CN116299665A - LSTM surface wave inversion method, device and medium - Google Patents

Abstract

The invention discloses an LSTM surface wave inversion method, device and medium. The method includes determining the fuzzy stratum parameter interval of the work area to be inverted; randomly generating different models according to the fuzzy stratum parameter interval, and using the generalized reflection-transmission coefficient The theoretical surface wave dispersion curve of each model is calculated by the method, and the training sample data pair is constructed; the training sample data pair is preprocessed; the LSTM network based on the FHLV loss function is constructed, and the training sample data pair is based on the preprocessed training sample data pair. The LSTM network is trained and the trained model is saved; the unsupervised learning is used to automatically pick up the dispersion curve of the actual dispersion imaging data, and the trained model is used to predict the dispersion curve after the uniform dimension size, and the near-surface Shear wave velocity model. The invention improves processing efficiency and inversion accuracy, and is suitable for large-scale data processing.

Description

LSTM surface wave inversion method, device and medium

technical field

The invention belongs to the field of seismic data processing for near-surface exploration, and in particular relates to an LSTM surface wave inversion method, device and medium.

Background technique

The surface wave analysis method is widely used to establish the near-surface shear wave velocity structure. Regardless of whether it is an active source or a passive source surface wave analysis method, the near-surface S-wave velocity structure is obtained by inverting the surface wave dispersion curve. The inversion of surface wave dispersion curve is a typical highly nonlinear, multi-parameter, multi-extreme geophysical inversion problem. The commonly used surface wave dispersion curve inversion methods are roughly divided into two categories, one is local linearization methods, such as least squares method, damped least squares inversion, etc.; due to the forward modeling of Rayleigh waves in layered media The dispersion equation is a nonlinear function, so when the initial model is not selected, it is difficult for this kind of local linearization method to find the global optimal solution of the objective function. The other is the global nonlinear optimization method, commonly used are genetic algorithm and simulated annealing inversion. This type of algorithm can avoid the local linearization inversion method's dependence on the initial model to a certain extent, but in practical applications, the local search ability is not strong and it takes a long time. Therefore, surface wave inversion methods need to be explored in new fields.

At present, various traditional surface wave dispersion curve inversion methods have certain limitations. Deep learning has the ability to solve many nonlinear problems. In recent years, machine learning and deep learning have shown great potential when applied to various geophysical research problems, providing automation performance in some tasks. Therefore, in order to solve the problems of low inversion efficiency and poor inversion effect in traditional surface wave exploration, a LSTM surface wave inversion method based on FHLV loss function is proposed, which improves the processing efficiency and inversion accuracy, and is suitable for large-scale Scale data processing.

Contents of the invention

The purpose of the present invention is to overcome the technical problems proposed by the background technology, and propose an LSTM surface wave inversion method, device and medium. Specifically, it is a situation where a single data demander generates a perceived data demand, and multiple data owners compete to participate in the sharing task qualification. In this method, blockchain technology is used to solve the trust problem brought by a trusted third party. The incentive mechanism is designed based on the reverse auction model to help miners screen out data owners with irrational quotations, reduce the workload of subsequent verification of data quality levels, and improve the performance of the incentive model. A softmax regression algorithm is used to calculate the quality level of the perceived data. Finally, the value of the data is calculated through the data owner's quotation and the data quality level, and the rewards are distributed according to different data values, so as to encourage data owners to upload reasonable-priced, high-quality and reliable data.

Concrete technical scheme of the present invention is as follows:

According to a first aspect of the present invention, an LSTM surface wave inversion method is provided, the method comprising:

Determine the range of fuzzy formation parameters in the work area to be inverted;

Randomly generate different models according to the fuzzy formation parameter interval, calculate the theoretical surface wave dispersion curve of each model by using the generalized reflection-transmission coefficient method, and construct the training sample data pair;

Preprocessing the training sample data pair;

Construct an LSTM network based on the FHLV loss function, train the LSTM network and save the trained model based on the preprocessed training sample data pair;

The unsupervised learning is used to automatically pick up the dispersion curve from the actual dispersion imaging data, and the trained model is used to predict the dispersion curve after the unified dimension, and the near-surface shear wave velocity model is obtained.

