural network, and the second feature extraction layer is established based on an LSTM model.

[0028] The present invention has the following advantages:

[0029] 1. The present invention converts the underwater acoustic signal into an underwater acoustic spectrum graph, and feeds the underwater acoustic spectrum graph into an underwater acoustic spectrum segmentation model for processing. The underwater acoustic spectrum segmentation model adopts deep learning technology, and can segment and mark the available spectrum in the underwater acoustic spectrum graph based on the feature information in the underwater acoustic spectrum graph, thereby realizing the extraction and segmentation of the available frequency bands in the underwater acoustic signal.

[0030] 2. The underwater acoustic spectrum segmentation model in the present invention can enhance features through the time dependency relationship of two learnings, thereby improving the segmentation effect of the underwater acoustic spectrum, and uses multiple different convolution kernels to perform convolution operations on the underwater acoustic spectrum graph in the preprocessing layer, which can adapt to the data and underwater environment corresponding to different underwater acoustic signals.

[0031] 3. The present invention simulates and calculates the size of the convolution kernel, the sliding window size M and the step size H through a genetic algorithm, which can optimize the parameter setting of the underwater acoustic spectrum segmentation model and improve the accuracy of the underwater acoustic spectrum segmentation model.

[0032] 4. The present invention also performs real-time training on the underwater acoustic spectrum segmentation model, which can make the underwater acoustic spectrum segmentation model adapt to changes in the underwater environment and improve the accuracy of the actual use of the underwater acoustic spectrum segmentation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A flowchart of the underwater acoustic spectrum segmentation communication method based on deep learning adopted in an embodiment of the present invention.

[0034] Figure 2 Schematic diagram of the structure of the underwater acoustic spectrum segmentation model used in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to enable persons skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0036] Embodiment 1, underwater acoustic spectrum segmentation communication method based on deep learning, such as Figure 1 As shown, including:

[0037] Acquire an underwater acoustic signal. The underwater acoustic signal is a sound wave signal obtained from a receiver during underwater acoustic communication. This sound wave signal is mixed with a lot of environmental noise. To continue communication, it is necessary to remove the environmental noise in this underwater acoustic signal and extract the available spectrum in the underwater acoustic signal to perform subsequent communication;

[0038] The underwater acoustic signal is converted into an underwater acoustic spectrum through the Mel-frequency cepstrum coefficient. The underwater acoustic spectrum is a two-dimensional matrix, with the horizontal axis representing time and the vertical axis representing frequency. The gray value in the underwater acoustic spectrum represents the signal strength. Converting the underwater acoustic signal into an underwater acoustic spectrum can make the characteristics of various frequency bands clearer. For example, the required available spectrum may appear as a short-time pulse signal within a certain specific frequency range, while the ship noise may appear as a continuous spectrum within a lower frequency range, which can help the subsequent identification and segmentation of the available spectrum.

[0039] The underwater acoustic spectrum graph is sent to the underwater acoustic spectrum segmentation model for processing to obtain an underwater acoustic spectrum annotation graph, which includes an available spectrum area and a noise spectrum area. The underwater acoustic spectrum annotation graph segments and marks the available spectrum and the noise spectrum in the underwater acoustic signal, constructs the available spectrum area and the noise spectrum area, and can segment the spectrum of the underwater acoustic signal through the underwater acoustic spectrum annotation graph;

[0040] Mapping the available spectrum area in the underwater acoustic spectrum annotation diagram to the underwater acoustic signal frequency band, the underwater acoustic signal frequency band includes the frequency interval corresponding to the available spectrum. Since the available spectrum area in the underwater acoustic spectrum annotation diagram corresponds to a time and frequency range on the underwater acoustic spectrum diagram, a corresponding frequency interval can be allocated to each available spectrum area, thereby realizing mapping to the underwater acoustic signal frequency band;

[0041] All underwater acoustic signal frequency bands are allocated to different network channels for transmission to achieve underwater acoustic spectrum segmentation communication.

