A method and system for rayleigh wave dispersion inversion fusing physical information

Abstract

The application belongs to the technical field of depth inversion, and provides a Rayleigh wave dispersion inversion method and system fusing physical information, which adaptively generates a multi-type layered geological model based on dynamic Markov decision according to the conditional probability relationship of the previous layer parameter; a fast vectorization Rayleigh wave forward algorithm is used to calculate the base order and first-order dispersion curve corresponding to the geological model, and construct dispersion sample data; a dispersion curve joint inversion model based on deep learning is constructed, the base order dispersion curve, the first-order dispersion curve and the corresponding mask information are taken as inputs, a loss function with physical information constraint is introduced to train the dispersion curve joint inversion model; the trained dispersion curve joint inversion model is used to process target geological data, and the layered transverse wave velocity and layer thickness parameters of the underground medium are obtained, and Rayleigh wave dispersion curve inversion is realized; the application can enhance the constraint on the real stratum relationship, and realize efficient and accurate stratum velocity and layer thickness inversion.

Description

Technical Field

[0001] This invention belongs to the field of deep inversion technology, specifically relating to a Rayleigh wave dispersion inversion method and system that integrates physical information. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Rayleigh surface waves are waves generated by the interaction of longitudinal and transverse waves propagating inside the Earth during an earthquake. Their propagation speed is slightly lower than that of transverse waves, and they often have the largest amplitude and longest duration in seismic records. Therefore, they are considered a strong interference wave. They have dispersion characteristics and medium sensitivity in layered media and are therefore applied in the fields of quality inspection and exploration.

[0004] Currently, in seismic exploration and geological engineering, the ability to quickly and accurately obtain stratigraphic structural parameters has a direct impact on construction safety, engineering design, and geological risk assessment. Rayleigh wave dispersion curve inversion, as the primary means of obtaining stratigraphic shear wave velocity (Vs) and layer thickness (H), directly affects the reliability of the inversion results due to its accuracy and efficiency. However, with the increasing complexity of exploration tasks, the growing amount of data, and the existence of multimodal dispersion information, traditional inversion methods face challenges such as strong dependence on initial models, slow convergence speed, long computation time, and limited inversion accuracy, making it difficult to meet the needs of rapid and intelligent analysis of complex strata.

[0005] In recent years, with the rapid development of artificial intelligence and machine learning, deep learning has been widely used in the field of geophysical exploration due to its powerful nonlinear mapping capabilities and good generalization performance.

[0006] However, current Rayleigh wave inversion research still faces several challenges: First, the real sample dataset is too small. Existing studies use a large number of simulated samples for training, but the correlation between these simulated data and actual strata is limited, resulting in limited generalization ability of the model in real-world scenarios. Second, only the fundamental dispersion curve is used for end-to-end mapping, ignoring the non-uniqueness of dispersion curve inversion. Third, some studies only focus on the mapping from dispersion curves to velocity structures, without simultaneously considering the influence of stratum thickness, leading to insufficient adaptability of the model under complex geological conditions. Fourth, the loss function only considers the error between data, ignoring the continuity of soil layer sequence information, failing to capture the development and evolution patterns of soil layers, resulting in problems such as poor interpretability and generalization ability of the model. Summary of the Invention

[0007] To address the aforementioned problems, this invention proposes a Rayleigh wave dispersion inversion method and system that integrates physical information. In terms of data generation, this invention employs dynamic Markov decision-making to construct an inversion velocity model, making it more consistent with actual geological characteristics. In terms of inversion, it uses deep learning to capture the temporal dependencies between local features and modes, and designs a custom loss function that integrates soil impedance ratio physical information and data errors to enhance the constraint on the actual stratigraphic relationship, thereby achieving efficient and accurate inversion of stratigraphic velocity and layer thickness.

[0008] According to some embodiments, the present invention adopts the following technical solution:

[0009] A Rayleigh wave dispersion inversion method that integrates physical information includes the following steps: First, set the range of physical property parameters of the initial strata as the initial state for model generation. Based on the conditional probability relationship of the parameters of the previous strata, adaptively generate multiple types of layered geological models based on dynamic Markov decision.

[0010] Based on multi-type layered geological models, the corresponding basic order and first-order dispersion curves are calculated using the fast vectorized Rayleigh wave forward modeling algorithm to construct dispersion sample data.

[0011] Based on the corresponding geological model and its corresponding dispersion sample data, a training sample library containing stratigraphic parameters and dispersion characteristics is constructed.

[0012] A deep learning-based joint inversion model of dispersion curves is constructed. The basic order dispersion curves, first order dispersion curves and corresponding mask information in the training sample library are used as multi-channel inputs of the model. The joint inversion model of dispersion curves is trained using a loss function that incorporates physical information constraints.

