A Machine Learning Inversion Method and System Based on Joint Constraints of HVSR and Dispersion Curves
By using a machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves, the problems of large computational load, long time consumption and insufficient adaptability to noise interference in shallow shear wave velocity inversion are solved, and stable and accurate prediction of underground medium structure parameters is achieved.
Patent Information
- Application Number
- CN202611307518.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-27
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-08-27
AI Technical Summary
Existing shallow shear wave velocity inversion methods suffer from high computational complexity and time consumption when dealing with the inhomogeneity and complexity of subsurface media. They are also difficult to adapt to the frequency band variations, frequency gaps, and noise interference of actual data. Furthermore, they are difficult to distinguish the contributions of different observation information to layer thickness, layer interface depth, and shear wave velocity, which affects the stability and accuracy of the inversion results.
A machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves is adopted. By acquiring and processing dispersion curves and HVSR curves, dispersion information, HVSR information and effective frequency band range information are generated. The prediction stability and accuracy of underground medium structure parameters are improved by using baseline prediction results and gated residual correction.
While maintaining the stability of the basic constraints of the dispersion curve, the supplementary constraints of the HVSR curve are introduced, which improves the prediction stability and accuracy of the underground medium structure parameters and solves the problem of unstable and inaccurate inversion results in the existing technology.
Smart Images

Figure CN122836832B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geological exploration technology, specifically to a machine learning inversion method and system based on the collaborative constraints of HVSR curves and dispersion curves. Background Technology
[0002] Shallow shear wave velocity is a crucial parameter for shallow surface site characterization, engineering geophysical exploration, and seismic site response analysis, and is typically obtained through inversion from geophysical observation data. However, shallow subsurface media exhibit significant inhomogeneity and complexity in related technologies. The inversion process suffers from high computational complexity and time consumption, insufficient adaptability to frequency band variations, missing frequencies, and noise interference in actual data, and difficulty in distinguishing the contributions of different observational information to parameters such as layer thickness, inter-layer interface depth, and shear wave velocity, thus affecting the stability and accuracy of the inversion results. Summary of the Invention
[0003] This invention is based on the inventor's discoveries and understanding of the following facts and problems:
[0004] In related technologies, inversion methods for shallow shear wave velocities mainly include linearized iterative methods, global search methods, and machine learning methods. Linearized iterative methods typically rely on an initial model; when the initial model differs significantly from the actual subsurface structure, it is prone to getting trapped in local optima, leading to biased inversion results. While global search methods can search for a better model within a preset parameter space, they usually require extensive model sampling and forward modeling, resulting in high computational costs and long processing times, making them unsuitable for rapid inversion. Although machine learning methods can learn the nonlinear mapping relationship between observed curves and subsurface medium structural parameters through training data and achieve rapid prediction after training, existing methods lack adaptability to frequency band variations, frequency gaps, and noise interference in actual observation data. Furthermore, they do not fully utilize the complementary relationships between multi-source observation information, making it difficult to distinguish the effective contributions of different observation information to different subsurface medium structural parameters, which can easily affect the stability and accuracy of the inversion results.
[0005] To address the technical problems mentioned above, this invention provides a machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves, and a machine learning inversion system based on the joint constraints of HVSR curves and dispersion curves, thereby improving the stability and accuracy of underground medium structure parameter prediction.
[0006] In some embodiments, a machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves includes:
[0007] S1, obtain the dispersion curve and HVSR curve corresponding to the object to be inverted;
[0008] S2, process the dispersion curve and the HVSR curve to obtain dispersion information, HVSR information and effective frequency band range information;
[0009] S3. Based on the dispersion information and the effective frequency band range information, perform initial prediction of the underground medium structure parameters corresponding to the object to be inverted, and obtain the baseline prediction result.
[0010] S4, Based on the dispersion information, the HVSR information and the effective frequency band range information, generate correction information, and perform gated weighting on the correction information to obtain the gated residual correction amount used to correct the baseline prediction result;
[0011] S5. Based on the baseline prediction results and the gated residual correction amount, determine the target values of the underground medium structure parameters corresponding to the object to be inverted;
[0012] The underground medium structural parameters include at least layer thickness parameters and shear wave velocity parameters.
[0013] In some embodiments, S2 includes:
[0014] Adaptive frequency band filtering is performed on the dispersion curve and the HVSR curve to obtain the effective input frequency band of dispersion and the effective input frequency band of HVSR.
[0015] Based on the effective input frequency band of the dispersion, the dispersion curve is resampled and encoded to obtain the dispersion sequence input;
[0016] Based on the effective input frequency band of the HVSR, the HVSR curve is resampled and encoded to obtain the HVSR sequence input;
[0017] Feature extraction is performed on the dispersion sequence input to obtain the dispersion information, and feature extraction is performed on the HVSR sequence input to obtain the HVSR information;
[0018] Based on the effective input frequency band of the dispersion and the effective input frequency band of the HVSR, the frequency band coding feature is generated, and the frequency band coding feature serves as the effective frequency band range information;
[0019] The frequency band coding features include the upper and lower frequency limits of the effective input frequency band of the HVSR, as well as the upper and lower frequency limits of the effective input frequency band of the dispersion band.
[0020] In some embodiments, the adaptive bandwidth filtering of the dispersion curve and the HVSR curve to obtain the effective input bandwidth of dispersion and the effective input bandwidth of HVSR includes:
[0021] Obtain the original available frequency band of the HVSR curve and the original available frequency band of the dispersion curve;
[0022] Determine the main peak frequency of HVSR within the original available frequency band of the HVSR curve;
[0023] The upper limit frequency of the effective input frequency band of the HVSR is determined based on the amplitude attenuation characteristics to the right of the main peak frequency of the HVSR.
[0024] The lower limit frequency of the effective input frequency band of the dispersion is determined based on the main peak frequency of the HVSR and the original available frequency band of the dispersion curve.
[0025] Wherein, the effective input frequency band of the HVSR and the effective input frequency band of the dispersion have an overlapping frequency range, or the frequency interval between the effective input frequency band of the HVSR and the effective input frequency band of the dispersion is less than or equal to a preset interval threshold;
[0026] And / or, the step of resampling and encoding the dispersion curve according to the effective input frequency band of the dispersion to obtain a dispersion sequence input, and resampling and encoding the HVSR curve according to the effective input frequency band of the HVSR to obtain an HVSR sequence input, includes:
[0027] The dispersion curve located within the effective input frequency band of the dispersion and the HVSR curve located within the effective input frequency band of the HVSR are respectively resampled to a preset number of sampling points;
[0028] The normalized frequency value and the corresponding curve value of each sampling point are used as the features of the sampling point;
[0029] The dispersion sequence input and the HVSR sequence input are generated based on the features of multiple sampling points.
[0030] In some embodiments, the step of extracting features from the dispersion sequence input to obtain the dispersion information, and extracting features from the HVSR sequence input to obtain the HVSR information, includes:
[0031] The sequence input to be processed is sequentially subjected to multi-scale convolution processing, channel recalibration processing, residual connection processing, and group normalization processing to obtain the corresponding curve features;
[0032] Wherein, when the input sequence to be processed is the dispersion sequence input, the corresponding curve feature is the dispersion information; when the input sequence to be processed is the HVSR sequence input, the corresponding curve feature is the HVSR information.
[0033] In some embodiments, the multi-scale convolution processing includes:
[0034] Multiple convolutional kernels of different sizes are used to perform convolution processing on the sequence input to be processed, resulting in curve response features at multiple scales;
[0035] By splicing together the curve response features at multiple scales, the spliced channel features are obtained.
[0036] The spliced channel features are linearly combined using 1×1 convolution to obtain the fused curve intermediate features.
[0037] The sequence input to be processed is either the dispersion sequence input or the HVSR sequence input.
[0038] In some embodiments, the channel recalibration process includes:
[0039] Global average pooling is performed on the curve features to be recalibrated to obtain the channel description vector;
[0040] Calculate the channel weight corresponding to each channel based on the channel description vector;
[0041] The curve features to be recalibrated are weighted channel by channel according to the channel weights to obtain the recalibrated curve features.
[0042] In some embodiments, S3 includes:
[0043] Based on a preset baseline prediction multilayer perceptron network, the dispersion information and the effective frequency band range information are processed to obtain the baseline prediction result;
[0044] The S4 includes:
[0045] Based on a preset residual correction multilayer perceptron network, the residual correction amount is obtained by operating on the dispersion information, the HVSR information, and the frequency band coding features;
[0046] Based on a pre-defined gated multilayer perceptron network, gate weights are obtained by operating on the dispersion information, the HVSR information, and the frequency band coding features.
[0047] The gated residual correction amount is obtained by weighting the residual correction amount based on the gated weights.
[0048] In some embodiments, the residual correction includes residual components corresponding to each layer thickness parameter and each shear wave velocity parameter, and the gating weight includes weight components corresponding to each layer thickness parameter and each shear wave velocity parameter.
[0049] Each parameter in the underground medium structure parameters includes at least the layer thickness parameter and the shear wave velocity parameter;
[0050] The weighting of the residual correction amount based on the gate weights includes:
[0051] The corresponding residual components are weighted based on the weighted components to obtain the gated residual correction components corresponding to each parameter item.
[0052] In some embodiments, before obtaining the dispersion curve and HVSR curve corresponding to the object to be inverted, the method further includes: constructing and training a collaborative constraint model;
[0053] The collaborative constraint model is used to obtain the underground medium structure parameters corresponding to the object to be inverted based on the dispersion curve and the HVSR curve.
[0054] The construction and training of the collaborative constraint model includes:
[0055] Multiple original layered stratigraphic models are randomly generated, wherein the number of layers in the original layered stratigraphic models is variable;
[0056] Each of the original layered stratigraphic models is mapped to a standardized stratigraphic model with a fixed output dimension;
[0057] Forward modeling is performed on each of the standardized stratigraphic models to obtain the training dispersion curve and training HVSR curve corresponding to each of the standardized stratigraphic models;
[0058] Training samples are constructed based on the training dispersion curve, the training HVSR curve, and the corresponding underground medium structure parameters; wherein the training dispersion curve and the training HVSR curve are used as training input data, and the corresponding underground medium structure parameters are used as training labels.
[0059] The collaborative constraint model is trained based on the training samples.
[0060] According to a second aspect of this application, the present invention also provides a machine learning inversion system based on the joint constraints of HVSR curves and dispersion curves, comprising:
[0061] The curve acquisition unit is used to acquire the dispersion curve and HVSR curve corresponding to the object to be inverted;
[0062] The feature extraction unit is used to process the dispersion curve and the HVSR curve to obtain dispersion information, HVSR information and effective frequency band range information;
[0063] The baseline prediction unit is used to perform initial prediction of the underground medium structure parameters corresponding to the object to be inverted based on the dispersion information and the effective frequency band range information, and to obtain the baseline prediction result.
[0064] The correction calculation unit is used to generate correction information based on the dispersion information, the HVSR information and the effective frequency band range information, and to perform gated weighting on the correction information to obtain the gated residual correction amount used to correct the baseline prediction result.
[0065] The fusion output unit is used to determine the target value of the underground medium structure parameter corresponding to the object to be inverted based on the baseline prediction result and the gated residual correction amount.
[0066] The underground medium structural parameters include at least layer thickness parameters and shear wave velocity parameters.
[0067] In summary, the machine learning inversion method and system based on the collaborative constraints of HVSR curves and dispersion curves provided by this invention have at least the following technical effects: This invention first uses dispersion information and effective frequency band range information to form a baseline prediction result, and then uses HVSR information, dispersion information, and effective frequency band range information to form a gated residual correction amount. This allows HVSR information to participate in the inversion in a controlled manner, forming "baseline prediction result + gated residual correction amount" to obtain the target value. While maintaining the stability of the basic constraints of the dispersion curve, it introduces the HVSR curve to supplement the constraints of underground medium structure parameters, thereby improving the stability and accuracy of underground medium structure parameter prediction. Attached Figure Description
[0068] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:
[0069] Figure 1 This is a flowchart illustrating the machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves in an embodiment of the present invention.
[0070] Figure 2 This is a schematic diagram of the frequency band range of the dispersion curve and the HVSR curve in the machine learning inversion method based on the joint constraints of the HVSR curve and the dispersion curve in an embodiment of the present invention.
[0071] Figure 3 This is a schematic diagram of some models and their corresponding dispersion curves and HVSR curves in the machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves in the embodiments of the present invention; wherein, (a) is a composite HVSR curve; (b) is a composite dispersion curve; and (c) is a shear wave velocity model.
[0072] Figure 4 This is a schematic diagram of a multilayer perceptron network in the machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves in this embodiment of the invention.
[0073] Figure 5 This is a comparison chart of the results of machine learning inversion based on HVSR curve and dispersion curve co-constraint and Monte Carlo inversion for two test samples in this embodiment of the invention.
[0074] Among them, (a) is a comparison of the HVSR curve of the test sample 1 with the forward model HVSR curve of the prediction model and the optimal Monte Carlo model; (b) is a comparison of the Rayleigh wave dispersion curve of the test sample 1 with the forward model dispersion curve of the prediction model and the optimal Monte Carlo model; (c) is a comparison of the Vs model of the test sample 1, the prediction Vs model and the optimal Monte Carlo Vs model; (d) is a comparison of the HVSR curve of the test sample 2 with the forward model HVSR curve of the prediction model and the optimal Monte Carlo model; (e) is a comparison of the Rayleigh wave dispersion curve of the test sample 2 with the forward model dispersion curve of the prediction model and the optimal Monte Carlo model; and (f) is a comparison of the Vs model of the test sample 2, the prediction Vs model and the optimal Monte Carlo Vs model.
