Intelligent identification method for tmb surrounding rock classification based on penetration index and bayesian enhancement

Abstract

The application discloses a TBM surrounding rock classification intelligent identification method based on a penetration index and a Bayesian enhancement, and comprises the following steps: 1, surrounding rock classification; 2, constructing a primary surrounding rock classification model based on a penetration index; 3, constructing a primary surrounding rock classification model sample library; 4, training the primary surrounding rock classification model; 5, testing the primary surrounding rock classification model; 6, calculating a prior probability; 7, calculating a conditional probability; 8, tunnel excavation; 9, calculating a penetration index of an excavation section; 10, primary prediction of surrounding rock classification of the excavation section; and 11, Bayesian posterior estimation. The application can be calculated in real time according to on-site machine rock breaking and excavation monitoring data, can effectively reduce the cost of on-site geological exploration and indoor test, and the intelligent prediction method based on the Bayesian enhancement can obtain a posterior result of a working face position based on the prediction of a section that has been excavated, and further improve the credibility of the prediction result.

Description

Intelligent Identification Method for TBM Surrounding Rock Classification Based on Penetration Index and Bayesian Enhancement Technical Field

[0001] This invention relates to the field of TBM construction technology applications, and in particular to an intelligent identification method for TBM surrounding rock classification based on penetration index and Bayesian enhancement. Background Technology

[0002] Tunnel boring machines (TBMs) are widely used in urban underground space engineering and tunnel construction in many deep-buried areas in my country. Because the cutterhead and shield of a TBM provide a closed working space for workers during excavation, it offers a safe and friendly construction environment. However, the cutterhead and shield also hinder on-site geological engineers from conducting geological surveys and mapping of the tunnel surrounding rock, especially at the tunnel face. This leads to significant delays in face stability assessments and surrounding rock support classification. Furthermore, TBMs are highly sensitive to changes in the geological environment. Failure to promptly and accurately identify adverse geological conditions and their types at the tunnel face can easily lead to engineering disasters such as machine jamming, while also increasing the time required for on-site support material preparation and coordination, and reducing construction efficiency.

[0003] Traditional geological surveys of the TBM excavation face primarily rely on advanced drilling and geological mapping of the excavation face during intermittent machine downtime to obtain rock physical, mechanical, and joint parameters. These methods depend on laboratory physical and mechanical strength testing, which is time-consuming and costly, and involves lengthy result analysis, failing to meet the needs of rapid on-site diagnosis and prediction in tunnel construction. Furthermore, these tests can severely disrupt the normal TBM tunneling process, reducing the machine's excavation speed.

[0004] Therefore, there is an urgent need to develop a method for classifying and identifying the surrounding rock at the tunnel face during TBM construction, so as to achieve low-cost, high-quality, and rapid identification of the surrounding rock while minimizing the impact on the construction process. Summary of the Invention

[0005] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by providing a TBM (Tunnel Boring Machine) surrounding rock classification and identification method based on penetration index and Bayesian enhancement. This method first extracts a penetration index reflecting the difficulty of rock excavation based on the physical laws governing the interaction between the cutterhead and the rock mass during TBM construction. This index can be calculated in real-time based on on-site machine rock-breaking and tunneling monitoring data, effectively reducing the costs of on-site geological exploration and laboratory testing. Furthermore, the index extraction method from on-site monitoring data minimizes its impact on the construction process. In addition, to further eliminate the influence of monitoring data quality, a Bayesian enhancement-based intelligent prediction method can obtain a posteriori result of the tunnel face location based on predictions of the already excavated section, thereby improving the reliability of the prediction results. This technical solution can quickly and accurately identify the surrounding rock classification at the tunnel face during TBM construction and excavation, ensuring construction safety and guiding the selection of on-site support measures.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A smart identification method for TBM surrounding rock classification based on penetration index and Bayesian enhancement includes the following steps.

[0008] Step 1: Classification of surrounding rock: Based on rock strength, integrity, structural surface condition, occurrence, and groundwater conditions, the surrounding rock is classified into P types; then, each type of surrounding rock is labeled with a numerical classification tag C.

[0009] Step 2: Construct a primary rock classification model based on penetration indices: For each tunneling segment during TBM construction, a primary rock classification model based on penetration indices is constructed using the random forest algorithm; the input layer of the primary rock classification model is the penetration index vector of the tunneling segment during TBM construction. The output layer is the surrounding rock classification label C for the corresponding tunneling section; where a is the linear fitting slope of the cutterhead single cutter thrust and penetration in the tunneling section; b is the initial penetration thrust of the cutterhead roller in the tunneling section. The linear fit between the single cutter thrust and penetration depth in the tunneling section is represented by TPI; TPI is the linear fit slope between the single cutter torque and penetration depth in the tunneling section. This indicates the goodness of linear fit between the single-blade torque and penetration depth in the tunneling section.

[0010] Step 3: Construct a primary rock classification model sample library, which includes the following steps:

[0011] Step 3-1: Collect effective tunneling segment data: Collect effective tunneling segment data for all types of surrounding rock in Step 1; wherein, the number of effective tunneling segments for each type of surrounding rock is not less than M; M≥100.

[0012] Step 3-2: Extracting Penetration Indicators for Each Excavation Section: For each excavation section of each type of surrounding rock, the penetration index vector is extracted, and the actual surrounding rock classification label C corresponding to each excavation section is recorded, thus obtaining a primary surrounding rock classification model sample library containing N sample data; each sample data includes...

[0013] Step 3-3, Sample Library Classification: Divide the sample library into two parts: training set and test set.

[0014] Step 4: Train the primary rock classification model: Use the training set obtained in Step 3 to train the primary rock classification model constructed in Step 2 to obtain the trained primary rock classification model.

[0015] Step 5: Test the surrounding rock classification model once: Use the penetration index vector of each sample data in the test set obtained in Step 3. The surrounding rock classification model trained in step 4 is then tested to obtain the surrounding rock classification prediction results.

[0016] Step 6: Calculate the prior probabilities: Using the primary rock classification model sample library constructed in Step 3, calculate the prior probabilities of stable and unstable surrounding rocks.

[0017] Step 7: Calculate the conditional probability: Compare and statistically analyze the initial rock classification prediction results obtained in Step 5 with the actual rock classification of each sample data in the test set to obtain the conditional probability.

[0018] Step 8, Tunnel Excavation: The TBM constructs the tunnel to be excavated and monitors the excavation data of the cutterhead thrust and torque of each excavation segment; let the current excavation segment be t, and record the excavation data of the t-m+1th excavation segment, the t-m+2th excavation segment, ..., the tth excavation segment.

