Surface wave frequency dispersion curve intelligent inversion method based on parameter normalization

Abstract

The invention discloses a surface wave frequency dispersion curve intelligent inversion method based on parameter normalization, and relates to the technical field of deep learning. The method comprises the following steps: firstly, generating a plurality of groups of training data pairs in a reference underground model space, then pre-training the training data pairs by using an improved Transform-based PADIT model, mapping an actually measured Rayleigh wave frequency dispersion curve into the reference underground model space, predicting the actually measured Rayleigh wave frequency dispersion curve by using the pre-trained PADIT model, and finally selecting a better inversion result by using a mismatch value. According to the method, the defect that in the prior art, a deep learning method depends on a specific scale sample is overcome, and the efficiency of seismic data Rayleigh wave frequency dispersion curve inversion work and the universality of the model are greatly improved; meanwhile, the deep learning model adopted by the invention can directly process frequency dispersion curve data with variable length, does not need to carry out preprocessing such as interpolation or truncation on input data, is more in line with the condition that the data length is variable in practical application, and has better flexibility and robustness.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, and in particular to an intelligent inversion method for surface wave dispersion curves based on parameter normalization. Background Technology

[0002] In the field of geophysical exploration, obtaining high-precision subsurface shear wave velocity (Vs) structures is fundamental for geological structural interpretation, resource exploration, engineering site evaluation, and hazard assessment. Seismic data is the most direct and abundant source of information revealing subsurface structures. In seismic data processing, surface waves (especially Rayleigh waves), as a wavefield component carrying a wealth of near-surface information, have dispersion characteristics (i.e., waves of different frequencies propagate at different speeds) that are closely related to the Vs structure of the subsurface medium. Therefore, analyzing surface wave dispersion curves extracted from active or passive source (e.g., natural earthquakes, background noise) seismic records and then inverting them to obtain the Vs structure is a widely used and cost-effective geophysical exploration technique. This technique is widely applied across multiple scales, from near-surface engineering investigations (e.g., site stability assessment, active fault detection) to deep crustal structure imaging (e.g., basin structure studies, Moho depth determination).

[0003] Traditional inversion methods are mainly divided into two categories: gradient-based local optimization methods and stochastic global optimization methods. The former converges quickly, but is sensitive to the initial model and is prone to getting trapped in local optima; the latter can perform a global search, but has extremely high computational cost and is time-consuming.

[0004] With the development of artificial intelligence technology, deep learning methods have been introduced into surface wave dispersion inversion. By using neural networks to learn the mapping relationship between dispersion data and velocity structure, rapid inversion can be achieved after training. However, existing deep learning inversion techniques have the following main drawbacks:

[0005] Strong scale dependence and poor generality: Existing models are usually trained on training data at specific depth and velocity scales. For example, a model trained for deep crustal research cannot be directly used for inversion problems in shallow engineering exploration. When faced with a new problem that does not match the scale of the training data, the training data must be regenerated and the model retrained. This process is time-consuming and labor-intensive, greatly limiting the model's generality and application efficiency.

[0006] Fixed input length limits flexibility: Most neural network models require a fixed dimension for the input features, while the number of data points (i.e., the length) of the experimental dispersion curves obtained through actual acquisition and processing often varies. This makes it difficult for existing models to directly handle dispersion curves of variable length, resulting in insufficient flexibility in practical applications.

[0007] The process is complex and relies on manual parameters: Traditional inversion methods and some deep learning applications require users to perform tedious manual parameterization, such as setting the number of underground layers and setting velocity constraints for each layer. These subjective choices can significantly affect the accuracy of the final results. Summary of the Invention

[0008] To address at least one of the aforementioned problems, this application provides an intelligent inversion method for surface wave dispersion curves based on parameter normalization.

