Method, medium and device for surface wave dispersion curve inversion based on sensitivity information constraint deep learning

Abstract

The present application belongs to the technical field of geophysical exploration data processing, and relates to a surface wave dispersion curve inversion method based on sensitivity information constraint deep learning, a medium and equipment, the method comprising constructing a Senformer network model, making a data set, optimizing a training process and inversion testing; the method is based on the Senformer network model introducing sensitivity information, by inputting the dispersion curve and its corresponding sensitivity features into the network model, using convolution layers and multi-layer Transformer encoders to extract local and global features, and combining cross attention mechanisms to realize effective constraint of sensitivity information on the inversion process. The present application proposes a SmoothL1 loss function based on sensitivity information weighting, which improves the stability and accuracy of the key horizon shear wave velocity inversion. Through large-scale forward data set training and optimization, the present application can efficiently and stably realize modeling of complex velocity structures, and has significant advantages in noise resistance, stability and inversion accuracy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of exploration geophysics, and particularly relates to a surface wave dispersion curve inversion method based on sensitivity information constraint deep learning, a medium and equipment. BACKGROUND

[0002] Surface wave exploration, as a rapidly developing near-surface exploration technology in recent years, has been widely used in the investigation of shallow and medium-deep underground velocity structure. Rayleigh wave is a typical surface wave propagating along the surface, which is formed by the vertical components of longitudinal and transverse waves interfering with each other on the free surface, and has the advantages of strong energy, slow attenuation and obvious dispersion characteristics. The Multi-channel Analysis of Surface Waves (MASW) method is the core theoretical basis of surface wave exploration, which has promoted the rapid development of shallow seismic exploration technology. The MASW method generally includes three key steps: field data acquisition, dispersion curve extraction and dispersion curve inversion. Among them, dispersion curve inversion is a highly nonlinear and complex process, and how to stably and accurately recover the underground transverse wave velocity structure from the dispersion curve is an important technical problem in current surface wave exploration research.

[0003] The existing dispersion curve inversion methods mainly include two categories: one is the local linear inversion method, such as the least squares method and the Levenberg-Marquardt (L-M) algorithm, etc. This kind of method has fast convergence speed, but it is strongly dependent on the initial model, and the partial derivative involved in the calculation and the inversion result are affected by the precision of the Jacobian matrix. The second is the nonlinear global optimization method, such as genetic algorithm, particle swarm optimization and simulated annealing algorithm, etc. This kind of method does not need to preset the initial model, but the inversion efficiency is low, and it is easy to appear convergence difficulty when the search space is set unreasonably. The above traditional inversion methods are all physically driven methods, and it is difficult to achieve an ideal balance between inversion accuracy and computational efficiency.

[0004] In recent years, the rapid development of deep learning technology in the field of earth science has provided a new solution for dispersion curve inversion. Unlike traditional methods, deep learning is a data-driven inversion strategy that can effectively improve inversion efficiency and reduce dependence on the initial model. Existing methods include inversion methods based on ANN, CNN, LSTM, etc. Although these methods have advantages in speed and computational efficiency, they still have problems such as multiple solutions, insufficient noise resistance, and poor inversion stability when dealing with highly nonlinear dispersion curve inversion problems due to the lack of physical constraints on surface wave propagation mechanisms.

[0005] Therefore, it is urgent to develop a deep learning surface wave dispersion curve inversion method combined with physical information constraints to improve the accuracy and stability of shallow surface transverse wave velocity modeling, thereby overcoming the limitations of traditional methods and existing deep learning methods. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a new surface wave dispersion curve inversion method based on sensitivity information constraint deep learning, a medium and equipment.

