A method for determining environmental parameters based on neural network

Abstract

A method for determining environmental parameters based on a neural network, specifically a method for determining various parameters of an underwater environment based on a neural network, in order to solve the problem of low accuracy of obtaining underwater environmental parameters by using an existing LPCC feature extraction method or an MFCC feature extraction method, the invention comprises S1, collecting underwater target radiation noise and corresponding environmental parameters in each time period, and dividing the underwater target radiation noise into low-frequency signals and high-frequency signals; S2, preprocessing the collected underwater target radiation noise in each time period; S3, establishing a network model, inputting the underwater target radiation noise and corresponding environmental parameters in each time period preprocessed in S2 into the network model for training until the loss loss remains unchanged, and obtaining a trained network model; S4, collecting the underwater target radiation noise of the environmental parameters to be determined, and after preprocessing in S2, inputting the noise into the trained network model obtained in S3 to obtain the environmental parameters. It belongs to the field of environmental parameter determination.

Description

Technical Field

[0001] The invention relates to a parameter determination method, in particular to a method for determining various parameters of an underwater environment based on a neural network, and belongs to the field of environmental parameters. Background Art

[0002] The LPCC feature extraction method (LPCC for short) or the MFCC feature extraction method (MFCC for short) can be used in conjunction with a convolutional neural network to extract the features of underwater target radiation noise. When determining underwater environmental parameters, although the LPCC feature extraction method has the advantages of small computational complexity and easy implementation, its disadvantage is poor anti-noise performance. The MFCC feature extraction method can better reflect the auditory characteristics of the human ear and is superior to the LPCC feature extraction method in terms of noise resistance, but its computational complexity and computational accuracy requirements are higher than the LPCC feature extraction method, and the calculation steps are more complicated. In addition, when MFCC processes high-frequency sounds, its accuracy is not as accurate as when processing low-frequency sounds. In summary, the accuracy of obtaining underwater environmental parameters using the existing LPCC feature extraction method or MFCC feature extraction method is low. Summary of the invention

[0003] In order to solve the problem of low accuracy of obtaining underwater environmental parameters by using the existing LPCC feature extraction method or MFCC feature extraction method, the present invention further proposes a method for determining environmental parameters based on a neural network.

[0004] The technical solution adopted by the present invention is:

[0005] A method for determining environmental parameters based on a neural network comprises the following steps:

[0006] S1, collecting underwater target radiation noise and corresponding environmental parameters in each time period, wherein the environmental parameters include temperature, depth and sound speed, and dividing the underwater target radiation noise in each time period into low-frequency signals and high-frequency signals;

[0007] S2, preprocessing the underwater target radiation noise collected in each time period respectively, to obtain the preprocessed underwater target radiation noise in each time period and the corresponding environmental parameters;

[0008] S3, establish a network model, input the underwater target radiation noise and corresponding environmental parameters of each period preprocessed in S2 into the network model for training until the loss remains unchanged, and obtain a trained network model;

[0009] S4, collecting the underwater target radiation noise of the environmental parameters to be determined, and after preprocessing in S2, inputting it into the trained network model obtained in S3 to obtain the environmental parameters.

[0010] Preferably, in S2, the underwater target radiation noise collected in each time period is preprocessed respectively, and the specific process is as follows:

[0011] S21, extracting the sound feature vector of the low-frequency signal obtained in S1 by using the MFCC feature extraction method to obtain the sound feature vector matrix of the low-frequency signal;

[0012] The high frequency signal obtained in S1 is extracted with the LPCC feature extraction method to obtain the sound feature vector matrix of the high frequency signal;

[0013] S22. Map the sound feature vector matrix of the low-frequency signal and the sound feature vector matrix of the high-frequency signal obtained in S21 respectively using a mapping interpolation algorithm to obtain a mapped sound feature vector matrix of the low-frequency signal and a mapped sound feature vector matrix of the high-frequency signal, and make the mapped sound feature vector matrix of the low-frequency signal and the mapped sound feature vector matrix of the high-frequency signal equal.