[0075] Figure 6 This is a schematic diagram of the structure of the machine learning inversion system based on the joint constraints of HVSR curves and dispersion curves provided in an embodiment of the present invention.
[0076] Figure 7 This is a schematic diagram of the hardware structure of the electronic device provided in an embodiment of the present invention.
[0077] Figure label:
[0078] 110. Curve acquisition unit; 120. Feature extraction unit; 130. Baseline prediction unit; 140. Correction calculation unit; 150. Fusion output unit;
[0079] 210. Processor; 220. Memory; 230. Communication interface; 240. Communication bus. Detailed Implementation
[0080] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.
[0081] refer to Figure 1 The present invention provides a machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves, which includes the following steps:
[0082] S1, obtain the dispersion curve and HVSR curve of the object to be inverted.
[0083] Specifically, the object to be inverted can be measured micro-motion data, active source surface wave data, passive source surface wave data, or curve data obtained after preprocessing such as denoising, smoothing, outlier removal, frequency normalization, and amplitude normalization. The dispersion curve can be a surface wave dispersion curve, used to characterize the variation of surface wave velocity with frequency; the HVSR curve can be a horizontal and vertical spectral ratio curve, used to characterize the variation of the horizontal and vertical component spectral ratios at different frequencies.
[0084] Among them, the dispersion curve and the HVSR curve reflect different response characteristics of the subsurface medium structure. The dispersion curve usually has a relatively stable constraint capability on the shear wave velocity structure, while the HVSR curve can provide supplementary information such as the main peak frequency, the main peak amplitude, and the post-peak attenuation. Thus, it can provide two types of complementary observation information for subsequent coordinated constraints, thereby improving the information basis for the inversion of subsurface medium structure parameters.
[0085] S2, process the dispersion curve and the HVSR curve to obtain dispersion information, HVSR information and effective frequency band range information.
[0086] Specifically, actual observed curves often suffer from problems such as inconsistent effective frequency bands, inconsistent number of frequency sampling points, inconsistent sampling intervals, and noise interference. The method provided by this invention can unify the curve input form through input processing, obtain curve shape expression through feature extraction, and improve the model's adaptability to actual observed data through effective frequency range information.
[0087] S3. Based on the dispersion information and the effective frequency band range information, perform initial prediction of the underground medium structure parameters corresponding to the object to be inverted, and obtain the baseline prediction result.
[0088] Specifically, the dispersion curve has a relatively stable basic constraint effect on the shallow shear wave velocity structure, and is suitable as the main source of information for initial prediction; the effective frequency band range information can represent the effective frequency range corresponding to the dispersion information, so that the model can distinguish the dispersion response under different frequency band conditions, thereby forming a relatively stable initial prediction of the underground medium structure parameters, providing a basic result for subsequent residual correction.
[0089] The baseline prediction result can be a parameter vector consisting of multiple layer thickness parameters and multiple shear wave velocity parameters.
[0090] S4. Based on the dispersion information, the HVSR information, and the effective frequency band range information, correction information is generated, and the correction information is gated and weighted to obtain the gated residual correction amount used to correct the baseline prediction result.
[0091] Specifically, while HVSR curves can provide supplementary information related to underground layered structures, their sensitivity to different parameter terms is inconsistent. Simply concatenating HVSR information with dispersion information for unified prediction may introduce invalid or noisy information. The method provided in this invention uses gated residuals to allow HVSR information to participate in correction in a controlled manner, thereby improving the stability and accuracy of collaborative constraints.
[0092] S5. Based on the baseline prediction results and the gated residual correction amount, determine the target values of the underground medium structure parameters corresponding to the object to be inverted.
[0093] Specifically, the baseline prediction results can provide stable initial inversion results, while the gated residual correction can provide controlled supplementary correction, so that the final output results have both the stabilizing constraint of the dispersion curve and the supplementary constraint of the HVSR curve, thereby improving the stability of the prediction of layer thickness parameters and shear wave velocity parameters.
[0094] The underground medium structural parameters include at least layer thickness parameters and shear wave velocity parameters. The layer thickness parameter characterizes the thickness of each layer of medium in the underground layered structure, and the shear wave velocity parameter characterizes the shear wave velocity of each layer of underground medium.
[0095] In summary, this invention employs a "baseline prediction result + gated residual correction" approach to obtain the target value. In related technologies, dispersion curves typically provide relatively stable constraints on the shear wave velocity structure, while HVSR curves can provide supplementary information such as peak frequency, peak amplitude, and post-peak attenuation. However, the sensitivity of HVSR curves to different parameter terms is inconsistent. Simply splicing the two types of curves together for inversion may introduce invalid or noise information into the prediction result. By first using dispersion information and effective bandwidth information to form a baseline prediction result, and then using HVSR information, dispersion information, and effective bandwidth information to form a gated residual correction, HVSR information can participate in the inversion in a controlled manner.
[0096] Therefore, the embodiments of the present invention can improve the stability and accuracy of underground medium structure parameter prediction by introducing HVSR curves to supplement the constraints of the dispersion curve basic constraints while maintaining the stability of the dispersion curve basic constraints.
[0097] In some embodiments, step S3 includes:
[0098] Based on a preset baseline prediction multilayer perceptron network, the dispersion information and the effective frequency band range information are processed to obtain the baseline prediction result.
[0099] Specifically, dispersion information reflects the shape of the dispersion curve, while effective bandwidth information reflects the effective input frequency range corresponding to the dispersion curve. Processing both using the baseline prediction multilayer perceptron network allows for simultaneous consideration of curve shape and bandwidth coverage differences during initial prediction, improving the adaptability of the baseline prediction results to different input bandwidth conditions. For example, within the baseline prediction multilayer perceptron network, the dispersion information vector and the bandwidth-coded feature vector from the effective bandwidth information can be concatenated along the feature dimension to form the baseline prediction input vector.
[0100] It should be noted that the baseline prediction multilayer perceptron network may include an input layer, one or more hidden layers, and an output layer. The hidden layers may employ ReLU, GELU, LeakyReLU, or other nonlinear activation functions. The output dimension of the baseline prediction result may be consistent with the output dimension of the subsurface medium structure parameters.
[0101] Step S4 includes:
[0102] Based on a preset residual correction multilayer perceptron network, the residual correction amount is calculated from the dispersion information, the HVSR information, and the effective frequency band range information.
[0103] Based on a pre-defined gated multilayer perceptron network, the gate weights are calculated from the dispersion information, the HVSR information, and the effective frequency band range information.
[0104] The gated residual correction amount is obtained by weighting the residual correction amount based on the gated weights.
[0105] Specifically, the dispersion information, the HVSR information, and the effective frequency band range information can be concatenated to form a residual correction input vector, which is then input into a residual correction multilayer perceptron network to output a residual correction amount corresponding to the dimension of the underground medium structure parameters. That is, the residual correction amount needs to simultaneously consider the dispersion curve information, the HVSR curve information, and the effective frequency band range information, so that the supplementary information provided by the HVSR information can act on the baseline prediction result in the form of residuals, rather than replacing the baseline prediction result, thereby maintaining the stability of the initial prediction.
[0106] For example, the residual correction amount can be a positive value, a negative value, or a value close to zero, which respectively represent increasing the correction, decreasing the correction, or basically no correction to the baseline prediction results.
[0107] Gating weights are used to control the degree to which residual corrections are applied to corresponding subsurface structural parameters. A larger gating weight indicates that the corresponding parameter utilizes more residual corrections; a smaller gating weight indicates that the corresponding parameter utilizes less residual correction. In related technologies, different subsurface structural parameters have varying sensitivities to the HVSR curve, and the same residual correction intensity cannot be applied to all parameters. Using spliced features that are the same as or similar to the residual correction input as gating inputs, and outputting gating weights through a gated multilayer perceptron network, allows the model to adaptively determine whether each parameter should fully utilize residual corrections based on the input features, improving the controllability of the residual correction process. Gating weights can be consistent with the dimension of the subsurface structural parameters or output as parameter groups.
[0108] Subsequently, the gate weights and residual corrections are multiplied element-wise, or weighted according to the parameter group, to obtain the gated residual correction. In related technologies, the residual correction itself represents the correction direction and magnitude, but whether it applies to a certain parameter needs to be controlled by the gate weights. The method of this invention can avoid the residual correction directly acting on all parameters, which would lead to erroneous corrections, thereby improving the stability of the final inversion result.
[0109] Baseline prediction primarily utilizes dispersion curves to form stable baseline results; residual correction mainly uses the joint information of HVSR curves and dispersion curves to provide supplementary corrections; gating control is used to determine whether these supplementary corrections should be fully applied to each parameter. Through the above execution process, the problem of unclear functional boundaries caused by a single network simultaneously undertaking initial prediction and supplementary correction functions can be avoided, resulting in a clearer technical division of labor in the collaborative constraint process.
[0110] Furthermore, the gated multilayer perceptron network can output gate weights using a sigmoid function, ensuring the gate weights are between 0 and 1. These gate weights represent the degree to which residual correction is applied. Without range constraints on the gate weights, the residual correction may be abnormally amplified. This embodiment of the invention, by limiting the gate weights to between 0 and 1, makes the residual correction process more controllable and reduces the risk of excessive shift in the inversion results due to residual correction.
[0111] In some embodiments, the baseline prediction multilayer perceptron network, the residual correction multilayer perceptron network, and the gated multilayer perceptron network can be set independently, or they can share a portion of the pre-feature mapping layer before setting the baseline prediction branch, residual correction branch, and gated weight branch separately. In related technologies, different network branches perform different functions, but their input features share certain commonalities. Setting them independently can enhance the functional differentiation of each branch, while sharing the pre-feature mapping layer can reduce the number of model parameters and improve training efficiency, thus providing a flexible implementation method for model deployment under different computing resource conditions.
[0112] In some embodiments, the residual correction includes residual components corresponding to each layer thickness parameter and each shear wave velocity parameter, and the gating weight includes weight components corresponding to each layer thickness parameter and each shear wave velocity parameter.
[0113] Each parameter in the underground medium structure parameters includes at least the layer thickness parameter and the shear wave velocity parameter;
[0114] The weighting of the residual correction amount based on the gate weights includes:
[0115] The corresponding residual components are weighted based on the weighted components to obtain the gated residual correction components corresponding to each parameter item.
[0116] Specifically, if the subsurface medium structure parameters include multiple layer thickness parameters and multiple shear wave velocity parameters, the residual correction can be a vector with the same dimension as the subsurface medium structure parameters, and the gating weights can also be vectors with the same dimension as the subsurface medium structure parameters. For each parameter, the residual component corresponding to that parameter is weighted using the weight component corresponding to that parameter to obtain the gating residual correction component corresponding to that parameter. This allows the residual correction to be applied according to the sensitivity differences of different parameter items, thereby improving the prediction stability of layer thickness parameters and shear wave velocity parameters.
[0117] In embodiments of the present invention, the residual components in the residual correction amount can first be determined, allowing the residual correction amount to provide an independent correction direction and magnitude for each parameter. For example, the residual correction multilayer perceptron network can output a residual vector with the same dimension as the subsurface medium structure parameter. Each component in this residual vector corresponds to a subsurface medium structure parameter, such as a layer thickness parameter or a shear wave velocity parameter.
[0118] Then, the weight components in the gating weights are determined so that each parameter has an independent residual adoption level, improving the differential correction capability. For example, the gating multilayer perceptron network output can be a gating weight vector with the same dimension as the residual correction amount. Each weight component corresponds to one residual component.
[0119] Finally, the corresponding residual components are weighted based on each weighted component to obtain the gated residual correction component for each parameter item. This reduces the risk of incorrect correction of insensitive parameter items and improves the prediction stability of layer thickness and shear wave velocity parameters. For example, in an embodiment of the present invention, the i-th weighted component can be multiplied by the i-th residual component to obtain the i-th gated residual correction component.
[0120] In some embodiments, the gating weights can also be set according to parameter groups. For example, weights can be set separately for the layer thickness parameter group and the shear wave velocity parameter group, or weights can be set separately for shallow, middle and deep parameters. This can reduce the output dimension and model complexity of the gating weights while still retaining the ability to differentiate between different parameter types or different depth ranges.
[0121] In the method provided in the embodiments of the present invention, the gate weight and the residual correction amount can be multiplied element by element to obtain the gated residual correction amount; then the baseline prediction result is added to the gated residual correction amount to obtain the target value of the underground medium structure parameter corresponding to the object to be inverted.
[0122] Specifically, if the baseline prediction result, residual correction, and gating weights are all vectors of the same dimension, then element-wise multiplication represents the application of the i-th gating weight to the i-th residual component, resulting in the i-th gating residual correction component. Subsequently, the i-th baseline prediction parameter is added to the i-th gating residual correction component to obtain the target value of the i-th subsurface medium structure parameter.
[0123] Through the above fusion process, supplementary corrections can be introduced while maintaining the stability of the baseline prediction results, avoiding excessive shift of some parameters caused by the direct superposition of residual corrections.