[0013] The target geological data is processed using the trained dispersion curve joint inversion model to obtain the layered shear wave velocity and layer thickness parameters of the subsurface medium, thereby realizing the Rayleigh wave dispersion curve inversion.

[0014] As an alternative implementation method, the process of adaptively generating multiple types of layered geological models based on dynamic Markov decision-making according to the conditional probability relationship of the parameters of the previous layer includes: setting the initial state of velocity structure data, including the range of shear wave velocity, total thickness, number of layers and other relevant physical parameters of the first layer, and constructing the state space S;

[0015] Based on historical data or prior models, a transition probability matrix P is defined for each state in the state space S, which describes the possible distribution of the current state to the next state.

[0016] The next layer of parameters for generating velocity structure data is determined based on a dynamic Markov decision strategy.

[0017] Repeat the above action selection and dynamic generation steps until the preset requirements are met, and use the final generated velocity structure data sequence as the output result.

[0018] As a further defined implementation, the initial state of the velocity structure data is set, including the range of shear wave velocities of the first layer, total thickness, number of layers, and other relevant physical parameters. The process of constructing the state space S includes the range of shear wave velocities Vs of the first layer of the formation, total thickness H, number of layers N, and density. And the P-wave velocity VP, where the range of the S-wave velocity Vs of the first stratum, the total thickness H, and the number of layers N are set according to the actual situation, and the density And the longitudinal wave velocity VP is calculated.

[0019] As a further defined implementation, the process of defining a transition probability matrix P for each state in the state space S based on historical data or a priori models includes setting an initial probability distribution according to typical site types, so that the initial strata generated by the model have a real engineering background; wherein the thickness of each layer is... , The generation strategy is as follows:

[0020] Define the normalized order vector:

[0021] ;

[0022] The initial weights for layer thickness are adjusted by adding a small random perturbation:

[0023] ;

[0024] in This indicates a tendency for deeper layers to be thicker, reflecting the effect of sedimentary accumulation. This represents a perturbation used to simulate the non-uniformity of layer thickness;

[0025] Final thickness of each layer:

[0026] ;

[0027] As a further defined implementation, the process of determining the next layer parameters for generating the velocity structure data based on the dynamic Markov decision strategy includes: when the number of layers is greater than or equal to 2, randomly generating coefficients. ,in ,according to Different options are chosen depending on the specific circumstances:

[0028] when Choose to continue the trend in increasing layers:

[0029] ;

[0030] in , This results in greater shallow disturbances and smaller deep disturbances, which better reflects the true temporal characteristics of stratigraphy. Indicates the first Number of floors;

[0031] when A high-speed mezzanine is chosen to achieve the jump:

[0032] ;

[0033] when Choose a weak interlayer to achieve the jump descent:

[0034] ;

[0035] in, Indicates the first Number of floors For the first Shear wave velocity of the layer, For the first The transverse wave velocity of the layer above the layer, , This results in greater shallow disturbances and smaller deep disturbances, which better reflects the true temporal characteristics of stratigraphy. For the first Minimum shear wave velocity of the layer, For the first Maximum shear wave velocity of the layer For the first The thickness of the layer above the layer.

[0036] As a further defined implementation method, the transverse wave velocity of the half-space layer is set according to the actual site conditions.

[0037] As an alternative implementation method, the process of constructing a deep learning-based dispersion curve joint inversion model includes: constructing a dispersion curve joint inversion model, wherein the dispersion curve joint inversion model is a Convolutional Neural Network (CNN)-Long Short-Term Memory (LSTM) model optimized using the Quantum Particle Swarm Optimization (QPSO) algorithm. In the CNN-LSTM model, the CNN is used to extract local features of the input data, and the LSTM model is used to model the dynamic evolution law of the dispersion curve with frequency change, so as to realize the continuous expression of the relationship between modes and layers.

[0038] As a further defined implementation, the convolutional neural network includes convolutional layers, batch normalization layers, and max pooling layers. The convolutional layers slide convolution kernels on the input dual-channel dispersion curve to perform one-dimensional convolution operations on local regions, generating feature maps and achieving weighted aggregation of local features. After the convolution output, the number of features in the dispersion curve increases, and the information dimension is expanded. The batch normalization layer is used to perform normalization processing to stabilize the training process and accelerate convergence. An activation function is introduced to give the network nonlinear representation capabilities, enabling it to express complex mapping relationships.

[0039] Max pooling layers are used to downsample convolutional features, compress sequence length, and retain the features that are most significantly correlated with dispersion curves and velocity structures.