Further, the range of fuzzy formation parameters in the work area to be inverted is determined by using known logging information or by existing inversion methods.

Further, formation parameters include formation shear wave velocity, formation longitudinal wave velocity, formation thickness and formation density of each layer, wherein the formation compression wave velocity is 2.4-3 times of formation shear wave velocity in the same layer, and formation density is 1.5-2Kg/ m ³ , the fuzzy formation parameter interval is the fuzzy interval range of each layer of formation parameters, which is determined by the minimum and maximum values of each layer parameter in the extracted one-dimensional velocity structure at different positions, and the starting point of the fuzzy formation parameter interval is The range is 40%-80% of the minimum value, and the end range is 120%-140% of the maximum value. The larger the value of the formation parameter, the smaller the ratio of the corresponding start range and end range. The one-dimensional velocity structure of different positions Obtained based on known logging information or existing inversion methods.

Further, the preprocessing of the training sample data pair includes:

The training sample data pair is normalized, and the formation shear wave velocity and formation thickness in the training sample data pair are column vectorized.

Further, the LSTM network based on the FHLV loss function includes three layers of hidden layers, each layer of hidden layers corresponds to 256 LSTM units, skip connections are used between each LSTM unit, and a full connection is connected to the output of the last LSTM unit layer to control the output dimensions.

Further, based on the preprocessed training sample data pair, the LSTM network is trained and the trained model is saved, including:

The theoretical surface wave dispersion curve is input into the LSTM network, and the corresponding formation shear wave velocity and formation thickness are output to complete the training of the LSTM network;

The loss function of the LSTM network is:

Loss1=mean(abs(v_label-v))

Loss2=mean(abs(h_label-h))

Among them, Loss1 is the speed loss, Loos2 is the thickness loss, FHLV is the actual loss function during training, label is the label, mean is the average operator, abs is the absolute value operator, v is the formation shear wave velocity, h is the formation thickness, e is Ideal velocity loss, set to 15-20% of the normalized minimum layer velocity, wh is the weight decay coefficient, set to 0.01-0.1.

Further, the use of unsupervised learning to automatically pick up the dispersion curve of the actual dispersion imaging data includes:

Determine the surface wave dispersion energy matrix according to the actual dispersion imaging data;

Normalizing the surface wave dispersion energy matrix, using the coordinates of points greater than the first dispersion energy threshold as the first dispersion energy point;

clustering the first dispersion energy points;

The dispersion energy corresponding to the clustered points is compared, and the phase velocity of the frequency vector corresponding to the local peak is screened to obtain the final dispersion curve.

According to a second aspect of the present invention, an LSTM surface wave inversion device is provided, the device includes a processor, and the processor is configured to:

Determine the range of fuzzy formation parameters in the work area to be inverted;

Randomly generate different models according to the fuzzy formation parameter interval, calculate the theoretical surface wave dispersion curve of each model by using the generalized reflection-transmission coefficient method, and construct the training sample data pair;

Preprocessing the training sample data pair;

Construct an LSTM network based on the FHLV loss function, train the LSTM network and save the trained model based on the preprocessed training sample data pair;

The unsupervised learning is used to automatically pick up the dispersion curve from the actual dispersion imaging data, and the trained model is used to predict the dispersion curve after the unified dimension, and the near-surface shear wave velocity model is obtained.

Further, formation parameters include formation shear wave velocity, formation longitudinal wave velocity, formation thickness and formation density of each layer, wherein the formation compression wave velocity is 2.4-3 times of formation shear wave velocity in the same layer, and formation density is 1.5-2Kg/ m ³ , the fuzzy formation parameter interval is the fuzzy interval range of each layer of formation parameters, which is determined by the minimum and maximum values of each layer parameter in the extracted one-dimensional velocity structure at different positions, and the starting point of the fuzzy formation parameter interval is The range is 40%-80% of the minimum value, and the end range is 120%-140% of the maximum value. The larger the value of the formation parameter, the smaller the ratio of the corresponding start range and end range. The one-dimensional velocity structure of different positions Obtained based on known logging information or existing inversion methods.

