A shallow sea geoacoustic parameter inversion method based on feedback neural network model

Abstract

The present invention belongs to the field of marine engineering, and specifically relates to a shallow sea geoacoustic parameter inversion method based on a feedback neural network model, comprising: constructing a shallow sea acoustic field forward model, obtaining the sound pressure value of each point in the fluid, and forming a sound pressure data set p; constructing a BP neural network model; training and correcting the BP neural network model in S2 to make it meet the needs of geoacoustic parameter inversion under the shallow sea acoustic field forward model; substituting the measured sound pressure into the BP neural network model to obtain the values ​​of various geoacoustic parameters to be inverted in the preset environment model. By adjusting the neuron weights and thresholds in the constructed model, the entire neural network model can quickly approach the mapping relationship between the measured data and the geoacoustic parameters to be inverted. Under the same accuracy requirements, the inversion efficiency is higher than that of the existing optimization algorithm, and the determined neural network model can be directly used for solving the same type of problems, avoiding repeated calculations, and greatly improving its application efficiency and application prospects.

Description

Technical Field

[0001] The invention belongs to the field of marine engineering, and in particular relates to a shallow sea geoacoustic parameter inversion method based on a feedback neural network model. Background Art

[0002] Geoacoustic Parameters are parameters that describe the acoustic characteristics of the seabed, including medium sound velocity (including sound attenuation) and medium density. These two types of geoacoustic parameters are important physical parameters for effectively studying the sound propagation characteristics of the shallow seabed. How to efficiently obtain shallow sea geoacoustic parameters has always been a classic and hot issue in the field of domestic hydroacoustic research.

[0003] Since the above-mentioned geoacoustic parameters are difficult to measure directly and over a large area, the acoustic method can be used to quickly and efficiently invert and obtain the geoacoustic parameters of a large sea area, which has attracted extensive attention and has important research significance and application value. In recent years, various shallow sea geoacoustic parameter inversion methods based on various sound field propagation characteristics have emerged, such as the geoacoustic parameter inversion method using propagation loss, the geoacoustic parameter inversion method using the arrival time of the acoustic signal, and the inversion method using the waveguide dispersion characteristics.

[0004] However, the above-mentioned geoacoustic parameter inversion methods mainly focus on the selection of the forward model in the inversion problem, and then solve the objective function through various classical optimization algorithms, such as genetic algorithm (GA), simulated annealing algorithm (SA), etc., to obtain the results of the parameters to be inverted. When various classical optimization algorithms are applied, the iterative optimization calculation between the input data and the optimal inversion solution not only consumes a lot of computing time, but also easily falls into the local optimal solution.

[0005] Based on the above situation, the applicant proposed a shallow water geoacoustic parameter inversion method based on a feedback neural network model, in order to achieve efficient and accurate inversion of shallow water geoacoustic parameters. Summary of the invention

[0006] In order to make up for the deficiencies of the prior art, the present invention provides a technical solution for a shallow water geoacoustic parameter inversion method based on a feedback neural network model.

[0007] Since shallow sea geoacoustic parameters are important environmental parameters that determine the distribution characteristics of the sound field in shallow sea environments, changes in seabed geoacoustic parameters will have a significant impact on the distribution characteristics of the underwater sound pressure field. Therefore, the geoacoustic parameters can be inverted from shallow sea sound pressure field measurement data.

[0008] The traditional shallow sea geoacoustic parameter inversion method based on sound pressure data adopts the method of multiple matching optimization between the measured sound pressure data and the simulated values ​​of the sound field model. That is, through various optimization algorithms, a set of geoacoustic parameters that best matches the measured sound pressure is searched in the simulated sound pressure data corresponding to multiple sets of geoacoustic parameters as the inversion result. However, the optimization algorithm is prone to fall into the local optimal solution during application, and each optimization process needs to be cyclically brought into the sound field forward model for iterative calculation, which greatly increases the calculation time. The BP neural network model starts from the perspective of machine learning, and repeatedly trains the neural network model between the sound pressure data set and the geoacoustic parameters to be inverted, so that it approximates the complex nonlinear mapping between the two. In the later stage, only the measured sound pressure data needs to be substituted into the trained model to obtain the corresponding parameters to be inverted, avoiding repeated iterative calculations in the use of the optimization algorithm. It can not only greatly shorten the calculation time, but also has strong robustness.

