Rock mechanical parameter prediction method based on improved convolutional neural network

Abstract

The invention discloses a rock mechanical parameter prediction method based on an improved convolutional neural network, and belongs to the technical field of geological exploration. The method comprises the steps of firstly collecting a logging curve of a target mining area and core experiment data to construct an initial data set; sensitive logging curves are screened through data cleaning, normalization and correlation analysis to serve as feature input; combining logging data of a plurality of continuous depth points into a feature matrix as a training sample, and introducing context information; the core of the method is to construct a convolutional neural network model using an asymmetric convolution kernel, fully learn spatial structure features of logging data through longitudinal and transverse differential feature extraction capability of the convolutional neural network model, and establish a complex nonlinear mapping relationship between the logging data and rock mechanical parameters; the model is optimized in combination with an Adam optimizer, learning rate scheduling and an early stop mechanism, and an optimal model is determined through grid search and K-fold cross validation. Therefore, efficient, accurate and automatic prediction of the full-well-section rock mechanical parameters is realized, and the identification efficiency and the geological adaptability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, and specifically to a method for predicting rock mechanical parameters based on an improved convolutional neural network. Background Technology

[0002] In coalbed methane exploration and development, the rock mechanics parameters of coal-bearing strata are the core geological basis for accurate reservoir evaluation and efficient stimulation. Accurate mechanics parameter profiles are not only fundamental for classifying coal seam structure and identifying high-permeability "sweet spots," but also a decisive factor in optimizing fracturing design. In reservoir fracturing operations, whether targeting the coal seam itself or employing roof fracturing technology to avoid coal dust damage, success hinges on accurate assessment of the mechanical zones. The prediction of construction pressure, the control of fracture geometry, and the optimization of completion schemes are all directly governed by the formation mechanical behavior revealed by these parameters, ultimately determining the final development effect and gas production life of the reservoir by influencing the complexity of the seepage network formed by fracturing. Furthermore, CO2 sequestration and enhanced recovery (CO2-ECBM) technology demonstrates enormous potential, which also profoundly relies on a thorough understanding of the rock mechanics properties of coal-bearing strata. Predicted rock mechanics parameters provide crucial inputs for assessing caprock integrity, predicting reservoir mechanical stability under long-term injection pressure, optimizing CO2 injection pressure and rate, and preventing microseismic events or CO2 leakage risks.

[0003] Traditionally, key rock mechanics parameters of coal-bearing strata, such as elastic modulus, Poisson's ratio, and compressive strength, heavily rely on core experiments, which are costly and difficult to implement for continuous evaluation across the entire well section. Continuous prediction based on well logging data has thus become a crucial technical approach for achieving precise quantitative evaluation of reservoirs. However, constrained by the complexity of coalbed methane geological conditions and the ambiguity of well logging responses, traditional empirical formulas and polynomial regression methods generally suffer from limitations such as cumbersome characteristic engineering, weak ability to characterize nonlinear relationships, and insufficient model generalization. Therefore, developing a rock mechanics parameter prediction technology that can deeply mine well logging information, possesses good geological adaptability, and achieves high prediction accuracy is not only crucial for the efficient development of conventional coalbed methane but also a key theoretical support and core link connecting and promoting the development of next-generation low-carbon geological technologies such as CO2-ECBM and carbon dioxide geological storage. Summary of the Invention

[0004] Technical Problem: The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for predicting rock mechanics parameters based on an improved convolutional neural network. By constructing a neural network model with asymmetric convolutional kernels, it can adapt to the spatial structure of well logging data. When predicting parameters at a specific depth point, it can effectively integrate the contextual information of adjacent well logging points above and below, thereby accurately establishing a complex nonlinear mapping between well logging data and rock mechanics parameters.

[0005] Technical solution: The present invention provides a method for predicting rock mechanical parameters based on an improved convolutional neural network, comprising the following steps: (a) Data preparation: Collect logging curves and core logging data from core boreholes in the target mining area, and obtain rock mechanical parameters of different lithological samples through experimental testing, and establish an initial data table containing logging data and corresponding rock mechanical parameters; (b) Data preprocessing: The logging curves are cleaned and normalized to eliminate noise, dimensional and scale differences, and the processed data from all boreholes are merged to construct the initial dataset; (c) Feature selection: Correlation analysis is performed on the logging curves and rock mechanics parameters in the initial dataset. The logging curves that are sensitive to the response of rock mechanics parameters are selected as the feature curves for rock mechanics parameter prediction. The logging curves other than the feature curves in the initial dataset are removed to form the training dataset. (d) Sample construction: From the training dataset, read the feature curve data of multiple adjacent logging depth points with the same lithology and continuous sampling depth, combine them into the logging feature matrix of the training sample, and use the rock mechanics parameters corresponding to the center depth point of the training sample as the label of the training sample. (e) Model construction: A rock mechanics parameter prediction model is constructed using a convolutional neural network with asymmetric convolutional kernels. The rock mechanics parameter prediction model includes at least sequentially connected convolutional layers, pooling layers, and fully connected layers. The convolutional layers are configured to use asymmetric convolutional kernels of different sizes, with the well logging feature matrix as input. (f) Model optimization: The Adam optimizer is used to optimize the configuration of the rock mechanics parameter prediction model, and the ReduceLROnPlateau learning rate scheduler and early stopping mechanism are combined to optimize the training process and prevent overfitting. (g) Model training and evaluation: The optimal combination of hyperparameters is found in the preset hyperparameter space by using the grid search method, and the performance of the rock mechanics parameter prediction model is evaluated by combining K-fold cross-validation to determine the final trained rock mechanics parameter prediction model. (h) Parameter prediction: Perform data preprocessing steps on the logging curves of the core borehole to be predicted, construct the sample to be predicted, input it into the trained rock mechanics parameter prediction model, and output the corresponding rock mechanics parameter prediction results.

