Tunnel rock mass quality identification method and system based on multi-scale drilling model fusion

Abstract

The present application relates to the technical field of geological prediction, in particular to a tunnel rock mass quality identification method and system based on multi-scale drilling model fusion, specifically comprising the following steps: S1, collecting and preprocessing the field drilling data to obtain a test set; S2, inputting the test set into a multi-scale fusion model to obtain a tunnel rock mass quality identification result, the tunnel rock mass quality identification result being one of the prediction labels, and the prediction labels including relatively complete, relatively broken, broken and fissure; in the training process of the multi-scale fusion model, the field drilling data is collected, preprocessed and segmented in multiple scales, and the multi-scale model fusion is performed by using the error reciprocal weighted average method in the training process. The effect of weakening the shortcomings of each scale model, optimizing the prediction effect and enhancing the practicality of the technology is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of advanced geological prediction, in particular to a tunnel rock mass quality identification method and system based on multi-scale drilling model fusion. BACKGROUND

[0002] Advanced geological prediction is one of the important means to ensure the safety of tunnel construction, which can explore the geological structure, adverse geological body type and special rock mass properties in front of the working face, and provide important reference for tunnel excavation method and support measures. According to the working principle, tunnel advanced geological prediction can be divided into two categories: geophysical method and geological analysis method. The former such as geological radar and seismic wave detects the physical property change of unexcavated rock mass to make corresponding geological prediction, which has the disadvantage of significant multiple solutions; advanced drilling belongs to the geological analysis method, which uses drilling equipment to drill into the rock mass in front of the working face to directly reveal the strength, integrity and other data of unexcavated rock mass, as well as the properties of underground water, karst, soft interlayer and other geological bodies, which is one of the most direct, effective and reliable advanced geological prediction methods and is widely used in engineering practice.

[0003] With the development of digital drilling technology, since the 1970s, multifunctional drilling machines carrying multiple sensors have appeared, which can record and monitor the drilling parameters such as drilling rate, rotation speed, torque and thrust force in real time, which lays an important data foundation for the development of drilling interpretation technology. The research results of this period are mainly from abroad, covering integrity evaluation, rock mass quality evaluation, weak plane identification, karst development detection and other fields. However, the above researches only consider the role of a small number (1-2) of drilling parameters, which makes the interpretation reliability insufficient, and lacks corresponding research on the noise reduction of original drilling data.

[0004] Since the 21st century, advanced drilling prediction has been paid more and more attention in the field of tunnel karst detection at home and abroad, and drilling data interpretation has become the focus of scholars and engineers. In view of the problems existing in conventional drilling interpretation such as relying on the experience of technical personnel, non-uniform evaluation standard, insufficient use of data and low interpretation efficiency, the current research on drilling data interpretation, i.e. the existing technical scheme, can be divided into the following three categories:

[0005] Firstly, based on the analysis of rock breaking mechanism, this kind of interpretation mostly takes the energy analysis and mechanical balance of drilling process as the starting point. For example, based on the rock breaking mechanism of PDC drill bit, the quantitative relationship model between drilling parameters and rock uniaxial compressive strength and shear strength is derived by using energy analysis method;

[0006] Second, from the perspective of dilution principle, using traditional mathematical statistics as a tool for drilling interpretation analysis. For example, through statistical analysis of drilling data, an index for quantitative quality evaluation of rock mass is constructed; using clustering analysis, linear regression and other methods, a fitting equation of drilling rate and integrity in stratum is established;

[0007] In addition, with the development of information technology such as big data and artificial intelligence, the concept of machine learning has gradually penetrated into various fields, providing a new way for data analysis, so in recent years some researchers have introduced machine learning methods into tunnel drilling data interpretation, which is the third method. For example, the intelligent analysis of drilling test data and stratum identification method based on neural network can accurately identify different strata; using support vector machine (SVM) to study the advance classification of grade III to V surrounding rock; using machine learning to predict the cavity, collapse and fracture zone of the rock mass in front of the tunnel face.

[0008] In general, the third method, i.e. using machine learning model for advance drilling data interpretation, can better consider the nonlinear influence of various digital drilling indicators on the interpretation results, and has high interpretation efficiency, which is the current mainstream development technology. However, the main problem is that a single scale interpretation model is used, which has poor actual interpretation effect, i.e. the single scale model formed by dividing the drilling data at a certain specific division interval and training the data set is easily affected by data outliers and the limitation of data set sample size, resulting in poor interpretation effect in subsequent actual drilling data application. SUMMARY

[0009] In view of the above problem that the single scale interpretation model has poor actual interpretation effect, the present application constructs a multi-scale drilling interpretation model, and based on the multi-scale drilling interpretation model, a tunnel rock mass quality identification method and system based on multi-scale drilling model fusion are proposed.