[0009] In this application, the simulated forward sound pressure field data under a preset shallow sea environment model will be used as a training sound pressure data set to train a BP neural network model that satisfies the mapping relationship between the sound pressure field under the preset model and the geoacoustic parameters to be inverted. Finally, this model will be used to achieve efficient inversion acquisition of geoacoustic parameters under the preset shallow sea environment model.

[0010] A shallow sea geoacoustic parameter inversion method based on a feedback neural network model comprises:

[0011] S1 builds a shallow sea acoustic field forward model to obtain the acoustic pressure value of each point in the fluid and form an acoustic pressure data set p;

[0012] S2 builds BP neural network model;

[0013] S3 trains and corrects the BP neural network model in S2 to make it meet the needs of geoacoustic parameter inversion under the shallow sea acoustic field forward model;

[0014] S4 substitutes the measured sound pressure into the BP neural network model to obtain the values ​​of various geoacoustic parameters to be inverted in the preset environmental model.

[0015] Further, the S1 includes:

[0016] In the three-dimensional cylindrical coordinate system, a sound field model that meets the characteristics of the shallow sea environment is preset. In the model, the simple harmonic point sound source is located on the symmetry axis of the cylindrical coordinates. The three-dimensional problem is converted into a two-dimensional (r, z) plane for solution. z = 0 represents the sea surface, the sea surface is downwards as the positive direction of the depth z axis, and the positive r axis represents the direction of sound field propagation. In the model, the depth of the sea layer is set to H; the frequency is f 0 The sound source is located at the depth z of the seawater layer s The density and speed of sound in the seawater layer are ρ 1 and c 1 ; The seabed longitudinal wave speed, transverse wave speed, density, longitudinal wave speed attenuation and transverse wave speed attenuation are expressed as cp 、c s , b , α p , α s Indicates; let the displacement potential function in the model fluid layer be φ 1 , the sound pressure satisfies p=ρ 1 ω 2 φ 1 , angular frequency ω=2πf 0 , the sound pressure value at each point in the fluid is obtained by solving the displacement potential function, and the sound pressure field in the water layer is expressed as:

[0017]

[0018] Among them, Z 1 is the ordinary differential equation for depth z and horizontal wave number ξ, J 0 is the zero-order Bessel function;

[0019] Solve formula (3) to obtain the sound pressure data set p.

[0020] Further, the S2 includes:

[0021] The neural network input layer uses n different receiving positions (r i ,z i m groups of sound pressure data p = [p )(1≤i≤n) 1 (r 1 ,z 1 ),…,p j (r i ,z i ),…,p m (r n ,z n )] m×n As input data, the corresponding geoacoustic parameter Y = [c p ,c s ,ρ b ,α p ,α s ] m×5 As label data for model building.

[0022] Furthermore, the S2 further includes:

[0023] When constructing the BP neural network model, a single hidden layer is set; in the model, neurons in the same layer are not connected to each other, and there are two types of signal exchanges between layers. One is the working signal function, that is, the sound pressure field data p in the input layer. j (r i ,z i ) and the hyperparameter matrix [w,b], the signal is transmitted from the input layer to the output layer, and the expression is The other is the error signal E(m), which is the error function between the inversion result of the network model and the true value. It is transmitted layer by layer from the output end to the output end, and the expression is:

[0024] Where w = [w kv ,w vl ],w kv represents the weight from the input layer to the hidden layer, w vl represents the weight from the hidden layer to the output layer, b v represents the threshold of each neuron in the hidden layer, x is the input value of each neuron, Y r,e =[c p ,c s ,ρ b ,α p ,α s ], is the matrix composed of the parameters to be inverted, Y r represents the simulation value, Y e represents the inversion value, N represents the number of samples;

[0025] After the sound pressure data p is substituted into the input layer, the neurons in each layer are connected through the hyperparameter matrix [w, b] and the activation function f(x), and the inversion result Y is finally obtained through the hidden layer and the output layer. e .

