A fast matching field seafloor acoustic parameter inversion method based on machine learning

Abstract

A method for rapid matching field seabed acoustic parameter inversion based on machine learning, firstly, according to the environmental prior information and transceiver configuration of the experimental sea area, an acoustic propagation model is selected to establish a simulation environment; then, the target function value is calculated in combination with the measured sound field under different seabed acoustic parameters, and a data set is established based on the forward mapping for training and testing the machine learning model; finally, in the parameter update process of the matching field inversion, for the parameters before and after the disturbance, the machine learning model is first used to quickly screen the effective disturbance, and then the acoustic propagation model is used to judge whether to accept or reject, and after multiple rounds of parameter updates, it converges to the global optimum. The present invention introduces a pre-screening process based on machine learning in the traditional matching field inversion process, reduces the number of calls to the acoustic propagation model, and thus reduces the time consumption of inversion, and the further judgment of the acoustic propagation model ensures the inversion accuracy, which can improve the ability of rapid seabed characteristic analysis.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater acoustic detection, and in particular to a method for inverting seabed acoustic parameters of a fast matching field based on machine learning. Background Art

[0002] In shallow sea environments, part of the sound rays emitted by underwater sound sources are reflected at the junction of the water body and the seabed, carrying seabed information. The seabed acoustic parameter inversion method, referred to as the geoacoustic inversion method, obtains the seabed information in the received sound field through signal processing algorithms. Its main principle is to use the inversion method to estimate the seabed acoustic parameters, such as the sound velocity, density, and attenuation coefficient of the sediment layer, based on the propagation characteristics of the underwater acoustic signal according to the measured sound field. These parameters play a very important role in the prediction of sound propagation loss and the evaluation of sonar performance.

[0003] Matched field inversion is the most commonly used geoacoustic inversion method. Its basic principle is to match the receiving sound field generated by the acoustic propagation model with the measured sound field, and use the optimization algorithm to search for the seabed acoustic parameters with the highest matching degree in the high-dimensional parameter space. The acoustic propagation model is a numerical calculation model that describes the propagation characteristics of underwater acoustic signals. The receiving sound field generated by it is usually called the copy sound field. In the geoacoustic inversion problem, there are many parameters to be inverted, the search space is large and complex, and it increases geometrically with the increase of parameters, so that the matched field inversion needs to call the time-consuming acoustic propagation model multiple times, and it takes a long time to obtain the global optimal solution, and the inversion efficiency is low. When the seabed acoustic parameters are related to the distance, it is necessary to estimate them in real time based on the measured sound field obtained during navigation, so as to achieve fast and accurate marine environment modeling for further hydroacoustic applications. In such cases where there is a real-time requirement for geoacoustic inversion, it is particularly important to solve the problem of long matching field inversion time. Summary of the invention

[0004] The present invention aims to overcome the time-consuming shortcoming of traditional matching field inversion and provide a fast matching field seabed acoustic parameter inversion method based on machine learning.

[0005] The present invention introduces a pre-screening process based on machine learning into the traditional matching field inversion, which greatly reduces the inversion time without affecting the inversion accuracy. It includes the following steps:

[0006] S1. Obtain data set;

[0007] According to the prior information of the environment and the transceiver configuration of the experimental sea area, an acoustic propagation model is selected to establish a simulation environment, and the copy sound field is calculated under different seabed acoustic parameters. The former is the input of the machine learning model, recorded as features, and the latter is the model prediction reference, recorded as labels, to establish a data set;

[0008] S2, training and testing machine learning models;

[0009] Construct an objective function to measure the matching degree between the copied sound field and the measured sound field, convert the labels of the data set into the objective function values, and use the new data set to build a machine learning model for training and testing;

[0010] S3, introduce machine learning models into traditional matching field inversion;

[0011] In the traditional matching field inversion process, an optimization algorithm is used to search for the global optimum in the high-dimensional parameter space. The parameter update process is:

[0012]

