Fractured coal and rock mass strength intelligent prediction method based on image recognition

Abstract

The invention relates to a fractured coal and rock mass strength intelligent prediction method based on image recognition, and belongs to the technical field of rock mass mechanics prediction. The method comprises the following steps: (1) carrying out an indoor test according to different fracture parameters and lithology, and establishing an initial training data set; (2) introducing numerical simulation data, supplementing laboratory data through the simulation data, integrating the data into a data set, and constructing a sample database of coal and rock mass fractures and strength; (3) training a machine learning model by using the sample database, wherein the machine learning model is trained by adopting a gradient lifting decision tree model; (4) testing the precision of the machine learning model by using test set data; (5) acquiring a working face image; and (6) predicting the strength of the to-be-measured coal rock mass. According to the method, a traditional rock strength measurement method is improved, an accurate prediction model can be constructed by combining the nonlinear relationship between the geometrical characteristics and the strength of the coal-rock mass fractures with different lithology, and efficient prediction is realized.

Description

Technical Field

[0001] The invention relates to an intelligent prediction method for fractured coal rock mass strength based on image recognition, belonging to the technical field of rock mass mechanics prediction. Background Art

[0002] In rock mass engineering such as mines, tunnels, water conservancy and slopes, the mechanical parameters of coal rock mass are the basis and key to engineering design. Accurately measuring the strength of coal rock mass can help predict mine stability, evaluate support requirements, design tunnel layout, etc., which is crucial to ensure safe and efficient mining. At present, the research on the strength of coal rock mass mainly focuses on intact rocks, and these studies are usually obtained through a series of tests on intact rock specimens in the laboratory. However, the coal rock mass actually encountered in engineering usually forms a large number of cracks under various geomechanical effects, resulting in significant changes in its mechanical properties. The existence of cracks significantly weakens the overall strength, stiffness and stability of the rock mass, and the distribution morphology, density and length of the cracks also have extremely complex effects on the mechanical properties of the rock mass. At present, it is still difficult to directly apply the traditional intact rock mechanics research results to fractured coal rock mass, making the strength assessment of fractured coal rock mass a difficult point in engineering.

[0003] At present, the strength prediction of fractured coal-rock mass depends on the on-site measurement of the geometric parameters of the fractures (such as fracture length, width, inclination, density, etc.). However, the traditional on-site measurement method is usually time-consuming and labor-intensive, and relies on manual records by professionals, which is susceptible to human interference and leads to insufficient measurement accuracy. In addition, the cost of using on-site sampling and mechanical testing equipment is high, which makes it difficult to meet the needs of real-time prediction and wide application. Therefore, the engineering field urgently needs an efficient and convenient strength prediction method that can quickly predict the mechanical strength of fractured coal-rock mass without relying on complex on-site tests.

[0004] With the rapid development of computer vision and artificial intelligence technologies, methods based on image recognition and machine learning have gradually shown great application potential in the engineering field. Image recognition technology can efficiently extract the geometric characteristics of cracks on the surface of coal and rock masses, including key parameters such as the shape, length, width, and density of the cracks. These parameters can be used for strength analysis of cracked coal and rock masses, avoiding the errors and inefficiencies in traditional manual measurements. At the same time, machine learning algorithms can adaptively learn from a large amount of data and establish a nonlinear mapping relationship between crack characteristics and coal and rock strength. By using advanced machine learning models, the mechanical properties of coal and rock masses under different crack conditions can be effectively predicted, providing more accurate and efficient prediction results than traditional methods. Summary of the invention

[0005] In view of the deficiencies of the prior art, the present invention provides an intelligent prediction method for the strength of fractured coal rock mass based on image recognition, which solves the technical problem that the existing prediction methods do not consider the fractures of coal rock mass in engineering. It has the functions of automatic image acquisition, fracture feature extraction, machine learning model training and strength prediction. After the coal rock mass images taken on site are preprocessed, the geometric parameters of the coal rock mass fractures are obtained through image recognition and input into the model, thereby realizing the prediction of the strength of the coal rock mass. With the help of large-scale data sets, the machine learning model can continuously optimize the mapping relationship between fracture characteristics and strength, improve the prediction accuracy and generalization ability, and adapt to the coal rock mass strength prediction needs under different geological conditions. It can not only provide important mechanical parameters for engineering design, but also significantly improve the safety and economy of engineering construction.