[0019] Step 9: Calculate the penetration index of the tunneling segment: Based on the tunneling data recorded in Step 8, calculate the penetration index vector of the t-m+1 tunneling segment, the t-m+2 tunneling segment, ..., the t tunneling segment.

[0020] Step 10, Prediction of Surrounding Rock Classification in Tunneling Sections: Substitute the penetration index vectors of tunneling sections t-m+1, t-m+2, ..., t obtained in Step 9 into the primary surrounding rock classification model trained in Step 4 to obtain the surrounding rock classification sequence of tunneling sections t-m+1, t-m+2, ..., t. t-m+1 ,c t-m+2 ,...,c t ].

[0021] Step 11, Bayesian posterior estimation: Based on the surrounding rock classification sequence obtained in Step 10 [ct-m+1 ,c t-m+2 ,...,c t The prior probability obtained in step 6 and the conditional probability obtained in step 7 are used to obtain the posterior probability of the unstable category of the surrounding rock in the t-th tunneling section, and then the classification of the surrounding rock at the tunnel face is obtained.

[0022] In step 1, P = 2, representing stable and unstable surrounding rock respectively. According to the national standard "Code for Geological Investigation of Water Conservancy and Hydropower Projects GB50487-2008", the surrounding rock is divided into N = 5 types, namely Class I, II, III, IV and V. Among them, Class I, II and III are stable types of surrounding rock, and Class IV and V are unstable types of surrounding rock.

[0023] In step 1, the classification label for stable surrounding rock is marked as 0, and the classification label for unstable surrounding rock is marked as 1.

[0024] In step 8, the formula for calculating m is:

[0025]

[0026] In the formula, L is the length of the cutterhead shield; d is the maximum travel depth of a single tunneling section.

[0027] In step 7, the conditional probability is that the model predicts the surrounding rock category as c. i Under the condition that its true category is c j The probability of this is denoted as P(c j / c i ), then P(c j / c i The formula for calculating ) is:

[0028]

[0029] In the formula, N ji This indicates that the predicted surrounding rock category is c. i However, the actual surrounding rock category is C. j The number of samples at that time.

[0030] N i This indicates that the predicted surrounding rock category is c. i The total number of samples.

[0031] Step 11, the Bayesian posterior estimation method, includes the following steps.

[0032] Step 11-1: Calculate the predicted posterior probability P(c) of the (t-m+1)th tunnel segment. i / c j ) t-m+1 Specifically, it includes the following steps.

[0033] Step 11-1A: Determine the prior probability of the (t-m+1)th tunnel segment: The prior probability of the (t-m+1)th tunnel segment includes the probability of c t -m+1 The prior probability P(c) of opposite-type surrounding rocks i ) t-m+1 and c t-m+1 Prior probability of the corresponding surrounding rock Wherein, P(c i ) t-m+1 It was found in step 6;

[0034] Step 11-1B, Selecting the conditional probability of the (t-m+1)th tunnel segment: The prior probability of the (t-m+1)th tunnel segment includes P(c j / c i )and in, To be with c t-m+1 Under opposite types of surrounding rock conditions, the true category is c. j The probabilities are all obtained from step 7.

[0035] Step 11-1C, Calculate P(c) i / c j ) t-m+1 :P(c i / c j ) t-m+1 For the (t-m+1)th tunnel segment, and c t-m+1 The specific formula for calculating the posterior probability of the corresponding surrounding rock is as follows:

[0036]

[0037] Step 11-2: Calculate the prior probability of the (t-m+2)th tunnel segment: The prior probability of the (t-m+2)th tunnel segment includes the probability of c. t -m+2 The prior probability P(c) of the corresponding surrounding rock i ) t-m+2 and c t-m+2 Prior probability of opposite type of surrounding rock but:

[0038] P(c i ) t-m+2 =P(c i / c j ) t-m+1

[0039]

[0040] Step 11-3: Repeat steps 11-1B to 11-1C to calculate the distance between the (t-m+2)th tunnel segment and c. t-m+2 The posterior probability P(c) of the corresponding surrounding rock i / c j ) k-m+2 .

[0041] Step 11-4: Repeat steps 11-2 to 11-3 to calculate the posterior probability P(c) of the predicted surrounding rock type for the (t-m+3)th tunnel segment, ..., the tth tunnel segment. i / c j ) t-m+3 ... P(c i / c j ) t .

[0042] Step 11-5, Secondary determination of surrounding rock classification: Based on P(c) obtained in step 11-4 i / c j ) t The rock type with the highest predicted probability is used as the rock type of the face ahead of the current t-th tunnel segment.

[0043] In step 4, the optimal hyperparameter combination of the random forest is selected using a grid search-based cross-validation method.

[0044] Step 4, the method for selecting the optimal hyperparameter combination of the random forest through cross-validation, includes the following steps:

[0045] Step 4-1: Design hyperparameter combinations: For random forest, design h kinds of hyperparameter combinations.

[0046] Step 4-2: Divide the training set into n training subsets D using a bootstrap sampling method. i .

[0047] Step 4-3: Build a random forest: For each training subset D i Each independent decision tree for surrounding rock classification prediction is constructed, resulting in a total of n decision trees for surrounding rock classification prediction, forming a random forest;

[0048] Step 4-4, Hyperparameter Training: Using the random forest built in Step 4-3, the hyperparameters for each combination designed in Step 4-1 are trained. During training, the generation process of each surrounding rock classification prediction decision tree includes the following steps:

[0049] Step 4-4A: Calculate the node Gini index: Calculate the node Gini index at each node split; the formula for calculating the Gini index G(w) of node w is:

[0050]

[0051] In the formula, p c It represents the probability that the current node w belongs to the surrounding rock category c.

[0052] Step 4-4B: Calculate the optimal split point: For each node, calculate each feature f. j Given all possible split points, find the split point that maximizes the information gain; where f j If ∈[a,b,TPI], then the formula for calculating the optimal split point s is:

[0053]

[0054] In the formula, D t The sample set for the current node. and Based on f j The left and right samples after segmentation by the split point s of the feature; G(D t ), and They are nodes D respectively t Node D t and nodes The Gini index.

[0055] Step 4-4C, Recursive Splitting: Repeat steps 4-4A to 4-4B, selecting the best splitting node for the left and right nodes respectively, and recursively splitting until the stopping growth condition of the surrounding rock classification prediction decision tree is reached.

[0056] Steps 4-5: Select the optimal hyperparameters: Calculate the average accuracy of each combination of hyperparameters during validation, and select the set of hyperparameters with the highest average accuracy as the optimal hyperparameters for the random forest.