[0013] The acceptance condition is determined by the optimization algorithm. m and m′ represent the seabed acoustic parameters before and after the disturbance. The objective function values ​​f and f′ before and after the disturbance are calculated based on the acoustic propagation model. If Δf=f′-f satisfies the acceptance condition, the parameters are updated to the parameters after the disturbance, otherwise the parameters remain unchanged. After the machine learning model is introduced, the parameter update process becomes

[0014]

[0015] Among them, for the seabed acoustic parameters m and m′ before and after the disturbance, the objective function values ​​F and F′ before and after the disturbance are first quickly calculated based on the machine learning model, and pre-screening is performed according to ΔF=F′-F and the acceptance condition. If the pre-screening passes, the objective function values ​​f and f′ before and after the disturbance are further accurately calculated based on the acoustic propagation model. If Δf=f′-f satisfies the acceptance condition, it is updated to the parameters after the disturbance, otherwise the parameters remain unchanged;

[0016] S4, estimate seafloor acoustic parameters;

[0017] According to the measured acoustic field and parameter search interval, the global optimal parameters are estimated using a fast matching field seafloor acoustic parameter inversion method based on machine learning.

[0018] Preferably, in the matching field inversion, the objective function in step S2 is usually a phone-coherent objective function or a frequency-coherent objective function to measure the matching degree between the copied sound field and the measured sound field. When the spectrum of the transmitted signal is unknown, the phone-coherent objective function is usually used:

[0019]

[0020] Alternatively, when the uncertainty of the array element position is large, the objective function described in step S2 usually uses a frequency-coherent objective function that is insensitive to array shape mismatch:

[0021]

[0022] Where m=[m1 m2 ... m K ] T represents the parameter vector composed of K seafloor acoustic parameters, (·) T and(·) H denote transpose and conjugate transpose respectively. N F and N H are the frequency of the received sound field and the number of received array elements respectively. Let d n (f i ) indicates that the frequency at the nth array element is f i The measured sound pressure value, YesN H The array element is at frequency f i The measured sound pressure vector at is the nth array element in N F The measured sound pressure vector at the frequency. Let g n (f i , m) means that for a given parameter vector m, the frequency at the nth array element is f i The copy sound pressure value, is a given parameter vector m, N H The array element is at frequency f i The copy sound pressure vector at Given a parameter vector m, the nth element in N F The copied sound pressure vector at the frequency.

[0023] In the traditional matching field inversion process, the present invention utilizes machine learning to fit the objective function based on the acoustic propagation model for rapid parameter pre-screening, thereby reducing the number of calls to the acoustic propagation model when updating parameters. Without affecting the inversion accuracy, the inversion time is greatly reduced, thereby improving the ability to rapidly analyze seabed characteristics.

[0024] Compared with the traditional matching field inversion method, the present invention is characterized by:

[0025] 1) Use machine learning models to accelerate matching field inversion. The time-consuming traditional matching field inversion mainly comes from multiple calls to the acoustic propagation model. For multiple sets of seabed acoustic parameters, the acoustic propagation model needs to calculate the objective function value in sequence and each calculation is relatively time-consuming. Accelerating the calculation of the objective function value can effectively reduce the inversion time. Since the forward mapping from seabed acoustic parameters to objective function values ​​satisfies uniqueness and it is easy to obtain a large number of data sets from the simulation environment, the machine learning model is suitable for fitting the mapping relationship. The machine learning model is efficient in calculation and supports parallel input of multiple sets of parameters. It can also be further accelerated by a graphics processing unit (GPU), which can quickly calculate the objective function value and is used to accelerate matching field inversion.

