Interpolation method of three-dimensional shear wave velocity field in goaf based on micro-vibration detection technology

Abstract

The present invention discloses a three-dimensional shear wave velocity field interpolation method for goaf areas based on micro-motion detection technology, which includes: S1, data preparation; S2, network design; S3, data processing; S4, model training; S5, model verification; S6, model optimization; S7, model application. The artificial intelligence technology based on the three-dimensional convolutional neural network model has the characteristics of strong adaptability and strong nonlinear fitting ability. It can better handle complex terrain and scenes, and can automatically learn complex features in the underground velocity field, including the spatial distribution of velocity, geological structure, etc. It can improve the accuracy and robustness of interpolation through its powerful nonlinear fitting ability and adaptability, and improve the accuracy of velocity field establishment; compared with traditional interpolation methods, the above interpolation method can usually complete training and prediction in a shorter time, which improves calculation efficiency.

Description

Technical Field

[0001] The invention relates to the technical field of establishing three-dimensional shear wave velocity field in goaf area, and in particular to a three-dimensional shear wave velocity field interpolation method in goaf area. Background Art

[0002] Coal mining generally causes certain damage to the ecological environment. Common ones include various geological disasters induced by mining collapse, which will cause social security risks and property losses. Research on how to effectively solve the exploration problems of mining areas is of great significance to protecting people's property and personal safety. In the existing schemes, the exploration of mining areas mostly uses geophysical methods, such as high-density resistivity method, geological radar method, shallow seismic method and transient electromagnetic method, etc. The applicability and effect of various methods are also different. For example, the geological radar method has a shallow detection depth; the shallow seismic method is affected by obstacles and background noise in the city, and the construction is difficult; the electrical exploration is mainly affected by the surrounding electromagnetic field. If the interference is serious, the construction will be difficult.

[0003] At present, there is also a method of micro-vibration detection for goaf exploration, which has the characteristics of simple construction, low cost, and less impact from the surrounding environment, and has great advantages in underground space detection. This method is based on the theory of stationary random processes, extracts surface wave dispersion curves from micro-vibration signals, and obtains the shear wave velocity of the underground medium by inverting the dispersion curve, thereby performing lithology stratification and structural analysis. Among them, micro-vibration detection data can be used to establish the shear wave velocity field in the underground three-dimensional space.

[0004] In actual situations, some goafs usually have complex surface geological conditions and large terrain fluctuations. The traditionally collected micro-seismic data is directly used to establish the underground three-dimensional space shear wave velocity field, which has a large deviation from the actual situation. The main reasons are that the terrain fluctuations lead to discontinuous differences and the terrain boundary effect. For this, it is proposed to calculate the data by interpolation methods. The commonly used interpolation methods for three-dimensional spatial attribute data include trilinear interpolation, polynomial interpolation, radial basis function interpolation, Kriging interpolation, cubic spline interpolation, nearest neighbor interpolation, etc. After inquiry, it is found that the above-mentioned single interpolation calculation method is usually used in the interpolation calculation of underground three-dimensional space data in goafs. However, in complex terrain and scenes, a single interpolation method may not meet all needs, because different methods may show different advantages and disadvantages in different situations, and the accuracy and robustness of the interpolation calculation are insufficient, so that the accuracy of the velocity field establishment is still not high. In order to solve the problem of inaccurate velocity field establishment, the present invention proposes a three-dimensional shear wave velocity field interpolation method for goafs based on micro-seismic detection technology. Summary of the invention

[0005] The purpose of the present invention is to provide a three-dimensional shear wave velocity field interpolation method for goaf areas based on micro-motion detection technology, which can solve the problem of inaccurate establishment of three-dimensional shear wave velocity field.

[0006] To achieve the above purpose, a method for interpolating three-dimensional shear wave velocity field in goaf based on micro-motion detection technology is provided, which includes:

[0007] S1. Data preparation: Collect velocity field data and corresponding geological structure data from different geographical locations and geological environments as data sets, check and preprocess the data sets, and divide the data sets into training sets and validation sets;

[0008] S2. Network design: Design a three-dimensional convolutional neural network model to learn the interpolation law of the velocity field from the input geological structure data; the input layer of the three-dimensional convolutional neural network model can accept a three-dimensional array;

[0009] S3, data processing: preprocessing the training set, including normalization and standardization operations;

[0010] S4, model training: use the preprocessed training set to train the neural network model, and select the mean square error function to measure the difference between the predicted velocity field and the actual velocity field;

[0011] S5. Model validation: Use the validation set to validate the trained neural network model and use mean square error to evaluate the performance of the model;

[0012] S6, model optimization: adjust the optimization algorithm according to the verification results, and adjust the model parameters according to the gradient of the loss function to minimize the loss function value;

[0013] S7. Model application: Apply the optimized neural network model to the actual velocity field interpolation task to obtain the interpolation result of the three-dimensional shear wave velocity field.

[0014] According to the three-dimensional shear wave velocity field interpolation method of goaf area based on micro-motion detection technology, in S1, the geological structure data includes topographic data and geological structure data, the velocity field data includes x, y, z and V, the topographic data includes x, y and z, wherein x, y, z are the three-dimensional spatial coordinates of a single point, V is the shear wave velocity, and the geological structure data includes the absence of faults, rock type mutations, the presence or absence of large caves and the surface landform category.

