Lithology data drive identification method based on new Austrian tunnel excavation while-drilling data

Abstract

The invention discloses a lithology data drive identification method based on new Austrian tunnel excavation while-drilling data, and relates to the technical field of lithology identification. According to the method, an ADASYN oversampling method is adopted to process the MWD data, and the problem of multi-dimensional data imbalance of the MWD data is solved; hyper-parameter optimization is carried out on the lithology identification model by using a grid search algorithm, so that the generality capability and the identification capability of the identification model are further improved; and the optimized GS-ADASYN-XGBoost is selected to classify the measurement while drilling (MWD) data, so that the accuracy of an identification result can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of lithology identification, and in particular to a lithology data driven identification method based on New Austrian Tunneling Method (NATM) tunnel excavation while drilling data. Background Art

[0002] Measurement while drilling (MWD) technology can measure the operating parameters of the drilling rig in real time during the drilling process, thereby providing specific information on the characteristics of the rock mass. Using measurement while drilling parameters for rock identification is a comprehensive technology with the characteristics of high precision, multi-information, and integration, which can significantly improve the identification efficiency. Over the years, the measurement while drilling technology has been widely used in the rock excavation process. After years of development, the measurement while drilling technology has been gradually improved, which has promoted the real-time nature of data acquisition and the process of characterization research between drilling parameters and rock properties. In terms of revealing the relationship between drilling data and rock mass characteristics, it is mainly summarized from two aspects: methods and discrimination targets. At the method level, domestic and foreign scholars have conducted a lot of exploration. It is worth noting that the problem of data imbalance has received more and more attention in the current ML-based rock feature research. Due to the uncertainty of rock type distribution in actual geological conditions, the number of samples of different rocks will inevitably be unbalanced. When the class labels are unbalanced, two main ideas are used to circumvent this phenomenon, namely, increasing the number of minority samples by upsampling or reducing the number of majority samples by downsampling. To conduct in-depth analysis. However, some studies deal with it by expanding the number of minority class samples to the same as that of majority class samples. This can alleviate data imbalance and improve prediction performance to a certain extent. However, when the number of class labels is highly unbalanced, simply making the number of samples in the two classes the same may lead to unreliable prediction results. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a lithology data-driven identification method based on New Austrian Tunneling Method (NATM) tunnel excavation while drilling data, which solves the problem that the prior art cannot effectively alleviate the low prediction accuracy and reliability caused by data imbalance.

[0004] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0005] A lithology data-driven identification method based on New Austrian Tunneling Method (NATM) tunnel excavation while drilling data is provided, which includes the following steps:

[0006] S1, real-time collection of MWD data of the excavation process of the new Austrian tunnel to be identified, i.e., multi-dimensional unbalanced data;

[0007] S2. Preprocess the multidimensional unbalanced data to obtain preprocessed multidimensional unbalanced data;

[0008] S3, using ADASYN oversampling algorithm to process the preprocessed multidimensional unbalanced data to obtain oversampled MWD data;

[0009] S4, constructing a lithology identification model;

[0010] S5. Input the oversampled MWD data into the lithology identification model for processing to obtain lithology identification results.

[0011] Furthermore, the multi-dimensional imbalance data in step S1 includes drilling depth, driving speed, impact pressure, driving pressure, rotation pressure, water pressure and water flow.

[0012] Further, step S3 includes the following steps:

[0013] S3-1, taking the preprocessed multi-dimensional unbalanced data as sample data, and counting the data volume of each type of sample data;

[0014] S3-2, determine whether the amount of each type of sample data is less than the quantity threshold; if so, determine that type of sample data is a minority type sample; otherwise, determine that type of sample data is a majority type sample;

[0015] S3-3, using the K nearest neighbor algorithm to determine the k nearest neighbor sample data corresponding to each minority class sample in the minority class data space composed of the minority class samples;

[0016] S3-4, according to the formula:

[0017]

[0018] Get the density r of each minority class sample i ; where n i represents the i-th nearest neighbor sample data, and k represents the total number of nearest neighbor sample data;

[0019] S3-5, according to the formula:

[0020]

[0021] Get the number of synthetic samples G that need to be generated i ; Among them, N maj Represents the total amount of sample data for most categories, N min represents the total amount of minority class sample data, β represents the proportion of generating new samples, and ∑(·) represents the summation function;

[0022] S3-6, based on the number of synthetic samples G to be generated i , generate corresponding synthetic samples for each minority category sample through random linear interpolation;

[0023] S3-7, merge each minority class sample and the corresponding synthetic sample respectively;

[0024] Among them, the oversampled MWD data includes majority category samples and merged minority category samples.

