Underwater sound target identification method and device and computer readable storage medium

Abstract

The invention belongs to the technical field of target recognition, and relates to an underwater acoustic target recognition method and device and a computer readable storage medium, and the method comprises the steps: obtaining a plurality of same-type underwater acoustic signal samples from data sets corresponding to different sea areas; obtaining the prediction probability that each underwater acoustic signal sample belongs to each category; performing parameterization on the prediction probability of each underwater acoustic signal sample belonging to each category, and outputting the prediction probability confidence distribution uncertainty of each underwater acoustic signal sample; the prediction probability and the prediction probability confidence distribution uncertainty of the plurality of underwater acoustic signal samples belonging to each category are fused to obtain a comprehensive basic probability assignment and a comprehensive probability distribution uncertainty, the prediction category of the plurality of underwater acoustic signal samples is output, and the value of an underwater acoustic target identification loss function is calculated; and training the convolutional neural network module and the basic probability distribution module to obtain an underwater acoustic target recognition model. The underwater acoustic target recognition model of the scheme has high recognition accuracy for underwater acoustic signals of different sea areas.

Description

Technical Field

[0001] The present invention relates to the technical field of target recognition, and in particular to an underwater acoustic target recognition method, device, and computer-readable storage medium. Background Art

[0002] Underwater acoustic target recognition refers to analyzing and processing underwater sound signals to identify whether there are target objects or phenomena in the water, which plays an important role in fields such as marine scientific research and marine resource development. Traditional and mature intelligent underwater acoustic target recognition technologies mainly focus on feature selection, feature extraction, and classifier design. By extracting features, characteristic values that can represent the essence of the target are obtained, and then classification and recognition are performed in combination with a classifier.

[0003] In the prior art, various intelligent underwater acoustic target recognition methods have been proposed. For example, the patent with publication number CN118606804A proposes an underwater acoustic target recognition method based on a fuzzy support vector center. By collecting the ship radiated noise signals in the sea area to be recognized and extracting the statistical feature of the acoustic texture slices in different dimensions, the features are input into a fuzzy support vector machine for training. The trained fuzzy support vector is used to output the fuzzy support vector, and the distance from each statistical feature to the center of the fuzzy support vector is calculated to determine the membership degree distribution function. The membership degree distribution function is used to combine the statistical features in different dimensions to generate a new training set, and the fuzzy support vector machine is trained again to obtain the optimal fuzzy support vector machine, so as to classify the new ship radiated noise signals to identify whether there are targets in the sea area to be recognized. The patent with publication number CN118131200A discloses an underwater acoustic target recognition method based on metric learning. By collecting sonar signals in the sea area to be recognized, the time-frequency spectrum features of the sonar signals are calculated, and then an underwater acoustic signal dataset is generated. A deep residual network optimized by attention is constructed to achieve efficient signal amplitude-phase feature mining. Then, a loss function based on metric learning is used to mine the key recognition features in the underwater acoustic signals, so as to achieve high-accuracy and high-generalization target recognition for the sea area to be recognized. The patent with publication number CN106250848A discloses a one-class underwater acoustic target recognition method based on multi-model fusion. The features of the underwater acoustic target are extracted by using a deep autoencoder neural network, and then the Adaboost algorithm is used for one-class classifier ensemble learning to obtain a one-class underwater acoustic target recognition model, so as to directly extract the target features in the underwater acoustic signals to determine whether there are targets in the sea area to be recognized.

[0004] Due to the significant differences in the underwater acoustic propagation environment in different sea areas, for example, in shallow sea areas, the water layer is relatively thin and the seabed topography is complex, and underwater acoustic signals are affected by reflection, scattering, and absorption during propagation. In deep sea areas, the water layer is thick, the propagation distance is long, and the sound speed changes significantly with depth, and underwater acoustic signals are prone to distortion during propagation. Therefore, underwater acoustic signals in different sea areas are very different in characteristics such as frequency, amplitude, and phase. Existing underwater acoustic target recognition methods all use historical data of the sea area to be recognized to train the underwater acoustic target recognition model. Since the trained model can only specifically capture the characteristics of the underwater acoustic signals in the sea area to be recognized and cannot adapt to the unique characteristics of data in other sea areas, the accuracy of the recognition results of the model for underwater acoustic signals in other sea areas is relatively low.

[0005] Currently, there are two ways to solve this problem. One is to re-collect the underwater acoustic signal data of the sea area before target recognition in different sea areas and conduct targeted training on the model. However, this method will increase the computational resources and recognition time cost. The other is to directly fuse the data sets of different sea areas, and input the samples in the fused data set into the model for training in sequence. However, the difference in the feature distribution of data in different sea areas will cause the update direction of the model parameters to change frequently during training, which not only increases the computational resources and time consumption, but also the model cannot fully learn the characteristics of various data, and even may occur the situation where the model cannot converge, that is, the model cannot effectively identify whether there is a target in the underwater acoustic signal.

[0006] In summary, the existing underwater acoustic target recognition methods ignore the feature differences of underwater acoustic signals in different sea areas, resulting in the problem that it is impossible to balance the recognition efficiency and recognition accuracy when conducting target recognition in different sea areas. Summary of the Invention

[0007] Therefore, the technical problem to be solved by the present invention is to overcome the problem that the existing underwater acoustic target recognition methods ignore the feature differences of underwater acoustic signals in different sea areas, resulting in the problem that it is impossible to balance the recognition efficiency and recognition accuracy when conducting target recognition in different sea areas.