[0015] According to the three-dimensional shear wave velocity field interpolation method for goaf area based on micro-motion detection technology, in S2, the three-dimensional convolutional neural network model includes:

[0016] Input layer: receives velocity field data and geological structure data as input;

[0017] Multi-scale convolution layer: Use 3D convolution operations with convolution kernels of different scales to extract multi-scale feature parameters in the input data to capture the spatial characteristics of the velocity field in the goaf, and then concatenate the output features of different scales to form a multi-channel feature;

[0018] Residual connection module: The input tensor passes through a multi-scale convolution layer, and uses three convolution kernels of different sizes to perform convolution operations on the input tensor to obtain three new tensors, which are then concatenated to form a multi-channel feature to retain information from different scales; then, the multi-channel feature passes through a normalization layer to normalize each channel of the tensor to increase the stability and convergence speed of the model; then, the normalized tensor passes through an activation function layer, uses the ReLU function as the activation function, and performs a nonlinear transformation on each element in the tensor to increase the expressive power of the model; finally, the input is directly connected to the output, and the output of the residual block is obtained by adding the input vector and the output of the internal convolution network of the residual block, thereby forming a residual block;

[0019] Pooling layer: performs dimensionality reduction processing on the feature maps output by the multi-scale convolutional layer;

[0020] Attention mechanism layer: The attention mechanism assigns different weights to different parts of the input, allowing the model to focus on the more important parts when processing the input. This allows the model to use limited computing resources more efficiently and enhance the effect of network interpolation.

[0021] Fully connected layer: expands the feature map output by the attention mechanism layer into a one-dimensional vector so that it can be input into the fully connected layer and learn the relationship between features through the fully connected layer;

[0022] Output layer: output three-dimensional velocity field value results;

[0023] Among them, spatial characteristics include topography, geological structure, rock type and composition, groundwater and its flow, and geophysical characteristics; the generated feature map is obtained by processing the original data to represent the feature intensity or response of different locations or regions in the model.

[0024] According to the three-dimensional shear wave velocity field interpolation method of goaf area based on micro-seismic detection technology, in S2, the interpolation law of the velocity field is learned from the input geological structure data, including:

[0025] Multi-scale convolutional layers with kernel sizes of 3×3×3, 5×5×5, and 7×7×7 are used to extract the characteristic parameters of velocity fields and terrain data of different scales. Convolutional kernels of different scales work in parallel, and then their outputs are concatenated.

[0026] Use attention layers to assign different weights to different input regions;

[0027] The velocity field and terrain data are fed into the network as multiple channels; each channel represents a different type of data, allowing the network to learn the relevant features of the two types of data at the same time;

[0028] Use residual connection modules to enhance the neural network model's ability to extract complex feature parameters, help alleviate the gradient vanishing and explosion problems, and accelerate the training process;

[0029] Select pooling layers and regularization methods to reduce feature dimensions and improve model robustness, thereby improving the generalization and robustness of the neural network model;

[0030] The fully connected layer is used to convert the features extracted by the previous layers into the final prediction results, select the activation function to introduce nonlinearity, and help the model learn complex mapping relationships;

[0031] The interpolation result of the velocity field is obtained through the fully connected layer and activation function.

[0032] According to the three-dimensional shear wave velocity field interpolation method for goaf area based on micro-motion detection technology, in S4, the specific steps of model training include:

[0033] Batch data processing: input the training set into the neural network model in batches;

[0034] Forward propagation: For each batch of data, forward propagation calculations are first performed; in this step, the input data passes through each layer of the neural network to generate the model's predicted output;

[0035] Calculate the loss function: Use the mean squared error loss function to calculate the difference between the model's predicted output and the actual data;

[0036] Back propagation: After calculating the loss function, the gradient of the loss function to the model parameters is calculated through the back propagation algorithm;

[0037] Parameter optimization: Use the Adam or SGD optimization algorithm to update the weights and biases of the neural network model according to the calculated gradient to optimize the performance of the neural network model; adjust the parameters according to the direction and size of the gradient to gradually reduce the loss function value;

[0038] Repeat training: Repeat the above steps until the predetermined number of training rounds is reached or the condition for stopping training is met; at the end of each training round, the model will perform a complete training on the entire training set;

[0039] Monitoring and evaluation: During the training process, the validation set is used to monitor the performance of the neural network model. The generalization ability of the neural network model on unseen data is evaluated by calculating the loss value or evaluation index on the validation set. The evaluation indicators include accuracy and mean square error.

[0040] Model saving: After training, if the neural network model performs well on the validation set, save it as the final training model;

[0041] Among them, for the difference data between the model prediction output and the actual data, the back propagation algorithm is used to calculate the gradient, and the parameters of the neural network model are updated according to the gradient to reduce the value of the loss function.