[0010] In order to achieve the above invention purpose, the present application provides the following technical solutions:

[0011] The tunnel rock mass quality identification method based on multi-scale drilling model fusion specifically comprises the following steps:

[0012] S1, collecting and preprocessing the field drilling data to obtain a test set;

[0013] S2, inputting the test set into a multi-scale fusion model to obtain a tunnel rock mass quality identification result, the tunnel rock mass quality identification result being one of the prediction labels, the prediction labels including relatively complete, relatively broken, broken and fractured;

[0014] The training process of the multi-scale fusion model comprises:

[0015] S21, collecting and preprocessing field drilling data to obtain a training set, wherein the training set is segmented by multiple scales;

[0016] S22, training multiple single-scale interpretation models based on the training set and an XGBoost classifier, and evaluating the multiple single-scale interpretation models to obtain weights of the single-scale interpretation models;

[0017] S23, performing multi-scale model fusion by using an error reciprocal weighted average method based on the weights of the single-scale interpretation models to obtain a multi-scale fusion model.

[0018] As a preferred scheme of the present application, step S21 specifically comprises the following steps:

[0019] S211, obtaining original drilling data;

[0020] S212, performing outlier analysis on the original drilling data and eliminating outliers to obtain optimized original drilling data;

[0021] S213, performing multiple-scale segmentation on the optimized original drilling data to obtain data sets under different segmentation scales;

[0022] S214, performing feature correlation analysis on the data sets under different segmentation scales, eliminating parameters with high correlation to obtain a test set, wherein the test set comprises feature parameters RMT, DRM, TPM, TM and DRV.

[0023] As a preferred scheme of the present application, the multiple-scale segmentation on the optimized original drilling data in step S213 is equidistant segmentation of the optimized original drilling data according to the depth of the drilling rock mass, and the equidistant segmentation scales are 0.5 m, 1 m and 2 m.

[0024] As a preferred scheme of the present application, the prediction label output by the multi-scale fusion model in step S23 is represented by a formula as follows:

[0025] L = ω1L 0.5 + ω2L1+ ω3L2

[0026] wherein, L 0.5 , L1 and L2 represent the prediction label values of the scale models; ω1, ω2 and ω3 represent the weights of the scale models, and L is the prediction label output by the multi-scale fusion model.

[0027] As a preferred scheme of the present application, the calculation formula of the weights ω1, ω2 and ω3 of the scale models is as follows:

[0028] ω1 = (ε2 + ε3) / 2 (ε1 + ε2 + ε3)

[0029] ω2 = (ε1 + ε3) / 2 (ε1 + ε2 + ε3)

[0030] ω3 = (ε1 + ε2) / 2 (ε1 + ε2 + ε3)

[0031] In the formula, ω1, ω2 and ω3 represent the weights of the scale models under three kinds of split intervals; ε1, ε2 and ε3 represent the errors of the scale models under three kinds of split intervals.

[0032] As a preferred scheme of the present application, the values of ω1, ω2 and ω3 are 0.41, 0.34 and 0.25 respectively.

[0033] As a preferred scheme of the present application, in step S22, the plurality of single-scale interpretation models are evaluated, and the evaluation parameters include accuracy, precision, recall and harmonic mean;

[0034] The accuracy calculation formula is:

[0035] The precision calculation formula is:

[0036] The recall calculation formula is:

[0037] The harmonic mean calculation formula is:

[0038] In the formula, TP, TN, FP and FN respectively represent the number of positive samples predicted as positive by the model, the number of negative samples predicted as negative by the model, the number of negative samples predicted as positive by the model, and the number of positive samples predicted as negative by the model.

[0039] Based on the same concept, a tunnel rock mass quality identification system based on multi-scale drilling model fusion is also proposed, which comprises:

[0040] A drilling data acquisition module is configured to collect and preprocess field drilling data to obtain a test set.

[0041] A tunnel rock mass quality identification module is configured to input the test set into a multi-scale fusion model to obtain a tunnel rock mass quality identification result, wherein the tunnel rock mass quality identification result is one of the prediction labels, and the prediction labels include relatively complete, relatively broken, broken and fractured.

[0042] The training process of the multi-scale fusion model comprises:

[0043] S21, collecting and preprocessing field drilling data to obtain a training set, wherein the training set is subjected to data segmentation in multiple scales;

[0044] S22, based on the XGBoost classifier, a plurality of single-scale interpretation models are trained based on the training set, and each single-scale interpretation model is evaluated to obtain the weight of each single-scale interpretation model;

[0045] S23, based on the weight of the plurality of single-scale interpretation models, an error reciprocal weighted average method is used for multi-scale model fusion to obtain a multi-scale fusion model.

[0046] Based on the same concept, a tunnel rock mass quality identification device based on multi-scale drilling model fusion is also proposed, which comprises at least one processor and a memory in communication connection with the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the tunnel rock mass quality identification method based on multi-scale drilling model fusion according to any one of the above.

[0047] Based on the same concept, a computer readable storage medium is also proposed, which stores a computer program, and the computer program is executed by a processor to implement the tunnel rock mass quality identification method based on multi-scale drilling model fusion according to any one of the above.

