A rock failure prediction method and system

Abstract

The present application relates to a kind of rock failure prediction method and system, first utilize DIC technique in rock sample when rock compression test is carried out to collect the strain distribution image of rock sample;Strain distribution image is selected from the strain value between 0 and the strain distribution image of the strain distribution image of preset value and the strain distribution image of peak value destruction, and data set is divided into training set and test set;PreRNN++ prediction model is built, and the model is trained using training set, and the trained model is evaluated using test set, when the evaluation result meets preset requirement, the trained model is obtained, the image of peak value destruction is predicted using the trained model to be measured rock sample, and the axial deformation is calculated according to the predicted graph, and then the damage energy and damage time are estimated.The present application makes certain technical accumulation for subsequent prediction of major natural disasters, especially the prediction of surrounding rock instability and destruction.

Description

Technical Field

[0001] This application relates to the field of rock failure prediction, and in particular to a method and system for predicting rock failure. Background Technology

[0002] Over the past decade, with the continuous and rapid advancement of infrastructure construction nationwide, China's civil engineering has developed extremely quickly. With the large-scale construction of bridges and tunnels across the country, tunnels often need to traverse high-stress rock strata or fault fracture zones, and bridges frequently need to be erected on steep rocky mountains. Ensuring the stability and safety of structures throughout their entire life cycle has always been a key concern for researchers and engineers. Among these concerns, due to the unique properties of rock, sudden natural disasters often become a major factor restricting the safety of structures. Therefore, the instability and failure of rock and rock mass systems has always been a core issue in civil engineering research, and the prediction of rock mass system failure is of great significance and value for human disaster prevention and mitigation.

[0003] Since rock sample cracks continue to develop and eventually form the main fracture surface, in order to advance the research on the prediction and forecasting of rock mass systems, it is necessary to predict the location of the main fracture surface of the rock sample and the time and energy of failure to reach the peak strength, so as to make certain technical accumulation for the prediction of major natural disasters, especially the prediction of the instability and failure of the surrounding rock.

[0004] However, traditional rock sample failure detection mainly relies on acoustic emission (AE) testing. Acoustic emission (AE) is the phenomenon of transient stress waves released during the deformation and failure of rocks as primary defects and newly formed micro-fractures propagate, evolve, and fracture. AE technology can monitor the internal damage evolution and defects of rock samples in real time during deformation and failure. The test process begins by attaching sensors to the surface of the rock sample. Under the pressure of the testing machine, micro-cracks appear inside the rock sample, and rock particles release sound waves during the gradual fracture process due to the Kaiser effect. By using highly sensitive AE sensors, the sound waves generated by internal rock damage can be detected and collected. After initial noise reduction, signal analysis and other methods can be used to roughly infer the main process of rock failure. The advantage of this method is its high sensitivity to propagating defects. However, its disadvantages are also obvious: AE can only detect concentrated bursts of energy within the rock (i.e., the occurrence of crack propagation and particle fracture at a certain moment), and it cannot predict the crack propagation direction, failure time, or failure energy. Furthermore, AE testing is relatively cumbersome and inefficient due to the need to attach and adjust the sensors. Therefore, there is an urgent need for a rock failure prediction method to efficiently and accurately predict the location of the main fracture surface, axial deformation, failure time, and failure energy. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting rock failure, which can efficiently and accurately predict the location of the main fracture surface, axial deformation, failure time and failure energy of rock failure, thereby making certain technical contributions to the prediction of subsequent major natural disasters.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for predicting rock failure includes: Throughout the rock compression test on the rock sample, the strain distribution map on the surface of the rock sample was acquired using DIC technology. An image of the main fracture surface of the rock sample is selected from the strain distribution map and recorded as the first image; A dataset is selected from the first image and divided into a training set and a test set; the dataset includes a strain distribution map with strain values ​​between 0 and a preset value, and a strain distribution map at peak failure. A prediction model based on PreRNN++ is constructed, and the strain distribution map of strain values ​​between 0 and preset values ​​in the training set is used as input. The strain distribution map at peak failure in the training set is used as label to train the prediction model to obtain the first model. The first model is evaluated using the test set. If the evaluation result does not meet the preset requirements, the hyperparameters of the prediction model are adjusted, and the process returns to the step "and the prediction model is trained using the training set" until the evaluation result meets the preset requirements, thus obtaining the second model. The strain distribution map at the peak failure of the rock sample to be predicted is predicted using the second model, and a predicted image is obtained. Based on the positional changes of pixels in the predicted image along the pressure direction, the axial deformation of the rock sample to be predicted at the time of failure is calculated. Based on the axial deformation, the failure energy and failure time are predicted.