[0026] Furthermore, the S2 also includes: the number of neurons in each layer is set according to to determine;

[0027] Among them, n represents the number of nodes in the input layer, that is, the number of simulated sound pressure points, v represents the number of nodes in the hidden layer, l represents the number of nodes in the output layer, that is, the number of inverted geoacoustic parameters, and α is a constant coefficient.

[0028] Furthermore, the S2 also includes: in the BP neural network model, kv ,I vl are the hidden layer input data, the hidden layer output data, I vl The calculation formula is The inversion result Y e The calculation formula is

[0029] Furthermore, S2 also includes: updating parameters by gradient descent, and the design process is as follows:

[0030]

[0031]

[0032] Among them, the partial derivative of the weight parameter from the sound pressure data p to the hidden layer is Δwkv , the partial derivative of the weight between the hidden layer and the ground acoustic parameter Y is Δw vl , η is the learning rate. During the calculation process, the number of iterations t and the correction parameter w are continuously updated according to whether the E(m) value meets the set accuracy. kv 、w vl , such as the modified formula is w kv (t+1)=w kv (t)+Δw kv , w vl (t+1)=w vl (t)+Δw vl .

[0033] Furthermore, S3 also includes: mapping each group of parameters in the training set with the corresponding generated ambient sound pressure one by one and then inputting them into the model for training. When the error function E(m) reaches the set accuracy requirement, the training is completed.

[0034] The present invention proposes an inversion method for five geoacoustic parameters, namely, shallow seabed density, longitudinal wave sound velocity, transverse wave sound velocity, longitudinal wave sound velocity attenuation and transverse wave sound velocity attenuation. When the method is applied, the theoretical prediction value of the shallow sea sound pressure field is obtained by the fast field method, and then the relationship model between the predicted sound pressure field and the geoacoustic parameter value to be inverted is established according to the BP neural network model, and finally the measured sound pressure field data is brought into the neural network model to obtain the inversion result. The processing results of the simulation data and the water tank scaled experimental data show that the application of this method can accurately obtain the five seabed geoacoustic parameters of interest.

[0035] Compared with the various optimization algorithms in the prior art that are inefficient and prone to falling into local optimal solutions, the geoacoustic parameter inversion method based on the neural network model of the present invention adjusts the neuron weights and thresholds in the constructed model so that the entire neural network model quickly approximates the mapping relationship between the measured data and the geoacoustic parameters to be inverted. Under the same accuracy requirements, the inversion efficiency is higher than that of the existing optimization algorithm, and the determined neural network model can be directly used to solve the same type of problems, avoiding repeated calculations, greatly improving its application efficiency and application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flow chart of the present invention;

[0037] Figure 2 It is a schematic diagram of the forward model of shallow sea acoustic field in cylindrical coordinates;

[0038] Figure 3 It is a schematic diagram of the structure of the BP neural network model used for geoacoustic parameter inversion under a preset environment;

[0039] Figure 4 Schematic diagram of training error during training;

[0040] Figure 5 This is a schematic diagram of regression analysis during training;

[0041] Figure 6 The comparison chart of the inversion results of 200 test sets and the preset values;

[0042] Figure 7 Schematic diagram of the error fluctuation of each parameter during the test;

[0043] Figure 8 Schematic diagram of simulated TL curve under simulation conditions;

[0044] Fig. 9 Schematic diagram of BP inversion TL curve under simulation conditions;

[0045] Fig.10 It is a schematic diagram of the fitness curve in the process of solving the SA algorithm;

[0046] Fig.11 It is a schematic diagram of the fitness curve in the process of solving the BP neural network model;

[0047] Fig.12 This is a schematic diagram of the TL curve measured in the water tank;

[0048] Fig.13 It is a schematic diagram of TL curve inversion by BP neural network;

[0049] Fig.14 Schematic diagram of SA inversion TL curve. DETAILED DESCRIPTION

[0050] The present invention will be further described below in conjunction with the accompanying drawings.