[0048] Compared with the prior art, the beneficial effects of the present application are:

[0049] The tunnel rock mass quality identification method based on multi-scale drilling model fusion proposed in the present application has the effects of weakening the shortcomings of each scale model, optimizing the prediction effect, and enhancing the practicality of technology and engineering, and under the basic condition of meeting higher prediction accuracy, more detailed geological prediction information can be efficiently and conveniently provided. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 The flowchart of the tunnel rock mass quality identification method based on multi-scale drilling model fusion in embodiment 1 is shown in the figure;

[0051] Figure 2 The flowchart of the training process of the multi-scale fusion model in embodiment 1 is shown in the figure;

[0052] Figure 3 The table for explaining the confusion matrix in embodiment 1 is shown in the figure;

[0053] Figure 4 The specific flowchart from data collection and preprocessing to model construction, evaluation and fusion in embodiment 1 is shown in the figure;

[0054] Figure 5 The pseudo code diagram of GA in embodiment 2 is shown in the figure;

[0055] Figure 6 The visualization diagram of the original drilling data in embodiment 2 is shown in the figure;

[0056] Figure 7 The multiscale equidistant segmentation graph of the original drilling data in Example 2;

[0057] Figure 8 The feature correlation heat map in Example 2;

[0058] Figure 9 The prediction result confusion matrix heat map in Example 2;

[0059] Figure 10 The model interpretation effect of each scale and the model interpretation effect of multiscale model fusion in Example 2. DETAILED DESCRIPTION

[0060] The present application will be further described in conjunction with test examples and specific embodiments. However, it should not be understood that the scope of the above-mentioned subject matter of the present application is limited to the following examples, and any technology implemented based on the content of the present application falls within the scope of the present application.

[0061] Example 1

[0062] The tunnel rock mass quality identification method based on multiscale drilling model fusion, the flow chart is as shown in Figure 1 , and specifically includes the following steps:

[0063] S1, collecting and preprocessing the field drilling data to obtain a test set.

[0064] Based on a plurality of tunnels with different lithology, the original drilling data is collected and standardized noise reduction and preprocessing measures are developed. After multiscale (0.5m, 1m and 2m) segmentation, the training set for machine learning model learning and training and the test set for predicting the quality of tunnel rock mass are formed, and measures such as feature correlation analysis and one-hot encoding are taken to improve the quality of the data set.

[0065] S2, inputting the test set into the multiscale fusion model to obtain a tunnel rock mass quality identification result, the tunnel rock mass quality identification result being one of the prediction labels, and the prediction labels including relatively complete, relatively broken, broken and fractured.

[0066] The XGBoost integrated algorithm is used as a classifier to construct an automatic rock mass quality interpretation model, and the GA algorithm is used to realize automatic optimization of its super parameters, and the performance index of the constructed model is comprehensively evaluated to verify the accuracy and reliability of the model.

[0067] The training process flow chart of the multiscale fusion model is as shown in Figure 2 , and specifically includes the following steps:

[0068] S21, collecting and preprocessing the field drilling data to obtain a training set, and the training set is segmented by multiple scales.

[0069] S22, training each single scale interpretation model based on the XGBoost classifier and based on the training set, and evaluating a plurality of each scale interpretation models to obtain weights of a plurality of single scale interpretation models;

[0070] S23, performing multi-scale model fusion based on the weights of a plurality of single scale interpretation models by using an error reciprocal weighted average method to obtain a multi-scale fusion model.

[0071] The specific flowchart from data collection and preprocessing to model construction, evaluation and fusion is shown in Figure 3 .

[0072] Further, step S21 specifically comprises the following steps:

[0073] S211, obtaining original drilling data;

[0074] S212, performing outlier analysis on the original drilling data and eliminating outliers to obtain optimized original drilling data;

[0075] S213, performing multi-scale segmentation on the optimized original drilling data to obtain data sets under different segmentation scales;

[0076] S214, performing feature correlation analysis on the data sets under different segmentation scales, eliminating parameters with high correlation, and obtaining a test set, wherein the test set includes RMT, DRM, TPM, TM, and DRV feature parameters.

[0077] In step S213, the optimized original drilling data is segmented according to the depth of the drilling rock mass, and the scales of the equidistant segmentation are 0.5m, 1m and 2m.

[0078] Further, the prediction label output by the multi-scale fusion model in step S23 is represented by the formula:

[0079] L = ω1L 0.5 + ω2L1+ ω3L2

[0080] Wherein, L 0.5 , L1 and L2 represent the predicted label values of each scale model; ω1, ω2 and ω3 represent the weights of each scale model, and L is the predicted label output by the multi-scale fusion model.

[0081] Wherein, the calculation formula of the weights ω1, ω2 and ω3 of the scale model is:

[0082] ω1 = (ε2+ε3) / 2(ε1+ε2+ε3)

[0083] ω2 = (ε1 + ε3) / 2 (ε1 + ε2 + ε3)

[0084] ω3 = (ε1 + ε2) / 2 (ε1 + ε2 + ε3)

[0085] In the formula, ω1, ω2, and ω3 represent the weights of the scale models under three kinds of partition intervals; ε1, ε2, and ε3 represent the errors of the scale models under three kinds of partition intervals.

[0086] As a preferred solution, the values of ω1, ω2, and ω3 are 0.41, 0.34, and 0.25, respectively.

[0087] In addition, in step S22, the plurality of single-scale interpretation models are evaluated, and the evaluation parameters include accuracy, precision, recall, and harmonic mean;

[0088] The accuracy calculation formula is:

[0089] The precision calculation formula is:

[0090] The recall calculation formula is:

[0091] The harmonic mean calculation formula is:

[0092] In the formula, TP, TN, FP, and FN represent true positive (the number of positive samples predicted as positive by the model), true negative (the number of negative samples predicted as negative by the model), false positive (the number of negative samples predicted as positive by the model), and false negative (the number of positive samples predicted as negative by the model) in the predicted confusion matrix, respectively.