[0042] According to the three-dimensional shear wave velocity field interpolation method of goaf area based on micro-motion detection technology,

[0043] In S4, the multi-scale convolutional layer is represented as:

[0044] For input features , there are three convolution kernels of different scales , and Corresponding to the sizes of 3×3×3, 5×5×5 and 7×7×7 respectively, for each position , output feature map It can be calculated as follows:

[0045]

[0046]

[0047]

[0048] in, Represents the convolution operation, m and n are the position indexes of the convolution kernel;

[0049] The output features of different scales are concatenated to form a multi-channel feature to retain information from different scales. The final output feature map can be expressed as:

[0050]

[0051] Among them, the semicolon represents the splicing operation in the channel direction;

[0052] In S4, for the input features , whose size is , the pooling window size is , the step length is , the pooling layer can be expressed as:

[0053] For output features Any location , whose value is given by the following formula:

[0054]

[0055] in, , , The function is to take the maximum value of all elements in the window; in S4, the attention mechanism can be expressed as:

[0056] The input is a five-dimensional tensor ,in is the batch size, is the number of channels, , , The three-dimensional space coordinates of a single point; the input is transformed through a three-dimensional linear transformation , and get a new tensor :

[0057]

[0058] in, represents the transpose of the input;

[0059] Will Split into three parts , and , each part has channels, stacking these three transformed tensors together to form a new tensor ;

[0060]

[0061] from Extract features and aggregate them, expressed as:

[0062]

[0063] in, is the input tensor, whose size is , and They are The batch size and number of channels, , , The three-dimensional space coordinates of a single point. is the output feature vector, whose size is , and They are The batch size and number of channels. In this formula, X, Y, and Z represent all coordinate data in a group that participate in the calculation.

[0064] The aggregated features are transformed through a linear transformation Linear1, then through a nonlinear activation function, and then through another linear transformation Linear2 to obtain a shape of Tensor , normalized by the softmax function, and the attention weight matrix is ​​obtained , which can be expressed as:

[0065]

[0066]

[0067] Weighted combination results for:

[0068]

[0069] in, is the attention weight A sub-tensor in , corresponding to the The weight of the transformation, is the attention weight A sub-tensor in ;

[0070] The final weighted combination result for:

[0071]

[0072] in, is the attention weight A sub-tensor in , corresponding to the The weight of the transformation, is the attention weight A sub-tensor in ;

[0073] Finally, the output is transformed again through a linear , and get the final output :

[0074]

[0075] Here Represents a transpose operation to ensure that the output has the correct dimension order.

[0076] According to the three-dimensional shear wave velocity field interpolation method of goaf area based on micro-motion detection technology, in S5, the mean square error is calculated as follows:

[0077]

[0078] in, is the velocity field value predicted by the model, Vi is the actual velocity field value, and N is the number of samples.

[0079] According to the three-dimensional shear wave velocity field interpolation method of goaf area based on micro-vibration detection technology, in S6, in the Adam optimization algorithm, it is assumed that , , , Respectively represent the update of each value, and the formula is as follows:

[0080]

[0081]

[0082]

[0083] W and b appear before and after the equal sign because this is a recursive formula that describes the update process of the parameters in each iteration. The W and b on the left side of the equal sign represent the updated parameter values, while the W and b on the right side of the equal sign represent the parameter values ​​before the update.

[0084] in, Represents the weight The corrected value of the partial derivative of Indicates bias The corrected value of the partial derivative of Represents the weight The corrected value of the squared partial derivative of Indicates bias The corrected value of the squared partial derivative of and respectively control the decay rate of the exponentially weighted average of the first-order moment and the second-order moment, Represents the weight The partial derivative of Indicates bias The partial derivative of is the learning rate, which is used to control the size of the update. A very small constant is a small value introduced to avoid the denominator being zero. and is the first-order moment estimate without bias correction, and the weights are and bias At the current time step The weighted average of the gradients of , corresponding to, and is the unbias-corrected second-order moment estimate.

[0085] According to the three-dimensional shear wave velocity field interpolation method for goaf area based on micro-motion detection technology, in S7, the optimized neural network model is applied to the actual velocity field interpolation task, and the interpolation results of the three-dimensional shear wave velocity field are obtained, including:

[0086] Input the three-dimensional spatial position coordinates of the velocity field to be interpolated;

[0087] The neural network model takes these spatial position coordinates as input and predicts the output to obtain the velocity field value at the corresponding position through the learned interpolation law;

[0088] The output result is the interpolation result of the three-dimensional velocity field.

[0089] According to the three-dimensional shear wave velocity field interpolation method for goaf areas based on micro-motion detection technology, the characteristic parameters include the distance and height difference between two adjacent points.

[0090] Beneficial effects: As mentioned above, the artificial intelligence technology based on the three-dimensional convolutional neural network model has the characteristics of strong adaptability and strong nonlinear fitting ability. It can better handle complex terrain and scenes, and can automatically learn the complex features in the underground velocity field, including the spatial distribution of velocity, geological structure, etc. It can improve the accuracy and robustness of interpolation through its powerful nonlinear fitting ability and adaptability, and improve the accuracy of velocity field establishment; compared with traditional interpolation methods, the above interpolation method can usually complete training and prediction in a shorter time, thereby improving computational efficiency.

