A method for predicting noise characteristics based on LM-BP neural network model

Abstract

The present invention provides a method for predicting noise characteristics based on an LM-BP neural network model, comprising the following steps: S1. Collecting data: seawater temperature T, seawater salinity S, seawater depth z, sea surface wind speed v, sea surface rainfall rate r, and marine environmental noise data; S2. Creating a data set including the above five types of data and marine environmental noise data; S3. Using the five types of data collected in S1 as input, designing the topological structure of the LM-BP neural network model; S4. Training the LM-BP neural network model; S5. Predicting noise characteristics based on the LM-BP neural network model: inputting five types of parameters, namely, sea surface wind speed, sea surface rainfall rate, seawater temperature, salinity and depth, into the model generated in S4, and outputting noise characteristics. The model provided by the method has good accuracy in predicting the spectral level of marine environmental noise.

Description

Technical Field

[0001] The invention relates to the field of ocean noise characteristic prediction, and in particular to a method for predicting noise characteristics based on a LM-BP neural network model. Background Art

[0002] In the actual engineering application field, there are a large number of ocean noise signals in the marine environment, but there are not many studies on them, especially in the prediction of noise characteristics. In the past, most of them were predicted by using the simple normal wave theory to model the spectrum level of marine environmental noise. However, considering the complexity and variability of marine environmental noise characteristics, noise may also be affected by factors such as distant ships and underwater organisms, and the seabed parameters only use ideal conditions, it is impossible to accurately describe the source parameters and propagation conditions of marine environmental noise. Therefore, it is particularly important to improve the accuracy of predicting noise characteristics.

[0003] Due to the complexity and variability of the characteristics of ocean environmental noise, the noise may also be affected by factors such as distant ships and underwater organisms, and the seabed parameters only use ideal conditions, which makes it impossible to accurately describe the source parameters and propagation conditions of the ocean environmental noise. The method of using traditional theoretical models to predict the spectral level of ocean environmental noise is not accurate in predicting noise characteristics. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a method for predicting noise characteristics based on a LM-BP neural network model, comprising the following steps:

[0005] S1 collects data: seawater temperature T, seawater salinity S, seawater depth z, sea surface wind speed v, sea surface rainfall rate r, and ocean environmental noise data;

[0006] S2. Create a data set including: sea water temperature T, sea water salinity S, sea water depth z, sea surface wind speed v, sea surface rainfall rate r data, and ocean environmental noise data;

[0007] S3. Taking seawater temperature T, seawater salinity S, seawater depth z, sea surface wind speed v, and sea surface rainfall rate r as input, design the topological structure of the LM-BP neural network model;

[0008] S4. Training LM-BP neural network model;

[0009] S5. Predict noise characteristics based on the LM-BP neural network model: input the five parameters of sea surface wind speed, sea surface rainfall rate, sea water temperature, salinity and depth into the model generated in S4 and output the noise characteristics.

[0010] Preferably, in step S1, the seawater temperature T, seawater salinity S, and seawater depth z are collected in the following manner:

[0011] The temperature T, salinity S, and depth z data in .xml format are collected by the CTD sensor and converted into a .csv file;

[0012] The spatial grid resolution of the sea surface wind speed v and the sea surface rainfall rate r is 25km, and the time resolution is 1h of ERA5, from which the sea surface wind speed v and the sea surface rainfall rate r data are obtained;

[0013] The method for collecting the ocean environmental noise data is as follows:

[0014] The ocean ambient noise data collected by the hydrophone is stored in the .dat file format and processed using the Welch method. The Welch method is as follows:

[0015] Divide the data x(n) of length N into L segments, each segment has M data, and the i-th segment data is represented as: 000000000

[0016] x i (n)=x(n+iM-M),0≤n≤M,1≤i≤L

[0017] Then add the window function w(n) to each segment of data and find the periodogram of each segment. The power spectrum of the i-th segment is:

[0018]

[0019] Where U is the normalization factor,

[0020]

[0021] The approximate values ​​of the periodograms of each segment are considered to be unrelated, and the final power spectrum is estimated as:

[0022]

[0023] The noise power spectrum is estimated using the Blackman window. The window function width is 1 / 3 of the sampling time, and the data overlap rate is 48%. After obtaining the power spectrum, further processing is performed to obtain the 1 / 3 octave spectral level L. p data.