[0014] Preferably, in S22, the sound feature vector matrix of the low-frequency signal and the sound feature vector matrix of the high-frequency signal obtained in S21 are mapped by using a mapping interpolation algorithm to obtain a mapped sound feature vector matrix of the low-frequency signal and a mapped sound feature vector matrix of the high-frequency signal, and the mapped sound feature vector matrix of the low-frequency signal and the mapped sound feature vector matrix of the high-frequency signal are made equal. The specific process is:

[0015] S221, the sound feature vector matrix of the low-frequency signal and the sound feature vector matrix of the high-frequency signal obtained in S2 are collectively referred to as the original matrix A, and the sound feature vector matrix of the low-frequency signal and the sound feature vector matrix of the high-frequency signal after mapping are collectively referred to as the target matrix A′;

[0016] S222. Calculate the mapping factor:

[0017]

[0018] H f represents the horizontal mapping factor;

[0019] V f represents the vertical mapping factor;

[0020] srcWidth represents the length of the original matrix A;

[0021] srcHeight represents the width of the original matrix A;

[0022] dstWidth represents the length of the target matrix A′;

[0023] dstHeight represents the width of the target matrix A′;

[0024] S223, define the coordinates of the element points in the original matrix A as P 0 (x 0 ,y 0 ), the element point coordinates P in the target matrix A′ 1 (i,j), and calculate the coordinates P of the element point in the original matrix A 0 (x 0 ,y 0 ):

[0025]

[0026] S224, scaling the original matrix A, the specific process is:

[0027] Calculate the coordinates P of the element point from the original matrix A 0 (x 0 ,y 0 ) The coordinates of the four nearest points are P 2 (x 1 ,y 1 ), P 3 (x 2 ,y 1 ), P 4 (x 1 ,y 2 ), P 5 (x 2 ,y 2 );

[0028] S225, calculate the element point coordinates P in the target matrix A' 1 The value of (i,j):

[0029] Calculate P 2 , P 3 , P 4 , P 5 The weight of

[0030]

[0031] w 1 Indicates P 2 The weight of 2 Indicates P 3 The weight of 3 Indicates P 4 The weight of 4 Indicates P 5 The weight of

[0032] P 1 (i,j)=w 1 P 2 (x 1,y 1 )+w 2 P 4 (x 2 ,y 1 )+w 3 P 3 (x 1 ,y 2 )+w 4 P 5 (x 2 ,y 2 ) (4)

[0033] S226, coordinate mapping:

[0034]

[0035] A′ ij Represents any coordinate point in the target matrix A′;

[0036] i represents the horizontal axis;

[0037] j represents the ordinate;

[0038] A′ max Represents the maximum value in the target matrix A′;

[0039] A′ min Represents the minimum value in the target matrix A′;

[0040] Obtaining a sound feature vector matrix of a mapped low-frequency signal and a sound feature vector matrix of a mapped high-frequency signal;

[0041] S227, making the sound feature vector matrix of the mapped low-frequency signal equal to the sound feature vector matrix of the mapped high-frequency signal.

[0042] Preferably, the horizontal coordinate i and the vertical coordinate j in S227 are both non-negative numbers.

[0043] Preferably, in S3, a network model is established, and the underwater target radiation noise and corresponding environmental parameters preprocessed in S2 in each period are input into the network model for training until the loss remains unchanged, thereby obtaining a trained network model. The specific process is as follows:

[0044] S31, establishing a network model, the network model includes a network model A and a network model B in sequence;

[0045] S32, inputting the equal sound feature vector matrix of the mapped low-frequency signal and the sound feature vector matrix of the mapped high-frequency signal obtained in S227 into the network model A, performing weighted combination according to the convolution weighting method of the network model A, and obtaining a weighted combined sound feature vector matrix;

[0046] S33, inputting the weighted combined sound feature vector matrix into the network model B;

[0047] S34. Repeat S32-S33 until the loss remains unchanged and a trained network model is obtained.

[0048] Preferably, the network model A includes convolution and ReLU functions.

[0049] Preferably, the network model B is a ResNet network and a LeNet network.

[0050] Preferably, in S32, the sound feature vector matrix of the mapped low-frequency signal and the sound feature vector matrix of the mapped high-frequency signal obtained in S227 are input into the network model A, and weighted combination is performed according to the convolution weighted method of the network model A to obtain the sound feature vector matrix after weighted combination. The specific process is:

[0051] S321, respectively convolve the sound feature vector matrix of the mapped low-frequency signal and the sound feature vector matrix of the mapped high-frequency signal obtained in S227 to obtain the sound feature vector matrix of the convolved low-frequency signal and the sound feature vector matrix of the convolved high-frequency signal;

[0052] S322. Use the ReLU function to perform nonlinear mapping on the sound feature vector matrix of the low-frequency signal after convolution and the sound feature vector matrix of the high-frequency signal after convolution obtained in S321, linearly add the elements at corresponding positions in the sound feature vector matrix of the low-frequency signal after nonlinear mapping and the sound feature vector matrix of the high-frequency signal to obtain the sound feature vector matrix after weighted combination.