[0091] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] The present invention is further described below in conjunction with the accompanying drawings and embodiments;

[0093] Figure 1 is a flow chart of the method of the present invention;

[0094] Figure 2 is the ReLU function image;

[0095] Figure 3 Schematic diagram of residual block. DETAILED DESCRIPTION

[0096] This section will describe in detail the specific embodiments of the present invention. The preferred embodiments of the present invention are shown in the accompanying drawings. The purpose of the accompanying drawings is to supplement the description of the text part of the specification with graphics, so that people can intuitively and vividly understand each technical feature and the overall technical solution of the present invention, but it cannot be understood as a limitation on the scope of protection of the present invention.

[0097] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, connecting, etc. should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0098] Reference Figure 1-3 , a three-dimensional shear wave velocity field interpolation method for goaf area based on micro-motion detection technology includes the following contents:

[0099] S1. Data preparation: Collect velocity field data and corresponding geological structure data from different geographical locations and geological environments as data sets, check and preprocess the data sets, and divide the data sets into training sets and validation sets;

[0100] S2. Network design: Design a three-dimensional convolutional neural network model to learn the interpolation law of the velocity field from the input geological structure data; the input layer of the three-dimensional convolutional neural network model can accept a three-dimensional array;

[0101] S3, data processing: preprocessing the training set, including normalization and standardization operations;

[0102] S4, model training: use the preprocessed training set to train the neural network model, and select the mean square error function to measure the difference between the predicted velocity field and the actual velocity field;

[0103] S5. Model validation: Use the validation set to validate the trained neural network model and use mean square error to evaluate the performance of the model;

[0104] S6. Model optimization: adjust the optimization algorithm according to the verification results, and adjust the model parameters according to the gradient of the loss function (calculated using mean square error) to minimize the loss function value;

[0105] S7. Model application: Apply the optimized neural network model to the actual velocity field interpolation task to obtain the interpolation result of the three-dimensional shear wave velocity field.

[0106] As mentioned above, artificial intelligence technology based on the three-dimensional convolutional neural network model has the characteristics of strong adaptability and strong nonlinear fitting ability. It can better handle complex terrain and scenes, and can automatically learn complex features in the underground velocity field, including the spatial distribution of velocity, geological structure, etc. It can improve the accuracy and robustness of interpolation through its powerful nonlinear fitting ability and adaptability. Compared with traditional interpolation methods, the above interpolation method can usually complete training and prediction in a shorter time, which improves computational efficiency. Based on the above method, underground three-dimensional spatial data can be obtained efficiently, quickly and reliably, providing important support for the mining and safety monitoring of goaf areas.

[0107] In S1, in order to ensure that the model can accurately predict the geological structure from the velocity field data, velocity field data and geological structure data from different geographical locations and geological environments are collected as data sets. To ensure the quality of the data, the inspection and preprocessing of the data set include: checking whether there are missing values, outliers, etc., and performing necessary preprocessing, such as filling missing values ​​and removing outliers.

[0108] Specifically, in S1, geological structure data includes topographic data and geological structure data, velocity field data includes x, y, z and V, and topographic data includes x, y and z, where x, y and z are the three-dimensional spatial coordinates of a single point, and V is the shear wave velocity. The geological structure data specifically includes: whether there are faults, rock mutations, whether there are large caves, and surface landform types.

[0109] In S2, the designed 3D convolutional neural network model includes input layer, multi-scale convolution layer, residual connection module, pooling layer, attention mechanism layer, fully connected layer and output layer. Among them, the geological structure data is three-dimensional, so the input layer should accept a three-dimensional array, use 3D convolution kernels of different sizes (3×3×3, 5×5×5 and 7×7×7) to capture features at different scales, use convolution kernels of different sizes in parallel, and then splice the results to form a multi-scale feature representation. Add residual connection modules between convolution layers to help the model learn deeper features and alleviate the gradient vanishing problem. Use ReLU as the activation function to introduce nonlinearity to help the model learn more complex features. Introduce the attention mechanism to automatically adjust the degree of attention according to different areas of the input data, which helps the model focus on the most relevant geological attributes. Use multiple layers of fully connected layers to further extract features and finally output the interpolation results of the velocity field. Use L1 regularization to prevent overfitting, and discard a certain proportion of neuron outputs to further enhance the robustness of the model.

[0110] Specifically, the three-dimensional convolutional neural network model includes:

[0111] Input layer: receives geological structure data as input;

[0112] Multi-scale convolution layer: The characteristics of geological structure data are complex and have multi-scale features, while a single-size convolution kernel can only learn features of a specific scale. The multi-scale convolution layer is a network composed of convolution kernels of different sizes connected in parallel, which not only effectively solves the problem of selecting the size of the convolution kernel, but also can effectively capture the complex characteristics of seismic data. The 3D convolution operation using convolution kernels of different scales is used to extract multi-scale feature parameters in the input data to capture the spatial characteristics of the velocity field in the goaf, and then the output features of different scales are spliced ​​together to form a multi-channel feature;

[0113] Residual connection module: The input tensor is passed through a multi-scale convolution layer, and three convolution kernels of different sizes are used to perform convolution operations on the input tensor to obtain three new tensors. The three new tensors are then concatenated to form a multi-channel feature to retain information from different scales. Then, the multi-channel feature is passed through a normalization layer to normalize each channel of the tensor to increase the stability and convergence speed of the model. Next, the standardized tensor is passed through an activation function layer, using the ReLU function as the activation function to perform a nonlinear transformation on each element in the tensor to increase the expressive power of the model. Finally, the input is directly connected to the output. The output of the residual block is obtained by adding the input vector and the output of the internal convolution network of the residual block to form a residual block.

