A Deep Learning-Based Method for Marine Sound Source Localization and Marine Parameter Inversion

Abstract

This application discloses a deep learning-based method for ocean sound source localization and ocean parameter inversion. First, a marine environment dataset is generated using simulation software and hydrophone measurement data. Then, the dataset is used to initially train first and second residual neural network models. Next, the measured data is used to validate the initially trained first and second residual neural network models. Afterward, the sound source localization information obtained after initial training is used as input to the second residual neural network model, which predicts propagation loss. If the termination condition is not met, the predicted propagation loss is used as input to the first residual neural network model, and training continues. The output of the first residual neural network model is then used as input to the second residual neural network model, and training continues. This method reduces the error in the computational results.

Description

Technical Field

[0001] This application relates to the field of marine acoustics technology, and in particular to a deep learning-based method for marine sound source localization and marine parameter inversion, as well as a deep learning-based system for marine sound source localization and marine parameter inversion. Background Technology

[0002] Sound wave propagation in seawater is affected by parameters such as seawater density, temperature, surface and seabed reflection coefficients, seabed density, and attenuation, leading to inaccurate target identification and localization. Current methods primarily rely on direct computation of marine data acquired on-site for sound source localization, but these computations contain inherent errors. Therefore, a sound source localization method and model with minimal error is urgently needed. Summary of the Invention

[0003] Based on this, in order to reduce positioning errors and increase the accuracy of results, this application discloses the following technical solutions.

[0004] On the one hand, a deep learning-based method for ocean sound source localization and ocean parameter inversion is provided, including:

[0005] Step 100: Generate a marine environment dataset using simulation software and hydrophone measurement data;

[0006] Step 200: Initial training of the first residual neural network model and the second residual neural network model is performed using the dataset;

[0007] Step 300: Validate the first residual neural network model and the second residual neural network model that have completed initial training using the measured data;

[0008] Step 400: Use the sound source localization information obtained after the initial training as the input to the second residual neural network model;

[0009] Step 500: Predict the propagation loss using the second residual neural network model;

[0010] Step 600: If the termination condition is not met, the predicted propagation loss is used as the input of the first residual neural network model and the first residual neural network model is trained again. The output of the first residual neural network model during the process is used as the input of the second residual neural network model. Then, return to step 500 to continue training the second residual neural network model.

[0011] The termination condition includes the number of rounds of training to continue.

[0012] In one possible implementation, initial training of the first residual neural network model using the dataset includes:

[0013] The input dimension of the first residual neural network model is set to one dimension, the convolution kernel is set to one dimension, and the linear rectified function is used as the activation function.

[0014] The data range to which the sound source localization information belongs is obtained through the coarse division module of the first residual neural network model;

[0015] The interval is further subdivided by the subdivision module of the first residual neural network model, and the subdivision module is used to predict the sound source localization information.

[0016] In one possible implementation, the coarse segmentation module uses the cross-entropy loss function, the formula of which is:

[0017]

[0018] Where N is the total number of samples, n is the current number of samples, m is the number of categories, and y ic Let y be the sign function, where y is the true class of sample i equal to c. ic =1, otherwise y ic =0, p ic Let be the probability that sample i belongs to category c.

[0019] In one possible implementation, after the coarse classification module determines the data interval to which the sound source localization information belongs, the classification accuracy E of the interval is calculated using the following formula. Accu :

[0020]

[0021] Where N is the total number of samples, and Nc is the number of correctly classified samples.

[0022] In one possible implementation, initial training of the second residual neural network model using the dataset includes:

[0023] The input dimension of the second residual neural network model is set to one dimension, the convolution kernel is set to one dimension, and the linear rectified function is used as the activation function.

[0024] The second residual neural network model is trained using the dataset.

[0025] The relevant ocean parameters are predicted based on the output of the first residual neural network model.

[0026] On the other hand, a deep learning-based system for locating marine sound sources and inverting marine parameters is also provided, including:

[0027] The dataset generation module is used to generate marine environment datasets using simulation software and hydrophone measurement data;

[0028] An initial training module is used to perform initial training on the first residual neural network model and the second residual neural network model using the dataset;

[0029] The model validation module is used to validate the first residual neural network model and the second residual neural network model that have completed initial training using the measured data;

[0030] The propagation loss prediction module is used to take the sound source localization information obtained after the initial training as the input of the second residual neural network model, and predict the propagation loss through the second residual neural network model.