[0009] To achieve the above objectives, this application provides a smart inversion method for surface wave dispersion curves based on parameter normalization, comprising the following steps:

[0010] S1. Set up a reference underground model space with a reference depth and a reference shear wave velocity. The reference underground model space includes a semi-infinite layer extending infinitely downward below the reference depth, and n thin layers discretized to a fixed thickness above the reference depth. Set a constant density and Poisson's ratio for all layers. The shear wave velocity of the semi-infinite layer is the reference shear wave velocity.

[0011] S2. Under the preset shear wave velocity, multiple training samples are generated using the random layer allocation sampling method. Each training sample contains n shear wave velocities. Forward calculation is performed based on the training samples to obtain the first dispersion curve. Each training sample and its corresponding first dispersion curve constitute a training data pair.

[0012] S3. Construct the PADIT model and train the model based on the above training data to obtain the trained model; The PADIT model is based on the Transformer model. First, each data point in the first dispersion curve is transformed into a high-dimensional feature vector through the embedding layer, and a position encoding is added to each high-dimensional feature vector. Then, the sequence of high-dimensional feature vectors with position encoding is input into a multi-layer Transformer encoder, and a summing pooling layer is used to aggregate the data points corresponding to the first dispersion curve into a global descriptor with a fixed dimension.

[0013] S4. Obtain the measured Rayleigh wave dispersion curve of the reservoir to be inverted, and based on its maximum detection wavelength and maximum phase velocity, estimate the depth range and shear wave velocity range of the semi-infinite layer in the real underground model space according to experience. Generate multiple sets of depth and shear wave velocity parameter pairs within the aforementioned range through sampling. Map the parameter pairs to the benchmark underground model space through normalization operation to obtain multiple scale factor pairs. The scale factor pairs consist of depth normalization scale factor and velocity normalization scale factor.

[0014] S5. Based on multiple scaling factor pairs, the measured Rayleigh wave dispersion curve is transformed into the benchmark underground model space to obtain multiple second dispersion curves. The multiple second dispersion curves are substituted into the pre-trained PADIT model to obtain multiple normalized velocity profiles.

[0015] S6. Multiple normalized velocity profiles are inversely normalized based on scaling factor pairs to obtain multiple candidate inverted velocity profiles. Multiple third dispersion curves are generated by forward modeling multiple candidate inverted velocity profiles. The inversion result is selected based on the mismatch value between the third dispersion curve and the measured Rayleigh wave dispersion curve.

[0016] The beneficial effects of this invention are as follows:

[0017] 1. The method of this invention enables a single pre-trained model to handle inversion problems at any depth and velocity scale, from near-surface to deep crust. This overcomes the limitation of existing deep learning methods that rely on samples at specific scales, significantly improving the efficiency of inversion work and the versatility of the model. Users no longer need to perform tedious and subjective initial model parameterization settings; they only need to provide approximate range estimates of the half-space parameters. This greatly simplifies the inversion process, reduces biases introduced by human intervention, and makes the inversion results more objective and reliable.

[0018] 2. The deep learning model used in this invention can directly process dispersion curve data of variable length without the need for preprocessing such as interpolation or truncation of the input data. This is more in line with the situation of variable data length in actual applications and has better flexibility and robustness.

[0019] 3. It focuses more on the intrinsic, scale-independent physical mapping relationship between dispersion curves and velocity structures, thereby avoiding the learning of scale-dependent surface features, which improves the inversion accuracy and robustness of the model. Attached Figure Description

[0020] Figure 1 This is a comparison between the inversion results and measured values ​​of Example 1;

[0021] Figure 2 This is a comparison between the inversion results and measured values ​​of Example 2;

[0022] Figure 3 This is a comparison between the inversion results and measured values ​​of Example 3. Detailed Implementation

[0023] The technical solution of this application will be clearly described below with reference to the embodiments thereof. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are all within the protection scope of this application.

[0024] This application provides a method for intelligent inversion of surface wave dispersion curves based on parameter normalization, including the following steps:

[0025] S1. Set up a reference underground model space with a reference depth and a reference shear wave velocity. The reference underground model space includes a semi-infinite layer extending infinitely downward below the reference depth, and n thin layers discretized to a fixed thickness above the reference depth. Set a constant density and Poisson's ratio for all layers. The shear wave velocity of the semi-infinite layer is the reference shear wave velocity.

