A TBM tunneling surrounding rock perception method, device and storage medium based on feature alignment coupled TrAdaBoost

Abstract

The application discloses a TBM tunneling surrounding rock perception method and device based on feature alignment coupled TrAdaBoost and a storage medium. The method standardizes and pretreats TBM construction data of a source domain and a target domain, filters key general parameters as input features and unifies surrounding rock classification standards; a surrounding rock classification prediction pre-training model is established based on a Stacking integrated learning technology; when the target domain has no known label data, feature alignment is used to realize feature level domain adaptation, and an unsupervised transfer adaptation model is obtained; when the target domain has a small amount of known label data, a TrAdaBoost algorithm is used to adjust sample weights to realize instance level domain adaptation, and a supervised transfer adaptation model is obtained; input features of the target domain are input into the transfer adaptation model, and a surrounding rock classification prediction result is output. The application can effectively reduce feature domain deviation between the source domain and the target domain, improve pre-training model cross-project surrounding rock classification prediction accuracy, and provide a basis for TBM tunneling parameter optimization and support decision.

Description

Technical Field

[0001] This invention belongs to the field of intelligent construction technology of full-face tunnel boring machines (TBMs), specifically involving a TBM tunneling surrounding rock perception method based on feature alignment coupling TrAdaBoost, and also involving electronic devices and computer-readable storage media for implementing the method. Background Technology

[0002] Tunnel Boring Machines (TBMs) are widely used in the construction of long-distance tunnels in fields such as water conservancy, hydropower, transportation, and mining. During TBM excavation, timely and accurate perception of the surrounding rock conditions (rock classification) near the tunnel face is a crucial prerequisite for optimizing and adjusting excavation parameters, making support scheme decisions, and ensuring construction safety. Because the tunnel face is obscured by the cutterhead during TBM excavation, it is difficult to directly observe the surrounding rock conditions. Therefore, using the real-time excavation parameter data (rock-machine interaction data) collected by the TBM's onboard sensing system to invert and perceive the surrounding rock conditions has become an important technical means in the field of intelligent TBM construction.

[0003] Currently, most rock perception algorithms / models in the field of TBM intelligent construction are built based on historical data from single engineering application scenarios. On the one hand, due to the differences in engineering geology and construction equipment in different regions, the complexity of underground engineering problems, and the requirement for independent and identically distributed new samples when expanding the application of machine learning models, the application effect of pre-trained models based on specific engineering data in new engineering scenarios is often poor, and the model's generalization ability is insufficient. On the other hand, the data collected in the early stages of new engineering construction is limited and has poor representativeness, making it difficult to support the training of high-precision prediction models. my country has already constructed a large number of TBM tunnel projects, and their historical data contains rich empirical knowledge. Therefore, for algorithm models pre-trained based on historical data (source domain), how to effectively apply the pre-trained models in the source domain to new projects (target domain) through specific strategies is a research bottleneck that urgently needs to be overcome to improve the cross-engineering application performance of intelligent algorithms / models.

[0004] Existing research on TBM (Tunnel Boring Machine) rock sensing mainly focuses on using complete data from a single project to build intelligent models for training and testing. Research on cross-project applications of these models is still limited. Typical studies can be summarized into two categories: First, by introducing two additional parameters—the cutterhead diameter and the number of cutters—as new input features to the model, data from multiple projects in the source domain are integrated for model learning. This can improve the prediction performance of directly using a pre-trained model from the source domain for the target domain. Second, features reflecting the essential rock-machine interaction mechanism (such as the torque penetration index) are used. The fitting slope a and intercept b of thrust and penetration, single-blade torque, and single-blade thrust are used as physical shared features to replace the basic features collected by the equipment, which to some extent eliminates the inter-domain input feature distribution offset caused by cutterhead diameter, number of cutters, tunneling strategy, etc.