[0031] The iterative training module is used to continue training the first residual neural network model by taking the predicted propagation loss as the input of the first residual neural network model when the termination condition is not met, and to take the output of the first residual neural network model as the input of the second residual neural network model during the process, and to continue training the second residual neural network model through the propagation loss prediction module.

[0032] The termination condition includes the number of rounds of training to continue.

[0033] In one possible implementation, the initial training module performs initial training on the first residual neural network model through the following steps:

[0034] The input dimension of the first residual neural network model is set to one dimension, the convolution kernel is set to one dimension, and the linear rectified function is used as the activation function.

[0035] The data range to which the sound source localization information belongs is obtained through the coarse division module of the first residual neural network model;

[0036] The interval is further subdivided by the subdivision module of the first residual neural network model, and the subdivision module is used to predict the sound source localization information.

[0037] In one possible implementation, the coarse segmentation module uses the cross-entropy loss function, the formula of which is:

[0038]

[0039] Where N is the total number of samples, n is the current number of samples, m is the number of categories, and y ic Let y be the sign function, where y is the true class of sample i equal to c. ic =1, otherwise y ic =0, p icLet be the probability that sample i belongs to category c.

[0040] In one possible implementation, after the coarse classification module determines the data interval to which the sound source localization information belongs, the initial training module calculates the classification accuracy E of the interval using the following formula. Accu :

[0041]

[0042] Where N is the total number of samples, and Nc is the number of correctly classified samples.

[0043] In one possible implementation, the initial training module performs initial training on the second residual neural network model through the following steps:

[0044] The input dimension of the second residual neural network model is set to one dimension, the convolution kernel is set to one dimension, and the linear rectified function is used as the activation function.

[0045] The second residual neural network model is trained using the dataset.

[0046] The relevant ocean parameters are predicted based on the output of the first residual neural network model.

[0047] The method and system for ocean sound source localization and ocean parameter inversion based on deep learning disclosed in this application utilize deep learning neural network models for sound source localization, reducing the error of the calculation results. At the same time, this application uses a large-scale dataset to train the neural network, increasing the reliability of the calculation results. In addition, this application also uses a training method in which two neural network models are used as input and output to each other for neural network training, further increasing the reliability of sound source localization. Attached Figure Description

[0048] The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain and illustrate this application, and should not be construed as limiting the scope of protection of this application.

[0049] Figure 1 This is a flowchart illustrating an embodiment of the deep learning-based method for locating marine sound sources and inverting marine parameters disclosed in this application.

[0050] Figure 2 This is a schematic diagram of the structure of the first residual neural network model.

[0051] Figure 3 This is a structural block diagram of an embodiment of the deep learning-based marine sound source localization and marine parameter inversion system disclosed in this application. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings.

[0053] The following is for reference. Figures 1-3 This application describes in detail an embodiment of the deep learning-based method for ocean sound source localization and ocean parameter inversion. To address the problem of wide-area sound source localization, this embodiment employs a two-step localization strategy to achieve accurate positioning of the sound source target's depth and distance. For example... Figure 1 and Figure 2 As shown, the method disclosed in this embodiment includes steps 100 to 600.

[0054] Step 100: Generate a marine environment dataset using simulation software and hydrophone measurement data. A hydrophone, also known as a receiving transducer, is a transducer that converts acoustic signals into electrical signals and is used to receive acoustic signals in the water.

[0055] Marine environmental datasets can reach sizes of 10 15 This dataset, based on a normal mode underwater sound propagation model, is of immense magnitude and mainly includes: target sound source distance, sound source depth, complex sound pressure received by the hydrophone, seawater depth, seawater temperature, seawater salinity, sound source frequency, sedimentary P-wave velocity, sedimentary density, sedimentary P-wave attenuation, substrate thickness, substrate P-wave velocity, substrate density, and substrate P-wave attenuation. The complex sound pressure received by the hydrophone is the sound wave emitted from a sound source that reaches the receiving point after reflection, scattering, and absorption at the sea surface and seabed.