[0053] Preferably, the size of the convolution in S321 is 1×1×3.

[0054] Beneficial effects:

[0055] The present invention performs feature processing on the collected underwater target radiation noise to obtain the corresponding environmental parameters. First, the underwater target radiation noise is divided into a low-frequency signal and a high-frequency signal. Different optimal methods are used for sound feature extraction for different audio signals. Since the MFCC feature extraction method and the LPCC feature extraction method have different characteristics and different sound frequencies, the present invention uses the advantages of the two methods to extract the sound feature vector matrix of the low-frequency signal by the MFCC feature extraction method, and to extract the sound feature vector matrix of the high-frequency signal by the LPCC feature extraction method. Then, the two extracted sound feature vector matrices are respectively mapped into sound feature vector matrices with the same number of rows and columns and elements in the vectors between 0 and 255 by using a mapping interpolation algorithm. Then, the two sound feature vector matrices are weightedly combined by using a convolution weighted combination method. The weighted combined sound feature vector matrix is ​​input into a convolution neural network for feature extraction. Finally, the corresponding environmental parameters are obtained according to the extracted features. Compared with the prior art, the features extracted by the present invention have more accurate feature representation, that is, they can better improve the accuracy of underwater environmental parameters, solve the problems of poor noise resistance when the LPCC feature extraction method is used alone to determine the underwater environmental parameters, and complex calculation steps and low accuracy in processing high-frequency sounds when the MFCC feature extraction method is used alone to determine the underwater environmental parameters. At the same time, through multiple comparative experiments, it is found that the accuracy of the environmental parameters obtained by the present invention is significantly higher than that of any of the above two methods used alone. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of the present invention;

[0057] Figure 2 It is a flow chart of the convolution weighting method;

[0058] Figure 3 is an experimental result diagram of the embodiment, wherein the horizontal axis represents the classification category and the vertical axis represents the classification accuracy; DETAILED DESCRIPTION

[0059] Specific implementation method 1: Combination Figure 1-Figure 3 The present embodiment describes a method for determining environmental parameters based on a neural network, which comprises the following steps:

[0060] S1. Collect the underwater target radiation noise and corresponding environmental parameters in each time period, where the environmental parameters include temperature, depth and sound speed, and divide the underwater target radiation noise in each time period into low-frequency signals and high-frequency signals;

[0061] First, the underwater target radiation noise and the corresponding environmental parameters of each time period are collected, and the underwater target radiation noise is feature processed to obtain the corresponding environmental parameters. The environmental parameters collected at this time are convenient for subsequent model training. The present invention divides the underwater target radiation noise collected in each time period into low-frequency signals and high-frequency signals. The boundaries of low-frequency signals and high-frequency signals can be set by yourself, but generally the middle value of the sound frequency is taken as the boundary. The signal above the middle value is a high-frequency signal, and the signal below the middle value is a low-frequency signal. For example, the sound frequency of a certain section of underwater target radiation noise is 500Hz-10000 Hz, then the middle value of this sound frequency is 4750Hz, and the signal greater than or equal to 4750Hz is a high-frequency signal, and the signal less than 4750Hz is a low-frequency signal. In this way, it is convenient to use different optimal methods for feature extraction for different signals, ensuring the accuracy of the extracted features, thereby improving the accuracy of the environmental parameters. In the application, the type or name of the underwater target and the environmental parameters of the corresponding position of the underwater target in each time period can be obtained according to the underwater target radiation noise. The underwater environmental parameters include temperature, depth, sound speed and the time corresponding to the environmental parameters.

[0062] S2. Preprocess the underwater target radiation noise collected in each period respectively. The specific process is as follows:

[0063] S21, extracting the sound feature vector of the low-frequency signal obtained in S1 by using the MFCC feature extraction method to obtain the sound feature vector matrix of the low-frequency signal;

[0064] The high frequency signal obtained in S1 is extracted with the LPCC feature extraction method to obtain the sound feature vector matrix of the high frequency signal;

[0065] Since the accuracy of the feature plays an important role in determining the environmental parameters, the present invention uses different and optimal feature extraction methods for low-frequency signals and high-frequency signals to extract them respectively, ensuring the accuracy of the extracted low-frequency signal features and the extracted high-frequency signal features. All corresponding feature vectors are integrated into a feature matrix to facilitate subsequent calculation and understanding.