[0006] The technical solution of the present invention is as follows:

[0007] An intelligent prediction method for fractured coal rock mass strength based on image recognition, the steps are as follows:

[0008] (1) Conduct indoor experiments based on different fracture parameters and lithology to establish an initial training data set;

[0009] (2) Based on the initial data set, numerical simulation data was introduced, and the discrete element numerical simulation method was used to simulate the strength of coal rock mass under different fracture shapes, densities and distribution conditions. A large amount of simulation data was used to supplement the laboratory data, and the data were integrated into a data set to construct a sample database of coal rock mass fractures and strength;

[0010] (3) Use the sample database to train the machine learning model, and the machine learning model is trained using the Gradient Boosting Decision Tree (GBDT) model;

[0011] (4) Use test set data to test the accuracy of the machine learning model;

[0012] (5) Image acquisition of working face: At the construction site, high-resolution image acquisition equipment is used to obtain the surface image of the coal and rock mass to be tested. The acquired image is pre-processed and crack boundary detection is performed to record the crack length, crack width, crack angle, and crack density of the coal and rock mass at the working face.

[0013] (6) Predict the strength of the coal and rock mass to be tested, and perform adaptive learning, adjustment, and optimization by combining the engineering site data.

[0014] According to the preferred embodiment of the present invention, in step (1), specifically:

[0015] (11) Preparation of test specimens: rock blocks with different lithology and mechanical properties were collected from the project site and processed into 100 mm × 100 mm × 200 mm rectangular original rock specimens;

[0016] Cuboid casting specimens with different mechanical properties were cast using different proportions of cement, gypsum, river sand, borax and water, with a size of 100 mm × 100 mm × 200 mm;

[0017] Cast standard cylindrical specimens with a height of 100 mm and a diameter of 50 mm;

[0018] Some of the prepared rectangular original rock specimens were cored and subjected to uniaxial compression tests to obtain the mechanical properties of the specimens, including uniaxial compressive strength and elastic modulus, which were used for parameter calibration in subsequent simulations.

[0019] (12) preparing a large number of different through-cracks on the specimen obtained in step (11), wherein the crack characteristics include parameters such as length, width, angle, density, etc. of the cracks, and simulating the common coal rock crack morphology in engineering by precisely controlling these characteristics, thereby ensuring the diversity and representativeness of the test samples;

[0020] (13) Conduct uniaxial compression tests to measure the mechanical strength of the specimens under different fracture conditions. Record and organize the results of the uniaxial compression tests to form a data set containing fracture characteristic parameters, rock properties, and corresponding strength values. This data set is used as the initial training data set for the machine learning model to provide reliable input data for the model to learn fracture characteristics and rock properties.

[0021] According to the preferred embodiment of the present invention, in step (2), the specific method of data integration is:

[0022] (21) Integrate the indoor test data obtained in step (1) and the coal rock mass fracture parameter, lithology and strength data generated by numerical simulation, and clean the integrated data to remove possible outliers, noise data and missing data to ensure data quality;

[0023] (22) Data completion. In order to ensure the integrity of the data and the reliability of the analysis, it is necessary to reasonably complete the missing data. For coal rock mass fracture parameters, if the missing proportion of the parameters is less than 5%, the mean or median method is used for completion. If the parameter characteristics show normal distribution data, the mean is used for completion. For data with extreme values ​​or skewed distributions exceeding 20%, the median is used to reduce the impact of extreme values. The mean completion formula is:

[0024]

[0025] Among them, x iis the non-missing value of the feature, N is the number of non-missing values;

[0026] (23) The fracture parameters, lithology and strength data of coal rock mass are standardized so that they are within the same numerical range to facilitate the input of the machine learning model. Standardization can effectively reduce the differences between parameters of different dimensions and ensure that the model can accurately learn and process the relationship between each fracture parameter and strength value.

[0027] (24) Classify and label the standardized data, and combine them into different strength grade labels according to the different fracture parameters and lithology of the coal and rock mass. Through labeling and classification, support is provided for the classification and prediction functions of the subsequent models.

[0028] (25) The classified and labeled data are entered into the database to construct a sample database of coal rock fracture parameters and strength. The database adopts a structured data storage method, has efficient data query and retrieval functions, and can dynamically update data to adapt to different engineering needs. The database can also be seamlessly connected with the machine learning system to provide real-time data support for the model. In order to maintain the timeliness and applicability of the database, the database is regularly maintained and updated. By continuously adding new data in subsequent engineering practices, the coverage of the database will be more extensive. At the same time, with the advancement of numerical simulation and indoor test technology, more complex fracture characteristic data will be gradually introduced to ensure the continuous expansion and optimization of the database.