[0025] Furthermore, the lithology identification model in step S4 adopts an XGBoost model.

[0026] Furthermore, the training process of the lithology recognition model includes the following steps:

[0027] S5-1, obtaining training multidimensional imbalanced data and its labels, and oversampling using the same method as step S3 to obtain oversampled training MWD data;

[0028] S5-2. Use the GS grid search method to process the hyperparameters of the XGBoost model and obtain the corresponding hyperparameter combination scheme;

[0029] S5-3. Randomly select a hyperparameter combination and apply it to the XGBoost model to obtain the corresponding initial XGBoost model;

[0030] S5-4, input the oversampled training MWD data into the initial XGBoost model and process it using ten-fold cross validation to obtain ten initial indicator data values, namely MacroF1 values;

[0031] S5-5. Calculate the index mean of the ten initial index data values;

[0032] S5-6, repeating steps S5-3 to S5-6 until the mean values ​​of the indicators corresponding to all hyperparameter combinations are obtained;

[0033] S5-7. Select the best hyperparameters based on the mean of each indicator;

[0034] S5-8. Adjust the parameters of the initial training XGBoost model based on the optimal hyperparameters to obtain the trained and optimized GS-ADASYN-XGBoost model.

[0035] Furthermore, the specific process of the GS grid search method is:

[0036] Determine the hyperparameters to be optimized for the XGBoost model; set the value range of the hyperparameters to be optimized, and arrange and combine the values ​​of the hyperparameters to be optimized to generate a hyperparameter grid; initialize the XGBoost model and set the initial hyperparameters; use the GridSearchCV method to perform a grid search on the parameter grid to obtain the corresponding hyperparameter combination solution.

[0037] Further, step S5-4 includes the following steps:

[0038] S5-4-1, divide the oversampled training MWD data into ten subsets;

[0039] S5-4-2. Randomly select a subset as the validation subset, and the remaining nine subsets as the training subsets;

[0040] S5-4-3, input the training subset and its labels into the initial XGBoost model for training, and obtain the trained initial XGBoost model;

[0041] S5-4-4, input the verification subset into the trained initial XGBoost model to obtain the corresponding training recognition result;

[0042] S5-4-5. Calculate the corresponding initial indicator data value based on the training recognition result;

[0043] S5-4-6. Repeat steps S5-4-2 to S5-4-5 until ten initial indicator data values ​​are obtained.

[0044] Furthermore, the labels in step S5-1 are andesite, mixed andesite-sandstone, dolomite, sandstone, siltstone, and mudstone, which are set as category 0, category 1, category 2, category 3, category 4, and category 5, respectively.

[0045] The beneficial effects of the present invention are as follows: the method adopts the ADASYN oversampling method to process the MWD data of the measurement while drilling, thereby solving the problem of multi-dimensional data imbalance of the MWD data of the measurement while drilling; and uses the grid search algorithm to optimize the hyperparameters of the lithology identification model, thereby further improving the generalization and identification ability of the identification model; and selects the optimized GS-ADASYN-XGBoost to classify the MWD data of the measurement while drilling, thereby effectively improving the accuracy of the identification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a method flow chart;

[0047] Figure 2 is the distribution histogram of MWD excavation data; among them, (a) is the distribution histogram of drilling depth, (b) is the distribution histogram of advancement speed, (c) is the distribution histogram of impact pressure, (d) is the distribution histogram of advancement pressure, (e) is the distribution histogram of rotation pressure, (f) is the distribution histogram of water pressure, and (g) is the distribution histogram of water flow;

[0048] Figure 3 The MacroF1 curves of different models at different oversampling ratios;

[0049] Figure 4 Schematic diagram of the impact of different optimization algorithms on the MacroF1 score of lithology classification;