[0007] The present invention also provides a rock failure prediction system, comprising: The DIC acquisition module is used to acquire the strain distribution map on the surface of the rock sample using DIC technology throughout the entire process of rock compression testing. The main fracture surface selection module is used to select an image of the main fracture surface of the rock sample from the strain distribution map, which is denoted as the first image. The dataset acquisition module is used to select a dataset from the first image and divide the dataset into a training set and a test set; the dataset includes a strain distribution map with strain values ​​between 0 and a preset value and a strain distribution map at peak failure; The model training module is used to build a prediction model based on PreRNN++. It takes the strain distribution map of strain values ​​between 0 and preset values ​​in the training set as input and the strain distribution map at peak failure in the training set as labels to train the prediction model and obtain the first model. The model evaluation module is used to evaluate the first model using the test set. If the evaluation result does not meet the preset requirements, the hyperparameters of the prediction model are adjusted, and the process returns to the step "and train the prediction model using the training set" until the evaluation result meets the preset requirements, thus obtaining the second model. The prediction image acquisition module is used to predict the strain distribution map at the peak failure of the rock sample to be predicted using the second model, and obtain the prediction image. The axial deformation prediction module is used to calculate the axial deformation of the rock sample to be predicted when it fails, based on the positional changes of the pixels in the predicted image in the pressure direction. A destruction energy and destruction time prediction module is used to predict the destruction energy and destruction time based on the axial deformation.

[0008] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention provides a method and system for predicting rock failure. First, strain distribution images of rock samples are acquired during rock compression tests using DIC technology. A dataset is formed by selecting strain distribution maps with strain values ​​between 0 and a preset value, as well as the strain distribution map at peak failure. A PreRNN++ prediction model is built, using the strain distribution maps between 0 and the preset value as input and the strain distribution map at peak failure as training data. The trained model is then used to predict the peak failure image of the rock sample, and the axial deformation is calculated based on the predicted image, thereby estimating the failure energy and failure time. This invention can effectively predict the location of the main fracture surface and the axial deformation of the rock sample, and can also estimate the failure energy and failure time. Furthermore, the prediction results are directly presented in the image of the rock sample fracture surface, achieving the goal of directly and vividly displaying the location of the main fracture surface while maintaining high-efficiency prediction. This invention provides a certain technical foundation for the prediction of major natural disasters, especially the prediction of surrounding rock instability and failure. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of the rock failure prediction method provided in Embodiment 1 of the present invention; Figure 2 This is a flowchart of the DIC image acquisition process provided in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the DIC provided in Embodiment 1 of the present invention; Figure 4 This is a structural diagram of the Causal LSTM unit provided in Embodiment 1 of the present invention; Figure 5 This is a structural diagram of the GHU unit provided in Embodiment 1 of the present invention; Figure 6 This is a topology diagram of the PreRNN++ model provided in Embodiment 1 of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] Terminology Explanation: DIC: Digital Image Correlation (DIC), also known as digital speckle correlation, is a method that uses digital images of a specimen before and after deformation to obtain deformation information of the target area through correlation calculations.

[0013] RNN: Recurrent Neural Network (RNN) is a type of neural network used to process sequential data. Compared to general neural networks, it can handle data with varying sequences.

[0014] LSTM: Long Short-Term Memory (LSTM) is a special type of RNN neural network, mainly designed to solve the gradient vanishing and gradient explosion problems during long sequence training.

[0015] PreRNN++: A recurrent network architecture for spatiotemporal predictive learning, also known as a predictive RNN (PreRNN). This network uses some of the principles and framework of LSTM and is able to simultaneously memorize temporal and spatial information.

[0016] Uniaxial compression test of rock: The uniaxial compression test, also known as the "uniaxial test," is a type of rock compression test. It refers to a test in which a load is applied only along the axial direction to a cylindrical or prismatic rock sample. It is commonly used to determine the strength, deformation, and failure characteristics of rocks.

[0017] The main fracture surface of a rock sample refers to the fracture surface that begins when the rock is subjected to stress exceeding its strength, i.e., the stress difference exceeds the fracture strength. At the beginning of the fracture, micro-cracks first appear, and these micro-cracks gradually develop, connect with each other, and form a clear fracture surface.

[0018] Peak rock strength: Peak rock strength is the stress value corresponding to the highest point on the stress-strain curve of a rock sample. When the rock reaches this value, its internal structure has been damaged, deformation has intensified, and its load-bearing capacity has decreased significantly.