[0042] The present application converts the underwater acoustic signal into an underwater acoustic spectrum graph and feeds the underwater acoustic spectrum graph into an underwater acoustic spectrum segmentation model for processing. The underwater acoustic spectrum segmentation model adopts deep learning technology and can segment and mark the available spectrum in the underwater acoustic spectrum graph based on the feature information in the underwater acoustic spectrum graph, thereby realizing the extraction and segmentation of the available frequency bands in the underwater acoustic signal.

[0043] See also Figure 2 The underwater acoustic spectrum segmentation model includes a preprocessing layer, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer and a graph segmentation layer, wherein the preprocessing layer is used to perform convolution operations on the underwater acoustic spectrum graph using convolution kernels of different sizes to construct a fused feature graph. It should be noted that the sizes of the available spectrum area and the noise spectrum area in the underwater acoustic spectrum graph are not fixed, and will change with the data corresponding to the underwater acoustic signal and the underwater environment. Therefore, learning local features of different sizes through different convolution kernels can improve the accuracy of underwater acoustic spectrum segmentation; the first feature is established based on a convolutional neural network, which is used to perform convolution, pooling and full connection operations on the fused feature graph to construct an underwater acoustic spectrum feature vector. The convolution, pooling and full connection operations in the convolutional neural network can effectively learn the feature signals in the fused feature graph. Information; The second feature extraction layer is established based on the LSTM model, which is used to extract the time series features of the underwater acoustic spectrum feature vector to construct the underwater acoustic spectrum time series feature map. The time series feature extraction of the underwater acoustic spectrum feature vector can learn the local time dependency in the underwater acoustic signal and judge different spectra at the time series level; The third feature extraction layer is used to extract features from the underwater acoustic spectrum time series feature map using the self-attention mechanism to construct an enhanced feature map. The self-attention mechanism can further learn the long-term time dependency in the underwater acoustic signal and enhance the underwater acoustic spectrum time series feature map based on the long-term time dependency; The graph segmentation layer is used to perform deconvolution operations on the enhanced feature map to construct an underwater acoustic spectrum annotation map. The deconvolution operation can expand the size of the enhanced feature map so that the enhanced feature map is restored to the size of the underwater acoustic spectrum map.

[0044] The underwater acoustic spectrum graph is sent to the underwater acoustic spectrum segmentation model for processing to obtain an underwater acoustic spectrum annotation graph, which specifically includes the following steps:

[0045] In the preprocessing layer, the underwater acoustic spectrum is convolved by N convolution kernels of different sizes, and the N outputs of the underwater acoustic spectrum after the convolution operation are spliced ​​according to the channels to construct a fusion feature map, where N is 10. The sizes of the N convolution kernels are calculated by swarm optimization algorithm, which can correspond to available spectrum areas and noise spectrum areas of different sizes, so that the local features corresponding to the subsequent convolution operation are more in line with the actual area, improving the effect of feature learning.

[0046] In the first feature extraction layer, convolution, pooling and full connection operations are performed on the fused feature map to construct the underwater acoustic spectrum feature vector. It should be noted that the second feature extraction layer is also consistent with the convolutional neural network, including 5 convolution layers, two maximum pooling layers and two fully connected layers;

[0047] The second feature extraction layer includes an LSTM unit. In the second feature extraction layer, the feature vector of the underwater acoustic spectrum is selected with a sliding window of size M, and the step size is H, H is less than M, and several local feature vectors are generated. The specific operation is to start from the first data of the underwater acoustic spectrum feature vector, select M data in the future, construct a local feature vector, and then start from the Hth data of the underwater acoustic spectrum feature vector, select M data in the future, construct a local feature vector, and so on, to generate several local feature vectors; the local feature vectors are sent to the LSTM unit in the order of selection for processing The outputs of all LSTM units are spliced ​​from top to bottom in order to construct the underwater acoustic spectrum time series feature map. Since the underwater acoustic spectrum feature vector includes the high-level information of the underwater acoustic spectrum map, in order to strengthen the time series characteristics of the underwater acoustic signal, several local feature vectors are constructed by interception, and the adjacent local feature vectors overlap each other. The time series characteristics are extracted by LSTM units, which can strengthen the time series characteristics. The LSTM unit is set with reference to the LSTM model; the sliding window size M and step size H are also calculated by simulation of the swarm optimization algorithm;

