Grid processing method for sound propagation loss calculation

Abstract

The invention provides a grid processing method for sound propagation loss calculation, which comprises the following steps of: inputting parameters into a two-dimensional array in an underwater sound propagation model, and carrying out eigenvalue processing; determining a pseudo-color image corresponding to a two-dimensional array of an original sound propagation loss calculation result as a standard sound field image, and performing image similarity comparison; for two-dimensional arrays in the feature value group after sparse value extraction, the number of the two-dimensional arrays interpolated to each dimension through linear interpolation is consistent with that of the original arrays, and array similarity comparison is carried out; judging the size of the optimal grid, and determining the maximum sparse extraction value step length as the horizontal step length and the depth step length of the optimal calculation grid; and determining a training sample input into a machine learning algorithm, building a deep learning network, training the training sample of the machine learning algorithm, and outputting an optimal computational grid size. According to the technical scheme, by training and learning the calculation result of the sound propagation model, when the unknown sea area sound propagation model is calculated, the grid size is calculated, and the calculation efficiency of the sound propagation model is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of underwater sound propagation calculation, and in particular relates to a grid processing method for calculating sound propagation loss. Background Art

[0002] Sound propagation calculation models such as Bellhop and RAM are important tools for studying the propagation of various types of sound signals in the ocean. How to ensure the accuracy of their calculation results and reduce the calculation time has always been the focus of researchers. At present, there are two ways to improve the calculation efficiency of sound propagation models: first, starting from the hardware, improving computer performance, such as increasing running memory, etc.; second, modifying the source code of the sound propagation calculation model so that the sound propagation model can perform parallel calculations.

[0003] Among the input parameters of the sound propagation model, there are many parameters that will also affect the calculation time and calculation accuracy of the sound propagation calculation model, such as the number of sound rays emitted by the sound source and the calculation step in the Bellhop model; the pade series and calculation step in the RAM model. Comprehensively analyzing various sound propagation calculation models, it is found that the calculation step will affect the accuracy and calculation time of the calculation results of various sound propagation calculation models. When the calculation step is set to a small size, although the calculation result can be guaranteed to be more accurate, the influence of the computer running memory will cause the model calculation time to be longer and the real-time prediction efficiency to be reduced; when the calculation step is set to a large size, although the calculation time can be shortened, it is difficult to fully reflect the sound field characteristics in the target sea area, resulting in inaccurate calculation results and unable to provide a reference for the prediction of the sound field characteristics in the target sea area.

[0004] Therefore, how to provide a grid processing method for calculating sound propagation loss, which can realize the calculation of grid size when calculating the sound propagation model in unknown sea areas through training and learning the calculation results of the sound propagation model, and improve the calculation efficiency of the sound propagation model, has become a technical problem that needs to be solved urgently. Summary of the invention

[0005] The embodiment of the present invention provides a grid processing method for calculating sound propagation loss, which can calculate the grid size when calculating the sound propagation model in an unknown sea area by training and learning the calculation results of the sound propagation model, thereby improving the calculation efficiency of the sound propagation model.

[0006] In one embodiment of the present invention, a grid processing method for calculating sound propagation loss is provided, comprising:

[0007] S101, performing eigenvalue processing on the two-dimensional array input parameters in the underwater sound propagation model to obtain an eigenvalue group;

[0008] S102, based on the eigenvalue group of the underwater sound propagation model, taking the two-dimensional array of original sound propagation loss calculation results as the minimum step length calculation result, determining the corresponding pseudo-color image as the standard sound field image, and performing image similarity comparison;

[0009] S103, for the two-dimensional array in the eigenvalue group after sparse extraction, interpolate the number of each dimension to be consistent with the original array by linear interpolation, and compare the array similarity;

[0010] S104, determining the optimal grid size, and determining that the maximum sparse extraction step is the horizontal step dr and the depth step dz of the optimal calculation grid;

[0011] S105, forming a group of training samples based on the eigenvalue group and the grid size, performing traversal calculations, and obtaining training samples input into a machine learning algorithm;

[0012] S106. Build a deep learning network, train the training samples of the machine learning algorithm, and output an optimal computing grid size, where the optimal computing grid size includes a horizontal computing step size and a depth computing step size.