[0019] Peak failure time: The time required from the start of loading during the test to the peak strength of the rock.

[0020] Short-term forecast: During a uniaxial compression test of rock, a forecast is made of the failure time, the location of the main fracture surface, and the failure energy shortly before the rock fails.

[0021] Spatiotemporal sequence prediction: refers to the prediction of the state of a complex dynamic system in time and space.

[0022] Existing prediction methods are mostly inadequate for predicting the evolution of rock failure and the fracture surface. Many methods are also complex, costly, and the prediction results are not intuitive, with cumbersome manual operation procedures.

[0023] The purpose of this invention is to provide a rapid, accurate, and efficient method for predicting rock failure that can be used in indoor compression (tension) tests.

[0024] The instability and failure of rock and rock mass systems is a core scientific problem concerning major natural disasters. However, predicting and forecasting major natural disasters involving rock mass systems remains very difficult. To advance research on the prediction and forecasting of rock mass systems, this study focuses on the simplest rock testing system: the indoor uniaxial compression test. Combining DIC technology and machine learning algorithms, based on test data before the axial strain of standard rock samples reaches 0.5% and full strain maps acquired using DIC technology, a suitable spatiotemporal sequence model is constructed to predict the location of the main fracture surface and the axial deformation at the peak strength. Furthermore, the model incorporates stress-strain characteristics to predict the failure time and energy.

[0025] This invention proposes a method for predicting rock failure based on the simplest uniaxial compressive strength test. This invention can serve as a preliminary study for major geological hazards such as rockbursts, collapses, landslides, and earthquakes in the field of rock engineering.

[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] Example 1 This embodiment sequentially comprises seven processes: a uniaxial compression test of rock under constant displacement rate loading, DIC image acquisition, dataset establishment and preprocessing, PreRNN++ prediction model construction, model training, model evaluation, and practical application. The constant displacement rate loading rock compression test provides data support for this method. DIC technology is a non-contact strain measurement method that utilizes high-speed cameras and image processing technology to acquire strain information on the rock sample surface during the compression test, providing reliable information support for the subsequent prediction model. Dataset establishment and preprocessing aim to filter the feature images of the model's input and output; splitting the training and test sets according to a certain ratio prepares for training and testing the prediction model. The PreRNN++ prediction model is a Causal LSTM model incorporating cascaded dual memory, which effectively enhances the recursion depth between time steps and strengthens the connections between spatial features, exhibiting good performance in predicting fracture surfaces. After model establishment, the model needs to be trained using the training set, and the model's performance needs to be tested using the test set to evaluate its suitability for practical prediction and make corresponding adjustments. Since the ultimate goal of this method is to predict the fracture surface, fracture energy, and failure time, it is also necessary to establish the relationship between the three based on theoretical relationships to achieve practical prediction.

[0028] Specifically, this embodiment provides a method for predicting rock failure; please refer to [link / reference]. Figure 1 ,include: S1. During the entire process of rock compression testing on the rock sample, the strain distribution map on the surface of the rock sample is acquired using DIC technology. Step S1 requires performing a compression test on the rock sample and collecting strain distribution maps using DIC technology during the test.

[0029] To ensure good generalization ability of the model trained in subsequent processes, the compression test used 100 samples each of granite, sandstone, shale, quartzite, and gneiss. The samples were prepared as standard specimens with a diameter of 50 mm and a height of 100 mm. A constant displacement loading rate v0 was set between 0.02 and 0.05 mm / s. The test was performed until the rock sample reached its peak and then failed. The stress-strain characteristics throughout the process were recorded. Stress-strain data within the strain range of 0% to 0.5% were used to fit the data and obtain the expression. This is used for subsequent prediction of destructive energy.

[0030] DIC (Displacement Injection) technology can be used to acquire strain distribution maps on the surface of rock samples. Its basic principle is to obtain the displacement vector of a point by measuring the position change of the same pixel in adjacent speckle images, thereby obtaining the full-field displacement of the specimen surface. DIC technology has the advantages of full-field measurement, non-contact operation, and high precision. It has low environmental requirements, provides intuitive measurement results, and is suitable for dynamic measurements.

[0031] As an optional implementation, please refer to Figure 2 The process of acquiring strain distribution maps on the surface of the rock sample using DIC technology throughout the rock compression test specifically includes: S11. Spraying is performed on the surface of the rock sample to obtain scattered spots.

[0032] Specifically, you can first spray matte white paint onto the rock surface, and then spray black paint to create randomly scattered black spots on the rock surface.