[0008] To solve the above technical problem, the present invention provides an underwater acoustic target recognition method, including: Obtain multiple underwater acoustic signal samples with the same category from the data sets of different sea areas; Input each underwater acoustic signal sample into a convolutional neural network module, and output the prediction probability of each underwater acoustic signal sample belonging to each category; Input the prediction probability of each underwater acoustic signal sample belonging to each category into a basic probability assignment module, and perform parameterization operation on the prediction probability of each underwater acoustic signal sample belonging to each category by using the Dirichlet distribution, and output the uncertainty of the prediction probability confidence distribution of each underwater acoustic signal sample; The Dempster-Shafer evidence combination rule is used to fuse the prediction probabilities and the uncertainty of the prediction probability confidence distribution for each category of multiple underwater acoustic signal samples respectively, obtaining the comprehensive basic probability assignment and the comprehensive probability distribution uncertainty for each category of multiple underwater acoustic signal samples, thereby outputting the prediction categories of multiple underwater acoustic signal samples and calculating the value of the underwater acoustic target recognition loss function; Multiple underwater acoustic signal samples with the same category are retrieved from datasets in different sea areas again, and the convolutional neural network module and the basic probability assignment module are trained until the value of the underwater acoustic target recognition loss function is minimized, obtaining a trained underwater acoustic target recognition model.

[0009] Preferably, the construction process of the underwater acoustic target recognition loss function includes: Constructing the loss function for each underwater acoustic signal sample based on the prediction category and the true category of each underwater acoustic signal sample; the loss function is the mean squared error loss function, the cross-entropy loss function or the likelihood value loss function; Taking the optimization of the expectation of the loss function of each underwater acoustic signal sample with respect to its prediction probability confidence distribution as the objective, constructing the recognition loss function for each underwater acoustic signal sample; Based on the sum of the recognition loss functions of multiple underwater acoustic signal samples, the underwater acoustic target recognition loss function is obtained.

[0010] Preferably, the prediction probability confidence distribution of each underwater acoustic signal sample is expressed as: , where, represents the prediction probability confidence distribution of the th underwater acoustic signal sample; represents the prediction probability vector of the th underwater acoustic signal sample belonging to each category; represents the Dirichlet distribution parameter vector of the th underwater acoustic signal sample; represents the Dirichlet distribution normalization constant of the th underwater acoustic signal sample; represents the prediction probability that the th underwater acoustic signal sample belongs to the th category; represents the number of categories; represents the Dirichlet prediction parameter that the th underwater acoustic signal sample belongs to the th category; is the set composed of the prediction probabilities that the th underwater acoustic signal sample belongs to each category; When the loss function is the mean squared error loss function, the recognition loss function of each underwater acoustic signal sample is: , wherein, represents the recognition loss function of the th underwater acoustic signal sample; represents the mean square error loss function of the th underwater acoustic signal sample; represents the true probability vector of the th underwater acoustic signal sample belonging to each category; represents the true probability that the th underwater acoustic signal sample belongs to the th category; represents that the random variable follows a Dirichlet distribution with parameter vector ; represents the random variable; represents a Dirichlet distribution with parameter vector ; represents the expectation of the predicted probability that the th underwater acoustic signal sample belongs to the th category under the condition that the random variable follows a Dirichlet distribution with parameter vector ; represents the expectation of the square of the predicted probability that the th underwater acoustic signal sample belongs to the th category under the condition that the random variable follows a Dirichlet distribution with parameter vector ; When the loss function is the cross - entropy loss function, the underwater acoustic target recognition loss function is: , wherein, represents the cross - entropy loss function of the th underwater acoustic signal sample; When the loss function is the likelihood value loss function, the underwater acoustic target recognition loss function is: , wherein, represents the likelihood value loss function of the th underwater acoustic signal sample.

[0011] Preferably, after obtaining the underwater acoustic target recognition loss function, it further includes: Taking the minimization of the difference between the predicted probability confidence distribution of each underwater acoustic signal sample and the collapsed model distribution as the objective, constructing the KL - divergence loss of each underwater acoustic signal sample; obtaining the target KL - divergence loss function based on the sum of the KL - divergence losses of multiple underwater acoustic signal samples; Constructing an evidence collapse regularization term based on the target KL - divergence loss function and the warm - start parameter; The loss function of underwater acoustic target recognition is corrected by using the evidence collapse regular term, and then the convolutional neural network module and the basic probability assignment module are trained by using the corrected loss function of underwater acoustic target recognition.

[0012] Preferably, the corrected loss function of underwater acoustic target recognition is: , where, represents the corrected loss function of underwater acoustic target recognition; represents the loss function of underwater acoustic target recognition; represents the number of underwater acoustic signal samples; represents the warm start parameter; represents the target KL divergence loss function; represents the -th KL divergence loss of the underwater acoustic signal sample; represents the -th Dirichlet distribution parameter vector of the underwater acoustic signal sample after removing the Dirichlet prediction parameters of the categories with non-zero prediction probabilities to obtain a new Dirichlet distribution parameter vector; represents the -th prediction probability confidence distribution when the Dirichlet distribution parameter vector of the underwater acoustic signal sample is ; represents the all-ones Dirichlet distribution.

[0013] Preferably, the comprehensive basic probability assignment of multiple underwater acoustic signal samples belonging to each category is expressed as: , where, represents the comprehensive basic probability assignment of multiple underwater acoustic signal samples belonging to the -th category; represents the prediction probability that the first underwater acoustic signal sample belongs to the -th category; represents the prediction probability that the second underwater acoustic signal sample belongs to the -th category; represents the prediction probability that the N-th underwater acoustic signal sample belongs to the -th category; , represents the number of categories; The uncertainty of the comprehensive probability distribution is expressed as: , where, represents the uncertainty of the comprehensive probability distribution; represents the normalization coefficient; Represents the uncertainty of the predicted probability confidence distribution of the first underwater acoustic signal sample; Represents the uncertainty of the predicted probability confidence distribution of the second underwater acoustic signal sample; Represents the uncertainty of the predicted probability confidence distribution of the Nth underwater acoustic signal sample.

