A rock breaking efficiency prediction method based on image recognition and machine learning

Abstract

A rock breaking efficiency prediction method based on image recognition and machine learning, acquires a rock surface image, and applies image processing technology for grayscale processing, then performs pixel classification and image segmentation to clearly divide different minerals into several independent structure graphs; modeling the rock material according to the mineral composition, and importing the boundary information of the several structure graphs into the model to generate several large particle filling rock blocks with the same size as the several structure graphs, and calibrating the mesoscopic parameter values according to the rock mechanics characteristics; a disc cutter rock breaking discrete element calculation model is established, and multiple intrusion rock breaking simulation tests are carried out on rock samples with different mineral contents; the simulation results are collected to construct a rock breaking efficiency data set; a prediction model is established based on the PSO-BPNN algorithm, and after training, the rock breaking efficiency prediction model is obtained; the established model is used for real-time prediction of rock breaking efficiency. The method can consider the influence of rock mineral composition on disc cutter rock breaking effect, and realize efficient and accurate prediction of disc cutter rock breaking efficiency by combining rock micro composition with macro rock breaking behavior.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent control, and particularly relates to a rock breaking efficiency prediction method based on image recognition and machine learning. BACKGROUND

[0002] A tunnel boring machine (TBM) has been widely applied in urban rail transit, underground engineering and mining, etc. Its efficient tunneling capacity makes it the main equipment for modern tunnel construction. A cutter is one of the core components of the TBM, and plays a crucial role in its tunneling efficiency and overall performance. The design, material selection and wear characteristics of the cutter directly affect the rock breaking capacity and construction safety of the TBM. However, when the TBM faces complex geological conditions, its rock breaking efficiency is comprehensively affected by multiple factors such as rock composition, structure and cutter parameters. Accurate prediction of the rock breaking efficiency is of great significance to improving construction efficiency, reducing costs and ensuring construction safety. Based on image recognition technology to extract rock composition, a large amount of rock breaking data is generated through numerical simulation, and machine learning methods are used to train and analyze these data to construct a rock breaking efficiency prediction model, which is of great significance to improving the tunneling efficiency.

[0003] The existing research methods usually analyze rock composition or tunneling parameters in isolation, and lack in-depth research on the interaction between the two. At present, there is no comprehensive rock breaking efficiency prediction method combining image recognition, numerical simulation and machine learning to support the whole process application from stratum exploration analysis to rock breaking efficiency prediction. Therefore, it is urgent to provide a calculation method integrating stratum lithology recognition technology and rock breaking effect evaluation to more comprehensively realize intelligent prediction of rock breaking efficiency and provide a scientific basis for actual engineering construction parameter optimization. SUMMARY

[0004] In view of the problems existing in the prior art, the application provides a rock breaking efficiency prediction method based on image recognition and machine learning, which can effectively solve the problem that the existing rock breaking technology cannot accurately design construction parameters for different rock types. By analyzing the influence of rock mineral composition on cutter rock breaking, the method can combine rock micro composition and macro rock breaking behavior to efficiently and accurately predict rock breaking efficiency.

[0005] To achieve the above purpose, the application provides a rock breaking efficiency prediction method based on image recognition and machine learning, comprising the following steps:

[0006] Step 1: Obtain a rock surface image, and apply image processing technology to perform grayscale processing on the obtained rock surface image, and then perform pixel classification and image segmentation to clearly divide different minerals into independent structural diagrams;

[0007] Step two: model the rock material based on the Voronoi Grain-Based Model model, and import the boundary information of several structure diagrams into the model to generate several large particle-filled rock blocks of the same size as the several structure diagrams, and connect the contact polygon centroids to form mineral crystal boundaries, while filling smaller particles in the polygon area; in the numerical simulation software, group the large particle-filled rock blocks according to the defined mineral layers, divide and adjust the mesoscopic mechanical properties of different mineral types, and calibrate a set of most suitable mesoscopic parameter values to accurately reflect the composition and microstructure characteristics of the rock;