[0024] Preferably, in step S3, the topological structure of the designed LM-BP neural network model is:

[0025] The model has 5 nodes in the input layer, corresponding to the sea surface wind speed v, sea surface rainfall rate r, sea water temperature T, salinity S and depth z. There are 2 hidden layers, the first layer has 12 nodes, the second layer has 8 nodes, and the output layer has 1 neuron.

[0026] The input vector is X = (x1, x2, x3, x4, x5) T , the hidden layer 1 input is A=(a1,a2…a 12 ) T , the output of hidden layer 1 is B = (b1, b2…b 12 ) T , the input of hidden layer 2 is C = (c1, c2, ... c8) T , the output of hidden layer 2 is D = (d1, d2, ... d8) T , the input of the output layer is e1, and the output of the output layer is f1;

[0027] The connection weight between the input layer and the hidden layer 1 is W = (w i,j )(i=1,2...5,j=1,2...12)

[0028] The connection weight between hidden layer 1 and hidden layer 2 is Y = (y j,k )(j=1,2...12,k=1,2...8)

[0029] The connection weight between hidden layer 2 and output layer is V = (v k,t )(k=1,2...8,t=1)

[0030] The mathematical relationship between the layers is:

[0031] For hidden layer 1,

[0032]

[0033] For hidden layer 2,

[0034]

[0035] For the output layer,

[0036]

[0037] For the transfer function between the input layer and the hidden layer 1, and between the hidden layer 1 and 2, the relu function is used; for the transfer function between the hidden layer and the output layer, the sigmoid function is used.

[0038] Preferably, the specific method of the S4 training model includes the following sub-steps:

[0039] S401. Data preprocessing

[0040] Shuffle processing: the sample data set is randomly shuffled;

[0041] Split processing: Split the data set into training set and test set in proportion

[0042] Normalization: Normalize the input data to the range of [-1, 1].

[0043] The normalization formula is as follows:

[0044]

[0045] Where: X * is the normalized data, X is the sample data, X max , X min are the maximum and minimum values ​​of X respectively.

[0046] S402. After the data set preprocessing is completed, the LM-BP neural network model is trained.

[0047] Set the number of training times, target error ε, random constant μ0, constant β, where 0<β<1, initialize the number of iterations k to 0, μ is the learning rate, initialize μ=μ0,

[0048] Initialize the parameter weights and threshold vectors, then input the sample data corresponding to the five parameters of sea surface wind speed v, sea surface rainfall rate r, sea water temperature T, salinity S and depth z into the input layer of the model for iterative training to determine the error index function E (ω k ) is smaller than before the iteration,

[0049] If it decreases, let μ = μ × β; if the error after iteration increases, let μ = μ / β, until the updated learning rate reduces the error;

[0050] The iteration is as follows:

[0051] Assume ω k is the column vector consisting of the weights and thresholds obtained at the kth iteration of the model. According to Newton's method, the weight iteration formula can be obtained:

[0052] ω k＋1 =ω k -H k -1 G k

[0053] In the formula, H k is the Hessian matrix of the multidimensional vector, G k is the error function E k Gradient

[0054] The error function E in the model iteration processk Converted to the following form:

[0055]

[0056] In the formula, e represents the error between the model output value and the true value, e = [e1, e2, e3, … e l ] T

[0057] According to the above formula, the mean square error E k The gradient of

[0058]

[0059] Where J is the Jacobian matrix, expressed as follows

[0060]

[0061] At this time, the Hessian matrix H k Can be transformed into

[0062]

[0063] Among them, I is the unit matrix, μ is the learning rate, 0<μ<1.

[0064] Therefore, after improving the BP neural network gradient descent method based on the LM algorithm, the weight correction formula is as follows:

[0065] Δω=-[J T J+μI] -1 J T e

[0066] ω k+1 =ω k -[J T J+μI] -1 J T e

[0067] Preferably, in S402, the number of training times is 1000 and the target error ε is 0.001.