[0014] Preferably, inputting each underwater acoustic signal sample into the convolutional neural network module includes: Framing, windowing, and pre-emphasizing each underwater acoustic signal sample to obtain a target underwater acoustic signal sample; Performing short-time Fourier transform on each target underwater acoustic signal sample to obtain the spectrogram of the target underwater acoustic signal sample; Inputting the spectrogram of each target underwater acoustic signal sample into the convolutional neural network module.

[0015] Preferably, the steps of performing underwater acoustic recognition using the underwater acoustic target recognition model include: Obtaining the underwater acoustic signal to be recognized, inputting the underwater acoustic signal to be recognized into the convolutional neural network module in the underwater acoustic recognition model, and outputting the predicted probabilities of the underwater acoustic signal to be recognized belonging to various categories; Inputting the predicted probabilities of the underwater acoustic signal to be recognized belonging to various categories into the basic probability assignment module in the underwater acoustic recognition model, and performing parametric operation on the predicted probabilities of the underwater acoustic signal to be recognized belonging to various categories using the Dirichlet distribution, and outputting the uncertainty of the predicted probability confidence distribution of the underwater acoustic signal to be recognized; Based on the predicted probabilities of the underwater acoustic signal to be recognized belonging to various categories and the uncertainty of the predicted probability confidence distribution, determining whether there is a target in the underwater acoustic signal to be recognized.

[0016] The present invention also provides an underwater acoustic target recognition device, including: A data acquisition module, configured to acquire multiple underwater acoustic signal samples with the same category from datasets in different sea areas; A category probability prediction module, configured to input each underwater acoustic signal sample into the convolutional neural network module and output the predicted probabilities of each underwater acoustic signal sample belonging to various categories; An uncertainty acquisition module, configured to input the predicted probabilities of each underwater acoustic signal sample belonging to various categories into the basic probability assignment module, and perform parametric operation on the predicted probabilities of each underwater acoustic signal sample belonging to various categories using the Dirichlet distribution, and output the uncertainty of the predicted probability confidence distribution of each underwater acoustic signal sample; An evidence fusion module, which is used to fuse the prediction probabilities of multiple underwater acoustic signal samples belonging to various categories and the uncertainty of the prediction probability confidence distribution respectively by using the Dempster-Shafer evidence synthesis rule, so as to obtain the comprehensive basic probability assignment and the comprehensive probability distribution uncertainty of multiple underwater acoustic signal samples belonging to various categories, thereby outputting the prediction categories of multiple underwater acoustic signal samples and calculating the value of the underwater acoustic target recognition loss function; A model training and acquisition module, which is used to re-obtain multiple underwater acoustic signal samples of the same category from datasets in different sea areas, and train the convolutional neural network module and the basic probability assignment module until the value of the underwater acoustic target recognition loss function is the smallest, so as to obtain a trained underwater acoustic target recognition model.

[0017] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned underwater acoustic target recognition method are implemented.

[0018] The underwater acoustic target recognition method provided by this application has the following beneficial effects: 1. The prediction probabilities of each underwater acoustic signal sample of the same category from different sea areas are parameterized by using the Dirichlet distribution, and the uncertainty of the prediction probability confidence distribution of each underwater acoustic signal sample is output. The uncertainty of each underwater acoustic signal sample's prediction probability is described by using this uncertainty, considering the uncertainty that may be brought by the data characteristic differences in different sea areas; then, the prediction probabilities of each underwater acoustic signal sample belonging to various categories are used as evidence amounts, and the uncertainty of its prediction probability confidence distribution is used as the uncertainty information of the evidence amount. The Dempster-Shafer evidence synthesis rule is introduced to fuse the evidence amounts and uncertainty information of multiple underwater acoustic signal samples respectively, so as to obtain the comprehensive probability assignment and the comprehensive probability distribution uncertainty of multiple underwater acoustic signals of the same category in different sea areas belonging to various categories. Since the comprehensive probability assignment represents the prediction probability that multiple underwater acoustic signal samples belong to various categories together, and the comprehensive probability distribution uncertainty represents the prediction probability uncertainty caused by integrating the data characteristic differences in multiple sea areas, the evidence fusion strategy is used to fuse the characteristic information of underwater acoustic signal samples in different sea areas and the prediction result uncertainty caused by the characteristic information differences; finally, the prediction categories of multiple underwater acoustic signal samples are directly output based on the comprehensive probability assignment and the comprehensive probability distribution, and the model is iteratively trained by constructing a loss function, so that the model can capture the common characteristics of underwater acoustic signals of the same category in different sea areas during the training process. The model obtained after one training has high target recognition accuracy when dealing with underwater acoustic signals in different sea areas.

[0019] 2. In a conventional classification model, the predicted probability distribution output by the classification model is a function of the classification model parameters. Therefore, directly optimizing the predicted probability distribution output using the mean square error loss function, cross-entropy loss function, or likelihood value loss function can complete the optimization of the classification model parameters. In this application, a basic probability assignment module is introduced, and the output layer of the convolutional neural network is parameterized using the Dirichlet distribution. Therefore, during the training process, by optimizing the expectation of the mean square error loss function, cross-entropy loss function, or likelihood value loss function with respect to the Dirichlet distribution, an underwater acoustic target recognition loss function can be constructed, which can guide the model to focus on specific features and overall distribution rules in underwater acoustic signal samples, so that the model can fully learn the common probability distribution characteristics of underwater acoustic signal samples of the same category from different sea areas, and further improve the accuracy of the model in underwater acoustic target recognition in different sea areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to make the content of the present invention easier to be clearly understood, the following further details the present invention according to specific embodiments of the present invention in conjunction with the drawings, where: Figure 1 is a flowchart of the underwater acoustic target recognition method provided by this application; Figure 2 is a schematic diagram of the parameter dependency relationship of the underwater acoustic target recognition model provided by this application; Figure 3 is a schematic diagram of the underwater acoustic target recognition principle provided by this application; Figure 4 is a schematic diagram of the structure of the underwater acoustic target recognition device provided by this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The following further illustrates the present invention in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the embodiments cited are not intended to limit the present invention.