[0050] Figure 5(a) is a schematic diagram of the AUC curve;

[0051] Figure 5(b) is a schematic diagram of the AUC curve;

[0052] Figure 5(c) is a schematic diagram of the AUC curve. DETAILED DESCRIPTION

[0053] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0054] like Figure 1 As shown, a lithology data driven identification method based on New Austrian Tunneling Method (NATM) tunnel excavation while drilling data includes the following steps:

[0055] S1, real-time collection of MWD data of the excavation process of the new Austrian tunnel to be identified, i.e., multi-dimensional unbalanced data;

[0056] The multi-dimensional imbalance data in step S1 include drilling depth, driving speed, impact pressure, driving pressure, rotation pressure, water pressure and water flow.

[0057] S2. Preprocess the multidimensional unbalanced data to obtain preprocessed multidimensional unbalanced data;

[0058] The preprocessing is normalization, which converts the multi-dimensional imbalanced data into the range of [0,1] to ensure that the data are on the same scale. This can subsequently improve the prediction accuracy of the machine learning model and accelerate the optimization process.

[0059] S3, using ADASYN oversampling algorithm to process the preprocessed multidimensional unbalanced data to obtain oversampled MWD data;

[0060] There are two approaches to the problem of imbalanced category data: Data level: mainly undersampling and oversampling. Algorithm level: considering the cost difference optimization algorithm of different misclassification situations, mainly based on cost sensitive learning algorithm (Cost Sensitive Learning), such as Adacost algorithm. This method performs sampling processing from the data perspective. Since the ratio of the maximum and minimum number of lithology categories is about 16:1 (Andesite:Chlorite), the data distribution has very few Chlorite samples, undersampling may lead to data loss, which is not conducive to model learning, so oversampling is selected.

[0061] When oversampling the number of minority class samples, the key parameter is the oversampling ratio proportion. Different values ​​of the oversampling ratio proportion will affect the prediction accuracy. The main principle of the oversampling ratio proportion to control the oversampling process is: define the largest number of majority class samples n_maj, then match each minority class sample n_min with it, and oversample according to the set ratio. Specifically, the number of minority class samples generated is the oversampling ratio proportion multiplied by the difference between the majority class samples and the minority class samples. The corresponding formula is as follows:

[0062] generate_n i =proportion×(n_maj-n_min)

[0063] Among them, generate_n i Represents the number of samples generated by the i-th minority class.

[0064] Step S3 includes the following steps:

[0065] S3-1, taking the preprocessed multi-dimensional unbalanced data as sample data, and counting the data volume of each type of sample data;

[0066] S3-2, determine whether the amount of each type of sample data is less than the quantity threshold; if so, determine that type of sample data is a minority type sample; otherwise, determine that type of sample data is a majority type sample;

[0067] S3-3, using the K nearest neighbor algorithm to determine the k nearest neighbor sample data corresponding to each minority class sample in the minority class data space composed of the minority class samples;

[0068] S3-4, according to the formula:

[0069]

[0070] Get the density r of each minority class sample i ; where n i represents the i-th nearest neighbor sample data, and k represents the total number of nearest neighbor sample data;

[0071] S3-5, according to the formula:

[0072]

[0073] Get the number of synthetic samples G that need to be generated i ; Among them, N maj Represents the total amount of sample data for most categories, N min represents the total amount of minority class sample data, β represents the proportion of generating new samples, and ∑(·) represents the summation function;

[0074] S3-6, based on the number of synthetic samples G to be generated i , generate corresponding synthetic samples for each minority category sample through random linear interpolation;

[0075] S3-7, merge each minority class sample and the corresponding synthetic sample respectively;

[0076] Among them, the oversampled MWD data includes majority category samples and merged minority category samples.

[0077] S4, constructing a lithology identification model;

[0078] The lithology identification model in step S4 adopts the XGBoost model. The XGBoost sub-model includes multiple decision tree units, which are used as weak learners. The XGBoost sub-model uses the gradient descent architecture to enhance the learner. In each iteration, XGBoost fits a new classifier based on the residual of the previous round to improve the model performance. To avoid overfitting, XGBoost introduces regularization and weighted leaf node statistics.