[0013] Furthermore, the two-dimensional array input parameters in the underwater sound propagation model are subjected to eigenvalue processing to obtain an eigenvalue group, including:

[0014] In the underwater sound propagation model, the input parameters include: input frequency, seabed topography undulation, maximum sea depth, sound velocity profile, three types of seabed acoustic parameters and calculation distance; wherein the input frequency, calculation distance, maximum sea depth, three types of seabed acoustic parameters including the seabed sound velocity, seabed density and seabed sound attenuation coefficient are single values;

[0015] The seabed topography and sound velocity profile are array matrices. The two-dimensional array of the seabed topography is converted into a characteristic value that can describe the topography undulation. Let σ represent the topography undulation. The first dimension of the two-dimensional array is the horizontal distance r, and the second dimension is the sea depth H corresponding to each horizontal distance. The standard deviation of the two-dimensional array is calculated as the topography undulation σ:

[0016]

[0017] Among them, N is the length of the array, H n is the sea depth corresponding to each distance, and mean(H) is the mean sea depth.

[0018] Further, the corresponding pseudo-color image is determined as a standard sound field image, and image similarity comparison is performed, including:

[0019] Given two images x and y, the structural similarity of the two images is calculated as follows

[0020]

[0021] Among them, μx and μy represent the means of image x and image y respectively, σx and σy represent the variances of image x and image y respectively, and σxy represents the covariance of image x and image y;

[0022] The SSIM method measures image similarity from three aspects: brightness, contrast, and structure:

[0023]

[0024] Among them, C1, C1 and C1 are constants, take C1 = (K1*L)2, C2 = (K2*L)2, C3 = 2C2, K1 = 0.01, K2 = 0.03, L = 255, then

[0025] SSIM(X,Y)=l(X,Y)*c(X,Y)*s(X,Y)

[0026] The SSIM method is used to obtain the similarity between the standard pseudo-color image and the pseudo-color images of different sparse value arrays. The closer the SSIM index is to 1, the higher the similarity between the two images.

[0027] Furthermore, the two-dimensional array in the eigenvalue group after sparse extraction is interpolated to the same number in each dimension as the original array by linear interpolation, and the array similarity is compared, including:

[0028] When sparse extraction is performed on the two-dimensional array in the eigenvalue group, the step size is changed until it cannot be sparse anymore; extraction, interpolation and comparison with the original array require cyclic calculation, and each calculation needs to determine whether sparse extraction can continue. If so, interpolation is performed, and then the array similarity is calculated, and the corresponding array similarity is saved; if the maximum limit is exceeded, the loop is terminated;

[0029] The matrix similarity between the two-dimensional array matrix of the sound propagation loss after each sparse extraction of the two-dimensional array in the eigenvalue group and the original two-dimensional array matrix of the sound propagation loss is calculated.

[0030] Furthermore, the optimal grid size is determined to determine the maximum sparse sampling step size as the horizontal step size dr and the depth step size dz of the optimal calculation grid, including:

[0031] Given a picture similarity threshold, compare it with the calculated picture similarity results one by one to get the sparse sampling step size corresponding to the picture similarity threshold;

[0032] Given a matrix similarity threshold, compare it with the calculated matrix similarity results one by one to obtain the sparse sampling step size corresponding to the matrix similarity threshold;

[0033] The sparse sampling step that satisfies the matrix similarity and the sparse sampling step that satisfies the image similarity are combined, and the maximum sparse sampling step that satisfies the two is the horizontal step dr and the depth step dz of the optimal calculation grid.

[0034] Furthermore, a set of training samples is formed based on the eigenvalue group and the grid size, and the training samples input into the machine learning algorithm are obtained by traversal calculation, including:

[0035] Before each cycle calculation, first determine whether the input acoustic parameter variable exceeds the preset range. If not, calculate the sound field calculation result under the minimum calculation step length and save it as a two-dimensional array, and then enter the next cycle calculation; if all four input variables exceed the preset range, the calculation ends, and the sound field calculation result training sample library is built;

[0036] Furthermore, the minimum calculation step size includes a horizontal step size and a vertical step size;

[0037] In the training samples, the parameters are ranked as follows: seabed topography undulation, maximum sea depth, maximum calculation distance, seabed sound speed, seabed acoustic attenuation coefficient, seabed density, sound speed profile type, calculation frequency and optimal grid size.