[0029] Preferably, in step (3), the steps of establishing the gradient boosting decision tree model are as follows:

[0030] (311) In GBDT, decision trees are usually selected as base learners. Each tree is fitted on the residual of the previous tree. By gradually correcting the error, the overall model accuracy is ultimately improved.

[0031] (312) Definition of loss function: The goal of the gradient boosting decision tree model is to minimize the loss function by gradient descent. For regression problems, the loss function uses the mean square error (MSE) or mean absolute error (MAE). The model calculates the gradient at each step to minimize the loss.

[0032] (313) Set model hyperparameters: learning rate, number of trees (i.e., number of base learners), maximum tree depth, and subsampling rate.

[0033] Preferably, according to the present invention, in step (3), the gradient boosting decision tree model training steps are as follows:

[0034] (321) Initialization model: At the beginning of training, the predicted value is initialized with the mean value of coal rock strength as the prediction result of the first weak learner. The calculation formula is:

[0035]

[0036] in, N is the number of samples in the training set, y i is the true target value of the sample;

[0037] (322) Calculate the residual: In each iteration t, calculate the residual r between the current model prediction value and the actual value i （t） , the residual represents the prediction error of the current model. For each sample i, the residual is defined as:

[0038]

[0039] in, is the predicted value of the model at round t-1, r i （t） is the target that needs to be fitted in this round;

[0040] In round t, the new predicted value Obtained by the following formula:

[0041]

[0042] Among them, η is the learning rate, which controls the step size of each iteration, and f t (x i ) is the prediction result of the decision tree in round t, that is, the tree used to fit the residual;

[0043] (323) Fitting residuals: Taking the residuals as the target value, a new decision tree is trained to fit the current residuals. The prediction results of the decision tree are used to correct the prediction error of the model.

[0044] (324) Update the model prediction value: multiply the prediction result of the newly trained decision tree by the learning rate and add it to the overall prediction value of the model to gradually reduce the prediction error;

[0045] (325) Repeat steps (322)-(324) until a predetermined number of trees is reached or the loss function no longer decreases significantly;

[0046] (326) Hyperparameter tuning: After training is completed, the model’s hyperparameters (such as learning rate, number and depth of trees, etc.) are adjusted through grid search or random search to further improve model accuracy.

[0047] According to the preferred embodiment of the present invention, in step (4), specifically:

[0048] (41) Prepare the test set data, ensuring that the test set is independent of the training set data and contains data that the model has not seen. The test set should be representative and reflect the data distribution in actual applications. Use the same standardized parameters to process the test set data to ensure data consistency, input the feature data of the test set into the trained gradient boosting decision tree model, and obtain the predicted value of the gradient boosting decision tree model for each sample. ;

[0049] (42) Collect the true target value y of the test set and calculate the evaluation indicators mean square error MSE, mean absolute error MAE and determination coefficient R 2 ;

[0050] MSE is used to measure the square difference between the model prediction value and the true value. The smaller the MSE, the closer the model prediction is to the true value.

[0051]

[0052] Where N is the number of test set samples, y i is the true value of the sample, is the predicted value;

[0053] MAE is the absolute difference between the predicted value of the calculation model and the true value. MAE reflects the actual size of the model error;

[0054]

[0055] R 2 It is an indicator to measure the degree of model fit, with a value range of 0 to 1. The closer the value is to 1, the stronger the ability of the model to explain the data variance;

[0056]

[0057] in, is the true mean of the test set, R 2 The higher it is, the better the model can explain the data.

[0058] (43) If the prediction accuracy requirement is met, the model training is completed and the prediction result is output; if not, return to step (3), adjust the model's hyperparameters, and train the model again until the prediction accuracy requirement is met.

[0059] According to the preferred embodiment of the present invention, in step (5), the collected image is preprocessed, and the specific steps are as follows:

[0060] (1) Perform image denoising and grayscale processing on the coal and rock mass images in sequence;

[0061] (2) Perform Geodesic Shadow Removal smoothing;

[0062] (3) Crack enhancement processing and boundary detection.