[0079] S5. Input the oversampled MWD data into the lithology identification model for processing to obtain lithology identification results.

[0080] The training process of the lithology recognition model includes the following steps:

[0081] S5-1, obtaining training multidimensional imbalanced data and its labels, and oversampling using the same method as step S3 to obtain oversampled training MWD data;

[0082] The labels in step S5-1 are andesite, mixed andesite-sandstone, dolomite, sandstone, silty shale, and mudstone, which are set as category 0, category 1, category 2, category 3, category 4, and category 5, respectively.

[0083] S5-2. Use the GS grid search method to process the hyperparameters of the XGBoost model and obtain the corresponding hyperparameter combination scheme;

[0084] The specific process of GS grid search method is:

[0085] Determine the hyperparameters to be optimized for the XGBoost model; set the value range of the hyperparameters to be optimized, and arrange and combine the values ​​of the hyperparameters to be optimized to generate a hyperparameter grid; initialize the XGBoost model and set the initial hyperparameters; use the GridSearchCV method to perform a grid search on the parameter grid to obtain the corresponding hyperparameter combination solution.

[0086] S5-3. Randomly select a hyperparameter combination and apply it to the XGBoost model to obtain the corresponding initial XGBoost model;

[0087] S5-4, input the oversampled training MWD data into the initial XGBoost model and process it using ten-fold cross validation to obtain ten initial indicator data values, namely MacroF1 values;

[0088] Step S5-4 includes the following steps:

[0089] S5-4-1, divide the oversampled training MWD data into ten subsets;

[0090] S5-4-2. Randomly select a subset as the validation subset, and the remaining nine subsets as the training subsets;

[0091] S5-4-3, input the training subset and its labels into the initial XGBoost model for training, and obtain the trained initial XGBoost model;

[0092] S5-4-4, input the verification subset into the trained initial XGBoost model to obtain the corresponding training recognition result;

[0093] S5-4-5. Calculate the corresponding initial indicator data value based on the training recognition result;

[0094] S5-4-6. Repeat steps S5-4-2 to S5-4-5 until ten initial indicator data values ​​are obtained.

[0095] S5-5. Calculate the index mean of the ten initial index data values;

[0096] S5-6, repeating steps S5-3 to S5-6 until the mean values ​​of the indicators corresponding to all hyperparameter combinations are obtained;

[0097] S5-7. Select the best hyperparameters based on the mean of each indicator;

[0098] S5-8. Adjust the parameters of the initial training XGBoost model based on the optimal hyperparameters to obtain the trained and optimized GS-ADASYN-XGBoost model.

[0099] In practical applications, the values ​​of multidimensional unbalanced data (MWD data) corresponding to different lithologies are different, so the MWD data are input into the XGBoost model, and the MWD data are paired with the MWD data values ​​corresponding to each lithology to obtain the lithology classification results.

[0100] In one embodiment of the present invention, a drilling while drilling data experiment is carried out on a tunnel excavation surface at a certain location, and the MWD excavation data obtained includes 7216-dimensional data groups, each of which includes 7 feature data and 1 label value. Figure 2 As shown in the figure, most of the tunnel drilling depths are concentrated between 0 and 5 meters, and few exceed 4 meters. When recording other characteristic data, the depth value is fixed, resulting in a higher frequency of the same value and a more uniform distribution. As the drilling depth increases, the resistance of the acquisition equipment to the borehole increases, so the frequency gradually decreases. The frequency distribution of the propulsion speed, impact pressure, propulsion pressure, rotary pressure and water flow is roughly normal.

[0101] As shown in Table 1, the number of lithologies corresponding to this dataset of MWD tunneling data shows that the total amount of data is 7216, including 2573 andesites, 2024 mixed andesite-sandstone, 1425 dolomites, 702 sandstones, 333 silty shales, and 159 blue mudstones. Among them, andesite accounts for the highest proportion, and blue mudstone accounts for the lowest proportion, with a difference of 16 times. There are 4597 andesite and mixed andesite-sandstone, accounting for about 63.7% of the total dataset, while the other four types of rock samples account for less than 40%, and category 5 accounts for only 2.2%. The data distribution is extremely unbalanced, and it is obviously unreasonable to directly perform classification tasks.