[0114] Pooling layer: It performs dimensionality reduction processing on the features output by the multi-scale convolutional layer to reduce the complexity of the model, thereby reducing the complexity of subsequent calculations. At the same time, by reducing the number of network parameters, the pooling layer can prevent overfitting to a certain extent and improve the generalization ability of the model;

[0115] Attention mechanism layer: The attention mechanism assigns different weights to different parts of the input, allowing the model to focus on the more important parts when processing the input. This allows the model to use limited computing resources more efficiently and enhance the effect of network interpolation.

[0116] Fully connected layer: The feature map output by the attention mechanism layer is expanded into a one-dimensional vector so that it can be input into the fully connected layer. The main purpose of the fully connected layer is to learn the combination relationship between different features. Each neuron is connected to all neurons in the previous layer, which means it can learn the relationship between all input features;

[0117] Output layer: Output three-dimensional velocity field value results.

[0118] The captured spatial features include topography, geological structure, rock type and composition, groundwater and its flow, and geophysical characteristics. The generated feature map is generated by processing raw data, such as topographic data, geological structure data, velocity field data, etc., to represent the feature intensity or response of different locations or regions in the model. The relationship between features mainly refers to the various connections and influences that may exist between different features in the data. It can be direct (such as linear or nonlinear relationship), indirect (influence transmitted through other features), or statistically correlated (positive or negative).

[0119] In S2, “learning the interpolation law of the velocity field from the input geological structure data” specifically includes:

[0120] Multi-scale convolutional layers with kernel sizes of 3×3×3, 5×5×5, and 7×7×7 are used to extract the characteristic parameters of velocity fields and terrain data of different scales. Convolutional kernels of different scales work in parallel, and then their outputs are concatenated.

[0121] The attention mechanism layer can adaptively assign different weights to different input regions, thereby better capturing important feature information;

[0122] The velocity field and terrain data are fed into the network as multiple channels; each channel represents a different type of data, which allows the network to learn the relevant features of the two types of data at the same time;

[0123] Use residual connection modules to enhance the ability of neural network models to extract complex feature parameters, and can help alleviate the gradient disappearance / explosion problem and accelerate the training process;

[0124] Select pooling layers and regularization methods, such as Max Pooling, Average Pooling or Global Average Pooling and L1, L2 regularization or Dropout, to reduce feature dimensions and improve model robustness. Improve the generalization ability and robustness of the neural network model;

[0125] The fully connected layer is used to convert the features extracted by the previous layers into the final prediction results. It selects appropriate activation functions, such as ReLU, Leaky ReLU, or Swish, to introduce nonlinearity and help the model learn complex mapping relationships.

[0126] The interpolation result of the velocity field is obtained through the fully connected layer and activation function.

[0127] The interpolation rule is the regular pattern of velocity field interpolation learned by the 3D convolutional neural network model from the input geological structure data, including feature parameter extraction, multi-channel input or fusion strategy, residual connection module, pooling layer and regularization method, fully connected layer and activation function. In addition, the size and number of convolution kernels should be determined according to the data characteristics to capture information at different scales.

[0128] In S3, preprocessing the training set is an important step to improve the stability and convergence speed of neural network model training. Reasonable preprocessing can not only help the model converge faster, but also improve the generalization ability of the model. The purpose of normalization is to scale the data to a fixed range, usually the interval [0, 1]. The advantage of doing so is that it can avoid the impact of too large or too small values ​​on model training, and it also helps to improve the convergence speed of the model.

[0129] In S4, when training the neural network model, it is crucial to select the mean square error function to measure the difference between the predicted velocity field and the actual velocity field. It includes loading the preprocessed training dataset; defining the model architecture, including the input layer, convolution layer, pooling layer, attention mechanism layer, fully connected layer, etc.; setting the early stopping method, when the performance on the validation set does not improve within a certain round, terminate the training early to prevent overfitting.

[0130] Specifically, in S4, the steps of model training include:

[0131] Batch data processing: The training set is input into the neural network model in batches; each batch contains a certain number of data samples, which helps to speed up the training process and improve the stability of the neural network model;

[0132] Forward propagation: For each batch of data, forward propagation calculation is first performed; in this step, the input data passes through each layer of the neural network (convolutional layer, pooling layer, fully connected layer, etc.) to generate the model's predicted output;

[0133] Calculate the loss function: Use the mean squared error loss function to calculate the difference between the model's predicted output and the actual data;

[0134] Back propagation: After calculating the loss function, the gradient of the loss function to the model parameters is calculated through the back propagation algorithm;

[0135] Parameter optimization: Use the Adam or SGD optimization algorithm to update the model's weights and biases according to the calculated gradients to gradually optimize the model's performance; adjust the parameters according to the direction and size of the gradient to gradually reduce the loss function value;

[0136] Repeat training: Repeat the above steps until the predetermined number of training rounds (epochs) is reached or the condition for stopping training is met (for example, the loss function has reached the minimum); at the end of each training round, the model will perform a complete training on the entire training set;

[0137] Monitoring and evaluation: During the training process, use the validation set to monitor the performance of the model; evaluate the generalization ability of the model on unseen data by calculating the loss value or other evaluation indicators (such as accuracy, mean square error, etc.) on the validation set;

[0138] Model saving: After training, if the model performs well on the validation set (for example, the calculation deviation is less than 2%), save it as the final training model for subsequent application and evaluation.