[0040] As an alternative implementation, the Long Short-Term Memory network model includes an input gate. Forgotten Gate Output gate Candidate memory units With memory unit Establish a forget gate between the phase velocity and the VS layer velocity structure:

[0041] ;

[0042] At the current frequency point, through the input gate Controlling the introduction of new information while considering the temporary state of memory units. Update to cell state :

[0043] ;

[0044] ;

[0045] ;

[0046] Calculate the phase velocity and VS layer velocity structure of the network output at the current moment:

[0047] ;

[0048] ;

[0049] in, , , These represent the forget gate, input gate, and output gate, respectively; subscript characters. This indicates that the variable is located in the current training step. In the middle, subscript character This indicates that the variable is located in the previous training step. middle; , , This represents the weight matrix of each gate; , , , Indicates the deviation of each gate; This indicates the output of the layer above the hidden unit; Indicates hidden cell input; express Activation function This represents the hyperbolic tangent function.

[0050] As an alternative implementation, the loss function incorporating physical information constraints is as follows: a custom loss function combining geological physics knowledge and data is constructed, incorporating the soil layer impedance ratio as prior information, and using the degree of abrupt change in the medium properties of adjacent layers as a weighting coefficient to calculate the error in shear wave velocity. For the , The impedance ratio of a soil layer is defined as:

[0051] ;

[0052] in and They represent the first The first sample The density and shear wave velocity of the soil layer, when When the deviation from 1 exceeds the set value, it indicates a significant speed change, and a higher error weight is assigned.

[0053] The loss function is:

[0054] ;

[0055] ;

[0056] ;

[0057] ;

[0058] in, This indicates the shear wave velocity loss in the current batch; This indicates the thickness loss in the current batch; and These are the weighting coefficients for shear wave velocity and depth; Batch size; This is the [number]th batch of the current batch. One sample; The soil layer resistance ratio; For the first The first sample The actual value velocity of the layer; For the first The first sample The speed of the layer's predicted values; The number of layers is the thickness. For the first The first sample The actual thickness of the layer; for The first sample The predicted thickness of the layer.

[0059] A Rayleigh wave dispersion inversion system that integrates physical information includes:

[0060] The geological model construction module is configured to set the initial stratigraphic property parameter range as the initial state for model generation based on prior geological information, and adaptively generate multiple types of layered geological models based on dynamic Markov decision based on the conditional probability relationship of the parameters of the previous stratigraphic layer; the dispersion curve calculation module is configured to calculate the corresponding basic order and first order dispersion curves based on the multiple types of layered geological models using the fast vectorized Rayleigh wave forward modeling algorithm, and construct dispersion sample data.

[0061] The training sample library construction module is configured to construct a training sample library containing stratigraphic parameters and dispersion characteristics based on the corresponding geological model and its corresponding dispersion sample data.

[0062] The dispersion curve joint inversion model training module is configured to construct a dispersion curve joint inversion model based on deep learning. The module uses the basic dispersion curve, the first-order dispersion curve and the corresponding mask information in the training sample library as the multi-channel input of the model, and uses a loss function that introduces physical information constraints to train the dispersion curve joint inversion model.

[0063] The dispersion curve joint inversion module is configured to process the target geological data using the trained dispersion curve joint inversion model to obtain the layered shear wave velocity and layer thickness parameters of the subsurface medium, thereby realizing the Rayleigh wave dispersion curve inversion.

[0064] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0065] The data generation of this invention is no longer random and does not conform to the real physical laws, but rather a controllable and systematic dynamic Markov decision mechanism, which enables stratigraphic evolution to have a "conditional probability logic chain" and can automatically generate multiple types of geological models such as incremental, interlayer, and heterogeneous types, greatly improving the consistency between data distribution and real site and enhancing the generalization ability of the inversion model.

[0066] This invention enables fully automated construction of the generative model, forward modeling, and sample library. By combining the geological model generated by dynamic Markov decision with the fast vector forward modeling algorithm, it forms an automatic closed-loop data construction, easily supporting tens of thousands to millions of sample expansions, reducing manual modeling bias, and lowering the cost of data preparation.

[0067] This invention uses both the fundamental and first-order dispersion curves as input, adjusts the weights using masking information, and simultaneously captures the velocity variation characteristics of shallow and deep layers through joint inversion. It not only fully utilizes information from different modes but also adds inversion constraints, reducing the non-uniqueness problem of dispersion curves, thereby improving the accuracy and reliability of formation parameter prediction.

[0068] This invention implements a fast inversion method based on a QPSO-CNN-LSTM deep learning model. QPSO is introduced for global hyperparameter search, resulting in more stable convergence of network parameter combinations. CNN extracts local spatial features, while LSTM captures the temporal dependence of modalities on frequency variations. These two methods complement each other, improving the interpretability and fitting accuracy of dispersion curve inversion. Inversion is performed on all test samples, resulting in a very short average inversion time per sample, significantly reducing time costs.