[0033] S12. During the entire rock compression test, images of the rock sample are acquired using a camera to obtain the target image.

[0034] Specifically, light sources and high-resolution cameras (4096×3000PX, 30fps) can be set up in four mutually perpendicular directions on the rock sample (rock specimen), with a total of two sets, to acquire images of the entire experimental process. This embodiment does not limit the number of camera sets; those skilled in the art can increase or decrease the number of camera sets according to actual needs.

[0035] S13. Using the speckle as a reference, calculate the full-field strain displacement of the target image using the DIC algorithm.

[0036] As an optional implementation, step S13 specifically includes: S131. Divide the target image into multiple sub-regions.

[0037] S132. Obtain the coordinates of all pixels in each sub-region before and after deformation.

[0038] S133. Calculate the cross-correlation matrix of all sub-regions before and after deformation based on the coordinates.

[0039] S134. Solve the cross-correlation matrix using the nonlinear least squares method to obtain the full-field strain displacement of the target image.

[0040] like Figure 3 As shown, with each calculation point as the center, a certain area on the target image is selected as the sub-region of the current calculation point. Through the pixels in the sub-region, a certain cross-correlation function is constructed, and nonlinear least squares solution is performed to calculate displacement and strain data.

[0041] Assuming that the displacement of a pixel in the current target image undergoes a linear first-order transformation relative to the displacement of a pixel in the corresponding sub-region of the reference image, the following pixel coordinate relationship can be obtained: in and It is the initial reference subset point and coordinate, and It is the center of the initial reference subset and coordinates and It is the final subset point and coordinate,( i , j () is an index used to indicate the relative position of a subset point to its center. and These are the displacements in the x and y directions, respectively. These are the x and y coordinates of the center pixel of the sub-region, respectively. These represent the horizontal and vertical displacements of the center pixel of the sub-region, respectively.

[0042] Then, a cross-correlation matrix is ​​constructed between the reference image sub-region and the current deformed image sub-region. The similarity of the gray values ​​of each pixel in the two sub-regions is compared using the traditional covariance correlation function formula: Correlation coefficient Represents the point before deformation The sub-region centered and the point after deformation The correlation between the central sub-regions When = 1, it indicates that the two sub-regions are completely correlated; M Indicates the image before deformation as a point The sub-region size parameter is centered, and the sub-region size is... pixel× pixel; and These represent the grayscale values ​​of a point in the image before and after deformation, respectively. and These represent the average grayscale values ​​of the images before and after deformation, respectively.

[0043] Subpixel-precision displacement calculation is achieved by minimizing the correlation coefficient using a nonlinear optimization method. The optimization process mainly includes the following steps: (1) Initial displacement value calculation. This step is mainly achieved through integer pixel search to obtain the initial displacement value of nonlinear optimization. In addition, the initial displacement value can also be obtained through methods such as fast Fourier transform and feature matching.

[0044] (2) Iteratively solve the subpixel displacement. Perform a Taylor expansion of the nonlinear equation, and then use the iterative least squares method to solve for the subpixel displacement.

[0045] (3) Subpixel coordinate grayscale interpolation. In the process of iteratively solving the subpixel displacement, it is necessary to obtain the grayscale value of the subpixel position region in the speckle image, which requires interpolation using the grayscale values ​​of the surrounding integer pixel coordinates.

[0046] Finally, the reliability-guided DIC algorithm is adopted, which spreads the reliability-guided initial value between neighboring calculation points to achieve accurate and efficient full-field displacement calculation.

[0047] S14. Based on the full-field strain displacement, obtain the strain distribution map on the surface of the rock sample.

[0048] After obtaining the total displacement, the strain distribution diagram is obtained by using the formula: strain = displacement / length.

[0049] S2. Select an image of the main fracture surface of the rock sample from the strain distribution map and record it as the first image.

[0050] The strain distribution map of the rock sample surface obtained in step S14 contains strain distribution maps at multiple angles. It is necessary to select images from these images that include those captured by the camera at the main fracture surface for subsequent model training.

[0051] S3. Select a dataset from the first image and divide the dataset into a training set and a test set; the dataset includes a strain distribution map of strain values ​​between 0 and a preset value and a strain distribution map at peak failure.

[0052] Specifically, starting from strain 0, ten DIC full-field strain distribution maps at locations with adjacent strain differences of 0.05% were taken as input to the model. The DIC full-strain map at peak failure was taken as the model output. Since there were five types of rock samples, a total of 500 datasets were obtained. The spatial resolution of all images was converted to 200×400 dpi. The datasets of the five different lithologies were split into training and testing sets in an 8:2 ratio, resulting in 400 training sets and 100 testing sets.

