A Rock Failure Pattern Recognition Method Based on Image Recognition Algorithm
By using a CNN model based on image recognition algorithms to monitor the rock failure process in real time, the problem of speed and accuracy in the identification of rock mass failure modes in existing technologies has been solved, and real-time monitoring and accurate early warning of rock failure modes have been achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- UNIV OF SCI & TECH LIAONING
- Filing Date
- 2023-03-02
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies are insufficient for quickly and in real-time identification of rock mass failure modes, resulting in high labor costs and time consumption, and failing to meet the needs of rapid engineering assessment.
A method based on image recognition algorithms is adopted, which uses high-speed cameras to capture real-time photos of rock failure processes, uses a CNN model for digital image correlation processing, and combines attention mechanism and residual network module to extract rock displacement field features, establish a labeled dataset and train the model to realize real-time monitoring and identification of rock failure modes.
It improves the accuracy and robustness of rock failure mode recognition, enables real-time monitoring of rock failure, provides early warning of rock instability, and reduces labor and time costs.
Smart Images
Figure CN116129274B_ABST
Abstract
Description
Technical Field
[0001] This solution belongs to the field of rock engineering technology, specifically involving a rock damage mode recognition method based on image recognition algorithms. Background Technology
[0002] The mechanical properties of rock masses have a significant impact on engineering construction. In particular, the presence of weak structural planes in rock masses not only weakens the mechanical properties of the rock mass but also easily forms through-slip surfaces. Therefore, developing a rapid identification method for rock mass structure failure modes is an urgent need in the field of rock engineering.
[0003] Current research on rock mass failure modes mainly focuses on laboratory tests and numerical simulations. Laboratory tests primarily determine the failure mode of rock mass structures through scaled-down model experiments. Numerical simulations mainly employ continuum mechanics methods (finite element method, finite difference method) or discontinuous media mechanics methods (discrete element method, discontinuous deformation analysis method) to conduct numerical tests and determine the failure mode of rock mass structures. However, laboratory tests not only require significant time and effort, but also may not be suitable for some complex models. While numerical simulation has significant advantages in modeling, it still requires considerable time and cannot meet the needs of rapid assessment in engineering. Therefore, developing a rapid identification method for rock failure modes is both a key focus and a challenge in the prevention and control of rock engineering disasters.
[0004] Application number CN202111328786.6 discloses a rapid identification method for rock mass structure failure modes, comprising: determining the maximum and minimum principal stresses of the rock mass environment through geostress measurement; determining the cohesion and internal friction angle of the rock through direct shear tests; conducting rebound and shear tests on typical structural planes in the rock mass to determine the rebound strength and basic friction angle of the structural planes; statistically analyzing the distribution of structural planes to determine the surface roughness and the angle β between the normal direction of the structural plane and the direction of the maximum principal stress; assuming that the structural planes satisfy the Barton-Bandis yield criterion to obtain the structural plane failure criterion; assuming that the rock satisfies the Mohr-Colomb yield criterion to obtain the rock failure criterion; substituting the rock shear strength parameters into the rock failure criterion and the physical and mechanical property parameters of the structural planes into the structural plane failure criterion to comprehensively determine the rock mass structure failure mode;
[0005] This scheme targets the failure modes of rock masses. After obtaining the stress state and physical and mechanical parameters of the rock and structural surfaces through field tests, the failure mode can be determined through simple algebraic calculations. It boasts advantages such as ease of operation, advanced technology, high calculation accuracy, and strong engineering applicability. However, rock mass failure caused by complex environmental factors has a significant impact on engineering projects. For example, the Dagangshan Hydropower Station, which has already been built, recently experienced landslides in its reservoir area due to an earthquake. Rapidly identifying the failure modes of the slope rock mass would play a crucial role in post-disaster reconstruction and early warning of major disasters. However, due to the lack of devices and intelligent methods capable of real-time analysis of rock mass failure modes, analysis still relies on manual labor, resulting in significant labor and time costs for identification. Therefore, exploring intelligent real-time monitoring and identification methods for rock failure modes has important scientific significance and engineering application value. Summary of the Invention
[0006] This solution provides a rock damage pattern recognition method based on image recognition algorithms, which can realize real-time monitoring and damage pattern recognition of rock damage.