[0068] Traditional BP neural networks usually use the gradient descent method to correct weights and thresholds. However, the gradient descent method is prone to produce multiple local optimal points on the error surface during the correction process. Once it falls into the local optimal point during the training process, it will affect the model training results and reduce the accuracy of the results. The Levenberg-Marquardt algorithm (LM algorithm for short) combines the Newton method with the gradient descent method on the basis of the gradient descent method, which can achieve adaptive convergence direction and greatly improve the convergence speed. Combining the LM algorithm with the BP neural network can improve the way of correcting weights and thresholds using the gradient descent method in the traditional BP neural network, which can not only effectively improve the network convergence speed, but also effectively avoid the target from falling into the local optimum. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the specific implementation of the present invention or the technical solution in the prior art, the following will briefly introduce the drawings required for the specific implementation or the prior art description. The following text will describe some specific embodiments of the present invention in detail in an exemplary but not restrictive manner with reference to the drawings. The same reference numerals in the drawings indicate the same or similar components or parts. In the drawings:

[0070] Figure 1 The ten-fold performance evaluation of the single hidden layer structure model of the present invention is

[0071] Figure 2 The ten-fold performance evaluation of the double hidden layer structure model of the present invention is

[0072] Figure 3 The neural network optimization algorithm structure of the present invention is

[0073] Figure 4 This is the LM-BP neural network program flow chart of the present invention.

[0074] Figure 5 The correlation coefficient of the test data and the correlation coefficient of the total data of model 2 described in the present invention are

[0075] Figure 6 Comparison of the predicted values ​​and expected values ​​of the model 2 described in the present invention

[0076] Figure 7 The error distribution between the predicted value and the expected value of model 2 described in the present invention is DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0078] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0079] A method for detecting and predicting noise characteristics based on a BP neural network improved by the LM algorithm solves the problem of low accuracy in predicting the spectrum level of ocean environmental noise using traditional theoretical models. The method comprises the following steps:

[0080] S1. Collect data;

[0081] S101. The main factors affecting the spectrum characteristics of marine environmental noise include sea surface wind speed v, sea surface rainfall rate r, sea water sound speed C ω ,

[0082] C ω =1449.2+4.6T-0.055T 2 +0.00029T 3 +(1.34-0.01T)(S-35)+0.016z

[0083] Its value depends on the seawater temperature T, salinity S and depth z. Considering that certain errors will occur in the process of calculating the speed of sound, in order to avoid errors, the data collected here are T, S, and z. The data collected by the underwater glider can be unpacked through the specified protocol. The unpacking process is to convert the temperature T, salinity S, and depth z data in .xml format collected from the CTD sensor carried by the glider into a .csv file. At the same time, the ERA5 (comprehensive reanalysis data) with a spatial grid resolution of 25km and a time resolution of 1h downloaded from the European Centre for Medium-Range Weather Forecasts (ECWF) is used to obtain the sea surface wind speed v and sea surface rainfall rate r data;

[0084] S102. The ocean environment noise data collected by the underwater glider through the hydrophone is stored in the .dat file format and processed using the Welch method, that is, the data x(n) of length N is divided into L segments, each segment has M data, and the i-th segment data is represented as: 000000000

[0085] x i (n)=x(n+iM-M),0≤n≤M,1≤i≤L

[0086] Then add the window function w(n) to each segment of data and find the periodogram of each segment. The power spectrum of the i-th segment is:

[0087]

[0088] In the formula, U is called the normalization factor,

[0089]

[0090] The approximate values ​​of the periodograms of each segment are considered to be unrelated, and the final power spectrum is estimated as:

[0091]

[0092] The algorithm can effectively reduce the variance of spectrum estimation and improve the estimation accuracy. However, in this process, the selection of window function type, window function width and data overlap rate will directly affect the result of power spectrum estimation. Therefore, after comparing and analyzing various windows, the Blackman window is used to estimate the noise power spectrum. The window function width is 1 / 3 of the sampling time, and the data overlap rate is 48%. Since the power spectrum is the sound intensity in a 1Hz bandwidth, for a certain center frequency, the center frequency of 1kHz is taken here, combined with the octave formula and the center frequency f of the frequency band c , upper limit frequency f h and the lower frequency f l Relationship

[0093]

[0094]

[0095] Where n is the octave, take 1 / 3, so f h With f c The ratio is 2 1 / 6 , f c With f l The ratio is also 2 1 / 6 .