[0038] Further, a deep learning network is built, the training samples of the machine learning algorithm are trained, and the optimal computing grid size is output, the optimal computing grid size includes a horizontal computing step size and a depth computing step size, including:

[0039] Initialize the connection weights, generate a set of relatively small non-zero values, and randomly assign them to each neuron as initial weights;

[0040] Determine the topological parameters of the BP network and the number of sample training;

[0041] Determine sample pairs for input training based on the provided sample set;

[0042] Information is transmitted forward, and error information is obtained based on the network information;

[0043] The error signal is back propagated. According to the error signal, the connection weight threshold of each layer is updated to determine whether the number of training times has been reached. If it is less than the total number of samples, go to the next step, otherwise re-determine the sample pair of the input sample;

[0044] Count the total error of the BP network. If the result reaches the expected accuracy, stop. Otherwise, go to the third step and start a new round of training.

[0045] Furthermore, the training sample consists of seven input parameters and one output result, and the seven input parameters and data types are respectively:

[0046] Sound frequency: single floating point number;

[0047] Seafloor topography undulation: a single floating point number obtained by preprocessing the seafloor topography data;

[0048] Maximum sea depth: single floating point number;

[0049] Sound velocity profile type: an integer obtained by judging the sound velocity profile data;

[0050] Seabed sound speed: single float;

[0051] Seafloor density: single float;

[0052] Seabed sound attenuation coefficient: single floating point number;

[0053] Compute distance: a single floating point number.

[0054] The beneficial effects brought by the present invention are as follows:

[0055] It can be seen from the above scheme that the embodiment of the present invention provides a grid processing method for calculating sound propagation loss, and performs eigenvalue processing on the two-dimensional array input parameters in the underwater sound propagation model to obtain an eigenvalue group; determine the pseudo-color image corresponding to the two-dimensional array of the original sound propagation loss calculation result as the standard sound field image, and compare the image similarity; for the two-dimensional array in the eigenvalue group after sparse extraction, the number of interpolated to each dimension is consistent with the original array through the linear interpolation method, and compare the array similarity; judge the optimal grid size, and determine the maximum sparse extraction step size as the horizontal step size and depth step size of the optimal calculation grid; determine the training samples input to the machine learning algorithm, build a deep learning network, train the training samples of the machine learning algorithm, and output the optimal calculation grid size. The technical solution of the present invention realizes the calculation of the unknown sea area sound propagation model by training and learning the calculation results of the sound propagation model, and improves the calculation efficiency of the sound propagation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A flow chart showing a grid processing method for calculating sound propagation loss according to an embodiment of the present invention;

[0057] Figure 2 A flowchart of obtaining the sparsity of a training sample network of a grid processing method for calculating sound propagation loss according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] In the embodiment of the present invention, when calculating and predicting sound propagation in different ocean environments, the design of the sparsity of the spatial calculation grid directly affects the calculation time and calculation accuracy. At the same time, various parameter settings during the calculation will also have a certain impact on the calculation time and calculation accuracy.

[0060] If you want to obtain an optimal computational grid judgment algorithm that has a wide range of applications and is applicable to most sea areas, a lot of calculations and judgments are required. Therefore, it is necessary to generate training samples for the deep learning network, which is the main component of the algorithm, the training sample preprocessing and the learning part of the deep learning network. After the above training sample preprocessing and the learning part of the learning network are built, the optimal computational grid size can be predicted, which is the prediction part of the algorithm.

[0061] In the embodiment of the present invention, the algorithm is designed into two parts, one part is the data preprocessing part; the other part is the construction of the deep learning network, which includes two modules: learning and prediction.

[0062] like Figure 1 to Figure 2 As shown, Figure 1 A flow chart showing a grid processing method for calculating sound propagation loss according to an embodiment of the present invention; Figure 2 A flowchart of obtaining the sparsity of a training sample network of a grid processing method for calculating sound propagation loss according to an embodiment of the present invention is shown.