[0069] This invention employs a dual-drive loss function of knowledge acquisition and physics to make model predictions more consistent with the actual geological structure. It introduces soil impedance ratio and velocity-thickness coupling relationship as physical priors, so that the model no longer simply fits the data, but adheres to the hard constraints of geophysics, reducing common inversion problems such as non-physical interpretation and spurious smoothing.

[0070] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0071] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0072] Figure 1 The image shows a generated sample dataset in one embodiment, where (a) is a shear wave velocity model, (b) is a fundamental dispersion curve, and (c) is a first-order dispersion curve.

[0073] Figure 2 This is a system interface diagram of the Rayleigh wave inversion dispersion curve in one embodiment;

[0074] Figure 3 This is a flowchart of the inversion method in one embodiment;

[0075] Figure 4This is a diagram of the QPSO optimization steps in one embodiment;

[0076] Figure 5 This is a schematic diagram of a pop-up window for jumping to the optimization convergence curve in one embodiment.

[0077] Figure 6 This is a schematic diagram of a deep learning LSTM unit structure in one embodiment.

[0078] Figure 7 This is a schematic diagram of the training and verification convergence curve in one embodiment. Detailed Implementation

[0079] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0080] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0081] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0082] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0083] Example 1

[0084] As described in the background section, existing sample dataset generation methods suffer from limitations such as high volatility and a tendency to produce results that do not conform to actual geological conditions. The non-uniqueness of dispersion curve inversion is also a key concern in current inversion methods. Current inversion algorithms rely on initial models, resulting in slow inversion speeds. In addition, valuable prior geological information about the site's history cannot be fully utilized.

[0085] This embodiment provides a Rayleigh wave dispersion inversion method that integrates physical information, including the following steps:

[0086] First, the range of physical property parameters of the initial strata is set as the initial state for model generation. Based on the conditional probability relationship of the parameters of the previous strata, multi-type layered geological models are adaptively generated based on dynamic Markov decision-making.

[0087] Based on multi-type layered geological models, the corresponding basic order and first-order dispersion curves are calculated using the fast vectorized Rayleigh wave forward modeling algorithm to construct dispersion sample data.

[0088] Based on the corresponding geological model and its corresponding dispersion sample data, a training sample library containing stratigraphic parameters and dispersion characteristics is constructed.

[0089] A deep learning-based joint inversion model of dispersion curves is constructed. The basic order dispersion curves, first order dispersion curves and corresponding mask information in the training sample library are used as multi-channel inputs of the model. The joint inversion model of dispersion curves is trained using a loss function that incorporates physical information constraints.

[0090] The target geological data is processed using the trained dispersion curve joint inversion model to obtain the layered shear wave velocity and layer thickness parameters of the subsurface medium, thereby realizing the Rayleigh wave dispersion curve inversion.

[0091] The process of adaptively generating multi-type layered geological models based on dynamic Markov decision-making, according to the conditional probability relationship of the parameters of the previous layer, includes: setting the initial state of velocity structure data, including the range of shear wave velocity, total thickness, number of layers and other relevant physical parameters of the first layer, and constructing the state space S;

[0092] Based on historical data or prior models, a transition probability matrix P is defined for each state in the state space S, which describes the possible distribution of the current state to the next state.

[0093] The next layer of parameters for generating velocity structure data is determined based on a dynamic Markov decision strategy.

[0094] Repeat the above action selection and dynamic generation steps until the preset speed structure layer number, data length or other generation accuracy requirements are met;

[0095] The final generated velocity structure data sequence is used as the output.

[0096] In this embodiment, an initial state space S is constructed, which is determined by the shear wave velocity of the first layer of the formation. Range, total thickness H, number of layers N, density and longitudinal wave velocity constitute.

[0097] The range of shear wave velocity Vs, total thickness H, and number of layers N for the first layer of the formation are set according to the actual situation. The P-wave velocity of each layer is approximated using the current mainstream empirical formula. The specific formula is as follows:

[0098]

[0099] in For the first Layer shear wave velocity, For the first Longitudinal wave velocity, From the first floor to the second floor The cumulative depth of the bottom boundary of the layer.

[0100] The density of each layer is approximated using current mainstream empirical formulas. The specific formula is as follows:

[0101] (2)

[0102] in For the first Layer density, For the first Longitudinal wave velocity.

[0103] Setting the initial state. Number of site layers in this embodiment. Set to 10 layers, total depth Set to 30m, minimum layer thickness The first-layer shear wave velocity is 0.8m. Within the constraint range (100, 300) m / s, where the thickness of each layer is... , The generation strategy is as follows:

[0104] Define the normalized order vector:

[0105] (3)

[0106] The initial weights for layer thickness are adjusted by adding a small random perturbation:

[0107] (4)

[0108] in This indicates a tendency for deeper layers to be thicker, reflecting the effect of sedimentary accumulation. This represents a perturbation used to simulate the non-uniformity of layer thickness.