[0026] 2) Use machine learning models for pre-screening to reduce the impact on inversion accuracy. Traditional matching field inversion calculations are time-consuming. If a machine learning model is used instead of the acoustic propagation model of formula (1) to calculate the objective function value, the inversion speed is improved, but the calculation error of the objective function value will affect the inversion accuracy. In order to achieve accelerated inversion without affecting the inversion accuracy, the present invention first uses a machine learning model to quickly calculate the objective function value of the parameters before and after the disturbance in the parameter update process of the matching field inversion. If the difference between the two meets the acceptance condition, the parameters are updated using formula (1), otherwise the parameters remain unchanged, as shown in formula (2). In the new parameter update process, the first step is called a pre-screening process based on machine learning, which can quickly reject invalid parameter disturbances. If the pre-screening is passed, it is regarded as a valid disturbance. Only valid disturbances need to call the acoustic propagation model to calculate the objective function value for judging acceptance or rejection. The pre-screening process based on machine learning reduces the number of calls to the acoustic propagation model, thereby reducing the time consumption of inversion. The further judgment of the acoustic propagation model guarantees the inversion accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is an overall flow chart of the method of the present invention.

[0028] Figure 2 This is a schematic diagram of the experimental environment.

[0029] Figure 3 It is a schematic diagram of the machine learning model structure.

[0030] Figure 4 It is a curve chart of the loss value changes on the training set and the test set.

[0031] Figure 5 It is a distribution diagram of the objective function values ​​of various parameters under the rapid matching field seabed acoustic parameter inversion method based on machine learning and the traditional matching field inversion method. DETAILED DESCRIPTION

[0032] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0033] Reference Figure 1 ,The specific implementation steps of the fast matching field seabed acoustic parameter inversion method based on machine learning are as follows:

[0034] S1. Obtain data set;

[0035] Assume that the experimental environment is Figure 2 As shown in Figure 1, before the experiment, the acoustic propagation model was selected to establish the simulation environment based on the known prior information of the water body, the prior information of the seabed, and the geometric position relationship between the sound source and the horizontal linear array (HLA) in the experimental sea area. The various seabed acoustic parameters to be estimated include the thickness of the sediment layer hs , speed of sound c s , density ρ s , attenuation coefficient α s and the base sound velocity c b , density ρ b , respectively set the search interval and uniformly collect parameter vectors. In the simulation environment, each set of seabed acoustic parameter vectors corresponds to a set of copied sound fields, with the former as features and the latter as labels to form a data set.

[0036] S2, training and testing machine learning models;

[0037] In matching field inversion, the array element coherence objective function or frequency coherence objective function is usually selected to measure the matching degree between the copied sound field and the measured sound field. When the transmitted signal spectrum is unknown, the array element coherence objective function is usually used.

[0038]

[0039] When the uncertainty of array element positions is large, a frequency-coherent objective function that is insensitive to array shape mismatch is usually used.

[0040]

[0041] Where m=[m1 m2 ... m K ] T represents the parameter vector composed of K seafloor acoustic parameters, (·) T and(·) H denote transpose and conjugate transpose respectively. N F and N H are the frequency of the received sound field and the number of received array elements respectively. Let d n (f i ) indicates that the frequency at the nth array element is f i The measured sound pressure value, YesN H The array element is at frequency f i The measured sound pressure vector at is the nth array element in N F The measured sound pressure vector at the frequency. Let g n (f i , m) means that for a given parameter vector m, the frequency at the nth array element is f i The copy sound pressure value, is a given parameter vector m, N H The array element is at frequency f i The copy sound pressure vector at Given a parameter vector m, the nth element in N F The copied sound pressure vector at the frequency. Figure 2In the experimental environment, the spectrum of the transmitted signal is unknown but the array position information is known, so the array element coherence objective function is selected. For the measured sound field with a signal-to-noise ratio of 20dB, the label of the data set is converted into the objective function value using formula (3), thereby forming a new data set. The data set is divided into a training set and a test set in a ratio of 4:1. The former is used to train the machine learning model, and the latter is used to evaluate the generalization ability of the model and adjust the hyperparameters. The feature m in the data set is converted to the label E PC The forward mapping of (m) is regarded as a regression problem. Considering the size of the data set, the calculation speed and the fitting accuracy, we construct Figure 3 The Deep Neural Network (DNN) shown in Figure 1 is used to divide the training set into multiple batches and input them into the DNN for forward propagation. The mean squared error (MSE) is used to calculate the loss value. After reverse gradient propagation, the adaptive moment estimation (Adam) is used to update the DNN parameters until convergence. The loss value changes on the training set and the test set are shown in Figure 11. Figure 4 As shown in the figure, the loss values ​​on the training set and the test set continue to decrease with the increase in the number of training times, and finally converge to a lower value. The changes of the two are basically the same, but the former is always slightly lower than the latter, indicating that the fitting and generalization effects of DNN are good. As an approximation of the forward mapping, DNN always has calculation errors, but it supports parallel calculation of multiple sets of parameters and GPU acceleration, and the calculation speed is much faster than the former.

