A method and device for classifying the causes of formation anisotropy based on a neural network

Abstract

The present invention provides a neural network-based classification method and device for the causes of formation anisotropy, which relates to the fields of applied geophysics and petroleum exploration and development. The method includes: using a combination of formation parameters as a sample feature of a dispersion data set; calculating a dispersion equation through a numerical algorithm, solving a theoretical dispersion curve as a label of a dispersion data set; constructing a dispersion calculation forward model, correcting individual erroneous samples in the dispersion data set; using a combination of formation parameters as a sample feature, and using the corrected dispersion data set as a label to update the dispersion calculation forward model; extracting the dispersion curve of measured acoustic wave data; using the dispersion calculation forward model to obtain an initial predicted dispersion curve, establishing an objective function, and updating the optimal solution for dispersion data fitting based on a leapfrog algorithm iteration; and using the fitted optimal solution to classify the causes of formation anisotropy. The technical solution of the present invention overcomes the problems of long calculation time and poor algorithm stability in the classification method of dispersion curves in the prior art.

Description

Technical Field

[0001] The present invention relates to the fields of applied geophysics and oil exploration and development, and particularly relates to a method and device for classifying the causes of formation anisotropy based on a neural network. Background Art

[0002] The discrimination of phenomena such as borehole wall fractures, gas leakage, and tool eccentricity, as well as the determination of anisotropic regions caused by intrinsic anisotropy, stress, or fractures, can be evaluated by distinguishing and analyzing dispersion curves, so as to realize downstream applications such as borehole stability analysis, horizontal stress determination, and geological model calibration. Winkler et al. determined the cause of formation anisotropy through the crossing phenomenon of fast and slow flexural wave dispersion curves; Plona et al. combined on-site cross-dipole acoustic data to identify mechanical damage and in-situ stress using the dispersion characteristics of flexural waves. Therefore, correctly distinguishing the causes of anisotropy that induce dispersion curves is a key step in guiding petrophysical and geomechanical applications. However, existing traditional methods for classifying dispersion curves have defects such as long calculation time and poor algorithm stability, especially when dealing with noisy on-site data, and their disadvantages are particularly obvious, seriously restricting their application and development.

[0003] The development of machine learning technology has facilitated efficient processing of dispersion data, and its good adaptability and generalization ability have laid a solid foundation for the self-adjustment of model performance and accurate prediction of new data. Assous (2018) and Liang (2021) studied methods for calculating dispersion curves using neural networks, and promoting this method to the processing of cross-dipole flexural wave dispersion data is of great significance for the practical application of this technology.

[0004] Therefore, there is a need for a method and device for classifying the causes of formation anisotropy based on a neural network that can accurately classify the causes of formation anisotropy and improve the processing speed of cross-dipole flexural wave dispersion. Summary of the Invention

[0005] The main object of the present invention is to provide a method and device for classifying the causes of formation anisotropy based on a neural network to solve the problems of long calculation time and poor algorithm stability in existing methods for classifying dispersion curves.

[0006] To achieve the above object, the present invention provides a method for classifying the causes of formation anisotropy based on a neural network, which specifically includes the following steps:

[0007] S1, according to the dipole flexural wave logging model in a transversely isotropic formation and the equivalent tool theory, perform sensitivity analysis of different formation parameters, divide the parameter range and step size, and use the formation parameter combination as the sample feature of the dispersion data set.

[0008] S2. Calculate the dispersion equation through a numerical algorithm, and use the theoretical dispersion curves of the dipole flexural wave logging model in the VTI formation defined by each formation parameter combination obtained in step S1 as the labels of the dispersion data set.

[0009] S3. Use the multi-layer perceptron MLP of the fully connected neural network model to construct a forward dispersion calculation model. Set the maximum threshold of the difference between the numerical algorithm and the dispersion data set calculated by the forward dispersion calculation model, compare the error sizes of the training set and the test set, and correct individual error samples in the dispersion data set.