[0026] Specifically, in this step, the reference underground model space is a semi-infinite space: it has depth but no length or width is set. Regarding the specific parameters of the reference underground model space, its reference depth can be set to 100m, or other values ​​such as 20m, 50m, 80m, 200m, 500m, etc. The 100m value here is only for ease of subsequent calculations. Its reference shear wave velocity can be set to 500–2000 m / s, such as 800m / s, 1000m / s, 1200m / s, 1500m / s, 1800m / s. In this embodiment, the reference shear wave velocity is selected as 1000m / s.

[0027] The number of thin layers can be set according to the reference depth, and each thin layer has the same thickness. The thickness of the thin layer is usually 1 / 200 to 1 / 50 of the reference depth. The thinner the thin layer, the higher the resolution, but the higher the computational cost. Therefore, to balance resolution and computational cost, the thickness of the thin layer can be set to 1 / 100 of the reference depth. For example, when the reference depth is 100m, the thin layer thickness can be set to 0.5 to 2m, such as 0.5m, 0.8m, 1m, 1.2m, 1.5m, 1.8m, etc. In this embodiment, the thin layer thickness is set to 1m.

[0028] For Rayleigh wave dispersion curves, the main influencing factors are shear wave velocity and layer thickness, while the secondary influencing factors are density and Poisson's ratio. The influence of density and Poisson's ratio is relatively small compared to shear wave velocity and layer thickness. Therefore, for each thin layer in the benchmark underground model space, a constant density and Poisson's ratio can be set. The specific values ​​can be set based on measured values. In this embodiment, the density is set to 2000 kg / m³. 3 The Poisson's ratio is 0.33. Of course, those skilled in the art can modify the density and Poisson's ratio according to the actual situation.

[0029] S2. Under the preset shear wave velocity, multiple training samples are generated using the random layer allocation sampling method. Each training sample contains n shear wave velocities. Forward calculation is performed based on the training samples to obtain the first dispersion curve. Each training sample and its corresponding first dispersion curve constitute a training data pair.

[0030] In this step, the so-called preset shear wave velocity refers to the range of shear wave velocities composed of the minimum possible shear wave velocity and the shear wave velocity of the semi-infinite layer below the reference depth. In this embodiment, the minimum possible shear wave velocity is not less than 1 / 20 of the reference shear wave velocity, and can usually be set to 1 / 20, 1 / 15, 1 / 10, etc. of the reference shear wave velocity. In this embodiment, the minimum possible shear wave velocity is set to 50 m / s, which is 1 / 20 of the reference shear wave velocity. The reference shear wave velocity is the shear wave velocity of the semi-infinite layer extending infinitely downward below the reference depth, which is 1000 m / s. In the reference underground model space above the reference depth, the shear wave velocity of the uppermost thin layer is the smallest, and the shear wave velocity of the lowermost thin layer is the largest. The shear wave velocity of the thin layers in the middle increases from top to bottom. This setting method has the same monotonicity as the shear wave velocity variation in conventional strata (without low-velocity interlayers). Those skilled in the art can set the velocity of each thin layer according to the methods of existing technology.

[0031] To avoid premature convergence of velocity values ​​due to existing methods, this embodiment proposes a random layer-level allocation sampling method to configure shear wave velocities for each thin layer in the middle. The operation steps of the random layer-level allocation sampling method are as follows:

[0032] S21. Set the shear wave velocity of the uppermost and lowermost thin layers above the reference depth in the reference underground model space;

[0033] S22. Randomly select a thin layer in the middle, and perform uniform random sampling based on the shear wave velocity of the two adjacent thin layers with assigned velocities to assign shear wave velocity to the thin layer.

[0034] S23. Repeat S22 until all thin layers are configured with shear wave velocity;

[0035] S24. Repeat S21 to S23 until multiple training samples are obtained, the number of which is 500,000 to 5,000,000.