[0066] S22, mapping the sound feature vector matrix of the low-frequency signal and the sound feature vector matrix of the high-frequency signal obtained in S21 respectively by using a mapping interpolation algorithm to obtain a mapped sound feature vector matrix of the low-frequency signal and a mapped sound feature vector matrix of the high-frequency signal, and making the mapped sound feature vector matrix of the low-frequency signal and the mapped sound feature vector matrix of the high-frequency signal equal, the specific process is:

[0067] S221, the sound feature vector matrix of the low-frequency signal and the sound feature vector matrix of the high-frequency signal obtained in S2 are collectively referred to as the original matrix A, and the sound feature vector matrix of the low-frequency signal and the sound feature vector matrix of the high-frequency signal after mapping are collectively referred to as the target matrix A′;

[0068] S222. Calculate the mapping factor:

[0069]

[0070] H f represents the horizontal mapping factor;

[0071] V f represents the vertical mapping factor;

[0072] srcWidth represents the length of the original matrix A;

[0073] srcHeight represents the width of the original matrix A;

[0074] dstWidth represents the length of the target matrix A′;

[0075] dstHeight represents the width of the target matrix A′;

[0076] S223, define the coordinates of the element points in the original matrix A as P 0 (x 0 ,y 0 ), the element point coordinates P in the target matrix A′ 1 (i,j), and calculate the coordinates P of the element point in the original matrix A 0 (x 0 ,y 0 ):

[0077]

[0078] S224, scaling the original matrix A, the specific process is:

[0079] Calculate the coordinates P of the element point from the original matrix A 0 (x 0 ,y 0 ) The coordinates of the four nearest points are P 2 (x 1 ,y 1 ), P 3 (x 2 ,y 1 ), P 4 (x 1 ,y 2 ), P 5 (x 2 ,y 2 );

[0080] S225, calculate the element point coordinates P in the target matrix A' 1 The value of (i,j):

[0081] Calculate P 2 , P 3 , P 4 , P 5 The weight of

[0082]

[0083] w 1 Indicates P 2 The weight of 2 Indicates P 3 The weight of 3 Indicates P 4 The weight of 4 Indicates P 5 The weight of

[0084] P 1 (i,j)=w 1 P 2 (x 1 ,y 1 )+w 2 P 3 (x 2 ,y 1 )+w 3 P 4 (x 1 ,y 2 )+w 4 P 5 (x 2 ,y 2 ) (4)

[0085] S226, coordinate mapping:

[0086]

[0087] A′ ij Represents any coordinate point in the target matrix A′;

[0088] i represents the abscissa, j represents the ordinate, and both i and j are non-negative numbers;

[0089] A′ max Represents the maximum value in the target matrix A′;

[0090] A′ min Represents the minimum value in the target matrix A′;

[0091] Obtaining a sound feature vector matrix of a mapped low-frequency signal and a sound feature vector matrix of a mapped high-frequency signal;

[0092] S227, making the sound feature vector matrix of the mapped low-frequency signal equal to the sound feature vector matrix of the mapped high-frequency signal.

[0093] This step is based on the definition of the original matrix A and the target matrix A' in S221, the mapping factor obtained in S222, and the element point coordinates P in the original matrix obtained in S223. 0 (x 0 ,y 0 ) and the element point coordinates P in the target matrix obtained in S225 1 (i,j) Calculate the coordinate mapping between the original matrix A and the target matrix A′ in S226. The existing convolutional neural network structure has a very good processing effect on the image vector with a fixed input size. Inspired by this, the present invention uses a mapping interpolation algorithm to map the audio feature vector to a size that is most suitable for the convolutional neural network input, which not only makes the feature accuracy higher when performing feature extraction, but also can better combine the convolution weighting method with the convolutional neural network. The purpose of the mapping interpolation algorithm is to map the sound feature vector matrix of the low-frequency signal and the sound feature vector matrix of the high-frequency signal obtained in S226 to a sound feature vector matrix with the same number of rows and columns, and the elements in the vector are all between 0-255.

[0094] S3, establish a network model, input the underwater target radiation noise and corresponding environmental parameters in each period preprocessed in S2 into the network model for training until the loss remains unchanged, and obtain a trained network model. The specific process is:

[0095] S31, establishing a network model, the network model includes a network model A and a network model B in sequence;

[0096] S32, input the equal sound feature vector matrix of the mapped low-frequency signal and the sound feature vector matrix of the mapped high-frequency signal obtained in S227 into the network model A, perform weighted combination according to the convolution weighted method of the network model A, and obtain the sound feature vector matrix after weighted combination, the specific process is:

[0097] The network model A includes convolution and ReLU functions;

[0098] S321, respectively convolve the sound feature vector matrix of the mapped low-frequency signal and the sound feature vector matrix of the mapped high-frequency signal obtained in S227, the size of the convolution is 1×1×3, to obtain the sound feature vector matrix of the convolved low-frequency signal and the sound feature vector matrix of the convolved high-frequency signal;

[0099] S322. Use the ReLU function to perform nonlinear mapping on the sound feature vector matrix of the low-frequency signal after convolution and the sound feature vector matrix of the high-frequency signal after convolution obtained in S321, linearly add the elements at corresponding positions in the sound feature vector matrix of the low-frequency signal after nonlinear mapping and the sound feature vector matrix of the high-frequency signal to obtain the sound feature vector matrix after weighted combination.