[0022] Please refer to Figure 1 , Figure 1 shown is a flowchart of the underwater acoustic target recognition method provided by this application, and the method specifically includes: S10: Obtain multiple underwater acoustic signal samples of the same category from datasets in different sea areas.

[0023] Specifically, multiple underwater acoustic signal samples of the same category mean that multiple underwater acoustic signal samples all belong to the category with a target or all belong to the category without a target.

[0024] S20: Input each underwater acoustic signal sample into the convolutional neural network module, and output the predicted probability that each underwater acoustic signal sample belongs to each category.

[0025] Exemplarily, the convolutional neural network module in the embodiments of the present application is a ResNet18 deep residual network. The ResNet18 deep residual network solves the problems of gradient disappearance and performance saturation in deep networks by introducing residual connections. Each residual block contains multiple convolutional layers and a skip connection. This design enables the network to maintain good feature extraction performance while deepening. At the same time, the ResNet18 deep residual network has 18 convolutional layers, which enables it to extract rich features from underwater acoustic signals. Combining the characteristics of the residual connection allowing the network to learn the identity mapping link where the gradient disappears, and finally using batch normalization to stabilize the training process and improve the convergence speed of the network. In addition, compared with other deep neural networks, the ResNet18 deep residual network is relatively lightweight and can be better applied to resource-constrained devices.

[0026] S30: Input the prediction probabilities of each underwater acoustic signal sample belonging to various categories into the basic probability assignment module, and use the Dirichlet distribution to perform parametric operations on the prediction probabilities of each underwater acoustic signal sample belonging to various categories, and output the uncertainty of the prediction probability confidence distribution of each underwater acoustic signal sample.

[0027] S40: Use the Dempster-Shafer evidence synthesis rule to fuse the prediction probabilities and the uncertainty of the prediction probability confidence distribution of multiple underwater acoustic signal samples belonging to various categories respectively, obtain the comprehensive basic probability assignments and the comprehensive probability distribution uncertainties of multiple underwater acoustic signal samples belonging to various categories, thereby output the prediction categories of multiple underwater acoustic signal samples, and calculate the value of the underwater acoustic target recognition loss function.

[0028] S50: Re-obtain multiple underwater acoustic signal samples of the same category from the datasets in different sea areas, and train the convolutional neural network module and the basic probability assignment module until the value of the underwater acoustic target recognition loss function is minimized, and obtain the trained underwater acoustic target recognition model.

[0029] Furthermore, the specific implementation steps of inputting each underwater acoustic signal sample into the convolutional neural network in step S20 include: S200: Perform frame division, windowing, and pre-emphasis processing on each underwater acoustic signal sample to obtain the target underwater acoustic signal sample.

[0030] S201: Perform short-time Fourier transform on each target underwater acoustic signal sample to obtain the spectrogram of the target underwater acoustic signal sample.

[0031] S202: Input the spectrogram of each target underwater acoustic signal sample into the convolutional neural network module.

[0032] By selecting appropriate windows and window lengths to perform preprocessing such as framing, windowing, and pre-emphasis on underwater acoustic signal samples, the data quality can be ensured; by using the short-time Fourier transform to process the preprocessed target underwater acoustic signal samples, the spectrogram of each target underwater acoustic signal sample can be obtained. Since the spectrogram contains rich time-frequency information, using the spectrogram as the input of the convolutional neural network module is beneficial for the convolutional neural network to better extract the feature information of the underwater acoustic signal samples, and thus more accurately output the prediction probabilities of their belonging to different categories.

[0033] Due to the complex and changeable marine environment, the sound propagation is severely affected by the complex marine environment. Therefore, the prediction probabilities output by the convolutional neural network often have uncertainties. In practical application scenarios, if the uncertainties of the output results cannot be scientifically evaluated, it will seriously affect the accuracy of the recognition results. Therefore, after obtaining the prediction probabilities of each underwater acoustic signal sample belonging to each category, this application assumes that the prediction probability distribution of the underwater acoustic signal samples follows the Dirichlet distribution prior, and by maximizing the probability likelihood of the observed data, outputs the posterior distribution of the prediction probabilities, so as to learn the uncertainties of the prediction probabilities for use in measuring the uncertainties of underwater acoustic target recognition. This application uses the Dirichlet distribution to parameterize the output layer of the convolutional neural network module, uses the parameterized result as the evidence quantity, and realizes the estimation of the uncertainties of the prediction probabilities, thereby effectively assisting underwater acoustic target recognition using a scientific uncertainty measurement method.

[0034] Furthermore, after obtaining the prediction probabilities of each underwater acoustic signal sample belonging to each category and the uncertainty of the prediction probability confidence distribution, this application introduces an evidence fusion strategy to fuse the evidence quantities and uncertainties of underwater acoustic signal samples from different sources, enabling the model to learn the common features of underwater acoustic signal samples of the same category from different sources, thereby improving the accuracy of the model for underwater acoustic target recognition in different sea areas.

[0035] The evidence theory is an imprecise reasoning theory, also known as the Dempster-Shafer evidence combination rule (D-S evidence theory), belonging to the category of artificial intelligence, with the ability to process uncertain information. Its main feature is that it satisfies weaker conditions than Bayesian probability theory and has the ability to directly express "uncertainty" and "unknown". By using the D-S evidence theory to fuse the prediction probabilities of underwater acoustic signal samples from different sources belonging to different categories and the uncertainty of their prediction probability confidence distributions, considering the conflict of the prediction probability distributions between different underwater acoustic signal samples, and then using the normalization factor to adjust to obtain the final comprehensive probability distribution, enabling the model to determine the prediction categories of multiple underwater acoustic signal samples based on the fused comprehensive probability distribution.