[0010] S4. Use the formation parameter combination in step S1 as the sample feature and the corrected dispersion data set in step S3 as the label to update the forward dispersion calculation model.

[0011] S5. Use the Prony method or the weighted spectrum WSS method to extract the dispersion curve of the measured acoustic wave data, and mark and extract the dispersion data through the density-based clustering method.

[0012] S6. According to the known formation parameters at the current formation depth point and the range of formation parameters to be inverted, form different groups of formation parameter combinations. Use the forward dispersion calculation model trained in step S4 to obtain the initial predicted dispersion curve, establish an objective function, and iterate based on the leapfrog algorithm to update the optimal solution that fits the dispersion data processed in step S5.

[0013] S7. Use the optimal solution fitted in step S6 to classify the origin of formation anisotropy.

[0014] Furthermore, the sensitivity calculation formula in step S1 is:

[0015] (1);

[0016] Where represents the parameter sensitivity, represents the formation parameter, represents the phase velocity of the wave at different frequencies, is the angular frequency.

[0017] Specifically, in step S2, the specific expression for solving the theoretical dispersion curve of the dipole flexural wave logging model in the VTI formation defined by each formation parameter combination obtained in step S1 is:

[0018] (2);

[0019] Where is the dispersion matrix using the equivalent instrument theory, is the wave number, is the well diameter, is the formation density, and are the fluid density and fluid velocity respectively, is the Thomsen parameter, is the Poisson's ratio, is the equivalent instrument radius.

[0020] Furthermore, step S3 specifically includes the following steps:

[0021] S3.1, construct a forward dispersion calculation model using the multi-layer perceptron MLP of the fully connected neural network model. The forward dispersion calculation model includes: an input layer, three hidden layers, and an output layer connected in sequence; the forward dispersion calculation model is expressed as:

[0022] (3);

[0023] Among them, represents the dispersion curve calculated by the forward dispersion calculation model, represents the formation parameters, 、 respectively represent the weight and bias of the th layer network, where , represents the activation function, The specific expression of is:

[0024] (4).

[0025] S3.2, subtract the dispersion data set label obtained through step S2 from the data set label obtained by the forward dispersion calculation model trained by the input formation parameter combination, and determine whether it is an error sample by the set maximum threshold of the difference.

[0026] S3.3, correct the error sample.

[0027] S3.4, input the corrected sample into the forward dispersion calculation model for training.

[0028] S3.5, divide the training set and the test set, and judge whether the training set error is less than the test set error. If it is greater than the test set error, continue to execute step S3.2. If it is less than the test set error, the correction ends.

[0029] Furthermore, the cosine annealing algorithm is used in step S4 to iterate the parameters of the forward dispersion calculation model:

[0030] (5);

[0031] Among them, is the current batch learning rate, is the annealing cycle number, is the annealing rate, and are the maximum and minimum values of the learning rate respectively, 、 are the number of training rounds after the last restart and the given restart interval rounds respectively.

[0032] Furthermore, step S5 specifically includes the following steps:

[0033] S5.1, Use the Prony or weighted spectrum WSS method to extract the dispersion data of the measured acoustic wave data. The expression of Prony is as follows:

[0034] (6);

[0035] Where, represents the spectral prediction value of the th receiver, is a complex exponential term, is an imaginary number, is the interval between receivers, is the slowness spectrum, is the amplitude of the th mode shape, is the total number of mode shapes.

[0036] The specific expression of the weighted spectrum method is as follows:

[0037] (7);

[0038] Where, is the spectral coherence function, is the slowness, is the spectrum of the th receiver, is 's conjugate, is the number of receivers, and .

[0039] S5.2, Use the DBSCAN method to mark and extract the dispersion data. The specific expression is as follows:

[0040] (8);

[0041] Where, is the extracted dispersion data, is the number of clusters, and are different cluster samples, and are the average distances from the samples within the cluster to the cluster center, is the distance between the cluster centers, and is the cluster sample set.