[0006] In step (a), the rock mechanical parameters include elastic modulus, Poisson's ratio and compressive strength. Based on the core logging data, rock mechanical parameter values ​​are assigned to different logging depth points in the logging curve data to obtain the initial data for each core borehole. The initial data includes the logging sampling depth and the corresponding logging curve data and rock mechanical parameters.

[0007] In step (b), data cleaning: First, missing values ​​and outliers in the logging curves are removed, and the average logging data of nearby depth points is used to fill them; then, wavelet transform is used to denoise the filled logging curves, specifically including: using the sym8 wavelet basis for 5-level decomposition, and using the heursure adaptive threshold rule and semi-soft threshold function to process high-frequency coefficients; finally, the denoised logging curves are obtained through wavelet reconstruction. Normalization is a process of normalizing the logging curves of each core borehole separately: normalization is a maximum-minimum normalization, which scales the data to the interval [0, 1] by subtracting the minimum value from each logging curve and dividing by the difference between the maximum and minimum values, thus eliminating the dimensional and scale differences between different logging curves; finally, the initial data tables of each core borehole after logging curve normalization are merged into an initial dataset.

[0008] In step (c), correlation analysis is performed based on data from all well logging sampling depths. Kendall's coefficient (τ) and mutual information (MI) are used to evaluate the overall correlation between each type of well logging curve and rock mechanics parameters. A higher correlation indicates greater sensitivity to rock mechanics parameters. Based on the analysis results, the five well logging curve types with the highest overall correlation to rock mechanics parameters are selected as fixed characteristic curve combinations. These five types are: density logging (DEN), sonic transit time logging (AC), caliper logging (CAL), neutron logging (CNL), and natural gamma logging (GR). All other well logging curves besides these five types are removed from the initial dataset to form the training dataset.

[0009] In step (d), in the training dataset, each well logging sampling depth point is listed in a row, totaling [number missing]. n Each row contains logging data for 5 characteristic curves; Well logging feature matrix construction method: From the training dataset, sequentially from row 3 to row 4... Each row is considered a candidate center point, and the candidate center point is evaluated. The rock mechanics parameter values ​​are consistent with those of the two preceding and two following adjacent sampling depths, and these sampling depths are continuous; if the condition is met, then the values ​​are read from the first sampling depth. Arrive at the All characteristic curve data of the row are used to form a 5x5 matrix as a training sample well logging feature matrix, and candidate center points are selected. The rock mechanics parameters are used as labels for the training samples.

[0010] In step (e), the well logging feature matrix of the training samples is input into the convolutional layer of the rock mechanics parameter prediction model. The convolutional layer includes a 3×1 asymmetric convolutional kernel and a 3×2 asymmetric convolutional kernel. For the 3×1 asymmetric convolutional kernel, the stride is set to 1. For the 3×2 asymmetric convolutional kernel, the padding is set to (0,1) and the stride is set to 1. The 3×1 asymmetric convolutional kernel in the convolutional layer performs convolution calculation on the well logging feature matrix of each training sample to obtain a 3×5 feature map, and the 3×2 asymmetric convolutional kernel performs convolution calculation on each input well logging feature matrix to obtain a 3×6 feature map. The pooling layer is set to max pooling, using 2×1... A pooling kernel with a stride of 1 reduces the dimensionality of a 3×5 feature map to a 2×5 pooling feature map. A pooling kernel with a stride of 2×2 and a stride of 2×1 reduces the dimensionality of a 3×6 feature map to a 2×3 pooling feature map. The 64 2×5 pooling feature maps are flattened into a 640-length one-dimensional feature vector, and the 64 2×3 pooling feature maps are flattened into a 384-length one-dimensional feature vector. The 640-length and 384-length one-dimensional feature vectors are concatenated and input into a fully connected layer with 64 neurons. The fully connected layer connects forward to the concatenated one-dimensional feature vector and backward to the outputs of three rock mechanics parameters. Through relevant calculations, a mapping relationship between the one-dimensional feature vector and the outputs of the rock mechanics parameters is established. The 3×5 and 3×6 feature maps are activated by ReLU before being input to the pooling layer, and then regularized by Dropout with a Dropout ratio of 0.25. The concatenated one-dimensional feature vector is activated by ReLU before being input to the fully connected layer (8), and then regularized by Dropout with a Dropout ratio of 0.25.

[0011] In step (f), the parameters of the rock mechanics parameter prediction model are optimized using the Adam optimizer. The initial learning rate of the Adam optimizer is set to 0.001. The parameters of the ReduceLROnPlateau learning rate scheduler are configured as follows: mode is set to 'max', indicating that the monitoring index is maximized, factor is set to 0.5, and patience is set to 10. The early stopping mechanism sets the patience value to 20.