[0042] S3, introduce machine learning models into traditional matching field inversion;

[0043] The basic principle of matched field inversion is to match the copy sound field generated by the acoustic propagation model with the measured sound field, and use the optimization algorithm to search for the seabed acoustic parameters with the highest matching degree in the high-dimensional parameter space. This process is equivalent to starting from a random initial point in the high-dimensional parameter space and continuously updating the parameters until they converge to the global optimal point. The parameter update process of traditional matched field inversion is

[0044]

[0045] Among them, m and m′ represent the acoustic parameters of the seabed before and after the disturbance. The objective function values ​​f and f′ before and after the disturbance are calculated based on acoustic propagation. If Δf=f′-f satisfies the acceptance condition, the parameters are updated to the parameters after the disturbance, otherwise the parameters remain unchanged. In order to quickly converge to the global optimum, the hybrid optimization algorithm commonly used in matching field inversion, Adaptive Simplex Simulated Annealing (ASSA) algorithm, is used this time. This optimization algorithm uses the Metropolis criterion as the acceptance condition. If Δf<0 or e -Δf / T>ε, the disturbance is accepted, otherwise the parameters remain unchanged, where T represents the annealing temperature of the ASSA algorithm, and the random number ε~U(0,1). By introducing the pre-screening process based on machine learning, the parameter update process becomes

[0046]

[0047] Among them, for the seabed acoustic parameters m and m′ before and after the disturbance, the objective function values ​​F and F′ before and after the disturbance are first quickly calculated based on the machine learning model, and pre-screening is performed according to ΔF=F′-F and the Metropolis criterion. If the pre-screening passes, the objective function values ​​f and f′ before and after the disturbance are further accurately calculated based on the acoustic propagation model. If Δf=f′-f satisfies the Metropolis criterion, it is updated to the parameters after the disturbance, otherwise the parameters remain unchanged. The introduction of pre-screening processing reduces the number of calls to the acoustic propagation model, thereby reducing the time consumption of inversion, and further judgment of the acoustic propagation model ensures the accuracy of the inversion.

[0048] S4, estimate seafloor acoustic parameters;

[0049] According to the measured acoustic field and parameter search interval, the distribution of the objective function values ​​of each parameter under the fast matching field seabed acoustic parameter inversion method based on machine learning and the traditional matching field inversion method is as follows: Figure 5 As shown in the figure, the former is named DNN-assisted ASSA, and the latter is named ASSA. The time consumption and estimation results of the two types of inversion methods are shown in Table 1. The estimation results are basically the same, but the time consumption of the former is reduced by 55% compared with the latter. Therefore, compared with the traditional matching field inversion method, the present invention has similar inversion accuracy, but the inversion time consumption is greatly reduced, which can improve the ability of rapid seabed characteristic analysis.