[0051] See also Figure 1 , a shallow sea geoacoustic parameter inversion method based on a feedback neural network model, comprising the following steps:

[0052] S1 Shallow Sea Acoustic Field Forward Model

[0053] Considering the shallow sea environment, the marine environment can generally be approximated as a horizontal layered structure consisting of a seawater layer and a semi-infinite seafloor layer. For this reason, a sound field model that conforms to the characteristics of the shallow sea environment is preset in a three-dimensional cylindrical coordinate system. In the model, the simple harmonic point sound source is located on the cylindrical coordinate symmetry axis. Considering that the influence of the seabed shear wave speed on the sound propagation in the shallow sea waveguide environment cannot be ignored, the seawater layer and the seafloor layer are approximated as a uniform isotropic fluid medium and an elastic medium, respectively. Due to the axial symmetry of the cylindrical coordinate system, the three-dimensional problem can be transformed into a two-dimensional (r, z) plane for solution, z = 0 represents the sea surface, the sea surface is downwards in the positive direction of the depth z axis, and the positive r axis represents the direction of sound field propagation.

[0054] In the model, the depth of the seawater layer is set to H; the frequency is f0 The sound source is located at the depth z of the seawater layer s The density and speed of sound in the seawater layer are ρ 1 and c 1 ; The seabed longitudinal wave speed, transverse wave speed, density, longitudinal wave speed attenuation and transverse wave speed attenuation are expressed as c p 、c s , b , α p , α s It indicates that the above five parameters are the seabed geoacoustic parameters to be inverted in the present invention.

[0055] Under the wave theory, all physical quantities in the above model can be uniformly represented by the displacement potential function. Suppose the displacement potential function in the fluid layer of the model is φ 1 , and the sound pressure of the research object of the present invention satisfies p=ρ 1 ω 2 φ 1 (angular frequency ω=2πf 0 ), the sound pressure value at each point in the fluid can be obtained by solving the displacement potential function. Since the displacement potential function in the fluid layer satisfies:

[0056]

[0057] Its form is:

[0058]

[0059] where Z is the ordinary differential equation of depth z and horizontal wave number ξ, J 0 is the zero-order Bessel function.

[0060] According to the above derivation results, the acoustic pressure field in the water layer can be expressed as:

[0061]

[0062] For solving formula (3), the Normal Mode Method (NMM) and the Fast Field Method (FFM) can generally be used to solve formula (3). For shallow sea environments, FFM converts the integral in formula (3) into the form of Fourier transform for direct solution, which is more suitable for rapid calculation of shallow sea sound fields. Therefore, in the present invention, FFM is selected to perform forward simulation on the sound pressure field in the above parameterized model. After the sound pressure data set p is calculated, it is substituted into the BP neural network for model training to establish a model that can reflect the following: Figure 2 A BP neural network model of the mapping relationship between the underwater sound pressure field and the geoacoustic parameters to be inverted in a preset shallow sea environment.

[0063] S2 builds BP neural network

[0064] Drawing on the geoacoustic parameter inversion method using acoustic pressure field data, when studying the shallow sea geoacoustic parameter inversion method based on the BP neural network model, the neural network input layer uses n different receiving positions (r i ,z i m groups of sound pressure data p = [p )(1≤i≤n) 1 (r 1 ,z 1 ),…,p j (r i ,z i ),…,p m (r n ,z n )] m×n As input data, the corresponding geoacoustic parameter Y = [c p ,c s ,ρ b ,α p ,α s ] m×5 As label data, the model is constructed. Considering the complexity and computing power of the preset ocean environment model, a single hidden layer is set when constructing the BP neural network model. In the model, neurons in the same layer are not connected to each other. There are two types of signal exchanges between layers. One is the working signal function, that is, the sound pressure field data p in the input layer. j (r i ,z i ) and the hyperparameter matrix [w,b], the signal is transmitted from the input layer to the output layer in the forward direction, as shown in formula (4), where w = [w kv ,w vl ],w kv represents the distance from the input layer to the hidden layer, w vl represents the weight from the hidden layer to the output layer, b represents the threshold of the hidden layer neurons, and b v The other is the error signal E(m), which is the error function between the inversion result of the network model and the true value. In the present invention, the design is given by the mean square error function (MSE), which is transmitted from the output end to the output end layer by layer.

[23] .

[0065] Forward activation function f(x):

[0066]

[0067] Reverse error transfer function E(m) - Mean Square Error (MSE):

[0068]

[0069] Where x is the input value of each neuron, Y r,e =[c p ,c s ,ρ b ,α p ,α s ], is the matrix composed of the parameters to be inverted, Y r represents the simulation value, Y e represents the inversion value, and N represents the number of samples.