[0063] Figure 1 A grid processing method for calculating sound propagation loss, comprising:

[0064] S101, performing eigenvalue processing on the two-dimensional array input parameters in the underwater sound propagation model to obtain an eigenvalue group;

[0065] S102, based on the eigenvalue group of the underwater sound propagation model, taking the two-dimensional array of original sound propagation loss calculation results as the minimum step length calculation result, determining the corresponding pseudo-color image as the standard sound field image, and performing image similarity comparison;

[0066] S103, for the two-dimensional array in the eigenvalue group after sparse extraction, interpolate the number of each dimension to be consistent with the original array by linear interpolation, and compare the array similarity;

[0067] S104, determining the optimal grid size, and determining that the maximum sparse extraction step is the horizontal step dr and the depth step dz of the optimal calculation grid;

[0068] S105, forming a group of training samples based on the eigenvalue group and the grid size, performing traversal calculations, and obtaining training samples input into a machine learning algorithm;

[0069] S106. Build a deep learning network, train the training samples of the machine learning algorithm, and output an optimal computing grid size, where the optimal computing grid size includes a horizontal computing step size and a depth computing step size.

[0070] In the embodiment of the present invention, the two-dimensional array input parameters in the underwater sound propagation model are subjected to eigenvalue processing to obtain an eigenvalue group, including:

[0071] In the underwater sound propagation model, the input parameters include: input frequency, seabed topography undulation, maximum sea depth, sound velocity profile, three types of seabed acoustic parameters and calculation distance; wherein the input frequency, calculation distance, maximum sea depth, three types of seabed acoustic parameters including the seabed sound velocity, seabed density and seabed sound attenuation coefficient are single values;

[0072] The seabed topography and sound velocity profile are array matrices. The two-dimensional array of the seabed topography is converted into a characteristic value that can describe the topography undulation. Let σ represent the topography undulation. The first dimension of the two-dimensional array is the horizontal distance r, and the second dimension is the sea depth H corresponding to each horizontal distance. The standard deviation of the two-dimensional array is calculated as the topography undulation σ:

[0073]

[0074] Among them, N is the length of the array, H n is the sea depth corresponding to each distance, and mean(H) is the mean sea depth.

[0075] In the deep learning network, the input parameters cannot be matrices or arrays, but must be eigenvalues. The eight parameters to be input in this study are input frequency, seabed topography undulation, maximum sea depth, sound speed profile, three types of seabed acoustic parameters and calculated distance. Among them, the input frequency, three types of seabed acoustic parameters (seabed sound speed, seabed density and seabed sound attenuation coefficient), calculated distance and maximum sea depth are all single values. It can be considered that the above six input parameters are already in the form of eigenvalues, but the seabed topography and sound speed profile are both array matrices. Therefore, it is necessary to convert the array matrices of the two inputs, seabed topography and sound speed profile, into eigenvalues.

[0076] Training samples are needed to train the deep learning network. As mentioned above, the purpose of this study is to design a prediction algorithm that predicts the optimal grid size during calculation through eight input parameters, namely dr (horizontal calculation step) and dz (depth calculation step) in the sound propagation model. Therefore, each group (eight) of eigenvalues ​​and the corresponding dr and dz need to be determined in the training samples.

[0077] In the prediction module, although the prediction data only requires eight input parameters, the seabed topography and sound speed profile also need to be eigenvalued. The processing method is consistent with the training sample preprocessing method.

[0078] In one embodiment of the present invention, determining a corresponding pseudo-color image as a standard sound field image and performing image similarity comparison includes:

[0079] Given two images x and y, the structural similarity of the two images is calculated as follows

[0080]

[0081] Among them, μx and μy represent the means of image x and image y respectively, σx and σy represent the variances of image x and image y respectively, and σxy represents the covariance of image x and image y;

[0082] The SSIM method measures image similarity from three aspects: brightness, contrast, and structure:

[0083]

[0084] Among them, C1, C1 and C1 are constants, take C1 = (K1*L)2, C2 = (K2*L)2, C3 = 2C2, K1 = 0.01, K2 = 0.03, L = 255, then

[0085] SSIM(X,Y)=l(X,Y)*c(X,Y)*s(X,Y)

[0086] The SSIM method is used to obtain the similarity between the standard pseudo-color image and the pseudo-color images of different sparse sampling arrays. The closer the SSIM index is to 1, the higher the similarity between the two images. The image similarity between the sparse pseudo-color image of the sound field after each sparse sampling and the original pseudo-color image drawn from the original data is calculated.