[0007] To achieve the above objectives, this solution provides a rock damage pattern recognition method based on image recognition algorithms, comprising the following steps:
[0008] S1. Place the sample and position the high-speed camera in the direction of the free surface of the sample.
[0009] S2. The sample is loaded to failure using a press, and photos of the failure process are captured in real time using a high-speed camera.
[0010] S3. Use data analysis software to perform digital image correlation processing on the damage photos to obtain the sample displacement field characteristics and determine the failure mode based on the displacement field characteristics, and establish a labeled dataset.
[0011] S4. Use the labeled dataset from S3 to train and build a CNN model, and determine the model parameters;
[0012] The structure of the CNN model mentioned above includes an input layer, a convolutional layer, an activation layer, a pooling layer, and a fully connected layer;
[0013] Data input layer: used for data input.
[0014] Convolutional layers: utilize convolutional kernels for feature extraction and feature mapping;
[0015] Activation layer: Performs a non-linear mapping on the output of the convolutional layer, i.e., activation;
[0016] Pooling layers: Pooling layers are sandwiched between consecutive convolutional layers to compress the amount of data and parameters, reducing overfitting;
[0017] Fully connected layer: Refits at the end of the CNN to reduce the loss of feature information;
[0018] The convolutional layer has two convolutional kernels, and an attention mechanism module is provided between the two convolutional kernels. The attention mechanism module includes a global pooling operation, four fully connected layers, and a sigmoid function.
[0019] Residual network modules are provided before and after the two convolutional kernels, and the residual network modules add the images before and after the convolution operation;
[0020] S5. Using a trained CNN model, intelligently identify the newly acquired digital images and output the failure mode of the sample.
[0021] The beneficial effects of this scheme are as follows: This scheme adds an attention mechanism module between the two convolutional kernels, including global pooling, four fully connected layers, and a sigmoid function, which enables the CNN network to have better feature extraction capabilities. At the same time, this scheme performs residual operations before and after the two convolutional kernels, i.e., residual network modules, which add the images before and after the convolutional operation to avoid gradient explosion and vanishing problems. The attention mechanism module and residual network module significantly improve the recognition rate of rock damage patterns, and have better robustness and generalization effect in rock damage pattern recognition, thereby improving the accuracy of rock damage pattern recognition. It can also realize real-time monitoring of rock damage.
[0022] Furthermore, protective measures are installed in front of the high-speed camera to protect operators from being injured by flying rock fragments.
[0023] Furthermore, in step S1, the protective material is a transparent acrylic sheet. Acrylic sheets have high impact resistance, sixteen times that of ordinary glass, and are transparent, making them very suitable as a barrier for rock samples.
[0024] Furthermore, in step S2, a high-speed camera tracks and records the deformation information of measuring points on the rock surface in real time. The deformation information is the displacement value of a measuring point at a certain moment. Without the need for stress information, the displacement information of characteristic measuring points can be obtained through non-contact measurement technology, allowing direct prediction of the rock's failure state and failure mode, making the operation convenient.
[0025] Furthermore, in step S3, the sample displacement field characteristics include the displacement vector field of each measuring point on the rock over time, the displacement increment, the average value, variance, and coefficient of variation of the displacement increment. Based on the rock displacement and deformation information from multiple measuring points, and using the displacement increment coefficient of variation calculation method, the precursor information of rock failure can be accurately captured according to the coefficient of variation change curve, thereby achieving the function of real-time early warning of rock failure. In addition, based on the characteristics of the displacement vectors on both sides of the macroscopic crack, the type of crack can be determined, thereby achieving the function of determining the rock failure mode.
[0026] Furthermore, in step S3, the labeled dataset includes a lower limit 'a' of the variation of the displacement increment coefficient of each measuring point, determined based on the calculated range of the coefficient of variation values. The stability of the rock is judged based on the displacement increment coefficient of variation value 'a'. When the displacement increment coefficient of variation is less than 'a', the rock is in the pre-linear elastic and plastic deformation stage, indicating a stable state without failure. When the displacement increment coefficient of variation equals 'a', it indicates that the rock deformation has entered the post-plastic deformation stage, signifying the formation of numerous cracks within the rock and its tendency towards instability. The type of crack is determined based on the displacement vector field characteristics on both sides of the macroscopic crack: a shear crack is present when the displacement vectors on both sides are parallel, and a tension crack is present when the displacement vectors intersect. When the displacement increment coefficient of variation is greater than 'a', it indicates that the rock is nearing failure and instability. Based on the magnitude of the displacement increment coefficient of variation, the state of the rock system can be determined, accurately predicting the final failure time of the rock with high accuracy. Based on the displacement vector characteristics, the rock failure mode can be determined with accurate results.