[0096] The 1 / 3oct spectral level can be obtained by the following formula:

[0097]

[0098] S2. Creation of dataset;

[0099] Furthermore, the step S2 comprises the following steps:

[0100] The step S2 comprises the following steps:

[0101] After screening, a total of 1245 data sets were obtained, with a data set split ratio of 8:2, including 996 training samples and 249 test samples. The training samples and test samples were evaluated respectively, and the evaluation parameters included mean, mean square error, minimum, median, and maximum. The evaluation results are shown in Table 1 and Table 2. Comparing the data in Table 1 and Table 2, it can be seen that there is no significant difference between the training data and the test sample data, and the test sample can effectively verify the accuracy of the model.

[0102] Table 1 Training sample data evaluation

[0103]

[0104] Table 2 Test sample data evaluation

[0105]

[0106] S3. Design topological structure and establish model;

[0107] S301. The input parameters in the data set are v, r, T, S, z, and the output parameter is L p Therefore, the input layer nodes are set to 5 and the output layer nodes are set to 1. The selection of the specific number of hidden layers and the number of nodes in each layer requires further analysis.

[0108] Generally speaking, the number of nodes in the hidden layer can be determined using the following empirical formula:

[0109]

[0110] Among them, h is the number of hidden layer nodes, m is the number of input layer nodes, n is the number of output layer nodes, and a is the adjustment constant, which ranges from 1 to 10;

[0111] The value of S302.m is 5, the value of n is 1, and the value of h ranges from 3 to 12. The number of nodes is mainly selected for analysis: 4, 6, 8, 10, and 12. In addition, considering that the number of hidden layers should not be too many, the analysis is mainly focused on the cases of 1 and 2 layers. Networks with different topological structures are numbered according to the number of nodes in the input layer, hidden layer, and output layer. For example, when the neural network structure has a single hidden layer and the number of hidden layer nodes is 4, it is named 5-4-1;

[0112] S303. The situation when the hidden layer is 1 is analyzed, and the model performance is quantitatively analyzed using four evaluation parameters: (mean absolute error) MAE, (mean relative error) MRE, (mean square error) MSE, and (root mean square error) RMSE. The analysis results are shown in Table 3 Performance parameter evaluation of single hidden layer structure model.

[0113] Table 3 Performance parameter evaluation of single hidden layer structure model

[0114]

[0115] analyze Figure 1 It can be seen that under the single-layer hidden layer structure, as the number of hidden layer nodes gradually increases, the values ​​of MAE, MSE, RMSE, and MRE all decrease. Combining the data in Table 2 to compare the values ​​of the parameters of No. 5-12-1 and No. 5-4-1, it can be seen that compared with No. 5-4-1, No. 5-12-1 has a decrease of 19.45% in MAE, 38.38% in MSE, 21.54% in RMSE, and 19.03% in MRE. It can be seen that increasing the number of hidden layer nodes can effectively improve the performance of the neural network model. It can be seen that when the neural network model has a single-layer hidden layer structure, the performance is best when the number of nodes is 12;

[0116] S304. Analyze the double hidden layer structure. In the double hidden layer structure, the number of nodes in the first hidden layer is 12 according to the above analysis results, and the number of nodes in the second hidden layer is 4, 6, 8, 10, and 12 respectively for analysis. The model number naming rules are the same as above. The analysis results are shown in Table 4 Performance Parameter Evaluation of Double Hidden Layer Structure Model;

[0117] Table 4 Performance parameter evaluation of the double hidden layer structure model

[0118]

[0119] analyze Figure 2 It can be seen that when the model adopts a double-layer hidden layer structure, as the number of hidden layer nodes increases, the four values ​​of MAE, MSE, RMSE, and MRE do not simply increase or decrease, but first decrease and then increase. When the number of nodes in the second layer of the hidden layer is 8, the performance of model 5-12-8-1 is the best. Further analysis of the data in Table 3 shows that the MAE value of model 5-12-4-1 is 1.4818 and the MRE value is 2.325%, which is the largest among the five models; the MSE value of model 5-12-12-1 is 3.8026 and the RMSE value is 1.9411, which is the largest among the five models, and the MAE and MRE values ​​of model 5-12-8-1 are reduced by 1.68% and 4.17% respectively compared with model 5-12-4-1, and the MSE and RMSE values ​​are reduced by 2.87% and 1.34% respectively compared with model 5-12-12-1. This result shows that the more hidden layer nodes, the better. Too complex topology will reduce model performance. Therefore, it is very important to choose an appropriate topology. According to the analysis results, the neural network topology will adopt the model 5-12-8-1. That is, the input layer of the model has 5 parameters, namely sea surface wind speed v, sea surface rainfall rate r, sea water temperature T, salinity S and depth z. There are 2 hidden layers, the first layer has 12 nodes, and the second layer has 8 nodes. The topology is as follows: Figure 3 shown.