[0087] In one embodiment of the present invention, the two-dimensional array in the eigenvalue group after sparse extraction is interpolated to a number in each dimension consistent with the original array by a linear interpolation method, and the array similarity is compared, including:

[0088] When sparse extraction is performed on the two-dimensional array in the eigenvalue group, the step size is changed until it cannot be sparse anymore; extraction, interpolation and comparison with the original array require cyclic calculation, and each calculation needs to determine whether sparse extraction can continue. If so, interpolation is performed, and then the array similarity is calculated, and the corresponding array similarity is saved; if the maximum limit is exceeded, the loop is terminated;

[0089] The matrix similarity between the two-dimensional array matrix of the sound propagation loss after each sparse extraction of the two-dimensional array in the eigenvalue group and the original two-dimensional array matrix of the sound propagation loss is calculated.

[0090] In another embodiment of the present invention, the optimal grid size is determined, and the maximum sparse extraction step size is determined to be the horizontal step size dr and the depth step size dz of the optimal calculation grid, including:

[0091] Given a picture similarity threshold, compare it with the calculated picture similarity results one by one to get the sparse sampling step size corresponding to the picture similarity threshold;

[0092] Given a matrix similarity threshold, compare it with the calculated matrix similarity results one by one to obtain the sparse sampling step size corresponding to the matrix similarity threshold;

[0093] The sparse sampling step that satisfies the matrix similarity and the sparse sampling step that satisfies the image similarity are combined, and the maximum sparse sampling step that satisfies the two is the horizontal step dr and the depth step dz of the optimal calculation grid.

[0094] In another embodiment of the present invention, a group of training samples is formed based on the eigenvalue group and the grid size, and the training samples input into the machine learning algorithm are obtained by traversal calculation, including:

[0095] Before each cycle calculation, first determine whether the input acoustic parameter variable exceeds the preset range. If not, calculate the sound field calculation result under the minimum calculation step length and save it as a two-dimensional array, and then enter the next cycle calculation; if all four input variables exceed the preset range, the calculation ends, and the sound field calculation result training sample library is built;

[0096] In yet another embodiment of the present invention, the minimum calculation step size includes a horizontal step size and a vertical step size;

[0097] In the training samples, the parameters are ranked as follows: seabed topography undulation, maximum sea depth, maximum calculation distance, seabed sound speed, seabed acoustic attenuation coefficient, seabed density, sound speed profile type, calculation frequency and optimal grid size.

[0098] In another embodiment of the present invention, a deep learning network is built, training samples of the machine learning algorithm are trained, and an optimal computing grid size is output, wherein the optimal computing grid size includes a horizontal computing step size and a depth computing step size, including:

[0099] Initialize the connection weights, generate a set of relatively small non-zero values, and randomly assign them to each neuron as initial weights;

[0100] Determine the topological parameters of the BP network and the number of sample training;

[0101] Determine sample pairs for input training based on the provided sample set;

[0102] Information is transmitted forward, and error information is obtained based on the network information;

[0103] The error signal is back propagated. According to the error signal, the connection weight threshold of each layer is updated to determine whether the number of training times has been reached. If it is less than the total number of samples, go to the next step, otherwise re-determine the sample pair of the input sample;

[0104] Count the total error of the BP network. If the result reaches the expected accuracy, stop. Otherwise, go to the third step and start a new round of training.

[0105] In the embodiment of the present invention, the preset range includes: whether the frequency exceeds the frequency range of 10 Hz to 10 kHz, whether the seabed topography exceeds a given longitude and latitude range, etc.