[0100] S33, inputting the weighted combined sound feature vector matrix into the network model B, and outputting the accuracy of each environmental parameter;

[0101] Network model B is ResNet network and LeNet network.

[0102] S34. Repeat S32-S33 until the loss remains unchanged and a trained network model is obtained.

[0103] Since the characteristics of MFCC and LPCC are different for each other and for sound frequency, MFCC can well represent the sound characteristics of low-frequency signals for sound frequency, while LPCC can more accurately represent the sound characteristics of high-frequency signals. That is, the present invention uses MFCC to represent the sound characteristics of low-frequency signals, and LPCC to represent the sound characteristics of high-frequency signals. However, since underwater sound signals are complex and changeable, some high-frequency signals of sound account for the main part, while some low-frequency signals of sound account for the main part. At this time, how to allocate the weights of the two to combine and input them into the convolutional neural network for feature extraction is a question worth thinking about. If the initial weights are only randomly assigned for linear superposition, the prior art cannot accurately calculate the importance of the high-frequency signal part and the low-frequency signal part, and the underwater sound signal is nonlinear. If it is only linearly superposed, the accuracy of feature extraction will not be very high, so the present invention proposes a convolution weighting method as above. This step belongs to the preprocessing step of determining environmental parameters. The sound is converted into a feature vector using the convolution weighting method as above, and then the feature vector is used as the input of the existing convolutional neural network to achieve the purpose of classification. The weighted combined feature vector matrix is ​​input into the convolutional neural network for training. The input is the feature vector after convolution obtained in S321, and the output is the accuracy of each environmental parameter. The existing convolutional neural network needs to be retrained for different inputs to ensure that the loss loss is continuously reduced during the training process, so that the optimal classification accuracy can be obtained in the end. Until the loss loss remains unchanged, a trained network model is obtained to determine the underwater environmental parameters.

[0104] S4, collecting the underwater target radiation noise of the environmental parameters to be determined, and after preprocessing in S2, inputting it into the trained network model obtained in S3 to obtain the environmental parameters.

[0105] Example:

[0106] S1. Collect the underwater target radiation noise and corresponding environmental parameters in each time period, where the environmental parameters include temperature, depth, sound speed, and time corresponding to the environmental parameters, and divide the underwater target radiation noise in each time period into low-frequency signals and high-frequency signals;

[0107] S2, the sound feature vector of the low-frequency signal obtained in S1 is extracted by using the MFCC feature extraction method, and the sound feature vector matrix of the low-frequency signal is obtained. For the convenience of understanding in subsequent calculations, it is referred to as the MFCC feature matrix. The sound feature vector of the high-frequency signal obtained in S1 is extracted by using the LPCC feature extraction method, and the sound feature vector matrix of the high-frequency signal is obtained. Similarly, it is referred to as the LPCC feature matrix. Since the principles of the MFCC feature extraction method and the LPCC feature extraction method are different, the number of rows and columns of the MFCC feature matrix and the LPCC feature matrix are different.

[0108] S3, mapping the MFCC feature matrix and the LPCC feature matrix obtained in S2 respectively by using a mapping interpolation algorithm to obtain a mapped MFCC feature matrix and a LPCC feature matrix, so that the mapped MFCC feature matrix and the LPCC feature matrix have the same number of rows and columns, and the elements in the vector are all between 0 and 255;

[0109] S31. Calculate the mapping factor:

[0110] The mapping factors include the horizontal mapping factor H f and the vertical mapping factor V f , according to the size of the MFCC feature matrix and the MFCC target matrix, the specific values ​​of the horizontal mapping factor of the MFCC feature matrix and the vertical mapping factor of the MFCC feature matrix are calculated; at the same time, according to the size of the LPCC feature matrix and the LPCC target matrix, the specific values ​​of the horizontal mapping factor of the LPCC feature matrix and the vertical mapping factor of the LPCC feature matrix are calculated. Since the present invention includes two feature matrices (MFCC feature matrix and LPCC feature matrix), there is a set of mapping factors for each feature matrix, but since the calculation process of each feature matrix is ​​consistent, and in order to enable those skilled in the art to understand the present invention more clearly, the MFCC feature matrix and the LPCC feature matrix are collectively referred to as the original matrix A, and the horizontal mapping factor H is used uniformly. f and the vertical mapping factor V f Represents the mapping factor corresponding to each feature matrix, and defines that the coordinates of the element points in the above original matrix A (MFCC feature matrix and LPCC feature matrix) are all P 0 (x 0 ,y 0), the two feature matrices after the mapping interpolation algorithm are collectively referred to as the target matrix A′, that is, the target matrix A′ includes the MFCC target matrix and the LPCC target matrix, and the corresponding element point coordinates in the target matrix A′ are defined as P 1 (i,j), the mapping formula is as follows (6):

[0111]

[0112] Among them, srcWidth represents the length of the original matrix A, srcHeight represents the width of the original matrix A, dstWidth represents the length of the target matrix A′, dstHeight represents the width of the target matrix A′, and the length and width of each target matrix A′ can be defined as specific values ​​that are most suitable for the subsequent network structure.

[0113] S32, calculate the target point in the original matrix A:

[0114] Since the coordinates of the element points in the target matrix A′ are likely not integers after being mapped back to the original matrix A, it is necessary to first calculate the coordinates of the element points P in the target matrix A′. 1 (i,j) corresponds to the coordinates of the element point P in the original matrix A 0 (x 0 ,y 0 ), the calculation method is as follows:

[0115]

[0116] S33, scaling the original matrix A, the specific process is:

[0117] Calculate the coordinates P of the element point from the original matrix A 0 (x 0 ,y 0 ) The coordinates of the four nearest points are P 2 (x 1 ,y 1 ), P 3 (x 2 ,y 1 ), P 4 (x 1 ,y 2 ), P 5 (x 2 ,y 2 ).

[0118] S34, calculate the element point coordinates P in the target matrix A' 1 The value of (i,j):

[0119] Calculate P 2 , P 3 , P 4 , P5 The weight of

[0120]

[0121] w 1 Indicates P 2 The weight of 2 Indicates P 3 The weight of 3 Indicates P 4 The weight of 4 Indicates P 5 The weight of

[0122] P 1 (i,j)=w 1 P 2 (x 1 ,y 1 )+w 2 P 3 (x 2 ,y 1 )+w 3 P 4 (x 1 ,y 2 )+w 4 P 5 (x 2 ,y 2 ) (9)

[0123] S35, coordinate mapping:

[0124] According to the above steps S31-S34, let the scaled MFCC feature matrix be M, M becomes the MFCC target matrix M' after coordinate mapping, and the scaled LPCC feature matrix be L, L becomes the LPCC target matrix L' after coordinate mapping. Since the purpose of the mapping interpolation algorithm is to map M and L to feature matrices with the same number of rows and columns, and the elements in the vector are all between 0 and 255, so let the mapped M' and L' be equal, set m as the length of the target matrix, n as the width of the target matrix, then any coordinate point M' in M' ij As shown in formula (10), any coordinate point L′ in L′ ij As shown in formula (11):

[0125]

[0126]

[0127] Where, 0≤i≤m, 0≤j≤n, i represents the horizontal coordinate; j represents the vertical coordinate; M′ max represents the maximum value in M′, M′ min represents the minimum value in M′; L′ maxrepresents the maximum value in L′, L′ min represents the minimum value in L′.

[0128] S4, performing weighted combination of the MFCC target matrix and the LPCC target matrix obtained in S3 using a convolution weighted method to obtain a weighted combined feature matrix;