[0093] For example, a binary classification problem is taken as an example, and a more specific explanation is shown in Table 1. Figure 3 Figure 3 The confusion matrix is explained as shown in Table 1.

[0094] Corresponding to the present application, because it is a four-class multi-classification problem, positive and negative are a relative concept. Taking the predicted label “more complete” as an example, the actual or predicted label is “more complete” and is positive, and the actual or predicted label is other class labels and is negative. In this way, TP, TN, FP, and FN for the “more complete” label can be calculated, and then TP, TN, FP, and FN for the other three labels are calculated according to the same method, and then the average of the four classes is calculated to obtain the overall TP, TN, FP, and FN for the multi-classification problem, and finally the PRE and REC are calculated by bringing the formula.

[0095] Example 2

[0096] According to​Figure 4 As shown, a specific tunnel rock mass quality identification method based on multi-scale drilling model fusion specifically includes the following steps:

[0097] 1. Introduction of the model and algorithm used

[0098] (1) XGBoost machine learning model

[0099] The decision tree (Decsion tree) model is one of the most commonly used models in the field of machine learning due to its good interpretability and other advantages, and some ensemble algorithms based on decision trees have also been studied and valued, and XGBoost is one of them. As shown in equation 1, XGBoost corrects the prediction results of a large number of single decision trees through serial training, and finally obtains a more accurate result.

[0100]

[0101] In the formula: represents the prediction result of the entire model on this sample, K represents the total number of decision trees, f k represents the kth decision tree, x i represents the feature vector corresponding to sample i.

[0102] XGBoost introduces an objective function Obj to measure the operation efficiency of the algorithm, and the formula is shown in equation (2):

[0103]

[0104] In the formula: n is the total number of decision trees; is the loss function part, which is used to measure the gap between the predicted label and the true label; is the regularization term used to represent the model complexity.

[0105] In terms of specific classification model performance evaluation, the accuracy (ACC), precision (PRE), recall (REC) and harmonic mean (F1) four indicators are used for comprehensive evaluation to consider the model performance deviation caused by factors such as sample imbalance. The specific calculation formulas are shown in equations (3) to (6):

[0106]

[0107]

[0108]

[0109]

[0110] In the formula: TP, TN, FP, FN represent the number of true positive, true negative, false positive and false negative samples in the prediction confusion matrix respectively.

[0111] (2) GA genetic algorithm

[0112] Genetic algorithm (Genetic Algorithm) as a kind of random search algorithm inspired by the natural selection and natural genetic mechanism of biological world, it simulates the reproduction, crossing and gene mutation phenomenon occurred in the process of natural selection and natural genetic, using genetic operators (including selection, crossing and mutation operation) to combine these individuals, produce a new generation of candidate population, and repeat the process until the convergence condition is met.

[0113] The GA pseudo code is shown in Figure 5 , Figure 5 In the formula: maxf(X) represents the maximum constraint optimization problem to be solved, d is the size or dimension of the problem, N is the initial population individual, p c and p m The cross probability and mutation probability in the crossing and mutation process, MaxIt is the iteration evolution number of the algorithm, P1(t) and P2(t) are temporary populations generated by iteration, (Z i (t)) and B(t) are the individuals and the optimal individual in the genetic evolution after the tth iteration.

[0114] (3) Error reciprocal method

[0115] In the prediction practice, there are usually different prediction methods for the same problem. The combination of these different methods with appropriate weighting coefficients can obtain the combined prediction method. The combination prediction can reduce the influence of random factors of each single method, and comprehensively utilize the information provided by various methods, so the combination prediction can more effectively improve the prediction accuracy than the single method. Determining the weighting coefficient of each single prediction method is the key of the combination prediction.

[0116] The error reciprocal method is a method for determining the weighting coefficient in the combination prediction. In the error reciprocal method, the prediction model with smaller error is given larger weight, so the overall error of the combined prediction model can be significantly reduced. To determine the weight coefficient, the error reciprocal method is used to assign the weight to the model, and the formula is shown in formula (7)-(9):

[0117] ω1=(ε2+ε3) / 2(ε1+ε2+ε3) (7)

[0118] ω2=(ε1+ε3) / 2(ε1+ε2+ε3) (8)

[0119] ω3=(ε1+ε2) / 2(ε1+ε2+ε3) (9)

[0120] where ω1, ω2, and ω3 represent the weights of each scale model under 0.5m, 1m, and 2m partitioning interval; ε1, ε2, and ε3 represent the error of each scale model under 0.5m, 1m, and 2m partitioning interval.

[0121] 2. Raw drilling data preprocessing

[0122] 2.1 Raw drilling data source situation

[0123] Guangxi is located in the southeast edge of Yunnan-Guizhou Plateau, the overall landform is mountainous and hilly basin, and the tunnel site area is developed with unfavorable geological bodies such as karst, soft interlayer, etc. In order to ensure construction safety, Casagrande C6-2 type multifunctional drill is used for advanced drilling geological prediction on site.