[0008] Step three: establish a rock breaking model of a rolling cutter, use the GBM model to randomly generate several division areas filled with particles according to the particle size of the rock minerals, and fill the division areas with spherical particles, group and name each division area according to the content proportion of the mineral composition of the rock sample, and then assign different mineral compositions and corresponding mesoscopic parameters according to the grouping results; after the particles are generated, apply confining pressure to them, use a wall to simulate the cross-sectional shape of the rolling cutter, and conduct multiple intrusion simulation tests on rock samples with different mineral contents, and change the mineral content, rolling cutter spacing, and penetration depth to sequentially combine the intrusion simulation tests to achieve the destruction within the mineral crystals;

[0009] Step four: collect the simulation results of each simulation according to the numerical simulation process, record the mineral content of the rock sample, the rolling cutter spacing, the penetration depth, and the rock breaking efficiency, and construct a data set for subsequent analysis and modeling;

[0010] Step five: establish a prediction model based on the PSO-BPNN algorithm; divide the constructed data set into a training set and a test set according to a set proportion, use the training set to train the back propagation neural network based on particle swarm optimization, and at the same time, use the PSO algorithm to optimize the weights and biases of the back propagation neural network to enable it to more effectively learn the complex relationship between the input and the output during the training process; by adjusting the speed and position of the particles, the PSO algorithm constantly seeks the optimal solution of the fitness, thereby improving the performance of the model and obtaining a rock breaking efficiency prediction model; input the test data into the rock breaking efficiency prediction model to obtain the corresponding rock breaking efficiency prediction results, and use the set evaluation index to evaluate the prediction accuracy and reliability of the model;

[0011] Step six: prediction of rock breaking efficiency; during the rock breaking operation, real-time collection of rock surface images is performed, the rock surface images are preprocessed and then input into the rock breaking efficiency prediction model, the rock breaking efficiency prediction model is used to predict the rock breaking efficiency, and the prediction results are output.

[0012] As a preferred, in step five, the data set is randomly divided into a training set and a test set according to a ratio of 7:3.

[0013] As a preferred, in step one, the rock surface image is grayed by using Image J digital image processing analysis software.

[0014] As a preferred, in step two, the mesoscopic parameter value is calibrated by using "trial and error calibration method" and through continuous adjustment and trial calculation.

[0015] As a preferred, in step three, the rock mineral content, cutter spacing and penetration in each simulation process are collected, and the specific energy SE is calculated according to formula (1) by using the work done by the cutter and the rock breaking area;

[0016]

[0017] In the formula, W is the work done by the cutter in the process of breaking rock, i.e. external force work; V is the volume of rock produced by the rock breaking test; Wv is the work done by the penetration force of the cutter; W r is the work done by the rolling force of the cutter, when only considering the invasion of the cutter head, W r = 0; S is the rock breaking area;

[0018] As a preferred, in step five, the process of obtaining the corresponding rock breaking efficiency prediction result is as follows:

[0019] A1: obtaining the output h of the hidden layer according to formula (2);

[0020] h = f(W ih ·X + b h ) (2);

[0021] In the formula, W ih is the weight matrix from the input layer to the hidden layer, X is the input vector, b h is the bias of the hidden layer, and f(·) is the activation function tansing, whose output range is [-1, 1];

[0022] A2: obtaining the prediction result y of the output layer according to formula (3);

[0023] y = g(W ho ·h + b o ) (3);

[0024] In the formula, W ho is the weight matrix from the hidden layer to the output layer, b o is the bias of the output layer, and g(·) is the activation function purelin of the output layer, whose output range is (-∞, +∞).

[0025] As a preferred, in step five, the process of optimizing the weights of the back propagation neural network by using PSO algorithm is as follows:

[0026] B1: Calculate the mean square error L according to formula (4);

[0027]

[0028] where N is the number of test samples, y i is the actual output, is the predicted output;

[0029] B2: Derive formula (4) to obtain the gradient of the loss with respect to the weight according to formula (5)

[0030]

[0031] where, is the gradient of the loss with respect to the output, is the gradient of the output with respect to the weight;

[0032] B3: Obtain the updated weight W according to formula (6);

[0033]

[0034] where η is the learning rate, which controls the step size of weight update.