[0036] The above sampling method effectively avoids the bias problem that may occur with simple sampling methods, where the velocity values ​​converge prematurely to the half-space velocity, thus ensuring the diversity and unbiasedness of the training samples throughout the parameter space.

[0037] After obtaining multiple training samples, forward modeling is performed on each training sample to obtain multiple first dispersion curves.

[0038] S3. Construct the PADIT model and train the model based on the above training data to obtain the trained model; The PADIT model is based on the Transformer model. First, each data point in the first dispersion curve is transformed into a high-dimensional feature vector through the embedding layer, and a position encoding is added to each high-dimensional feature vector. Then, the sequence of high-dimensional feature vectors with position encoding is input into a multi-layer Transformer encoder, and a summing pooling layer is used to aggregate the data points corresponding to the first dispersion curve into a global descriptor with a fixed dimension.

[0039] In existing technologies, conventional models, such as the Transformer model and BPNN, require inputs of equal length: if the input length is less than the average length, it needs to be padded with special characters to represent whitespace; if the input length exceeds the average length, it needs to be truncated. However, the special characteristics of dispersion curves make it difficult to sample conventional model inputs. Therefore, in this embodiment, the inventors have made an improvement.

[0040] First, the inventors set up an embedding layer to transform each data point in the first dispersion curve into a high-dimensional feature vector. Considering the characteristics of the dispersion curve, a positional encoding was added to each data point. Subsequently, after processing by a multi-layer Transformer encoder of the Transformer model, a summing pooling layer was set up to aggregate the output vector into a single, fixed-dimensional global descriptor. This descriptor condenses the essential information of the entire dispersion curve, and its dimension is independent of the original length of the input curve. Finally, a regression head maps the global descriptor to n shear wave velocities. Combined with the positional encoding of the n shear wave velocities, a curve showing the variation of shear wave velocity with depth is obtained, namely the Vs (shear wave velocity) velocity profile.

[0041] Meanwhile, in this step, for the model itself, the root mean square percentage error (RMSE) is used as the loss function. Compared to the conventional loss function, the RMSE gives equal importance to the inversion accuracy of both shallow low-speed and deep high-speed components. The conventional loss function, however, uses the absolute error as the loss function value. This has the disadvantage of using the same error value for both cases with low and high actual speeds, which is inconsistent with reality. The AdamW optimizer and cosine annealing learning rate scheduling strategy are used to optimize the model parameters. Data augmentation techniques are employed to enhance the model's robustness and generalization ability; these data augmentation techniques refer to randomly cropping the training data into segments of different lengths.

[0042] S4. Obtain the measured Rayleigh wave dispersion curve of the reservoir to be inverted, and based on its maximum detection wavelength and maximum phase velocity, estimate the depth range and shear wave velocity range of the semi-infinite layer in the real underground model space according to experience. Generate multiple sets of depth and shear wave velocity parameter pairs within the aforementioned range through sampling. Map the parameter pairs to the benchmark underground model space through normalization operation to obtain multiple scale factor pairs. The scale factor pairs consist of depth normalization scale factor and velocity normalization scale factor.

[0043] In this step, the depth of the semi-infinite subsurface layer and the range of shear wave velocities are first estimated based on the maximum detection wavelength and maximum phase velocity of the measured Rayleigh wave dispersion curve. This operation is a standard procedure in the field, and therefore its specific method will not be elaborated here. The measured Rayleigh wave dispersion curve can be obtained from seismic data.

[0044] After calculating the depth and shear wave velocity ranges, logarithmic uniform sampling is performed within these ranges to obtain multiple sets of depth and shear wave velocity parameter pairs. For each parameter pair, combined with the baseline subsurface model space in S1, a normalization operation is performed, and the scaling factor pair is calculated using the following formula: α H =H0 / H hs α V =V S0 / V S,hs In the formula, α H and α V H represents the depth normalization scaling factor and the velocity normalization scaling factor, respectively; H0 represents the reference depth of the reference subsurface model space; H hs V represents the depth value of the parameter pair; S0 V represents the reference shear wave velocity in the reference underground model space. S,hs This represents the shear wave velocity of the parameter pair.