[0129] Due to the respective characteristics of MFCC and LPCC and the differences in sound frequency, MFCC can well represent the low-frequency characteristics of sound, while LPCC can more accurately represent the high-frequency characteristics of sound, that is, the present invention utilizes the above advantages of MFCC and LPCC to extract different sound frequencies respectively, which can ensure the accuracy of sound characteristics and provide a basis for accurately obtaining various environmental parameters. Since underwater sound signals are complex and changeable, some high-frequency sound signals account for the main part, while some low-frequency sound signals account for the main part. At this time, how to allocate the weights of the two to combine and input them into the convolutional neural network for feature extraction is a question worth considering. If the initial weights are only randomly assigned for linear superposition, the prior art cannot accurately calculate the importance of the high-frequency signal part and the low-frequency signal part, and the underwater sound signal is nonlinear. If it is only linearly superposed, the accuracy of feature extraction will definitely not be particularly high. Therefore, in response to the above problems, the present invention proposes a convolution weighting method, firstly, the MFCC target matrix M′ processed by the mapping interpolation algorithm and the LPCC target matrix L′ processed by the mapping interpolation algorithm are respectively subjected to convolutions of different sizes of 1×1×3 to obtain the convolved MFCC target matrix M′ and LPCC target matrix L′, and then a ReLU function is applied to each of them for nonlinear mapping, and the elements at corresponding positions in the MFCC target matrix M′ and the LPCC target matrix after nonlinear mapping are linearly added to obtain the sound feature vector matrix after weighted combination. The ReLU function enables the feature combination to have the characteristics of nonlinear combination. The two 1×1×3 convolutions can also represent the initial weights on different channels of MFCC and LPCC, and the final weights on different channels can be finally determined from the back propagation process of the convolutional neural network to achieve the purpose of nonlinear combination of features.

[0130] S5, input the weighted combined feature matrix obtained in S4 into the trained convolutional neural network, and output the accuracy of each environmental parameter. Determine the underwater environmental parameters according to the accuracy of each environmental parameter.

[0131] Among them, the two convolution operations S4 and S5 not only increase the number of channels of the feature matrix but also do not change the length and width of the feature matrix. It can also be understood as allocating weights from different channels to achieve better feature representation. The weighted combined feature matrix obtained in S4 is compared with the separately classified LPCC feature matrix and the separately classified MFCC feature matrix. The method steps are as follows:

[0132] Step 1: Input the weighted combined feature matrix obtained in S4 into a convolutional neural network and output the classification accuracy;

[0133] Step 2: Input the separately classified LPCC feature matrix into the convolutional neural network and output the classification accuracy;

[0134] Step 3: Input the separately classified MFCC feature matrix into the convolutional neural network and output the classification accuracy;

[0135] Step 4: Compare the accuracy of each environmental parameter outputted in step 1 with that in step 2 and step 3 respectively.

[0136] In order to determine the advantages of the present invention, multiple convolutional neural networks with different structures, such as ResNet network and LeNet network, were used for comparison. Although the convolutional neural network through which the three features pass each time is the same, after the above-mentioned multiple comparisons, the accuracy of each environmental parameter obtained by the weighted combination of the feature matrix of the present invention is significantly higher than the accuracy of each environmental parameter obtained by the separately classified LPCC feature matrix and the accuracy of each environmental parameter obtained by the separately classified MFCC feature matrix.

Claims

1. A method for determining environmental parameters based on a neural network, characterized in that: It includes the following steps: S1. Collect the underwater target radiation noise and corresponding environmental parameters in each time period, where the environmental parameters include temperature, depth and sound speed, and divide the underwater target radiation noise in each time period into low-frequency signals and high-frequency signals; S2, preprocessing the underwater target radiation noise collected in each time period respectively, to obtain the preprocessed underwater target radiation noise in each time period and the corresponding environmental parameters; S3, establish a network model, input the underwater target radiation noise and corresponding environmental parameters of each period preprocessed in S2 into the network model for training until the loss remains unchanged, and obtain a trained network model; S4, collecting the underwater target radiation noise of the environmental parameters to be determined, and after preprocessing in S2, inputting it into the trained network model obtained in S3 to obtain the environmental parameters; In S2, the underwater target radiation noise collected in each time period is preprocessed respectively, and the specific process is as follows: S21, extracting the sound feature vector of the low-frequency signal obtained in S1 by using the MFCC feature extraction method to obtain the sound feature vector matrix of the low-frequency signal; The high frequency signal obtained in S1 is extracted with the LPCC feature extraction method to obtain the sound feature vector matrix of the high frequency signal; S22. Map the sound feature vector matrix of the low-frequency signal and the sound feature vector matrix of the high-frequency signal obtained in S21 respectively using a mapping interpolation algorithm to obtain a mapped sound feature vector matrix of the low-frequency signal and a mapped sound feature vector matrix of the high-frequency signal, and make the mapped sound feature vector matrix of the low-frequency signal and the mapped sound feature vector matrix of the high-frequency signal equal.