[0124] In some embodiments, the fusion method between the baseline prediction result and the gated residual correction can also include weighted summation, summation after residual normalization, or summation within a constraint range. For example, a non-negative constraint can be applied to the layer thickness parameter after summation, or a velocity range constraint can be applied to the shear wave velocity parameter. In this embodiment of the invention, by applying physical constraints to the fusion output result, the probability of negative layer thickness, unreasonable shear wave velocity, or abnormal interface structure can be reduced, thereby improving the physical rationality of the inversion result.
[0125] In some embodiments, step S2 includes:
[0126] Adaptive bandwidth selection is performed on the dispersion curve and HVSR curve to obtain the effective input bandwidth of dispersion and HVSR.
[0127] Based on the effective input frequency band of dispersion, the dispersion curve is resampled and encoded to obtain the dispersion sequence input;
[0128] Based on the effective input frequency band of HVSR, the HVSR curve is resampled and encoded to obtain the HVSR sequence input;
[0129] Feature extraction is performed on the input dispersion sequence to obtain dispersion information, and feature extraction is performed on the input HVSR sequence to obtain HVSR information;
[0130] Based on the effective input frequency band of dispersion and the effective input frequency band of HVSR, frequency band coding features are generated, and the frequency band coding features serve as the effective frequency band range information.
[0131] The frequency band coding features include the upper and lower limit frequencies of the effective input frequency band of the HVSR, as well as the upper and lower limit frequency coding of the effective input frequency band of the dispersion.
[0132] Specifically, the effective input frequency bands for the two types of curves are first determined based on the available frequency band of the curve itself, the main peak characteristics of HVSR, the post-peak attenuation characteristics, and the available frequency range of the dispersion curve. Then, within the effective input frequency bands, the dispersion curve and the HVSR curve are resampled respectively, ensuring that the input curves of different samples have a uniform length. Afterwards, the resampled dispersion curve and HVSR curve are input into the feature extraction network respectively to obtain dispersion information and HVSR information. Simultaneously, frequency band coding features are generated based on the effective input frequency bands of the two types of curves.
[0133] Specifically, based on the main peak position of the HVSR curve, the post-peak attenuation characteristics, the original available frequency band of HVSR, and the original available frequency band of the dispersion curve, the effective input frequency bands of the two types of curves are determined. The effective frequency bands are dynamically selected according to the curves themselves to reduce the impact of invalid and noise frequency bands on subsequent feature extraction. Curve data within the effective input frequency band of the dispersion curve is extracted and resampled into a preset number of sampling points. The normalized frequency value and corresponding curve value of each sampling point are then used as the sampling point features, thereby converting the dispersion curve into a sequence input with a fixed length and fixed feature dimension, which meets the input requirements of the neural network. Curve data within the effective input frequency band of the HVSR curve is extracted and resampled and encoded in a similar manner to the dispersion curve, so that both the HVSR curve input format and the dispersion curve input format meet the fixed dimension requirements, which is convenient for processing by the dual-branch network. The one-dimensional convolutional feature extraction branches corresponding to the dispersion sequence input and the HVSR sequence input are respectively used to obtain their respective curve features. Finally, the upper and lower limits of the two types of effective input frequency bands are normalized or embedded to form frequency band coding features, which enable the subsequent baseline prediction, residual correction and gating weight generation processes to perceive the difference in input frequency coverage.
[0134] The embodiments of the present invention can solve the problems of inconsistent input formats and loss of frequency band information by combining adaptive frequency band selection, resampling coding, feature extraction and frequency band coding feature generation.
[0135] In some embodiments, feature extraction may not be limited to one-dimensional convolutional neural networks, but may also employ recurrent neural networks, Transformer encoders, temporal attention networks, fully connected networks, or combinations thereof. In related technologies, dispersion curves and HVSR curves both belong to sequential data arranged along the frequency direction, and different neural network structures can extract sequence relationships from different angles. Embodiments of this invention can select a suitable feature extraction network based on data scale, computational resources, and the complexity of curve features, thereby improving the flexibility of the method.
[0136] In some embodiments, the adaptive bandwidth filtering of the dispersion curve and the HVSR curve to obtain the effective input bandwidth of dispersion and the effective input bandwidth of HVSR includes:
[0137] Obtain the original available frequency band of the HVSR curve and the original available frequency band of the dispersion curve;
[0138] Determine the main peak frequency of HVSR within the original available frequency band of the HVSR curve;
[0139] The upper limit frequency of the effective input frequency band of the HVSR is determined based on the amplitude attenuation characteristics to the right of the main peak frequency of the HVSR.
[0140] The lower limit frequency of the effective input frequency band of the dispersion is determined based on the main peak frequency of the HVSR and the original available frequency band of the dispersion curve.
[0141] Wherein, the effective input frequency band of the HVSR and the effective input frequency band of the dispersion have an overlapping frequency range, or the frequency interval between the effective input frequency band of the HVSR and the effective input frequency band of the dispersion is less than or equal to a preset interval threshold.
[0142] Specifically, the original usable frequency bands of the HVSR curve and dispersion curve can be read first. The original usable frequency bands can be determined based on the data acquisition range, signal-to-noise ratio, frequency response range, curve continuity, or manual preprocessing results. Then, the main peak frequency of the HVSR curve is identified within the original usable frequency band. The main peak frequency of the HVSR curve can be determined by finding the point of maximum amplitude, or by identifying local peaks after smoothing the curve.
[0143] After determining the main peak frequency of the HVSR, the upper limit frequency of the effective input frequency band of the HVSR is determined based on the amplitude attenuation characteristics to the right of the main peak. This upper limit frequency can be the frequency corresponding to when the amplitude to the right of the main peak attenuates to a preset proportion of the peak value, or it can be determined based on the slope change after the peak, the continuity of the curve, or the signal-to-noise ratio threshold. Then, the lower limit frequency of the effective input frequency band of the dispersion curve is determined based on the main peak frequency of the HVSR and the original available frequency band, so that the effective input frequency band of the HVSR and the effective input frequency band of the dispersion curve have a frequency connection.
[0144] In related technologies, the high-frequency band of the HVSR curve is easily affected by noise, while the main peak and a certain range after the peak usually contain more information related to the underground medium structure. At the same time, if the dispersion curve and the HVSR curve are too far apart in the input frequency range, the constraints provided by the two types of curves may be discontinuous, which is not conducive to cooperative constraints.
[0145] This step allows for the dynamic determination of the input frequency range for each sample, improving the effectiveness of the input frequency band and reducing the impact of noise bands on the inversion results.
[0146] In some embodiments, the frequency corresponding to the attenuation of the amplitude to 30% of the peak value on the right side of the HVSR main peak can be used as the upper limit frequency of the effective input frequency band of the HVSR. After the attenuation to a certain proportion on the right side of the HVSR main peak, the structural response related to the main peak in the curve gradually weakens, and high-frequency noise or invalid information may increase. By determining the cutoff frequency based on the attenuation ratio after the peak, this embodiment of the invention can retain the information of the HVSR main peak and its effective attenuation segment, while reducing the entry of high-frequency invalid information into the model.
[0147] In some embodiments, the effective input frequency band of the HVSR and the effective input frequency band of the dispersion curve may have overlapping frequency ranges; if they cannot overlap, a preset interval threshold can be set so that the frequency interval between them does not exceed the threshold. In related technologies, if the frequency ranges of the two types of curves are too far apart, it may lead to discontinuity in frequency coverage in the cooperative constraint. The embodiments of the present invention can enhance the complementarity of the HVSR curve and the dispersion curve in the frequency range by overlapping frequency bands or connecting with small intervals, thereby improving the continuity of the cooperative constraint input information.
[0148] In some embodiments, the step of resampling and encoding the dispersion curve according to the effective input frequency band of the dispersion to obtain a dispersion sequence input, and resampling and encoding the HVSR curve according to the effective input frequency band of the HVSR to obtain an HVSR sequence input, includes:
[0149] The dispersion curve located within the effective input frequency band of the dispersion and the HVSR curve located within the effective input frequency band of the HVSR are respectively resampled to a preset number of sampling points;
[0150] The normalized frequency value and the corresponding curve value of each sampling point are used as the features of the sampling point;
[0151] The dispersion sequence input and the HVSR sequence input are generated based on the features of multiple sampling points.
[0152] Specifically, for dispersion curves, curve data within the effective input frequency band of dispersion can be extracted first. Then, several sampling frequency points are re-determined within this frequency band according to preset sampling rules. Based on the curve values of the original dispersion curve at the corresponding frequency positions, the curve values corresponding to each sampling point are obtained through interpolation, fitting, or nearest-neighbor methods, thus forming the resampled dispersion curve. For HVSR curves, the same or similar methods can be used to re-determine several sampling frequency points within the effective input frequency band of HVSR and obtain the HVSR curve values corresponding to each sampling point, thus forming the resampled HVSR curve.
[0153] For any sampling point, the frequency value corresponding to that sampling point can be normalized to obtain a normalized frequency value; simultaneously, the dispersion curve value or HVSR curve value corresponding to that sampling point is used as the curve value. Thus, each sampling point can consist of at least two features: a normalized frequency value and a corresponding curve value. The normalized frequency value characterizes the relative frequency position of the sampling point within the effective input frequency band, while the curve value characterizes the magnitude of the observed response at that frequency position.
[0154] For example, the sampling point features of multiple sampling points within the effective input frequency band of the dispersion sequence can be arranged in ascending order of frequency to form a dispersion sequence input; similarly, the sampling point features of multiple sampling points within the effective input frequency band of the HVSR sequence can be arranged in ascending order of frequency to form an HVSR sequence input. Thus, both the dispersion sequence input and the HVSR sequence input can be represented as two-dimensional sequence data, with one dimension corresponding to the sampling point sequence and the other dimension corresponding to the sampling point features.
[0155] In some optional embodiments, the preset number can be 100 sampling points. In this case, each sampling point includes two features: a normalized frequency value and a corresponding curve value. Therefore, the dispersion curve and the HVSR curve can each form a 100×2 sequence input. The 100 sampling points can control the network input dimension and computational load while maintaining sufficient curve morphology information, allowing the input data to express both local details and overall trends of the curve, while reducing the computational burden of model training and inference.
[0156] In some alternative embodiments, the preset number can also be 64, 128, 200, or other numbers. For example, the curve frequency range, frequency resolution, and computing resources vary in different application scenarios, and the sequence input length can be flexibly adjusted according to the target inversion accuracy and computing resources to improve the applicability of the method.
[0157] In some alternative embodiments, the normalized frequency values can be obtained through linear normalization or logarithmic frequency normalization. For example, the lower limit of the effective input frequency band can be mapped to 0, the upper limit of the effective input frequency band can be mapped to 1, and intermediate frequency values can be mapped to the range of 0 to 1. This reduces the frequency scale differences between different samples through frequency normalization, thereby improving the stability of model training.
[0158] In some embodiments, the frequency band coding features include upper and lower frequency limits of the HVSR effective input frequency band, and upper and lower frequency limit coding of the dispersion effective input frequency band.
[0159] Specifically, the lower limit of the effective input frequency band of HVSR, the upper limit of the effective input frequency band of HVSR, the lower limit of the effective input frequency band of dispersion, and the upper limit of the effective input frequency band of dispersion can be used as basic frequency band boundary parameters. These frequency boundary parameters can be directly normalized and used as frequency band coding features, or they can be processed through logarithmic transformation, standardization, fully connected mapping, or position coding to form frequency band coding features.
[0160] In related techniques, although the resampled dispersion sequence input and HVSR sequence input have a fixed length, a fixed-length sequence cannot fully represent the true frequency range covered by the original curve. Even if different samples are resampled to the same number of sampling points, their frequency coverage may differ. Band coding features can supplement this frequency range information.
[0161] By leveraging frequency band coding features, subsequent baseline prediction, residual correction, and gating weight generation processes can simultaneously utilize curve morphology information and effective frequency band range information, thereby improving the model's adaptability to different frequency band input conditions.
[0162] In some embodiments, the step of extracting features from the dispersion sequence input to obtain the dispersion information, and extracting features from the HVSR sequence input to obtain the HVSR information, includes:
[0163] The sequence input to be processed is sequentially subjected to multi-scale convolution processing, channel recalibration processing, residual connection processing, and group normalization processing to obtain the corresponding curve features;
[0164] Wherein, when the input sequence to be processed is the dispersion sequence input, the corresponding curve feature is the dispersion information; when the input sequence to be processed is the HVSR sequence input, the corresponding curve feature is the HVSR information.
[0165] Specifically, when the input sequence to be processed is a dispersive sequence, the corresponding curve feature is dispersive information; when the input sequence to be processed is an HVSR sequence, the corresponding curve feature is HVSR information. Dispersive and HVSR sequence inputs can be input into two separate one-dimensional convolutional neural network branches. The two branches can have the same structure, or different kernel sizes, number of channels, and network depths can be set according to the differences between the two types of curves.
[0166] In related technologies, although both dispersion curves and HVSR curves are frequency sequence data, they have different physical meanings. Dispersion curves emphasize the trend of velocity change with frequency, while HVSR curves emphasize spectral information such as the main peak, peak amplitude, and post-peak attenuation. Extracting features from the two types of curves separately can avoid the simple mixing of curves with different physical meanings in the early feature extraction stage.