[0007] The present application is implemented in a surface wave dispersion curve inversion method based on sensitivity information constraint deep learning, which comprises the following contents:

[0008] Step 1, constructing a Senformer network model: constructing the input layer, feature encoding module, feature extraction module and output layer of the network model; the input of the network model has two types of information, the first type of information is the dispersion curve, containing frequency and phase velocity, and the second type of information is the sensitivity information; in the feature encoding module, frequency position coding is introduced, using the sine, cosine and triangular function position coding method, so that the network can flexibly process dispersion curve input of different lengths;

[0009] The feature extraction module is constructed: the local and global features of the dispersion curve are learned using convolution and multi-layer Transformer encoder, and the cross-attention mechanism is introduced to help the neural network learn the nonlinear mapping between the dispersion curve and the shear wave velocity, and the inversion process is constrained by the sensitivity information;

[0010] Further, the cross-attention module receives two kinds of input information of dispersion curve features and sensitivity features, takes the dispersion curve features as the query vector (Query), and takes the sensitivity features as the key (Key) and value (Value), to realize the dynamic interaction and weighting between them.

[0011] Constructing a sensitivity information weighted SmoothL1 loss function: using the SmoothL1 loss function combining the advantages of L1 and L2 loss, reducing the numerical instability when approaching zero, and at the same time taking the average sensitivity of each stratum as the weight of the loss, introducing the sensitivity information of the layer dimension;

[0012] For each sample, the loss function formula is:

[0013] ;

[0014] Where D is the total number of layers, is the average sensitivity of the i-th layer, is the predicted shear wave velocity of the neural network, is the true shear wave velocity, is the smoothing parameter of the SmoothL1 loss, used to control the transition between L1 and L2 loss, and the value is 1.0;

[0015] Step 2, making a data set: generate a shallow surface shear wave velocity model, calculate the surface wave dispersion curve using the Knopoff forward method, and then calculate the sensitivity information using the following formula:

[0016] ;

[0017] where c represents the phase velocity, f represents the frequency, Vsi represents the shear wave velocity of the i-th layer, represents the perturbation ratio, which is generally set to 1%, represents the sensitivity of the dispersion curve to the perturbation of the shear wave velocity of the i-th layer.

[0018] Further, the data set is divided into a training set, a validation set and a test set according to the ratio of 8:1:1, and 10% Gaussian random noise is added to the training set using a data augmentation strategy.

[0019] Step 3, optimizing the training process: using the data set of step 2 to train the Senformer network model constructed in step 1, in the training process, using AdamW optimizer combined with cosine annealing learning rate scheduling strategy (CosineAnnealingLR) to optimize the training process, and the cosine annealing learning rate scheduler gradually reduces the learning rate from the initial value to the minimum value according to the cosine curve; the SmoothL1 loss function is the quantitative index for parameter optimization and training of the Senformer network model, in the model training process, the SmoothL1 loss function is used to quantify the difference between the network inversion result and the true shear wave velocity, the Senformer network model calculates the difference, and iteratively updates the network parameters using the back propagation algorithm and the gradient descent strategy, thereby continuously improving the inversion accuracy;

[0020] Step 4, using the test set data to perform inversion test on the Senformer network model optimized in step 3.

[0021] The application also provides a computer readable storage medium, which stores a computer program. The computer program, when loaded and executed by a processor, is used to implement the surface wave dispersion curve inversion method based on the Senformer network model.

[0022] The application also provides a computer device, which includes a memory and a processor, and the memory stores a computer program. When the processor executes the computer program, the Senformer network model based surface wave dispersion curve inversion method can be run.

[0023] The application has the following advantages due to the above technical solutions:

[0024] The application provides a surface wave dispersion curve inversion method based on sensitivity information constraint deep learning, which can effectively improve the precision of surface wave dispersion curve inversion and enhance the stability and reliability of shallow surface transverse wave velocity modeling.

[0025] The application innovatively constructs a Senformer network model with sensitivity as a physical constraint. The network extracts local and global features of the dispersion curve through a convolution layer and a Transformer encoder, which can enhance the feature recognition ability of the neural network for the dispersion curve.

[0026] In addition, the application uses a cross-attention mechanism to embed the sensitivity information, and uses a sensitivity weighted SmoothL1 loss function in the training process, so that the layer with high sensitivity is more effectively constrained in the inversion process, thereby improving the inversion precision and result stability of the key layer.