[0057] In step 4, when the primary rock classification model is training and outputting the rock classification labels of the tunneling section, a prediction model probability threshold is set. The primary rock classification model compares the probability of each rock classification with the prediction model probability threshold, and takes the rock classification with a probability greater than the prediction model probability threshold as the output of the primary rock classification model. The prediction model probability threshold can be adjusted to eliminate the problem of imbalance between samples of different rock categories.

[0058] Step 4, the method for adjusting the probability threshold of the prediction model, includes the following steps:

[0059] Step 4A-1: Calculate the maximum ratio of sample categories: Calculate the number of samples of the largest category T among the samples of different surrounding rock categories in the training set. max and the minimum number of samples in each category Tmin The ratio T is calculated using the following formula:

[0060]

[0061] Step 4A-2, Imbalance Assessment of Surrounding Rock Category Samples: Compare the ratio T calculated in Step 4A-1 with the set ratio threshold Y. When T≤Y, the surrounding rock category samples are considered balanced, and the probability threshold of the prediction model does not need to be adjusted; otherwise, the surrounding rock category samples are considered imbalanced, and proceed to Step 4A-3.

[0062] Step 4A-3: Determine the initial threshold for the prediction model probability: Set the initial threshold for the prediction model probability of both stable and unstable surrounding rock to the average of the two.

[0063] Step 4A-4, Prediction Model Probability Threshold Adjustment: Calculate the percentage K of stable and unstable surrounding rock samples in the total sample, and then adjust K relative to... Compare and, based on the comparison results, [do something]. The adjustments are as follows:

[0064] A. Regarding The initial probability thresholds for the prediction model are sequentially increased according to the following order: Δs, 2Δs, 3Δs, ..., RΔs, based on the surrounding rock category; where:

[0065] B. Regarding The initial probability thresholds of the prediction model for the surrounding rock category are sequentially decreased in the order of Δs, 2Δs, 3Δs, ..., RΔs; where: At the same time, keep all greater than The increase in the value of the surrounding rock category is less than The reduction values ​​for the surrounding rock category are the same.

[0066] Steps 4A-5: F1 Value Evaluation: For each increase or decrease in the probability threshold of the prediction model, perform an F1 value evaluation and construct a curve showing the change of the F1 value with the probability threshold of the prediction model.

[0067] Step 4A-6: Select the optimal prediction model probability threshold: Take the prediction model probability threshold corresponding to the maximum F1 value in the curve of F1 value changing with the prediction model probability threshold as the optimal prediction model probability threshold for the primary rock classification model.

[0068] The present invention has the following beneficial effects:

[0069] (1) This invention uses a Bayesian reinforcement ensemble learning method to intelligently identify the surrounding rock classification during TBM construction. The prediction results can evaluate the surrounding rock classification of the tunnel boring machine, especially providing real-time early warning and forecast when encountering extremely unstable sections, which can effectively reduce the risk of project instability and reduce accidents. In addition, the surrounding rock classification evaluation results can provide guidance for the selection of on-site surrounding rock support methods, which is conducive to further reducing support costs and improving construction efficiency.

[0070] (2) The excavability index parameters are extracted from the monitoring data of the interaction between the TBM and the rock machine during the TBM construction process and used as the input parameters of the intelligent prediction model. The measurement of relevant field monitoring data can rely on intelligent sensor devices. Compared with indoor tests, advanced drilling and other methods, the data acquisition cost is lower and the impact on normal construction organization is smaller.

[0071] (3) Bayesian update is used to further estimate the probability sequence prediction results of the machine learning model posteriorly. The Bayesian enhancement method can effectively reduce the variability of prediction results caused by the quality of monitoring data, and at the same time considers the influence of monitoring data before the current tunneling section, thus improving the reliability of the prediction results. Attached Figure Description

[0072] Figure 1 is a flowchart of the intelligent identification method for TBM surrounding rock classification based on penetration index and Bayesian enhancement.

[0073] Figure 2 shows the penetration index extracted from the location of the Chuojiliao Tunnel at chainage 57123; where (a) is the single-blade thrust-penetration index; and (b) is the single-blade torque-penetration index.

[0074] Figure 3 is a schematic diagram of the optimal segmentation features and score selection in the decision tree generation process.

[0075] Figure 4 is a comparison of prediction performance metrics using the adjusted prediction model probability threshold.

[0076] Figure 5 is a flowchart of the posterior estimation of probability results for enhanced ensemble learning based on the Bayesian method.

[0077] Figure 6 shows the relationship between the classification probability of the surrounding rock in the section from chainage 26+030 to 25+980 of the Yinchuo-Jiliao No. 3 Tunnel and the chainage.

[0078] Figure 7 is a comparison of the accuracy of the algorithm using random forest and the Bayesian augmentation algorithm in this invention. Detailed Implementation

[0079] The present invention will now be described in further detail with reference to the accompanying drawings and specific preferred embodiments.

[0080] This invention provides a detailed description of the on-site monitoring data collected during the construction process of the No. 6 branch tunnel of the No. 2 section of the Inner Mongolia Yinchuo Jiliao Project.

[0081] As shown in Figure 1, a smart identification method for TBM surrounding rock classification based on penetration index and Bayesian enhancement includes the following steps.

[0082] Step 1: Classification of surrounding rocks: Based on rock strength, integrity, structural surface condition, occurrence, and groundwater conditions, the surrounding rocks are preferentially classified into two types, P=2, namely stable and unstable. Then, each type of surrounding rock is labeled with a numerical classification tag C.

[0083] In this embodiment, key rock-breaking parameters during the construction of the Yinchuo-Jiliao 2-6 branch tunnel were collected. Key TBM excavation parameters, such as cutterhead thrust, torque, penetration depth, and cutterhead rotation speed, were collected at a sampling frequency of 1 Hz / s. The tunnel is intended for inter-regional water conveyance. According to the national standard "Code for Geological Investigation of Water Conservancy and Hydropower Projects GB50487-2008," the surrounding rock is classified into N=5 types: Class I, II, III, IV, and V. Classes I, II, and III are considered stable, while Classes IV and V are considered unstable. Furthermore, stable surrounding rock is labeled with 0, and unstable surrounding rock is labeled with 1.