[0012] In step (g), a grid search method is used to find the optimal hyperparameter combination of the rock mechanics parameter prediction model. Before obtaining the optimal hyperparameters, the value range of each hyperparameter that needs to be tuned in the hyperparameter space is first defined. The hyperparameter space includes at least the number of convolutional kernels, the number of neurons in the fully connected layer, and the learning rate of the prediction model. The hyperparameters are combined in sequence and the rock mechanics parameter prediction model is trained. The trained rock mechanics parameter prediction model is evaluated, and finally the optimal hyperparameter combination of the rock mechanics parameter prediction model is returned. K-fold cross-validation was used to evaluate the performance of a rock mechanics parameter prediction model. The training dataset contained logging data from K wells. Each time, samples from K-1 wells were extracted from the training dataset as the current training data. One well was selected from the training dataset to test the rock mechanics parameter prediction model's performance. From the remaining wells, 80% of the samples were extracted using stratified sampling to form the training set. The SMOTE algorithm was used to balance the training set before it was used to train the rock mechanics parameter prediction model. The remaining 20% ​​of the training data was used as the validation set to validate the rock mechanics parameter prediction model. This process was repeated K times, each time using a different well, until data from all K wells were used as the test set once. The mean squared error (MSE) and coefficient of determination (R²) of the rock mechanics parameter prediction model's prediction results on the test set samples were then calculated. 2 The average value is used as the evaluation result of the prediction model for rock mechanical parameters under the hyperparameter combination; the evaluation result with the minimum mean square error (MSE) and the coefficient of determination (R²) is selected. 2 The hyperparameter combination that reaches its maximum value is taken as the optimal model hyperparameter combination, and the corresponding model is the final rock mechanics parameter prediction model. The number of training iterations for the rock mechanics parameter prediction model was set to 500. The parameters of the convolutional neural network rock mechanics parameter prediction model were updated using mini-batch stochastic gradient descent and backpropagation. The mini-batch size was set to 128 during training. The weights and biases of the convolutional neural network model were updated in each iteration, which gradually reduced the prediction error of the rock mechanics parameter prediction model, and finally the trained convolutional neural network rock mechanics parameter prediction model was obtained.

[0013] In step (h), the characteristic curve of the core borehole to be predicted is processed by the data preprocessing in step (b) to obtain the dataset to be predicted. A continuous five-line characteristic curve data is read from the dataset to be predicted to form a sample to be predicted. The sample to be predicted is input into the trained rock mechanics parameter prediction model. The rock mechanics parameter prediction result is the predicted value of the rock mechanics parameter at the depth point corresponding to the middle row of the five-line characteristic curve data.

[0014] Beneficial Effects: By adopting the above technical solution, this invention overcomes the shortcomings of traditional well logging prediction techniques for rock mechanics parameters, such as low prediction efficiency, insufficient utilization of well logging curve information, and reliance on complex feature engineering. By constructing a neural network model with an asymmetric convolutional kernel, it can adapt to the spatial structure of well logging data. When predicting parameters at a specific depth point, it can effectively integrate the contextual information of adjacent well logging points, thereby accurately establishing the complex nonlinear mapping between well logging data and rock mechanics parameters. Its core lies in constructing a convolutional neural network model based on an asymmetric convolutional kernel, effectively capturing the local spatial features of well logging curve data and the correlation of its neighboring points, and then deeply mining the complex nonlinear mapping relationship between it and rock mechanics parameters. It aims to overcome the limitations of traditional methods and achieve rapid, efficient, and highly generalizable intelligent prediction of rock mechanics parameters under complex geological conditions. First, well logging curve preprocessing removes outliers and noise to ensure data validity; then, correlation analysis is used to screen sensitive curves, enabling the model to focus on key features. Secondly, by constructing a feature matrix from multiple consecutive rows of data, contextual information from neighboring points is introduced into single-point prediction, thereby achieving accurate identification of the rock mechanical parameters of coal-bearing strata. Using 3×1 and 3×2 asymmetric convolution kernels, differentiated feature extraction is achieved in both vertical and horizontal dimensions: the vertical dimension adaptively adjusts the contribution weights of each depth point, while the horizontal dimension fully preserves the collaborative variation characteristics between multiple curves, perfectly utilizing the spatial structure of well logging data. Finally, by fully learning the spatial variation patterns of the data, the model establishes an accurate nonlinear mapping relationship, significantly improving prediction accuracy. The method is simple to operate, has good performance, and has wide applicability in this technical field. The main advantages of this invention compared to existing technologies are: ① It has achieved rapid, efficient and intelligent identification of rock mechanical parameters throughout the entire well section; ② It can automatically capture complex mapping relationships, reducing reliance on manual feature engineering; ③ It integrates multiple sensitivity curves and makes full use of their spatial structure characteristics; ④ When identifying a single point, it can effectively refer to the logging data of the adjacent points above and below, making the judgment more accurate; ⑤ The model can be trained and optimized to adapt to different geological backgrounds and has strong generalization ability. Attached Figure Description

[0015] Figure 1 This is a flowchart of the rock mechanics parameter prediction method based on an improved convolutional neural network according to the present invention.

[0016] Figure 2 This is a diagram of the improved convolutional neural network rock mechanics parameter prediction model of the present invention.

[0017] Figure 3 This is a diagram of the model verification method of the present invention.

[0018] In the figure: 1. Well logging feature matrix; 2. Convolutional layer; 3-1. 3×1 asymmetric convolution kernel; 3-2. 3×2 asymmetric convolution kernel; 4-1. 3×5 feature map; 4-2. 3×6 feature map; 5. Pooling layer; 6-1. 2×1 pooling kernel with a step size of 1; 6-2. 2×2 pooling kernel with a step size of 2×1; 7-1. 2×5 pooling feature map; 7-2. 2×3 pooling feature map; 8. Fully connected layer; 9-1. 640-length one-dimensional feature vector; 9-2. 384-length one-dimensional feature vector; 10. Rock mechanics parameter output. Detailed Implementation