[0070] After the sound pressure data p is substituted into the input layer, the neurons in each layer are connected through the hyperparameter matrix [w, b] and the activation function f(x), and the inversion result Y is finally obtained through the hidden layer and the output layer. e

[24] The number of neurons in each layer can be determined according to formula (6):

[0071]

[0072] Where n represents the number of nodes in the input layer, i.e. the number of simulated sound pressure points, v represents the number of nodes in the hidden layer, l represents the number of nodes in the output layer, i.e. the number of inverted geoacoustic parameters, and α is a constant coefficient. kv ,I vl are the hidden layer input data, the hidden layer output data, and the inversion result Y e The calculation process is given by formula (7)-formula (8):

[0073]

[0074]

[0075] Since the network input error is a function of the weights and thresholds of each layer, the error E(m) can be changed by adjusting the weights. Obviously, the principle of adjusting the weights is to continuously reduce the error, so the weights should be proportional to the gradient descent of the error. Therefore, the gradient descent method is used to update the parameters. The design process is given by formula (9)-formula (10):

[0076]

[0077]

[0078] The partial derivative of the weight parameter from the sound pressure data p to the hidden layer is Δw kv , the partial derivative of the weight between the hidden layer and the ground acoustic parameter Y is Δw vl, η is the learning rate. During the calculation process, the number of iterations t and the correction parameter w are continuously updated according to whether the E(m) value meets the set accuracy. kv 、w vl , as shown in formula (11)-formula (12):

[0079] w kv (t+1)=w kv (t)+Δw kv (11)

[0080] w vl (t+1)=w vl (t)+Δw vl (12)

[0081] Since the present invention simplifies the seabed into two layers to establish a BP neural network inversion model, the input layer in the network structure is the simulated sound pressure data p, a set of sound pressure data p(r i ,z i ), i = 1, 2, 3, ... n (the number of measured sound pressure points n = 720), the output layer is the 5 geoacoustic parameters to be inverted, according to formula (6), it can be determined that α = -15, the hidden layer is v = 9 neurons, and the output layer is l = 5 neurons. The number of elements in the hyperparameter matrix (w, b) is 720 × 9 + 9 × 5 + 9 = 6534, indicating that when this neural network model is constructed by the gradient descent method, 6534 weight parameters need to be adjusted to make the neural network model approach the complex mapping relationship between input and output, thereby realizing the inversion calculation. For the shallow sea environment model preset by the present invention, the constructed BP neural network model structure is as follows Figure 3 shown.

[0082] S3BP model training dataset

[0083] Considering the variation range of shallow seabed geoacoustic parameters, the parameter training range of the BP neural network model for geoacoustic parameter inversion under the preset environment is set as shown in Table 1. The simulated sound pressure field data is set to the sound source depth z s =20m, receiving depth z r = 10m, seawater depth H = 100m A set of horizontally evenly spaced To collect the sound pressure field, each receiving point is 2m apart, and a total of 720 receiving points are set.

[0084] Table 1 BP neural network model training parameter setting range

[0085]

[0086] The model training samples used are 2200 sets of sound pressure data randomly generated within the search range of Table 1, and 2000 sets of data are randomly divided into training sets and the remaining 200 sets of data are test sets. Each set of parameters in the training set is mapped one by one with the corresponding generated ambient sound pressure and then put into the model for training. When the error function E(m) reaches the set accuracy requirement, the training is completed. In the training of the preset model of the present invention, the MSE value changes with the number of iterations and the normalized regression line diagram of the training data are respectively represented by Figure 4 , Figure 5 Given.

[0087] Under the above simulation conditions, after the training is completed, the value of E(m) reaches 10 after 10 iterations. -3 The setting accuracy is as follows: Figure 5 As shown in the figure, most of the normalized output values ​​in training have a good fit with the target values, and most of the output values ​​are scattered around the fitting line, indicating that the error reduction rate and training effect of the entire neural network in training are considerable, and a BP neural network model that meets the accuracy of geoacoustic parameter inversion is efficiently constructed.