[0084] Step 2: Construct a primary rock classification model based on penetration indices: For each tunneling segment during TBM construction, a primary rock classification model based on penetration indices is constructed using the random forest algorithm; the input layer of the primary rock classification model is the penetration index vector of the tunneling segment during TBM construction. The output layer is the surrounding rock classification label C for the corresponding tunneling section; where a is the linear fitting slope of the cutterhead single cutter thrust and penetration in the tunneling section; b is the initial penetration thrust of the cutterhead roller in the tunneling section. The linear fit between the single cutter thrust and penetration depth in the tunneling section is represented by TPI; TPI is the linear fit slope between the single cutter torque and penetration depth in the tunneling section. This indicates the goodness of linear fit between the single-blade torque and penetration depth in the tunneling section.

[0085] Step 3: Construct a primary rock classification model sample library, which includes the following steps:

[0086] Step 3-1: Collect effective tunneling segment data: Collect effective tunneling segment data for all types of surrounding rock in Step 1; wherein, the number of effective tunneling segments for each type of surrounding rock is not less than M; M≥100.

[0087] In this embodiment, the on-site rock breaking parameters of the 6th branch tunnel of the Yinchuo-Jiliao No. 2 section include M=5797 complete and effective rock breaking tunneling sections. The data for each tunneling section includes the total cutterhead thrust, cutterhead torque, propulsion speed, and cutterhead rotation speed sequence.

[0088] Step 3-2: Extracting Penetration Indicators for Each Excavation Section: For each excavation section of each type of surrounding rock, the penetration index vector is extracted, and the actual surrounding rock classification label C corresponding to each excavation section is recorded, thus obtaining a primary surrounding rock classification model sample library containing N sample data; each sample data includes... Penetration sequence calculation. Based on the advance speed v and cutterhead rotation speed n, the penetration depth p, which corresponds to the depth of rock penetration when the cutterhead rotates once, is calculated. The calculation formula is as follows:

[0089]

[0090] First, the total thrust sequence and torque sequence of the cutterhead are divided by the number of cutters on the TBM cutterhead to obtain the thrust and torque of a single cutter. Then, linear fitting is performed on the thrust and penetration of a single cutter, and the torque and penetration of a single cutter for each tunneling section to obtain the rock mass penetration index.

[0091] In this embodiment, the single-blade thrust, single-blade torque, and penetration depth of the effective rock-breaking extreme in the tunnel section of the No. 6 branch tunnel of the Yinchuo-Jiliao 2nd section, at the pile number 57123, are linearly fitted to obtain the rock mass penetration index. Based on this, combined with the classification tags of 5797 tunnel sections collected in this project, an initial database is constructed, as shown in Table 1 below.

[0092] Table 1. Examples of database entries for the Yinchuo-Jiliao 2-6 branch tunnel excavation section.

[0093]

[0094] Step 3-3, Sample Library Classification: Divide the sample library into two parts: training set and test set.

[0095] In this invention, a 10-fold algorithm is used to divide the sample library into 10 folds, with 9 folds used for model training and the remaining 1 fold used for validation.

[0096] In the No. 6 tunnel of the Yinchuo-Jiliao Project Section 2, the sample size ratio of rock mass penetration index under stable and unstable working conditions is 4949:848. Since the sample sizes differ between the two working conditions, this technical solution adopts a stratified sampling method based on working conditions for data partitioning. The specific steps are as follows:

[0097] First, 90% of the 4949 samples in the working condition 1 surrounding rock stability category, i.e., 4454 samples, are randomly selected as the working condition 1 training set, and the remaining 495 samples are used as the working condition 1 test set.

[0098] Then, 90% of the 848 samples in the stable and unstable rock category of working condition 2, i.e., 763 samples, were randomly selected as the training set for working condition 2, and the remaining 85 samples were selected as the test set for working condition 2.

[0099] The training sets from the two operating conditions described above are combined to form the total training set, and the test sets from the two operating conditions are combined to form the total test set. The final total training set contains 5217 samples, and the test set contains 580 samples.

[0100] Step 4: Train the primary rock classification model: Use the training set obtained in Step 3 to train the primary rock classification model constructed in Step 2 to obtain the trained primary rock classification model.

[0101] In this invention, a cross-validation method based on grid search is preferred to select the optimal combination of hyperparameters for random forest.

[0102] Step 4, the method for selecting the optimal hyperparameter combination of the random forest through cross-validation, includes the following steps:

[0103] Step 4-1: Design hyperparameter combinations: For random forest, design h kinds of hyperparameter combinations.

[0104] In this embodiment, the designed hyperparameter combinations are: number of decision trees [50, 100, 200, 300], minimum number of sample splits [14, 28, 42, 56], maximum number of features [2, 3, 4, 5], and node splitting criteria [gini, entropy]. Here, gini represents the Gini index, and entropy represents the cross-entropy value.

[0105] Step 4-2: Divide the training set into n training subsets D using a bootstrap sampling method. i .

[0106] Step 4-3: Build a random forest: For each training subset D i Each independent decision tree for surrounding rock classification and prediction is constructed, resulting in a total of n decision trees for surrounding rock classification and prediction, forming a random forest.

[0107] Step 4-4, Hyperparameter Training: Using the random forest built in Step 4-3, the hyperparameters of each combination designed in Step 4-1 are trained.

[0108] As shown in Figure 3, the generation process of each surrounding rock classification prediction decision tree during training includes the following steps.

[0109] Step 4-4A: Calculate the node Gini index: Calculate the node Gini index at each node split; the formula for calculating the Gini index G(w) of node w is:

[0110]

[0111] In the formula, p c It represents the probability that the current node w belongs to the surrounding rock category c.

[0112] Step 4-4B: Calculate the optimal split point: For each node, calculate each feature f. j Given all possible split points, find the split point that maximizes the information gain; where f j If ∈[a,b,TPI], then the formula for calculating the optimal split point s is:

[0113]

[0114] In the formula, D t The sample set for the current node. and Based on f j The left and right samples after segmentation by the split point s of the feature; G(D t ), and G(D) tR ) are nodes D t Node D t and nodes The Gini index.

[0115] In this embodiment, for each node, each feature f is calculated. j For all possible split points ∈ [a,b,TPI], find the split point that maximizes information gain. Taking TPI as an example, the initial minimum Gini value split point, Gini(TPI) = 0.27, is at TPI = 3.95, which is also less than the Gini values ​​of features a and b, Gini(a) = 0.37 and Gini(b) = 0.42. Therefore, the optimal split point is at TPI = 3.95.

[0116] Step 4-4C, Recursive Splitting: Repeat steps 4-4A to 4-4B, selecting the best splitting node for the left and right nodes respectively, and recursively splitting until the stopping growth condition of the surrounding rock classification prediction decision tree is reached.