[0019] An embodiment of the present invention will be further described below with reference to the accompanying drawings: like Figure 1 As shown, the rock mechanics parameter prediction method based on an improved convolutional neural network of the present invention preprocesses the logging data of cored boreholes, selects logging curves sensitive to rock mechanics parameter responses as feature curves for rock mechanics parameter prediction, and reads multiple rows of feature curve data with the same lithology and continuous sampling depth to form a sample logging feature vector 1. A convolutional neural network rock mechanics parameter prediction model based on asymmetric convolution kernels is established and optimized parameters are configured. The optimal rock mechanics parameter prediction model is obtained by training with training data. The rock mechanics parameter prediction result of the borehole to be predicted is obtained by preprocessing the logging data of the borehole to be predicted, constructing samples, and inputting them into the rock mechanics parameter prediction model. The specific steps are as follows: (a) Data preparation: Collect logging curves and core logging data from core drilling boreholes in the target mining area. Based on the core logging data, count the types of rocks that appear in the boreholes in the mining area. Take core samples of various types of rocks for rock mechanics parameter testing. Obtain three rock mechanics parameters: elastic modulus, Poisson's ratio, and compressive strength of core samples of various lithologies. Assign rock mechanics parameter values ​​to different logging depths based on the vertical variation information of lithology in the borehole recorded in the core logging data. Supplement the borehole logging curve data to obtain the initial data for each core drilling borehole. The initial data includes the logging sampling depth and the corresponding logging curve data and rock mechanics parameters, thereby establishing an initial data table containing logging data and corresponding rock mechanics parameters. (b) Data preprocessing: The logging curves in the initial data table are cleaned and normalized to eliminate noise, dimensional and scale differences, and all processed data from boreholes are merged to construct the initial dataset. The data cleaning process is as follows: First, null and outlier values ​​in the logging curves are removed, that is, individual null or outlier values ​​in the logging curve data are removed and filled with the mean of logging data from nearby depth points. Then, wavelet transform is used to denoise the filled logging curves. That is, if long segments of null or outlier values ​​are encountered, the null or outlier segments are directly removed. There are also some high-frequency noise data in the logging curves. In order to eliminate the high-frequency noise in the logging curves, the logging curve data is decomposed into 5 levels using the sym8 wavelet basis. Each level of decomposition transforms the logging curve into a low-frequency approximate curve and a high-frequency coefficient curve. After processing the decomposed high-frequency coefficient curves using the heursure adaptive threshold rule and semi-soft threshold function, some noise in the high-frequency coefficients is filtered out. Then, wavelet reconstruction is performed to merge the low-frequency approximate curve and the processed high-frequency coefficient curve to obtain the denoised logging curves. Because different logging curves have different dimensions, the scale of the same logging curve may also differ between different wells. To eliminate the differences in dimensions and scale between logging curves, normalization is performed on the logging curves. Normalization is performed separately for each core borehole, using maximum-minimum normalization. Each logging curve is scaled to the range [0, 1] by subtracting the minimum logging value and then dividing by the difference between the maximum and minimum logging values, as shown in Equation 1 below: Equation (1) In the formula: x is the original well logging curve data, These are normalized well logging curve data. It is the maximum value of the well logging curve data. It is the minimum value of the well logging curve data; Finally, the data tables of each core borehole after logging curve normalization were merged into the initial dataset.

[0020] (c) Feature Selection: Correlation analysis was performed on the logging curves and rock mechanics parameters in the initial dataset. Five logging curves that were sensitive to the rock mechanics parameters were selected as feature curves for rock mechanics parameter prediction. Logging curves other than the feature curves in the initial dataset were removed to form the training dataset. The correlation analysis was based on data from all logging sampling depth points, using Kendall's τ and mutual information (MI) to evaluate the correlation between each type of logging curve and the rock mechanics parameters. The sign of Kendall's τ indicates whether the logging curve and the rock mechanics parameters are positively or negatively correlated. The larger the absolute value of τ, the higher the sensitivity of the logging curve to changes in rock mechanics parameters. The mutual information (MI) value represents the richness of rock mechanics parameter information contained in the logging curve. The larger the MI value, the richer the amount of coal body structure information contained in the logging curve. Based on the analysis results, five logging curves that are sensitive to rock mechanics parameter responses are selected as a fixed combination of feature curves for rock mechanics parameter prediction. The five logging curve types include density logging (DEN), sonic transit time logging (AC), caliper logging (CAL), neutron logging (CNL), and natural gamma logging (GR). All logging curves other than the five types are removed from the initial dataset to form the training dataset.

[0021] (d) Sample Construction: From the training dataset, feature curve data from multiple adjacent logging depth points with the same lithology and continuous sampling depth are read and combined to form a logging feature matrix 1 for a training sample. The rock mechanics parameters corresponding to the center depth point of the logging feature matrix are used as the labels of the training samples. The training dataset is obtained through well logging. The sampling method is as follows: formation data are collected at fixed intervals of 0.05m in the vertical direction of drilling. The logging values ​​of 5 feature curves DEN, AC, CAL, CNL, and GR are recorded at each sampling depth point. The training dataset is sorted by depth. The row corresponds to a depth of m( i The training dataset contains (i = 1, 2, ..., n) and contains a total of (i = 1, 2, ..., n). n Line logging data; the reading process starts from the 3rd row (depth = 0.10m) of the training dataset and continues to the next row. n -2 rows (depth = (n-2) × 0.05m), each row is used as a candidate center point, and the candidate center point is judged. The rock mechanics parameter values ​​are consistent with those of the two preceding and two following adjacent sampling depths, and these sampling depths are continuous; if the condition is met, then the values ​​are read from the first sampling depth. Arrive at the All characteristic curve data of the row are used to form a 5x5 matrix as a training sample well logging feature matrix 1, and candidate center points are selected. The rock mechanics parameters are used as labels for the training samples.