[0048] In the third feature extraction layer, G self-attention mechanism operations are performed on the underwater acoustic spectrum time series feature map. The self-attention mechanism operation here is set with reference to the Transformer model. All outputs of the self-attention mechanism operation are spliced ​​according to the channel and nonlinear changes are performed to obtain an enhanced feature map. Through the G self-attention mechanism operation test, the long-term dependency relationship in the underwater acoustic spectrum time series feature map is learned from G dimensions;

[0049] In the graph segmentation layer, the enhanced feature map is deconvolved to construct an underwater acoustic spectrum annotation map.

[0050] The underwater acoustic spectrum segmentation model in the present application can enhance features through the time dependency relationship of two learnings, thereby improving the segmentation effect of the underwater acoustic spectrum, and uses multiple different convolution kernels to perform convolution operations on the underwater acoustic spectrum graph in the preprocessing layer, which can adapt to the data and underwater environment corresponding to different underwater acoustic signals.

[0051] Training the underwater acoustic spectrum segmentation model includes the following steps:

[0052] A number of underwater acoustic spectrum segmentation training samples are obtained. The underwater acoustic spectrum segmentation training samples are spectrum diagrams converted from the acquired underwater acoustic signals. The underwater acoustic spectrum segmentation training samples are segmented and labeled. The segmentation and labeling operations are performed by developers. All segmented and labeled underwater acoustic spectrum segmentation training samples are combined into an underwater acoustic spectrum segmentation training set. The underwater acoustic spectrum segmentation model is then trained with the underwater acoustic spectrum segmentation training set. The accuracy of the underwater acoustic spectrum segmentation model is calculated. If the accuracy of the underwater acoustic spectrum segmentation model is higher than a preset value, which is set by the developer, the trained underwater acoustic spectrum segmentation model is output; otherwise, the underwater acoustic spectrum segmentation model is trained with the underwater acoustic spectrum segmentation training set, and the parameter values ​​in the underwater acoustic spectrum segmentation model are adjusted.

[0053] The size of N convolution kernels, sliding window size M and step size H are simulated and calculated by the swarm optimization algorithm, which specifically includes the following steps:

[0054] Construct several parameter simulation individuals, and combine all constructed parameter simulation individuals into a parameter simulation population set, set the maximum number of iterations, and store the values ​​corresponding to the sizes of N convolution kernels, the sliding window size M, and the step size H in the parameter simulation individuals. When constructing the parameter simulation individuals, each value in the parameter simulation individuals is a random number within the corresponding value range.

[0055] The fitness corresponding to the parameter simulation individual is calculated for each parameter simulation individual in the parameter simulation population set. The fitness corresponding to the parameter simulation individual is calculated as follows: the underwater acoustic spectrum segmentation model is set with the numerical value in the parameter simulation individual, and then the underwater acoustic spectrum segmentation model is trained with the underwater acoustic spectrum segmentation training set to obtain the accuracy A of the underwater acoustic spectrum segmentation model, and the fitness corresponding to the parameter simulation individual is calculated δ=αA+β / B, where B is the floating-point operation number of the set underwater acoustic spectrum segmentation model, α and β are the accuracy weight coefficient and the floating-point operation weight coefficient respectively. The fitness setting corresponding to the parameter simulation individual can improve the accuracy of the underwater acoustic spectrum segmentation model under the constraint of the floating-point operation number;

[0056] With the assistance of the fitness corresponding to all parameter simulation individuals, the parameter simulation population set is iteratively updated through a genetic algorithm, during which selection, recombination and mutation operations are performed. The selection operation can be to select a group of parameter simulation individuals with the highest fitness in the parameter simulation population set as the parent set; the recombination operation can be to perform multi-point crossover transformation on any two parameter simulation individuals in the parent set, that is, to randomly select multiple positions in two parameter simulation individuals for numerical exchange; the mutation operation can be to randomly select a part of parameter simulation individuals and perform numerical changes on a random position in the selected parameter simulation individuals;

[0057] When the number of iterations reaches the maximum number of iterations, the parameter simulation individuals whose fitness is greater than the fitness threshold are output as the sizes of N convolution kernels, the sliding window size M and the step size H.