[0167] In this embodiment, the input sequence to be processed is first subjected to multi-scale convolution, the features after multi-scale convolution are then recalibrated by channels, the features after channel recalibration are then processed by residual connections, and the residual fused features are then processed by group normalization to obtain the corresponding curve features. Thus, by combining multi-scale convolution, channel recalibration, residual connections and group normalization, curve morphology information at different scales can be extracted, the effective channel response can be enhanced, the information transmission of deep networks can be improved, and the feature distribution can be stabilized.
[0168] In some embodiments, the order of multi-scale convolution processing, channel recalibration processing, residual connection processing, and group normalization processing can be adjusted according to the network structure. For example, group normalization can be placed after convolution processing and before the activation function; residual connections can also be combined with normalization processing to form residual blocks. In related technologies, the combination of convolution, activation, normalization, and residual connections may differ in different network structures. The embodiments of the present invention improve the flexibility of feature extraction network structure design while maintaining functions such as multi-scale feature extraction, channel weight adjustment, improved information transmission, and stable feature distribution.
[0169] In some embodiments, the multi-scale convolution processing includes:
[0170] Multiple convolutional kernels of different sizes are used to perform convolution processing on the sequence input to be processed, resulting in curve response features at multiple scales;
[0171] By splicing together the curve response features at multiple scales, the spliced channel features are obtained.
[0172] The spliced channel features are linearly combined using 1×1 convolution to obtain the fused curve intermediate features.
[0173] The sequence input to be processed is either the dispersion sequence input or the HVSR sequence input.
[0174] Specifically, multiple parallel convolutional branches can be set up, each using a convolutional kernel of a different size. Different sized kernels have different receptive fields, thus capturing curve variations within different frequency neighborhoods; for example, local details, second-scale variations, and overall trends can be extracted simultaneously. The outputs of each convolutional branch are then concatenated along the channel dimension to obtain concatenated channel features. Subsequently, a 1×1 convolution is used to fuse the concatenated channels to obtain the intermediate features of the curve.
[0175] In related technologies, single-scale convolutions struggle to simultaneously capture local peak details, curve slope variations, and overall trends across a wide frequency band. Multi-scale convolutions can extract complementary features across different receptive fields, while 1×1 convolutions can integrate multi-scale channels.
[0176] This step provides more complete intermediate curve features for subsequent channel recalibration, residual connection, and group normalization.
[0177] In some embodiments, multiple convolutional kernels of different sizes may include two, three, or more kernels; the kernel size can be 3, 5, 7, 9, etc., and can also be set according to the sequence length and frequency resolution. Multi-scale fusion methods are not limited to 1×1 convolution; weighted summation, attention fusion, gated fusion, or fully connected mapping after concatenation can also be used. In related technologies, the suitable receptive field and fusion method may differ under different numbers of sampling points and different frequency resolutions. Embodiments of this invention can select a suitable multi-scale convolution configuration based on curve sampling conditions and computational resources to improve the extraction capability of local details and overall trends of the curve.
[0178] In some embodiments, using multiple convolution kernels of different sizes to perform convolution processing on the sequence input to be processed may include:
[0179] The first-scale convolution kernel is used to extract local amplitude fluctuations and detailed changes near the peak.
[0180] The second-scale convolution kernel is used to extract the curve slope variation features;
[0181] The overall trend features within a wider frequency band are extracted using a third-scale convolution kernel.
[0182] Specifically, for HVSR sequence input, the first-scale convolutional kernel can capture local amplitude changes near the main peak, the second-scale convolutional kernel can capture slope changes in the rising segment before the peak and the decay segment after the peak, and the third-scale convolutional kernel can capture the overall spectral shape in the low-frequency to high-frequency range. For dispersion sequence input, the first-scale convolutional kernel can capture local velocity changes, the second-scale convolutional kernel can capture the slope of the dispersion curve, and the third-scale convolutional kernel can capture the overall dispersion trend.
[0183] In related technologies, the structural parameters of underground media may be correlated with local details, local slopes, and overall trends of curves. Extracting only single-scale features may result in the loss of information on local peaks, slope variations, or overall trends.
[0184] This step enhances the integrity of dispersion and HVSR information, thereby improving the accuracy of subsequent parameter predictions.
[0185] In some embodiments, the first scale, second scale, and third scale are relative concepts, and the specific convolutional kernel size can be determined based on the input sequence length, sampling density, and computational resources. Dilated convolution, deformable convolution, or self-attention mechanisms can also be used to expand the receptive field to extract multi-scale curve responses. In related technologies, a fixed convolutional kernel size may not be suitable for all frequency sampling conditions. The embodiments of this invention, by flexibly adjusting the receptive field range, can improve the network's adaptability to curves with different frequency resolutions and avoid the loss of local information or broadband trend information.
[0186] In some embodiments, the channel recalibration process includes: performing global average pooling on the curve features to be recalibrated to obtain a channel description vector; calculating the channel weights corresponding to each channel based on the channel description vectors; and performing channel-by-channel weighting on the curve features to be recalibrated based on the channel weights to obtain the recalibrated curve features.
[0187] Specifically, the curve features to be recalibrated can be the intermediate features of the fused curve obtained through multi-scale convolution processing. Global average pooling along the frequency sequence direction yields the overall response level of each channel across the entire frequency range. Then, channel weights are calculated based on the channel description vectors, and these weights are used to weight each channel.
[0188] In related techniques, different channels obtained from multi-scale convolution contribute differently to the prediction of subsurface medium structure parameters. Some channels may primarily respond to effective peaks, slopes, or trend information, while others may primarily respond to noise or repetitive information. Channel recalibration can enhance effective channels and suppress low-contribution channels.
[0189] This step can improve the quality of curve features and enhance the effectiveness of subsequent residual connection and group normalization processing.
[0190] In some embodiments, the channel description vector can be sequentially input into two fully connected layers, a ReLU activation function, and a Sigmoid activation function to obtain the channel weights corresponding to each channel. In related technologies, there may be a nonlinear relationship between channel importance and underground medium structural parameters, which a single linear mapping cannot fully express. This invention generates channel weights through two fully connected layers and a nonlinear activation function, which can improve the ability of channel recalibration to distinguish between effective and noisy channels.
[0191] In some embodiments, channel weights can also be generated using attention mechanisms, convolutional attention modules, gating networks, or other learnable weight generation methods. Global average pooling can also be replaced by global max pooling, a combination of average and max pooling, or frequency-weighted pooling. In related technologies, effective curve responses in different data may exhibit global average responses, local peak responses, or responses in specific frequency bands. Embodiments of the present invention can also enhance the adaptability of channel recalibration to different types of effective curve responses by selecting different pooling and weight generation methods.
[0192] In some embodiments, residual connection processing includes: obtaining the input features of residual connection processing and the recalibrated curve features; when the number of channels of the two features is the same, fusing the two features to obtain residual fusion features; when the number of channels of the two features is different, performing channel mapping on the input features of residual connection processing using 1×1 convolution, and then fusing the channel-mapped input features with the recalibrated curve features to obtain residual fusion features.
[0193] Specifically, the input features for residual connection processing can be the features before entering the current feature processing block, and the recalibrated curve features can be the features output by the current feature processing block. When the number of channels is the same, they can be directly added; when the number of channels is different, the input features can be mapped to the same number of channels as the output features through a 1×1 convolution before fusion.
[0194] In related technologies, deepening convolutional networks may lead to difficulties in information transfer or gradient propagation. Residual connections can preserve the original response information in the input features, enabling the network to avoid losing effective information while learning new features.
[0195] This step can improve the training stability and information transmission capability of deep feature extraction networks.
[0196] In some embodiments, when the frequency sequence lengths of the input features processed by the residual connection are inconsistent with those of the recalibrated curve features, padding, pooling, or interpolation can be used to make them consistent in the frequency dimension before fusion. In related technologies, convolution processing, pooling processing, or different network branches may cause inconsistencies in the frequency sequence lengths of the input and output features. This embodiment of the invention, through frequency dimension alignment, ensures stable execution of the residual connection and avoids feature fusion failure due to dimension mismatch.
[0197] In some embodiments, the fusion method can be addition, weighted addition, concatenation mapping, or other feature fusion methods. Channel mapping methods are not limited to 1×1 convolution; they can also be linear mapping, projection transformation, channel padding, or channel pruning. In related technologies, the number of channels, sequence length, and semantic level of input and output features may differ in different network structures. Embodiments of this invention can select a suitable residual fusion method based on feature dimension matching, thereby improving the applicability of residual connection structures.
[0198] In some embodiments, the group normalization process includes: grouping the residual fusion features to obtain multiple channel groups; calculating the mean and variance of the feature values within each channel group; normalizing the feature values within the corresponding channel group based on the mean and variance to obtain the group-normalized curve features; wherein the group-normalized curve features are used as the corresponding curve features.
[0199] Specifically, the channels of the residual fusion features can be divided into several channel groups. Within each channel group, the mean and variance are calculated, and the feature values within that group are normalized accordingly. After normalization, affine transformations can be performed using learnable scaling and translation parameters to enhance the network's expressive power.
[0200] In related technologies, variations in training batch size, sample frequency range, and curve shape can all lead to instability in feature distribution. Group normalization does not depend on batch statistics and is suitable for training scenarios with small or varying batch sizes.
[0201] This step stabilizes the feature distribution, improving model training stability and convergence.
[0202] In some embodiments, the maximum number of groups for group normalization can be set to 8. If the number of channels is divisible by 8, the channel features are divided into 8 channel groups. If the number of channels is not divisible by 8, the largest integer less than 8 that is divisible by the number of channels is selected as the actual number of groups, and grouping is performed according to the actual number of groups. In related technologies, group normalization requires dividing channels into several groups; if the number of groups is not divisible by the number of channels, stable grouping is difficult. This embodiment of the invention, by adaptively determining the actual number of groups based on the number of channels, enables stable execution of group normalization for different channel scales.
[0203] In some embodiments, group normalization can also be replaced by layer normalization, instance normalization, batch normalization, or a combination of the above normalization methods. In related technologies, the suitable normalization method may differ depending on the training batch size, sample distribution, and network structure. Embodiments of the present invention can select a suitable normalization method according to the training conditions to improve the stability of feature distribution and model convergence performance.
[0204] In some embodiments, before obtaining the dispersion curve and HVSR curve corresponding to the object to be inverted, the method further includes: constructing and training a collaborative constraint model;
[0205] The collaborative constraint model is used to obtain the underground medium structure parameters corresponding to the object to be inverted based on the dispersion curve and the HVSR curve.
[0206] The construction and training of the collaborative constraint model includes:
[0207] Multiple original layered stratigraphic models are randomly generated, wherein the number of layers in the original layered stratigraphic models is variable;
[0208] Each of the original layered stratigraphic models is mapped to a standardized stratigraphic model with a fixed output dimension;
[0209] Forward modeling is performed on each of the standardized stratigraphic models to obtain the training dispersion curve and training HVSR curve corresponding to each of the standardized stratigraphic models;
[0210] Training samples are constructed based on the training dispersion curve, the training HVSR curve, and the corresponding underground medium structure parameters; wherein the training dispersion curve and the training HVSR curve are used as training input data, and the corresponding underground medium structure parameters are used as training labels.
[0211] The collaborative constraint model is trained based on the training samples.
[0212] Specifically, multiple original layered stratigraphic models can be randomly generated, with a variable number of layers. Each original layered stratigraphic model is mapped to a standardized stratigraphic model with a fixed output dimension. Forward modeling is performed on each standardized stratigraphic model to obtain training dispersion curves and training HVSR curves. Training samples are constructed based on the training dispersion curves, training HVSR curves, and corresponding subsurface medium structure parameters. The training dispersion curves and training HVSR curves serve as training input data, and the corresponding subsurface medium structure parameters serve as training labels. Finally, the collaborative constraint model is trained based on the training samples.
[0213] By randomly generating layered stratigraphic models and performing forward modeling to generate curves, large-scale supervised training samples can be constructed. However, in related technologies, the number of layers in the original layered stratigraphic model is variable, and directly using it as network labels leads to inconsistent output dimensions. Therefore, it is necessary to map it to a standardized stratigraphic model with fixed output dimensions.
[0214] This training process can solve the problems of insufficient labeled samples and inconsistent label dimensions, providing a data foundation for supervised training of collaborative constraint models.
[0215] In some embodiments, the collaborative constraint model may include multiple functional units. These multiple functional units can be implemented by a single holistic neural network or by a combination of multiple sub-networks. In related technologies, the collaborative constraint process includes multiple functional stages such as input processing, feature extraction, prediction, correction, and fusion. The embodiments of the present invention implement the collaborative constraint model through modularization or holistic implementation, allowing for flexible selection between algorithmic clarity, training efficiency, and deployment convenience.
[0216] In some embodiments, generating multiple original layered stratigraphic models includes: determining the number of layers in the original layered stratigraphic model within a preset number of layers; determining the shear wave velocity parameters of each layer in the original layered stratigraphic model according to preset velocity constraints; wherein the preset velocity constraints include at least one of the following: a first layer shear wave velocity range constraint, an inter-layer shear wave velocity difference constraint in the intermediate layer, and a half-space layer shear wave velocity range constraint.