[0036] Specifically, the combined basic probability assignment of multiple underwater acoustic signal samples belonging to each category is expressed as: , wherein, represents the combined basic probability assignment that multiple underwater acoustic signal samples belong to the th category; represents the predicted probability that the first underwater acoustic signal sample belongs to the th category; represents the predicted probability that the second underwater acoustic signal sample belongs to the th category; represents the predicted probability that the Nth underwater acoustic signal sample belongs to the th category; , represents the number of categories; The uncertainty of the combined probability distribution is expressed as: , wherein, represents the uncertainty of the combined probability distribution; represents the normalization coefficient; represents the uncertainty of the confidence distribution of the predicted probability of the first underwater acoustic signal sample; represents the uncertainty of the confidence distribution of the predicted probability of the second underwater acoustic signal sample; represents the uncertainty of the confidence distribution of the predicted probability of the Nth underwater acoustic signal sample.

[0037] Taking two underwater acoustic signal samples A and B from different sea areas as an example, where the categories of the underwater acoustic signal samples include the presence of a target and the absence of a target, the combined basic probability assignment for them belonging to the category of the presence of a target is expressed as: , The combined basic probability assignment for them belonging to the category of the absence of a target is expressed as: , The uncertainty of the combined probability distribution is expressed as: , wherein, represents the predicted probability that the underwater acoustic signal sample A belongs to the category of the presence of a target; represents the predicted probability that the underwater acoustic signal sample B belongs to the category of the presence of a target; represents the predicted probability that the underwater acoustic signal sample A belongs to the category of the absence of a target; represents the predicted probability that the underwater acoustic signal sample B belongs to the category of the absence of a target; represents the uncertainty of the confidence distribution of the predicted probability of the underwater acoustic signal sample A; represents the uncertainty of the confidence distribution of the predicted probability of the underwater acoustic signal sample B; represents the normalization coefficient, used to ensure , representing the complete distribution of beliefs and uncertainties.

[0038] Furthermore, Figure 2 The following is a diagram showing the parameter dependency of the underwater acoustic recognition model provided by this application. In a conventional classification model, the predicted probability distribution output by the classification model is a function of the classification model parameters. Therefore, directly optimizing the output predicted probability distribution using the mean squared error loss function, cross-entropy loss function, or likelihood value loss function can complete the optimization of the classification model parameters. However, in this application, a basic probability assignment module is introduced, and the output layer of the convolutional neural network is parameterized using the Dirichlet distribution. Therefore, during the training process, the model parameters are optimized by optimizing the expectations of the mean squared error loss function, cross-entropy loss function, or likelihood value loss function with respect to the Dirichlet distribution. Specifically, the construction process of the underwater acoustic target recognition loss function includes: Construct the loss function for each underwater acoustic signal sample based on the predicted category and the true category of each underwater acoustic signal sample; the loss function is the mean squared error loss function, cross-entropy loss function, or likelihood value loss function.

[0039] Taking the optimization of the expectation of the loss function of each underwater acoustic signal sample with respect to its predicted probability confidence distribution as the goal, construct the recognition loss function for each underwater acoustic signal sample.

[0040] Based on the sum of the recognition loss functions of multiple underwater acoustic signal samples, obtain the underwater acoustic target recognition loss function.

[0041] Specifically, the predicted probability confidence distribution of each underwater acoustic signal sample is expressed as: , where, represents the predicted probability confidence distribution of the th underwater acoustic signal sample; represents the predicted probability vector of the th underwater acoustic signal sample belonging to each category; represents the Dirichlet distribution parameter vector of the th underwater acoustic signal sample; represents the Dirichlet distribution normalization constant of the th underwater acoustic signal sample; represents the predicted probability that the th underwater acoustic signal sample belongs to the th category; represents the number of categories; represents the Dirichlet prediction parameter that the th underwater acoustic signal sample belongs to the th category; is the The set composed of the prediction probabilities of each underwater acoustic signal sample belonging to various categories, The sum of all prediction probabilities in it is 1.

[0042] When the loss function is the mean squared error loss function, the recognition loss function of each underwater acoustic signal sample is: , where, represents the recognition loss function of the th underwater acoustic signal sample; represents the mean squared error loss function of the th underwater acoustic signal sample; represents the true probability vector of the th underwater acoustic signal sample belonging to various categories; represents the true probability that the th underwater acoustic signal sample belongs to the th category; represents that the random variable follows the Dirichlet distribution with the parameter vector ; represents the random variable; represents the Dirichlet distribution with the parameter vector ; represents that under the condition that the random variable follows the Dirichlet distribution with the parameter vector , the expectation of the prediction probability that the th underwater acoustic signal sample belongs to the th category; represents that under the condition that the random variable follows the Dirichlet distribution with the parameter vector , the expectation of the square of the prediction probability that the th underwater acoustic signal sample belongs to the th category.

[0043] When the loss function is the cross-entropy loss function, the underwater acoustic target recognition loss function is: , where, represents the cross-entropy loss function of the th underwater acoustic signal sample.

[0044] When the loss function is the likelihood value loss function, the underwater acoustic target recognition loss function is: , where, represents the likelihood value loss function of the th underwater acoustic signal sample.

[0045] Further, during the model training process, we hope that when the underwater acoustic signal samples are misrecognized, the uncertainty of the predicted probability confidence distribution is relatively high, so as to reflect that the current prediction result is unreliable. Therefore, based on the underwater acoustic target recognition loss function, this application introduces an evidence collapse regularization term. Since in the actual application scenario, the underwater acoustic signals usually clearly belong to the target presence class or the target absence class, and the collapsed model distribution is more in line with this actual situation. Therefore, after using the evidence collapse regularization term to correct the underwater acoustic target recognition loss function, it can not only fit the consistency between the predicted probability and the label, but also make the predicted probability confidence distribution output by the model close to the collapsed model distribution, so that the model outputs a more definite predicted category for the underwater acoustic signal samples, rather than outputting an ambiguous probability distribution, reducing the uncertainty of the recognition result.