[0063] Preferably, in step (6), the fracture parameters and lithology collected from the on-site images are input into the machine learning model to predict the mechanical strength of the coal and rock mass to be tested;

[0064] To improve the accuracy of the prediction, the prediction results are compared with the engineering site data, and the model parameters are automatically adjusted based on the feedback from the engineering site data to adapt to the new data input. This adaptive adjustment mechanism can ensure that the model is continuously optimized with the accumulation of engineering practice, improve the accuracy and reliability of the prediction, and meet the application needs in different engineering environments.

[0065] The beneficial effects of the present invention are:

[0066] (1) The present invention improves the traditional rock strength measurement method. The strength prediction method based on machine learning can combine the nonlinear relationship between the geometric characteristics of coal rock cracks of different lithologies and their strength to construct an accurate prediction model. This method does not rely on traditional mechanical tests, but uses crack characteristic data and lithology to train machine learning models, thereby realizing the prediction of the strength of fractured coal rock mass. Compared with traditional methods, the machine learning prediction method has a higher degree of automation and accuracy, and can extract characteristic patterns from a large amount of engineering site data and perform adaptive adjustments to achieve efficient prediction.

[0067] (2) The present invention realizes an efficient means of data acquisition. The geometric parameters of coal rock fractures can be automatically extracted through image recognition technology. These data will be automatically input into the machine learning model for training. Image recognition technology has the advantages of non-contact, high degree of automation, and fast processing speed. It can quickly and non-contactly obtain fracture parameters, provide rich data support for the training of machine learning models, greatly improve data collection efficiency, and reduce human intervention and measurement errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 1 is a reference schematic diagram of the values ​​of crack variable parameters in a uniaxial compression test in an embodiment of the present invention;

[0069] Figure 2 It is a flowchart of the present invention;

[0070] Figure 3 is the machine learning prediction network diagram of the present invention;

[0071] Figure 4 It is an implementation flow chart of the present invention. DETAILED DESCRIPTION

[0072] The present invention will be further described below by way of embodiments in conjunction with the accompanying drawings, but is not limited thereto.

[0073] Embodiment 1:

[0074] like Figure 1-Figure 4 As shown, this embodiment provides a method for intelligently predicting the strength of fractured coal and rock mass based on image recognition, and the steps are as follows:

[0075] (1) Conduct indoor tests based on different fracture parameters and lithology to establish an initial training data set. Specifically:

[0076] (11) Preparation of test specimens: rock blocks with different lithology and mechanical properties were collected from the project site and processed into 100 mm × 100 mm × 200 mm rectangular original rock specimens;

[0077] Cuboid casting specimens with different mechanical properties were cast using different proportions of cement, gypsum, river sand, borax and water, with a size of 100 mm × 100 mm × 200 mm;

[0078] Cast standard cylindrical specimens with a height of 100 mm and a diameter of 50 mm;

[0079] Some of the prepared rectangular original rock specimens were cored and subjected to uniaxial compression tests to obtain the mechanical properties of the specimens, including uniaxial compressive strength and elastic modulus, which were used for parameter calibration in subsequent simulations.

[0080] (12) preparing a large number of different through-cracks on the specimen obtained in step (11), wherein the crack characteristics include parameters such as length, width, angle, density, etc. of the cracks, and simulating the common coal rock crack morphology in engineering by precisely controlling these characteristics, thereby ensuring the diversity and representativeness of the test samples;

[0081] (13) Conduct uniaxial compression tests to measure the mechanical strength of the specimens under different fracture conditions. Record and organize the results of the uniaxial compression tests to form a data set containing fracture characteristic parameters, rock properties, and corresponding strength values. This data set is used as the initial training data set for the machine learning model to provide reliable input data for the model to learn fracture characteristics and rock properties.

[0082] (2) Based on the initial data set, numerical simulation data was introduced, and the discrete element numerical simulation method was used to simulate the strength of coal rock mass under different fracture shapes, densities and distribution conditions. A large amount of simulation data was used to supplement the laboratory data, and the data were integrated into a data set to construct a sample database of coal rock mass fractures and strength;

[0083] The specific methods of data integration are:

[0084] (21) Integrate the indoor test data obtained in step (1) and the coal rock mass fracture parameter, lithology and strength data generated by numerical simulation, and clean the integrated data to remove possible outliers, noise data and missing data to ensure data quality;