[0042] Furthermore, step S6 specifically includes the following steps:

[0043] S6.1. The objective function is:

[0044] (9);

[0045] Wherein, is the dispersion data processed through step S5, is the numerical value of the dispersion curve predicted by the neural network, is the number of frequency calculations, is the frequency calculation range, is the calculated frequency.

[0046] S6.2. The specific calculation formula of the leapfrog algorithm update strategy is:

[0047] (10);

[0048] (11);

[0049] Wherein, and respectively represent the optimal solution and the worst solution in the th population, is the updated optimal solution, is the jump step size, and are respectively the minimum value and the maximum value of the jump step size, is the updated new individual.

[0050] The present invention also provides a formation anisotropy origin classification device based on a neural network, including: at least one processor and a memory;

[0051] The memory stores computer execution instructions;

[0052] At least one processor executes the computer execution instructions stored in the memory, so that the at least one processor executes the formation anisotropy origin classification method based on the neural network.

[0053] The present invention has the following beneficial effects:

[0054] The present invention takes eight related parameters such as well diameter and instrument as sample features, and the dispersion curve as the sample label. According to the equivalent instrument theory model, a dispersion data set is made, and after forward training, a forward model for dispersion calculation based on machine learning is obtained. The forward model for dispersion calculation and the leapfrog algorithm are used to invert the target parameters, fit the dispersion curve extracted from the actual waveform, and achieve accurate classification of the causes of formation anisotropy. Compared with the traditional dispersion processing and classification methods, the present invention greatly improves the processing speed of the cross-dipole flexural wave dispersion, and the feasibility and effectiveness of the present invention are proved by combining on-site data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] 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 required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts. In the drawings:

[0056] Figure 1 A flowchart of a method for classifying the causes of formation anisotropy based on a neural network according to the present invention is shown.

[0057] Figure 2 A dispersion sensitivity analysis diagram of the dipole flexural wave fast formation under the VTI formation is shown.

[0058] Figure 3 A dispersion sensitivity analysis diagram of the dipole flexural wave slow formation under the VTI formation is shown.

[0059] Figure 4 Dipole fast and slow shear wave waveform diagrams are shown.

[0060] Figure 5 A fitting diagram of the actual dispersion data of the dipole fast and slow shear waves is shown.

[0061] Figure 6 A classification result diagram of the causes of formation anisotropy is shown. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0062] The technical solutions of the present invention will be clearly and completely described below with reference to the drawings. Obviously, the described embodiments are some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0063] Embodiment 1

[0064] As Figure 1A method for classifying the causes of formation anisotropy based on a neural network specifically includes the following steps:

[0065] S1. According to the dipole flexural wave logging model in a transversely isotropic formation and the equivalent instrument theory, perform sensitivity analysis of different formation parameters, divide the parameter range and step size, and use the formation parameter combination as the sample feature of the dispersion data set. Among them, the dipole flexural wave logging model refers to the logging model constructed by studying the propagation characteristics of flexural waves excited by a dipole sound source under formation parameter conditions; the equivalent instrument theory is to consider the elastic effect of the logging instrument and introduce instrument model parameters to simulate its influence on logging data.

[0066] Follow the principle of "the greater the parameter sensitivity, the relatively denser the value" to divide the parameter range and step size. As shown in Figure 2 and Figure 3 the sensitivity analysis diagram, the shear wave velocity shows high sensitivity in both fast formations and slow formations. Therefore, a smaller step size is adopted to divide more values within the shear wave velocity range. After taking values for the 8 formation parameters in the dipole flexural wave logging model at appropriate step sizes, perform permutation and combination among the parameters to determine the formation parameter combination as the sample feature of the dispersion data set.

[0067] S2. Calculate the dispersion equation through a numerical algorithm, and solve the theoretical dispersion curve of the dipole flexural wave logging model in the VTI formation defined by each formation parameter combination obtained in step S1 as the label of the dispersion data set.

[0068] S3. Use the multi-layer perceptron MLP of the fully connected neural network model to construct a forward dispersion calculation model, set the maximum threshold of the difference between the numerical algorithm and the dispersion data set calculated by the forward dispersion calculation model, compare the error sizes of the training set and the test set, and correct individual error samples in the dispersion data set.