[0088] After S4 completes the training of the BP neural network model in the preset shallow sea environment, in actual application, it only needs to substitute the measured sound pressure into the BP neural network model to obtain the values ​​of various geoacoustic parameters to be inverted in the preset environmental model.

[0089] The verification process of the above BP neural network model is as follows:

[0090] In order to verify the feasibility of the constructed BP neural network model in practical applications, the simulated sound pressure data and the measured sound pressure data of the scaled experiment will be used to verify the BP neural network model established above, and the performance of the BP neural network model and the classical optimization algorithm in the inversion of geoacoustic parameters will be compared and analyzed.

[0091] 1. Simulation data verification

[0092] After completing the BP neural network model training for the inversion of five types of geoacoustic parameters under the preset model, the reliability of the trained neural network model was first verified using simulation data. In the verification, 200 groups of data randomly generated and divided within the parameter range of Table 1 were selected as the test set, and the geoacoustic parameters of 200 groups of sound pressure field data were inverted through the trained BP neural network model. In order to quantify the error between the inversion results of each parameter and the preset true value, the performance function R2 was introduced to numerically represent the overlap between the inversion value and the true value. The closer the R2 value is to 1, the closer the inversion result is to the preset true value.

[0093]

[0094] Figure 6After processing 200 sets of acoustic pressure field data, a comparison chart between 5 types of inversion results and preset true values ​​is given. In the figure, "*" represents the simulated true value, "o" represents the inversion result, and the y-axis is the search range of each parameter to be inverted. Due to the different sensitivities of various geoacoustic parameters, although the R2 values ​​of the comparison results of various geoacoustic parameters are different, the overlap of each set of data is above 99.00%, which proves the accuracy of the simulation inversion results.

[0095] based on Figure 6 The degree of coincidence between the inversion value and the preset value is given intuitively. Figure 7 The absolute error between the inverted value and the preset value of each parameter in the validation set is given by Figure 7 It can be clearly observed that the error trend of each parameter in the inversion model is below 0.1 during verification. b 、c p 、c s The result error is always maintained below 0.01, and the inversion effect is excellent. p , α s The error change is relatively large, but in the prediction results, the parameter α with the largest error change p The maximum error is only 0.065, and there is no large error fluctuation. It can be seen that the constructed BP neural network model has good and stable prediction performance for shallow seabed geoacoustic parameter inversion, and the calculation efficiency is high, and the prediction results are highly reliable. Figure 8 , 9 The analysis of different sensitivities of seabed parameters shows that the robustness of this BP neural network for inversion of five types of parameters is: p 、c s , b >α p , α s .

[0096] Figure 8 , 9 A comparison chart is given between the transmission loss curve (TL) calculated by setting the true value of the geoacoustic parameter and the TL curve calculated by applying the inversion result. From the comparison, it can be seen that the distribution characteristics of the two curves are basically the same, which further proves the accuracy of the geoacoustic parameter inversion result under the preset model obtained by the BP neural network model inversion of the present invention.

[0097] In order to further discuss the application prospect of the BP neural network model of the present invention in the inversion of geoacoustic parameters, the BP neural network model and the classic SA algorithm are used to compare the inversion of the same inversion problem. To ensure the comparability of the results, the loss function E(m) in the BP neural network model is used as the cost function in the SA algorithm. The parameter settings for the comparison of the two algorithms during the calculation process are shown in Table 2. The adaptation process of the loss function and cost function in the two algorithms is shown in Table 2. Fig.10 , 11 The error comparison and efficiency comparison of the inversion results are given in Table 3 and Table 4 respectively.

[0098] Table 2 Inversion algorithm parameter settings

[0099]

[0100]

[0101] Depend on Fig.10 The change of the objective function value of the SA algorithm in the iteration shows that although it quickly drops to 10 during the optimization process -2 The accuracy is close to 10, but it takes a total of 730 iterations to reach the preset accuracy of 10. -3 Accuracy requirements, Fig.11 The BP neural network model only iterated 10 steps to reach the preset accuracy requirements and complete the construction of the inversion model. In terms of the number of model iterations, the computational efficiency of the neural network is much higher than the inversion performed by the optimization algorithm alone. After repeated calculations, the comparison of the inversion calculation time and the number of model iterations of the two algorithms for geoacoustic parameters under simulation conditions is given in Table 3.