[0085] (22) Data completion. In order to ensure the integrity of the data and the reliability of the analysis, it is necessary to reasonably complete the missing data. For coal rock mass fracture parameters, if the missing proportion of the parameters is less than 5%, the mean or median method is used for completion. If the parameter characteristics show normal distribution data, the mean is used for completion. For data with extreme values ​​or skewed distributions exceeding 20%, the median is used to reduce the impact of extreme values. The mean completion formula is:

[0086]

[0087] Among them, x i is the non-missing value of the feature, N is the number of non-missing values;

[0088] (23) The fracture parameters, lithology and strength data of coal rock mass are standardized so that they are within the same numerical range to facilitate the input of the machine learning model. Standardization can effectively reduce the differences between parameters of different dimensions and ensure that the model can accurately learn and process the relationship between each fracture parameter and strength value.

[0089] (24) Classify and label the standardized data, and combine them into different strength grade labels according to the different fracture parameters and lithology of the coal and rock mass. Through labeling and classification, support is provided for the classification and prediction functions of the subsequent models.

[0090] (25) The classified and labeled data are entered into the database to construct a sample database of coal rock fracture parameters and strength. The database adopts a structured data storage method, has efficient data query and retrieval functions, and can dynamically update data to adapt to different engineering needs. The database can also be seamlessly connected with the machine learning system to provide real-time data support for the model. In order to maintain the timeliness and applicability of the database, the database is regularly maintained and updated. By continuously adding new data in subsequent engineering practices, the coverage of the database will be more extensive. At the same time, with the advancement of numerical simulation and indoor test technology, more complex fracture characteristic data will be gradually introduced to ensure the continuous expansion and optimization of the database.

[0091] (3) Use the sample database to train the machine learning model, and the machine learning model is trained using the Gradient Boosting Decision Tree (GBDT) model;

[0092] The steps to build a gradient boosting decision tree model are as follows:

[0093] (311) In GBDT, decision trees are usually selected as base learners. Each tree is fitted on the residual of the previous tree. By gradually correcting the error, the overall model accuracy is ultimately improved.

[0094] (312) Definition of loss function: The goal of the gradient boosting decision tree model is to minimize the loss function by gradient descent. For regression problems, the loss function uses the mean square error (MSE) or mean absolute error (MAE). The model calculates the gradient at each step to minimize the loss.

[0095] (313) Set model hyperparameters: set the learning rate to 0.001, the number of trees to 300, the maximum depth of the tree to 4, and the subsampling rate to 0.8.

[0096] The training steps of the gradient boosting decision tree model are as follows:

[0097] (321) Initialization model: At the beginning of training, the predicted value is initialized with the mean value of coal rock strength as the prediction result of the first weak learner. The calculation formula is:

[0098]

[0099] Where N is the number of samples in the training set, y i is the true target value of the sample;

[0100] (322) Calculate the residual: In each iteration t, calculate the residual r between the current model prediction value and the actual value i （t） , the residual represents the prediction error of the current model. For each sample i, the residual is defined as:

[0101]

[0102] in, is the predicted value of the model at round t-1, r i （t） is the target that needs to be fitted in this round;

[0103] In round t, the new predicted value Obtained by the following formula:

[0104]

[0105] Among them, η is the learning rate, which controls the step size of each iteration, and f t (x i) is the prediction result of the decision tree in round t, that is, the tree used to fit the residual;

[0106] (323) Fitting residuals: Taking the residuals as the target value, a new decision tree is trained to fit the current residuals. The prediction results of the decision tree are used to correct the prediction error of the model.

[0107] (324) Update the model prediction value: multiply the prediction result of the newly trained decision tree by the learning rate and add it to the overall prediction value of the model to gradually reduce the prediction error;

[0108] (325) Repeat steps (322)-(324) until a predetermined number of trees is reached or the loss function no longer decreases significantly;

[0109] (326) Hyperparameter tuning: After training is completed, the model’s hyperparameters (such as learning rate, number and depth of trees, etc.) are adjusted through grid search or random search to further improve model accuracy.

[0110] (4) Use the test set data to test the accuracy of the machine learning model. Specifically:

[0111] (41) Prepare the test set data, ensuring that the test set is independent of the training set data and contains data that the model has not seen. The test set should be representative and reflect the data distribution in actual applications. Use the same standardized parameters to process the test set data to ensure data consistency, input the feature data of the test set into the trained gradient boosting decision tree model, and obtain the predicted value of the gradient boosting decision tree model for each sample. ;

[0112] (42) Collect the true target value y of the test set and calculate the evaluation indicators mean square error MSE, mean absolute error MAE and determination coefficient R 2 ;

[0113] MSE is used to measure the square difference between the model prediction value and the true value. The smaller the MSE, the closer the model prediction is to the true value.