[0035] As a preferred, in step five, the process of adjusting the velocity and position of the particle is as follows:

[0036] C1: Update the velocity of the particle according to formula (7)

[0037]

[0038] where, is the velocity of particle i at time t, w is the inertia weight, which controls the inertia of the particle motion, c1 is the individual learning factor, which controls the degree of movement of the particle to its best position, c2 is the group learning factor, which controls the degree of movement of the particle to the group best position, r1 and r2 are random numbers in the range of [0, 1], p i is the best position of particle i, is the position of particle i at time t, g is the best position of the group;

[0039] C2: Update the position of the particle according to formula (8)

[0040]

[0041] In the present application, the rock surface image obtained is processed based on image processing technology, different mineral types of particles can be grouped through preliminary analysis, so that several structure diagrams can be accurately obtained, which is beneficial to subsequent construction of an accurate rock sample model; the rock material is modeled by using numerical simulation, and the mesoscopic parameter value is calibrated, so that the rock sample model constructed can accurately reflect the composition and microstructure characteristics of the rock; the rolling cutter rock breaking model is used to perform rolling cutter penetration test based on the calibrated mesoscopic parameters, and the rock mineral content, rolling cutter spacing and penetration depth are combined for test, so that a large amount of numerical simulation data can be accurately obtained, which is beneficial to construction of a high-precision rock breaking efficiency prediction model. Machine learning technology is used to learn and train the obtained numerical simulation data, which can effectively capture the complex relationship between rock composition, cutter spacing, penetration and rock breaking efficiency, so as to realize more accurate rock breaking efficiency prediction and provide a scientific basis for optimization of penetration and other construction parameters.

[0042] The present application provides a comprehensive and efficient rock breaking efficiency prediction method based on image recognition and machine learning, which effectively solves the problem that the existing rock breaking technology cannot accurately design construction parameters for different rock types. By analyzing the influence of rock mineral composition on rolling cutter rock breaking, the method can combine rock microstructure and macro rock breaking behavior to accurately predict rock breaking efficiency. The method not only can more accurately predict rock breaking efficiency, but also can further optimize the cutter spacing and penetration to achieve higher construction efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The flowchart of the present application;

[0044] Figure 2 The rock surface image obtained in the present application is shown in the schematic diagram;

[0045] Figure 3 The structure diagram of the rock surface image in the present application;

[0046] Figure 4 The modeling diagram of the rock structure in the present application;

[0047] Figure 5 The structure schematic diagram before the rolling cutter rock breaking in the present application;

[0048] Figure 6 The structure schematic diagram after the rolling cutter rock breaking in the present application;

[0049] Figure 7 The fitting effect diagram of the BP neural network in the present application. DETAILED DESCRIPTION

[0050] The invention will now be further described with reference to the accompanying drawings.

[0051] like Figure 1 As shown, this invention provides a method for predicting rock-breaking efficiency based on image recognition and machine learning, comprising the following steps:

[0052] Step 1: As Figure 2 As shown, rock surface images are acquired, and image processing techniques are applied to convert the acquired rock surface images to grayscale. Then, pixel classification and image segmentation are performed to clearly divide different minerals into several independent structural images, such as... Figure 3 As shown;

[0053] Specifically, high-definition cameras are used to capture images of the rock surface. Based on this, digital image processing technology is introduced to conduct a preliminary analysis of the rock sample's composition without the need for additional physical information, and to determine the mineral composition of the rock sample. The obtained different composition maps are then used in the subsequent numerical simulation modeling process, which helps to increase the overall reference value of the subsequent model.

[0054] Step Two: As Figure 4 As shown, the rock material is modeled based on the Voronoi Grain-Based Model (GBM). The boundary information of several structural diagrams is imported into the model to generate several large-grain-filled rock blocks of the same size as the structural diagrams. These blocks are then connected to the centroids of contact polygons to form mineral crystal boundaries. Simultaneously, smaller grains are filled into the polygonal regions. In the numerical simulation software, the large-grain-filled rock blocks are grouped according to the defined mineral layers. The micromechanical properties of different mineral types are divided and adjusted to calibrate a set of the most suitable micro-parameter values ​​to accurately reflect the composition and microstructure characteristics of the rock.

[0055] Specifically, the boundary information of several structural diagrams is imported into PFC2D. For the generated large-particle-filled rock blocks, the light gray particles are feldspar, the gray particles are quartz, and the black particles are mica. As a further optimization, the contents of feldspar, quartz, and mica are 60%, 35.3%, and 4.7%, respectively.