2. A method for determining environmental parameters based on a neural network according to claim 1, characterized in that: In S22, the sound feature vector matrix of the low-frequency signal and the sound feature vector matrix of the high-frequency signal obtained in S21 are mapped by using a mapping interpolation algorithm to obtain a mapped sound feature vector matrix of the low-frequency signal and a mapped sound feature vector matrix of the high-frequency signal, and the mapped sound feature vector matrix of the low-frequency signal and the mapped sound feature vector matrix of the high-frequency signal are made equal. The specific process is as follows: S221, the sound feature vector matrix of the low-frequency signal and the sound feature vector matrix of the high-frequency signal obtained in S2 are collectively referred to as the original matrix A, and the sound feature vector matrix of the low-frequency signal and the sound feature vector matrix of the high-frequency signal after mapping are collectively referred to as the target matrix A′; S222. Calculate the mapping factor: H f represents the horizontal mapping factor; V f represents the vertical mapping factor; srcWidth represents the length of the original matrix A; srcHeight represents the width of the original matrix A; dstWidth represents the length of the target matrix A′; dstHeight represents the width of the target matrix A′; S223. Define the coordinates of the element point in the original matrix A as P0(x0,y0), the coordinates of the element point in the target matrix A′ as P1(i,j), and calculate the coordinates of the element point P0(x0,y0) in the original matrix A: S224, scaling the original matrix A, the specific process is: Calculate the coordinates of the four points closest to the element point coordinate P0(x0,y0) of the original matrix A, which are P2(x1,y1), P3(x2,y1), P4(x1,y2), and P5(x2,y2); S225. Calculate the value of the element point coordinate P1(i,j) in the target matrix A′: Calculate the weights of P2, P3, P4, and P5 w1 represents the weight of P2; w2 represents the weight of P3; w3 represents the weight of P4; w4 represents the weight of P5; P1(i,j)=w1P2(x1,y1)+w2P4(x2,y1)+w3P3(x1,y2)+w4P5(x2,y2) (4) S226, coordinate mapping: A i ' j Represents any coordinate point in the target matrix A′; i represents the horizontal axis; j represents the ordinate; A′ max Represents the maximum value in the target matrix A′; A′ min Represents the minimum value in the target matrix A′; Obtaining a sound feature vector matrix of a mapped low-frequency signal and a sound feature vector matrix of a mapped high-frequency signal; S227, making the sound feature vector matrix of the mapped low-frequency signal equal to the sound feature vector matrix of the mapped high-frequency signal.

3. A method for determining environmental parameters based on a neural network according to claim 2, characterized in that: In S227, both the horizontal coordinate i and the vertical coordinate j are non-negative numbers.

4. The method for determining environmental parameters based on a neural network according to claim 2, characterized in that: In S3, a network model is established, and the underwater target radiation noise and corresponding environmental parameters preprocessed in S2 at each time period are input into the network model for training until the loss remains unchanged, thereby obtaining a trained network model. The specific process is as follows: S31, establishing a network model, the network model includes a network model A and a network model B in sequence; S32, inputting the equal sound feature vector matrix of the mapped low-frequency signal and the sound feature vector matrix of the mapped high-frequency signal obtained in S227 into the network model A, performing weighted combination according to the convolution weighting method of the network model A, and obtaining a weighted combined sound feature vector matrix; S33, inputting the weighted combined sound feature vector matrix into the network model B; S34. Repeat S32-S33 until the loss remains unchanged and a trained network model is obtained.

5. A method for determining environmental parameters based on a neural network according to claim 4, characterized in that: The network model A includes convolution and ReLU functions.

6. A method for determining environmental parameters based on a neural network according to claim 4, characterized in that: Network model B is ResNet network and LeNet network.

7. A method for determining environmental parameters based on a neural network according to claim 5 or 6, characterized in that: In the S32, the sound feature vector matrix of the mapped low-frequency signal and the sound feature vector matrix of the mapped high-frequency signal obtained in S227 are input into the network model A, and weighted combination is performed according to the convolution weighting method of the network model A to obtain the sound feature vector matrix after weighted combination. The specific process is as follows: S321, respectively convolve the sound feature vector matrix of the mapped low-frequency signal and the sound feature vector matrix of the mapped high-frequency signal obtained in S227 to obtain the sound feature vector matrix of the convolved low-frequency signal and the sound feature vector matrix of the convolved high-frequency signal; S322. Use the ReLU function to perform nonlinear mapping on the sound feature vector matrix of the low-frequency signal after convolution and the sound feature vector matrix of the high-frequency signal after convolution obtained in S321, linearly add the elements at corresponding positions in the sound feature vector matrix of the low-frequency signal after nonlinear mapping and the sound feature vector matrix of the high-frequency signal to obtain the sound feature vector matrix after weighted combination.

8. A method for determining environmental parameters based on a neural network according to claim 7, characterized in that: The size of the convolution in S321 is 1×1×3.

Images