[0106] Furthermore, the training sample consists of seven input parameters and one output result, and the seven input parameters and data types are respectively:

[0107] Sound frequency: single floating point number;

[0108] Seafloor topography undulation: a single floating point number obtained by preprocessing the seafloor topography data;

[0109] Maximum sea depth: single floating point number;

[0110] Sound velocity profile type: an integer obtained by judging the sound velocity profile data;

[0111] Seabed sound speed: single float;

[0112] Seafloor density: single float;

[0113] Seabed sound attenuation coefficient: single floating point number;

[0114] Compute distance: a single floating point number.

[0115] The embodiment of the present invention provides a grid processing method for calculating sound propagation loss. The method comprises the following steps: performing eigenvalue processing on a two-dimensional array input parameter in an underwater sound propagation model to obtain an eigenvalue group; determining a pseudo-color image corresponding to the two-dimensional array of the original sound propagation loss calculation result as a standard sound field image, and performing image similarity comparison; interpolating the two-dimensional array in the eigenvalue group after sparse extraction to a number consistent with the original array in each dimension through a linear interpolation method, and performing array similarity comparison; determining an optimal grid size, and determining that the maximum sparse extraction step is the horizontal step and the depth step of the optimal calculation grid; determining a training sample for inputting a machine learning algorithm, building a deep learning network, training the training sample of the machine learning algorithm, and outputting an optimal calculation grid size.

[0116] The technical solution of the present invention, through training and learning the calculation results of the sound propagation model, realizes the calculation of the grid size when calculating the sound propagation model in unknown sea areas, thereby improving the calculation efficiency of the sound propagation model.

[0117] The above are preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A grid processing method for calculating sound propagation loss, characterized in that: The method comprises: S101, performing eigenvalue processing on the two-dimensional array input parameters in the underwater sound propagation model to obtain an eigenvalue group; S102, based on the eigenvalue group of the underwater sound propagation model, taking the two-dimensional array of original sound propagation loss calculation results as the minimum step length calculation result, determining the corresponding pseudo-color image as the standard sound field image, and performing image similarity comparison; S103, for the two-dimensional array in the eigenvalue group after sparse extraction, interpolate the number of each dimension to be consistent with the original array by linear interpolation, and compare the array similarity; S104, determining the optimal grid size, and determining that the maximum sparse extraction step is the horizontal step dr and the depth step dz of the optimal calculation grid; S105, forming a group of training samples based on the eigenvalue group and the grid size, performing traversal calculations, and obtaining training samples input into a machine learning algorithm; S106. Build a deep learning network, train the training samples of the machine learning algorithm, and output an optimal computing grid size, where the optimal computing grid size includes a horizontal computing step size and a depth computing step size.

2. A grid processing method for calculating sound propagation loss according to claim 1, characterized in that: The two-dimensional array input parameters in the underwater sound propagation model are processed by eigenvalue to obtain an eigenvalue group, including: In the underwater sound propagation model, the input parameters include: input frequency, seabed topography undulation, maximum sea depth, sound velocity profile, three types of seabed acoustic parameters and calculation distance; wherein the input frequency, calculation distance, maximum sea depth, three types of seabed acoustic parameters including the seabed sound velocity, seabed density and seabed sound attenuation coefficient are single values; The seabed topography and sound velocity profile are array matrices. The two-dimensional array of the seabed topography is converted into a characteristic value that can describe the topography undulation. Let σ represent the topography undulation. The first dimension of the two-dimensional array is the horizontal distance r, and the second dimension is the sea depth H corresponding to each horizontal distance. The standard deviation of the two-dimensional array is calculated as the topography undulation σ: Among them, N is the length of the array, H n is the sea depth corresponding to each distance, and mean(H) is the mean sea depth.

3. A grid processing method for calculating sound propagation loss according to claim 1, characterized in that: Determine the corresponding pseudo-color image as the standard sound field image and perform image similarity comparison, including: Given two images x and y, the structural similarity of the two images is calculated as follows Among them, μx and μy represent the means of image x and image y respectively, σx and σy represent the variances of image x and image y respectively, and σxy represents the covariance of image x and image y; The SSIM method measures image similarity from three aspects: brightness, contrast, and structure: Among them, C1, C1 and C1 are constants, take C1 = (K1*L)2, C2 = (K2*L)2, C3 = 2C2, K1 = 0.01, K2 = 0.03, L = 255, then SSIM(X,Y)=l(X,Y)*c(X,Y)*s(X,Y) The SSIM method is used to obtain the similarity between the standard pseudo-color image and the pseudo-color images of different sparse value arrays. The closer the SSIM index is to 1, the higher the similarity between the two images.