[0027] Furthermore, in step S3, the lower limit α of the coefficient of variation of the displacement increment of the measuring point is between 0.1 and 0.4 for rocks of different lithologies and scales.
[0028] Furthermore, in step S4, the number of each pixel point generated by the digital image should be no less than 20, and the total area of the measurement point control area should cover 1 / 3 to 2 / 3 of the rock surface area.
[0029] Furthermore, it also includes a computer device for rock damage prediction and damage pattern recognition, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-described method when executing the computer program. Attached Figure Description
[0030] Figure 1 This is a schematic diagram of the method steps in an embodiment of the present invention. Detailed Implementation
[0031] The following detailed description illustrates the specific implementation method:
[0032] The basic implementation examples are as follows: Figure 1 As shown:
[0033] A method for identifying rock damage patterns based on image recognition algorithms includes the following steps:
[0034] S1. Place the sample. Position the high-speed camera in the free surface direction of the sample and install a protective shield in front of the camera. The protective material is a transparent acrylic sheet. Transparent acrylic sheets have strong impact resistance, sixteen times that of ordinary glass, and are transparent, making them very suitable as a barrier for rock samples.
[0035] S2. The sample is loaded to failure using a pressure machine, and a high-speed camera captures real-time images of the failure process. The high-speed camera tracks and records the deformation information of measuring points on the rock surface in real time; the deformation information is the displacement value of a certain measuring point at a certain moment. Without stress information, the displacement information of characteristic measuring points is obtained through non-contact measurement technology, allowing direct prediction and identification of the rock failure state and failure mode. The operation is convenient.
[0036] S3. Using data analysis software, digital image correlation processing is performed on the photographs of the failure process to obtain the sample displacement field characteristics. Based on these characteristics, the failure mode is determined, and a labeled dataset is established. The sample displacement field characteristics include the displacement vector, displacement increment, mean, variance, and coefficient of variation of each measuring point on the rock over time. Based on the rock displacement field and deformation information of the characteristic measuring points, and using the displacement increment coefficient of variation calculation method, the precursor information of rock failure can be accurately captured according to the coefficient of variation change curve, thereby achieving real-time early warning of rock failure. The crack type is determined based on the displacement field direction characteristics on both sides of the macroscopic crack, thus realizing the identification of the rock failure mode.
[0037] The labeled dataset includes: digital images, failure modes, and a lower limit 'a' for the variation of the displacement increment coefficient of each measuring point, determined based on the calculated range of these values. The stability of the rock is then assessed based on the displacement increment coefficient of variation 'a'. When the displacement increment coefficient of variation is less than 'a', the rock is in the pre-linear elastic and plastic deformation stage, indicating a stable state without failure. When the displacement increment coefficient of variation equals 'a', it indicates that the rock deformation has entered the post-plastic deformation stage, suggesting the formation of numerous internal cracks and a tendency towards instability. Furthermore, the type of crack is determined based on the characteristics of the displacement vector field on both sides of the macroscopic crack: parallel displacement vectors indicate a shear crack, while intersecting vectors indicate a tension crack. When the displacement increment coefficient of variation is greater than 'a', it indicates that the rock is nearing failure and instability. Based on the magnitude of the displacement increment coefficient of variation, the state of the rock system can be determined, enabling precise prediction of the final failure mode and high accuracy. The lower limit 'a' for the displacement increment coefficient of variation ranges from 0.1 to 0.4 for rocks of different lithologies and scales.
[0038] S4. Use the labeled dataset from S3 to train and build a CNN model, and determine the model parameters; the lower limit 'a' of the coefficient of variation of the displacement increment at the measurement point is determined for rocks of different lithologies and scales.
[0039] S5. Using a trained CNN model, intelligently identify the newly acquired digital images and output the failure mode of the sample.