[0045] S5. Based on multiple scaling factor pairs, the measured Rayleigh wave dispersion curve is transformed into the benchmark underground model space to obtain multiple second dispersion curves. The multiple second dispersion curves are substituted into the pre-trained PADIT model to obtain multiple normalized velocity profiles.

[0046] In this step, firstly, based on the scaling factor pair of S4, each data point on the measured Rayleigh wave dispersion curve is transformed to the reference subsurface model space: f n =(α V / α H )×f,V R,n =α V ×V R In the formula, f and V R f represents the frequency and phase velocity at each data point on the measured Rayleigh wave dispersion curve, respectively; n and VR,n α represents the frequency and phase velocity of each data point on the second dispersion curve, respectively; H and α V These represent the depth normalization scaling factor and the velocity normalization scaling factor, respectively; subsequently, the frequencies and phase velocities of data points on multiple second dispersion curves are combined to form a second dispersion curve.

[0047] Each scaling factor pair corresponds to a second dispersion curve. After obtaining the second dispersion curves for all scaling factor pairs, these curves are substituted into the pre-trained model in S3 for calculation, ultimately yielding a normalized velocity profile with a fixed resolution. Here, the fixed resolution refers to the number of thin layers in S1.

[0048] From this perspective, since the dimensions of the second dispersion curve and the first dispersion curve may differ significantly, as mentioned above, existing models require inputs of equal length, making it difficult for them to solve this problem. However, this embodiment of the invention, by setting up an embedding layer, a summing pooling layer, and a regression head, enables the model to handle data calculations with different input dimensions, thus broadening its application scope.

[0049] S6. Multiple normalized velocity profiles are inversely normalized based on scaling factor pairs to obtain multiple candidate inverted velocity profiles. Multiple third dispersion curves are generated by forward modeling multiple candidate inverted velocity profiles. The inversion result is selected based on the mismatch value between the third dispersion curve and the measured Rayleigh wave dispersion curve.

[0050] In this step, the inverse normalization refers to inverse normalizing the scale using a scaling factor. Specifically, the formula can be found as follows: z = z n / α H and V S =V S,n / α V In the formula, z and Vs represent the denormalized depth and the denormalized shear wave velocity, respectively. n and V S,n These represent the depth of the normalized velocity profile and the shear wave velocity of the normalized velocity profile, respectively.

[0051] Meanwhile, the method for selecting inversion results based on mismatch values ​​is as follows: select the result with the smallest mismatch value as the inversion result output, which is the optimal output; or, according to the set mismatch threshold, select the inversion result output that is less than the mismatch threshold. The mismatch threshold is calculated as follows: the square root of the reduced chi-square statistic of the third dispersion curve and the measured Rayleigh wave dispersion curve is less than 1. The inversion result selected by the mismatch threshold can be used for uncertainty analysis.

[0052] To further illustrate the technical solutions of the embodiments of the present invention, specific examples are given below.

[0053] First, the baseline depth of the underground model space is set at 100 meters, the baseline shear wave velocity at 1000 m / s, and the density of the underground model space is limited to 2000 kg / m³. 3 With a Poisson's ratio of 0.33, the strata above the reference depth in the underground model space were discretized into 100 thin layers, each with a thickness of 1 meter.

[0054] Subsequently, the speed range was set to 50–1000 m / s, and one million training samples were generated using a random stratification sampling method. Forward modeling was used to calculate the first dispersion curve corresponding to each sample, and each training sample and its corresponding first dispersion curve constituted a training data pair.

[0055] Subsequently, the PADIT model was trained using training data, resulting in a pre-trained PADIT model.

[0056] Example 1: The target dispersion curve is low-frequency data (approximately 0.03-0.12Hz), with a phase velocity as high as 3000-3800m / s, reflecting a deep crustal structure at the kilometer level. Based on this, the estimated semi-infinite layer depth range is [20, 80] kilometers, and the semi-space shear wave velocity range is [3800, 4800] m / s.