[0109] Final thickness of each layer:

[0110] (5)

[0111] when At that time, randomly generated ,in ,according to Different options are chosen depending on the specific circumstances:

[0112] Choose to continue the trend in incremental layers.

[0113] (6)

[0114] in , This results in greater shallow disturbances and smaller deep disturbances, which better reflects the true temporal characteristics of stratigraphy. Indicates the first Number of floors.

[0115] Choosing a high-speed mezzanine to achieve a jump

[0116] (7)

[0117] Selecting a weak interlayer to achieve a jump descent

[0118] (8)

[0119] In this embodiment, the basic order and first order dispersion curves of the various formation models generated above are calculated using the fast vector transfer method. The two dispersion curves correspond to the same velocity structure, thereby improving the accuracy and stability of the inversion model and enhancing the constraint on the formation velocity structure.

[0120] In this embodiment, the generated various stratigraphic models are matched one-to-one with the calculated basic order and first-order dispersion curves to jointly construct a sample dataset, such as... Figure 1 As shown.

[0121] This example provides a QPSO-optimized CNN-LSTM Rayleigh wave dispersion curve inversion method for accurately and quickly inverting formation Vs and thickness H from the primary and first-order dispersion curves. The specific inversion process is as follows: Figure 3 As shown, it includes:

[0122] The QPSO optimization module uses the Quantum Particle Swarm Optimization (QPSO) algorithm to adaptively optimize the model's hyperparameters in order to obtain the optimal model performance.

[0123] Its optimization steps are as follows Figure 4 As shown, it includes the following:

[0124] (1) Set the initial parameters of QPSO: the calculation dimension is 6, the population size is 40, the number of iterations is 50, and the shrinkage coefficient is... It is 0.5;

[0125] (2) The particle swarm is randomly initialized in the parameter space, and each particle has its own position;

[0126] (3) The mean squared error (MSE) is used as the fitness of the particles. The fitness value is calculated for each particle and measured by the prediction error of the model on the validation set.

[0127] (4) Calculate the optimal average position of the population according to the QPSO formula, update the position by combining the current particle's historical optimal position and the population's optimal position, recalculate the fitness value at the new position, and if there is an improvement, update the corresponding individual optimal position and the population's optimal position.

[0128] (5) When the maximum number of iterations or the fitness convergence threshold is reached, the algorithm stops and outputs the optimal solution.

[0129] Parameter customization and display module: such as Figure 2 As shown, this embodiment allows for flexible customization of the number of particles, iterations, number of optimization samples, and shrinkage / expansion coefficients of QPSO according to actual needs. Through personalized parameter settings, it adapts to the training requirements of different scenarios, optimizes the model training effect, and displays the optimal hyperparameter settings in real time based on the QPSO optimization results.

[0130] Convergence Curve Display Module: After the QPSO optimization process is completed, the module outputs the optimization convergence curve, such as... Figure 5 As shown, the fitness change trend corresponding to the optimal hyperparameter combination in each iteration can be intuitively displayed; through the convergence curve, it is clear to observe whether the optimization process is stable, whether there is oscillation, and whether the algorithm finally reaches convergence.

[0131] CNN Local Feature Learning Module: Convolutional layers are primarily used to extract local features from the input data. By sliding convolutional kernels along the dual-channel dispersion curve of the input, one-dimensional convolution operations are performed on local regions, generating feature maps and achieving weighted aggregation of local features. After convolution, the number of dispersion curve features increases, expanding the information dimension. These features are then processed by batch normalization layers to stabilize the training process and accelerate convergence. The introduction of activation functions then endows the network with non-linear representation capabilities, enabling it to express complex mapping relationships. Finally, max-pooling layers downsample the convolutional features, compressing the sequence length while retaining the features most significantly correlated with the dispersion curve and velocity structure, thereby enhancing the translation invariance of features and improving the generalization ability of the deep learning model. LSTM Long Short-Term Memory Network Module: Consists of the input gate... Forgotten Gate Output gate Candidate memory units With memory unit It consists of five parts that work together to selectively memorize and update key information. The forget gate determines which information needs to be forgotten from the unit state, the input gate updates the storage unit, and the output gate determines the output of the LSTM hidden layer unit in this training step, such as... Figure 6 As shown.

[0132] The calculations for these gates are as follows:

[0133] (1) Establish a forget gate between phase velocity and VS layer velocity structure

[0134] (9)

[0135] (2) At the current frequency point, first through the input gate Controlling the introduction of new information while considering the temporary state of memory units. Update to cell state .