Further, the LSTM network based on the FHLV loss function includes three layers of hidden layers, each layer of hidden layers corresponds to 256 LSTM units, skip connections are used between each LSTM unit, and a full connection is connected to the output of the last LSTM unit layer to control the output dimensions.

Further, the processor is further configured to:

The theoretical surface wave dispersion curve is input into the LSTM network, and the corresponding formation shear wave velocity and formation thickness are output to complete the training of the LSTM network;

The loss function of the LSTM network is:

Loss1=mean(abs(v_label-v))

Loss2=mean(abs(h_label-h))

Among them, Loss1 is the speed loss, Loos2 is the thickness loss, FHLV is the actual loss function during training, label is the label, mean is the average operator, abs is the absolute value operator, v is the formation shear wave velocity, h is the formation thickness, e is Ideal velocity loss, set to 15-20% of the normalized minimum layer velocity, wh is the weight decay coefficient, set to 0.01-0.1.

Further, the processor is further configured to:

Determine the surface wave dispersion energy matrix according to the actual dispersion imaging data;

Normalizing the surface wave dispersion energy matrix, using the coordinates of points greater than the first dispersion energy threshold as the first dispersion energy point;

clustering the first dispersion energy points;

The dispersion energy corresponding to the clustered points is compared, and the phase velocity of the frequency vector corresponding to the local peak is screened to obtain the final dispersion curve.

According to a third aspect of the present invention, there is provided a computer-readable storage medium, on which computer-readable instructions are stored, and when the computer-readable instructions are executed by a processor of a computer, the computer is made to perform various implementations of the present invention. The LSTM surface wave inversion method described in the example.

The LSTM surface wave inversion method, device and medium provided by various embodiments of the present invention improve processing efficiency and inversion accuracy, and are suitable for large-scale data processing. The formation in the same area has continuity, and the inversion of the current dispersion curve can provide a reference for subsequent inversion of adjacent dispersion curves, and the LSTM network has memory characteristics, which is suitable for feature extraction of one-dimensional data, so the LSTM network is used to solve the problem of surface wave The dispersion curve inversion problem has good applicability. The LSTM network can learn the nonlinear mapping between the surface wave dispersion curve and the near-surface shear wave velocity structure by mining a large number of data features. Thickness is not sensitive, and it is often difficult for networks based on conventional loss functions to accurately predict shear wave velocity and formation thickness at the same time. For this reason, the FHLV loss function is designed to improve the learning ability of the network for thickness parameters. This loss function can maintain the optimal balance of formation velocity and formation thickness, so that the network converges better and the inversion accuracy is higher. The LSTM based on the FHLV loss function is formed. The surface wave inversion method finally obtains the high-precision near-surface shear wave velocity structure.

Description of drawings

In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings that need to be used in the description of the specific embodiments or the prior art. Throughout the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, elements or parts are not necessarily drawn in actual scale.

Fig. 1 is a processing flowchart of the LSTM surface wave inversion method based on the FHLV loss function.

Figure 2 is a schematic diagram of a theoretical two-dimensional velocity model composed of 2000 one-dimensional velocity structures.

Fig. 3 is a schematic diagram of the model, wherein (a)-(c) respectively represent a schematic diagram of a real two-dimensional shear wave velocity model, a schematic diagram of a two-dimensional shear wave velocity model predicted by MAE, and a schematic diagram of a two-dimensional shear wave velocity model predicted by the present invention.

Fig. 4 is a schematic diagram of predicted absolute difference, wherein (a) represents the predicted absolute difference of MAE, and (b) represents the predicted absolute difference of the present invention.