[0102] Table 3 Comparison of inversion algorithm efficiency

[0103]

[0104] Table 4 Comparison of inversion results of various parameters

[0105]

[0106] From the simulation comparison, we can see that the error between the inversion results obtained by using the SA algorithm for optimization calculation and the BP neural network model and the true value is ≤10 -3 Compared with the true value, the inversion results obtained by the two methods are only slightly different from the true value. p The inversion accuracy of the BP neural network model is higher than that of the BP neural network model. s , b , α p and α sThe accuracy of the inversion results is worse than that of the BP network model, and the BP neural network is more stable in controlling relative errors and has smaller fluctuations.

[0107] The present invention simultaneously utilizes the BP neural network model and the classical SA algorithm to respectively perform geoacoustic multi-parameter inversion. Theoretically, the BP neural network model has parallelism, can simultaneously train five types of inverted parameters and acoustic field data, directly establish a neural network between the two, and directly obtain the inversion result through the network, while the optimization algorithm performs random perturbations within the search range set by the parameters, then matches the data to be solved, and then screens the optimal solution. From the error analysis, the SA algorithm has differences in the resolution of multiple parameters in the multi-parameter solution problem, so the solution accuracy is inconsistent, and there is a problem of too long calculation time; while the BP neural network model only uses 2000 groups of data for training, and the number of iteration steps in the operation is 1.3% of the SA algorithm, which has achieved considerable accuracy requirements. Although there are also similar differences in the robustness of the inversion of different parameters, the overall inversion parameters can reach consistent accuracy.

[0108] 2. Measured data verification

[0109] Based on the simulation verification of the accuracy and applicability of the research method, the feasibility of the inversion method of the present invention in practical application is further verified by combining the experimental data of the anechoic pool. The experiment was carried out in an anechoic pool, using a uniform and high-hardness PVC board (Polyvinyl Chloride Polymer, with a measured density of 1.20g / cm -3 ) "semi-infinite elastic seafloor"; the sound source depth z in the experiment s =87mm, receiving depth z r =84mm, water depth H = 182mm, speed of sound in water c 1 The empirical formula of the speed of sound is used to calculate c by considering the water temperature of 11.5℃ under standard atmospheric pressure. 1 =1450.212m / s; keep the sound source fixed in the process, transmit f=135kHz pulse signal, place the receiving hydrophone on a movable frame, the acquisition card sampling frequency fs=20MHz, the hydrophone moves away from the sound source, and records data every time it moves 2mm. The propagation loss in the water tank environment is measured as follows Fig.12 shown.

[0110] The BP neural network model established in the present invention and the classic SA algorithm are used to perform parameter inversion on the measured data of the water tank. Table 5 shows the search ranges set in the inversion and the final inversion results of the two methods:

[0111] Table 5 Inversion results of measured data

[0112]