[0056] Step three: Establish a rock breaking model of a rolling cutter, use the GBM model to randomly generate a number of division areas filled with particles according to the particle size of the rock minerals, and fill the division areas with spherical particles, and respectively group and name each division area according to the content proportion of the mineral composition of the rock sample, and then respectively assign different component mineral compositions and corresponding mesoscopic parameters according to the grouping results; After the particles are generated, a confining pressure is applied to them, a wall is used to simulate the cross-sectional shape of the rolling cutter, and a plurality of intrusion simulation tests are performed on rock samples with different mineral contents, and the mineral content, rolling cutter spacing, and penetration depth are changed in turn to combine the intrusion simulation tests to achieve the destruction of the mineral crystals, as shown in Figure 5 and Figure 6 ;

[0057] Step four: According to the numerical simulation process, collect the simulation results each time, record the mineral content of the rock sample, the rolling cutter spacing, the penetration depth, and the rock breaking efficiency, and build a data set for subsequent analysis and modeling; wherein the mineral content includes the content proportion of corresponding feldspar, quartz, mica, etc.

[0058] As a further optimization, the data set is normalized using a normalization function to normalize the data to the range of 0-1 to improve the training effect, and the normalized parameters are saved for subsequent reverse normalization processing;

[0059] Step five: Establish a prediction model based on the PSO-BPNN algorithm, which is a hybrid algorithm combining the particle swarm optimization algorithm (PSO, Particle Swarm Optimization) and the backpropagation neural network (BPNN, Backpropagation Neural Network). PSO is a swarm intelligence-based optimization algorithm that simulates the behavior of bird foraging, and particles find the global optimal solution through mutual cooperation and information sharing. BPNN is a commonly used artificial neural network that adjusts parameters through error backpropagation algorithm to optimize network weights and biases, thereby minimizing output error. By combining PSO and BPNN, the PSO-BPNN algorithm can effectively improve the accuracy of model prediction; in PSO-BPNN, the PSO algorithm is used to optimize the initial weights and biases of BPNN to avoid local optimal problems caused by random initialization, and then BPNN is trained and predicted to ensure that the obtained model has excellent prediction accuracy.

[0060] For the optimization problem of the cutter rock breaking process, the number of input layer neurons corresponds to the rock mineral content, cutter spacing and penetration depth, and the output layer has one neuron, which outputs the rock breaking efficiency as the optimization target. In the three-layer neural network, the number of hidden layer neurons is initially determined based on an empirical formula, and the performance of the model corresponding to different numbers of hidden layer neurons is evaluated by mean square error (MSE). Based on this, the optimal number of hidden layer neurons is determined by comparing the training error (MSE) by cycling different numbers of hidden layer nodes. The optimal number of hidden layer neurons is a certain value (which can be determined by tuning), and finally the optimal number of hidden layer nodes is used to build the final BP neural network.

[0061] The specific formula for the number of hidden layer neurons is: where hiddennum is the number of hidden layer neurons, n is the number of input layer neurons, preferably 1, m is the number of output layer neurons, preferably 5, and a is a constant with a value range of 1-10.

[0062] The constructed data set is divided into training set and test set according to the set proportion. The training set is used to provide data for the model to learn, so as to obtain the required machine learning prediction model, i.e. the rock breaking efficiency prediction model. The test set is used to judge the accuracy of the generated prediction model.

[0063] The training set is used to train the back propagation neural network (BPNN) based on particle swarm optimization (PSO). At the same time, the PSO algorithm is used to optimize the weights and biases of the back propagation neural network, so that it can more effectively learn the complex relationship between the input and output during the training process. By adjusting the speed and position of the particles, the PSO algorithm constantly seeks the optimal solution of the fitness, thereby improving the performance of the model and obtaining the rock breaking efficiency prediction model. The test data is input into the rock breaking efficiency prediction model to obtain the corresponding rock breaking efficiency prediction result, and the set evaluation index is used to evaluate the prediction accuracy and reliability of the model.

[0064] Step six: prediction of rock breaking efficiency. During the rock breaking operation, the rock surface image is collected in real time. After preprocessing the rock surface image, it is input into the rock breaking efficiency prediction model to realize the prediction of the rock breaking efficiency of the cutter and output the prediction result.

[0065] As a preferred, in step five, the data set is randomly divided into training set and test set according to the ratio of 7:3.