[0102] Table 1

[0103] Lithology category Sample size Proportion / % andesite Category 0 2573 35.7 sandstone Category 1 702 9.7 Silty shale Category 2 333 4.6 Dolomites Category 3 1425 19.7 mixed andesitic-sandstone Category 4 2024 28 Green mudstone Category 5 159 2.2

[0104] For this dataset of MWD tunneling data, there are 6 types of rock categories. As shown in Table 1, only andesite accounts for more than 30% and can be used as the majority class, so 5 types of lithology data are used as minority classes, i.e., i = {1, 2, 3, 4, 5}. When the oversampling ratio is 1.0, it means that the ratio of minority class samples to majority class samples is 1:1. As the ratio increases, the original samples of minority class samples only account for a small part, and most of them will be generated false samples. At this time, further learning the law between data will be meaningless. However, if the ratio is too low, the number of minority class samples will not change much, and the effect will be minimal. Therefore, the oversampling ratio is taken in the interval of [0.001, 1.0], as shown in Table 2. Considering that oversampling is performed first and then the training set and test set are divided, false data will be generated in the test set. The model will achieve the classification purpose by learning the oversampling law, which will affect the effect of the oversampling ratio. Therefore, the training set and the test set are divided first, and then only the training set is oversampled.

[0105] Table 2

[0106]

[0107] The SVM classifier, RF random forest and XGBoost models were selected for comparison, the RS random grid search algorithm was selected for comparison with the GS grid search algorithm, and the SMOTE oversampling method was selected for comparison with the ADASYN oversampling algorithm.

[0108] The SVM classifier, RF random forest, XGBoost model, SMOTE oversampling method, and ADASYN oversampling algorithm are combined to obtain Models 1 to 6, namely SMOTE-SVM, ADASYN-SVM, SMOTE-XGBoost, ADASYN-XGBoost, SMOTE-RF, and ADASYN-RF.

[0109] As shown in Table 3, the MacroF1 scores of models 1 to 6 under different oversampling ratios; as shown in Table 4, the MacroF1 scores of the models classified without oversampling. It can be seen that the classification effect of the model can be effectively improved by using MWD data after oversampling.

[0110] Table 3

[0111]

[0112]

[0113] Table 4

[0114] Classifier SVM XGBoost RF <![CDATA[MacroF1]]> 0.60 0.79 0.88

[0115] like Figure 3As shown in the figure, the performance improvement effect of the SVM classifier is the most obvious, and it generally shows a trend of increasing first and then decreasing. The overall MacroF1 of SMOTE-SVM is higher than that of ADASYN-SVM. For the XGBoost classifier, as the oversampling ratio changes, the MacroF1 of SMOTE-XGBoost and ADASYN-XGBoost fluctuates. When oversampling is not used, the MacroF1 value of the RF classifier is 0.88, which performs well on this data set, and oversampling does not improve its performance much. In summary, oversampling the data can improve the classification performance to a certain extent, but the improvement effect depends on the specific classifier. In addition, the MacroF1 of each classifier is analyzed to achieve the optimal oversampling ratio. When the oversampling ratio of SMOTE-SVM is 0.03, 0.05 and 0.1, the MacroF1 reaches 0.68, and the optimal ratio is selected as 0.05. When the ratio is 0.5 and 0.6, the MacroF1 score of the SMOTE-XGBoost classifier is the highest, and the optimal oversampling ratio is selected as 0.5. The optimal oversampling ratios of ADASYN-SVM, ADASYN-XGBoost, SMOTE-RF and ADASYN-RF are 0.05, 0.2, 0.03 and 0.4 respectively.

[0116] SVM classifier, RF random forest, XGBoost model and MOTE oversampling method, ADASYN oversampling algorithm and RS random grid search algorithm, GS grid search algorithm are combined to obtain RS-SMOTE-XGBoost model, GS-SMOTE-XGBoost model, RS-ADASYN-XGBoost model, GS-ADASYN-XGBoost model, RS-SMOTE-SVM classifier, GS-SMOTE-SVM classifier, RS-ADASYN-SVM classifier, GS-ADASYN-SVM classifier, RS-SMOTE-RF random forest, GS-SMOTE-RF random forest, RS-ADASYN-RF random forest, GS-ADASYN-RF random forest. The present invention is a GS-ADASYN-XGBoost model. Wherein, for RS random grid search algorithm, the number of random samplings is set to 200. To reduce time cost, each model is consistent in the hyperparameter optimization interval in grid search and random search. Ten-fold cross validation is used in the search process to ensure the accuracy and credibility of the result.