Fig. 5 is the comparison figure of MAE, the one-dimensional velocity structure predicted by the present invention and the real velocity structure, (a) is the 420th track inversion result comparison figure, (b) is the 1000th track inversion result comparison figure, (c ) is the comparison chart of the 1000th track inversion results.

Figure 6 is a schematic diagram of the actual data, where (a) represents the positive offset part of a single shot, and (b) represents the corresponding high-resolution dispersion image.

Fig. 7 is the comparison chart of the forward modeling dispersion curve and the observation dispersion curve of the inversion result of the present invention, wherein (a) is the comparison chart of the 60th track inversion result, (b) is the comparison chart of the 100th track inversion result, (c ) is the comparison chart of the 260th track inversion results.

Fig. 8 is a verification diagram of the inversion effect, wherein (a) is a schematic diagram of the two-dimensional shear wave velocity model obtained by the inversion of the present invention, (b) is a comparison diagram between the effect of the present invention and conventional methods and logging data, and (c) is A Schematic diagram of drilling stratification in the region.

Fig. 9 is the LSTM deep neural network structure used for surface wave dispersion curve inversion in the present invention, X0, ... Xn are the dispersion curve points, H0, ... Hn are the output results of the LSTM network, P1, ... Pn are after the control dimension The network output of , Pn represents the predicted model parameters, and the dimensionality change is explained on the left.

Fig. 10 is a flowchart of unsupervised learning of the present invention to automatically pick up dispersion curves, wherein (a) represents a normalized dispersion energy map, (b) represents a point with a larger energy of interest, and (c) represents a point after clustering, (d) shows the energy peak point corresponding to the frequency vector picked up, and (e) shows the final picture of the picked-up effect.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below, obviously, the described embodiments are only some of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

In order to make the purpose, technical solution and advantages of the present invention more clear, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the invention.

The present invention will be described further in conjunction with accompanying drawing now.

Due to its low-cost and non-destructive characteristics, the surface wave imaging method has shown great potential in urban near-surface exploration, and has attracted extensive attention from geophysicists. Its core is to construct an accurate near-surface shear-wave velocity model, and the accurate inversion of the surface-wave dispersion curve is undoubtedly the basis for the construction of a near-surface shear-wave velocity model. Here, a surface wave inversion method through the fusion of deep learning network and improved loss function is provided, specifically an LSTM surface wave inversion method.

The basic technical principle of the present invention is: use the known logging information or through the conventional inversion method to determine the fuzzy formation parameter interval of the work area to be inverted, randomly generate a large number of different models in the fuzzy interval of the formation model parameters, and use the generalized reflection-transmission coefficient The corresponding theoretical surface wave dispersion curve is calculated by using the method, so as to construct a large number of sample data pairs. Through the LSTM network based on the FHLV loss function, the sample data pairs are trained to learn the nonlinear mapping between the surface wave dispersion curve and the near-surface shear wave velocity structure, and the inversion data is predicted to obtain an accurate near-surface shear wave velocity model. It is difficult for the network based on the conventional loss function to learn the characteristics of the thickness parameter to make it fully converge, resulting in inaccurate inversion of the thickness parameter of the multi-layer model and the fine-layer model. The FHLV loss function is the core of the proposed method, which consists of two parts: velocity loss and thickness loss. It optimizes the learning of thickness parameters by the network, thereby improving the overall prediction accuracy and better solving the problem of poor thickness prediction. Finally, an LSTM surface wave inversion method based on the FHLV loss function was formed to obtain a high-precision near-surface shear wave velocity structure.

Please refer to Figure 1, which is a flowchart of an LSTM surface wave inversion method. The specific processing steps of this method are divided into the following steps:

1) Use known logging information or conventional inversion methods to determine the range of fuzzy formation parameters in the work area to be inverted.

Set the number of layers of the stratum model. The determined stratum parameters are mainly the stratum shear wave velocity VS and the stratum thickness H of each layer. The corresponding stratum longitudinal wave velocity VP is set to 2.4-3 times VS, and the stratum density is set to 1.5-2Kg/m3. The range of the fuzzy interval of stratigraphic parameters in each layer is determined by the minimum and maximum values of each layer parameter in the extracted one-dimensional velocity structure at different positions, the starting range is 40%-80% of the minimum value, and the end range is 120% of the maximum value %-140%, the larger the formation parameter value, the smaller the ratio, and the one-dimensional velocity structure at different positions corresponds to the known logging information or the inversion results of other methods.