[0040] The structure of a CNN model includes an input layer, convolutional layers, activation layers, pooling layers, and fully connected layers.
[0041] Data input layer: used for data input.
[0042] Convolutional layers: utilize convolutional kernels for feature extraction and feature mapping;
[0043] Activation layer: Performs a non-linear mapping on the output of the convolutional layer, i.e., activation;
[0044] Pooling layers: Pooling layers are sandwiched between consecutive convolutional layers to compress the amount of data and parameters, reducing overfitting;
[0045] Fully connected layer: Refits at the end of the CNN to reduce the loss of feature information;
[0046] The convolutional layer has two convolutional kernels, and an attention mechanism module is located between the two convolutional kernels. The attention mechanism module includes a global pooling operation, four fully connected layers, and a sigmoid function.
[0047] Residual network modules are placed before and after the two convolutional kernels. The residual network modules add the images before and after the convolution operation.
[0048] It also includes a computer device for rock failure prediction and failure pattern recognition, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.
[0049] The specific implementation method is as follows:
[0050] Step 1: Using the failure test results of two sets of granite specimens (50mm*50mm*100mm and 200mm*200mm*400mm) as samples, the specimens were loaded to failure using a pressure machine, and real-time prediction was performed. Sample data includes the test time-load curve and the coefficient of variation curve of the displacement increment at the measuring points.
[0051] Step 2: For the rock surface under uniaxial compression, use digital image correlation to track and record the deformation information of 20 measuring points generated by a random algorithm. The total area of the measuring point control area accounts for 1 / 3 of the rock surface area. For the recorded measuring point deformation information, calculate the displacement increment of each measuring point over time, as well as the mean, variance, and coefficient of variation of the displacement increment.
[0052] Step 3: Based on the calculated range of the coefficient of variation of displacement increment at each measuring point, determine the lower limit of the coefficient of variation of displacement increment at the measuring point as 'a', and judge the stability of the rock based on the coefficient of variation of displacement increment 'a'; the value of 'a' is between 0.1 and 0.4.
[0053] Step 4: Use data analysis software to perform digital image correlation processing on the photos of the failure process to obtain the displacement field characteristics of the sample and determine the failure mode based on the displacement field characteristics. Then, establish a data label set based on the range of variation of the coefficient of variation of displacement increment at each measuring point in Step 3.
[0054] Step 5: Train the CNN model using the labeled dataset and determine the model parameters. The entire network uses two consecutive 3x3 convolutional kernels. Stacking these small convolutional kernels increases network depth and improves learning ability. All pooling layers use max pooling with a 2x2 kernel size and a stride of 2. After each max pooling operation, the image's length and width are halved, resulting in smaller rock feature images and clearer rock crack types, thus reducing overfitting and improving model generalization. The pooling layers all use max pooling (2x2). Through repeated convolution and pooling operations, the rock damage image is obtained.
[0055] Step Six: The trained CNN model is used to identify the granite digital image to be tested and determine the granite test location. The granite image to be tested is input into the trained CNN model. The input granite image passes through the granite measurement point extraction network to form a feature image. The range of variation values of the displacement increment of each measurement point is input into the pooling layer for normalization processing to obtain a displacement increment variation value 'a' of a unified dimension. The feature vector is input into the fully connected layer to obtain the rock fracture type and determine the rock failure mode. When the displacement increment variation coefficient of the measurement point is less than 'a', the rock is in the pre-linear elastic and plastic deformation stage, and the rock is in a stable state: no failure has occurred. When the displacement increment variation coefficient of the measurement point is equal to 'a', it indicates that the rock deformation has entered the post-plastic deformation stage, indicating that a large number of cracks have begun to be generated inside the rock, and the rock state tends to be unstable. The type of crack is determined according to the displacement vector field characteristics on both sides of the macroscopic crack: when the displacement vectors on both sides of the crack are parallel, it is a shear crack; when the displacement vectors on both sides of the crack intersect, it is a tension crack. When the displacement increment variation coefficient of the measurement point is greater than 'a', it indicates that the rock is close to failure and instability. Based on the magnitude of the displacement increment variation coefficient, the state of the rock system can be determined, and the prediction time of the final rock failure mode can be accurately determined in advance, which improves the accuracy of rock failure mode identification and realizes real-time monitoring of rock failure.