[0069] S4. Use the formation parameter combination in step S1 as the sample feature and the corrected dispersion data set in step S3 as the label to update the forward dispersion calculation model.

[0070] S5. Use the Prony method or the weighted spectrum WSS method to extract the dispersion curve of the measured acoustic wave data, and mark and extract the dispersion data through the density-based clustering method.

[0071] S6. According to the known formation parameters at the current formation depth point and the range of formation parameters to be inverted, including: Thomsen parameters and Poisson ratio, form different groups of formation parameter combinations, use the forward dispersion calculation model trained in step S4 to obtain the initial predicted dispersion curve, establish an objective function, and iterate based on the leapfrog algorithm to update the optimal solution that fits the dispersion data processed in step S5, that is, obtain the value of the current formation depth point, Values and the fitted dispersion curves.

[0072] S7. Classify the causes of formation anisotropy using the optimal solution fitted in step S6.

[0073] Based on processing time-domain acoustic waveforms, effectively extracting and marking dispersion data, the present invention trains a forward dispersion calculation model to rapidly calculate the theoretical dispersion curves of the dipole flexural wave logging model in VTI formations, and combines the leapfrog algorithm to invert the objective function and fit the actual dispersion data to complete the classification of the causes of anisotropic formations. Compared with traditional dispersion extraction and classification methods, the present invention significantly improves the processing efficiency of acoustic logging data and can better meet the timeliness requirements of on-site production.

[0074] Specifically, the sensitivity analysis of different formation parameters according to the dipole acoustic logging model in a transversely isotropic formation in step S1 is as follows:

[0075] According to the sensitivity analysis of parameters, the influence of different formation parameters on the dispersion curve generally follows the principle that "the greater the sensitivity of the formation parameter, the relatively denser the value", and the step size and parameter range are delimited.

[0076] The sensitivity analysis diagrams are as shown in Figure 2 and Figure 3 In the fast formation as shown in Figure 2 , for the dipole flexural wave in a VTI medium, at low frequencies, the sensitivity is mainly controlled by the shear wave velocity, and as the frequency increases, the fluid velocity dominates. In the slow formation as shown in Figure 3 , the dipole flexural wave has a high sensitivity to the VTI formation over the entire frequency range.

[0077] Perform sensitivity analysis on the 8 input formation parameters. The sensitivity calculation formula is:

[0078] (1);

[0079] where represents the parameter sensitivity, represents the formation parameter, represents the phase velocity of the wave at different frequencies, is the angular frequency.

[0080] As shown in Table 1, based on sensitivity analysis, appropriate step sizes and parameter ranges are divided. Through permutation and combination, the sample feature combinations of the dipole flexural wave dispersion dataset are obtained. The formation shear wave velocity shows high sensitivity in both fast formations and slow formations. Therefore, a smaller step size is adopted to divide more values within the formation shear wave velocity range. However, the equivalent instrument model modulus is at a low value in the entire frequency band in both fast formations and slow formations and does not change with frequency. Therefore, this is set as a fixed value and is not used as a sample feature parameter.

[0081] Table 1 Sample Feature Table of Dipole Flexural Wave Dispersion Dataset

[0082]

[0083] Specifically, in step S2, the borehole flexural wave excited by the dipole acoustic logging tool is a dispersive guided wave propagating along the borehole wall. The specific expression of the theoretical dispersion curve of the dipole flexural wave logging model in the VTI formation defined by each formation parameter combination obtained in step S1 is:

[0084] (2);

[0085] Among them, is the dispersion matrix using the equivalent instrument theory, is the wave number, is the borehole diameter, is the formation density, and are the fluid density and fluid velocity respectively, is the Thomsen parameter, is the Poisson's ratio, is the equivalent instrument radius.