[0022] use Representative center point The five characteristic curve data, , Representative center point Given the rock mechanics parameters, determine the following conditions: ① The center point The label is the same as the labels of the two points before it and the two points after it, that is... This ensures that the data in each logging feature matrix 1 originates from logging data of the same lithology; ② The sampling depth of these five lines of well logging data is continuous. In the experiment, the continuity was indexed. Indirect judgment ensures that the data of each logging feature matrix 1 comes from the same rock formation, avoiding the mixing of logging data from different rock formations; If the first If a row satisfies both of the above conditions, then read from the first row. Arrive at the The characteristic curve data of the row constitutes the first row. Well logging feature matrix 1 for each sample Its size is 5×5, and the center point is... tags Assign this sample As shown in equation (2) below; the final total obtained 1 valid sample ; Equation (2) (e) Model Construction: A rock mechanics parameter prediction model is constructed using a convolutional neural network with asymmetric kernels, such as... Figure 2As shown, the rock mechanics parameter prediction model includes at least a convolutional layer 2, a pooling layer 5, and a fully connected layer 8. Convolutional layer 2 is configured to use asymmetric convolutional kernels of different sizes to receive and process the logging feature matrix 1 in single-channel image form. The logging feature matrix 1 of the training samples is input into the convolutional layer 2 of the rock mechanics parameter prediction model. Convolutional layer 2 includes a 3×1 asymmetric convolutional kernel 3-1 and a 3×2 asymmetric convolutional kernel 3-2. For the 3×1 asymmetric convolutional kernel 3-1, the stride is set to 1. For the 3×2 asymmetric convolutional kernel 3-2, the padding is set to 0 or 1, and the stride is set to 1. The 3×1 asymmetric convolutional kernel 3-1 in convolutional layer 2 performs convolution calculations on the logging feature matrix 1 of each training sample to obtain a 3×5 feature map 4-1. The 3×2 asymmetric convolutional kernel 3-2 performs convolution calculations on each input logging feature matrix 1 to obtain... The feature map 4-2 is reduced to a size of 3×6. Pooling layer 5 is set to max pooling. A pooling kernel 6-1 with a size of 2×1 and a stride of 1 is used to reduce the size of the feature map 4-1 to a size of 2×5. A pooling kernel 6-2 with a size of 2×2 and a stride of 2×1 is used to reduce the size of the feature map 4-2 to a size of 2×3. The 64 pooling feature maps 7-1 of size 2×5 are flattened into a one-dimensional feature vector 9-1 of length 640. The 64 pooling feature maps 7-2 of size 2×3 are flattened into a one-dimensional feature vector 9-2 of length 384. The one-dimensional feature vector of length 640 and the one-dimensional feature vector of length 384 are concatenated and input into a fully connected layer 8 containing 64 neurons. The fully connected layer 8 is connected forward to the concatenated one-dimensional feature vector and backward to the three rock mechanics parameters 10 output. The mapping relationship between the one-dimensional feature vector and the rock mechanics parameter output 10 is established through relevant calculations.

[0023] The rock mechanics parameter prediction model's convolutional layer 2 accepts a 5×5 single-channel image as input and processes a 5×5 well logging feature matrix 1. Convolutional layer 2 contains 64 3×1 asymmetric convolutional kernels 3-1 and 64 3×2 asymmetric convolutional kernels 3-2. For the 3×2 asymmetric convolutional kernels 3-2, the well logging feature matrix 1 is padded with "0"s on both sides before convolution to ensure that each column value of the well logging feature matrix 1 participates in the calculation consistently during convolution. The sliding step size for both types of convolutional kernels is 1. This design aims to predict the depth center point. When determining the rock mechanical parameters, the main basis is the center point of the depth measurement. The logging values ​​are effectively integrated with the logging values ​​of adjacent points above and below. , , , The well logging information; the feature extraction mechanism of the rock mechanical parameter prediction model is as follows: For the input feature matrix When the convolution kernel scans horizontally, each data element participates in the operation the same number of times; during vertical movement, the logging data in the middle row of the matrix... Each data element participates in three convolution calculations, and the logging values ​​of adjacent rows are... , Each participated twice, and the logging values ​​of the next adjacent row were... , Each participates once; this design adaptively adjusts the contribution weight of different rows of logging values ​​to feature extraction by the number of times the convolution kernel covers the data. That is, the influence of logging values ​​farther away from the center point is gradually weakened, thereby better matching the spatial characteristics of the logging feature matrix 1 data.

[0024] In convolutional layer 2, a 3×1 asymmetric convolution kernel 3-1 performs convolution calculation on the well logging feature matrix 1 of each input sample, resulting in 64 3×5 feature maps 4-1. A 3×2 asymmetric convolution kernel 3-2 performs convolution calculation on the well logging feature matrix 1 of each input sample, resulting in 64 3×6 feature maps 4-2, providing multi-level feature representations for the rock mechanical parameter prediction model.

[0025] Pooling layer 5 is set to max pooling. A 2×1 pooling kernel 6-1 with a stride of 1 is used to reduce the dimensionality of the 3×5 feature map 4-1 to a 2×5 pooled feature map 7-1. A 2×2 pooling kernel 6-2 with a stride of 2×1 is used to reduce the dimensionality of the 3×6 feature map 4-2 to a 2×3 pooled feature map 7-2. These 64 2×5 pooled feature maps 7-1 are then flattened by a flattening layer into a one-dimensional feature vector 9-1 of length 640. The 2×3 pooled feature map 7-2 is flattened by a flattening layer and transformed into a 384-length one-dimensional feature vector 9-2. The 640-length one-dimensional feature vector 9-1 and the 384-length one-dimensional feature vector 9-2 are concatenated and input into a fully connected layer 8 containing 64 neurons. The fully connected layer 8 is connected forward to the concatenated one-dimensional feature vector and backward to the three rock mechanics parameter outputs 10. The mapping relationship between the one-dimensional feature vector and the rock mechanics parameter outputs 10 is established through relevant calculations. To improve the nonlinear fitting ability of the rock mechanics parameter prediction model, a ReLU activation function was added to the convolutional layer 2 of the rock mechanics parameter prediction model, along with an L2 regularization and Dropout layer. The L2 regularization value was 0.001, and the Dropout ratio was 0.25. A ReLU activation function and a Dropout layer were added after the fully connected layer 8, with a Dropout ratio of 0.25.