[0217] Specifically, the number of layers in the original layered stratigraphic model can be randomly determined within a preset range. Too few layers make it difficult for the model to describe the main velocity changes; too many layers increase the number of parameters, potentially exacerbating inversion non-uniqueness. Therefore, controlling model complexity by presetting the number of layers allows for a balance between geological expressiveness and training stability.
[0218] Furthermore, the shear wave velocity parameters of each layer can be determined based on preset velocity constraints. Velocity constraints can avoid unreasonable velocity abrupt changes caused by completely random generation, while allowing the existence of local low-velocity or high-velocity interlayers, making the training model both diverse and geologically plausible.
[0219] This step can improve the rationality of the training sample space and reduce the interference of unreasonable layer models on network training.
[0220] In some embodiments, the number of layers in the original layered stratigraphic model can be randomly determined between 5 and 10 layers. In related technologies, too few layers make it difficult to describe the main velocity changes of shallow subsurface structures, while too many layers increase parameter space complexity and exacerbate inversion non-uniqueness. By limiting the number of layers to a reasonable range, the embodiments of the present invention can achieve a balance between the expressive power of stratigraphic structure and the stability of model training.
[0221] In some embodiments, the preset velocity constraints include at least one of the following: a first-layer shear wave velocity range constraint, a shear wave velocity difference constraint between adjacent layers in the intermediate layer, and a shear wave velocity range constraint in the half-space layer. For example, the first-layer shear wave velocity can be set in the range of 100 m / s to 300 m / s; the shear wave velocity difference between adjacent layers in the intermediate layer (excluding the first layer and the half-space layer) can be set in the range of -100 m / s to 200 m / s; and the shear wave velocity in the half-space layer can be set in the range of the maximum shear wave velocity of the overlying strata plus 200 m / s to 1600 m / s. In related technologies, completely random generation of shear wave velocities can easily produce velocity models that do not conform to geological laws. By constraining the shallow velocity, the velocity difference between intermediate layers, and the velocity range in the half-space layer, the embodiments of the present invention can reduce unreasonable velocity abrupt changes, while allowing the existence of low-velocity or high-velocity interlayers, thereby improving the geological rationality of the training samples.
[0222] In some embodiments, mapping each original layered stratigraphic model to a standardized stratigraphic model with a fixed output dimension includes: determining the total thickness of multiple finite layers in the standardized stratigraphic model based on the depth range of the original layered stratigraphic model; determining the layer thickness of multiple finite layers according to a preset layer thickness allocation method; and mapping the shear wave velocity parameters in the original layered stratigraphic model to the standardized stratigraphic model according to the sequence relationship from shallow to deep in the original layered stratigraphic model; wherein the standardized stratigraphic model includes multiple finite layers and a half-space layer.
[0223] Specifically, the half-space top boundary depth in the original layered stratigraphic model can be extracted and used as the total thickness of multiple finite layers in the standardized stratigraphic model. Then, the thickness of multiple finite layers is determined according to a preset layer thickness allocation method. Subsequently, the original shear wave velocity parameters are mapped to the finite layers in the standardized stratigraphic model according to the shallow-to-deep sequence relationship of the original layered stratigraphic model.
[0224] In related technologies, the number of parameters in a variable-layer original layered stratigraphic model is not fixed, making it difficult to use as training labels for a neural network with a fixed output dimension. Standardization mapping can convert original models with different layers into a unified output dimension while preserving the stratigraphic sequence as much as possible.
[0225] This step resolves the conflict between variable-layer models and fixed-output-dimension networks, ensuring a unified training label structure.
[0226] In some embodiments, a standardized stratigraphic model may include nine finite layers and one half-space layer, thereby forming a ten-layer standardized stratigraphic model. In related technologies, fixed-layer models facilitate the formation of a unified output dimension, while nine finite layers can describe layered structures within a finite depth range, and the half-space layer can represent the underlying medium. Embodiments of this invention can convert the variable-layer original model into fixed-dimensional training labels, facilitating supervised training of neural networks.
[0227] In some embodiments, the thickness of multiple finite layers can be allocated according to a geometric progression, such that the layer thickness increases with depth. The common ratio of the geometric progression can be set to 1.05 to 1.50, for example, 1.05, 1.15, 1.2, 1.3, 1.5, etc. In related technologies, shallow structures are generally more sensitive to high-frequency responses and require higher resolution; deep structures correspond to low-frequency responses and can use larger layer thicknesses. The embodiments of this invention, through geometric progression layer thickness allocation, can balance shallow resolution and deep coverage, improving the expressive rationality of standardized stratigraphic models.
[0228] In some embodiments, shear wave velocity parameters of the same layer in the original layered stratigraphic model can be mapped to one or more adjacent finite layers in the standardized stratigraphic model, without cross-layer combination of non-adjacent layers in the original layered stratigraphic model. In related technologies, cross-layer combination may disrupt the original shallow-to-deep sequence relationship of the strata, leading to the generation of an unreasonable standardized model. This embodiment of the invention, by maintaining adjacent mapping and sequence consistency, can preserve the original stratigraphic structure order during fixed output dimension conversion.
[0229] In some embodiments, the preset layer thickness allocation method can also be equal thickness allocation, segmented incremental allocation, adaptive allocation based on target resolution, or depth-weighted allocation. In related technologies, the requirements for standardized layer thickness distribution vary depending on the detection depth, target resolution, and regional geological conditions. Embodiments of the present invention can flexibly set the standardized stratigraphic model structure according to the application scenario, improving the applicability of the standardized mapping method.
[0230] In some embodiments, forward modeling is performed on each standardized stratigraphic model to obtain the training dispersion curve and training HVSR curve corresponding to each standardized stratigraphic model, including:
[0231] The P-wave velocity and density parameters required for forward modeling are determined based on the shear wave velocity parameters of the standardized stratigraphic model.
[0232] Based on the layer thickness parameters, shear wave velocity parameters, p-wave velocity parameters, and density parameters of the standardized stratigraphic model, the training dispersion curve and training HVSR curve corresponding to the standardized stratigraphic model are calculated.
[0233] Specifically, the P-wave velocity and density parameters can be calculated empirically based on the shear wave velocity parameter. Then, the layer thickness, shear wave velocity, P-wave velocity, and density parameters are combined to form the complete layered medium parameters. Based on these complete layered medium parameters, the training dispersion curve and training HVSR curve corresponding to this standardized stratigraphic model are calculated.
[0234] In related technologies, forward modeling of dispersion curves and HVSR curves typically requires not only layer thickness and shear wave velocity parameters, but also physical properties such as p-wave velocity and density parameters. If the forward modeling parameters are incomplete, it is difficult to generate physically consistent training curves.
[0235] This step allows the standardized stratigraphic model to generate its corresponding training dispersion curve and training HVSR curve, providing input samples for supervised learning.
[0236] In some embodiments, the Thomson-Haskell transfer matrix method can be used to calculate the training dispersion curve corresponding to the standardized stratigraphic model. In related technologies, the Thomson-Haskell transfer matrix method is suitable for calculating the surface wave propagation characteristics in layered media, and can obtain the surface wave dispersion response based on parameters such as layer thickness, shear wave velocity, p-wave velocity, and density. Embodiments of this invention can generate training dispersion curves that match the physical properties of the layered structure for each standardized stratigraphic model.
[0237] In some embodiments, the diffusion field assumption can be used to calculate the training HVSR curve corresponding to the standardized stratigraphic model. In related technologies, the diffusion field assumption can be used to approximate the horizontal and vertical spectral ratio responses based on layered medium parameters. Embodiments of this invention can generate training HVSR curves that match the training dispersion curves for the same standardized stratigraphic model, thereby constructing paired joint training inputs.
[0238] In some embodiments, the training dispersion curve can also be calculated using the reflection-transmission matrix method, the fast matrix method, or other methods suitable for surface wave forward modeling in layered media; the training HVSR curve can also be calculated using the theoretical ellipticity method, the micro-motion field simulation method, or other HVSR forward modeling approximation methods. Different forward modeling algorithms in related technologies differ in terms of computational efficiency, stability, and applicable assumptions. Embodiments of this invention can select a suitable forward modeling method based on computational resources and application scenarios, thereby improving the flexibility of training sample generation.
[0239] In some embodiments, after obtaining the training dispersion curve and the training HVSR curve, the method further includes: setting input frequency ranges for the training dispersion curve and the training HVSR curve respectively, and extracting curve data within the corresponding frequency band; and superimposing noise on the extracted curve data to obtain input data for training.
[0240] Specifically, input frequency ranges can be set for the training dispersion curve and the training HVSR curve, respectively, based on the complete forward modeling curve. These input frequency ranges can be set randomly or according to preset candidate frequency bands, curve quality, or sensitive frequency bands. After extracting the curve data within the corresponding frequency band, noise can be superimposed to simulate environmental interference, instrument errors, or curve extraction errors in actual observation data.
[0241] In related technologies, the usable frequency range of the two types of curves in actual observation data is usually not fixed, and there is varying degrees of noise. If only complete and noise-free theoretical forward modeling curves are used to train the model, the trained model will not be able to adapt to real data.
[0242] By using frequency band truncation and noise superposition, the consistency between training data and actual observation data can be improved, thereby enhancing the generalization ability of the collaborative constraint model.
[0243] In some embodiments, the input frequency range can be set according to the sensitivity differences between the two types of data. For example, the input frequency range of the HVSR curve can be randomly determined within the low-frequency candidate range corresponding to the HVSR curve, and the input frequency range of the dispersion curve can be randomly determined within the high-frequency candidate range corresponding to the dispersion curve. In related technologies, the HVSR curve typically provides a more obvious structural response in the main peak and low-frequency band, while the dispersion curve has a better ability to distinguish shallow velocity changes in the higher frequency band. By distinguishing the candidate frequency bands of the two types of curves, the embodiments of the present invention can make the training input more consistent with the sensitivity differences between the two types of curves, thereby improving the ability of the collaborative constraint model to utilize effective information.
[0244] In some embodiments, the noise can be Gaussian noise, uniform noise, frequency-dependent noise, random missing perturbation, amplitude perturbation, or measurement error perturbation. In related technologies, noise in actual observed curves comes from various sources, including random environmental interference, instrument response, frequency-dependent errors, or curve extraction errors. By setting different types of noise perturbations, embodiments of the present invention can enhance the robustness of the model to different noise scenarios.
[0245] In some embodiments, training the collaborative constraint model based on training samples includes: inputting training input data from the training samples into the collaborative constraint model to obtain predicted underground medium structure parameters; calculating a loss value based on the difference between the predicted underground medium structure parameters and the training labels; and updating the model parameters of the collaborative constraint model based on the loss value.
[0246] Specifically, the training input data includes training dispersion curves and training HVSR curves, and may also include frequency band coding features generated from their effective frequency bands. Training labels include subsurface medium structure parameters corresponding to the standardized stratigraphic model, such as layer thickness parameters and shear wave velocity parameters. After the training input data is input into the collaboratively constrained model, the model sequentially performs input processing, feature extraction, baseline prediction, residual correction, gated weighting, and fusion output to obtain the predicted subsurface medium structure parameters. Then, the loss value is calculated based on the error between the predicted subsurface medium structure parameters and the training labels, and the model parameters are updated through backpropagation.
[0247] In related technologies, training the feature extraction network, baseline prediction network, residual correction network, and gating network separately may prevent these components from co-optimizing around the final prediction target of the underground medium structure parameters. Updating the model parameters uniformly using the final prediction error allows each processing stage to optimize together around the same inversion objective.
[0248] This training method can improve the overall training effect and prediction accuracy of the collaborative constraint model.
[0249] In some embodiments, the loss value can be obtained based on the weighted error of multiple parameter items in the underground medium structure parameters, with different parameter items corresponding to different weights. In related technologies, the layer thickness parameter and the shear wave velocity parameter have different physical dimensions, numerical ranges, and prediction difficulties. By setting weights for different parameter items, the embodiments of the present invention can balance the contribution of each parameter item to the training objective and avoid the dominance of a certain type of parameter error in model training.
[0250] In some embodiments, training samples can be divided into training and test sets using random sampling. For example, 70% of the training samples can be allocated to the training set, and 30% to the test set. In related technologies, without a test set, it is difficult to evaluate the model's inversion performance on unseen samples. This embodiment of the invention, by dividing the samples into training and test sets, allows for separate use for model training and performance verification, improving the reliability of model evaluation.
[0251] In some embodiments, the loss value can be mean squared error, mean absolute error, Huber loss, weighted loss, or a combination of the above. Model parameter updates can employ Adam, SGD, RMSProp, or other gradient optimization algorithms. Different objective functions and optimization algorithms in related technologies vary in terms of outlier sensitivity, convergence speed, and training stability. Embodiments of this invention can select appropriate training strategies based on training data characteristics and model convergence to improve model training effectiveness.
[0252] In some embodiments, during the construction and training of the collaborative constraint model, a weighted mean square error can be used as the objective function for model training. The objective function measures the difference between the predicted subsurface structure parameters output by the collaborative constraint model and the corresponding real subsurface structure parameters in the training labels, and updates the model parameters of the collaborative constraint model based on this difference.
[0253] Specifically, the objective function can be expressed as:
[0254] (Formula 1)
[0255] Where L represents the loss value; N represents the total number of output parameter items in the underground medium structure parameters; For the first The actual values of each output parameter For the first Predicted values for each output parameter; This represents the weight coefficient corresponding to the i-th output parameter item.