[0086] A total of 8 parameters are required for the dispersion equation to be solved. In step one of the present invention, 725,760 parameter combinations are obtained to define the dispersion equation. Through step S2, the theoretical dispersion curves of the dipole flexural wave logging model in the VTI formation defined by each formation parameter combination corresponding to the dispersion equation are solved. Each dispersion curve consists of the flexural wave velocities corresponding to 76 frequency points evenly distributed from 0.1 kHz to 15.1 kHz.

[0087] Specifically, due to the complexity and multiple solutions of the dispersion equation, during the numerical algorithm for calculating the dispersion equation in step S2, the algorithm needs to be continuously adjusted to ensure the correctness of the result.

[0088] Step S3 specifically includes the following steps:

[0089] S3.1. The present invention uses 725,760 sample features and labels in the dataset to train a fully-connected neural network model with three hidden layers, namely the dispersion calculation forward model. The trained dispersion calculation forward model is used to replace the time-consuming and unstable dispersion search mode in step S2.

[0090] The multi-layer perceptron (MLP) is a classic fully-connected neural network model, consisting of an input layer, hidden layers, and an output layer. An activation function is used for non-linear transformation in each hidden layer. The neural network model of the present invention is set as a fully-connected neural network with three hidden layers, and each hidden layer contains 1,200 neurons. This model can be regarded as a function with eight formation parameters as independent variables and the dispersion curve of the VTI formation flexural wave as the dependent variable.

[0091] The dispersion calculation forward model is constructed by using the fully-connected neural network model multi-layer perceptron MLP. The dispersion calculation forward model includes: an input layer, three hidden layers, and an output layer connected in sequence; the dispersion calculation forward model is expressed as:

[0092] (3);

[0093] Where, represents the dispersion curve calculated by the dispersion calculation forward model, represents the formation parameters, , respectively represent the weight and bias of the th layer network, where , represents the activation function, The specific expression of

[0094] (4).

[0095] S3.2. Subtract the dispersion dataset labels obtained through step S2 from the dataset labels obtained by the dispersion calculation forward model trained with the input formation parameter combinations, and determine whether it is an incorrect sample by setting the maximum threshold of the difference.

[0096] S3.3. Correct the incorrect samples.

[0097] S3.4. Input the corrected samples into the neural network for training.

[0098] S3.5. Divide the training set and the test set, and determine whether the training set error is less than the test set error. If it is greater than the test set error, continue to execute step S3.2. If it is less than the test set error, the correction ends.

[0099] In step S4, the cosine annealing algorithm is used to iteratively calculate the forward model parameters of dispersion. The training set and the validation set are divided in a ratio of 95:5. The cosine annealing algorithm gradually reduces the learning rate by using the cosine function, smooths the change of the learning rate, and prevents gradient oscillation. The specific expression of the algorithm principle is:

[0100] (5);

[0101] Among them, is the batch learning rate at the current moment , is the number of annealing cycles, is the annealing rate, and are the maximum and minimum values of the learning rate respectively, , are the number of training rounds after the last restart and the given restart interval rounds respectively.

[0102] Since the objective optimization function may have multiple local optimal solutions, this method of hot restart makes it possible to jump out of the local optimal solution, thereby improving the global search ability of network training convergence.

[0103] Specifically, step S5 specifically includes the following steps:

[0104] S5.1, using the Prony or weighted spectrum WSS method to extract the dispersion data of the measured acoustic wave data. The expression of Prony is as follows:

[0105] (6);

[0106] Among them, represents the spectral prediction value of the th receiver, is a complex exponential term, is an imaginary number, is the interval between receivers, is the slowness spectrum, is the th mode amplitude, is the total number of modes.

[0107] The specific expression of the weighted spectrum method is as follows:

[0108] (7);

[0109] Among them, is the spectral coherence function, is the slowness, is the spectrum of the th receiver, is The conjugate of is the number of receivers, and .

[0110] S5.2. Use the DBSCAN method to label and extract dispersion data, and then remove noise. The specific expression is as follows:

[0111] (8);

[0112] Where is the extracted dispersion data, is the number of clusters, and are samples of different clusters, and are the average distances from the samples within the cluster to the cluster center, is the distance between the cluster centers, and are the cluster sample sets.