[0117] Steps 4-5: Select the optimal hyperparameters: Calculate the average accuracy of each combination of hyperparameters during validation, and select the set of hyperparameters with the highest average accuracy as the optimal hyperparameters for the random forest.

[0118] In this embodiment, the accuracy is highest when the hyperparameter combination is [100,28,3,'gini'], so the model trained with this hyperparameter combination is selected for model testing.

[0119] Furthermore, when training and outputting the surrounding rock classification labels for the tunneling section, the primary surrounding rock classification model sets a prediction model probability threshold. The primary surrounding rock classification model compares the probability of each surrounding rock classification with the prediction model probability threshold, and takes the surrounding rock classification that is greater than the prediction model probability threshold as the output of the primary surrounding rock classification model. The prediction model probability threshold can be adjusted to eliminate the problem of imbalance between samples of different surrounding rock categories.

[0120] The method for adjusting the probability threshold of the above prediction model includes the following steps.

[0121] Step 4A-1: Calculate the maximum ratio of sample categories: Calculate the number of samples of the largest category T among the samples of different surrounding rock categories in the training set. max and the minimum number of samples in each category T min The ratio T is calculated using the following formula:

[0122]

[0123] In this embodiment, the number T of samples of the largest category among samples of different surrounding rock categories in the training set is calculated. max =4989 and the minimum number of samples in each category T min The ratio T of 848 is:

[0124]

[0125] Step 4A-2, Imbalance Assessment of Surrounding Rock Category Samples: Compare the ratio T calculated in Step 4A-1 with the set ratio threshold Y. When T≤Y, the surrounding rock category samples are considered balanced, and the probability threshold of the prediction model does not need to be adjusted; otherwise, the surrounding rock category samples are considered imbalanced, and proceed to Step 4A-3.

[0126] In this embodiment, the set ratio threshold Y is preferably set to 2, 5.88>2, that is, T>Y, and proceed to step 4A-3.

[0127] Step 4A-3: Determine the initial threshold for the prediction model probability: Set the initial threshold for the prediction model probability of both stable and unstable surrounding rock to the average of the two.

[0128] Step 4A-4, Prediction Model Probability Threshold Adjustment: Calculate the percentage K of stable and unstable surrounding rock samples in the total sample, and then adjust K relative to... Compare and, based on the comparison results, [do something]. The adjustments are as follows:

[0129] A. Regarding The initial probability thresholds for the prediction model are sequentially increased according to the following order: Δs, 2Δs, 3Δs, ..., RΔs, based on the surrounding rock category; where:

[0130] B. Regarding The initial probability thresholds of the prediction model for the surrounding rock category are sequentially decreased in the order of Δs, 2Δs, 3Δs, ..., RΔs; where: At the same time, keep all greater than The increase in the value of the surrounding rock category is less than The reduction values ​​for the surrounding rock category are the same.

[0131] In this embodiment, stable category samples account for 0.855% of the total samples, and unstable category samples account for 0.145% of the total samples.

[0132] Furthermore, if Δs = 0.1 is preferred, then the increases or decreases in Δs, 2Δs, 3Δs, ..., RΔs are 0.2, 0.3, 0.4, and 0.5, respectively.

[0133] Steps 4A-5, F1 score evaluation: For each increase or decrease in the probability threshold of the prediction model, an F1 score evaluation is performed, and a curve showing the change of the F1 score with the probability threshold of the prediction model is constructed; the F1 score is the harmonic mean score of the model precision and recall, which is the existing technology.

[0134] Step 4A-6: Select the optimal prediction model probability threshold: Take the prediction model probability threshold corresponding to the maximum F1 value in the curve of F1 value changing with the prediction model probability threshold as the optimal prediction model probability threshold for the primary rock classification model.

[0135] To further illustrate the impact of different prediction probability thresholds on model performance, Figure 4 shows the curves of prediction accuracy, precision, recall, and F1 score under unstable states with varying probability thresholds. The results show that as the threshold increases, precision gradually increases while recall decreases in unstable states. This indicates that increasing the threshold increases the reliability of the prediction results but also increases the false negative rate. Furthermore, accuracy, F1 score, and weighted F1 score initially increase and then decrease with increasing threshold, reaching their maximum at a threshold of 0.4. This suggests that adjusting the probability threshold can improve prediction performance under imbalanced samples. Therefore, adjusting the probability threshold is an effective method to eliminate sample imbalance, and the optimal probability threshold should be used for prediction on the test set.

[0136] Therefore, in this embodiment, the optimal prediction model probability threshold of the primary surrounding rock classification model is 0.4.

[0137] Step 5: Test the surrounding rock classification model once: Use the penetration index vector of each sample data in the test set obtained in Step 3. The surrounding rock classification model trained in step 4 is then tested to obtain the surrounding rock classification prediction result.

[0138] Confusion matrix calculation

[0139] The optimal threshold parameter was used to evaluate the performance of the surrounding rock classification prediction model on the test set. The average prediction probability of a sample under different surrounding rock classifications in the leaf nodes of each decision tree was used as the prediction probability, and the category corresponding to the maximum probability was taken as the prediction result. A confusion matrix describing the accuracy of the test set was obtained based on the predicted category and the true label.

[0140]

[0141] In the formula, N ij This represents the number of samples in the test set where the true rock class is i, but which are predicted to be class j.

[0142] Taking the binary classification of surrounding rock as an example, unstable rock masses are positive samples (1), and stable rock masses are negative samples (0), resulting in the following confusion matrix:

[0143]

[0144] Since the random forest method based on ensemble learning only considers the excavability index of the current tunnel segment for classification and prediction, the prediction results are affected by the data quality of the current tunnel segment, and therefore the reliability of the results needs to be improved. Bayesian inference based on more observation data can improve the confidence of the results by relying on the information of the already excavated sections, and its process is shown in Figure 5.

[0145] Step 6: Calculate the prior probabilities: Using the primary rock classification model sample library constructed in Step 3, calculate the prior probabilities of stable and unstable surrounding rocks.

[0146] In this embodiment, the total number of samples in the database is 5797, of which 4949 are in the stable region and 848 are in the unstable region. Therefore, the prior probabilities for the two operating conditions are as follows:

[0147] P(c0) = 4949 / 5797 = 0.85

[0148] P(c1) = 848 / 5797 = 0.15

[0149] Step 7: Calculate the conditional probability: Compare and statistically analyze the initial rock classification prediction results obtained in Step 5 with the actual rock classification of each sample data in the test set to obtain the conditional probability.