[0057] The final result is as follows Figure 1 As shown in a and b in the figure, the blue line represents the result with the minimum mismatch value, the light blue area represents the inversion result based on the mismatch threshold selection, and the vertical line represents the measured value. It can be seen that the method of the embodiment of the present invention has high accuracy when applied to the dispersion curve corresponding to the low frequency band and high phase velocity region.

[0058] Example 2: The target dispersion curve is mid-frequency data (approximately 3-15Hz), with a phase velocity range of 500-1300 m / s, reflecting a structure at a depth of several hundred meters. Based on this, the estimated depth range of the semi-infinite layer is [100, 400] meters, and the half-space shear wave velocity range is [1300, 2000] m / s.

[0059] The final result is as follows Figure 2 As shown in figures a and b, the blue line represents the result with the minimum mismatch value, the light blue area represents the inversion result based on the mismatch threshold selection, and the vertical line represents the measured value. It can be seen that the method of this embodiment of the invention has high accuracy when applied to the dispersion curve corresponding to the mid-frequency band and the mid-phase velocity region.

[0060] Example 3: The target dispersion curve is high-frequency data (approximately 20-50Hz), with a phase velocity range of 200-350m / s, used to detect near-surface shallow engineering sites at the 10-meter level. Based on this, the estimated semi-infinite layer depth range is [8, 20] meters, and the half-space shear wave velocity range is [350, 800] m / s.

[0061] The final result is as follows Figure 3 As shown in the figure, the blue line represents the result with the minimum mismatch value, the light blue area represents the inversion result based on the mismatch threshold selection, and the vertical line represents the measured value. It can be seen that the method of the embodiment of the present invention has high accuracy when applied to the dispersion curve corresponding to the high frequency band and low phase velocity region.

[0062] In summary, the method of this embodiment can be applied to the inversion of dispersion curves across multiple frequency bands and velocity scales, and the inversion results are highly accurate, providing data support for actual production.

[0063] It should be understood that this application is not limited to the methods described above, and various modifications and changes can be made without departing from its scope. The true scope is indicated by this application.

Claims

1. A parameter normalization-based intelligent inversion method for surface wave dispersion curves, characterized in that, The method comprises the following steps: S1, setting a reference subsurface model space with a reference depth and a reference shear wave velocity, the reference subsurface model space comprising an infinite downward extending semi-infinite layer below the reference depth and n thin layers above the reference depth discretized into fixed thickness, and setting constant density and Poisson's ratio for all layers, and the shear wave velocity of the semi-infinite layer being the reference shear wave velocity; S2, generating a plurality of training samples under the condition of a preset shear wave velocity using a random layer position allocation sampling method, each training sample comprising n shear wave velocities, performing forward calculation based on the training samples to obtain a first dispersion curve, and each training sample and the corresponding first dispersion curve constituting a training data pair; S3, constructing a PADIT model, and training the model based on the plurality of training data pairs to obtain a trained model; the PADIT model taking a Transformer model as a reference, first converting each data point in the first dispersion curve into a high-dimensional feature vector through an embedding layer, adding position encoding to each high-dimensional feature vector, then inputting the high-dimensional feature vector sequence with position encoding into a multi-layer Transformer encoder, and aggregating the data points corresponding to the first dispersion curve into a global descriptor with fixed dimensions through a summation pooling layer; S4, obtaining a measured Rayleigh wave dispersion curve of a reservoir to be inverted, and based on the maximum detection wavelength and the maximum phase velocity, estimating the depth range and shear wave velocity range of the semi-infinite layer in the real subsurface model space according to experience, and generating a plurality of parameter pairs of depth and shear wave velocity in the aforementioned range through sampling, mapping the parameter pairs to the reference subsurface model space through normalization operation, and obtaining a plurality of scale factor pairs, the scale factor pair consisting of a depth normalization scale factor and a velocity normalization scale factor; S5, based on the plurality of scale factor pairs, transforming the measured Rayleigh wave dispersion curve into the reference subsurface model space to obtain a plurality of second dispersion curves, and substituting the plurality of second dispersion curves into the pre-trained PADIT model to obtain a plurality of normalized velocity profiles; S6, based on the scale factor pairs, denormalizing the plurality of normalized velocity profiles to obtain a plurality of candidate inversion velocity profiles, generating a plurality of third dispersion curves through forward calculation of the plurality of candidate inversion velocity profiles, and selecting an inversion result based on the mismatch value between the third dispersion curve and the measured Rayleigh wave dispersion curve.