[0124] According to the advanced drilling operation, four tunnels (two of limestone and two of sandstone) are obtained, and more than 20,000 pieces of raw drilling data are collected, covering a total length of about 400 meters (the sampling frequency of digital drilling system is about 50 pieces / meter). During the drilling prediction process, the technical personnel mainly interpret and mark four types of quality rock mass, i.e. "relatively complete", "relatively broken", "broken" and "fracture / cavity", according to the raw drilling data and the site conditions (among them, the unfavorable geological types such as argillaceous filling cave, soft interlayer and dissolution fracture can be further judged according to the position, distribution, drilling backwater color, etc. of "fracture / cavity").

[0125] 2.2 Analysis of abnormal values of raw drilling data

[0126] Under the digital drilling system, each piece of raw drilling data mainly includes four drilling indicators in addition to the depth, i.e. the pushing speed, pushing force, torque and rotating speed. The representative raw drilling data (10m of a certain tunnel from YK2+621 to YK2+611, as shown in FIG. 1) are selected for feature analysis, and the abnormal values mainly have the following three aspects: Figure 6

[0127] (1) Sampling stage: the digital drilling system shows obvious stage division in the raw drilling data sampling process, i.e. the rising section in the early drilling stage and the normal sampling stage afterwards, wherein the rising section is usually concentrated in the range of 0-1m footage, which mainly involves the drill emptying, initial concrete spraying and the rock mass section affected by the drilling and blasting disturbance;

[0128] (2) Large dispersion degree of individual indicators: the values of pushing speed and torque show a large dispersion, which is also the main criterion of the technical personnel in the conventional interpretation process;

[0129] ​(3) Different types of outliers exist: During the normal sampling phase, there are two main types of outliers. One is that the drilling index appears null due to the operator's operation and mechanical factors. The other is that the propulsion speed sampling value will experience a very brief surge due to the rod changing operation.

[0130] 2.3 Preprocessing measures for raw drilling data

[0131] (1) Multi-scale segmentation

[0132] To mitigate the adverse effects of surging outliers and fully reflect the potential relationship between dispersion and drilled rock mass quality, multi-scale equidistant segmentation and secondary index calculation were performed. Specifically, considering practical engineering experience and forecast accuracy requirements, the equidistant segmentation scale *d* was set to 0.5m, 1m, and 2m. Secondary calculations were then performed on the data within each segment of the index to form secondary indices, which serve as features for the machine learning dataset. Specifically, the mean and variance were selected as secondary indices for the following reasons:

[0133] ① Mean x: The original drilling data of different rock mass qualities all have a certain range of values. The mean is an important manifestation of this range. Moreover, by taking the mean, the impact of abnormal data within the segmentation interval on the overall true data can be reduced, thereby achieving further noise reduction of the original drilling data.

[0134] ② Variance S n 2 By calculating the variance of the original drilling data within the segmentation interval, the dispersion of the original drilling data of rock masses of different qualities can be reflected more scientifically and reasonably, and the regularity of the relationship between the dispersion and the quality of the drilled rock mass can be discovered and quantitatively expressed.

[0135] The mean and variance within each segment are calculated as shown in equations (10) to (11):

[0136]

[0137]

[0138] In the formula: n is the number of samples within the segment; x i and These represent the values ​​of each sample and the overall sample mean.

[0139] by Figure 6 Taking the propagation speed of subgraph (a) as an example, the effect of the above noise reduction operation is as follows: Figure 7 As shown.

[0140] After denoising and multi-scale segmentation, the original drilling data forms a machine learning dataset corresponding to three segmentation scales of 0.5 m, 1 m and 2 m, with sample data of 968, 472 and 235 respectively. All datasets contain 9 features (RMT is the lithology category, DRM is the average push speed, TPM is the average push speed, TM is the average torque, RM is the average rotation speed, DRV is the push speed variance, TPV is the push speed variance, TV is the torque variance, and RV is the rotation speed variance) and 4 labels (more complete, more broken, broken, and fracture / cave). Taking the 0.5 m scale segmented dataset as an example, the feature description is shown in Table 1.

[0141] Table 1: Dataset features and their numerical descriptions

[0142] Features Min Max Median Mean RMT / / / / DRM 12.39 207.80 31.44 49.16 TPM 96.35 109.09 102.39 103.03 TM 35.14 86.77 49.51 52.25 RM 64.60 182.01 144.99 139.61 DRV 1.24 2555.04 37.43 178.46 TPV 0.52 398.52 1.35 3.79 TV 0.17 153.11 3.38 10.02 RV 0.84 260.41 13.65 26.57

[0143] (2) Correlation analysis of dataset features

[0144] To reduce the feature dimension of the dataset and speed up the subsequent interpretation model fitting, the correlation analysis of the 8 secondary indicators of the machine learning dataset is performed to eliminate the indicators with high correlation. The correlation coefficient p between the features is calculated as shown in equation (12):

[0145]

[0146] In the formula, a i and b i are the sample values of the two features; a and b are the sample means of the two features. The value of the correlation coefficient p is between -1 and +1, close to +1 indicates strong positive correlation, close to -1 indicates strong negative correlation, and 0 indicates no correlation.