[0117] Table 5

[0118]

[0119] As shown in Table 5, the default hyperparameters and optimization rules of each classifier are set. For the SVM classifier, the optimal hyperparameter combination of SMOTE and ADASYN oversampling is the same on this dataset. However, for the other two meta-classifiers, different oversampling methods produce different optimal hyperparameter combinations. The potential reason is that different classifiers have different sensitivities to the data, which affects the selection of hyperparameters.

[0120] like Figure 4 As shown in the figure, the SVM classifier performs the worst among the three, but hyperparameter optimization has the most obvious effect on it. Among the different classifiers, the degree of performance improvement of GS and RS varies significantly. Specifically, the random search of the SVM classifier is better than the grid search, while the opposite is true for the XGBoost classifier. The effects of the two search methods of the RF classifier are almost the same. In addition, the MacroF1 score of the RF classifier with SMOTE oversampling decreases after hyperparameter optimization, which may be due to the randomness and calculation errors in model training.

[0121] As shown in Figures 5(a), 5(b) and 5(c), twelve models are used to classify the oversampled MWD excavation data set, and the AUC value curves of three of the models are shown. The inflection points of each curve are close to the upper left corner, indicating that these models have high prediction accuracy for the NATM tunnel MWD data set. The AUC values ​​of each lithology category are all above 0.95, further verifying the reliability of this method.

[0122] As shown in Table 6, the F1 score and MacroF1 of all models on different rock categories are calculated. It can be seen that the overall classification performance of the GS-ADASYN-XGBoost model is the best, with a MacroF1 value of 0.91 and an F1 score of 0.77 for the few-class sample data, which is 2.03 times higher than that of the GS-SMOTE-SVM classifier and better than the other 11 models, further proving that the classification result of the GS-ADASYN-XGBoost model adopted in the present invention is the best.

[0123] Table 6

[0124]

[0125]

[0126] Based on the above content, the GS-ADASYN-XGBoost data-driven model obtained after ADASYN oversampling and grid search GS optimization is used to identify the lithology of the test set. Finally, the lithology corresponding to the test MWD data can be obtained. The recognition performance is evaluated by the MacroF1 value, which is as high as 91% in this optimization model.

[0127] In summary, the present invention adopts the ADASYN oversampling method to process the MWD data of while drilling measurement, which solves the problem of multidimensional data imbalance of the MWD data of while drilling measurement; and uses the grid search algorithm to optimize the hyperparameters of the lithology identification model, which further improves the generalization and recognition ability of the identification model; selects the optimized GS-ADASYN-XGBoost to classify the MWD data of while drilling measurement, which can effectively improve the accuracy of the identification results.

Claims

1. A lithology data driven identification method based on New Austrian Tunneling Method (NATM) tunnel excavation while drilling data, characterized by: The following steps are involved: S1, real-time collection of MWD data of the excavation process of the new Austrian tunnel to be identified, i.e., multi-dimensional unbalanced data; S2. Preprocess the multidimensional unbalanced data to obtain preprocessed multidimensional unbalanced data; S3, using ADASYN oversampling algorithm to process the preprocessed multidimensional unbalanced data to obtain oversampled MWD data; S4, constructing a lithology identification model; S5. Input the oversampled MWD data into the lithology identification model for processing to obtain lithology identification results.

2. The lithology data driven identification method based on New Austrian Tunneling Method (NATM) tunnel excavation while drilling data according to claim 1 is characterized by: The multi-dimensional imbalance data in step S1 includes drilling depth, driving speed, impact pressure, driving pressure, rotation pressure, water pressure and water flow.