[0120] The output layer is 1 neuron, and the input vector is X = (x1, x2, x3, x4, x5) T , the hidden layer 1 input is A=(a1,a2…a 12 ) T , the output of hidden layer 1 is B = (b1, b2…b 12 ) T , the input of hidden layer 2 is C = (c1, c2, ... c8) T , the output of hidden layer 2 is D = (d1, d2, ... d8) T , the input of the output layer is e1, and the output of the output layer is f1;

[0121] The connection weight between the input layer and the hidden layer 1 is W = (w i,j )(i=1,2...5,j=1,2...12)

[0122] The connection weight between hidden layer 1 and hidden layer 2 is Y = (y j,k )(j=1,2...12,k=1,2...8)

[0123] The connection weight between hidden layer 2 and output layer is V = (v k,t )(k=1,2...8,t=1)

[0124] The mathematical relationship between the layers is:

[0125] For hidden layer 1,

[0126]

[0127] For hidden layer 2,

[0128]

[0129] For the output layer,

[0130]

[0131] For the transfer function between the input layer and the hidden layer 1, and between the hidden layer 1 and 2, the relu function is used; for the transfer function between the hidden layer and the output layer, the sigmoid function is used;

[0132] S4. Training model;

[0133] S401. The LM-BP neural network model is constructed based on the 5-12-8-1 topology structure obtained from the previous section. Before training the model, the data needs to be preprocessed, which mainly includes shuffling, segmentation and normalization. Shuffling refers to randomly shuffling the sample data set, which can reduce the probability of overfitting during the training process. The segmentation process uses the K-fold cross-validation method to divide the sample data set, thereby realizing the adjustment of the network parameters. There are a total of 1245 samples in the data set. 245 samples are taken out as the test set according to the ratio of 8:2. The remaining 1000 samples are randomly shuffled and divided into 10 groups, each with 100 data. Nine groups of data are taken as training sets in turn, and the remaining group is taken as the validation set. After the training is completed in turn, parameter optimization and 10 network errors are achieved. The average of these 10 network errors is taken. The above four error evaluation parameters are used as indicators to evaluate the performance of the network model under the selected hyperparameters (learning rate, number of iterations, number of neural network layers, activation function, etc.). By comparing the average network errors under different hyperparameters, the corresponding minimum is selected as the optimal hyperparameter of the model. Normalization refers to normalizing the input data to the interval [-1, 1], which can eliminate the dimensional differences caused by different input parameters and effectively improve the convergence rate:

[0134] The normalization formula is as follows:

[0135]

[0136] Where: X * is the normalized data, X is the sample data, X max , X min are the maximum and minimum values ​​of X respectively.

[0137] S402. After the data set is preprocessed, the LM-BP neural network model is trained, the hyperparameters are set, the number of training times is 1000, the target error ε is 0.001, k=0, constants μ0 and β (0<β<1), where k is the number of iterations, initialized to 0, μ0 is the learning rate, initialize μ=μ0, and initialize the parameter weights and threshold vectors, then input the sample data for iterative training, and judge the error index function E(ω k ) is smaller than before the iteration. If it is smaller, set μ = μ × β; if the error after the iteration increases, set μ = μ / β to make the algorithm more inclined to the gradient descent method until the updated learning rate reduces the error. The iterative principle in the middle is as follows:

[0138] Assume ω k is the column vector consisting of the weights and thresholds obtained at the kth iteration of the model. According to Newton's method, the weight iteration formula can be obtained:

[0139] ω k+1 =ω k -H k -1 G k

[0140] In the formula, H k is the Hessian matrix of the multidimensional vector, G k is the error function E k gradient.

[0141] The error function E in the model iteration process k Converted to the following form:

[0142]

[0143] In the formula, e represents the error between the model output value and the true value, e = [e1, e2, e3, … e l ] T .