[0066] As a preferred, in step one, the rock surface image is grayed by using Image J digital image processing analysis software.

[0067] The particle contact model between and among minerals is PBM (Parallel-bonded model), and the mesoscopic mechanical properties of different mineral types are divided. As a preferred, in step two, the mesoscopic parameter value is calibrated by using the "trial and error calibration method" and through continuous adjustment and trial calculation.

[0068] As a preferred, in step three, the rock mineral content, cutter spacing and penetration of each simulation process are collected, and the specific energy SE is calculated by using the cutter work and rock breaking area according to formula (1); the specific energy is the power required for cutting unit volume of rock, which can be used to evaluate the rock breaking efficiency of the cutter, and the smaller the specific energy of the cutter for breaking unit volume of rock, the higher the rock breaking efficiency;

[0069]

[0070] In the formula, W is the work done by the cutter in the process of breaking rock, that is, the external force work; V is the volume of rock produced by the rock breaking test; Wv is the work done by the penetration force of the cutter; W r is the work done by the rolling force of the cutter, when only the cutter head invasion is considered, W r = 0; S is the rock breaking area;

[0071] Forward propagation is the way adopted by the neural network in the process of calculating the output, which converts the input data into the final output through the layers of the network step by step, and through the introduction of the activation function, the neural network can model complex nonlinear relationships. As a preferred, in step five, the process of obtaining the corresponding rock breaking efficiency prediction result is as follows:

[0072] A1: obtain the output h of the hidden layer according to formula (2);

[0073] h = f (W ih ·X + b h ) (2);

[0074] In the formula, W ih is the weight matrix from the input layer to the hidden layer, X is the input vector, b h is the bias of the hidden layer, and f(·) is the activation function tansing, whose output range is [-1, 1]; tansig helps the model to learn nonlinear relationships and is suitable for the hidden layer.

[0075] A2: obtain the prediction result y of the output layer according to formula (3);

[0076] y = g (W ho ·h + bo ) (3) ;

[0077] where W ho is the weight matrix from the hidden layer to the output layer, b o is the bias of the output layer, and g(·) is the activation function purelin of the output layer with an output range (-∞, +∞). The purelin directly outputs the predicted value, which is suitable for the output layer of the regression task. The activation function not only introduces non-linear characteristics to the neural network, but also affects the training and performance of the network. Reasonable selection and application of the activation function is an important part of building an efficient neural network.

[0078] Backpropagation is a process of updating network weights by calculating the gradient of the loss function. The process of calculating the gradient by the mean square error is one of the core steps of the backpropagation algorithm. Specifically, in the neural network, the predicted output of the network is first obtained by forward propagation, and compared with the actual target output. The mean square error as a loss function can be used to quantify the difference between the predicted result and the actual result. As a preferred, in step five, the process of optimizing the weights of the backpropagation neural network by using the PSO algorithm is as follows:

[0079] B1: Calculate the mean square error L according to formula (4);

[0080]

[0081] where N is the number of test samples, y i is the actual output, is the predicted output;

[0082] B2: By taking the derivative of the mean square error, the gradient of the loss function with respect to each weight can be obtained. The gradient represents the sensitivity of the loss function to the change of the weight, that is, under the current weight setting, what effect will the increase or decrease of the weight have on the loss function. Taking the derivative of formula (4), the gradient of the loss with respect to the weight is obtained according to formula (5)

[0083]

[0084] where is the gradient of the loss with respect to the output, is the gradient of the output with respect to the weight;

[0085] B3: After obtaining the gradient, the weight can be updated according to the gradient descent method. Specifically, the updated weight W is obtained according to formula (6). The weight W is a parameter that connects neurons and determines the strength of signal transmission;

[0086]

[0087] In the formula, η is the learning rate, which controls the step size of weight updates. This process allows the model to gradually adjust the weights in each iteration to reduce prediction errors, thereby effectively improving the model's accuracy and generalization ability. By repeatedly executing this process, the neural network can continuously optimize its parameters, ultimately achieving superior performance.