[0050] Table 1 Time consumption and estimation results of two types of inversion methods

[0051]

[0052] The present invention provides a method for rapid matching field seabed acoustic parameter inversion based on machine learning. First, according to the environmental prior information and transceiver configuration of the experimental sea area, an acoustic propagation model is selected to establish a simulation environment; then, the target function value is calculated in combination with the measured sound field under different seabed acoustic parameters, and a data set is established based on the forward mapping for training and testing the machine learning model; finally, in the parameter update process of the matching field inversion, for the parameters before and after the disturbance, the machine learning model is first used to quickly screen the effective disturbance, and then the acoustic propagation model is used to judge whether to accept or reject it, and after multiple rounds of parameter updates, it converges to the global optimum. The present invention introduces a pre-screening process based on machine learning in the traditional matching field inversion process, reduces the number of calls to the acoustic propagation model, and thus reduces the time consumption of the inversion, and the further judgment of the acoustic propagation model ensures the inversion accuracy, which can improve the ability of rapid seabed characteristic analysis.

[0053] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for rapid matching field seafloor acoustic parameter inversion based on machine learning, comprising the following steps: S1. Obtain data set; According to the experimental prior information, the acoustic propagation model is selected to establish the simulation environment, and the receiving sound field, also known as the copy sound field, is calculated under different seabed acoustic parameters. The former is the input of the machine learning model, recorded as features, and the latter is the model prediction reference, recorded as labels, to establish a data set; S2, training and testing machine learning models; Construct an objective function to measure the matching degree between the copied sound field and the measured sound field, convert the labels of the data set into the objective function values, and use the new data set to build a machine learning model for training and testing; S3. Introducing machine learning models into traditional matching field inversion In the traditional matching field inversion process, an optimization algorithm is used to search for the global optimum in the high-dimensional parameter space. The parameter update process is: Among them, the Metropolis criterion is used as the acceptance condition; m and m′ represent the seabed acoustic parameters before and after the disturbance. The objective function values ​​f and f′ before and after the disturbance are calculated based on the acoustic propagation model. If Δf=f′-f satisfies the acceptance condition, the parameters are updated to the parameters after the disturbance, otherwise the parameters remain unchanged. After the machine learning model is introduced, the parameter update process becomes Among them, for the seabed acoustic parameters m and m′ before and after the disturbance, the objective function values ​​F and F′ before and after the disturbance are first quickly calculated based on the machine learning model, and pre-screening is performed according to ΔF=F′-F and the acceptance condition. If the pre-screening passes, the objective function values ​​f and f′ before and after the disturbance are further accurately calculated based on the acoustic propagation model. If Δf=f′-f satisfies the acceptance condition, it is updated to the parameters after the disturbance, otherwise the parameters remain unchanged; S4, estimate seafloor acoustic parameters; According to the measured acoustic field and parameter search interval, the global optimal parameters are estimated using a fast matching field seafloor acoustic parameter inversion method based on machine learning.

2. The method for rapid matching field seabed acoustic parameter inversion based on machine learning according to claim 1, characterized in that: When the spectrum of the transmitted signal is unknown, the objective function described in step S2 uses a phone-coherent objective function to measure the matching degree between the copied sound field and the measured sound field: Where m = [m1 m2 ... m K ] T represents the parameter vector composed of K seafloor acoustic parameters, (·) T and(·) H denote transpose and conjugate transpose respectively, N F and N H are the frequency of the received sound field and the number of received array elements respectively, let d n (f i ) indicates that the frequency at the nth array element is f i The measured sound pressure value, YesN H The array element is at frequency f i The measured sound pressure vector at g n (f i ,m) means that for a given parameter vector m, the frequency at the nth array element is f i The copy sound pressure value, is a given parameter vector m, N H The array element is at frequency f i The copied sound pressure vector at .

3. The method for rapid matching field seabed acoustic parameter inversion based on machine learning according to claim 2, characterized in that: When the uncertainty of the array element position is large, the objective function described in step S2 uses a frequency-coherent objective function that is insensitive to array shape mismatch to measure the matching degree between the copied sound field and the measured sound field: in, is the nth array element in N F The measured sound pressure vector at the frequency, Given a parameter vector m, the nth element in N F The copied sound pressure vector at the frequency.

Images