[0144] According to the above formula, the mean square error E k The gradient of

[0145]

[0146] Where J is the Jacobian matrix, expressed as follows

[0147]

[0148] At this time, the Hessian matrix H k Can be transformed into

[0149]

[0150] Among them, I is the unit matrix, μ is the learning rate, 0<μ<1.

[0151] Therefore, after improving the BP neural network gradient descent method based on the LM algorithm, the weight correction formula is as follows:

[0152] Δω=-[J T J+μI] -1 J T e

[0153] ω k+1 =ω k -[J T J+μI] -1 J T The specific training steps of the eLM-BP neural network model are as follows: Figure 4 As shown;

[0154] Considering the impact of learning rate on model performance, the models under different learning rates are further discussed. The test parameter evaluation results under different learning rates are shown in Table 5.

[0155] Table 5 Test parameter evaluation under different learning rates

[0156]

[0157] By analyzing the test parameters of models 1 to 8 in Table 5, it is found that when the learning rate is 0.001, the MAE is 1.4370, the MRE is 2.227%, the MSE is 3.5616, and the RMSE is 1.8872. The values ​​of the four parameters are the smallest in Table 5, that is, model 2 has the best performance. Figure 5 It can be seen that the test correlation coefficient of model 2 is 0.88, and the total model coefficient is 0.86, which has a high correlation. Figure 6 is the comparison between the predicted value and the expected value of model 2, Figure 6 It can be seen that the model prediction value matches the expected value well, and the overall prediction performance is good. Figure 7 The error distribution between the predicted value and the expected value of all test set data of model 2 shows that the error is mainly distributed in the interval [-3,4], and the distribution is relatively concentrated, which shows that the model has good stability. According to the analysis results, model 5-12-8-1 has the best performance and the best prediction effect when the learning rate μ is 0.001.

[0158] S5. Prediction and evaluation model;

[0159] In order to further evaluate the prediction performance of the LM-BP neural network model and the noise spectrum level theoretical model, all 249 samples in the test set were selected, and the four evaluation parameters of MAE, MRE, MSE, and RMSE were used to evaluate the prediction data of the two models. The evaluation results are shown in Table 6. Among them, the MAE, MRE, MSE, and RMSE values ​​of the LM-BP neural network model prediction data are 1.4370, 2.227%, 3.5616, and 1.8872, respectively, which are 20.97%, 22.67%, 31.69%, and 17.35% higher than those of the theoretical model. It can be seen that the performance of the neural network model optimized by the LM-BP algorithm is significantly improved compared with the theoretical model, indicating that this model has better accuracy in predicting noise characteristics than the traditional noise theoretical model, and the effect is significant.

[0160] Table 6 Evaluation of prediction data of two models

[0161]