[0026] (f) Model Optimization: The Adam optimizer was configured for the convolutional neural network rock mechanics parameter prediction model, combined with the ReduceLROnPlateau learning rate scheduler and an early stopping mechanism to optimize the training process and prevent overfitting. The initial learning rate of the Adam optimizer was set to 0.001. The parameters of the ReduceLROnPlateau learning rate scheduler were configured as follows: mode='max', factor=0.5, patience=10; the patience value of the early stopping mechanism was set to 20. To improve the training efficiency and final performance of the rock mechanics parameter prediction model, the learning rate of the rock mechanics parameter prediction model was set to 0.001. The ReduceLROnPlateau learning rate scheduler was added to the rock mechanics parameter prediction model to optimize the training process. The monitoring mode of the ReduceLROnPlateau learning rate scheduler was set to "max", the decay factor was set to 0.5, and the patience value was set to 10 to achieve dynamic adjustment of the learning rate. At the same time, an early stopping mechanism was used to prevent overfitting of the rock mechanics parameter prediction model, and the patience value of the early stopping mechanism was set to 20.

[0027] (g) Model Training and Evaluation: The optimal hyperparameter combination is found in the pre-defined hyperparameter space using a grid search method to determine the final optimal rock mechanics parameter prediction model. The performance of the rock mechanics parameter prediction model is evaluated using K-fold cross-validation to determine the final trained rock mechanics parameter prediction model. Before obtaining the optimal hyperparameters, the value range of each hyperparameter that needs to be tuned is first defined. The hyperparameter space includes at least the number of convolutional kernels, the number of neurons in the fully connected layer, and the learning rate. The hyperparameters are combined sequentially, the rock mechanics parameter prediction model is trained, and the trained rock mechanics parameter prediction model is evaluated. Finally, the optimal hyperparameter combination of the rock mechanics parameter prediction model is returned. K-fold cross-validation was used to evaluate the performance of rock mechanics parameter prediction models. The K-fold cross-validation method is as follows: Figure 3 As shown, K represents the number of core boreholes collected for geological exploration in the target mining area. The training dataset contains logging data from K wells. In each experiment, samples from K-1 wells are extracted from the training dataset as training data. 80% of the samples from the training dataset are selected using stratified sampling to form the training set. After equalization using the SMOTE algorithm, these samples are used to train the rock mechanics parameter prediction model. The remaining 20% ​​of the training data is used as the validation set to verify the rock mechanics parameter prediction model. The remaining sample from the last well is used as the test set to test the performance of the rock mechanics parameter prediction model. The experiment is repeated K times, with no repetition of test wells used in each experiment, until the test set has traversed all well data. The mean squared error (MSE) and coefficient of determination (R²) of the rock mechanics parameter prediction model's prediction results on the test set samples are then taken from the K experiments.2 The average value is used as the evaluation result of the rock mechanics parameter prediction model under the hyperparameter combination; the evaluation result with the minimum mean square error (MSE) and the coefficient of determination (R²) is selected. 2 The hyperparameter combination at its maximum value is taken as the optimal hyperparameter combination for predicting rock mechanics parameters, and the corresponding rock mechanics parameter prediction model is the final rock mechanics parameter prediction model.

[0028] A rock mechanics parameter prediction model was trained with 500 training iterations. The parameters of the convolutional neural network (CNN) rock mechanics parameter prediction model were updated using mini-batch stochastic gradient descent and backpropagation. The mini-batch size was set to 128 samples. In each iteration, the weights and biases of the CNN model were updated, gradually reducing the prediction error. The final trained CNN rock mechanics parameter prediction model was obtained. The validation set was used to monitor the prediction performance after each iteration, while the test set was used to test the actual prediction performance of the model after training.

[0029] (h) Parameter prediction: Perform data preprocessing steps on the logging curves of the core boreholes to be predicted, construct the samples to be predicted, input them into the trained rock mechanics parameter prediction model, and output the corresponding rock mechanics parameter prediction results. After the characteristic curves of the core borehole to be predicted are preprocessed in step (b), a dataset to be predicted is obtained. Five consecutive rows of characteristic curve data are read from this dataset to form a sample to be predicted. This sample is then input into the trained rock mechanics parameter prediction model. The predicted rock mechanics parameters are the predicted values ​​of the rock mechanics parameters at the depth points corresponding to the middle row of the five rows of characteristic curve data. Specific steps are as follows: 1) Before determining the rock mechanics parameters of the core borehole to be predicted, the characteristic curves of the borehole to be predicted are first subjected to step (b) data cleaning and normalization to obtain the dataset to be predicted. ,in For the first ( Five characteristic curve data points at each sampling depth; 2) Construct the samples to be predicted using the sliding window method: Window size: Fixed at 5 consecutive depth points, meaning each read... to ; Sliding step size: 1 depth point, meaning the next window starts from... to ; Sample composition: Each sample contains feature data from 5 depth points, predicting the intermediate depth point. Rock mechanical parameters; 3) Forecast range: First valid sample: window ,predict ; Last valid sample: window ,predict ; Finally obtained to common Prediction results for each depth point.