[0046] Specifically, after obtaining the underwater acoustic target recognition loss function, it further includes: Taking the minimization of the difference between the predicted probability confidence distribution of each underwater acoustic signal sample and the collapsed model distribution as the goal, constructing the KL divergence loss of each underwater acoustic signal sample; obtaining the target KL divergence loss function based on the sum of the KL divergence losses of multiple underwater acoustic signal samples.

[0047] Constructing an evidence collapse regularization term based on the target KL divergence loss function and the warm start parameter.

[0048] Using the evidence collapse regularization term to correct the underwater acoustic target recognition loss function, and then using the corrected underwater acoustic target recognition loss function to train the convolutional neural network module and the basic probability assignment module.

[0049] Specifically, the corrected underwater acoustic target recognition loss function is: , where, represents the corrected underwater acoustic target recognition loss function; represents the underwater acoustic target recognition loss function; represents the number of underwater acoustic signal samples; represents the warm start parameter; represents the target KL divergence loss function; represents the KL divergence loss of the th underwater acoustic signal sample; represents the new Dirichlet distribution parameter vector obtained by removing the Dirichlet prediction parameters of the categories with non-zero predicted probabilities from the Dirichlet distribution parameter vector of the th underwater acoustic signal sample; represents the Dirichlet distribution parameter vector of the th underwater acoustic signal sample is The Dirichlet distribution representing all 1s.

[0050] As Figure 3 Shown is a schematic diagram of the underwater acoustic target recognition principle provided by this application. It specifically includes three parts: data preprocessing, deep evidence learning structure, and decision fusion strategy based on evidence theory. By preprocessing multiple underwater acoustic signal samples from different sources, obtaining their spectrograms and respectively inputting them into a convolutional neural network, the prediction probabilities of each underwater acoustic signal sample belonging to various categories are output. Then, through the basic probability assignment module, the uncertainty of the prediction probability confidence distribution of each underwater acoustic signal sample is obtained. The prediction probabilities of each underwater acoustic signal sample belonging to various categories are used as evidence amounts, and the uncertainty information of the prediction probability confidence distribution of each underwater acoustic signal sample is used as the uncertainty information of the evidence amounts. Finally, an evidence fusion strategy is introduced to fuse the evidence amounts and uncertainty information of multiple underwater acoustic signal samples respectively, obtaining comprehensive evidence amounts and uncertainty information, thereby obtaining the prediction categories of multiple underwater acoustic signal samples. Since the model continuously learns the common characteristics of underwater acoustic signals of the same category in different sea areas during the training process, the trained model has high target recognition accuracy when dealing with complex marine environments, providing new ideas and technical means for solving the problems existing in underwater target recognition and having high application value.

[0051] The embodiments of this application also provide specific experimental examples to verify the effectiveness of the above method. Specifically, actual ship radiated noise data under actual sea conditions collected 3 times in 3 different marine environments in an underwater acoustic database is obtained. Among them, the data in 3 different marine environments collected in the same batch is a dataset, and a total of three datasets, A, B, and C, are included. This application randomly selects data with the same sample number from the three datasets A, B, and C it collected. The specific data description is shown in Table 1: Table 1

[0052] Based on the above data, the embodiments of this application conduct specific target recognition experimental tests, and the specific steps are as follows: 1. Draw a dataset of underwater acoustic signal spectrograms, specifically including: 1) Split all long underwater acoustic signals into audio signals of one second each; 2) Perform short-time Fourier transform and size transformation on each audio signal to obtain the spectrogram corresponding to each audio signal. The size of the spectrogram is 496*369*3. Taking the audio signal per second as the unit, use a 256-point window for Fourier transform, and set the number of overlapping samples and the number of points for Fourier transform to 128 and 256 points respectively; 3) Label the data in the dataset with the presence or absence of targets respectively; 4) Divide the dataset to distinguish between the training set and the test set. Specifically, for the data with and without the target, 80% of the data is randomly selected as the training set, and the remaining 20% is used as the test set.

[0053] 2. Model construction set probability prediction fusion, specifically including: 1) Use the spectrogram of the underwater acoustic signal in each dataset as the input, and use the deep neural network DNN as the convolutional neural network module to output the prediction probability that each underwater acoustic signal belongs to each category, that is, the evidence vector e. Connect the evidence vector e to the parameters of the Dirichlet distribution in the basic probability assignment module through subjective logic, calculate its prediction probability confidence distribution and its uncertainty as subjective opinions, and finally use the Dempster-Shafer evidence synthesis rule to fuse the prediction probabilities and the prediction probability confidence distribution uncertainties of the underwater acoustic signals of the same category in different datasets to obtain a fused opinion; 2) Adjust the traditional mean square error loss function to the form of integration on the Dirichlet distribution to obtain the underwater acoustic target recognition loss function, merge the KL divergence term into the underwater acoustic target recognition loss function, and regulate the prediction probability distribution by punishing the evidence generated by the wrong label; 3) Obtain the value of the underwater acoustic target recognition loss function through the forward propagation of the model to measure the matching degree between the model prediction result and the actual result, continuously correct the model parameters to minimize the value of the underwater acoustic target recognition loss function, and update the weights of each module in the model through the Adam optimizer.

[0054] 3. For the 3 datasets, perform deep evidence fusion pairwise to obtain multiple fused datasets, and use each fused dataset to train and test the model respectively.