[0113] Specifically, step S6 specifically includes the following steps:

[0114] S6.1. The objective function is:

[0115] (9);

[0116] Where is the dispersion data after being processed in step S5, is the numerical value of the dispersion curve predicted by the neural network, is the number of frequency calculations, is the frequency calculation range, is the calculated frequency.

[0117] S6.2. The specific calculation formula for the shuffled frog leaping algorithm update strategy is:

[0118] (10);

[0119] (11);

[0120] Where and respectively represent the optimal solution and the worst solution in the th population, is the updated optimal solution, is the jump step size, and are respectively the minimum value and the maximum value of the jump step size, is the updated new individual.

[0121] This heuristic optimization algorithm combines local search and global search, updates the optimal individual according to the fitness value of the solution, and when the local search reaches a certain threshold, integrates and updates the information within the group, and iteratively obtains the global optimal solution.

[0122] Figure 4 The fast and slow shear wave waveforms obtained from the orthogonal dipole four-component data are processed by step S5 to obtain Figure 5 the fast and slow shear wave in-situ dispersion data in, which are scatter points. By fitting the in-situ dispersion data through step S6, the fast and slow shear wave dispersion curves are obtained, which are dotted lines. This in-situ example is used to verify the effectiveness of the method for extracting waveforms from full-wave train signals to mark the dipole flexural wave dispersion data and automatically obtain the data processing method for fitting the dispersion curve.

[0123] Based on the dispersion characteristics of dipole flexural waves, anisotropic formation classification is carried out for different phenomenon characteristics such as dispersion curve coincidence, intersection, and mid-frequency band splitting at each depth point. For example Figure 6 As shown, at different depth segments, the fast and slow shear wave dispersion curves after in-situ data processing and fitting are colored with different colors to accurately and efficiently classify the dispersion types. This method for classifying the origin of formation anisotropy based on neural networks has the characteristics of high-efficiency dispersion processing and high-precision parameter inversion.

[0124] Embodiment 2

[0125] An apparatus for classifying the origin of formation anisotropy based on neural networks includes: at least one processor and a memory.

[0126] The memory stores computer-executable instructions.

[0127] At least one processor executes the computer-executable instructions stored in the memory, so that at least one processor executes a method for classifying the origin of formation anisotropy according to the present invention.

[0128] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions, or substitutions made by those skilled in the art within the scope of the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A neural network-based classification method for formation anisotropy, characterized in that: The specific steps include: S1, based on the dipole flexural wave logging model in transversely isotropic formations and the equivalent instrument theory, sensitivity analysis of different formation parameters is performed, parameter ranges and step sizes are divided, and formation parameter combinations are used as sample features of dispersion data sets; S2, calculating the dispersion equation by numerical algorithm, solving the theoretical dispersion curve of the dipole bending wave logging model in the VTI formation defined by each formation parameter combination obtained in step S1 as a label of the dispersion data set; S3, using a fully connected neural network model multi-layer perceptron MLP to build a dispersion calculation forward model, set the maximum threshold of the difference between the dispersion data set calculated by the numerical algorithm and the dispersion calculation forward model, compare the error size of the training set and the test set, and correct individual error samples in the dispersion data set; S4, using the formation parameter combination in step S1 as a sample feature and the corrected dispersion data set in step S3 as a label, updating the dispersion calculation forward model; S5, extracting the dispersion curve of the measured acoustic wave data using the Prony method or the weighted spectrum WSS method, and marking and extracting the dispersion data using a density-based clustering method; S6, forming different groups of formation parameter combinations according to the known formation parameters at the current formation depth point and the range of formation parameters that need to be inverted, using the dispersion calculation forward model trained in step S4 to obtain an initial predicted dispersion curve, establishing an objective function, and updating the optimal solution for fitting the dispersion data processed in step S5 based on the iteration of the leapfrog algorithm; S7, classifying the causes of formation anisotropy using the optimal solution fitted in step S6.