[0053] S4. Build a prediction model based on PreRNN++, and use the strain distribution map of strain values ​​between 0 and preset values ​​in the training set as input, and use the strain distribution map at peak failure in the training set as labels to train the prediction model to obtain the first model.

[0054] The PreRNN++ prediction model can simultaneously capture temporal and spatial variations because it employs cascaded dual-memory Causal LSTM units, which enhances spatial correlation and recursion depth between time steps, resulting in excellent learning of the deformation process of the fracture surface in both time and space. The structure of the Causal LSTM unit is as follows: Figure 4 The key equation is as follows: in, Represents time memory, Represents spatial memory. Indicates the time step. Indicates the first A hidden layer, Represents the Gate of Oblivion Indicates the input gate. Indicates the input modulation gate. W1-W5 represent output gates and convolutional filters, respectively. Both are 1×1 convolutional filters. This indicates the final output.

[0055] On the other hand, to avoid the gradient vanishing problem that easily occurs during model training, the Gradient Highway Unit (GHU unit) is introduced. The structure of the GHU unit is as follows: Figure 5 The key equation is as follows: in, Indicates opening and closing the door. Indicates input conversion. Indicates a hidden state.

[0056] To adapt to the prediction of the fracture surface, a prediction model based on PreRNN++ was established, such as... Figure 6The model includes an input layer, hidden layers, and a recovery layer. The hidden layer consists of four Causal LSTM layers and a Gradient Highway Unit (GHU) layer in a cascaded structure, with the GHU layer inserted between the first and second Causal LSTM layers. Each Causal LSTM layer contains 11 Causal LSTM units, and each GHU layer contains 11 GHU units. The first 10 units of the first Causal LSTM layer are connected to the input layer, and the last unit of the fourth Causal LSTM layer is connected to the recovery layer. The first Causal LSTM layer, the GHU layer, the fourth Causal LSTM layer, and the recovery layer all have 64 hidden states, and use 3×3, 1×1, 5×5, and 1×1 convolutional kernels respectively. The second and third Causal LSTM layers each have 128 hidden states and use 5×5 convolutional kernels.

[0057] After the model is built, 400 training sets are put into the model for parameter training until the loss value and prediction accuracy tend to stabilize.

[0058] S5. Evaluate the first model using the test set. If the evaluation result does not meet the preset requirements, adjust the hyperparameters of the prediction model and return to step "train the prediction model using the training set" until the evaluation result meets the preset requirements to obtain the second model.

[0059] Specifically, 100 test sets were placed on the trained model to predict the rupture surface, and the model was evaluated using three metrics: Mean Square Error (MSE), Structural Similarity Index (SSIM), and Peak Signal to Noise Ratio (PSNR).

[0060] The mean squared error (MSE) determines the difference between two images by measuring the mean squared value of the pixel differences between the generated image and the target image. Its definition is as follows: in, Represents the target image. Represents the predicted image. This represents the x and y positions of each pixel in each image. The height of the sample image. The width of the sample image.

[0061] Structural similarity, based on the sensitivity of human visual perception, evaluates the similarity between the predicted image and the target image from three aspects: brightness, contrast, and structure. Its definition is as follows: in, They represent , The mean, They represent , variance express covariance, Represents a constant.

[0062] Peak signal-to-noise ratio (PSNR) is an error based on the corresponding pixel, and its definition is as follows: Where max represents the maximum pixel value in the image.

[0063] Meanwhile, the prediction results of axial deformation are evaluated using relative error, which is defined as follows: in, L 1 This represents the actual axial height of the rock sample at peak failure. This represents the predicted axial height of the rock sample at peak failure.

[0064] If the prediction results meet the accuracy requirements, the model is suitable for predicting the fracture surface at peak failure. If not, the hyperparameters of the model should be adjusted and the model retrained.

[0065] S6. The strain distribution map at the peak failure of the rock sample to be predicted is predicted using the second model to obtain the predicted image.

[0066] The trained and evaluated model can be used to predict the peak fracture surface of a rock sample, provided that some data of the sample is known, including its volume and height. Furthermore, it can determine the stress within a certain stress range of the sample. strain However, due to the lack of information about the relationship between the two rock samples, it is impossible to know the information of the main fracture surface, the energy of failure, and the time of failure at the peak failure time. Therefore, it is necessary to use a pre-trained PreRNN++ prediction model (second model) to predict the rock sample, thereby obtaining the information of the main fracture surface, the energy of failure, and the time of failure at the peak failure time of the rock sample.