[0088] Using the particle's current velocity and historical best position; as a preferred option, in step five, the process of adjusting the particle's velocity and position is as follows:

[0089] C1: Update the particle velocity according to formula (7)

[0090]

[0091] In the formula, Let be the velocity of particle i at time t, w be the inertia weight controlling the particle's motion inertia, c1 be the individual learning factor controlling the degree to which the particle moves towards its optimal position, c2 be the group learning factor controlling the degree to which the particle moves towards the group's optimal position, r1 and r2 be random numbers in the range [0, 1], and p be the velocity of particle i at time t. i For the optimal position of particle i, Let be the position of particle i at time t, and g be the optimal position of the group;

[0092] C2: Update the particle position according to formula (8)

[0093]

[0094] This invention continuously seeks the globally optimal particle position by iteratively updating the particle's velocity and position, and records the convergence curve. The error of the PSO-BP neural network is calculated, simulation predictions are performed on the test set, and the normalized data is denormalized back to the actual values. The fitting effect is as follows: Figure 7 As shown, the regression analysis graph is divided into training, validation, testing, and overall fit effects. The horizontal axis represents the "target value," indicating the actual output of the data, while the vertical axis represents the "output," indicating the model's predicted output value. The solid line represents the regression line between the predicted and actual results, while the dashed line represents the ideal Y=T line, i.e., the case where the predicted result is completely consistent with the actual result. The correlation coefficient R in the graph represents the degree of fit on the training set; a value close to 1 indicates that the model fits well on the training set.

[0095] In the present application, the rock surface image is processed based on image processing technology, different mineral types of particles can be grouped through preliminary analysis, so that several structure diagrams can be accurately obtained, which is beneficial to subsequent construction of accurate rock sample model; numerical simulation is used to model the rock material and calibrate the mesoscopic parameter value, so that the constructed rock sample model can accurately reflect the composition and microstructure characteristics of the rock; the rolling cutter rock breaking model is used to perform rolling cutter penetration test based on the calibrated mesoscopic parameters, and the rock mineral content, rolling cutter spacing and penetration depth are changed for combination test, so that a large amount of numerical simulation data can be accurately obtained, which is beneficial to construction of high-precision rock breaking efficiency prediction model. Machine learning technology is used to learn and train the obtained numerical simulation data, which can effectively capture the complex relationship between rock composition, cutter spacing, penetration and rock breaking efficiency, so as to realize more accurate rock breaking efficiency prediction and provide a scientific basis for optimizing the penetration and other construction parameters.

[0096] The present application provides a comprehensive and efficient rock breaking efficiency prediction method based on image recognition and machine learning, which effectively solves the problem that the existing rock breaking technology cannot accurately design construction parameters for different rock types, and can combine rock micro composition and macro rock breaking behavior to accurately predict rock breaking efficiency. This method not only can more accurately predict rock breaking efficiency, but also can further optimize the cutter spacing and penetration according to the prediction results to achieve higher construction rock breaking efficiency.