[0139] In S4, the multi-scale convolutional layer is represented as:

[0140] For input features , there are three convolution kernels of different scales , and Corresponding to the sizes of 3×3×3, 5×5×5 and 7×7×7 respectively, for each position , output feature map It can be calculated as follows:

[0141]

[0142]

[0143]

[0144] in, Represents the convolution operation, m and n are the position indexes of the convolution kernel;

[0145] The output features of different scales are concatenated to form a multi-channel feature to retain information from different scales. The final output feature map can be expressed as:

[0146]

[0147] Among them, the semicolon represents the splicing operation in the channel direction;

[0148] In S4, for the input features , whose size is , the pooling window size is , the step length is , the pooling layer can be expressed as:

[0149] For output features Any location , whose value is given by the following formula:

[0150]

[0151] in, , , The function is to take the maximum value of all elements in the window; in S4, the attention mechanism can be expressed as:

[0152] The input is a five-dimensional tensor ,in is the batch size, is the number of channels, , , The three-dimensional space coordinates of a single point; the input is transformed through a three-dimensional linear transformation , and get a new tensor :

[0153]

[0154] in, represents the transpose of the input;

[0155] Will Split into three parts , and , each part has channels, stacking these three transformed tensors together to form a new tensor ;

[0156]

[0157] First from Extract features and aggregate them, expressed as:

[0158]

[0159] in, is the input tensor, whose size is , and They are The batch size and number of channels, , , The three-dimensional space coordinates of a single point. is the output feature vector, whose size is , and They are The batch size and number of channels. In this formula, X, Y, and Z represent all coordinate data in a group that participate in the calculation.

[0160] The aggregated features are then transformed through a linear transformation Linear1, then through a nonlinear activation function (GELU), and then through another linear transformation Linear2. Finally, a shape of Tensor , normalized by the softmax function, and the final attention weight matrix is ​​obtained , which can be expressed as:

[0161]

[0162]

[0163] The final weighted combination result for:

[0164]

[0165] in, is the attention weight A sub-tensor in , corresponding to the The weight of the transformation, is the attention weight A sub-tensor in ;

[0166] Finally, the output is transformed again through a linear , and get the final output :

[0167]

[0168] Here Represents a transpose operation to ensure that the output has the correct dimension order.

[0169] In S5, analyze the loss function value on the validation set to determine whether the model is overfitting or underfitting. In S6, adjust the optimization algorithm based on the validation results. You can choose Adam or SGD, and adjust the model parameters based on the gradient of the loss function to minimize the loss function value. Model parameters include learning rate, etc. To adjust the learning rate, you can try different initial learning rates or use a learning rate scheduling strategy.

[0170] In S5, the mean square error is calculated as follows:

[0171]

[0172] in, is the velocity field value predicted by the model, Vi is the actual velocity field value, and N is the number of samples.

[0173] In S6, in the Adam optimization algorithm, it is assumed that , , , Respectively represent the update of each value, and the formula is as follows:

[0174]

[0175]

[0176]

[0177] W and b appear before and after the equal sign because this is a recursive formula that describes the update process of the parameters in each iteration. The W and b on the left side of the equal sign represent the updated parameter values, while the W and b on the right side of the equal sign represent the parameter values ​​before the update.

[0178] in, Represents the weight The corrected value of the partial derivative of Indicates bias The corrected value of the partial derivative of Represents the weight The corrected value of the squared partial derivative of Indicates bias The corrected value of the squared partial derivative of and respectively control the decay rate of the exponentially weighted average of the first-order moment and the second-order moment, Represents the weight The partial derivative of Indicates bias The partial derivative of is the learning rate, which is used to control the size of the update. A very small constant is a small value introduced to avoid the denominator being zero. and is the first-order moment estimate without bias correction, and the weights are and bias At the current time step The weighted average of the gradients of , corresponding to, and is the unbias-corrected second-order moment estimate.

[0179] The mean square error is used to reflect the difference between the estimator and the estimated quantity, and the difference is used as the verification result, based on which the model is optimized. In S6, after minimizing the loss function value, the optimized neural network model can meet the actual use requirements. The loss function is the mean square error calculation.

[0180] In S4, for the data that produces the difference, the back propagation algorithm is used to calculate the gradient, and the parameters of the model are updated according to the gradient to reduce the value of the loss function. This process is iterative, and the model is optimized through multiple iterations to better fit the training data. The loss function value is used in the verification process to determine whether the generalization ability and prediction performance of the model meet the expected requirements.

[0181] In S5, the mean square error data itself is not the "validation result" described in step S6. The mean square error data is an indicator used to measure the performance of the model on the validation set. The validation result is a combination of multiple indicators and evaluation results obtained during the entire validation process.