[0147] According to the calculated correlation, the heat map is drawn as shown in Figure 8 . Setting ±0.6 as the feature correlation analysis and screening threshold, it can be found that RMT and RM, TV and RV have strong positive correlation, with values of 0.68 and 0.78 respectively, and RMT and TV, DRV and TPM, as well as TPM and TPV have strong negative correlation, with values of -0.71, -0.76, -0.65 and 0.62 respectively. Considering the above, the features RM, TPV, TV and RV are removed, and the final machine learning dataset retains 5 features, which are RMT, DRM, TPM, TM and DRV.

[0148] (3) One-hot encoding of dataset features

[0149] To make the machine learning model identify non-numerical features, the feature RMT is one-hot encoded. Compared with traditional numerical encoding (such as encoding limestone and sandstone as "0" and "1" respectively), one-hot encoding can be expressed in the form of a matrix, thereby avoiding the influence of the mathematical numerical properties of numerical encoding on subsequent machine model training. The one-hot encoding of limestone and sandstone is [0, 1] and [1, 0] respectively.

[0150] 3. Single scale interpretation model construction and evaluation

[0151] (1) GA-XGBoost automatic interpretation model construction

[0152] To achieve efficient and accurate automatic interpretation of drilling data, GA-XGBoost model is formed by taking XGBoost as the classifier and combining GA algorithm. GA can efficiently determine its preferred super parameter combination. The specific settings are as follows: in terms of population size, the initial population size is 10 and the population iteration is 20 times; in terms of population fitness measurement index, it is defined as 5-fold cross-validation ACC, and the calculation formula is shown in formula (3) above; in terms of individual mating form, uniform crossover is adopted, and the individual mutation probability is set to 0.1 to increase species diversity.

[0153] In terms of super parameter value range, according to related research on XGBoost model, the most important 6 super parameters are selected as optimization objects, and the specific super parameters and their optimization space are as follows: n_estimators, max_depth are (10, 200, 1) and (3, 10, 1) respectively, learning_rate, min_child_weight, subsample and colsample_bytree are all (0, 1, 0.01), the first two values in the parentheses are the minimum and maximum values respectively, and the last one is the value particle density. The iteration convergence condition is shown in formula (13).

[0154] fit(n) max -fit(n-1) max = 0 (13)

[0155] In the formula: fit(n) max is the fitness value of the current optimal individual; fit(n-1) max is the fitness value of the previous optimal individual. That is, when the fitness value of the optimal individual in the population does not change during iteration, it is considered that the iteration converges.

[0156] The information carried by the optimal individuals of each scale after iteration (i.e. the XGBoost model super parameter combination) is selected as shown in Table 2 below.

[0157] Table 2 Data set features and their numerical description

[0158] Hyperparameters 0.5m 1.0m 2.0m n_estimators 106 83 72 max_dept 7 8 8 learning_rate 0.21 0.30 0.50 min_child_weight 0.45 0.13 0.53 subsample 0.68 0.45 0.37 colsample_bytree 0.95 0.91 0.72

[0159] (2) Each scale interpretation model performance comprehensive evaluation

[0160] In order to reasonably evaluate the interpretation model, the ratio of training set and prediction set is set to 7:3, and the comprehensive index of the model is evaluated according to formula (3) to formula (6). At the same time, the evaluation results of the XGBoost model without hyperparameter optimization are taken as the comparison (taking the default hyperparameters of the model), and the results are shown in table 3.

[0161] Table 3 Data set characteristics and their digital description

[0162]

[0163] From table 3, it can be seen that the model performance after parameter optimization is generally higher than that of the XGBoost model without parameter optimization and using default parameters. In order to further explore the classification and prediction of specific models, the heat map of the prediction confusion matrix of each scale model is output, and the results are shown in Figure 9 .

[0164] Figure 9 The horizontal coordinate is the actual sample label, and the vertical coordinate is the predicted sample label. Matching the above-mentioned second point rule, the error prediction of each scale classification model is mainly concentrated in predicting "more complete" as "more broken", and with the increase of the segmentation scale, this prediction trend is more and more obvious. The main reasons are as follows: first, according to engineering practice experience, there is a large classification difficulty between the two types of rock mass, which leads to the mislabeling of the collected original drilling data; second, with the increase of the segmentation scale, the mean and variance calculated in the "more complete" segmentation paragraph are more likely to be confused with "more broken", so that the prediction accuracy decreases. From the perspective of safety bias in tunnel engineering construction, this prediction trend is acceptable.

[0165] 4, Multi-scale drilling interpretation model fusion

[0166] When collecting data, in order to improve the quality of data, more drilling data with less noise and obvious label characteristics are selected, which affects the interpretation effect of the model in actual and continuous original drilling data. Therefore, in order to improve the engineering practicability of the model, the present application proposes an automatic drilling rock mass quality interpretation model based on error reciprocal weighted average multi-scale model fusion, which can greatly fuse the interpretation effects of three scales, so that the final interpretation can improve the prediction integrity while ensuring the accuracy to optimize the interpretation effect. The fusion rules set follow the following two points:

[0167] (1) The four interpretation labels are coded as quantitative values (coded as 1, 2, 3, 4), and the average calculation of the predicted labels is carried out according to formula (14):

[0168] L = ω1L 0.5 + ω2L1+ ω3L2 (14)

[0169] In the formula: L 0.5 , L1 and L2 represent the predicted label values of each scale model; ω1, ω2 and ω3 represent the weights of each scale model.