[0162] The above descriptions are only some specific implementation methods of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any person familiar with the art within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for predicting noise characteristics based on LM-BP neural network model, characterized in that: The following steps are involved: S1 collects data: seawater temperature T, seawater salinity S, seawater depth z, sea surface wind speed v, sea surface rainfall rate r, and ocean environmental noise data; S2. Create a data set including: sea water temperature T, sea water salinity S, sea water depth z, sea surface wind speed v, sea surface rainfall rate r data, and ocean environmental noise data; S3. Taking seawater temperature T, seawater salinity S, seawater depth z, sea surface wind speed v, and sea surface rainfall rate r as input, design the topological structure of the LM-BP neural network model; S4. Training LM-BP neural network model; S5. Predict noise characteristics based on LM-BP neural network model: input the five parameters of sea surface wind speed, sea surface rainfall rate, sea water temperature, salinity and depth into the model generated in S4, and output the noise characteristics; In step S3, the topological structure of the designed LM-BP neural network model is: The model has 5 nodes in the input layer, corresponding to the sea surface wind speed v, sea surface rainfall rate r, sea water temperature T, salinity S and depth z. There are 2 hidden layers, the first layer has 12 nodes, the second layer has 8 nodes, and the output layer is 1 neuron. The input vector is X = (x1, x2, x3, x4, x5) T , the hidden layer 1 input is A=(a1,a2…a 12 ) T , the output of hidden layer 1 is B = (b1, b2…b 12 ) T , the input of hidden layer 2 is C = (c1, c2, ... c8) T , the output of hidden layer 2 is D = (d1, d2, ... d8) T , the input of the output layer is e1, and the output of the output layer is f1; The connection weight between the input layer and the hidden layer 1 is W = (w i,j )(i=1,2...5,j=1,2...12) The connection weight between hidden layer 1 and hidden layer 2 is Y = (y j,k )(j=1,2...12,k=1,2...8) The connection weight between hidden layer 2 and output layer is V = (v k,t )(k=1,2...8,t=1) The mathematical relationship between the layers is: For hidden layer 1, For hidden layer 2, For the output layer, For the transfer function between the input layer and the hidden layer 1, and between the hidden layer 1 and 2, the relu function is used; for the transfer function between the hidden layer and the output layer, the sigmoid function is used; The specific method of training the model in S4 includes the following sub-steps: S401. Data preprocessing Shuffle processing: the sample data set is randomly shuffled; Split processing: Split the data set into training set and test set in proportion Normalization: Normalize the input data to the range of [-1, 1]. The normalization formula is as follows: Where: X * is the normalized data, X is the sample data, X max , X min are the maximum and minimum values ​​of X respectively; S402. After the data set preprocessing is completed, the LM-BP neural network model is trained. Set the number of training times, target error ε, random constant μ0, constant β, where 0<β<1, initialize the number of iterations k to 0, μ is the learning rate, initialize μ=μ0, Initialize the parameter weights and threshold vectors, then input the sample data corresponding to the five parameters of sea surface wind speed v, sea surface rainfall rate r, sea water temperature T, salinity S and depth z into the input layer of the model, perform iterative training, and judge the error index function E(ω k ) is smaller than before the iteration, If it decreases, let μ = μ × β; if the error after iteration increases, let μ = μ / β, until the updated learning rate reduces the error; The iteration is as follows: Assume ω k is the column vector consisting of the weights and thresholds obtained at the kth iteration of the model, and the weight iteration formula is obtained according to Newton's method: oh k+1 =ω k -H k -1 G k In the formula, H k is the Hessian matrix of the multidimensional vector, G k is the error function E k Gradient The error function E in the model iteration process k Converted to the following form: In the formula, e represents the error between the model output value and the true value, e = [e1, e2, e3, … e l ] T According to the above formula, we know the mean square error E k The gradient of Where J is the Jacobian matrix, expressed as follows At this time, the Hessian matrix H k Transformed into Where I is the identity matrix, μ is the learning rate, 0<μ<1; Therefore, after improving the BP neural network gradient descent method based on the LM algorithm, the weight correction formula is as follows: Give = [J] T J+μI] -1 J T e oh k+1 =ω k -[J T J+μI] -1 J T e.

2. A method for predicting noise characteristics based on LM-BP neural network model as claimed in claim 1, characterized in that: In step S1: The seawater temperature T, seawater salinity S, and seawater depth z are collected in the following manner: The temperature T, salinity S, and depth z data in .xml format are collected by the CTD sensor and converted into a .csv file; The spatial grid resolution of the sea surface wind speed v and the sea surface rainfall rate r is 25km, and the time resolution is 1h of ERA5, from which the sea surface wind speed v and the sea surface rainfall rate r data are obtained; The method for collecting the ocean environmental noise data is as follows: The ocean ambient noise data collected by the hydrophone is stored in the .dat file format and processed using the Welch method. The Welch method is as follows: Divide the data x(n) of length N into L segments, each segment has M data, and the i-th segment data is represented as: 000000000 x i (n)=x(n+iM-M),0≤n≤M,1≤i≤L Then add the window function w(n) to each segment of data and find the periodogram of each segment. The power spectrum of the i-th segment is: Where U is the normalization factor, The approximate values ​​of the periodograms of each segment are considered to be unrelated, and the final power spectrum is estimated as: The noise power spectrum is estimated using the Blackman window. The window function width is 1 / 3 of the sampling time, and the data overlap rate is 48%. After obtaining the power spectrum, further processing is performed to obtain the 1 / 3 octave spectral level L. p data.

3. A method for predicting noise characteristics based on LM-BP neural network model as claimed in claim 2, characterized in that: In S402, the number of training times is 1000, and the target error ε is 0.001.

Images