Claims

1. A rock mechanics parameter prediction method based on an improved convolutional neural network, characterized by The method comprises the following steps: (a) Data preparation: collecting the logging curves and core logging data of the core drilling holes in the target mining area, and obtaining the rock mechanics parameters of different lithology samples through experimental testing, to establish an initial data table containing the logging data and corresponding rock mechanics parameters; (b) Data preprocessing: performing data cleaning and normalization processing on the logging curves to eliminate noise, dimension and scale differences, and merging the processed data of all drilling holes to construct an initial data set; (c) Feature selection: performing correlation analysis on the logging curves and rock mechanics parameters in the initial data set, selecting the logging curves sensitive to the response of the rock mechanics parameters as the feature curves for predicting the rock mechanics parameters, and removing the logging curves other than the feature curves from the initial data set to form a training data set; (d) Sample construction: reading the feature curve data of multiple adjacent logging depth points with the same lithology and continuous sampling depth from the training data set, combining to form a logging feature matrix (1) of the training sample, and taking the rock mechanics parameters corresponding to the center depth point of the training sample as the label of the training sample; (e) Model construction: constructing a rock mechanics parameter prediction model using a convolutional neural network with asymmetric convolution kernels, the rock mechanics parameter prediction model at least comprising sequentially connected convolution layers (2), pooling layers (5) and fully connected layers (8); the convolution layers (2) are configured to use asymmetric convolution kernels of different sizes, with the logging feature matrix (1) as the input; (f) Model optimization: optimizing the rock mechanics parameter prediction model configuration using the Adam optimizer, and combining the ReduceLROnPlateau learning rate scheduler and the early stopping mechanism to optimize the training process and prevent overfitting; (g) Model training and evaluation: finding the optimal hyperparameter combination in the preset hyperparameter space using the grid search method, and evaluating the performance of the rock mechanics parameter prediction model combined with K-fold cross-validation to determine the final trained rock mechanics parameter prediction model; (h) Parameter prediction: performing the data preprocessing step on the logging curves of the core drilling hole to be predicted, constructing a to-be-predicted sample, inputting it into the trained rock mechanics parameter prediction model, and outputting the corresponding rock mechanics parameter prediction result.

2. The rock mechanics parameter prediction method based on the improved convolutional neural network according to claim 1, characterized in that: In step (a), the rock mechanics parameters include elastic modulus, Poisson's ratio and compressive strength. The rock mechanics parameter values are assigned to different logging depth points of the logging curve data according to the core logging data, to obtain the initial data of each core drilling hole. The initial data includes the logging sampling depth, the corresponding logging curve data and the rock mechanics parameters. 3.The rock mechanics parameter prediction method based on the improved convolutional neural network according to claim 1, characterized in that: In step (b), data cleaning: first, remove the null values and outliers in the logging curve, and fill in the mean value of the logging data of the adjacent depth points; then, use wavelet transform to denoise the filled logging curve, which includes: using sym8 wavelet basis for 5-layer decomposition, and using heursure adaptive threshold rule and semi-soft threshold function to process high-frequency coefficients, and finally obtaining the denoised logging curve through wavelet reconstruction. The normalization processing is to normalize the logging curves of each coring borehole respectively: the normalization is the maximum-minimum normalization, each logging curve is subtracted by the minimum value and divided by the difference between the maximum value and the minimum value, the data is scaled to the interval of [0, 1], and the dimensional and scale differences between different logging curves are eliminated; finally, the initial data tables of each coring borehole after the logging curve normalization processing are merged into an initial data set. 4.The rock mechanics parameter prediction method based on the improved convolutional neural network according to claim 1, characterized in that: In step (c), the correlation analysis is based on the data of all logging sampling depth points, uses Kendall's τ and mutual information MI two indexes to evaluate the overall correlation between each type of logging curve and the rock mechanics parameters, and the higher the correlation between the logging curve and the rock mechanics parameters, the more sensitive it is to the rock mechanics parameters; according to the analysis result, the five types of logging curves with the highest overall correlation with the rock mechanics parameters are selected as the fixed feature curve combination, and the five types of logging curves are: density logging curve DEN, acoustic time difference logging curve AC, caliper logging CAL, neutron logging CNL and natural gamma logging GR; all other logging curves except the five types of logging curves are removed from the initial data set to form a training data set.

5. The rock mechanics parameter prediction method based on the improved convolutional neural network according to claim 4, characterized in that: In step (d), in the training data set, each well logging sampling depth point is listed as a row, and there are n rows in total, each of which contains logging data of 5 characteristic curves; The well logging feature matrix (1) is constructed in the following manner: from the training data set, each row is taken as a candidate center point in turn from the third row to the last row, and it is determined whether the rock mechanics parameter value of the candidate center point is consistent with the rock mechanics parameter values of the two adjacent sampling depth points before and after the candidate center point and whether the sampling depth points are continuous. The well logging feature matrix (1) is constructed in the following manner: from the training data set, each row is taken as a candidate center point in turn from the third row to the last row, and it is determined whether the rock mechanics parameter value of the candidate center point is consistent with the rock mechanics parameter values of the two adjacent sampling depth points before and after the candidate center point and whether the sampling depth points are continuous.​ If the condition is met, all the characteristic curve data from the first row to the first row are read, a 5-row 5-column matrix is formed as a well logging characteristic matrix (1) of a training sample, and the rock mechanics parameters of the candidate center point are taken as labels of the training sample.