[0056] The target sound source distance can range from 0 to 50 kilometers, with a variation range of 1 kilometer; the sound source depth can range from 30 to 80 meters, with a variation range of 1 meter; the seawater depth can range from 200 to 300 meters, with a variation range of 2 meters; the seawater temperature can range from 5 to 25 degrees Celsius, with a variation range of 0.2 degrees Celsius; the seawater salinity can range from 34.5 to 36‰, with a variation range of 0.1‰; the sound source frequency can range from 100 to 500 Hz, with a variation range of 10 Hz; and the longitudinal wave velocity in the sedimentary layer can range from 1450 to 1550 meters per second, with a variation range of 1... The substrate thickness can range from 1 to 20 meters, with a variation range of 1 meter; the substrate P-wave velocity can range from 1600 to 1800 meters per second, with a variation range of 1 meter per second; the substrate density can range from 1.8 to 1.9 grams per cubic centimeter, with a variation range of 0.01 grams per cubic centimeter; and the substrate P-wave attenuation can be 0.05 dB / γ.

[0057] Step 200: Initial training of the first residual neural network model and the second residual neural network model is performed using the dataset.

[0058] The first residual neural network model can be a ResNet-101 neural network, and the second residual neural network model can be a ResNet-50 neural network. ResNet (Residual Network) is a residual neural network, while 101 and 50 represent the depth (number of layers). The bottleneck structure is introduced by using ResNet-101 and ResNet-50 neural networks to avoid gradient vanishing in multiple convolutional layers.

[0059] In one possible implementation, the initial training of the first residual neural network model using the dataset specifically includes the following steps A1 to A3.

[0060] Step A1: Set the input dimension of the first residual neural network model to one dimension, set the convolution kernel to one dimension, remove the pooling layer, and use the rectified linear unit (ReLU) as the activation function. The rectified linear unit (ReLU) typically refers to a nonlinear function represented by the ramp function and its variants.

[0061] Step A2: The data range to which the sound source localization information belongs is obtained through the coarse division module of the first residual neural network model.

[0062] The simulation software-generated data in the dataset serves as the training set for the neural network, while the measured data acts as the validation set. This step uses the training set from the dataset. The coarse-grained segmentation module can employ a ResNet101 network structure. Since the first residual neural network model is responsible for sound source localization, the input to the coarse-grained segmentation module consists of the complex sound pressure received by the hydrophone in the dataset, along with the seawater depth, seawater temperature, seawater salinity, sound source frequency, sedimentary P-wave velocity, sedimentary density, sedimentary P-wave attenuation, substrate thickness, substrate P-wave velocity, substrate density, and substrate P-wave attenuation. The output is the data range containing the sound source localization information, which includes the target sound source distance and sound source depth.

[0063] The ResNet-101-1 coarse-mapping module can determine the range to which the sound source localization information belongs. For example, if the overall data range for the distance to the target sound source is 0-50 kilometers, the aforementioned complex sound pressure data can be input into the ResNet-101-1 model to obtain the data range in which the distance to the target sound source is located.

[0064] Assuming the target sound source distance is predefined into five intervals: [0,10), [10,20), [20,30), [30,40), and [40,50], then [0,10) represents class 1, [10,20) class 2, [20,30) class 3, [30,40) class 4, and [40,50] class 5. The ResNet-101-1 model will output the specific interval category. The same applies to the output sound source depth; the ResNet-101-1 model outputs the interval category for sound source depth.

[0065] Understandably, the "-1" in ResNet-101-1 indicates that the model is used for coarse segmentation. The segmentation of interval categories can be equal, such as the five equally divided intervals mentioned above for the distance to the target sound source, or it can be unequal, and the number of interval categories can change depending on the data items. This step is used for coarse localization, i.e., horizontal distance localization. Therefore, multiple intervals are divided, and localization within the divided intervals is considered correct, resulting in the output of the result.

[0066] In one possible implementation, the coarse segmentation module uses the cross-entropy loss function, the formula of which is:

[0067]

[0068] Where N is the total number of samples, n is the current number of samples, m is the number of categories, and y ic Let y be the sign function, where y is the true class of sample i equal to c. ic =1, otherwise y ic =0, p ic Let be the probability that sample i belongs to category c.