[0058] The present application uses a genetic algorithm to simulate and calculate the size of the convolution kernel, the sliding window size M and the step size H, which can optimize the parameter setting of the underwater acoustic spectrum segmentation model and improve the accuracy of the underwater acoustic spectrum segmentation model.

[0059] Since the underwater environment is constantly changing, this application also includes real-time training for the underwater acoustic spectrum segmentation model. The specific steps are as follows:

[0060] The water acoustic spectrum segmentation real-time training data is constructed by combining the water acoustic spectrum graph and the corresponding water acoustic spectrum annotation graph each time the water acoustic spectrum segmentation model is used, and a random part of the water acoustic spectrum segmentation real-time training samples is corrected. The correction operation can be manually corrected by the annotation personnel, and the corrected water acoustic spectrum segmentation real-time training data is used as the water acoustic spectrum segmentation real-time training samples until the water acoustic spectrum segmentation real-time training samples reach the preset upper limit. The preset upper limit is set in advance, and all the water acoustic spectrum segmentation real-time training samples are combined into a water acoustic spectrum segmentation real-time training set, and then the water acoustic spectrum segmentation model is trained in real time through the water acoustic spectrum segmentation real-time training set. It should be added that the learning rate of the real-time training will be lower than the learning rate of the initial training of the water acoustic spectrum segmentation model to achieve fine-tuning of the parameters.

[0061] The present application also performs real-time training on the underwater acoustic spectrum segmentation model, which can enable the underwater acoustic spectrum segmentation model to adapt to changes in the underwater environment and improve the accuracy of the actual use of the underwater acoustic spectrum segmentation model.

[0062] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A deep learning-based underwater acoustic spectrum segmentation communication method, characterized in that: include: Acquire hydroacoustic signals; Generate a water acoustic spectrum diagram through the Mel frequency cepstrum coefficient of the water acoustic signal; The underwater acoustic spectrum graph is sent to an underwater acoustic spectrum segmentation model for processing to obtain an underwater acoustic spectrum annotation graph, wherein the underwater acoustic spectrum annotation graph includes an available spectrum region and a noise spectrum region; The available spectrum area in the hydroacoustic spectrum annotation diagram is mapped to the hydroacoustic signal frequency band; Allocate all underwater acoustic signal frequency bands to different network channels for transmission to achieve underwater acoustic spectrum segmentation communication; The underwater acoustic spectrum segmentation model includes a preprocessing layer, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer and a graph segmentation layer; The underwater acoustic spectrum graph is sent to the underwater acoustic spectrum segmentation model for processing to obtain an underwater acoustic spectrum annotation graph, which specifically includes the following steps: In the preprocessing layer, the underwater acoustic spectrum is convolved by N convolution kernels of different sizes, and the N outputs of the underwater acoustic spectrum after the convolution operation are spliced ​​according to the channel to construct a fusion feature map, where N is 10, and the size of the N convolution kernels is calculated by simulation using the swarm optimization algorithm; In the first feature extraction layer, convolution, pooling and full connection operations are performed on the fused feature map to construct an underwater acoustic spectrum feature vector; The second feature extraction layer includes an LSTM unit. In the second feature extraction layer, a sliding window of size M is used to select the feature vector of the underwater acoustic spectrum, and the step size is H, H is less than M, and a number of local feature vectors are generated; the local feature vectors are sent to the LSTM unit in the order of selection for processing, and the outputs of all LSTM units are spliced ​​in order from top to bottom to construct a time series feature map of the underwater acoustic spectrum; In the third feature extraction layer, the self-attention mechanism operation is performed G times for the underwater acoustic spectrum time series feature map, and all the outputs of the self-attention mechanism operation are spliced ​​according to the channel and nonlinearly changed to obtain an enhanced feature map; In the graph segmentation layer, the enhanced feature graph is deconvolved to construct an underwater acoustic spectrum annotation graph; The size of N convolution kernels, sliding window size M and step size H are simulated and calculated by the swarm optimization algorithm, which specifically includes the following steps: Construct several parameter simulation individuals, and combine all constructed parameter simulation individuals into a parameter simulation population set, set the maximum number of iterations, and store the values ​​corresponding to the sizes of N convolution kernels, the sliding window size M, and the step size H in the parameter simulation individuals; The fitness corresponding to the parameter simulation individual is calculated for each parameter simulation individual in the parameter simulation population set. The fitness corresponding to the parameter simulation individual is calculated as follows: the underwater acoustic spectrum segmentation model is set with the numerical value in the parameter simulation individual, and then the underwater acoustic spectrum segmentation model is trained with the underwater acoustic spectrum segmentation training set to obtain the accuracy A of the underwater acoustic spectrum segmentation model, and the fitness corresponding to the parameter simulation individual is calculated δ=αA+β / B, where B is the floating-point operation number of the set underwater acoustic spectrum segmentation model, and α and β are the accuracy weight coefficient and the floating-point operation weight coefficient respectively; With the help of the fitness corresponding to all parameter simulation individuals, the parameter simulation population set is iteratively updated through the group optimization algorithm; When the number of iterations reaches the maximum number of iterations, the parameter simulation individuals whose fitness is greater than the fitness threshold are output as the sizes of N convolution kernels, the sliding window size M and the step size H.