2) A large number of different models are randomly generated in the fuzzy interval of formation model parameters, and the corresponding theoretical surface wave dispersion curves are calculated by the generalized reflection-transmission coefficient method, thereby constructing a large number of sample data pairs;

3) Construct an LSTM network based on the FHLV loss function, preprocess a large number of sample data pairs generated and send them to the network for training, and save the trained model;

The LSTM based on the FHLV loss function takes the LSTM unit module as the main body, and there are three hidden layers in total, each layer corresponds to 256 LSTM units, with skip connections between, and a fully connected layer connected to the output of the last LSTM unit Control the output dimension; input the surface wave dispersion curve into the LSTM, output the corresponding formation model shear wave velocity and formation thickness, and complete the training of the LSTM; during the training process, the error between the output formation model parameters and the sample label is determined by the FHLV The loss function is defined by the loss function. The FHLV loss function is the core of the proposed method. It consists of two parts: the speed loss Loss1 and the thickness loss Loss2. It optimizes the learning of the thickness parameter by the network, thereby improving the overall prediction accuracy. It is characterized by the following:

Loss1=mean(abs(v_label-v))

Loss2=mean(abs(h_label-h))

label is the label, mean is the average operator, abs is the absolute value operator, v is the formation shear wave velocity, h is the formation thickness, e is the ideal velocity loss, usually set to 15-20% of the normalized minimum layer velocity, wh is the weight decay coefficient, usually set to 0.01-0.1.

4) Use the dispersion imaging method to obtain the actual surface wave dispersion curve, and use the unsupervised clustering method to automatically pick up the dispersion curve, and then use the trained neural network to perform inversion to obtain the near-surface shear wave velocity model.

Since picking up the dispersion curve is to extract the energy peak of the corresponding frequency vector from the perspective of human thinking, the density-based DBSCAN clustering method is mainly used to automatically pick up the dispersion curve, and the surface wave dispersion energy matrix is mainly normalized to select Among them, the coordinates of the points with larger energy are clustered by using the density-based DBSCAN clustering method to select the coordinates of the dispersed energy points, and then the dispersed energy corresponding to the clustered points is compared, and the frequency vector corresponding to the local peak value is screened. phase velocity to get the final dispersion curve.

Through the processing of the above specific steps, the difficult problem of using deep learning to perform high-precision surface wave inversion has been realized.

In order to verify the LSTM surface wave inversion method and its inversion effect based on the FHLV loss function, the training process and inversion results of the theoretical two-dimensional velocity model and the seismic surface wave data of urban near-surface exploration in a western region A are taken as examples. for analysis.

Figure 2 is a 6-layer theoretical two-dimensional shear wave velocity model composed of 2000 one-dimensional velocity structures. Extract 5 channels of one-dimensional velocity structure in the model (the corresponding logging data in the actual work area or inversion results of conventional methods), and determine the range of model parameters. The model parameter ranges of its training set and test set are shown in Table 1. 60,000 sets of model parameters were randomly generated within its interval, and the corresponding fundamental-order dispersion curves were calculated to obtain 60,000 pairs of samples. Under the same number of iterations, MAE loss function and FHLV loss function are used for training and prediction respectively. All 60,000 pairs of samples are used for training, and the dispersion curves corresponding to 2,000 one-dimensional velocity structures in the two-dimensional model are used for testing.