[0150] The above conditional probability is for the model to predict the surrounding rock category as c. i Under the condition that its true category is c j The probability of this is denoted as P(c j / c i In this embodiment, i = 0 or 1, j = 0 or 1, then P(c j / c i The formula for calculating ) is:

[0151]

[0152] In the formula, N ji This indicates that the predicted surrounding rock category is c. i However, the actual surrounding rock category is C. j The number of samples at that time.

[0153] N i This indicates that the predicted surrounding rock category is c. i The total number of samples.

[0154] In this embodiment, the confusion matrix obtained in step 5 is used to calculate the following:

[0155] P(c1 / c0) = 69 / 495 = 0.16

[0156]

[0157] P(c0 / c1) = 20 / 85 = 0.24

[0158]

[0159] Step 8, Tunnel Excavation: The TBM constructs the tunnel to be excavated and monitors the excavation data of the cutterhead thrust and torque of each excavation segment; let the current excavation segment be t, and record the excavation data of the t-m+1th excavation segment, the t-m+2th excavation segment, ..., the tth excavation segment.

[0160] The formula for calculating m above is:

[0161]

[0162] In the formula, L is the length of the cutterhead shield; d is the maximum travel depth of a single tunneling section.

[0163] In this embodiment, the parameter m is set based on the length of the cutterhead shield L = 6m and the maximum advance d = 1.8m of a single tunneling section, calculated by the following formula:

[0164]

[0165] Step 9: Calculate the penetration index of the tunneling segment: Based on the tunneling data recorded in Step 8, calculate the penetration index vector of the t-m+1 tunneling segment, the t-m+2 tunneling segment, ..., the t tunneling segment.

[0166] In this embodiment, since m = 4, the penetration index vectors of the (t-3)th tunneling segment, the (t-2)th tunneling segment, the (t-1)th tunneling segment, and the tth tunneling segment are calculated.

[0167] Step 10, Prediction of Surrounding Rock Classification in Tunneling Sections: Substitute the penetration index vectors of tunneling sections t-m+1, t-m+2, ..., t obtained in Step 9 into the primary surrounding rock classification model trained in Step 4 to obtain the surrounding rock classification sequence of tunneling sections t-m+1, t-m+2, ..., t. t-m+1 ,c t-m+2 ,...,c t ].

[0168] In this embodiment, the surrounding rock classification sequence of tunneling segment t-3, tunneling segment t-2, tunneling segment t-1, and tunneling segment t is...

[0169] [c t-3 ,c t-2 ,c t-1 ,c t ] = [1,1,1,1].

[0170] Step 11, Bayesian posterior estimation: Based on the surrounding rock classification sequence obtained in Step 10 [c t-m+1 ,c t-m+2 ,...,c t The prior probability obtained in step 6 and the conditional probability obtained in step 7 are used to obtain the posterior probability of the unstable category of the surrounding rock in the t-th tunneling section, and then the classification of the surrounding rock at the tunnel face is obtained.

[0171] The Bayesian posterior estimation method described above includes the following steps.

[0172] Step 11-1: Calculate the predicted posterior probability P(c) of the (t-m+1)th tunnel segment. i / c j ) t-m+1 Specifically, it includes the following steps:

[0173] Step 11-1A: Determine the prior probability of the (t-m+1)th tunnel segment: The prior probability of the (t-m+1)th tunnel segment includes the probability of c t -m+1 Conversely, the prior probability P(c) of the surrounding rock i ) t-m+1 and c t-m+1 Prior probability of the corresponding type of surrounding rock Wherein, P(c i ) t-m+1 It was found in step 6;

[0174] In this embodiment, since the (t-3)th tunneling segment is connected to c t-3 =1. The prior probability of the opposite type of surrounding rock, i.e., stable surrounding rock, is as follows:

[0175] P(c i ) t-3 =P(c0) = 4949 / 5797 = 0.85

[0176] Step 11-1B, Selecting the conditional probability of the (t-m+1)th tunnel segment: The prior probability of the (t-m+1)th tunnel segment includes P(c j / c i )and in, To be with c t-m+1 Under opposite types of surrounding rock conditions, the true category is c. j The probabilities are all obtained from step 7.

[0177] In this embodiment, the prior probability of the (t-3)th tunnel segment includes:

[0178] P(c1 / c0) = 69 / 495 = 0.16

[0179]

[0180] Step 11-1C, Calculate P(c) i / c j ) t-m+1 :P(c i / c j ) t-m+1 For the (t-m+1)th tunnel segment, and c t-m+1 The specific formula for calculating the posterior probability of the corresponding surrounding rock is as follows:

[0181]

[0182] In this embodiment, the (t-3)th tunnel segment and c t-3 The posterior probability of the surrounding rock corresponding to 1 is:

[0183]

[0184] Step 11-2: Calculate the prior probability of the (t-m+2)th tunnel segment: The prior probability of the (t-m+2)th tunnel segment includes the probability of c. t -m+2 The prior probability P(c) of the corresponding surrounding rock i ) t-m+2 and c t-m+2 Prior probability of opposite type of surrounding rock but:

[0185] P(c i ) t-m+2 =P(c i / c j ) t-m+1

[0186]

[0187] In this embodiment, the (t-3)th tunneling segment and c t-2 =1. The prior probability of the opposite type of surrounding rock, i.e., stable surrounding rock, is:

[0188] Penetration index P(c) i ) t-2 =P(c0 / c1) t-3 =0.544

[0189]

[0190] Step 11-3: Repeat steps 11-1B to 11-1C to calculate the distance between the (t-m+2)th tunnel segment and c. t-m+2 The posterior probability P(c) of the corresponding surrounding rock i / c j ) k-m+2 .

[0191] Based on the posterior probability when the (k-3)th prediction result is 1:

[0192]

[0193] P(c0) t-1 =P(c0 / c1) t-2 =0.20

[0194] Step 11-4: Repeat steps 11-2 to 11-3 to calculate the posterior probability P(c) of the predicted surrounding rock type for the (t-m+3)th tunnel segment, ..., the tth tunnel segment. i / cj ) k-m+3 ... P(c i / c j ) t .

[0195] Based on the posterior probability when the (k-2)th prediction result is 1:

[0196]

[0197] P(c0) t =P(c0 / c1) t-1 =0.05

[0198] Based on the posterior probability when the (k-1)th prediction result is 1:

[0199]

[0200] Step 11-5, Secondary determination of surrounding rock classification: Based on P(c) obtained in step 11-4 i / c j ) t The rock type with the highest predicted probability is used as the rock type of the face ahead of the current t-th tunnel segment.