[0055] In the embodiments of the present application, the ordinary ResNet18 deep residual network and the underwater acoustic target recognition model provided by the present application are respectively used to perform recognition experiments on the data in the fused dataset. The recognition result accuracies of the two models and the uncertainty of the recognition results provided by the model of the present application are shown in Table 2: Table 2

[0056] As can be seen from the data in Table 2, for the underwater acoustic signal datasets integrated with data from different sea areas, the accuracy of the recognition results of the underwater acoustic recognition model provided by this application is higher than 98.5% for all, and the highest reaches 99.87%. However, the accuracy of the underwater acoustic target recognition results of the traditional recognition model for the merged dataset is relatively low. This indicates that the underwater acoustic target recognition method provided by this application improves the ability to scientifically evaluate the uncertainty of the model recognition results by introducing the amount of evidence and uncertainty information, and by fusing the amount of evidence and uncertainty of underwater acoustic signals from different sources, the model can fully learn the common characteristics of data of the same category in different sea areas, improving the accuracy and reliability of underwater acoustic target recognition.

[0057] Based on the underwater acoustic target recognition method provided in the above embodiments, an embodiment of this application further provides an underwater acoustic target recognition device, as Figure 4 shown. Specifically, the device includes: A data acquisition module 10, configured to obtain multiple underwater acoustic signal samples with the same category from datasets in different sea areas.

[0058] A category probability prediction module 20, configured to input each underwater acoustic signal sample into a convolutional neural network module and output the prediction probability of each underwater acoustic signal sample belonging to each category.

[0059] An uncertainty acquisition module 30, configured to input the prediction probability of each underwater acoustic signal sample belonging to each category into a basic probability assignment module, perform a parameterization operation on the prediction probability of each underwater acoustic signal sample belonging to each category using the Dirichlet distribution, and output the uncertainty of the prediction probability confidence distribution of each underwater acoustic signal sample.

[0060] An evidence fusion module 40, configured to use the Dempster-Shafer evidence combination rule to fuse the prediction probabilities and the uncertainties of the prediction probability confidence distributions of multiple underwater acoustic signal samples belonging to each category respectively, obtain the comprehensive basic probability assignment and the comprehensive probability distribution uncertainty of multiple underwater acoustic signal samples belonging to each category, thereby output the prediction categories of multiple underwater acoustic signal samples, and calculate the value of the underwater acoustic target recognition loss function.

[0061] A model training and acquisition module 50, configured to re-obtain multiple underwater acoustic signal samples with the same category from datasets in different sea areas, and train the convolutional neural network module and the basic probability assignment module until the value of the underwater acoustic target recognition loss function is the smallest, to obtain a trained underwater acoustic target recognition model.

[0062] An embodiment of this application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned underwater acoustic target recognition method are implemented.

[0063] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) that contain computer-usable program code.

[0064] The present application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0065] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0066] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0067] Obviously, the above embodiments are only examples given for clear illustration and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. And the obvious changes or modifications derived therefrom are still within the protection scope of the present invention.

Claims

1. A method for underwater acoustic target recognition, characterized in that: include: Acquire multiple samples of underwater acoustic signals of the same category from data sets in different sea areas; Input each water acoustic signal sample into the convolutional neural network module, and output the predicted probability that each water acoustic signal sample belongs to each category; The predicted probability of each water acoustic signal sample belonging to each category is input into the basic probability distribution module, and the predicted probability of each water acoustic signal sample belonging to each category is parameterized using Dirichlet distribution, and the predicted probability confidence distribution uncertainty of each water acoustic signal sample is output; The Dempster-Shafer evidence synthesis rule is used to fuse the predicted probabilities of multiple water acoustic signal samples belonging to each category and the predicted probability confidence distribution uncertainty, respectively, to obtain the comprehensive basic probability assignment and comprehensive probability distribution uncertainty of multiple water acoustic signal samples belonging to each category, thereby outputting the predicted categories of multiple water acoustic signal samples and calculating the value of the water acoustic target recognition loss function; Re-obtain multiple hydroacoustic signal samples of the same category from data sets in different sea areas, and train the convolutional neural network module and the basic probability distribution module until the value of the hydroacoustic target recognition loss function is minimized, thus obtaining a trained hydroacoustic target recognition model.

2. The underwater acoustic target recognition method according to claim 1, characterized in that: The construction process of the underwater acoustic target recognition loss function includes: A loss function is constructed for each water acoustic signal sample based on the predicted category and the true category of each water acoustic signal sample; the loss function is a mean square error loss function, a cross entropy loss function or a likelihood loss function; Taking the expected optimization of the loss function of each underwater acoustic signal sample with respect to its predicted probability confidence distribution as the goal, the recognition loss function of each underwater acoustic signal sample is constructed; The underwater acoustic target recognition loss function is obtained by summing up the recognition loss functions of multiple underwater acoustic signal samples.

3. The underwater acoustic target recognition method according to claim 2, characterized in that: The predicted probability confidence distribution of each hydroacoustic signal sample is expressed as: ， in, Indicates The predicted probability confidence distribution of the hydroacoustic signal samples; Indicates The predicted probability vector of each category of the underwater acoustic signal samples; Indicates Dirichlet distribution parameter vector of the underwater acoustic signal samples; Indicates The normalization constant of the Dirichlet distribution of the underwater acoustic signal samples; Indicates The hydroacoustic signal samples belong to The predicted probability of each category; Indicates the number of categories; Indicates The hydroacoustic signal samples belong to Dirichlet prediction parameters for each class; For the The set of predicted probabilities that each underwater acoustic signal sample belongs to each category; When the loss function is the mean square error loss function, the recognition loss function of each underwater acoustic signal sample is: ， in, Indicates The recognition loss function of the underwater acoustic signal samples; Indicates The mean square error loss function of the underwater acoustic signal samples; Indicates The true probability vector of the underwater acoustic signal samples belonging to each category; Indicates The hydroacoustic signal samples belong to The true probability of each category; It means that the random variable obeys the parameter vector The Dirichlet distribution of represents a random variable; The parameter vector is The Dirichlet distribution of Indicates that the random variable obeys the parameter vector Under the condition of Dirichlet distribution, The hydroacoustic signal samples belong to The expected predicted probability of each category; Indicates that the random variable obeys the parameter vector Under the condition of Dirichlet distribution, The hydroacoustic signal samples belong to The expectation of the square of the predicted probability of each category; When the loss function is the cross entropy loss function, the loss function for underwater acoustic target recognition is: ， in, Indicates Cross entropy loss function of water acoustic signal samples; When the loss function is the likelihood loss function, the underwater acoustic target recognition loss function is: ， in, Indicates Likelihood loss function of a sample of underwater acoustic signal.