[0182] In S7, the optimized neural network model is applied to the actual velocity field interpolation task, and the interpolation results of the three-dimensional shear wave velocity field are obtained, including:

[0183] Input the three-dimensional spatial position coordinates of the velocity field to be interpolated;

[0184] The neural network model takes these spatial position coordinates as input and predicts the output to obtain the velocity field value at the corresponding position through the learned interpolation law;

[0185] The output result is the interpolation result of the three-dimensional velocity field.

[0186] Among them, in the above scheme, the characteristic parameters include the distance and height difference between two adjacent points.

[0187] In summary, the above methods include data preparation, network design, training, and optimization. The main purpose is to establish a deep learning model that can accurately predict the underground velocity field, and emphasize the importance of model optimization to improve the performance and generalization ability of the model. In complex terrain and scenes, a single interpolation method may not meet all requirements. Therefore, it is necessary to comprehensively consider the characteristics of terrain data, computing requirements, and accuracy requirements, and select the most appropriate interpolation method to obtain more accurate interpolation results.

[0188] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge scope of ordinary technicians in the technical field without departing from the purpose of the present invention.

Claims

1. A three-dimensional shear wave velocity field interpolation method for goaf areas based on micro-vibration detection technology, characterized in that: include: S1. Data preparation: Collect velocity field data and corresponding geological structure data from different geographical locations and geological environments as data sets, check and preprocess the data sets, and divide the data sets into training sets and validation sets; S2. Network design: Design a three-dimensional convolutional neural network model to learn the interpolation law of the velocity field from the input geological structure data; the input layer of the three-dimensional convolutional neural network model can accept a three-dimensional array; S3, data processing: preprocessing the training set, including normalization and standardization operations; S4, model training: use the preprocessed training set to train the neural network model, and select the mean square error function to measure the difference between the predicted velocity field and the actual velocity field; S5. Model validation: Use the validation set to validate the trained neural network model and use mean square error to evaluate the performance of the model; S6, model optimization: adjust the optimization algorithm according to the verification results, and adjust the model parameters according to the gradient of the loss function to minimize the loss function value; S7, model application: apply the optimized neural network model to the actual velocity field interpolation task to obtain the interpolation result of the three-dimensional shear wave velocity field; Among them, the interpolation law of learning the velocity field from the input geological structure data includes: Multi-scale convolutional layers with kernel sizes of 3×3×3, 5×5×5, and 7×7×7 are used to extract the characteristic parameters of velocity fields and terrain data of different scales. Convolutional kernels of different scales work in parallel, and then their outputs are concatenated. Use attention layers to assign different weights to different input regions; The velocity field and terrain data are fed into the network as multiple channels; each channel represents a different type of data, allowing the network to learn the relevant features of the two types of data at the same time; Use residual connection modules to enhance the neural network model's ability to extract complex feature parameters, help alleviate the gradient vanishing and explosion problems, and accelerate the training process; Select pooling layers and regularization methods to reduce feature dimensions and improve model robustness, thereby improving the generalization and robustness of the neural network model; The fully connected layer is used to convert the features extracted by the previous layers into the final prediction results, select the activation function to introduce nonlinearity, and help the model learn complex mapping relationships; The interpolation result of the velocity field is obtained through the fully connected layer and activation function.

2. The method for interpolating three-dimensional shear wave velocity field in goaf based on micro-vibration detection technology according to claim 1 is characterized in that: In S1, the geological structure data includes topographic data and geological structure data, the velocity field data includes x, y, z and V, the topographic data includes x, y and z, where x, y and z are the three-dimensional spatial coordinates of a single point, V is the shear wave velocity, and the geological structure data includes the presence or absence of faults, rock type mutations, the presence or absence of large caves and the type of surface landforms.

3. The method for interpolating three-dimensional shear wave velocity field in goaf based on micro-vibration detection technology according to claim 2 is characterized in that: In S2, the 3D convolutional neural network model includes: Input layer: receives velocity field data and geological structure data as input; Multi-scale convolution layer: Use 3D convolution operations with convolution kernels of different scales to extract multi-scale feature parameters in the input data to capture the spatial characteristics of the velocity field in the goaf, and then concatenate the output features of different scales to form a multi-channel feature; Residual connection module: The input tensor passes through a multi-scale convolution layer, and uses three convolution kernels of different sizes to perform convolution operations on the input tensor to obtain three new tensors, which are then concatenated to form a multi-channel feature to retain information from different scales; then, the multi-channel feature passes through a normalization layer to normalize each channel of the tensor to increase the stability and convergence speed of the model; then, the normalized tensor passes through an activation function layer, uses the ReLU function as the activation function, and performs a nonlinear transformation on each element in the tensor to increase the expressive power of the model; finally, the input is directly connected to the output, and the output of the residual block is obtained by adding the input vector and the output of the internal convolution network of the residual block, thereby forming a residual block; Pooling layer: performs dimensionality reduction processing on the feature maps output by the multi-scale convolutional layer; Attention mechanism layer: The attention mechanism assigns different weights to different parts of the input, allowing the model to focus on the more important parts when processing the input. This allows the model to use limited computing resources more efficiently and enhance the effect of network interpolation. Fully connected layer: expands the feature map output by the attention mechanism layer into a one-dimensional vector so that it can be input into the fully connected layer and learn the relationship between features through the fully connected layer; Output layer: output three-dimensional velocity field value results; Among them, spatial characteristics include topography, geological structure, rock type and composition, groundwater and its flow, and geophysical characteristics; the generated feature map is obtained by processing the original data to represent the feature intensity or response of different locations or regions in the model.