[0256] The underground medium structural parameters include at least layer thickness parameters and shear wave velocity parameters. Correspondingly, the actual values... The predicted value can be the actual value of the layer thickness parameter or the actual value of the shear wave velocity parameter corresponding to the training label. The objective function can be used to output the corresponding predicted values of layer thickness parameters or shear wave velocity parameters from the collaborative constraint model. Therefore, the objective function can simultaneously constrain the prediction errors of both layer thickness and shear wave velocity parameters, allowing the collaborative constraint model to be trained around the overall prediction accuracy of the subsurface medium structure parameters.
[0257] In some optional embodiments, the layer interface depth parameter can be obtained by accumulating the layer thickness parameters along the depth direction, or it can be used as an equivalent representation of the layer thickness parameter in training. For example, when the standardized stratigraphic model includes multiple finite layers and one half-space layer, the corresponding layer interface depth parameter can be calculated based on the layer thickness parameters of the multiple finite layers; alternatively, the layer interface depth parameter can be used as part of the output parameter item in the training label. In this case, the output parameter item may include the layer interface depth parameter and the shear wave velocity parameter. For example, in an embodiment where the standardized stratigraphic model includes 9 finite layers and 1 half-space layer, the output parameter item may include the depth parameters of the interfaces of layers 1 to 9 and the shear wave velocity parameters of layers 1 to 10, where N can be 19. The above values are only one specific implementation method and do not limit the number of layers or the number of output parameter items in the standardized stratigraphic model.
[0258] Set the weighting coefficient W iThe reason lies in the fact that the physical dimensions, numerical ranges, and inversion difficulties of different underground media structural parameters vary. For example, layer thickness, inter-layer interface depth, and shear wave velocity parameters differ in numerical scale. If unweighted mean square error is directly used, parameters with larger numerical ranges may dominate in the objective function, thereby weakening the constraint effect of other parameters on model training. Therefore, by setting weight coefficients for different output parameter terms, the contributions of various parameter terms to the training objective can be balanced.
[0259] In one embodiment of the present invention, the construction process of the collaborative constraint model may include the following steps:
[0260] Randomly generated original layered stratigraphic models: A range of values for the shear wave velocity parameters in the stratigraphic model is defined, and 30,000 horizontally layered stratigraphic models are randomly generated within this range. To include different types of shallow subsurface models, the model is configured with a variable number of layers, ranging from 5 to 10. The range of shear wave velocity parameters is shown in Table 1. By constraining the velocity difference between adjacent layers in the intermediate layer, unreasonable velocity abrupt changes can be effectively reduced, while allowing the existence of local low-velocity and high-velocity interlayers. The half-space shear wave velocity range aims to ensure that the half-space velocity is higher than that of the overlying strata, generating a reasonable geological model.
[0261] Table 1. Range of Shear Wave Velocities
[0262]
[0263] Construct a 10-layer standardized parameter model.
[0264] The original stratigraphic models of layers 5–10 generated in step 1.1 are uniformly mapped to a 10-layer standardized parameter model to meet the requirement of fixed output dimension of the neural network. The 10-layer standardized parameter model includes 9 finite layers and a half-space. The specific process is as follows:
[0265] Define the number of layers in the original model. Let the total number of layers in the original stratigraphic model be... ,in The original stratigraphic model includes, from top to bottom: The finite layer and the first A half-layer space.
[0266] Extract the half-space top boundary depth from the original stratigraphic model. Using the stated half-space top boundary depth... The total thickness of the first nine finite layers in the standardized parameter model.
[0267] Standardized layer thicknesses are allocated geometrically. In the 10-layer standardized parameter model, the thicknesses of the first nine finite layers are allocated geometrically, with the tenth layer set as a half-space layer. The common ratio of this geometric progression is r, where r is a layer thickness growth factor greater than 1, used to control the gradual increase of layer thickness with depth. r can take values from 1.05 to 1.50, with r=1.15 being preferred. The thickness of each standardized finite layer is:
[0268] (Formula 2)
[0269] And satisfy:
[0270] (Formula 3)
[0271] Mapping Shear Wave Velocity Parameters. After constructing the 10-layer thickness parameters, the shear wave velocity parameters in the original stratigraphic model need to be mapped to the normalized model to obtain the corresponding 10-layer shear wave velocity parameters. During the mapping process, parameters are assigned according to the original stratigraphic sequence from shallowest to deepest. The shear wave velocity of a certain layer in the original model can be randomly mapped to one or more adjacent normalized layers. However, during the mapping process, non-adjacent original layers are not combined across layers to maintain the original stratigraphic sequence relationship and avoid unreasonable velocity model rearrangement after normalization.
[0272] Calculate the P-wave velocity and density parameters.
[0273] Based on the generated shear wave velocity, the longitudinal wave velocity of each layer is further calculated. and density Among them, the longitudinal wave velocity Calculate using the following formula:
[0274] (Formula 4)
[0275] In the formula, For transverse wave velocity, and The unit is km / s. Density Calculate using the following formula:
[0276] (Formula 5)
[0277] In the formula, The unit is g / .
[0278] Construct a standardized stratigraphic model dataset.
[0279] The standardized layer thickness parameters, shear wave velocity parameters, p-wave velocity parameters, and density parameters are combined to construct a standardized stratigraphic model dataset.
[0280] For each generated stratigraphic model, this invention calculates its corresponding fundamental dispersion curve and HVSR curve using the Thomson-Haskell method and the diffusion field assumption. Then, based on the complete forward modeling curve, different input frequency ranges are randomly set for the dispersion curve and HVSR curve, and curve data within the corresponding frequency bands are extracted. Figure 2 This processing method simulates situations where the available frequency range in actual acquired data is not fixed. Based on the sensitivity differences between the two types of data, the HVSR curve mainly contains the lower frequency band, while the dispersion curve mainly contains the higher frequency band. 5Hz is used as the dividing point between the low-frequency and high-frequency ranges, where the low-frequency range is 0.4 to 5Hz and the high-frequency range is above 5Hz. Finally, Gaussian noise is superimposed on the generated curves to obtain the input dataset for neural network training. Based on this, the dataset is divided into a training set and a test set using random sampling, with the training set accounting for 70% and the test set accounting for 30%. Figure 3 Some models and their corresponding dispersion curves and HVSR curves are shown, where (a) is the composite HVSR curve; (b) is the composite dispersion curve; and (c) is the shear wave velocity model.
[0281] Before inputting the dispersion curve and HVSR curve into the neural network, this invention first reads the original available frequency bands of the two types of curves in each sample, wherein the original available frequency bands of the HVSR curve are denoted as […]. , The original usable frequency band of the surface wave dispersion curve is denoted as []. , Subsequently, the main peak frequency is identified within the original usable frequency band of the HVSR curve, and the upper limit of the effective frequency band of the HVSR curve is determined based on the amplitude attenuation characteristics to the right of the main peak. Simultaneously, the lower limit of the effective frequency band of the dispersion curve is determined based on its relative position to the original usable frequency band of the surface wave dispersion curve, ensuring that the effective frequency band of the dispersion curve lies to the right of the main peak frequency of the HVSR. Further, by setting frequency band overlap constraints between the two types of curves, the effective input frequency band of the HVSR and the effective input frequency band of the dispersion curve corresponding to each sample are obtained. This process does not use a fixed frequency range but dynamically adjusts the upper and lower frequency limits for different samples to retain effective model information and reduce the influence of invalid or noisy frequency bands. In related technologies, the effective frequency band of actual data is not fixed, and the HVSR curve is more susceptible to noise in the high-frequency part. In contrast, the vicinity of the main peak usually contains more effective information related to deep structures. Specifically, this invention first identifies the main peak frequency of the HVSR curve, then determines the cutoff frequency corresponding to the amplitude attenuation to 30% of the peak value, and calculates the attenuation ratio. Then, update the upper limit frequency of the HVSR curve and the lower limit frequency of the dispersion curve. The updated upper limit of the effective frequency band of the HVSR curve is:
[0282] (Formula 6)
[0283] The lower bound of the effective bandwidth after the dispersion curve is updated is:
[0284] (Formula 7)
[0285] In the formula, This is the original maximum frequency of the HVSR. This represents the minimum frequency of the original dispersion curve. Simultaneously, to avoid gaps in the frequency band, a minimum overlap constraint is set:
[0286] (Formula 8)
[0287] This constraint ensures a certain overlap between the upper frequency limit of the HVSR curve and the lower frequency limit of the dispersion curve. After the above processing, this step, while preserving the effective information of the main peak, suppresses high-frequency noise interference and ensures the continuity of the frequency coverage of the two types of curves, thereby improving the stability of subsequent collaborative constraints.
[0288] The selected horizontal-vertical spectral ratio curves and surface wave dispersion curves were resampled into 100 sampling points each. Each sampling point consisted of a normalized frequency value and a corresponding curve value, thus forming a 100×2 sequence input for each type of curve. Subsequently, this sequence was flattened into a 200-dimensional input vector, and the upper and lower frequency band limits of the HVSR curves and surface wave dispersion curves were combined to form a 4-dimensional frequency band encoding vector, which was then input into the neural network model.
[0289] This invention inputs the preprocessed HVSR sequence and dispersion sequence into two independent one-dimensional convolutional neural network branches. Each branch employs a multi-scale convolutional structure, extracting multi-scale curve features through three parallel convolutional paths. The three paths use convolutional kernels of different sizes to extract curve response features within different frequency neighborhoods. Specifically, the first-scale convolutional branch extracts local amplitude fluctuations and detailed changes near peaks, the second-scale convolutional branch extracts curve slope changes, and the third-scale convolutional branch extracts the overall trend across a wider frequency band.
[0290] After completing the multi-scale feature extraction, the features output from each branch are concatenated, and a 1×1 convolution is used to linearly combine the concatenated channel features to integrate curve information at different scales and obtain more comprehensive curve features.
[0291] Subsequently, a channel recalibration module is introduced to adaptively adjust the importance of different channel features, thereby enhancing effective features and suppressing secondary features. Specifically, the channel recalibration module learns the corresponding channel weights based on the response intensity of each channel feature across the entire frequency sequence and its contribution to the prediction error of the model parameters. Channels with strong correlation to the prediction of layer interface depth and shear wave velocity are assigned larger weights, while channels with strong noise response, high repeatability, or small contribution to prediction are assigned smaller weights, thus enhancing effective features and suppressing secondary features.
[0292] Furthermore, this invention incorporates residual connections and group normalization operations into the feature extraction module. Residual connections are used to add the module's input features to the output features extracted by convolution, preserving the original frequency response information and improving gradient propagation. When the number of channels in the input and output features is inconsistent, a 1×1 convolution is first used to map the input features to the same channel dimension as the output features before addition. When the frequency sequence lengths are inconsistent, padding, pooling, or interpolation are used to ensure consistency in the frequency dimension. This avoids information transmission interruptions caused by feature dimension mismatch.
[0293] Group normalization is used to normalize the feature response curves after convolution. The feature response curves are the frequency direction responses automatically learned by the convolution kernel during training. Different channels can characterize the main peak shape, post-peak attenuation, spectral shape changes from low to high frequencies of the HVSR curve, as well as the phase velocity level, local slope changes, and overall dispersion trend of the surface wave dispersion curve. In this embodiment, the multi-scale convolution module of each curve branch outputs 128 channel features, with the three scale convolution branches outputting 42, 42, and 44 channel features respectively. These are then concatenated and fused into 128 channel features via a 1×1 convolution. The maximum number of groups for group normalization is set to 8, and the actual number of groups is determined based on the number of channels in the current layer. When the number of channels is not divisible by 8, the maximum number of groups less than 8 and divisible by that number is selected. Therefore, 42 channels are divided into 7 groups, 44 channels into 4 groups, and 128 channels into 8 groups.
[0294] Through the above-mentioned adaptive frequency band selection, unified resampling coding, multi-scale feature extraction, channel recalibration, residual connection and group normalization processing, the present invention can extract stable curve features when different samples have different effective frequency ranges, reduce the dependence of the neural network on fixed input frequency range and fixed sampling interval, and improve the adaptability to frequency band changes of actual observation data.
[0295] The principle of 1×1 convolution is as follows:
[0296] Convolution is primarily used to recombine concatenated channel features. After the features output from each scale convolution branch are concatenated along the channel dimension, a single frequency sampling point will contain features from multiple channels. Assume the... There are a total of 1 frequency points Each channel feature, with its feature value represented as:
[0297] (Formula 9)
[0298] In the formula, Indicating the splicing feature, the first At the nth frequency sampling point The characteristic values of each channel, , This represents the total number of channels. The convolutional kernel assigns a set of learnable weights to these channels and performs a weighted sum at that frequency point:
[0299] (Formula 10)
[0300] In the formula, For learnable weights, For bias terms, This refers to the output features obtained after recombination.
[0301] The channel relabeling steps are as follows:
[0302] Global average pooling needs to be performed on the curve features first:
[0303] (Formula 11)
[0304] In the formula, The curve features are obtained after multi-scale convolutional branch splicing and 1×1 convolutional channel integration. For channel indexing, The length of the feature sequence. For sequence position index. All channels' · Together they form the channel description vector Then calculate the weights of each channel:
[0305] (Formula 12)
[0306] In the formula, and This represents the learnable parameter matrix of the two fully connected transformations during the channel weight calculation process. It is the ReLU activation function. The activation function is Sigmoid. ReLU is used to introduce a nonlinear transformation, and Sigmoid is used to map the output to... The range is determined to generate recalibration weights for each channel. Finally, the channels are recalibrated.