4. The underwater acoustic target recognition method according to claim 3, characterized in that: After obtaining the underwater acoustic target recognition loss function, it also includes: The goal is to minimize the difference between the predicted probability confidence distribution of each hydroacoustic signal sample and the collapse model distribution, and construct the KL divergence loss of each hydroacoustic signal sample; the target KL divergence loss function is obtained based on the sum of the KL divergence losses of multiple hydroacoustic signal samples; Construct the evidence collapse regularization term based on the target KL divergence loss function and the hot start parameter; The evidence collapse regularization term is used to correct the loss function of underwater acoustic target recognition, and then the convolutional neural network module and the basic probability distribution module are trained using the corrected loss function of underwater acoustic target recognition.

5. The underwater acoustic target recognition method according to claim 4, characterized in that: The corrected loss function of underwater acoustic target recognition is: ， in, represents the corrected loss function for underwater acoustic target recognition; represents the loss function of underwater acoustic target recognition; Represents the number of hydroacoustic signal samples; Indicates hot start parameters; represents the target KL divergence loss function; Indicates KL divergence loss of underwater acoustic signal samples; Indicates that from The Dirichlet distribution parameter vector of underwater acoustic signal samples In , the new Dirichlet distribution parameter vector is obtained by removing the Dirichlet prediction parameters of the categories with prediction probabilities not equal to 0; Indicates The Dirichlet distribution parameter vector of the underwater acoustic signal samples is The predicted probability confidence distribution when ; represents a Dirichlet distribution with all ones.

6. The underwater acoustic target recognition method according to claim 1, characterized in that: The comprehensive basic probability assignment of multiple underwater acoustic signal samples belonging to each category is expressed as: ， in, Indicates that multiple underwater acoustic signal samples belong to The comprehensive basic probability assignment of each category; Indicates that the first underwater acoustic signal sample belongs to The predicted probability of each category; Indicates that the second underwater acoustic signal sample belongs to The predicted probability of each category; Indicates that the Nth underwater acoustic signal sample belongs to The predicted probability of each category; , Indicates the number of categories; The comprehensive probability distribution uncertainty is expressed as: ， in, Represents the uncertainty of the comprehensive probability distribution; represents the normalization coefficient; Represents the prediction probability confidence distribution uncertainty of the first hydroacoustic signal sample; Represents the prediction probability confidence distribution uncertainty of the second hydroacoustic signal sample; Represents the uncertainty of the predicted probability confidence distribution of the Nth hydroacoustic signal sample.

7. The underwater acoustic target recognition method according to claim 1, characterized in that: Inputting each hydroacoustic signal sample into the convolutional neural network module includes: Performing frame division, windowing and pre-emphasis processing on each underwater acoustic signal sample to obtain a target underwater acoustic signal sample; Performing short-time Fourier transform on each target water acoustic signal sample to obtain a spectrogram of the target water acoustic signal sample; The spectrogram of each target underwater acoustic signal sample is input into the convolutional neural network module.

8. The underwater acoustic target recognition method according to claim 1, characterized in that: The steps of using the underwater acoustic target recognition model for underwater acoustic recognition include: Acquire the water acoustic signal to be identified, input the water acoustic signal to be identified into the convolutional neural network module in the water acoustic recognition model, and output the predicted probability that the water acoustic signal to be identified belongs to each category; The predicted probability that the water acoustic signal to be identified belongs to each category is input into the basic probability distribution module in the water acoustic recognition model, and the predicted probability that the water acoustic signal to be identified belongs to each category is parameterized using Dirichlet distribution, and the predicted probability confidence distribution uncertainty of the water acoustic signal to be identified is output; Based on the predicted probability that the underwater acoustic signal to be identified belongs to each category and the uncertainty of the predicted probability confidence distribution, it is determined whether there is a target in the underwater acoustic signal to be identified.

9. An underwater acoustic target recognition device, characterized in that: include: A data acquisition module is used to acquire multiple hydroacoustic signal samples of the same category from data sets in different sea areas; A category probability prediction module is used to input each hydroacoustic signal sample into a convolutional neural network module and output the predicted probability that each hydroacoustic signal sample belongs to each category; The uncertainty acquisition module is used to input the predicted probability of each water acoustic signal sample belonging to each category into the basic probability distribution module, use Dirichlet distribution to perform parameterized operations on the predicted probability of each water acoustic signal sample belonging to each category, and output the predicted probability confidence distribution uncertainty of each water acoustic signal sample; The evidence fusion module is used to use the Dempster-Shafer evidence synthesis rule to fuse the predicted probabilities of multiple water acoustic signal samples belonging to each category and the predicted probability confidence distribution uncertainty, respectively, to obtain the comprehensive basic probability assignment and comprehensive probability distribution uncertainty of multiple water acoustic signal samples belonging to each category, thereby outputting the predicted categories of multiple water acoustic signal samples and calculating the value of the water acoustic target recognition loss function; The model training and acquisition module is used to re-acquire multiple hydroacoustic signal samples of the same category from data sets in different sea areas, train the convolutional neural network module and the basic probability distribution module until the value of the hydroacoustic target recognition loss function is minimized, and obtain a trained hydroacoustic target recognition model.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the underwater acoustic target recognition method according to any one of claims 1 to 8 are implemented.

Images