[0136] (10)

[0137] (11)

[0138] (12)

[0139] (3) Calculate the output values ​​of the network at the current time for the phase velocity and VS layer velocity structure.

[0140] (13)

[0141] (14)

[0142] in , , These represent the forget gate, input gate, and output gate, respectively; subscript characters. This indicates that the variable is located in the current training step. In the middle, subscript character This indicates that the variable is located in the previous training step. middle; , , This represents the weight matrix of each gate; , , , Indicates the deviation of each gate; This indicates the output of the layer above the hidden unit; Indicates hidden cell input; express Activation function This represents the hyperbolic tangent function.

[0143] Model training module: Based on training data, the system can quickly build a deep learning model integrating CNN and LSTM. This model fully utilizes the advantages of CNN in local spatial feature extraction; simultaneously, LSTM networks excel at processing time-series or sequential data, capable of modeling the dynamic evolution of dispersion curves with frequency variations, achieving continuous expression of inter-modal and inter-layer relationships. During training, the training loss curve and validation loss curve are displayed in real time, allowing for a direct view of the current convergence trend, fluctuations, and potential overfitting signs. This enables timely identification of problems and adjustment of optimization strategies before training is complete, improving the overall effectiveness and stability of training. Figure 7 As shown.

[0144] Single-step prediction module: Automatically reads the provided fundamental and first-order dispersion curve data, the data format of which is shown in Table 1.

[0145] Table 1

[0146]

[0147] This embodiment standardizes the input data based on the normalization parameters used during training to ensure consistent model predictions. It automatically handles missing or outlier values ​​and applies a mask to mask unreliable first-order dispersion points. The trained model is then invoked for prediction, and a trained CNN-LSTM network is used for forward inference on the normalized dispersion curves. The system supports single-sample or batch predictions without requiring manual adjustment of network parameters. In some embodiments, a profile comparison chart can be displayed to visually show the prediction profile, including layer thickness, velocity distribution, and trends. Users can select different depth ranges or samples for viewing, facilitating rapid evaluation of model prediction accuracy.

[0148] In summary, this embodiment provides a knowledge-physical dual-driven custom loss function to improve the physical accuracy and stability of the inversion, employing a custom loss function that combines stratigraphic physical knowledge with data. While ensuring data fitting, the soil layer impedance ratio is introduced as prior information, reflecting the degree of abrupt changes in the medium properties of adjacent layers, and is used as a weighting coefficient to calculate the error of the shear wave velocity (VS). For the first... The impedance ratio of a soil layer is defined as:

[0149] (15)

[0150] in and They represent the first The first sample The density and transverse wave velocity of the soil layer. When When the deviation from 1 is large, it indicates a significant change in speed, and the model will assign a higher error weight.

[0151] The loss function in this paper is: and It consists of two parts:

[0152] (16)

[0153] (17)

[0154] (18)

[0155] (19)

[0156] in, This indicates the shear wave velocity loss in the current batch; This indicates the thickness loss in the current batch; and These are the weighting coefficients for shear wave velocity and depth; Batch size; This is the [number]th batch of the current batch. One sample; The soil layer resistance ratio; For the first The first sample The actual value velocity of the layer; For the first The first sample The speed of the layer's predicted values; The number of layers is the thickness. For the first The first sample The actual thickness of the layer; for The first sample The predicted thickness of the layer.

[0157] This embodiment utilizes dynamic Markov decision-making to generate more comprehensive stratigraphic data with evolutionary patterns, thereby capturing the correlation of stratigraphic depth sequences.

[0158] This embodiment realizes the joint inversion of geological models using both basic and first-order dispersion curves, thus solving the problem of non-uniqueness in dispersion curve inversion.

[0159] This embodiment utilizes the QPSO-CNN-LSTM model to simultaneously invert speed and thickness, resulting in fast inversion speed, with each sample taking only about 0.3ms.

[0160] This embodiment uses the soil impedance ratio as a physical prior constraint to improve the model inversion accuracy and avoid non-physical phenomena. The average relative error of the predicted samples is between 5% and 7%, and the average relative error of the predicted samples with noise is between 8% and 9%.

[0161] This embodiment demonstrates that the present invention can achieve rapid, high-precision, and strongly constrained Rayleigh wave inversion, and is applicable to scenarios such as geological exploration, foundation testing, engineering planning, and risk assessment.

[0162] Example 2

[0163] A Rayleigh wave dispersion inversion system that integrates physical information includes:

[0164] The geological model building module is configured to set the range of physical property parameters of the initial strata as the initial state for model generation based on prior geological information, and adaptively generate multiple types of layered geological models based on dynamic Markov decision based on the conditional probability relationship of the parameters of the previous strata.

[0165] The dispersion curve calculation module is configured to calculate the corresponding basic and first-order dispersion curves based on multiple types of layered geological models and using the fast vectorized Rayleigh wave forward modeling algorithm to construct dispersion sample data.