2. The underwater acoustic spectrum segmentation communication method based on deep learning according to claim 1 is characterized in that: Training the underwater acoustic spectrum segmentation model includes the following steps: A number of underwater acoustic spectrum segmentation training samples are obtained, where the underwater acoustic spectrum segmentation training samples are spectrum diagrams converted from the acquired underwater acoustic signals. The underwater acoustic spectrum segmentation training samples are segmented and labeled, and all the segmented and labeled underwater acoustic spectrum segmentation training samples are combined into an underwater acoustic spectrum segmentation training set. The underwater acoustic spectrum segmentation model is then trained with the underwater acoustic spectrum segmentation training set, and the accuracy of the underwater acoustic spectrum segmentation model is calculated. If the accuracy of the underwater acoustic spectrum segmentation model is higher than a preset value, the trained underwater acoustic spectrum segmentation model is output; otherwise, the underwater acoustic spectrum segmentation model is trained with the underwater acoustic spectrum segmentation training set, and the parameter values ​​in the underwater acoustic spectrum segmentation model are adjusted.

3. The underwater acoustic spectrum segmentation communication method based on deep learning according to claim 2 is characterized in that: The population optimization algorithm adopts a genetic algorithm, which iteratively updates the parameter simulation population set, performing selection, recombination and mutation operations during the process. The selection operation is to select a group of parameter simulation individuals with the highest fitness in the parameter simulation population set as the parent set; the recombination operation is to perform multi-point crossover transformation on any two parameter simulation individuals in the parent set; the mutation operation is to randomly select a part of the parameter simulation individuals and make numerical changes to a random position in the selected parameter simulation individuals.

4. The underwater acoustic spectrum segmentation communication method based on deep learning according to claim 3 is characterized in that: It also includes real-time training for the underwater acoustic spectrum segmentation model. The specific steps are as follows: The water acoustic spectrum segmentation real-time training data is constructed by using the water acoustic spectrum graph and the corresponding water acoustic spectrum annotation graph each time the water acoustic spectrum segmentation model is used, and a random part of the water acoustic spectrum segmentation real-time training samples are corrected, and the corrected water acoustic spectrum segmentation real-time training data is used as the water acoustic spectrum segmentation real-time training samples. When the water acoustic spectrum segmentation real-time training samples reach a preset upper limit, all the water acoustic spectrum segmentation real-time training samples are combined into a water acoustic spectrum segmentation real-time training set, and then the water acoustic spectrum segmentation model is trained in real time using the water acoustic spectrum segmentation real-time training set.

5. The underwater acoustic spectrum segmentation communication method based on deep learning according to claim 4 is characterized in that: The first feature is built based on the convolutional neural network, and the second feature extraction layer is built based on the LSTM model.

Images