Table 1 Parameter ranges of test set and training set of low-speed thin-layer theory model

Figure 3 below shows the 2000 channel prediction results of the two networks respectively interpolated to form the two-dimensional near-surface shear wave velocity model of the network inversion. It can be observed that the inversion results of the two networks basically reflect the change trend of the model. In order to compare the inversion effect, it can be observed from the inversion error graphs in Figures 4(a) and 4(b) that the model inverted by the network based on the MAE loss function is quite different from the original model in some local places (especially located in the red box). In terms of the overall inversion effect, the absolute difference between the inversion result based on the FHLV loss function of the present invention and the real model is smaller, the performance is more consistent with the original model, and the inversion accuracy is higher. In order to better compare the details of the low-velocity thin-bed model, three inverted one-dimensional velocity structures were extracted from the test set for comparison with the model velocity structure, and their positions are located at the red lines I, II, and III in Fig. 3(a). From Figure 5(a)-(c), it can be found that the network inversion results based on the FHLV loss function of the present invention have achieved better results on the whole, especially the thickness parameter prediction accuracy is higher.

In order to further illustrate the effect of the method, the seismic surface wave data of urban near-surface exploration in an area A in the west are selected. Figure 6(q) shows a typical single-shot record, in which obvious surface wave information can be seen. Figure 6(b) shows the corresponding dispersion image, from which the distinct fundamental-order dispersion energy can be identified.

The near-surface shear-wave velocity structure within 1.2 km of the collected data was inverted, and a total of 300 sets of dispersion curves were picked up. The initial model was set according to the stratification ratio method, and the dispersion curves at five different positions were inverted to determine the range of formation parameters. The specific parameter ranges are shown in Table 2. Randomly generate 50,000 sets of model parameter structures within a given interval, calculate the corresponding theoretical dispersion curve, and form a large amount of training data. As shown in Figure 7, the picked 300 sets of dispersion curves are predicted, and the forward modeling data of the inversion results are in good agreement with the observation data, and the correlation is greater than 92%. Since each inverted 1D shear wave velocity structure reflects the subterranean structure below the receiver, the 1D inversion results of all the 300 picked dispersion curves are arranged on the corresponding coordinates, and interpolated and smoothed to obtain the final 2D Shear wave velocity model. As shown in Figure 8, it shows the inversion results shallower than 20m. Combined with the existing drilling data in this work area (the well position is located at the red line in Figure 8(a)), it can be compared with the damped least squares method inversion results. It is found that the network inversion results are basically consistent with the actual near-surface conditions, which further proves the accuracy and practicability of the automatic inversion method proposed by the present invention.

Table 2A Region actual data training set parameter range

Through theoretical data and two data experiments in area A, it is found that using the LSTM surface wave inversion method based on the FHLV loss function can obtain a more accurate near-surface shear wave velocity structure. Compared with conventional methods and logging data, it is found that the research method can effectively improve the inversion accuracy.

The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it can still be applied to the foregoing embodiments Modifications to the technical solutions described in the examples, or equivalent replacement of some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention, and they shall cover Within the scope of the claims and description of the present invention.

Claims (10)

1. A method of LSTM face wave inversion, the method comprising:

determining a fuzzy stratum parameter interval of a work area to be inverted;

according to the fuzzy stratum parameter interval, different models are randomly generated, theoretical plane wave dispersion curves of the models are calculated by using a generalized reflection-transmission coefficient method, and training sample data pairs are constructed;

preprocessing the training sample data pair;

constructing an LSTM network based on the FHLV loss function, training the LSTM network based on the preprocessed training sample data pair, and storing a trained model;

and (3) automatically picking up a dispersion curve from actual dispersion imaging data by using unsupervised learning, and predicting the dispersion curve with uniform dimension by using the trained model to obtain a near-surface transverse wave velocity model.

2. The method of claim 1, wherein the section of the fuzzy formation parameters of the work area to be inverted is determined using known logging information or according to existing inversion methods.