[0027] The application has achieved good inversion effect in both conventional models and complex velocity models, can realize high-precision and high-efficiency dispersion curve inversion, and has good applicability and engineering application value. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is a workflow diagram of the application;

[0029] Figure 2 is a structural schematic diagram of the Senformer network model;

[0030] Figure 3 is the transverse wave velocity profile and dispersion curve of three randomly selected samples, (a) and (b) are sample 1, (c) and (d) are sample 2, and (e) and (f) are sample 3;

[0031] Figure 4 is a loss function curve diagram;

[0032] Figure 5 is a comparison diagram of the inversion results of the test set, (a), (b) and (c) are the inversion results of the three conventional velocity increasing models, (d), (e) and (f) are the inversion results of the three high-speed hard layer models, and (g), (h) and (i) are the inversion results of the three low-speed soft layer models;

[0033] Figure 6 is a dispersion curve error distribution diagram of the inversion results of the test set. DETAILED DESCRIPTION

[0034] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to examples. It should be understood that the specific examples described herein are only used to explain the present application and not to limit the present application.

[0035] The present application provides a new surface wave dispersion curve inversion technology based on sensitivity information constraint deep learning, and constructs a Senformer network model and a sensitivity information weighted SmoothL1 loss function. The present application realizes high-precision dispersion curve inversion for simple velocity models and complex velocity models. The present application will be described in detail below with reference to the accompanying drawings.

[0036] As shown in Figure 1 , the specific steps of the surface wave dispersion curve inversion based on sensitivity information constraint deep learning proposed by the present application are as follows:

[0037] Step S10: Constructing a Senformer network model: as shown in Figure 2 , the Senformer network model includes an input layer, a feature coding module, a feature extraction module and an output layer. The input of the network model has two types of information, the first type of information is the dispersion curve, containing frequency f and phase velocity Vr, and the second type of information is the sensitivity information. The frequency position coding is introduced in the feature coding module, and the sine, cosine triangular function position coding method is used, and its mathematical expression is:

[0038] (1);

[0039] wherein, represents the frequency, and respectively represent the minimum and maximum values of the frequency range, which are used for logarithmic normalization; represents the coding dimension, represents the dimension index, and the value range is ; and respectively represent the position coding values of the frequency in the first and the second dimension.

[0040] Step S20: constructing a feature extraction module: in the feature extraction module, convolution and multi-layer Transformer encoder are used to learn the local and global features of the dispersion curve, and the feature recognition ability of the neural network for the dispersion curve is enhanced. In addition, the introduction of sensitivity information is the core design of the Senformer network model, and the sensitivity reflects the influence degree of the velocity perturbation of different layers on the phase velocity at each frequency of the dispersion curve. The specific implementation method is to introduce the cross-attention mechanism to help the neural network learn the nonlinear mapping between the dispersion curve and the S-wave velocity, and to constrain the inversion process through the sensitivity information, so as to improve the accuracy and stability of the inversion. Figure 2 The cross-attention module receives two kinds of input information of dispersion curve features and sensitivity features, takes the dispersion curve features as the query vector (Query), and takes the sensitivity features as the key (Key) and value (Value), realizes the dynamic interaction and weighting between them, and the cross-attention calculation process is as follows:

[0041] (2);

[0042] Wherein, represents the dimension of Key, is a common normalization function, which can assign weights to each Value by calculating the similarity between Query and Key. Through the introduction of sensitivity features, the cross-attention module can guide the model to pay attention to the dispersion curve information most related to the S-wave velocity change in the learning process.

[0043] Step S30: constructing a sensitivity information weighted SmoothL1 loss function. The loss function constructed is the quantitative index for parameter optimization and training of the Senformer network model. In the model training process, the loss function is used to quantify the difference between the network inversion result and the true S-wave velocity. The network model calculates the difference, iteratively updates the network parameters using the back propagation algorithm and gradient descent strategy, thereby continuously improving the inversion accuracy.