Claims

1. A rock breaking efficiency prediction method based on image recognition and machine learning, characterized in that, The method comprises the following steps: Step one: obtaining a rock surface image, and applying image processing technology to perform grayscale processing on the obtained rock surface image, then performing pixel classification and image segmentation to clearly divide different minerals into independent structure graphs; Step two: modeling the rock material based on a Voronoi Grain-Based Model, importing the boundary information of the structure graphs into the model, generating a plurality of large particle filled rock blocks with the same size as the structure graphs, connecting the polygon centroids to form mineral crystal boundaries, and filling smaller particles in the polygon area; grouping the large particle filled rock blocks according to the defined mineral layers in the numerical simulation software, dividing and adjusting the mesoscopic mechanical properties of different mineral types, and calibrating a set of most suitable mesoscopic parameter values to accurately reflect the composition and microstructure characteristics of the rock; Step three: establishing a rock breaking model, using the GBM model to randomly generate a plurality of division regions filled with particles according to the particle size of the rock minerals, and filling the division regions with spherical particles, grouping and naming each division region according to the content proportion of the rock sample mineral composition, and then assigning different mineral compositions and corresponding mesoscopic parameters according to the grouping results; after the particles are generated, a confining pressure is applied to the particles, the wall body simulates the cross-sectional shape of the roller cutter, and a plurality of intrusion simulation tests are performed on rock samples with different mineral contents, the rock mineral content, the roller cutter spacing and the penetration depth are changed, and the intrusion simulation tests are sequentially combined to realize the damage within the mineral crystal; Step four: collecting the simulation results of each simulation, recording the mineral content of the rock sample, the roller cutter spacing, the penetration depth and the rock breaking efficiency, and constructing a data set for subsequent analysis and modeling; Step five: establishing a prediction model based on the PSO-BPNN algorithm; dividing the constructed data set into a training set and a test set according to a set proportion, training the back propagation neural network based on the particle swarm optimization using the training set, and using the PSO algorithm to optimize the weights and biases of the back propagation neural network to enable it to more effectively learn the complex relationship between the input and the output during the training process; by adjusting the speed and position of the particles, the PSO algorithm constantly seeks the optimal solution of the fitness, thereby improving the performance of the model, and obtaining a rock breaking efficiency prediction model; inputting the test data into the rock breaking efficiency prediction model to obtain the corresponding rock breaking efficiency prediction result, and evaluating the prediction accuracy and reliability of the model using the set evaluation index; Step six: predicting the rock breaking efficiency; During the rock breaking operation, the rock surface image is collected in real time, and after the rock surface image is preprocessed, it is input into the rock breaking efficiency prediction model to realize the prediction of the rock breaking efficiency, and the prediction result is output. 2.The rock breaking efficiency prediction method based on image recognition and machine learning according to claim 1, characterized in that, In step five, the data set is randomly divided into a training set and a test set according to a 7:3 ratio. 3.The rock breaking efficiency prediction method based on image recognition and machine learning according to claim 1, characterized in that, In step one, the Image J digital image processing and analysis software is used to perform grayscale processing on the rock surface image. 4.The rock breaking efficiency prediction method based on image recognition and machine learning according to claim 1, characterized in that, In step two, the "trial and error calibration method" is used, and the mesoscopic parameter value is calibrated through continuous adjustment and trial calculation. 5.The rock breaking efficiency prediction method based on image recognition and machine learning according to claim 1, characterized in that, In step three, the content of each mineral of rock, the distance between the cutters and the penetration during each simulation process are collected, and the specific energy is calculated according to formula (1) using the work done by the cutters and the broken rock area ; （1）； wherein is the work done by the tool during rock breaking, i.e. the external force work; is the volume of rock produced by the rock breaking test; is the work done by the penetration force of the roller cutter; is the work done by the rolling force of the roller cutter, when only the effect of the cutter head penetration is considered, ; is the rock breaking area.

6. The rock breaking efficiency prediction method based on image recognition and machine learning according to claim 1, characterized in that, In step five, the process of obtaining the corresponding rock breaking efficiency prediction result is as follows: A1 : Obtain the output of the hidden layer according to formula (2) ; （2）； wherein, is a weight matrix from the input layer to the hidden layer, is an input vector, is a bias of the hidden layer, is an activation function tansing with an output range [-1, 1]; A2: Obtain the prediction result of the output layer according to formula (3) ; （3）； wherein, is a weight matrix from the hidden layer to the output layer, is a bias for the output layer, is an activation function purelin for the output layer with an output range .

7. The rock breaking efficiency prediction method based on image recognition and machine learning according to claim 1, characterized in that, In step five, the process of optimizing the weights of the back propagation neural network using the PSO algorithm is as follows: B1 : Calculate the mean square error according to formula (4) ; （4）； In the formula, N is the number of test samples, Y is the actual output, Y is the predicted output; B2: Derive the formula (4), and obtain the gradient of the loss to the weight according to the formula (5) ; （5）； wherein is the loss on the output, is the gradient of the output on the weights; B3: Obtain updated weights according to formula (6) ; （6）； In the formula, is a learning rate that controls the step size of the weight updates. 8.The rock breaking efficiency prediction method based on image recognition and machine learning according to claim 1, characterized in that, In step five, the process of adjusting the velocity and position of the particle is as follows: C1 : Update the velocity of the particle according to equation (7) ; （7）； wherein is the particle at time is the velocity of the particle at time is the inertia weight that controls the movement of the particle, is the individual learning factor that controls the extent to which the particle moves towards its best position, is the group learning factor that controls the extent to which the particle moves towards the group best position, and is a random number in the range [0, 1], is the best position of the particle , is the position of the particle at time , is the best position of the group; C2: update the position of the particle according to formula (8) ; （8）。

Images