[0069] The coarse-classification module of the first residual neural network model uses classification accuracy as the evaluation metric. The training process of the coarse-classification module can be carried out using the PyTorch framework, with data parallel training using 8 NVIDIA V100 GPUs. The initial learning rate can be set to 0.001, and after every 20 training epochs, the learning rate is reduced to 50% of the original. The SGD stochastic gradient descent algorithm is used as the optimizer.

[0070] In one possible implementation, after the coarse classification module determines the data interval to which the sound source localization information belongs, the classification accuracy E of the interval is calculated using the following formula. Accu :

[0071]

[0072] Where N is the total number of samples, and Nc is the number of correctly classified samples.

[0073] Classification accuracy E Accu This step is used to evaluate the performance of the coarse-classification module of the first residual neural network model. It evaluates the classification results of the coarse-classification module.

[0074] Step A3: The interval is further subdivided by the subdivision module of the first residual neural network model, and the subdivision module is used to predict the sound source localization information.

[0075] The purpose of further subdivision is to achieve precise localization within the interval predicted by the coarse subdivision module of the first residual neural network model. If the subdivision module is used alone for refined localization without first using the coarse subdivision module for rough localization, all the data needs to be used for training, which leads to high data complexity and potential interference between data, resulting in high network training costs and low localization accuracy. Therefore, this application adopts a two-step localization strategy: first, a localization interval is obtained through step A2, and then the subdivision network trained using the data of that interval is called in step A3 to reduce training difficulty and improve localization accuracy.

[0076] Ten ResNet-101-2 models can be used for precise localization. The input to the subdivision module is the complex sound pressure received by the hydrophone in the dataset, as well as the seawater depth, seawater temperature, seawater salinity, sound source frequency, sediment layer P-wave velocity, sediment layer density, sediment layer P-wave attenuation, basement thickness, basement P-wave velocity, basement density, and basement P-wave attenuation. The output is precise sound source localization information (target sound source distance and sound source depth).

[0077] Specifically, ten independent neural network models—ResNet-101-2-1-R, ResNet-101-2-1-D, ResNet-101-2-2-R, ResNet-101-2-2-D, ResNet-101-2-3-R, ResNet-101-2-3-D, ResNet-101-2-4-R, ResNet-101-2-4-D, ResNet-101-2-5-R, and ResNet-101-2-5-D—can be used to predict the sound source distance R and depth D within the subdivided intervals. In ResNet-101-2-1, "-2" indicates that this model is used for further subdivision, while "-1" indicates that the first of the five intervals is further subdivided. This step is used for precise localization, i.e., comprehensive localization of horizontal distance and depth. Performing the rough positioning in step A2 first, followed by the precise positioning in this step, can improve the accuracy of the positioning.

[0078] like Figure 2As shown, these 10 models are connected to the output of the coarse segmentation module. They have the same structure but use different training sets. For example, ResNet-101-2-1 uses data within a 0-10 km range from the hydrophone, corresponding to the first class interval divided by the coarse segmentation module; ResNet-101-2-2 uses data within a 10-20 km range from the hydrophone, corresponding to the second class interval divided by the coarse segmentation module; and so on for the other models. The training process of the sub-segmentation module can also use the PyTorch framework, employing 8 NVIDIA V100 GPUs for data parallel training. The initial learning rate can be set to 0.001, decreasing to 10% of the original rate every 20 training epochs. SGD stochastic gradient descent is used as the optimizer. Finally, the parameters of the training epochs with the minimum validation loss are selected as the final parameters for the sound source localization information prediction model.

[0079] The 10 models have the same structure, consisting of 101 layers of ResNet-101, including convolutional layer 1, 3 convolutional blocks 2, 4 convolutional blocks 3, 23 convolutional blocks 4, 3 convolutional blocks 5, and one fully connected layer. Their parameters are shown in Table 1 below.

[0080] Table 1 Structural parameters of the sub-modules

[0081]

[0082]

[0083] The formula for the width transformation of the input one-dimensional vector after convolution in the network is w = (W - K + 2P) / S + 1. Here, w is the width after convolution, W is the width before convolution, K is the kernel size, P is the padding, and S is the stride.