[0201] Based on the above process, the posterior probability of the stable category based on the four unstable prediction results [1,1,1,1] is 0.01, and the posterior probability of the unstable category based on the four unstable prediction results [1,1,1,1] is 0.99. Using the Bayesian augmentation method, the probability of the unstable section based on the excavated section is 0.99.

[0202] The Bayesian augmented ensemble learning method described above was applied to Tunnel No. 3 of Section 2 of the Yinchuo-Jiliao Water Diversion Project for verification. A length of approximately 50m was selected from chainage 26+030 to 25+980 for testing. The rock mass in this section was classified as Class V, indicating unstable rock mass. The predicted probabilities obtained using both random forest and Bayesian augmented learning methods are shown in Figures 6 and 7, respectively. It can be seen that the accuracy of the Bayesian augmented learning method is 92%, which is 3.37% higher than the accuracy of the random forest algorithm.

[0203] This invention enables rapid extraction of penetration indices reflecting the rock-machine interaction process from rock-breaking parameters monitored during tunnel boring machine (TBM) construction, achieving intelligent perception of surrounding rock classification and reducing the time and economic costs of geological exploration. Employing Bayesian theory to enhance the intelligent prediction model allows for the consideration of more observational data and probabilistic results from already excavated sections to predict the current tunneling section, effectively eliminating the impact of monitoring data quality on prediction results and further improving the reliability of surrounding rock classification results. This method significantly enhances the automation and intelligence of geological condition perception during mechanized tunnel construction, providing guidance for surrounding rock stability assessment and support method selection.

[0204] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A smart identification method for TBM surrounding rock classification based on penetration index and Bayesian enhancement, characterized in that: The process includes the following steps: Step 1, Rock Classification: Based on rock strength, integrity, structural surface condition, occurrence, and groundwater conditions, the surrounding rock is classified into P types; then, each type of surrounding rock is labeled with a numerical classification tag C; Step 2, Construction of a Primary Rock Classification Model Based on Penetration Indices: For each tunneling section during TBM construction, a primary rock classification model based on penetration indices is constructed using a random forest algorithm; the input layer of the primary rock classification model is the penetration index vector of the tunneling section during TBM construction. The output layer is the surrounding rock classification label C of the corresponding tunneling section; where a is the linear fitting slope of the cutterhead single cutter thrust and penetration in the tunneling section; b is the initial penetration thrust of the cutterhead roller in the tunneling section. The linear fit between the single cutter thrust and penetration depth in the tunneling section is represented by TPI; TPI is the linear fit slope between the single cutter torque and penetration depth in the tunneling section. This represents the linear fit goodness of the single-blade torque and penetration depth in the tunneling section; Step 3: Construct a primary rock classification model sample library, specifically including the following steps: Step 3-1: Collect effective tunneling section data: Collect effective tunneling section data for all types of surrounding rock in Step 1; wherein, the number of effective tunneling sections for each type of surrounding rock is not less than M; M≥100; Step 3-2: Extract penetration index of tunneling section: Extract the penetration index vector for each tunneling section of each type of surrounding rock, and record the actual surrounding rock classification label C corresponding to each tunneling section, thereby obtaining a primary rock classification model sample library containing N sample data; each sample data includes Step 3-3: Sample Library Classification: Divide the sample library into two parts: a training set and a test set; Step 4: Train the Primary Rock Classification Model: Use the training set obtained in Step 3 to train the primary rock classification model constructed in Step 2, obtaining the trained primary rock classification model; Step 5: Test the Primary Rock Classification Model: Use the ingress index vector of each sample data in the test set obtained in Step 3. The training model for the surrounding rock classification, along with the model trained in step 4, is tested to obtain the initial surrounding rock classification prediction results. Step 6: Calculate the prior probabilities: Using the sample library of the initial surrounding rock classification model constructed in step 3, calculate the prior probabilities of stable and unstable surrounding rocks. Step 7: Calculate the conditional probabilities: Compare and statistically analyze the initial surrounding rock classification prediction results obtained in step 5 with the actual surrounding rock classification of each sample data in the test set to obtain the conditional probabilities; where the conditional probability is the model's predicted surrounding rock category. Under the condition that its true category is The probability; Step 8, Tunnel excavation: The TBM constructs the tunnel to be excavated, and monitors the excavation data of the cutterhead thrust and torque of each excavation segment; Let the current excavation segment be t, and record the excavation data of the t-m+1th excavation segment, the t-m+2th excavation segment, ..., the tth excavation segment; where the parameter m is set based on the length of the cutterhead shield used in the tunnel excavation project and the maximum travel of the machine in a single excavation segment; where the calculation formula for m is: In the formula, L is the length of the cutterhead shield; d is the maximum travel distance of a single tunneling segment; Step 9: Calculate the penetration index of the tunneling segment: Based on the tunneling data recorded in Step 8, calculate the penetration index vectors for the (t-m+1)th, (t-m+2)th, ..., (t)th tunneling segments; Step 10: First-order prediction of surrounding rock classification for tunneling segments: Substitute the penetration index vectors obtained in Step 9 for the (t-m+1)th, (t-m+2)th, ..., (t)th tunneling segments into the first-order surrounding rock classification model trained in Step 4 to obtain the surrounding rock classification sequence for the (t-m+1)th, (t-m+2)th, ..., (t)th tunneling segments. Step 11, Bayesian posterior estimation: Based on the surrounding rock classification sequence obtained in step 10. The prior probability obtained in step 6 and the conditional probability obtained in step 7 are used to obtain the posterior probability of the unstable category of the surrounding rock in the t-th tunneling section, and then the classification of the surrounding rock at the tunnel face is obtained.

2. The intelligent identification method for TBM surrounding rock classification based on penetration index and Bayesian enhancement as described in claim 1, characterized in that: In step 1, the surrounding rock is divided into two types, C=2, namely stable surrounding rock and unstable surrounding rock. According to the national standard "Code for Geological Investigation of Water Conservancy and Hydropower Projects GB50487-2008", the surrounding rock is divided into five types, namely Class I, II, III, IV and V. Among them, Class I, II and III are stable types of surrounding rock, and Class IV and V are unstable types of surrounding rock.

3. The intelligent identification method for TBM surrounding rock classification based on penetration index and Bayesian enhancement as described in claim 2, characterized in that: In step 1, the classification label for stable surrounding rock is marked as 0, and the classification label for unstable surrounding rock is marked as 1.

4. The intelligent identification method for TBM surrounding rock classification based on penetration index and Bayesian enhancement as described in claim 1, characterized in that: In step 7, the conditional probability is denoted as... ,but The calculation formula is: In the formula, Indicates the predicted surrounding rock type is However, the actual surrounding rock type is The number of samples at that time; Indicates the predicted surrounding rock type is The total number of samples.