[0044] The loss function constructed by the application is different from the conventional loss function, and the sensitivity information weight is added, which more fully utilizes the physical information. For each sample, the loss function formula is:

[0045] (3);

[0046] Wherein, D is the total number of layers, is the average sensitivity of the i-th layer, is the S-wave velocity predicted by the neural network, is the true S-wave velocity, Smooth parameter for SmoothL1 loss, used to control the transition between L1 and L2 loss, generally taken as 1.0.

[0047] Step S40: Making data sets.

[0048] Fifty thousand shallow surface S-wave velocity models are generated, the surface wave dispersion curves are calculated using the Knopoff forward method, and the sensitivity information is calculated using the following formula (4). Fifty thousand pairs of data sets are divided into training set, validation set and test set according to the ratio of 8:1:1, i.e. forty thousand groups of data are used as training data, and five thousand groups of data are used as validation data and test data. Figure 3 Three groups of data randomly selected from them are shown, including conventional velocity increasing model, high-speed hard interlayer model and low-speed soft interlayer model, etc. In order to enhance the data features, data enhancement strategy is used in the training set, and 10% Gaussian random noise is added.

[0049] (4);

[0050] where c represents the phase velocity, f represents the frequency, Svi represents the S-wave velocity of the i-th layer, δ represents the perturbation ratio, generally set as 1%, Seni represents the sensitivity of the dispersion curve to the perturbation of the S-wave velocity of the i-th layer.

[0051] Step S50: Optimization of training process.

[0052] In the training process, AdamW optimizer combined with cosine annealing learning rate scheduling strategy (CosineAnnealingLR) is used to optimize the training process. AdamW optimizer can adaptively adjust the learning rate of each parameter, making the training process more stable. The cosine annealing learning rate scheduler gradually reduces the learning rate from the initial value to the minimum value according to the cosine curve. This smooth learning rate decay method effectively avoids the training instability phenomenon caused by rapid learning rate reduction, and helps the neural network to make more fine adjustment and update in the later training period.

[0053] In order to evaluate whether the addition of sensitivity information constraint can improve the inversion precision, the Senformer network is compared with the Dispformer network without embedding sensitivity information. The data sets are trained and tested in the Senformer network and the Dispformer network respectively. In the two training processes, the batch size is set to 64, the AdamW optimizer is combined with the cosine annealing learning rate scheduling strategy, the initial learning rate is 0.001, the Senformer network uses the sensitivity weighted SmoothL1 loss, and the Dispformer network uses the SmoothL1 loss. Figure 4The loss function curves of the training set and the validation set in the training process of the Senformer network are trained, and finally the training result of the 529th round is selected as the model parameter for inversion test in the test set.

[0054] The training environment of the present example is:

[0055] (1) Processor: 12th Gen Intel(R) Core(TM) i7-12700F;

[0056] (2) Memory: 64GB;

[0057] (3) System type: 64-bit operating system;

[0058] (4) Graphics card: NVIDIA GeForce RTX 4060, 8GB;

[0059] (5) Running platform: based on the PyTorch deep learning framework, using Jupyter Notebook as the development and training environment.

[0060] Step S60: Inversion test using test set data.

[0061] In order to test the stability and noise resistance of the method, 20% Gaussian random noise is added to the 5000 test set data. Then the test set data is input into the trained Senformer network and Dispformer network respectively, and nine groups of inversion results are extracted from the results, as shown in Figure 5 It can be found from the figure that the accuracy of the shear wave model inverted by the Senformer network is significantly higher than that of the Dispformer network. For special high-speed interlayer models and low-speed interlayer models, the Senformer network uses sensitivity information to constrain the inversion process, which can effectively improve the inversion ability of abnormal strata, and the inversion result is basically consistent with the true model. Then, the test set all inversion results are forward to obtain the corresponding dispersion curves, and the error between the input dispersion curves is calculated to obtain the probability density distribution diagram of the dispersion curve fitting error. As shown in Figure 6 For the inversion results of the Dispformer network, the mean and standard deviation of the dispersion curve fitting error are 2.93 m / s and 9.78 m / s, respectively, while the mean and standard deviation of the error of the Senformer network inversion result are -0.25 m / s and 4.50 m / s, respectively. These results show that the Senformer network proposed in the present application has higher accuracy and stability.