[0166] The training sample library construction module is configured to construct a training sample library containing stratigraphic parameters and dispersion characteristics based on the corresponding geological model and its corresponding dispersion sample data.

[0167] The dispersion curve joint inversion model training module is configured to construct a dispersion curve joint inversion model based on deep learning. The module uses the basic dispersion curve, the first-order dispersion curve and the corresponding mask information in the training sample library as the multi-channel input of the model, and uses a loss function that introduces physical information constraints to train the dispersion curve joint inversion model.

[0168] The dispersion curve joint inversion module is configured to process the target geological data using the trained dispersion curve joint inversion model to obtain the layered shear wave velocity and layer thickness parameters of the subsurface medium, thereby realizing the Rayleigh wave dispersion curve inversion.

[0169] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0170] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0171] 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 function 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 1 The function specified in one or more boxes.

[0172] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed 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.

[0173] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A Rayleigh wave dispersion inversion method integrating physical information, characterized in that, Includes the following steps: First, the initial range of physical property parameters of the initial strata is set as the initial state for model generation. Based on the conditional probability relationship of the parameters of the previous stratum, multi-type layered geological models are adaptively generated using dynamic Markov decision-making. The process includes: setting the initial state of velocity structure data and constructing the state space S; the state space S includes: the shear wave velocity V of the first stratum. s Range, total thickness H, number of layers N, density and longitudinal wave velocity V P The shear wave velocity V in the first layer of the strata s The range, total thickness H, and number of layers N are set according to the actual situation, and the density is... and longitudinal wave velocity V P Calculated; Based on historical data or prior models, a transition probability matrix P is defined for each state in the state space S, which describes the possible distribution of the current state to the next state. The next layer of parameters for generating velocity structure data is determined based on a dynamic Markov decision strategy. Repeat the above action selection and dynamic generation steps until the preset requirements are met, and use the final generated velocity structure data sequence as the output result. Based on multi-type layered geological models, the corresponding basic order and first-order dispersion curves are calculated using the fast vectorized Rayleigh wave forward modeling algorithm, and dispersion sample data are constructed. Based on the corresponding geological model and its corresponding dispersion sample data, a training sample library containing stratigraphic parameters and dispersion characteristics is constructed. A deep learning-based joint inversion model of dispersion curves is constructed. The basic order dispersion curves, first order dispersion curves and corresponding mask information in the training sample library are used as multi-channel inputs of the model. The joint inversion model of dispersion curves is trained using a loss function that incorporates physical information constraints. The target geological data is processed using the trained dispersion curve joint inversion model to obtain the layered shear wave velocity and layer thickness parameters of the subsurface medium, thereby realizing the Rayleigh wave dispersion curve inversion. The process of determining the parameters of the next layer for generating velocity structure data based on the dynamic Markov decision strategy includes: each layer has a thickness of... , The generation strategy is as follows: Define the normalized order vector: ； The initial weighting of layer thickness is adjusted by adding a small random perturbation: ； in This indicates a tendency for deeper layers to be thicker, reflecting the effect of sedimentary accumulation. This represents a perturbation used to simulate the non-uniformity of layer thickness; Final thickness of each layer: ； When the number of layers is greater than or equal to 2, the coefficients are generated randomly. ,in ,according to Different options are chosen depending on the specific circumstances: when Choose to continue the trend in increasing layers: ； in , This results in greater shallow disturbances and smaller deep disturbances, which better reflects the true temporal characteristics of stratigraphy. Indicates the first Number of floors; when A high-speed mezzanine is chosen to achieve the jump: ； when Choose a weak interlayer to achieve the jump descent: ； in, For the first Shear wave velocity of the layer, For the first The transverse wave velocity of the layer above the layer, For the first Minimum shear wave velocity of the layer, For the first Maximum shear wave velocity of the layer For the first The thickness of the layer above; The loss function incorporating physical information constraints is as follows: A custom loss function combining geological physics knowledge and data is constructed, introducing soil impedance ratio as prior information; the loss function L is: ； ； ； ； in, This indicates the shear wave velocity loss in the current batch; This indicates the thickness loss in the current batch; and These are the weighting coefficients for shear wave velocity and depth, respectively. Batch size; This is the [number]th batch of the current batch. One sample; The soil layer resistance ratio; For the first The first sample The actual value velocity of the layer; For the first The first sample The speed of the layer's predicted values; The number of layers is the thickness. For the j-th sample The actual thickness of the layer; For the first The first sample The predicted thickness of the layer.

2. The Rayleigh wave dispersion inversion method integrating physical information as described in claim 1, characterized in that, The process of defining a transition probability matrix P for each state in the state space S based on historical data or prior models includes setting an initial probability distribution according to typical site types, so that the initial strata generated by the model have a real engineering background.