5. The intelligent identification method for TBM surrounding rock classification based on penetration index and Bayesian enhancement according to claim 4, characterized in that: In step 11, the Bayesian posterior estimation method includes the following steps: Step 11-1, calculate the predicted posterior probability of the (t-m+1)th tunnel segment. Specifically, it includes the following steps: Step 11-1A, Determine the prior probability of the (t-m+1)th tunnel segment: The prior probability of the (t-m+1)th tunnel segment includes... Prior probability of opposite type of surrounding rock Wayo Prior probability of the corresponding surrounding rock ;in, It was found in step 6; Step 11-1B: Select the conditional probability of the (t-m+1)th tunnel segment: The prior probability of the (t-m+1)th tunnel segment includes... and ;in, To and The true category under opposite-type surrounding rock conditions is: The probabilities are all obtained from step 7; step 11-1C, calculation ： For the (t-m+1)th tunnel segment and The specific formula for calculating the posterior probability of the corresponding surrounding rock is as follows: Step 11-2: Calculate the prior probability of the (t-m+2)th tunnel segment: The prior probability of the (t-m+2)th tunnel segment includes... Prior probability of the corresponding surrounding rock Wayo Prior probability of opposite type of surrounding rock ,but: ； Step 11-3: Repeat steps 11-1B to 11-1C to calculate the tunneling section t-m+2. Posterior probability of the corresponding surrounding rock Step 11-4: Repeat steps 11-2 to 11-3 to calculate the posterior probability of the predicted surrounding rock type for the (t-m+3)th tunnel segment, ..., the tth tunnel segment. 、……、 Step 11-5, Secondary determination of surrounding rock classification: Based on the results obtained in step 11-4 The rock type with the highest predicted probability is used as the rock type of the face ahead of the current t-th tunnel segment.

6. The intelligent identification method for TBM surrounding rock classification based on penetration index and Bayesian enhancement according to claim 1, characterized in that: In step 4, the optimal hyperparameter combination of the random forest is selected using a grid search-based cross-validation method.

7. The intelligent identification method for TBM surrounding rock classification based on penetration index and Bayesian enhancement as described in claim 6, characterized in that: In step 4, the cross-validation method for selecting the optimal hyperparameter combination for the random forest includes the following steps: Step 4-1, Designing hyperparameter combinations: For the random forest, design h hyperparameter combinations; Step 4-2, Dividing the training set: Divide the total training set into n training subsets D using a bootstrap sampling method. i Step 4-3: Build a random forest: For each training subset D i Each independent rock classification prediction decision tree is constructed, resulting in a total of n rock classification prediction decision trees, forming a random forest; Step 4-4, Hyperparameter Training: Using the random forest built in Step 4-3, the hyperparameters for each combination designed in Step 4-1 are trained; During the training process, the generation process of each rock classification prediction decision tree includes the following steps: Step 4-4A, Calculating the node Gini index: When each node splits, its node Gini index is calculated; where, the Gini index of node w is... The calculation formula is: In the formula, p c This is the probability that the current node w belongs to the surrounding rock category c; Step 4-4B: Calculate the optimal split point: For each node, calculate each feature Given all possible split points, find the split point that maximizes information gain; where, Then the formula for calculating the optimal split point s is: In the formula, The sample set for the current node. and Based on The left and right samples after segmentation by the split point s of the feature; 、 and They are nodes ,node and nodes The Gini index; Step 4-4C, Recursive splitting: Repeat steps 4-4A to 4-4B, recursively splitting the left and right nodes by selecting the best splitting nodes respectively, until the stopping growth condition of the surrounding rock classification prediction decision tree is reached; Step 4-5, Selecting the optimal hyperparameters: Calculate the average accuracy of each combination of hyperparameters during validation, and select the set of hyperparameters with the highest average accuracy as the optimal hyperparameters for the random forest.

8. The intelligent identification method for TBM surrounding rock classification based on penetration index and Bayesian enhancement according to claim 1, characterized in that: In step 4, when the primary rock classification model is training and outputting the rock classification labels of the tunneling section, a prediction model probability threshold is set. The primary rock classification model compares the probability of each rock classification with the prediction model probability threshold, and takes the rock classification with a probability greater than the prediction model probability threshold as the output of the primary rock classification model. The prediction model probability threshold can be adjusted to eliminate the problem of imbalance between samples of different rock categories.

9. The intelligent identification method for TBM surrounding rock classification based on penetration index and Bayesian enhancement as described in claim 8, characterized in that: In step 4, the method for adjusting the probability threshold of the prediction model includes the following steps: Step 4A-1, Calculate the maximum ratio of sample classes: Calculate the number of samples of the largest class T in the different surrounding rock categories in the training set. max and the minimum number of samples in each category T min The ratio T is calculated using the following formula: Step 4A-2, Imbalance Assessment of Surrounding Rock Category Samples: Compare the ratio T calculated in Step 4A-1 with the set ratio threshold Y. When T ≤ Y, the surrounding rock category samples are considered balanced, and the prediction model probability threshold does not need to be adjusted; otherwise, the surrounding rock category samples are considered imbalanced, and proceed to Step 4A-3. Step 4A-3, Determine the Initial Probability Threshold of the Prediction Model: Set the initial probability threshold of the prediction model for each type of surrounding rock to the average value of the different categories. =1 / P; Step 4A-4, Prediction model probability threshold adjustment: Calculate the percentage K of stable and unstable surrounding rock samples in the total sample, and then compare K with... Compare and, based on the comparison results, [do something]. Adjustments will be made, specifically: A. For The surrounding rock category is used to determine the initial threshold of the prediction model probability, and then... 、 、 、……、 The order is as follows, increasing sequentially; where: B. For The surrounding rock category is used to determine the initial threshold of the prediction model probability, and then... 、 、 、……、 The order is to decrease sequentially; where: At the same time, keep all greater than The increase in the value of the surrounding rock category is less than The reduction values ​​for the surrounding rock category are the same; Step 4A-5, F1 value evaluation: For each increase or decrease in the prediction model probability threshold, an F1 value evaluation is performed, and a curve of the change of F1 value with the prediction model probability threshold is constructed; Step 4A-6, Selecting the optimal prediction model probability threshold: The prediction model probability threshold corresponding to the largest F1 value in the curve of the change of F1 value with the prediction model probability threshold is taken as the optimal prediction model probability threshold for a single surrounding rock classification model.

Images