[0056] The above descriptions are merely embodiments of the present invention, and common knowledge regarding specific structures and characteristics is not elaborated upon here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. A rock failure mode identification method based on an image recognition algorithm, characterized in that: Includes the following steps: S1. Place the sample and position the high-speed camera in the direction of the free surface of the sample. S2. The sample is loaded to failure using a press, and photos of the failure process are captured in real time using a high-speed camera. S3. Use data analysis software to perform digital image correlation processing on the damage photos to obtain the sample displacement field characteristics and determine the failure mode based on the displacement field characteristics, and establish a labeled dataset. S4. Use the labeled dataset from S3 to train and build a CNN model, and determine the model parameters; The structure of the CNN model mentioned above includes an input layer, a convolutional layer, an activation layer, a pooling layer, and a fully connected layer; Data input layer: used for data input; Convolutional layers: utilize convolutional kernels for feature extraction and feature mapping; Activation layer: Performs a non-linear mapping on the output of the convolutional layer, i.e., activation; Pooling layers: Pooling layers are sandwiched between consecutive convolutional layers to compress the amount of data and parameters, reducing overfitting; Fully connected layer: Refits at the end of the CNN to reduce the loss of feature information; The convolutional layer has two convolutional kernels, and an attention mechanism module is provided between the two convolutional kernels. The attention mechanism module includes a global pooling operation, four fully connected layers, and a sigmoid function. Residual network modules are provided before and after the two convolutional kernels, and the residual network modules add the images before and after the convolution operation; S5. Using a trained CNN model, intelligently identify the newly acquired digital images and output the failure mode of the sample.
2. The rock failure mode identification method based on image recognition algorithm according to claim 1, characterized in that: Protective measures were installed in front of the high-speed camera.
3. The rock failure mode identification method based on image recognition algorithm according to claim 2, characterized in that: In step S1, the protective material is a transparent acrylic sheet.
4. The rock failure mode identification method based on image recognition algorithm according to claim 1, characterized in that: In step S2, a high-speed camera tracks and records the deformation information of the measuring points on the rock surface in real time. The deformation information is the displacement value of a measuring point at a certain moment.
5. The rock failure mode identification method based on image recognition algorithm according to claim 1, characterized in that: In step S3, the sample displacement field characteristics include the displacement increment of each measuring point on the rock surface over time, the average value, variance, and coefficient of variation of the displacement increment.
6. The rock damage pattern recognition method based on image recognition algorithm according to claim 5, characterized in that: In step S3, the labeled dataset includes determining the lower limit 'a' of the variation of the displacement increment coefficient of each measuring point based on the calculated range of the coefficient of variation values, and judging the rock stability state based on the displacement increment coefficient of variation value 'a'. When the displacement increment coefficient of variation of the measuring point is less than 'a', the rock is in the pre-linear elastic and plastic deformation stage, and the rock is in a stable state: no failure has occurred. When the displacement increment coefficient of variation of the measuring point is equal to 'a', it indicates that the rock deformation has entered the post-plastic deformation stage, indicating that a large number of cracks have begun to be generated inside the rock, and the rock state tends to be unstable. The type of crack is judged based on the displacement vector field characteristics on both sides of the macroscopic crack: when the displacement vectors on both sides of the crack are parallel, it is a shear crack; when the displacement vectors on both sides of the crack intersect, it is a tension crack. When the displacement increment coefficient of variation of the measuring point is greater than 'a', it indicates that the rock is close to failure and instability.
7. A rock damage pattern recognition method based on an image recognition algorithm according to claim 6, characterized in that: In step S3, the lower limit of the coefficient of variation α of the displacement increment of the measuring point is between 0.1 and 0.4 for rocks of different lithologies and scales.
8. The rock damage pattern recognition method based on image recognition algorithm according to claim 1, characterized in that: In step S3, the number of pixels generated in the digital image should be no less than 20, and the total area of the measurement point control area should cover 1 / 3 to 2 / 3 of the rock surface area.
9. It also includes a computer device, characterized in that: The computer device is used for rock damage prediction and damage pattern recognition. The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method according to any one of claims 1-8.