[0170] According to the ACC index in Table 3, the errors of each scale model are 3.09%, 5.68% and 8.57% respectively. Then, the weights of each scale model are calculated by formula (7)-(9), and the final multi-scale model fusion calculation formula is shown in formula (15):

[0171] L = 0.41L 0.5 + 0.34L1+ 0.25L2 (15)

[0172] (2) From the perspective of tunnel construction safety, special rules are formulated, that is, when the model prediction label is "4" (i.e. "fracture / cave") at the scale of 0.5m, the fusion prediction label is "4" as the final label.

[0173] Example 3

[0174] The data of 1-20m in No.1 hole of a mountain tunnel ZK109+403 working face drilling are selected to illustrate the actual prediction effect of the model, and an automatic rock mass quality interpretation method based on error reciprocal weighted average of multi-scale model fusion is proposed according to the interpretation effect analysis, as shown in Figure 10 .

[0175] The original drilling data is shown in Figure 10 subgraphs (a)-(c), the automatic interpretation effect of each scale model is shown in Figure 10 subgraphs (d)-(f), and the multi-scale fusion interpretation effect proposed by the application is shown in Figure 10 subgraph (g). It can be found that the single scale model has the following two shortcomings when applied in the actual continuous original drilling data:

[0176] (1) 0.5m scale due to the segmentation interval is too small, resulting in the model on the two important features of DRV and DRV numerical changes too sensitive, resulting in two phenomena: ① a small number of original data points of the surge or sharp drop is prone to cause its prediction results in "more complete", "more broken" and "broken" three labels change, on the one hand, prone to cause the wrong judgment due to abnormal data, on the other hand, the integrity of the overall prediction effect is poor, such as 1-5m, 9.5-10m prediction effect; ② due to the above reasons, the prediction of "fracture / cavity" label is very accurate, some thin layer of weak interlayer or fracture / cavity can also be predicted, such as 12-13.5m. The former is not conducive to engineering practical application, and the latter is more conducive to tunnel construction safety.

[0177] (2) In 1m and 2m scale, due to the increase of segmentation interval, the prediction effect is better than 0.5m scale, and the corresponding disadvantage is that the prediction accuracy is insufficient, such as 1m scale 13-13.5m "fracture / cavity" label prediction, 2m scale 14-15m "fracture / cavity" label prediction.

[0178] In combination with the above two points, as Figure 10 As can be seen from subgraph (g), the multi-scale model fusion method based on error reciprocal weighted average proposed in the application can weaken the shortcomings of each scale model, optimize the prediction effect, and enhance the technical engineering practicability. Under the condition of meeting the basic condition of higher prediction accuracy, more detailed geological prediction information can be efficiently and conveniently provided.

[0179] Embodiment 3

[0180] Based on the same concept, a tunnel rock mass quality identification system based on multi-scale drilling model fusion is also proposed, comprising:

[0181] The drilling data acquisition module is used for collecting and preprocessing the field drilling data to obtain a test set;

[0182] The tunnel rock mass quality identification module is used for inputting the test set into the multi-scale fusion model to obtain a tunnel rock mass quality identification result, wherein the tunnel rock mass quality identification result is one of the prediction labels, and the prediction label includes more complete, more broken, broken and fracture;

[0183] The training process of the multi-scale fusion model comprises:

[0184] S21, collecting and preprocessing the field drilling data to obtain a training set, wherein the training set is segmented by multiple scales;

[0185] S22, based on the XGBoost classifier, a plurality of single-scale interpretation models are trained based on the training set, and each single-scale interpretation model is evaluated to obtain the weight of each single-scale interpretation model;

[0186] S23, based on the weight of the plurality of single-scale interpretation models, an error reciprocal weighted average method is used for multi-scale model fusion to obtain a multi-scale fusion model.

[0187] Based on the same idea, a tunnel rock mass quality identification device based on multi-scale drilling model fusion is also proposed, which comprises at least one processor and a memory in communication connection with the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute any one of the tunnel rock mass quality identification methods based on multi-scale drilling model fusion.

[0188] Based on the same idea, a computer readable storage medium is also proposed, which stores a computer program, and the computer program is executed by a processor to implement the tunnel rock mass quality identification method based on multi-scale drilling model fusion.