4. The method for interpolating three-dimensional shear wave velocity field in goaf based on micro-motion detection technology according to claim 3 is characterized in that: In S4, the specific steps of model training include: Batch data processing: input the training set into the neural network model in batches; Forward propagation: For each batch of data, forward propagation calculations are first performed; in this step, the input data passes through each layer of the neural network to generate the model's predicted output; Calculate the loss function: Use the mean squared error loss function to calculate the difference between the model's predicted output and the actual data; Back propagation: After calculating the loss function, the gradient of the loss function to the model parameters is calculated through the back propagation algorithm; Parameter optimization: Use the Adam or SGD optimization algorithm to update the weights and biases of the neural network model according to the calculated gradient to optimize the performance of the neural network model; adjust the parameters according to the direction and size of the gradient to gradually reduce the loss function value; Repeat training: Repeat the above steps until the predetermined number of training rounds is reached or the condition for stopping training is met; at the end of each training round, the model will perform a complete training on the entire training set; Monitoring and evaluation: During the training process, the validation set is used to monitor the performance of the neural network model. The generalization ability of the neural network model on unseen data is evaluated by calculating the loss value or evaluation index on the validation set. The evaluation indicators include accuracy and mean square error. Model saving: After training, if the neural network model performs well on the validation set, save it as the final training model; Among them, for the difference data between the model prediction output and the actual data, the back propagation algorithm is used to calculate the gradient, and the parameters of the neural network model are updated according to the gradient to reduce the value of the loss function.

5. The method for interpolating three-dimensional shear wave velocity field in goaf based on micro-motion detection technology according to claim 4 is characterized in that: In S4, the multi-scale convolutional layer is represented as: For input features , there are three convolution kernels of different scales , and Corresponding to the sizes of 3×3×3, 5×5×5 and 7×7×7 respectively, for each position , output feature map It can be calculated as follows: in, Represents the convolution operation, m and n are the position indexes of the convolution kernel; The output features of different scales are concatenated to form a multi-channel feature to retain information from different scales. The final output feature map can be expressed as: Among them, the semicolon represents the splicing operation in the channel direction; In S4, for the input features , whose size is , the pooling window size is , the step length is , the pooling layer can be expressed as: For output features Any location , whose value is given by the following formula: in, , , The function is to take the maximum value of all elements in the window; in S4, the attention mechanism can be expressed as: The input is a five-dimensional tensor ,in is the batch size, is the number of channels, , , is the three-dimensional space coordinate of a single point; the input is transformed through a three-dimensional linear transformation , and get a new tensor : in, represents the transpose of the input; Will Split into three parts , and , each part has channels, stacking these three transformed tensors together to form a new tensor ; from Extract features and aggregate them, expressed as: The aggregated features are transformed through a linear transformation Linear1, then through a nonlinear activation function, and then through another linear transformation Linear2 to obtain a shape of Tensor , normalized by the softmax function, and the attention weight matrix is ​​obtained , which can be expressed as: Weighted combination results for: in, is the attention weight A sub-tensor in , corresponding to the The weight of the transformation, is the attention weight A sub-tensor in ; Finally, the output is transformed again through a linear , and get the final output : Here Represents a transpose operation to ensure that the output has the correct dimension order.

6. The method for interpolating three-dimensional shear wave velocity field in goaf based on micro-vibration detection technology according to claim 5 is characterized in that: In S5, the mean square error is calculated as follows: in, is the velocity field value predicted by the model, Vi is the actual velocity field value, and N is the number of samples.

7. The method for interpolating three-dimensional shear wave velocity field in goaf based on micro-motion detection technology according to claim 5 is characterized in that: In S6, in the Adam optimization algorithm, it is assumed , , , Respectively represent the update of each value, and the formula is as follows: in, Represents the weight The corrected value of the partial derivative of Indicates bias The corrected value of the partial derivative of Represents the weight The corrected value of the squared partial derivative of Indicates bias The corrected value of the squared partial derivative of and respectively control the decay rate of the exponentially weighted average of the first-order moment and the second-order moment, Represents the weight The partial derivative of Indicates bias The partial derivative of is the learning rate, which is used to control the size of the update. A very small constant is a small value introduced to avoid the denominator being zero.

8. The method for interpolating three-dimensional shear wave velocity field in goaf based on micro-motion detection technology according to claim 5 is characterized in that: In S7, the optimized neural network model is applied to the actual velocity field interpolation task, and the interpolation results of the three-dimensional shear wave velocity field are obtained, including: Input the three-dimensional spatial position coordinates of the velocity field to be interpolated; The neural network model takes these spatial position coordinates as input and predicts the output to obtain the velocity field value at the corresponding position through the learned interpolation law; The output result is the interpolation result of the three-dimensional velocity field.

9. The method for interpolating three-dimensional shear wave velocity field in goaf based on micro-vibration detection technology according to claim 1 is characterized in that: The characteristic parameters include the distance and height difference between two adjacent points.

Images