3. The method of claim 1, wherein the formation parameters include formation shear wave velocity, formation longitudinal wave velocity, formation thickness, and formation density for each layer, wherein theThe longitudinal wave velocity of the stratum is 2.4-3 times of the transverse wave velocity of the stratum at the same stratum, and the density of the stratum is 1.5-2Kg/m ³ The fuzzy stratum parameter interval is a fuzzy interval range of stratum parameters of each layer, and is determined by the minimum value and the maximum value of the parameters of each layer in the extracted one-dimensional speed structures at different positions, the initial range of the fuzzy stratum parameter interval is 40-80% of the minimum value, the end range is 120-140% of the maximum value, the larger the numerical value of the stratum parameters is, the smaller the proportion of the corresponding initial range and the end range is, and the one-dimensional speed structures at different positions are obtained according to known logging information or the existing inversion method.

4. A method according to claim 3, wherein said pre-processing said training sample data pairs comprises:

normalizing the training sample data pair, and vectorizing the stratum transverse wave speed and stratum thickness column in the training sample data pair.

5. The method of claim 1, wherein the LSTM network based on FHLV loss function comprises three hidden layers, each hidden layer corresponds to 256 LSTM units, each LSTM unit uses a jump connection, and a full connection layer is connected to the output terminal of the last LSTM unit to control the output dimension.

6. The method of claim 5, wherein training the LSTM network and saving the trained model based on the preprocessed training sample data pair, comprises:

inputting a theoretical plane wave dispersion curve into the LSTM network, outputting a corresponding stratum transverse wave speed and stratum thickness, and finishing training of the LSTM network;

the loss function of the LSTM network is:

Loss1＝mean(abs(v_label-v))

Loss2＝mean(abs(h_label-h))

where los 1 is the velocity Loss, los 2 is the thickness Loss, FHLV is the actual Loss function during training, label is the label, mean is the average operator, abs is the absolute operator, v is the formation shear wave velocity, h is the formation thickness, e is the ideal velocity Loss, set to 15-20% of the normalized minimum layer velocity, wh is the weight attenuation coefficient, set to 0.01-0.1.

7. The method of claim 1, wherein automatically picking up the dispersion curve for the actual dispersion imaging data using unsupervised learning comprises:

determining a surface wave dispersion energy matrix according to actual dispersion imaging data;

normalizing the surface wave dispersion energy matrix, and taking the coordinates of points larger than a first dispersion energy threshold value as first dispersion energy points;

clustering the first scattered energy points;

and comparing the frequency dispersion energy corresponding to the clustered points, and screening the phase velocity of the frequency vector corresponding to the local peak value to obtain a final frequency dispersion curve.

8. An LSTM surface wave inversion apparatus, the apparatus comprising a processor configured to:

determining a fuzzy stratum parameter interval of a work area to be inverted;

according to the fuzzy stratum parameter interval, different models are randomly generated, theoretical plane wave dispersion curves of the models are calculated by using a generalized reflection-transmission coefficient method, and training sample data pairs are constructed;

preprocessing the training sample data pair;

constructing an LSTM network based on the FHLV loss function, training the LSTM network based on the preprocessed training sample data pair, and storing a trained model;

and (3) automatically picking up a dispersion curve from actual dispersion imaging data by using unsupervised learning, and predicting the dispersion curve with uniform dimension by using the trained model to obtain a near-surface transverse wave velocity model.

9. The apparatus of claim 8, wherein the formation parameters include formation shear wave velocity, formation longitudinal wave velocity, formation thickness, and formation density for each layer, wherein the formation longitudinal wave velocity is 2.4-3 times the formation shear wave velocity at the same layer, and the formation density is 1.5-2Kg/m ³ The fuzzy stratum parameter interval is a fuzzy interval range of stratum parameters of each layer, and is determined by the minimum value and the maximum value of the parameters of each layer in the extracted one-dimensional speed structures at different positions, the initial range of the fuzzy stratum parameter interval is 40-80% of the minimum value, the end range is 120-140% of the maximum value, the larger the numerical value of the stratum parameters is, the smaller the proportion of the corresponding initial range and the end range is, and the one-dimensional speed structures at different positions are obtained according to known logging information or the existing inversion method.

10. A computer readable storage medium having stored thereon computer readable instructions which, when executed by a processor of a computer, cause the computer to perform the method of any of claims 1-7.

Images