[0067] Based on the known strain distribution map of the rock sample within the range of 0~0.5% strain, it is input into the second model to obtain the image of the main fracture surface at peak failure.

[0068] S7. Calculate the axial deformation of the rock sample to be predicted when it fails, based on the positional changes of the pixels in the predicted image in the pressure direction; predict the failure energy and failure time based on the axial deformation.

[0069] Since the rock compression test involves applying a force along the vertical axis to the rock, the direction of pressure described here can refer to the positional change of a pixel along the vertical axis. It should also be noted that this positional change is relative to the state before pressure is applied.

[0070] It should also be noted that the data of the rock sample to be predicted, such as the model training data, is obtained by applying pressure at a constant displacement rate.

[0071] The axial deformation at failure can be obtained from the image of the main fracture surface: in, L The actual height of the rock sample is 10 cm in this embodiment.

[0072] Based on the stress within a certain stress range of the rock sample to be tested strain The relationship was established, and the expression was obtained by fitting the data using the least squares method. According to the definition of peak failure energy, the energy at failure can be estimated using the following formula, where V The volume of the rock sample to be tested: Considering the experiment was conducted at a constant displacement rate v 0 Loading allows us to calculate the time of failure: This embodiment uses five different lithological specimens. A standard specimen is loaded at a constant loading rate, and the stress-strain curves of the rock specimens throughout the entire process are recorded. The stress-strain relationship expression for strains from 0% to 0.5% is obtained. During the loading process, DIC (Diverterless Computation) technology is used to acquire full-strain maps of the specimens throughout the failure process. Ten DIC maps are constructed at 0.05% strain intervals, representing strains from 0% to 0.5%, as a data set. A total of 500 sets of experiments are completed, serving as the dataset for subsequent machine learning (400 sets for training and 100 sets for testing). A PreRNN++ prediction model incorporating a cascaded dual-memory Causal LSTM model is then built. The 400 training sets are fed into this model for parameter training until the loss value and prediction accuracy stabilize. This method uses four model evaluation metrics: mean squared error, structural similarity, peak signal-to-noise ratio, and relative squared error to evaluate the model's performance. After the prediction model is trained, the 100 test sets are used for model evaluation. If the evaluation results are good, the fracture surface is predicted and the axial deformation is determined. The failure energy and failure time can be calculated by integrating the stress-strain relationship expression obtained from the rock test. If the model evaluation results are poor, the hyperparameters are adjusted and the model is retrained.

[0073] The entire prediction process can effectively predict the location of the main fracture surface and the axial deformation of the rock sample, and can also estimate the failure energy and failure time. Furthermore, the prediction results in this embodiment are directly presented as images of the rock sample fracture surface, achieving the goal of directly and vividly displaying the location of the main fracture surface of the rock while maintaining high-efficiency prediction.

[0074] The advantages of this embodiment are as follows: (1) The method of combining DIC technology with spatiotemporal sequence model has a high degree of visualization and can intuitively display the fracture surface of rock in the form of images.

[0075] (2) The operation process is simple, convenient and quick. The prediction process itself will not interfere with the rock sample and can better predict the fracture surface and axial deformation of the rock. (3) It adopts non-contact operation, eliminating the need to attach sensors, and can improve detection efficiency; (4) A prediction model based on PreRNN++ was proposed, which enhances the spatiotemporal correlation of prediction compared with traditional machine learning prediction models, and solves the problem of gradient vanishing during backpropagation.

[0076] Example 2 This embodiment provides a rock failure prediction system, including: The DIC acquisition module M1 is used to acquire the strain distribution map on the surface of the rock sample using DIC technology throughout the entire process of rock compression test. The main fracture surface selection module M2 is used to select an image of the main fracture surface of the rock sample from the strain distribution map, which is denoted as the first image. The dataset acquisition module M3 is used to select a dataset from the first image and divide the dataset into a training set and a test set; the dataset includes a strain distribution map with strain values ​​between 0 and a preset value and a strain distribution map at peak failure; The model training module M4 is used to build a prediction model based on PreRNN++. It takes the strain distribution map of strain values ​​between 0 and preset values ​​in the training set as input and the strain distribution map at peak failure in the training set as labels to train the prediction model and obtain the first model. The model evaluation module M5 is used to evaluate the first model using the test set. If the evaluation result does not meet the preset requirements, the hyperparameters of the prediction model are adjusted, and the process returns to the step "and train the prediction model using the training set" until the evaluation result meets the preset requirements, thus obtaining the second model. The prediction image acquisition module M6 is used to predict the strain distribution map at the peak failure of the rock sample to be predicted using the second model, and obtain the prediction image. Axial deformation prediction model M7 is used to calculate the axial deformation of the rock sample to be predicted when it fails, based on the changes of pixels in the predicted image in the pressure direction. A destruction energy and destruction time prediction module is used to predict the destruction energy and destruction time based on the axial deformation.