4. A grid processing method for calculating sound propagation loss according to claim 1, characterized in that: For the two-dimensional array in the eigenvalue group after sparse extraction, the number of interpolated values ​​in each dimension is consistent with the original array through linear interpolation, and the array similarity is compared, including: When sparse extraction is performed on the two-dimensional array in the eigenvalue group, the step size is changed until it cannot be sparse anymore; extraction, interpolation and comparison with the original array require cyclic calculation, and each calculation needs to determine whether sparse extraction can continue. If so, interpolation is performed, and then the array similarity is calculated, and the corresponding array similarity is saved; if the maximum limit is exceeded, the loop is terminated; The matrix similarity between the two-dimensional array matrix of the sound propagation loss after each sparse extraction of the two-dimensional array in the eigenvalue group and the original two-dimensional array matrix of the sound propagation loss is calculated.

5. A grid processing method for calculating sound propagation loss according to claim 1, characterized in that: The optimal grid size is determined to determine the maximum sparse extraction step size as the horizontal step size dr and the depth step size dz of the optimal calculation grid, including: Given a picture similarity threshold, compare it with the calculated picture similarity results one by one to obtain the sparse sampling step size corresponding to the picture similarity threshold; Given a matrix similarity threshold, compare it with the calculated matrix similarity results one by one to obtain the sparse sampling step size corresponding to the matrix similarity threshold; The sparse sampling step that satisfies the matrix similarity and the sparse sampling step that satisfies the image similarity are combined, and the maximum sparse sampling step that satisfies the two is the horizontal step dr and the depth step dz of the optimal calculation grid.

6. A grid processing method for calculating sound propagation loss according to claim 1, characterized in that: A set of training samples is formed based on the eigenvalue group and the grid size, and the training samples input into the machine learning algorithm are obtained by traversal calculation, including: Before each cycle calculation, first determine whether the input acoustic parameter variable exceeds the preset range. If not, calculate the sound field calculation result under the minimum calculation step and save it as a two-dimensional array, and then enter the next cycle calculation; if all four input variables exceed the preset range, the calculation ends and the sound field calculation result training sample library is completed.

7. A grid processing method for calculating sound propagation loss according to claim 6, characterized in that: The minimum calculation step length includes: a horizontal step length and a vertical step length; In the training samples, the parameters are ranked as follows: seabed topography undulation, maximum sea depth, maximum calculation distance, seabed sound speed, seabed acoustic attenuation coefficient, seabed density, sound speed profile type, calculation frequency and optimal grid size.

8. A grid processing method for calculating sound propagation loss according to claim 1, characterized in that: Build a deep learning network, train the training samples of the machine learning algorithm, and output the optimal computing grid size, which includes the horizontal computing step size and the depth computing step size, including: Initialize the connection weights, generate a set of relatively small non-zero values, and randomly assign them to each neuron as initial weights; Determine the topological parameters of the BP network and the number of sample training; Determine sample pairs for input training based on the provided sample set; Information is transmitted forward, and error information is obtained based on the network information; The error signal is back propagated. According to the error signal, the connection weight threshold of each layer is updated to determine whether the number of training times has been reached. If it is less than the total number of samples, go to the next step, otherwise re-determine the sample pair of the input sample; Count the total error of the BP network. If the result reaches the expected accuracy, stop. Otherwise, go to the third step and start a new round of training.

9. A grid processing method for calculating sound propagation loss according to claim 8, characterized in that: The training sample consists of seven input parameters and one output result. The seven input parameters and data types are: sound frequency: single floating point number; Seafloor topography undulation: a single floating point number obtained by preprocessing the seafloor topography data; Maximum sea depth: single floating point number; Sound velocity profile type: an integer obtained by judging the sound velocity profile data; Seabed sound speed: single float; Seafloor density: single float; Seabed sound attenuation coefficient: single floating point number; Compute distance: a single floating point number.

Images