3. The Rayleigh wave dispersion inversion method integrating physical information as described in claim 1, characterized in that, The process of constructing a deep learning-based dispersion curve joint inversion model includes: constructing a dispersion curve joint inversion model, which is a convolutional neural network-long short-term memory network model optimized by quantum particle swarm optimization algorithm. In the convolutional neural network-long short-term memory network model, the convolutional neural network is used to extract local features of the input data, and the long short-term memory network model is used to model the dynamic evolution law of the dispersion curve with frequency change, so as to realize the continuous expression of the relationship between modes and layers.

4. The Rayleigh wave dispersion inversion method integrating physical information as described in claim 3, characterized in that, The convolutional neural network includes convolutional layers, batch normalization layers, and max pooling layers. The convolutional layers slide convolution kernels on the input dual-channel dispersion curve to perform one-dimensional convolution operations on local regions, generating feature maps and achieving weighted aggregation of local features. After the convolution output, the number of features in the dispersion curve increases, and the information dimension is expanded. The batch normalization layer is used to perform normalization processing to stabilize the training process and accelerate convergence. An activation function is introduced to give the network nonlinear representation capabilities, enabling it to express complex mapping relationships. Max pooling layers are used to downsample convolutional features, compress sequence length, and retain the features that are most significantly correlated with dispersion curves and velocity structures.

5. The Rayleigh wave dispersion inversion method integrating physical information as described in claim 3, characterized in that, The Long Short-Term Memory (LSTM) network model includes an input gate. Forgotten Gate Output gate Candidate memory units With memory unit Establish phase velocity and V S Forget gates between layer velocity structures: ； At the current frequency point, through the input gate Controlling the introduction of new information while considering the temporary state of memory units. Update to memory cell state : ； ； ； Calculate phase velocity and V S The output value of the network at the current moment for the layer velocity structure: ； ； in, , , These represent the forget gate, input gate, and output gate, respectively; subscript characters. This indicates that the variable is located in the current training step. In the middle, subscript character This indicates that the variable is located in the previous training step. middle; , , This represents the weight matrix of each gate; , , , Indicates the deviation of each gate; This indicates the output of the layer above the hidden unit; Indicates input; express Activation function This represents the hyperbolic tangent function.

6. The Rayleigh wave dispersion inversion method integrating physical information as described in claim 1, characterized in that, For the The impedance ratio of a soil layer is defined as: ； in and They represent the first The first sample The density and shear wave velocity of the soil layer, when When the deviation from 1 exceeds the set value, it indicates a significant speed change, and a higher error weight is given.

7. A Rayleigh wave dispersion inversion system that integrates physical information, characterized in that, Implementing the Rayleigh wave dispersion inversion method that integrates physical information as described in any one of claims 1-6, comprising: The geological model construction module is configured to set the initial range of physical parameters of the initial strata based on prior geological information as the initial state for model generation. Based on the conditional probability relationship of the parameters of the previous stratum, it adaptively generates multiple types of layered geological models using dynamic Markov decision-making. The process includes: setting the initial state of velocity structure data and constructing the state space S; the state space S includes: the shear wave velocity V of the first stratum. s Range, total thickness H, number of layers N, density and longitudinal wave velocity V P The shear wave velocity V in the first layer of the strata s The range, total thickness H, and number of layers N are set according to the actual situation, and the density is... and longitudinal wave velocity V P Calculated; Based on historical data or prior models, a transition probability matrix P is defined for each state in the state space S, which describes the possible distribution of the current state to the next state. The next layer of parameters for generating velocity structure data is determined based on a dynamic Markov decision strategy. Repeat the above action selection and dynamic generation steps until the preset requirements are met, and use the final generated velocity structure data sequence as the output result. The dispersion curve calculation module is configured to calculate the corresponding basic and first-order dispersion curves based on multiple types of layered geological models and using the fast vectorized Rayleigh wave forward modeling algorithm to construct dispersion sample data. The training sample library construction module is configured to construct a training sample library containing stratigraphic parameters and dispersion characteristics based on the corresponding geological model and its corresponding dispersion sample data. The dispersion curve joint inversion model training module is configured to construct a dispersion curve joint inversion model based on deep learning. The module uses the basic dispersion curve, the first-order dispersion curve and the corresponding mask information in the training sample library as the multi-channel input of the model, and uses a loss function that introduces physical information constraints to train the dispersion curve joint inversion model. The dispersion curve joint inversion module is configured to process the target geological data using the trained dispersion curve joint inversion model to obtain the layered shear wave velocity and layer thickness parameters of the subsurface medium, thereby realizing the Rayleigh wave dispersion curve inversion.

Images