[0084] Sample prediction bias E dev We obtain it from the following formula:

[0085]

[0086] Where pi is the predicted parameter and gi is the actual parameter. This formula represents the loss function of the first residual neural network, and negative feedback is used to adjust the training process by decreasing the loss function during training.

[0087] In one possible implementation, the initial training of the second residual neural network model using the dataset specifically includes the following steps B1 to B3.

[0088] Step B1: Set the input dimension of the second residual neural network model to one dimension, set the convolution kernel to one dimension, remove the pooling layer, and use the linear rectified function as the activation function.

[0089] Step B2 involves training the second residual neural network model using the dataset, specifically training the ResNet-50 neural network. Since the second residual neural network model is responsible for inversion, the data items involved include all data items except for sound source localization information (target sound source distance and sound source depth).

[0090] Step B3 involves predicting relevant oceanographic parameter values ​​based on the output of the first residual neural network model. The input to the second residual neural network model includes the sound source localization information output by the first residual neural network model, as well as the complex sound pressure, seawater depth, seawater temperature, seawater salinity, sound source frequency, sedimentary P-wave velocity, sedimentary density, basement thickness, basement P-wave velocity, and basement density received by the hydrophone in the dataset. The output relevant oceanographic parameter values ​​are sedimentary P-wave attenuation and basement P-wave attenuation.

[0091] The formula for calculating the sample prediction bias of the second residual neural network model is the same as that of the first residual neural network model, and will not be repeated here.

[0092] Step 300: Validate the first residual neural network model and the second residual neural network model that have completed initial training using the measured data.

[0093] This step uses the validation set in the dataset. Since the training process in step 200 uses simulation data generated by simulation software, it is necessary to use actual test data to further validate the trained model and verify the model's localization accuracy and reliability.

[0094] Step 400: Use the sound source localization information obtained after the initial training as the input to the second residual neural network model.

[0095] Step 500 involves predicting the propagation loss using the second residual neural network model, which is essentially performing ocean parameter inversion. The inverted ocean parameters, i.e., the predicted propagation loss, specifically include sediment layer P-wave attenuation and basement P-wave attenuation. Since sound waves are the only type of energy that can propagate over long distances in seawater, studying the propagation characteristics of sound waves in seawater is a prerequisite for predicting all other underwater acoustic phenomena, and the most direct parameter of sound wave propagation characteristics is propagation loss.

[0096] Step 600: If the termination condition is not met, the predicted propagation loss is used as the input of the first residual neural network model and the first residual neural network model is trained again. The output of the first residual neural network model during the process is used as the input of the second residual neural network model, and the process returns to step 500 to continue training the second residual neural network model.

[0097] The termination condition for continued training includes the number of training epochs. Specifically, using a stochastic gradient descent optimizer, the preset number of training epochs for the model can be set to 300. The final model is selected based on the loss function value and the validation machine evaluation metric for each training epoch.

[0098] In step 500, after the second residual neural network model predicts the P-wave attenuation of the sedimentary layer and the basement, it combines these two parameters with other parameters in the dataset to form a new dataset. Then, in step 600, this new dataset serves as the new input to the first residual neural network model for further training. During this training, the sound source localization information (target sound source distance and sound source depth) output by the first residual neural network model continues to serve as the new input to the second residual neural network model. The process then returns to step 500 for further training of the second residual neural network model, until a training iteration is deemed complete at a certain point in step 600. Therefore, during this training process, the outputs of the two neural network models serve as inputs to each other and iteratively train until the aforementioned termination condition is met.

[0099] Because there is a certain error between the simulation data and the real data, the results of the first residual neural network model and the second residual neural network model cannot be 100% the same when verified with the measured data. However, after many rounds of mutual iteration between steps 500 and 600, the dataset is closer to the actual data, and the model prediction is more accurate.

[0100] The deep learning-based method for ocean sound source localization and ocean parameter inversion disclosed in this embodiment utilizes neural network deep learning for sound source localization, reducing the error of the calculation results. At the same time, this application uses a large-scale dataset to train the neural network, increasing the reliability of the calculation results. In addition, this application also uses a training method in which two neural network models are mutually input and output to train the neural network, further increasing the reliability of sound source localization.