[0077] Optionally, the DIC acquisition module M1 specifically includes: The spraying submodule is used to spray the surface of the rock sample to obtain a speckled pattern. The acquisition submodule is used to acquire images of the rock sample using a camera throughout the entire rock compression test to obtain the target image; The DIC processing submodule is used to calculate the full-field strain displacement of the target image using the DIC algorithm with the speckle as a reference. The strain distribution map acquisition submodule is used to obtain the strain distribution map of the rock sample surface based on the full-field strain displacement.

[0078] Optionally, the DIC processing submodule specifically includes: A partitioning unit is used to divide the target image into multiple sub-regions; A coordinate acquisition unit is used to acquire the coordinates of all pixels in each sub-region before and after deformation. The cross-correlation matrix acquisition unit is used to calculate the cross-correlation matrix of all the sub-regions before and after deformation based on the coordinates. The least squares solution unit is used to solve the cross-correlation matrix using the nonlinear least squares method to obtain the full-field strain displacement of the target image.

[0079] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0080] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting rock failure, characterized in that, include: Throughout the rock compression test on the rock sample, the strain distribution map on the surface of the rock sample was acquired using DIC technology. An image of the main fracture surface of the rock sample is selected from the strain distribution map and recorded as the first image; A dataset is selected from the first image and divided into a training set and a test set; the dataset includes a strain distribution map with strain values ​​between 0 and a preset value, and a strain distribution map at peak failure. A prediction model based on PreRNN++ is constructed, and the strain distribution map of strain values ​​between 0 and preset values ​​in the training set is used as input. The prediction model is trained using the strain distribution map at peak failure in the training set as labels to obtain a first model. The prediction model based on PreRNN++ includes an input layer, a hidden layer, and a restoration layer. The hidden layer consists of four layers of Causal LSTM and Gradient Highway Unit layer in a cascaded structure. The Gradient Highway Unit layer is inserted between the first Causal LSTM layer and the second Causal LSTM layer. Each Causal LSTM layer contains 11 Causal LSTM units, and the Gradient Highway Unit layer contains 11 Gradient Highway Unit units. The first 10 units of the first Causal LSTM layer are connected to the input layer, and the last unit of the fourth Causal LSTM layer is connected to the restoration layer. The first model is evaluated using the test set. If the evaluation result does not meet the preset requirements, the hyperparameters of the prediction model are adjusted, and the prediction model is trained again using the training set until the evaluation result meets the preset requirements, thus obtaining the second model. The strain distribution map at the peak failure of the rock sample is predicted using the second model to obtain a predicted image; the axial deformation at the failure of the rock sample is calculated based on the positional changes of the pixels in the pressure direction in the predicted image. The specific method for predicting the failure energy and failure time based on the axial deformation is as follows: based on the stress-strain curve of the rock sample to be predicted, the relationship between stress σ and strain ε is fitted using the least squares method to obtain the stress-strain expression; Using formula Calculate the destructive energy U f V is the volume of the rock sample to be predicted, ΔL is the axial deformation, and L is the actual height of the rock sample to be predicted. The failure time t is calculated using the formula t = ΔL / v0; v0 is the constant displacement rate of the rock sample to be tested in a rock compression test.

2. The method according to claim 1, characterized in that, During the entire rock compression test of the rock sample, the strain distribution map on the surface of the rock sample is acquired using DIC technology, specifically including: Spraying was performed on the surface of the rock sample to obtain scattered spots; Throughout the rock compression test, images of the rock sample are acquired using a camera to obtain the target image; Using the speckle as a reference, the full-field strain displacement of the target image is calculated using the DIC algorithm; Based on the full-field strain displacement, the strain distribution map of the rock sample surface is obtained.

3. The method according to claim 2, characterized in that, The calculation of the full-field strain displacement of the target image using the DIC algorithm specifically includes: The target image is divided into multiple sub-regions; Obtain the coordinates of all pixels in each sub-region before and after deformation; Based on the coordinates, calculate the cross-correlation matrix of all the sub-regions before and after deformation; The cross-correlation matrix is ​​solved using the nonlinear least squares method to obtain the full-field strain displacement of the target image.