[0114]

[0115] Where N is the number of test set samples, y i is the true value of the sample, is the predicted value;

[0116] MAE is the absolute difference between the predicted value of the calculation model and the true value. MAE reflects the actual size of the model error.

[0117]

[0118] R 2It is an indicator to measure the degree of model fit, with a value range of 0 to 1. The closer the value is to 1, the stronger the ability of the model to explain the data variance;

[0119]

[0120] in, is the true mean of the test set, R 2 The higher it is, the better the model can explain the data.

[0121] (43) If the prediction accuracy requirement is met, the model training is completed and the prediction result is output; if not, return to step (3), adjust the model's hyperparameters, and train the model again until the prediction accuracy requirement is met.

[0122] (5) Image acquisition of working face: At the construction site, high-resolution image acquisition equipment is used to obtain the surface image of the coal and rock mass to be tested. The acquired image is pre-processed and crack boundary detection is performed to record the crack length, crack width, crack angle, and crack density of the coal and rock mass at the working face. The specific method is as follows:

[0123] (1) At the construction site, the surface image of the coal and rock mass to be tested was obtained by using a high-resolution image acquisition device, a Canon EOS 6D MARK Ⅱ full-frame SLR camera, equipped with a Canon 50mm f1.4 full-frame fixed-focus lens, a SLR camera tripod, etc. The aperture size was f=1.4, the sensitivity was iso=40000, and the exposure time was 1 / 100 s.

[0124] (2) A digital camera is placed 10 m in front of the work surface to take photos.

[0125] (3) The collected coal and rock mass images are subjected to image denoising and grayscale conversion processing in turn. The conversion formula is:

[0126]

[0127] Among them, R, G and B represent the pixel values ​​of the red, green and blue color channels respectively, and the converted grayscale value Gray Will be output as the final grayscale image;

[0128] (4) Perform Geodesic Shadow Removal smoothing, which includes five parts: mathematical morphology "closing", Gaussian smoothing, watershed, histogram equalization, contrast enhancement, and use the morphological "closing" function to eliminate cracks caused by the cracks having similar intensities to the shadow area. In the shadow area estimation, it can prevent the cracks from being counted as shadows. Apply a Gaussian smoothing filter to remove noise. The following three steps can be performed:

[0129] a. Use the watershed method to segment the image into a series of geodesic regions with different average illumination. ,by The area is shaded. The area is the background area. Here, .

[0130] b. Select the best reference area for brightness compensation and set a geodesic area as , the span of the grayscale value of a geodesic area is set to H, which means that the H value of the geodesic area corresponding to the penumbra area is usually larger than that of the background area. The image is roughly divided into the shadow area and the background area by the formula, where H S and H B Representing low and high value categories respectively, through the formula Selecting the best reference area ,in, and Respectively represent the mean and standard deviation of pixels in the image;

[0131] c. Pixels Assign to new value , where I i,j Represents the gray value at pixel (i, j), Indicates the result of brightness compensation, , , D S and D B denote the standard deviation of S and B respectively, and Represent the average gray value of S and B respectively. This algorithm can enhance the contrast of the image.

[0132] (5) Perform crack enhancement treatment and boundary detection, and record the crack characteristics of the coal and rock mass at the working face. The specific methods of crack enhancement treatment are as follows:

[0133] ① Get the Hessian matrix H of the image σ (x,y):

[0134]

[0135] Among them, G σ is a Gaussian filter of scale σ, I is the input image, and * represents the convolution operation;

[0136] ② For each pixel position (x, y) the Hessian matrix H σ (x, y) is subjected to eigenvalue decomposition to obtain eigenvalues ​​λ1 and λ2, where it is assumed that , for crack features, λ1 is usually larger and λ2 is smaller. Using this feature, the ratio of eigenvalues ​​is used to determine whether the position belongs to a crack;

[0137] ③ Use the enhanced response function based on the Hessian matrix to enhance the cracks. The commonly used response function is the FrangiVesselness Filter, which has a stronger response to slender structures and is defined as follows:

[0138]

[0139] in, is the eigenvalue ratio, which is used to measure the linearity of the feature. It indicates the strength and measures the significance of the structure. β and c are adjustment parameters, which are set to β=0.5 and c=15 through experiments.