3. The lithology data driven identification method based on New Austrian Tunneling Method (NATM) tunnel excavation while drilling data according to claim 1 is characterized by: The step S3 comprises the following steps: S3-1, taking the preprocessed multi-dimensional unbalanced data as sample data, and counting the data volume of each type of sample data; S3-2, determine whether the amount of each type of sample data is less than the quantity threshold; if so, determine that type of sample data is a minority type sample; otherwise, determine that type of sample data is a majority type sample; S3-3, using the K nearest neighbor algorithm to determine the k nearest neighbor sample data corresponding to each minority class sample in the minority class data space composed of the minority class samples; S3-4, according to the formula: Get the density ri of each minority class sample; where ni represents the i-th nearest neighbor sample data, and k represents the total number of nearest neighbor sample data; S3-5, according to the formula: Get the number of synthetic samples Gi that need to be generated; where Nmaj represents the total amount of majority class sample data, Nmin represents the total amount of minority class sample data, β represents the proportion of generating new samples, and ∑(·) represents the summation function; S3-6, based on the number of synthetic samples Gi that need to be generated, generate corresponding synthetic samples for each minority class sample by random linear interpolation; S3-7, merge each minority class sample and the corresponding synthetic sample respectively; Among them, the oversampled MWD data includes majority category samples and merged minority category samples.

4. The lithology data driven identification method based on New Austrian Tunneling Method (NATM) tunnel excavation while drilling data according to claim 2 or 3, characterized in that: The lithology identification model in step S4 adopts the XGBoost model.

5. The lithology data driven identification method based on New Austrian Tunneling Method (NATM) tunnel excavation while drilling data according to claim 4 is characterized by: The training process of the lithology identification model includes the following steps: S5-1, obtaining training multidimensional imbalanced data and its labels, and oversampling using the same method as step S3 to obtain oversampled training MWD data; S5-2. Use the GS grid search method to process the hyperparameters of the XGBoost model and obtain the corresponding hyperparameter combination scheme; S5-3. Randomly select a hyperparameter combination and apply it to the XGBoost model to obtain the corresponding initial XGBoost model; S5-4, input the oversampled training MWD data into the initial XGBoost model and process it using ten-fold cross validation to obtain ten initial indicator data values, namely MacroF1 values; S5-5. Calculate the index mean of the ten initial index data values; S5-6, repeating steps S5-3 to S5-6 until the mean values ​​of the indicators corresponding to all hyperparameter combinations are obtained; S5-7. Select the best hyperparameters based on the mean of each indicator; S5-8. Adjust the parameters of the initial training XGBoost model based on the optimal hyperparameters to obtain the trained and optimized GS-ADASYN-XGBoost model.

6. The lithology data driven identification method based on New Austrian Tunneling Method (NATM) tunnel excavation while drilling data according to claim 5 is characterized by: The specific process of the GS grid search method is: Determine the hyperparameters to be optimized for the XGBoost model; set the value range of the hyperparameters to be optimized, and arrange and combine the values ​​of the hyperparameters to be optimized to generate a hyperparameter grid; initialize the XGBoost model and set the initial hyperparameters; use the GridSearchCV method to perform a grid search on the parameter grid to obtain the corresponding hyperparameter combination solution.

7. The lithology data driven identification method based on New Austrian Tunneling Method (NATM) tunnel excavation while drilling data according to claim 5 is characterized by: The step S5-4 comprises the following steps: S5-4-1, divide the oversampled training MWD data into ten subsets; S5-4-2. Randomly select a subset as the validation subset, and the remaining nine subsets as the training subsets; S5-4-3, input the training subset and its labels into the initial XGBoost model for training, and obtain the trained initial XGBoost model; S5-4-4, input the verification subset into the trained initial XGBoost model to obtain the corresponding training recognition result; S5-4-5. Calculate the corresponding initial indicator data value based on the training recognition result; S5-4-6. Repeat steps S5-4-2 to S5-4-5 until ten initial indicator data values ​​are obtained.

8. The lithology data driven identification method based on New Austrian Tunneling Method (NATM) tunnel excavation while drilling data according to claim 5 is characterized by: The labels in step S5-1 are andesite, mixed andesite-sandstone, dolomite, sandstone, silty shale, and mudstone, which are set to category 0, category 1, category 2, category 3, category 4, and category 5 respectively.

Images