6. The rock mechanics parameter prediction method based on the improved convolutional neural network according to claim 5, characterized in that: In step (e), the well logging feature matrix (1) of the training sample is input into the convolution layer (2) of the rock mechanics parameter prediction model, and the convolution layer (2) includes an asymmetric convolution kernel (3-1) with a size of 3x1 and an asymmetric convolution kernel (3-2) with a size of 3x2. For the asymmetric convolution kernel (3-1) with a size of 3x1, the sliding step stride is set to 1. For the asymmetric convolution kernel (3-2) with a size of 3x2, the padding is set to (0, 1), and the stride is set to 1. The asymmetric convolution kernel (3-1) with a size of 3x1 in the convolution layer (2) performs convolution calculation on the well logging feature matrix (1) of each training sample to obtain a feature map (4-1) with a size of 3x5. The asymmetric convolution kernel (3-2) with a size of 3x2 performs convolution calculation on each input well logging feature matrix (1) to obtain a feature map (4-2) with a size of 3x6. The pooling layer (5) is set to maximum pooling, and the pooling kernel (6-1) with a size of 2x1 and a step of 1 is used to reduce the dimension of the feature map (4-1) with a size of 3x5 to a pooling feature map (7-1) with a size of 2x5. The pooling kernel (6-2) with a size of 2x2 and a step of 2x1 is used to reduce the dimension of the feature map (4-2) with a size of 3x6 to a pooling feature map (7-2) with a size of 2x3. The 64 pooling feature maps (7-1) with a size of 2x5 are converted into a one-dimensional feature vector (9-1) with a length of 640 through flattening. The 64 pooling feature maps (7-2) with a size of 2x3 are converted into a one-dimensional feature vector (9-2) with a length of 384 through flattening. The one-dimensional feature vector with a length of 640 and the one-dimensional feature vector with a length of 384 are spliced and input into the fully connected layer (8) containing 64 neurons. The fully connected layer (8) connects the spliced one-dimensional feature vector in front and connects three rock mechanics parameters (10) in back to output, and the mapping relationship between the one-dimensional feature vector and the rock mechanics parameter output (10) is established through correlation calculation.

7. The rock mechanics parameter prediction method based on the improved convolutional neural network according to claim 6, characterized in that: The feature map (4-1) with a size of 3x5 and the feature map (4-2) with a size of 3x6 are activated using the ReLU activation function before being input into the pooling layer (5), and then regularized using Dropout with a ratio of 0.

25. The one-dimensional feature vector after splicing is activated using the ReLU activation function before being input into the fully connected layer (8), and then regularized using Dropout with a ratio of 0.

25. 8.The rock mechanics parameter prediction method based on the improved convolutional neural network according to claim 1, characterized in that: In step (f), the parameters of the rock mechanics parameter prediction model are optimized using the Adam optimizer, and the initial learning rate of the Adam optimizer is set to 0.

001. The parameter configuration of the ReduceLROnPlateau learning rate scheduler is as follows: the mode is set to'max', indicating that the maximum monitoring index is monitored, the factor is set to 0.5, and the patience is set to 10. The patience of the early stopping mechanism is set to 20. 9.The rock mechanics parameter prediction method based on the improved convolutional neural network according to claim 1, characterized in that: In step (g), the grid search method is used to find the optimal hyperparameter combination of the rock mechanics parameter prediction model. Before obtaining the optimal hyperparameters, the value range of each hyperparameter to be optimized in the hyperparameter space is first defined. The hyperparameter space at least includes: the number of convolution kernels, the number of fully connected layer neurons, and the learning rate of the prediction model. The hyperparameters are combined in turn, the rock mechanics parameter prediction model is trained, and the trained rock mechanics parameter prediction model is evaluated. Finally, the optimal hyperparameter combination of the rock mechanics parameter prediction model is returned. The performance of the rock mechanics parameter prediction model is evaluated by K-fold cross-validation. There are K well logging data in the training data set. In each training, K-1 well samples are extracted from the training data set as the training data. The performance of the rock mechanics parameter prediction model is tested by using one well sample from the training data set. The remaining well samples are extracted by stratified sampling method to obtain 80% training set. The SMOTE algorithm is used to balance the training set, and the balanced training set is used to train the rock mechanics parameter prediction model. The remaining 20% samples in the training data set are used as the validation set to verify the rock mechanics parameter prediction model. The K times are repeated, and different test wells are used each time until the data of K wells are used as the test set once. The mean square error (MSE) and the determination coefficient (R 2 The average value is taken as the evaluation result of the rock mechanics parameter prediction model under the hyperparameter combination. The evaluation result with the minimum mean square error (MSE) and the determination coefficient (R 2 The hyperparameter combination with the maximum value is taken as the optimal model hyperparameter combination, and the corresponding model is the final rock mechanics parameter prediction model. The rock mechanics parameter prediction model is trained for 500 iterations. The small batch random gradient descent method and the back propagation algorithm are used to update the parameters of the convolutional neural network rock mechanics parameter prediction model. The Mini-batch Size of the rock mechanics parameter prediction model is set to 128 during training. The weights and bias of the convolutional neural network model are updated every iteration to gradually reduce the prediction error of the rock mechanics parameter prediction model, and finally a trained convolutional neural network rock mechanics parameter prediction model is obtained.

10. The rock mechanics parameter prediction method based on the improved convolutional neural network according to claim 1, characterized in that: In step (h), the characteristic curve of the coring borehole to be predicted is preprocessed in step (b) to obtain a prediction data set. Five consecutive rows of characteristic curve data are read from the prediction data set to form a prediction sample. The prediction sample is input into the trained rock mechanics parameter prediction model. The rock mechanics parameter prediction result is the rock mechanics parameter prediction value of the middle row of the five rows of characteristic curve data.

Images