[0140] ④Superimpose the crack enhancement response on the original image to obtain the enhanced image:

[0141]

[0142] Among them, I(x,y) is the original image, V final (x,y) is the enhancement response and α is the adjustment factor used to control the strength of the enhancement.

[0143] Boundary detection uses the Canny edge detection method, which usually performs better in complex images and is suitable for crack images with more noise. First, the gradient amplitude is calculated and direction At each pixel point, check whether its gradient amplitude is a local maximum along the gradient direction. If it is a local maximum, retain the pixel as an edge, otherwise suppress it. The non-maximum suppression formula is as follows:

[0144]

[0145] Define two thresholds, high and low, and mark pixels with gradient magnitudes higher than the high threshold as strong edges, and discard pixels with gradient magnitudes lower than the low threshold. Pixels between the high and low thresholds are retained as weak edges if they are connected to a strong edge. Connect the strong edge and the weak edge to get the final boundary.

[0146] (6) Input the fracture parameters and lithology collected from the field images into the machine learning model to predict the mechanical strength of the coal and rock mass to be tested;

[0147] To improve the accuracy of the prediction, the prediction results are compared with the engineering site data, and the model parameters are automatically adjusted based on the feedback from the engineering site data to adapt to the new data input. This adaptive adjustment mechanism can ensure that the model is continuously optimized with the accumulation of engineering practice, improve the accuracy and reliability of the prediction, and meet the application needs in different engineering environments.

Claims

1. An intelligent prediction method for fractured coal and rock mass strength based on image recognition, characterized in that: Here are the steps: (1) Conduct indoor experiments based on different fracture parameters and lithology to establish an initial training data set; (2) Based on the initial data set, numerical simulation data were introduced, and the discrete element numerical simulation method was used to simulate the strength of coal rock mass under different crack shapes, densities and distribution conditions. The laboratory data were supplemented by simulation data, and the data were integrated into a data set to construct a sample database of coal rock mass cracks and strength; (3) Use the sample database to train the machine learning model, and the machine learning model is trained using the gradient boosting decision tree model; (4) Use test set data to test the accuracy of the machine learning model; (5) Image acquisition of working face: At the construction site, high-resolution image acquisition equipment is used to obtain the surface image of the coal and rock mass to be tested. The acquired image is pre-processed and crack boundary detection is performed to record the crack length, crack width, crack angle, and crack density of the coal and rock mass at the working face. (6) Predict the strength of the coal and rock mass to be tested, and perform adaptive learning, adjustment, and optimization by combining the engineering site data.

2. The method for intelligent prediction of fractured coal and rock mass strength based on image recognition according to claim 1, characterized in that: In step (1), specifically: (11) Preparation of test specimens: rock blocks with different lithology and mechanical properties were collected from the project site and processed into 100 mm × 100 mm × 200 mm rectangular original rock specimens; Cuboid casting specimens with different mechanical properties were cast using different proportions of cement, gypsum, river sand, borax and water, with a size of 100 mm × 100 mm × 200 mm; Cast cylindrical specimens with a height of 100 mm and a diameter of 50 mm; Some of the prepared rectangular original rock specimens were cored and subjected to uniaxial compression tests to obtain the mechanical property parameters of the specimens, including uniaxial compressive strength and elastic modulus; (12) preparing different through-cracks on the specimen obtained in step (11), wherein the crack characteristics include the length, width, angle, and density of the cracks; (13) Conduct uniaxial compression tests to measure the mechanical strength of the specimens under different fracture conditions. Record and organize the results of the uniaxial compression tests to form a data set containing fracture characteristic parameters, lithology, and corresponding strength values. This data set is used as the initial training data set for the machine learning model.