4. The method according to claim 1, characterized in that, The step of selecting the dataset from the first image specifically includes: Multiple strain distribution maps with strain values ​​in the range of 0-0.5% were selected, using a strain value step size of 0.05%. Select the strain distribution diagram at the peak failure; The dataset is obtained from multiple strain distribution maps within the strain range of 0-0.5% and the strain distribution map at peak failure.

5. The method according to claim 1, characterized in that, The first layer of Causal LSTM, the GradientHighway Unit layer, the fourth layer of Causal LSTM, and the restoration layer all have 64 hidden states, and 3×3, 1×1, 5×5 and 1×1 convolutional kernels are selected in sequence. Both the second-layer Causal LSTM and the third-layer Causal LSTM have 128 hidden states and use 5×5 convolutional kernels.

6. The method according to claim 1, characterized in that, The method further includes evaluating the first model using three metrics: mean square error, structural similarity, and peak signal-to-noise ratio.

7. A rock failure prediction system, characterized in that, include: The DIC acquisition module is used to acquire the strain distribution map on the surface of the rock sample using DIC technology throughout the entire process of rock compression testing. The main fracture surface selection module is used to select an image of the main fracture surface of the rock sample from the strain distribution map, which is denoted as the first image. The dataset acquisition module is used to select a dataset from the first image and divide the dataset into a training set and a test set; the dataset includes a strain distribution map with strain values ​​between 0 and a preset value and a strain distribution map at peak failure; The model training module is used to build a prediction model based on PreRNN++. It takes the strain distribution map of strain values ​​between 0 and a preset value in the training set as input and the strain distribution map at peak failure in the training set as labels to train the prediction model, obtaining a first model. The PreRNN++-based prediction model includes an input layer, a hidden layer, and a restoration layer. The hidden layer consists of four layers of Causal LSTM and Gradient Highway Unit layers in a cascaded structure. The Gradient Highway Unit layer is inserted between the first and second Causal LSTM layers. Each Causal LSTM layer contains 11 Causal LSTM units, and each Gradient Highway Unit layer contains 11 Gradient Highway Unit units. The first 10 units of the first Causal LSTM layer are connected to the input layer, and the last unit of the fourth Causal LSTM layer is connected to the restoration layer. The model evaluation module is used to evaluate the first model using the test set. If the evaluation result does not meet the preset requirements, the hyperparameters of the prediction model are adjusted, and the prediction model is trained again using the training set until the evaluation result meets the preset requirements, thus obtaining the second model. The prediction image acquisition module is used to predict the strain distribution map at the peak failure of the rock sample to be predicted using the second model, and obtain the prediction image. The axial deformation prediction module is used to calculate the axial deformation of the rock sample to be predicted when it fails, based on the positional changes of the pixels in the predicted image in the pressure direction. The destruction energy and destruction time prediction module is used to predict the destruction energy and destruction time based on the axial deformation. Specifically, it uses the least squares method to fit the relationship between stress σ and strain ε based on the stress-strain curve of the rock sample to be predicted, and obtains the stress-strain expression. Using formula Calculate the destructive energy U f V is the volume of the rock sample to be predicted, ΔL is the axial deformation, and L is the actual height of the rock sample to be predicted. The failure time t is calculated using the formula t = ΔL / v0; v0 is the constant displacement rate of the rock sample to be tested in a rock compression test.

8. The system according to claim 7, characterized in that, The DIC acquisition module specifically includes: The spraying submodule is used to spray the surface of the rock sample to obtain a speckled pattern. The acquisition submodule is used to acquire images of the rock sample using a camera throughout the entire rock compression test to obtain the target image; The DIC processing submodule is used to calculate the full-field strain displacement of the target image using the DIC algorithm with the speckle as a reference. The strain distribution map acquisition submodule is used to obtain the strain distribution map of the rock sample surface based on the full-field strain displacement.

9. The system according to claim 8, characterized in that, The DIC processing submodule specifically includes: A partitioning unit is used to divide the target image into multiple sub-regions; A coordinate acquisition unit is used to acquire the coordinates of all pixels in each sub-region before and after deformation. The cross-correlation matrix acquisition unit is used to calculate the cross-correlation matrix of all the sub-regions before and after deformation based on the coordinates. The least squares solution unit is used to solve the cross-correlation matrix using the nonlinear least squares method to obtain the full-field strain displacement of the target image.

Images