[0189] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A tunnel rock mass quality identification method based on multi-scale drilling model fusion, characterized in that, Specifically comprising the following steps: S1, collecting and preprocessing the field drilling data to obtain a test set; S2, inputting the test set into a multi-scale fusion model to obtain a tunnel rock mass quality identification result, the tunnel rock mass quality identification result being one of the prediction labels, and the prediction labels including relatively complete, relatively broken, broken and fractured; The training process of the multi-scale fusion model comprises: S21, collecting and preprocessing the field drilling data to obtain a training set, and the training set being segmented by multiple scales; S22, training multiple single-scale interpretation models based on the training set and an XGBoost classifier, and evaluating the multiple single-scale interpretation models to obtain the weights of the single-scale interpretation models; S23, based on the weights of the single-scale interpretation models, performing multi-scale model fusion by using an error reciprocal weighted average method to obtain a multi-scale fusion model; Step S21 specifically comprises the following steps: S211, obtaining original drilling data; S212, performing outlier analysis on the original drilling data and eliminating outliers to obtain optimized original drilling data; S213, segmenting the optimized original drilling data by multiple scales to obtain data sets under different segmentation scales; S214, performing feature correlation analysis on the data sets under different segmentation scales, eliminating parameters with high correlation, and obtaining a training set, wherein the feature parameters included in the training set are lithology category, average pushing speed, average pushing speed, average torque, and pushing speed variance; The step of screening the feature parameters comprises: Obtaining initial features, including: lithology category, average pushing speed, average pushing speed, average torque, average rotation speed, pushing speed variance, pushing speed variance, torque variance, and rotation speed variance; First, calculate the correlation coefficients of the initial features, then select parameters with high positive correlation and parameters with high negative correlation as the screening threshold is 0.6, select representative parameters from the parameters with high positive correlation, and select representative parameters from the parameters with high negative correlation, to obtain 5 feature parameters; In step S22, the XGBoost is used as a classifier, and a GA-XGBoost model is formed by combining a GA algorithm; in the process of optimizing the hyperparameters of the XGBoost classifier by using the GA algorithm, 6 hyperparameters are selected as the optimization objects, and the specific hyperparameters and their optimization spaces are: n_estimators, max_depth, learning_rate, min_child_weight, subsample, and colsample_bytree.

2. The tunnel rock mass quality identification method based on multi-scale drilling model fusion according to claim 1, characterized in that, In step S213, the optimized original drilling data is segmented by multiple scales according to the depth of the drilling rock mass, and the scales of the equidistant segmentation are 0.5m, 1m and 2m.

3. The tunnel rock mass quality identification method based on multi-scale drilling model fusion according to claim 2, characterized in that, In step S23, the prediction label output by the multi-scale fusion model is represented by the following formula: wherein, L 0.5 , L 1 and L 2 represent the predicted label values of each scale model; , , and represent the weights of each scale model, and L is the predicted label output by the multi-scale fusion model.

4. The tunnel rock mass quality identification method based on multi-scale drilling model fusion as described in claim 3, wherein the weights of the scale models are... , ,and The calculation formula is: wherein , , represent the weights of each scale model under three partition intervals; , and represent the errors of each scale model under three partition intervals.

5. The tunnel rock mass quality identification method based on multi-scale drilling model fusion according to claim 4, characterized in that, , , and 0.41, 0.34, 0.25, respectively.

6. The tunnel rock mass quality identification method based on multi-scale drilling model fusion according to claim 1, characterized in that, In step S22, the plurality of single scale interpretation models are evaluated, and the evaluation parameters include accuracy, precision, recall, and harmonic mean; The accuracy calculation formula is: ; The precision calculation formula is: ; The recall rate is calculated by the formula: ; The harmonic mean is calculated by the formula: ; In the formula, TP, TN, FP, and FN represent the number of positive samples predicted as positive by the model, the number of negative samples predicted as negative by the model, the number of negative samples predicted as positive by the model, and the number of positive samples predicted as negative by the model in the prediction confusion matrix.

7. A tunnel rock mass quality identification system based on multi-scale drilling model fusion, characterized in that, The method comprises the following steps: The drilling data acquisition module is configured to collect and preprocess field drilling data to obtain a test set; The tunnel rock mass quality identification module is configured to input the test set into the multi-scale fusion model to obtain a tunnel rock mass quality identification result, wherein the tunnel rock mass quality identification result is one of prediction labels, and the prediction labels include relatively complete, relatively broken, broken, and fractured. The training process of the multi-scale fusion model comprises the following steps: S21, collecting and preprocessing field drilling data to obtain a test set, wherein the training set is segmented by multiple scales; S22, training a plurality of single scale interpretation models based on the training set by using an XGBoost classifier, and evaluating the plurality of single scale interpretation models to obtain the weights of the plurality of single scale interpretation models; S23, performing multi-scale model fusion by using an error reciprocal weighted average method based on the weights of the plurality of single scale interpretation models to obtain a multi-scale fusion model; The step of screening the feature parameters comprises the following steps: The initial features include lithology category, average pushing speed, average pushing speed, average torque, average rotation speed, pushing speed variance, pushing speed variance, torque variance, and rotation speed variance; The correlation coefficients of the initial features are calculated, then the parameters with high positive correlation and the parameters with high negative correlation are selected by taking 0.6 as a screening threshold, the representative parameters are selected from the parameters with high positive correlation, and the representative parameters are selected from the parameters with high negative correlation, thereby obtaining five feature parameters; In step S22, the GA-XGBoost model is formed by using XGBoost as a classifier and combining a GA algorithm; in the process of optimizing the hyperparameters of the XGBoost classifier by using the GA algorithm, six hyperparameters are selected as optimization objects, and the specific hyperparameters and their optimization spaces are as follows: n_estimators, max_depth, learning_rate, min_child_weight, subsample, and colsample_bytree.

8. A tunnel rock mass quality identification device based on multi-scale drilling model fusion, characterized in that, The at least one processor is in communication connection with the memory; The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the tunnel rock mass quality identification method based on multi-scale drilling model fusion according to any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the tunnel rock mass quality identification method based on multi-scale drilling model fusion according to any one of claims 1 to 6.

Images