3. The method for intelligent prediction of fractured coal and rock mass strength based on image recognition according to claim 2, characterized in that: In step (2), the specific method of data integration is: (21) Integrate the indoor test data obtained in step (1) and the coal rock mass fracture parameter, lithology and strength data generated by numerical simulation, and clean the integrated data; (22) Data completion. In order to ensure the integrity of the data and the reliability of the analysis, it is necessary to reasonably complete the missing data. For coal rock mass fracture parameters, if the missing proportion of the parameters is less than 5%, the mean or median method is used for completion. If the parameter characteristics show normal distribution data, the mean is used for completion. For data with extreme values ​​or skewed distributions exceeding 20%, the median is used to reduce the impact of extreme values. The mean completion formula is: Among them, x i is the non-missing value of the feature, N is the number of non-missing values; (23) Standardize the fracture parameters, lithology and strength data of coal and rock masses; (24) Classify and label the standardized data, and combine them into different strength grade labels according to the different fracture parameters and lithology of the coal and rock mass; (25) The classified and labeled data are entered into the database to construct a sample database of coal rock mass fracture parameters and strength.

4. The method for intelligent prediction of fractured coal and rock mass strength based on image recognition according to claim 3, characterized in that: In step (3), the steps for building the gradient boosting decision tree model are as follows: (311) Select decision tree as the base learner; (312) Definition of loss function: The goal of the gradient boosting decision tree model is to minimize the loss function by gradient descent. For regression problems, the loss function uses the mean square error (MSE) or mean absolute error (MAE). (313) Set model hyperparameters: learning rate, number of trees, maximum tree depth, subsampling rate.

5. The method for intelligent prediction of fractured coal and rock mass strength based on image recognition according to claim 4, characterized in that: In step (3), the training steps of the gradient boosting decision tree model are as follows: (321) Initialization model: At the beginning of training, the predicted value is initialized with the mean value of coal rock strength as the prediction result of the first weak learner. The calculation formula is: in, N is the number of samples in the training set, y i is the true target value of the sample; (322) Calculate the residual: In each iteration t, calculate the residual r between the current model prediction value and the actual value i （t） , the residual represents the prediction error of the current model. For each sample i, the residual is defined as: in, is the predicted value of the model at round t-1, r i （t） is the target that needs to be fitted in this round; In round t, the new predicted value Obtained by the following formula: Among them, η is the learning rate, which controls the step size of each iteration, and f t (x i ) is the prediction result of the decision tree in round t; (323) Fitting residuals: Take the residuals as the target value and train a new decision tree to fit the current residuals; (324) Update the model prediction value: multiply the prediction result of the newly trained decision tree by the learning rate and add it to the overall prediction value of the model to gradually reduce the prediction error; (325) Repeat steps (322)-(324) until a predetermined number of trees is reached or the loss function no longer decreases significantly; (326) Hyperparameter tuning: After training is complete, the model’s hyperparameters are adjusted through grid search or random search.

6. The method for intelligent prediction of fractured coal and rock mass strength based on image recognition according to claim 5, characterized in that: In step (4), specifically: (41) Input the feature data of the test set into the trained gradient boosting decision tree model to obtain the predicted value of each sample by the gradient boosting decision tree model. ; (42) Collect the true target value y of the test set and calculate the evaluation indicators mean square error MSE, mean absolute error MAE and determination coefficient R 2 ; MSE is used to measure the square difference between the model prediction value and the true value. The smaller the MSE, the closer the model prediction is to the true value. Where N is the number of test set samples, y i is the true value of the sample, is the predicted value; MAE is the absolute difference between the predicted value of the calculation model and the true value. MAE reflects the actual size of the model error. R 2 It is an indicator to measure the degree of model fit, with a value range of 0 to 1. The closer the value is to 1, the stronger the ability of the model to explain the data variance; in, is the true mean of the test set, R 2 The higher it is, the better the model can explain the data; (43) If the prediction accuracy requirement is met, the model training is completed and the prediction result is output; if not, return to step (3), adjust the model's hyperparameters, and train the model again until the prediction accuracy requirement is met.

7. The method for intelligent prediction of fractured coal and rock mass strength based on image recognition according to claim 6, characterized in that: In step (5), the collected image is preprocessed. The specific steps are as follows: (1) Perform image denoising and grayscale processing on the coal and rock mass images in sequence; (2) Perform Geodesic Shadow Removal smoothing; (3) Crack enhancement processing and boundary detection.

8. The method for intelligent prediction of fractured coal and rock mass strength based on image recognition according to claim 7, characterized in that: In step (6), the fracture parameters and lithology collected from the field images are input into the machine learning model to predict the mechanical strength of the coal rock mass to be tested; To improve the accuracy of the prediction, the prediction results are compared with the engineering site data, and based on the feedback from the engineering site data, the parameters of the model are automatically adjusted to adapt to the new data input.

Images