[0101] The following is for reference. Figure 3 This application describes in detail an embodiment of a deep learning-based ocean sound source localization and ocean parameter inversion system. This embodiment is a system for implementing the aforementioned ocean sound source localization and ocean parameter inversion method embodiments.

[0102] like Figure 3 As shown, the system disclosed in this embodiment mainly includes:

[0103] The dataset generation module is used to generate marine environment datasets using simulation software and hydrophone measurement data;

[0104] An initial training module is used to perform initial training on the first residual neural network model and the second residual neural network model using the dataset;

[0105] The model validation module is used to validate the first residual neural network model and the second residual neural network model that have completed initial training using the measured data;

[0106] The propagation loss prediction module is used to take the sound source localization information obtained after the initial training as the input of the second residual neural network model, and predict the propagation loss through the second residual neural network model.

[0107] The iterative training module is used to continue training the first residual neural network model by taking the predicted propagation loss as the input of the first residual neural network model when the termination condition is not met, and to take the output of the first residual neural network model as the input of the second residual neural network model during the process, and to continue training the second residual neural network model through the propagation loss prediction module.

[0108] The termination condition includes the number of rounds of training to continue.

[0109] In one possible implementation, the initial training module performs initial training on the first residual neural network model through the following steps:

[0110] The input dimension of the first residual neural network model is set to one dimension, the convolution kernel is set to one dimension, and the linear rectified function is used as the activation function.

[0111] The data range to which the sound source localization information belongs is obtained through the coarse division module of the first residual neural network model;

[0112] The interval is further subdivided by the subdivision module of the first residual neural network model, and the subdivision module is used to predict the sound source localization information.

[0113] In one possible implementation, the coarse segmentation module uses the cross-entropy loss function, the formula of which is:

[0114]

[0115] Where N is the total number of samples, n is the current number of samples, m is the number of categories, and y ic Let y be the sign function, where y is the true class of sample i equal to c. ic =1, otherwise y ic =0, p ic Let be the probability that sample i belongs to category c.

[0116] In one possible implementation, after the coarse classification module determines the data interval to which the sound source localization information belongs, the initial training module calculates the classification accuracy E of the interval using the following formula. Accu :

[0117]

[0118] Where N is the total number of samples, and Nc is the number of correctly classified samples.

[0119] In one possible implementation, the initial training module performs initial training on the second residual neural network model through the following steps:

[0120] The input dimension of the second residual neural network model is set to one dimension, the convolution kernel is set to one dimension, and the linear rectified function is used as the activation function.

[0121] The second residual neural network model is trained using the dataset.

[0122] The relevant ocean parameters are predicted based on the output of the first residual neural network model.

[0123] The division of modules, units, or components in this document is merely a logical functional division. In actual implementation, other division methods are possible; for example, multiple modules and / or units may be combined or integrated into another system. Modules, units, and components described as separate parts may or may not be physically separate. Components shown as units may or may not be physical units; that is, they may be located in a specific place or distributed across grid cells. Therefore, some or all of the units can be selected to implement the embodiment's scheme according to actual needs.

[0124] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based method for ocean acoustic source localization and ocean parameter inversion, characterized in that, include: Step 100: Generate a marine environment dataset using simulation software and hydrophone measurement data; Step 200: Initial training of the first residual neural network model and the second residual neural network model is performed using the dataset; Initial training of the first residual neural network model using the dataset includes: The input dimension of the first residual neural network model is set to one dimension, the convolution kernel is set to one dimension, and the linear rectified function is used as the activation function. The data range to which the sound source localization information belongs is obtained through the coarse division module of the first residual neural network model; The interval is further subdivided by the subdivision module of the first residual neural network model, and the subdivision module is used to predict the sound source localization information. The coarse segmentation module uses the cross-entropy loss function, the formula of which is: ； Where N is the total number of samples, n is the current number of samples, and m is the number of categories. Let c be the sign function, when the true class of sample i is equal to c. =1, otherwise =0, Let be the probability that sample i belongs to category c; The first residual neural network model uses a ResNet-101 neural network, and the second residual neural network model uses a ResNet-50 neural network; The input to the second residual neural network model includes the sound source localization information output by the first residual neural network model, as well as the complex sound pressure, seawater depth, seawater temperature, seawater salinity, sound source frequency, sedimentary P-wave velocity, sedimentary density, basement thickness, basement P-wave velocity, and basement density received by the hydrophone in the dataset; the output ocean parameter values ​​are sedimentary P-wave attenuation and basement P-wave attenuation. Step 300: Validate the first residual neural network model and the second residual neural network model that have completed initial training using the measured data; Step 400: Use the sound source localization information obtained after the initial training as the input to the second residual neural network model; Step 500: Predict the propagation loss using the second residual neural network model; Step 600: If the termination condition is not met, the predicted propagation loss is used as the input of the first residual neural network model and the first residual neural network model is trained again. The output of the first residual neural network model during the process is used as the input of the second residual neural network model. Then, return to step 500 to continue training the second residual neural network model. The termination condition includes the number of rounds of training to continue.