[0307] (Formula 13)
[0308] In the formula, ⊙ represents channel-by-channel weighting.
[0309] The residual connection formula is:
[0310] (Formula 14)
[0311] In the formula, These are the input features for the residual module. The output characteristics of the residual module, This indicates a projection transformation.
[0312] The group normalization formula is:
[0313] (Formula 15)
[0314] (Formula 16)
[0315] (Formula 17)
[0316] In the formula, Indicates the first number in a certain group Each input feature value For the current channel group, This represents the total number of characteristic elements participating in the statistics within this group. For the first The mean of group features, For the first Variance of group characteristics Represents the normalized eigenvalues. This represents the output feature value after affine transformation. and These are the scaling and translation parameters in group normalization, respectively, both of which are learnable parameters. The smallest positive constant introduced to ensure numerical stability.
[0317] The extracted dispersion information, HVSR information, and frequency band coding features are input into the collaborative constraint learning network. The dispersion information is... HVSR information is The frequency band coding features are These dimensions are the feature embedding dimensions in the network. The dispersion curve feature is set to 64 dimensions to represent its relatively stable frequency and velocity variation characteristics. The HVSR information is set to 128 dimensions to retain more complex curve information such as the main peak frequency, main peak amplitude, and slope variation. The frequency band coding feature is obtained by encoding the upper and lower limits of the input frequency band after adaptive filtering of the HVSR curve and dispersion curve, and is set to 32 dimensions.
[0318] dispersive information With frequency band coding features The data is then spliced together and input into a multilayer perceptron network to obtain the baseline prediction results for the underground model parameters:
[0319] (Formula 18)
[0320] In the formula, To predict the baseline multilayer perceptron network, This is the initial inversion result obtained solely based on dispersion and frequency band information. The initial result includes 9 layer interface depth parameters and 10 shear wave velocity parameters.
[0321] Based on the initial baseline prediction results, HVSR information Dispersion information and frequency band coding features The data is then concatenated and input into a multilayer perceptron network to obtain supplementary correction values.
[0322] (Formula 19)
[0323] In the formula, To correct the residuals in a multilayer perceptron network, This represents the supplementary correction amount for the initial baseline prediction result based on HVSR information. For different output parameters, the residual correction amount can be positive, negative, or close to zero, representing an increase, decrease, or essentially no correction based on the initial prediction result, respectively. For example, for the above sample, the HVSR curve has a significant main peak around 4 Hz, followed by a rapid decay. Based on the characteristics of the HVSR curve, dispersion curve, and frequency band coding, the residual correction network determines that the HVSR main peak information indicates the third layer interface may be deeper than the baseline prediction result; therefore, it outputs a residual correction amount corresponding to the depth of the third layer interface as +3 m. Simultaneously, the network determines that the fifth layer shear wave velocity may be slightly lower than the baseline prediction result; therefore, it outputs a residual correction amount corresponding to the fifth layer shear wave velocity as -50 m / s. At this point, the residual correction amount only represents a suggested correction magnitude and is not directly used as the final correction result.
[0324] HVSR information Dispersion information and frequency band coding features Inputting the multilayer perceptron network, we obtain the gating weights corresponding to each output parameter:
[0325] (Formula 20)
[0326] In the formula, For gated multilayer sensor networks, This represents the Sigmoid function. This represents the gating weight vector corresponding to each output parameter. The closer the gating weight is to 1, the more residual correction is used for that output parameter; the closer the gating weight is to 0, the less residual correction is used for that output parameter. For example, for the above sample, the HVSR main peak is clear, and this main peak has a strong constraint effect on the depth of the third layer interface. Therefore, the gating weight corresponding to the depth of the third layer interface output by the gating network is 0.8, indicating that this depth parameter uses 80% residual correction.
[0327] Step 6.5: Convert the baseline prediction results Correction amount of residuals after gate weighting By combining the results, the final predicted parameters of the underground model are obtained. :
[0328] (Formula 21)
[0329] In the formula, This is an element-wise multiplication operation. The final prediction result includes 9 layer interface depth parameters. and 10 shear wave velocity parameters .
[0330] Construct the objective function for model training, minimizing the predicted result based on the final underground model parameters. Compared with actual underground model parameters The difference between them is used to calculate the loss value, and the multilayer perceptron network is jointly optimized and trained.
[0331] A multilayer perceptron network is a type of feedforward neural network used to implement a nonlinear mapping between input features and output parameters (such as...). Figure 4 A multilayer perceptron network includes at least an input layer, one or more hidden layers, and an output layer. The input layer receives the input feature vector, the hidden layers perform layer-by-layer nonlinear transformations on the input features, and the output layer outputs the prediction result.
[0332] For the The calculation process for a layer can be represented as follows:
[0333] (Formula 22)
[0334] In the formula, Indicates the first The output features of the layer Indicates the first The output features of the layer Indicates the first The weight matrix of the layer, Indicates the first Layer bias terms, This represents the activation function, used to introduce non-linear mapping capabilities, including functions such as ReLU, GELU, and Sigmoid.
[0335] The calculation process of the output layer can be represented as follows:
[0336] (Formula 23)
[0337] In the formula, This indicates the network output result. This represents the output feature of the last hidden layer. This represents the weight matrix of the output layer. This represents the bias term of the output layer.
[0338] By combining multilayer linear transformations and nonlinear activations, multilayer perceptron networks can learn complex mapping relationships from input features, thereby enabling the prediction of underground model parameters. In this invention, the multilayer perceptron network is used to map dispersion information, HVSR information, and frequency band coding features, outputting layer thickness parameters and shear wave velocity parameters.
[0339] The inventors' experiments revealed that this invention compares and verifies the proposed machine learning inversion method with the Monte Carlo inversion method. To quantitatively evaluate the inversion results of this invention, the root mean square percentage error (RMSPE) is used to measure the fitting error between the true and predicted curves; a smaller RMSPE indicates a better curve fitting effect. For the subsurface velocity model, this invention uses area accuracy (AA) to evaluate the consistency between the predicted and true models. In the formula, This indicates the overall deviation between the predicted profile and the actual profile. This represents the area of the actual cross-section. The larger the AA value, the closer the predicted model is to the actual model, and the more accurate the model parameter prediction results are. The combination of these two values can comprehensively reflect the inversion performance of the method of this invention.
[0340] Figure 5Comparisons are presented between the machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves and the Monte Carlo inversion results for two test samples. Specifically, (a) compares the HVSR curve of sample 1 with the forward HVSR curves of the predicted model and the optimal Monte Carlo model; (b) compares the Rayleigh wave dispersion curve of sample 1 with the forward dispersion curves of the predicted model and the optimal Monte Carlo model; (c) compares the Vs model, the predicted Vs model, and the optimal Monte Carlo Vs model of sample 1; (d) compares the HVSR curve of sample 2 with the forward HVSR curves of the predicted model and the optimal Monte Carlo model; (e) compares the Rayleigh wave dispersion curve of sample 2 with the forward dispersion curves of the predicted model and the optimal Monte Carlo model; and (f) compares the Vs model, the predicted Vs model, and the optimal Monte Carlo Vs model of sample 2.
[0341] As shown in the figure, the forward modeling results of the HVSR curve and dispersion curve obtained by the method of this invention have a higher degree of agreement with the observation data, and the overall fitting effect is better than that of the Monte Carlo inversion results. The results indicate that this invention can effectively learn the nonlinear mapping relationship between observation data and subsurface velocity model parameters, achieving more efficient and accurate collaborative constraints on subsurface models.
[0342] in addition, Figure 1 The flowchart illustrating the machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves shown in the diagram is merely a representation of the logical order provided by one of the many embodiments of this invention. This order is intended to clearly illustrate the execution flow of the method in a specific application scenario. However, in practical applications, the flexibility and scalability of this invention allow those skilled in the art to adjust the order of the above steps based on different practical needs and environmental conditions, such as allowing parallel or concurrent relationships between some steps, which will not be elaborated upon here.
[0343] Based on the same inventive concept, this application also provides a machine learning inversion system based on the joint constraints of HVSR curves and dispersion curves. For example... Figure 6 As shown, the system includes a curve acquisition unit 110, a feature extraction unit 120, a baseline prediction unit 130, a correction amount calculation unit 140, and a fusion output unit 150. The curve acquisition unit 110, feature extraction unit 120, baseline prediction unit 130, correction amount calculation unit 140, and fusion output unit 150 can transmit signals to each other via wired or wireless means.
[0344] Specifically, the curve acquisition unit 110 is used to acquire the dispersion curve and HVSR curve corresponding to the object to be inverted. The feature extraction unit 120 is used to process the dispersion curve and the HVSR curve to obtain dispersion information, HVSR information, and effective frequency band range information. The baseline prediction unit 130 is used to perform initial prediction of the subsurface medium structure parameters corresponding to the object to be inverted based on the dispersion information and the effective frequency band range information to obtain the baseline prediction result. The correction amount calculation unit 140 is used to generate correction information based on the dispersion information, the HVSR information, and the effective frequency band range information, and to perform gated weighting on the correction information to obtain the gated residual correction amount used to correct the baseline prediction result. The fusion output unit 150 is used to determine the target value of the subsurface medium structure parameters corresponding to the object to be inverted based on the baseline prediction result and the gated residual correction amount. The subsurface medium structure parameters include at least layer thickness parameters and shear wave velocity parameters. The feature extraction unit 120, baseline prediction unit 130, correction calculation unit 140, and fusion output unit 150 can be integrated into the collaborative constraint model.
[0345] Modular configuration of curve acquisition, input processing and feature extraction, baseline prediction, gated residual correction and fusion output in related technologies facilitates the implementation of this collaborative constraint process in computing devices, servers, edge devices or cloud platforms.
[0346] This system enables automated collaborative constraint functions corresponding to the method implementation, thereby improving the efficiency of shallow surface and subsurface media structure parameter inversion.
[0347] In some embodiments, the feature extraction unit may further include a frequency band selection subunit, a resampling coding subunit, a curve feature extraction subunit, and a frequency band coding subunit. The frequency band selection subunit is used to determine the effective input frequency band for dispersion and the effective input frequency band for HVSR; the resampling coding subunit is used to generate the dispersion sequence input and the HVSR sequence input; the curve feature extraction subunit is used to generate dispersion information and HVSR information; and the frequency band coding subunit is used to generate frequency band coded features. In related technologies, the input processing and feature extraction process includes multiple processing steps with different functions. The embodiments of the present invention, through subunit division, make the system structure and method steps more clearly correspond, facilitating system implementation, maintenance, and modular deployment.
[0348] In some embodiments, the correction unit may further include a residual correction subunit, a gate weight generation subunit, and a gate weighting subunit. The residual correction subunit is used to generate the residual correction amount; the gate weight generation subunit is used to generate the gate weights; and the gate weighting subunit is used to weight the residual correction amount according to the gate weights to obtain the gated residual correction amount. In related technologies, the gated residual collaborative constraint process includes three functional steps: residual generation, gate generation, and weighted correction. This embodiment of the invention, through subunit division, clarifies the internal functional division of the system, ensuring consistency between the system structure and the core algorithm logic.
[0349] In some embodiments, the above-described units can be implemented as software modules, hardware modules, combined software and hardware modules, or neural network sub-models. The system can be deployed on local computing devices, edge computing devices, servers, or cloud platforms. In related technologies, different application scenarios have different requirements for computing resources, real-time performance, and deployment methods. The embodiments of the present invention can be flexibly deployed according to the needs of on-site inversion, batch data processing, or cloud computing, thereby improving the applicability of the system.
[0350] It should be noted that the machine learning inversion system based on the joint constraints of HVSR curves and dispersion curves provided in this application embodiment has the same implementation principle and technical effects as the aforementioned machine learning inversion method embodiment based on the joint constraints of HVSR curves and dispersion curves. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned machine learning inversion method embodiment based on the joint constraints of HVSR curves and dispersion curves. Furthermore, it is understood that the method steps in the model construction process can also be used in the method execution process, and will not be repeated here.
[0351] It should be noted that the technical effect of this application is not produced by a single isolated step, but is achieved by the cooperation of multiple links, including model training, input processing, feature extraction, gating residual collaborative constraints, and system output.
[0352] First, the training sample construction process addresses the problems of insufficient measured labeled samples, variable number of original model layers, and large differences between theoretical and actual observed curves by combining the following steps: "generation of a variable-layer original layered stratigraphic model—standardized mapping with fixed output dimensions—forward modeling to generate training dispersion curves and training HVSR curves—bandwidth truncation and noise superposition—joint training based on final parameter errors." This combination ensures that the collaboratively constrained model can obtain sufficient training samples while adapting to bandwidth variations and noise interference in real-world data.
[0353] Second, the input processing addresses the issues of different effective frequency bands, different numbers of sampling points, different sampling densities, and loss of frequency band information after resampling by combining "adaptive frequency band selection—resampled encoding—frequency band coding feature generation." Resampled encoding provides a unified network input format, while frequency band coding features preserve the input frequency range information. Together, they enable the model to perceive the frequency coverage differences of different samples under a fixed input dimension.