2. The neural network-based formation anisotropy genesis classification method according to claim 1, characterized in that: The sensitivity calculation formula in step S1 is: （1）； in, represents parameter sensitivity, represents the formation parameters, represents the phase velocity of the wave at different frequencies, is the angular frequency.

3. The neural network-based formation anisotropy genesis classification method according to claim 1, characterized in that: In step S2, the specific expression of the theoretical dispersion curve of the dipole flexural wave logging model in the VTI formation defined by each formation parameter combination obtained in step S1 is solved as follows: （2）； in, is the dispersion matrix using equivalent instrument theory, is the wave number, is the well diameter, is the formation density, and are the fluid density and fluid velocity, respectively, is the Thomsen parameter, is Poisson's ratio, is the equivalent instrument radius.

4. The neural network-based formation anisotropy genesis classification method according to claim 1, characterized in that: Step S3 specifically includes the following steps: S3.1, a fully connected neural network model multi-layer perceptron MLP is used to construct a dispersion calculation forward model. The dispersion calculation forward model includes: an input layer, three hidden layers and an output layer connected in sequence; the dispersion calculation forward model is expressed as: （3）； in, represents the dispersion curve calculated by the dispersion calculation forward model, represents the formation parameters, , Respectively represent The weights and biases of the layer network, where , represents the activation function, The specific expression is: （4）； S3.2, subtract the dispersion data set label obtained in step S2 from the data set label obtained by the dispersion calculation forward model trained by inputting the formation parameter combination, and determine whether it is an erroneous sample by setting the maximum difference threshold; S3.3, correcting erroneous samples; S3.4, inputting the corrected samples into the dispersion calculation forward model for training; S3.5, divide the training set and the test set, and determine whether the training set error is smaller than the test set error. If it is larger than the test set error, continue to step S3.

2. If it is smaller than the test set error, the correction is completed.

5. The neural network-based formation anisotropy genesis classification method according to claim 1, characterized in that: In step S4, the cosine annealing algorithm is used to iteratively calculate the forward model parameters: （5）； in, is the current batch learning rate, is the number of annealing cycles, is the annealing rate, and are the maximum and minimum values ​​of the learning rate, respectively. , They are the training round after the last restart and the round with a given restart interval, respectively.

6. The neural network-based formation anisotropy genesis classification method according to claim 1, characterized in that: Step S5 specifically includes the following steps: S5.1, use Prony or weighted spectrum WSS method to extract the dispersion data of measured sound wave data. The expression of Prony is as follows: （6）； in, Indicates The spectrum prediction value of each receiver is is a complex exponential term, is an imaginary number, is the spacing between receivers, is the slowness spectrum, For the The amplitude of the vibration mode, is the total number of vibration modes; The specific expression of the weighted spectrum method is as follows: （7）； in, is the spectral coherence function, For slowness, For the The spectrum of the receiver, for The conjugate of is the number of receivers, and ; S5.2, use the DBSCAN method to mark and extract dispersion data. The specific expression is as follows: （8）； in, is the extracted dispersion data, is the number of clusters, and are samples of different clusters, and is the average distance from the samples in the cluster to the cluster center, is the distance between cluster centers, and is the cluster sample set.

7. The neural network-based formation anisotropy genesis classification method according to claim 1, characterized in that: Step S6 specifically includes the following steps: S6.1, Objective Function for: （9）； in, is the dispersion data after processing in step S5, is the dispersion curve value predicted by the neural network, Count the number of frequencies, is the frequency calculation range, To calculate the frequency; S6.2, the specific calculation formula of the frog leaping algorithm update strategy is: （10）； （11）； in, and Respectively represent The best and worst solutions in a population, is the updated optimal solution, is the jump step length, and are the minimum and maximum values ​​of the jump step length, For updated new individuals.

8. A neural network-based formation anisotropy genesis classification device, characterized in that: include: at least one processor and memory; The memory stores computer-executable instructions; The at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor performs the neural network-based formation anisotropy genesis classification method as described in any one of claims 1 to 7.

Images