2. The method for locating marine sound sources and inverting marine parameters as described in claim 1, characterized in that, After the coarse classification module determines the data range to which the sound source localization information belongs, the classification accuracy of the range is calculated using the following formula. : ； Where N is the total number of samples, This is the number of correctly classified samples.

3. The method for locating marine sound sources and inverting marine parameters as described in claim 1, characterized in that, The initial training of the second residual neural network model using the dataset includes: The input dimension of the second residual neural network model is set to one dimension, the convolution kernel is set to one dimension, and the linear rectified function is used as the activation function. The second residual neural network model is trained using the dataset. The relevant ocean parameters are predicted based on the output of the first residual neural network model.

4. A deep learning-based system for locating marine sound sources and inverting marine parameters, characterized in that, include: The dataset generation module is used to generate marine environment datasets using simulation software and hydrophone measurement data; An initial training module is used to perform initial training on the first residual neural network model and the second residual neural network model using the dataset; Initial training of the first residual neural network model using the dataset includes: The input dimension of the first residual neural network model is set to one dimension, the convolution kernel is set to one dimension, and the linear rectified function is used as the activation function. The data range to which the sound source localization information belongs is obtained through the coarse division module of the first residual neural network model; The interval is further subdivided by the subdivision module of the first residual neural network model, and the subdivision module is used to predict the sound source localization information. The coarse segmentation module uses the cross-entropy loss function, the formula of which is: ； Where N is the total number of samples, n is the current number of samples, and m is the number of categories. Let c be the sign function, when the true class of sample i is equal to c. =1, otherwise =0, Let be the probability that sample i belongs to category c; The first residual neural network model uses a ResNet-101 neural network, and the second residual neural network model uses a ResNet-50 neural network; The input to the second residual neural network model includes the sound source localization information output by the first residual neural network model, as well as the complex sound pressure, seawater depth, seawater temperature, seawater salinity, sound source frequency, sedimentary P-wave velocity, sedimentary density, basement thickness, basement P-wave velocity, and basement density received by the hydrophone in the dataset; the output ocean parameter values ​​are sedimentary P-wave attenuation and basement P-wave attenuation. The model validation module is used to validate the first residual neural network model and the second residual neural network model that have completed initial training using the measured data; The propagation loss prediction module is used to take the sound source localization information obtained after the initial training as the input of the second residual neural network model, and predict the propagation loss through the second residual neural network model. The iterative training module is used to continue training the first residual neural network model by taking the predicted propagation loss as the input of the first residual neural network model when the termination condition is not met, and to take the output of the first residual neural network model as the input of the second residual neural network model during the process, and to continue training the second residual neural network model through the propagation loss prediction module. The termination condition includes the number of rounds of training to continue.

5. The deep learning-based marine sound source localization and marine parameter inversion system as described in claim 4, characterized in that, After the coarse classification module determines the data interval to which the sound source localization information belongs, the initial training module calculates the classification accuracy of the interval using the following formula. : ； Where N is the total number of samples, This is the number of correctly classified samples.

6. The deep learning-based marine sound source localization and marine parameter inversion system as described in claim 4, characterized in that, The initial training module performs initial training on the second residual neural network model through the following steps: The input dimension of the second residual neural network model is set to one dimension, the convolution kernel is set to one dimension, and the linear rectified function is used as the activation function. The second residual neural network model is trained using the dataset. The relevant ocean parameters are predicted based on the output of the first residual neural network model.

Images