[0062] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any modification, equivalent replacement and improvement within the technical range disclosed by the present application and within the spirit and principle of the present application should be covered in the protection scope of the present application.

Claims

1. A surface wave dispersion curve inversion method based on sensitivity information-constrained deep learning, characterized in that, The method includes the following: Step 1: Construct the Senformer network model: Construct the input layer, feature encoding module, feature extraction module, and output layer of the network model; The input of this network model has two types of information: the first type of information is the dispersion curve, which contains frequency and phase velocity, and the second type of information is sensitivity information; In the feature encoding module, frequency position encoding is introduced, using the sine and cosine trigonometric function position encoding method; The feature extraction module uses convolution and multi-layer Transformer encoders to learn local and global features of dispersion curves and introduces a cross-attention mechanism to constrain the inversion process through sensitivity information. Constructing a sensitivity-weighted SmoothL1 loss function: The SmoothL1 loss function combines the advantages of L1 and L2 losses, while using the average sensitivity of each layer as the weight of the loss, thus introducing layer-level sensitivity information. For each sample, the loss function formula is: ； Where D is the total number of floors. Let be the average sensitivity of the i-th layer. The transverse wave velocity predicted by the neural network. For the true transverse wave velocity, This is the smoothing parameter for SmoothL1 loss, used to control the transition between L1 and L2 losses, and its value is 1.0; Step 2: Create a dataset: Generate a shallow surface shear wave velocity model, calculate the surface wave dispersion curve using the Knopoff forward modeling method, and then calculate the sensitivity information using the following formula; ； Where c represents the phase velocity, Indicates frequency, Represents the transverse wave velocity of the i-th layer. This indicates the perturbation percentage, set to 1%. This indicates the sensitivity of the dispersion curve to the velocity perturbation of the i-th layer shear wave; Step 3: Optimize the training process: Use the dataset from Step 2 to train the Senformer network model constructed in Step 1. During training, the AdamW optimizer combined with a cosine annealing learning rate scheduling strategy is used to optimize the training process. The cosine annealing learning rate scheduler gradually decreases the learning rate from the initial value to the minimum value according to the cosine curve. The SmoothL1 loss function is a quantitative indicator for parameter optimization and training of the Senformer network model. During the training process of the Senformer network model, the SmoothL1 loss function is used to quantify the difference between the network inversion result and the true shear wave velocity. The Senformer network model calculates this difference and iteratively updates the network parameters using the backpropagation algorithm and gradient descent strategy. Step 4: Use the dataset created in Step 2 to perform an inversion test on the Senformer network model that has been optimized and trained in Step 3.

2. The surface wave dispersion curve inversion method based on sensitivity information-constrained deep learning according to claim 1, characterized in that, In step 1, the cross-attention mechanism receives two types of input information: dispersion curve features and sensitivity features. It uses the dispersion curve features as the query vector and the sensitivity features as the key and value to achieve dynamic interaction and weighting between the two.

3. The surface wave dispersion curve inversion method based on sensitivity information-constrained deep learning according to claim 1, characterized in that, A sensitivity-weighted SmoothL1 loss function is constructed, using the average sensitivity of each stratum as the weight of the loss, and layer-dimensional sensitivity information is introduced. The dataset is divided into training, validation and test sets in an 8:1:1 ratio. Data augmentation strategy is used in the training set, adding 10% Gaussian random noise.

4. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is loaded and executed by the processor, it is used to implement the surface wave dispersion curve inversion method based on sensitivity information constrained deep learning as described in any one of claims 1-3.

5. A computer device, the device comprising a memory and a processor, characterized in that, The memory stores a computer program that, when executed by the processor, enables the implementation of the surface wave dispersion curve inversion method based on sensitivity information-constrained deep learning as described in any one of claims 1-3.

Images