[0113] Figure 12-14 The propagation loss comparison curves of the BP neural network model and the SA inversion algorithm on the measured data are given. From the comparison curves in the figure, it can be seen that the inverse TL curves of the two methods are basically consistent with the measured TL curves. Combined with the inversion results given in Table 5, it can be seen that the BP neural network model and the SA algorithm have good performance on c p 、c s , b , α p , α s The inversion results of these five types of geoacoustic parameters are very close, which further verifies the applicability of the BP neural network model in the actual geoacoustic parameter inversion research. b =1.20g / cm -3 In the case of about 1.23 g / cm, the inversion of BP neural network model and SA inversion algorithm are 1.23 g / cm -3 and 1.21 g / cm -3 , the relative error values ​​are 2.5% and 0.83%, combined with Fig.10 and Fig.11 Although the BP neural network model is slightly lower in accuracy than the SA inversion algorithm, its efficiency is only 30% of the SA inversion algorithm. On the whole, the BP neural network algorithm will have a broader application prospect and development space in the inversion of geoacoustic parameters.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A shallow sea geoacoustic parameter inversion method based on a feedback neural network model, characterized in that: include: S1 builds a shallow sea acoustic field forward model to obtain the acoustic pressure values ​​at each point in the fluid, forming an acoustic pressure data set p, including: In the three-dimensional cylindrical coordinate system, a sound field model that meets the characteristics of the shallow sea environment is preset. In the model, the simple harmonic point sound source is located on the cylindrical coordinate symmetry axis. The three-dimensional problem is converted into a two-dimensional (r, z) plane for solution. z = 0 represents the sea surface, the sea surface is downward as the positive value direction of the depth z axis, and the positive r axis represents the direction of sound field propagation. In the model, the depth of the sea layer is set to H; the sound source with a frequency of f0 is located at the depth z of the sea layer. s The density and sound speed in the seawater layer are ρ1 and c1 respectively; the longitudinal wave sound speed, transverse wave sound speed, density, longitudinal wave sound speed attenuation and transverse wave sound speed attenuation of the sea bottom are c p 、c s , b , α p , α s Represents; Assume that the displacement potential function in the model fluid layer is Sound pressure meets , angular frequency ω=2πf0, the sound pressure value at each point in the fluid is obtained by solving the displacement potential function, and the sound pressure field in the water layer is expressed as: Where Z1 is the ordinary differential equation of depth z and horizontal wave number ξ, and J0 is the zero-order Bessel function; Solve formula (3) to obtain the sound pressure data set p; S2 builds a BP neural network model, including: The neural network input layer uses n different receiving positions (r i ,z i ) of m groups of sound pressure data p=[p1(r1,z1),…,p j (r i ,z i ),…,p m (r n ,z n )] m×n As input data, 1≤i≤n, and the corresponding geoacoustic parameter Y=[c p ,c s ,ρ b ,α p, α s ] m×5 As label data for model building; When constructing the BP neural network model, a single hidden layer is set; in the model, neurons in the same layer are not connected to each other, and there are two types of signal exchanges between layers. One is the working signal function, that is, the sound pressure field data p in the input layer. j (r i ,z i ) and the hyperparameter matrix [w,b], the signal is transmitted from the input layer to the output layer, and the expression is The other is the error signal E(m), which is the error function between the inversion result of the network model and the true value. It is transmitted layer by layer from the output end to the output end, and the expression is: Where w = [w kv ,w vl ],w kv represents the weight from the input layer to the hidden layer, w vl represents the weight from the hidden layer to the output layer, b v represents the threshold of each neuron in the hidden layer, x is the input value of each neuron, Y r,e =[c p ,c s ,ρ b ,α p, α s ], is the matrix composed of the parameters to be inverted, Y r represents the simulation value, Y e represents the inversion value, N represents the number of samples; After the sound pressure data p is substituted into the input layer, the neurons in each layer are connected through the hyperparameter matrix [w, b] and the activation function f(x), and the inversion result Y is finally obtained through the hidden layer and the output layer. e ; The number of neurons in each layer is based on to determine; Where n represents the number of nodes in the input layer, i.e., the number of simulated sound pressure points; v represents the number of nodes in the hidden layer; l represents the number of nodes in the output layer, i.e., the number of inverted geoacoustic parameters; and α is a constant coefficient; In the BP neural network model kv ,I vl are the hidden layer input data, the hidden layer output data, I vl The calculation formula is The inversion result Y e The calculation formula is The gradient descent method is used to update the parameters. The design process is as follows: Among them, the partial derivative of the weight parameter from the sound pressure data p to the hidden layer is Δw kv , the partial derivative of the weight between the hidden layer and the ground acoustic parameter Y is Δw vl , η is the learning rate. During the calculation process, the number of iterations t and the correction parameter w are continuously updated according to whether the E(m) value meets the set accuracy. kv 、w vl , such as the modified formula is w kv (t+1)=w kv (t)+Δw kv , w vl (t+1)=w vl (t)+Δw vl ; S3 trains and corrects the BP neural network model in S2 to make it meet the needs of geoacoustic parameter inversion under the shallow sea acoustic field forward model; S4 substitutes the measured sound pressure into the BP neural network model to obtain the values ​​of various geoacoustic parameters to be inverted in the preset environmental model.

2. The shallow sea geoacoustic parameter inversion method based on the feedback neural network model according to claim 1 is characterized in that: The S3 also includes: mapping each group of parameters in the training set with the corresponding generated ambient sound pressure one by one and then inputting them into the model for training. When the error function E(m) reaches the set accuracy requirement, the training is completed.

Images