[0354] Third, the feature extraction process, through a combination of "bi-branch feature extraction—multi-scale convolution—channel recalibration—residual connection—group normalization," solves the problems of single-scale convolution's difficulty in extracting complete curve shapes, different channel contributions, difficulties in information transmission in deep networks, and unstable feature distribution. This combination enables the model to extract local details, slope changes, overall trends, and effective channel responses from dispersion curves and HVSR curves respectively, improving the quality of curve feature representation.
[0355] Fourth, the collaborative constraint process addresses the issues of differing constraint capabilities between dispersion curves and HVSR curves, the potential for introducing invalid information through direct fusion, and the varying sensitivities of HVSR to different parameters. This approach combines "baseline prediction results formed from dispersion information and frequency band coding features—residual corrections formed from dispersion information, HVSR information, and frequency band coding features—residual corrections controlled by gating weights—fusion of baseline prediction results and gating residual corrections." This combination enables dispersion curves to provide stable basic predictions, HVSR curves to provide controlled supplementary corrections, and frequency band coding features to provide auxiliary information for the input frequency range, thereby improving the stability and accuracy of the inversion results.
[0356] Fifth, the system implementation, through modular configuration of curve acquisition, input processing and feature extraction, baseline prediction, gated residual correction, and fusion output, enables the above methods to be implemented in software, hardware, or a combination of both. This system can be applied to scenarios such as shallow surface structure interpretation, site classification, engineering geophysical evaluation, and seismic site response analysis.
[0357] In summary, this application improves the stability, accuracy, and applicability of the collaborative constraint of underground medium structure parameters by combining "training sample construction and standardized label generation, frequency band perturbation and noise simulation, adaptive input processing, bi-branch multi-scale feature extraction, and gated residual collaborative constraint" in situations where the effective frequency band of actual observation data is not fixed, frequency sampling is inconsistent, noise interference exists, and the two types of curves have different sensitivities to different parameters.
[0358] refer to Figure 7This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. The electronic device provided in this embodiment includes a processor 210 and a memory 220. The memory 220 stores machine-readable instructions executable by the processor 210. When the machine-readable instructions are executed by the processor 210, they perform the steps in the machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves described in any of the preceding embodiments. At least one processor 210 and one memory 220 are provided.
[0359] In this embodiment of the invention, the electronic device further includes a communication interface 230 and a communication bus 240, wherein the processor 210, the memory 220, and the communication interface 230 are interconnected via the communication bus 240. The communication bus 240 can be divided into an address bus, a data bus, a control bus, etc., for ease of representation. Figure 7 The communication bus 240 is represented by a single thick line, but this does not mean that there is only one communication bus 240 or only one type of communication bus 240. The processor 210 can also be called a controller, and there is no restriction on the name.
[0360] In this embodiment, memory 220 stores instructions executable by at least one processor 210. By executing the instructions stored in memory 220, at least one processor 210 can perform the steps in the machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves discussed above. Processor 210 can implement... Figure 7 The functions of each module in the device shown.
[0361] The processor 210 is the control center of the device. It can connect to various parts of the control device through various interfaces and lines. By running or executing instructions stored in the memory 220 and calling data stored in the memory 220, the various functions of the device and the data processing are monitored as a whole.
[0362] In one possible design, processor 210 may include one or more processing units. Processor 210 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor. In some embodiments, processor 210 and memory 220 may be implemented on the same chip or on separate chips.
[0363] Processor 210 can be a general-purpose processor, such as a central processing unit (CPU), digital signal processor, application-specific integrated circuit, field-programmable gate array or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the machine learning inversion method based on the collaborative constraints of HVSR curves and dispersion curves disclosed in the embodiments of this application can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.
[0364] Memory 220, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. Memory 220 may include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic memory, magnetic disk, optical disk, etc. Memory 220 can be any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. In the embodiments of this application, memory 220 may also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.
[0365] By designing and programming the processor 210, the code corresponding to the machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves described in the aforementioned embodiments can be embedded into the chip, thereby enabling the chip to execute the code during runtime. Figure 2 The steps of the machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves in the illustrated embodiment are described. How to design and program the processor 210 is a technique well-known to those skilled in the art and will not be elaborated upon here.
[0366] Furthermore, this application also provides a computer-readable storage medium storing computer-executable instructions. When executed by processor 210, these instructions are used to implement the machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves described in any of the preceding embodiments. Therefore, they will not be repeated here. Additionally, the beneficial effects of using the same method will not be repeated here either. For technical details not disclosed in the embodiments of the computer-readable storage medium involved in this invention, please refer to the description of the method embodiments of this invention.
[0367] In some possible implementations, various aspects of the machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves provided in this application can also be implemented in the form of a program product, which includes program code. When the program product is run on a device, the program code is used to cause the control device to perform the steps in the machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves based on various exemplary embodiments of this application described above.
[0368] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0369] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0370] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.
[0371] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment, causing a series of operational steps to be executed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0372] Additionally, any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order based on the functions involved, as should be understood by those skilled in the art to which embodiments of the invention pertain.
[0373] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. A machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves, characterized in that, include: S1, obtain the dispersion curve and HVSR curve corresponding to the object to be inverted; S2, process the dispersion curve and the HVSR curve to obtain dispersion information, HVSR information and effective frequency band range information; S3. Based on the dispersion information and the effective frequency band range information, perform initial prediction of the underground medium structure parameters corresponding to the object to be inverted, and obtain the baseline prediction result. S4, Based on the dispersion information, the HVSR information and the effective frequency band range information, generate correction information, and perform gated weighting on the correction information to obtain the gated residual correction amount used to correct the baseline prediction result; S5. Based on the baseline prediction results and the gated residual correction amount, determine the target values of the underground medium structure parameters corresponding to the object to be inverted; Among them, each parameter item in the underground medium structure parameters includes at least each layer thickness parameter and each shear wave velocity parameter; S4 includes the following steps: Based on a preset residual correction multilayer perceptron network, the residual correction amount is obtained by operating on the dispersion information, the HVSR information and the frequency band coding features. The residual correction amount includes residual components corresponding to each layer thickness parameter and each shear wave velocity parameter. Based on a pre-defined gated multilayer perceptron network, gate weights are obtained by processing the dispersion information, the HVSR information, and the frequency band coding features. The gate weights include weight components corresponding to each layer thickness parameter and each shear wave velocity parameter. The corresponding residual components are weighted based on the weighted components to obtain the gated residual correction components corresponding to each parameter item.
2. The machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves according to claim 1, characterized in that, S2 includes: Adaptive frequency band filtering is performed on the dispersion curve and the HVSR curve to obtain the effective input frequency band of dispersion and the effective input frequency band of HVSR. Based on the effective input frequency band of the dispersion, the dispersion curve is resampled and encoded to obtain the dispersion sequence input; Based on the effective input frequency band of the HVSR, the HVSR curve is resampled and encoded to obtain the HVSR sequence input; Feature extraction is performed on the dispersion sequence input to obtain the dispersion information, and feature extraction is performed on the HVSR sequence input to obtain the HVSR information; Based on the effective input frequency band of the dispersion and the effective input frequency band of the HVSR, the frequency band coding feature is generated, and the frequency band coding feature serves as the effective frequency band range information; The frequency band coding features include the upper and lower frequency limits of the effective input frequency band of the HVSR, as well as the upper and lower frequency limits of the effective input frequency band of the dispersion band.
3. The machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves according to claim 2, characterized in that, The adaptive bandwidth filtering of the dispersion curve and the HVSR curve to obtain the effective input bandwidth of dispersion and the effective input bandwidth of HVSR includes: Obtain the original available frequency band of the HVSR curve and the original available frequency band of the dispersion curve; Determine the main peak frequency of HVSR within the original available frequency band of the HVSR curve; The upper limit frequency of the effective input frequency band of the HVSR is determined based on the amplitude attenuation characteristics to the right of the main peak frequency of the HVSR. The lower limit frequency of the effective input frequency band of the dispersion is determined based on the main peak frequency of the HVSR and the original available frequency band of the dispersion curve. Wherein, the effective input frequency band of the HVSR and the effective input frequency band of the dispersion have an overlapping frequency range, or the frequency interval between the effective input frequency band of the HVSR and the effective input frequency band of the dispersion is less than or equal to a preset interval threshold; And / or, the step of resampling and encoding the dispersion curve according to the effective input frequency band of the dispersion to obtain a dispersion sequence input, and resampling and encoding the HVSR curve according to the effective input frequency band of the HVSR to obtain an HVSR sequence input, includes: The dispersion curve located within the effective input frequency band of the dispersion and the HVSR curve located within the effective input frequency band of the HVSR are respectively resampled to a preset number of sampling points; The normalized frequency value and the corresponding curve value of each sampling point are used as the features of the sampling point; The dispersion sequence input and the HVSR sequence input are generated based on the features of multiple sampling points.
4. The machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves according to claim 2, characterized in that, The step of extracting features from the dispersion sequence input to obtain the dispersion information, and extracting features from the HVSR sequence input to obtain the HVSR information, includes: The sequence input to be processed is sequentially subjected to multi-scale convolution processing, channel recalibration processing, residual connection processing, and group normalization processing to obtain the corresponding curve features; Wherein, when the input sequence to be processed is the dispersion sequence input, the corresponding curve feature is the dispersion information; when the input sequence to be processed is the HVSR sequence input, the corresponding curve feature is the HVSR information.
5. The machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves according to claim 4, characterized in that, The multi-scale convolution processing includes: Multiple convolutional kernels of different sizes are used to perform convolution processing on the sequence input to be processed, resulting in curve response features at multiple scales; By stitching together the curve response features at multiple scales, the stitched channel features are obtained. The spliced channel features are linearly combined using 1×1 convolution to obtain the fused curve intermediate features. The sequence input to be processed is either the dispersion sequence input or the HVSR sequence input.
6. The machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves according to claim 5, characterized in that, The channel recalibration process includes: Global average pooling is performed on the curve features to be recalibrated to obtain the channel description vector; Calculate the channel weight corresponding to each channel based on the channel description vector; The curve features to be recalibrated are weighted channel by channel according to the channel weights to obtain the recalibrated curve features.
7. The machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves according to claim 1, characterized in that, The S3 includes: Based on a preset baseline prediction multilayer perceptron network, the dispersion information and the effective frequency band range information are processed to obtain the baseline prediction result.
8. The machine learning inversion method based on the joint constraints of HVSR curves and dispersion curves according to any one of claims 1 to 7, characterized in that, Before obtaining the dispersion curve and HVSR curve corresponding to the object to be inverted, the following steps are also included: constructing and training a collaborative constraint model; The collaborative constraint model is used to obtain the underground medium structure parameters corresponding to the object to be inverted based on the dispersion curve and the HVSR curve. The construction and training of the collaborative constraint model includes: Multiple original layered stratigraphic models are randomly generated, wherein the number of layers in the original layered stratigraphic models is variable; Each of the original layered stratigraphic models is mapped to a standardized stratigraphic model with a fixed output dimension; Forward modeling is performed on each of the standardized stratigraphic models to obtain the training dispersion curve and training HVSR curve corresponding to each of the standardized stratigraphic models; Training samples are constructed based on the training dispersion curve, the training HVSR curve, and the corresponding underground medium structure parameters; wherein the training dispersion curve and the training HVSR curve are used as training input data, and the corresponding underground medium structure parameters are used as training labels. The collaborative constraint model is trained based on the training samples.
9. A machine learning inversion system based on the joint constraints of HVSR curves and dispersion curves, characterized in that, include: The curve acquisition unit is used to acquire the dispersion curve and HVSR curve corresponding to the object to be inverted; The feature extraction unit is used to process the dispersion curve and the HVSR curve to obtain dispersion information, HVSR information and effective frequency band range information; The baseline prediction unit is used to perform initial prediction of the underground medium structure parameters corresponding to the object to be inverted based on the dispersion information and the effective frequency band range information, and to obtain the baseline prediction result. The correction calculation unit is used to generate correction information based on the dispersion information, the HVSR information, and the effective frequency band range information, and to perform gated weighting on the correction information to obtain a gated residual correction amount used to correct the baseline prediction result. The correction calculation unit includes a residual correction subunit, a gated weight generation subunit, and a gated weighting subunit. The residual correction subunit is used to calculate the residual correction amount based on the dispersion information, the HVSR information, and the frequency band coding features according to a preset residual correction multilayer perceptron network. The residual correction amount includes residual components corresponding to each layer thickness parameter and each shear wave velocity parameter. The gated weight generation subunit is used to calculate the gated weights based on the dispersion information, the HVSR information, and the frequency band coding features according to a preset gated multilayer perceptron network. The gated weights include weight components corresponding to each layer thickness parameter and each shear wave velocity parameter. The gated weighting subunit is used to weight the corresponding residual components according to each weight component to obtain the gated residual correction component corresponding to each parameter item. The fusion output unit is used to determine the target value of the underground medium structure parameter corresponding to the object to be inverted based on the baseline prediction result and the gated residual correction amount. Among them, each parameter item in the underground medium structure parameters includes at least each layer thickness parameter and each shear wave velocity parameter.
Citation Information
Patent Citations
Global optimization method for surface wave dispersion and body wave spectral ratio joint inversion shallow velocity
CN109061731A
Shallow velocity model construction method based on joint inversion
CN119439254A