2. The method of claim 1, wherein, In S1, the reference depth in the reference subsurface model space is 100 m, and the reference shear wave velocity is 500-2000 m / s. The density and Poisson's ratio of each layer in the reference subsurface model space are the same.

3. The method of claim 2, wherein, In S2, in the preset shear wave velocity, the reference shear wave velocity is the shear wave velocity of the semi-infinite layer of the reference subsurface model space extending infinitely downward below the reference depth, and the shear wave velocity of the uppermost thin layer of the reference subsurface model space is not less than 1 / 20 of the reference shear wave velocity.

4. The method of claim 1, wherein, The random layer position allocation sampling method comprises the following steps: S21, setting the shear wave velocities of the uppermost thin layer and the lowermost thin layer above the reference depth of the reference subsurface model space; S22, randomly select a thin layer in the middle, based on the shear wave velocity of the two nearest thin layers with assigned velocity above and below the thin layer, perform uniform random sampling to assign the shear wave velocity to the thin layer; S23, repeat S22 until all thin layers are assigned with shear wave velocity; S24, repeat S21-S23 until a plurality of training samples are obtained, the number of training samples is 0.5 million to 5 million.

5. The method of claim 1, wherein, In S3, the PADIT model is used to optimize the model parameters with root mean square percentage error as the loss function, AdamW optimizer and cosine annealing learning rate scheduling strategy, and data enhancement technology is used to enhance the robustness and generalization ability of the model; the data enhancement technology refers to cutting the training data into different length fragments at random.

6. The method of claim 1, wherein, In S4, a plurality of parameter pairs are obtained by logarithmic uniform sampling method, and the number of parameter pairs is 10,000-160,000.

7. The method of claim 1, wherein, In S4, the calculation method of the scale factor pair is as follows: α H = H0 / H hs , α V = V S0 / V S,hs ; in the formula, α H and α V respectively represent the depth normalization scale factor and the velocity normalization scale factor; H0 represents a reference depth of the reference subsurface model space; H hs V represents a depth value of the parameter pair; V S0 V0 represents a reference shear wave velocity of the reference subsurface model space; V S,hs V represents a shear wave velocity of the parameter pair.

8. The method of claim 1, wherein, In S5, the method for obtaining the second dispersion curve is as follows: taking the scale factor pair, each data point on the measured Rayleigh wave dispersion curve is transformed into the reference subsurface model space by using the following formula: f n = (α V / α H ) x f, V R,n = α V x V R ; in the formula, f and V R respectively represent the frequency and phase velocity of each data point on the measured Rayleigh wave dispersion curve; f n and V R,n respectively represent the frequency and phase velocity of each data point on the second dispersion curve; α H and α V respectively represent the depth normalization scale factor and the velocity normalization scale factor; subsequently, the frequency and phase velocity of the data points on the plurality of second dispersion curves are combined into the second dispersion curve.

9. The method of claim 1, wherein, In S6, the inverse normalization refers to inverse normalization of the parameter pair by using the scale factor.

10. The method of claim 1, wherein, In S6, the result with the smallest mismatch value is selected as the inversion result output; or, according to the set mismatch threshold value, the inversion result smaller than the mismatch threshold value is selected as the output, and the calculation method of the mismatch threshold value is that the value of the square root of the reduced chi-square statistic of the third dispersion curve and the measured Rayleigh wave dispersion curve is less than 1.

Images