[0005] Overall, research on transfer learning for TBM tunneling rock perception and prediction models is still in its early exploratory stages. The two aforementioned approaches have not significantly improved the model's effectiveness in cross-engineering applications, and neither has developed a systematic and effective transfer learning strategy for the two typical working conditions: no known labeled data in the early stages of new engineering construction (unsupervised case) and a limited amount of known labeled data (supervised case). Therefore, proposing a universal transfer learning method to promote the generalization performance and widespread application of TBM tunneling rock perception models pre-trained with rich historical data in new engineering scenarios has strong practical engineering significance. Summary of the Invention

[0006] The purpose of this invention is to overcome the above-mentioned shortcomings of the prior art and provide a TBM tunneling surrounding rock perception method based on feature alignment coupling TrAdaBoost, so as to effectively reduce the feature domain offset between the source domain (existing engineering data) and the target domain (new engineering data), and solve the problem that the TBM tunneling surrounding rock perception intelligent model based on source domain data pre-trained has poor generalization ability and prediction performance in the target domain application scenario.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A TBM tunneling surrounding rock sensing method based on feature alignment coupled with TrAdaBoost includes the following steps:

[0009] Step S1: Obtain TBM construction data from the source and target domains. The source domain consists of TBM construction data from existing projects, and the target domain consists of TBM construction data from new projects. Perform data standardization preprocessing on the TBM construction data from the source and target domains to obtain effective rock breaking time series data. Calculate the statistical characteristic values ​​of each parameter in the stable rock breaking stage as representative values, and select key general parameters from the source and target domains as model input features. Unify the surrounding rock classification standards of the source and target domains, and use the unified surrounding rock classification as the model output feature.

[0010] Step S2: Normalize the input features of the model, encode the features of the surrounding rock classification label data, and establish source domain data sample set and target domain data sample set for TBM tunneling surrounding rock perception respectively.

[0011] Step S3: Select the base classifier and meta classifier, establish a TBM tunneling surrounding rock perception and prediction model based on Stacking ensemble learning technology, and use the source domain data sample set to train the model and optimize the hyperparameters to obtain a pre-trained model for surrounding rock classification and prediction.

[0012] Step S4: Perform transfer learning based on whether there is known surrounding rock classification label data in the target domain, and output the surrounding rock classification prediction result for the target domain:

[0013] If there is no known rock classification label data in the target domain, feature alignment is used to process the model input features of the source domain to achieve feature-level domain adaptation between the source and target domains. The hyperparameters of the rock classification prediction pre-training model are fixed, and the rock classification prediction pre-training model is iteratively trained using the source domain data after feature alignment to obtain an unsupervised transfer adaptation model. The model input features of the target domain are input into the unsupervised transfer adaptation model, and the rock classification prediction result of the target domain is output.

[0014] If the target domain has a small amount of known rock classification label data, feature alignment is used to process the model input features of the source domain to achieve feature-level domain adaptation between the source and target domains. The hyperparameters of the pre-trained rock classification prediction model are fixed, and the pre-trained model is iteratively trained using the feature-aligned source domain data to obtain an unsupervised transfer adaptation model. Then, the unsupervised transfer adaptation model is iteratively trained using the known label data of the target domain and the TrAdaBoost algorithm. During training, the weights of source and target domain samples are iteratively adjusted to increase the weight of source domain samples with distributions similar to the target domain and decrease the weight of source domain samples with distributions unrelated to the target domain, achieving instance-level domain adaptation between the source and target domains to obtain a supervised transfer adaptation model. The model input features of samples in the target domain other than the known label data are input into the supervised transfer adaptation model, and the rock classification prediction result for the target domain is output.

[0015] Preferably, in step S1, the data standardization preprocessing operation includes sequentially performing tunneling state discrimination, data change point identification, and outlier removal; the statistical feature value is the mean of each parameter in the stable rock breaking stage; the model input feature includes the propulsion speed. Cutter head speed Total propulsion Cutter head torque Torque penetration index and on-site penetration index There are a total of 6 key general parameters; among them, penetration per revolution Torque penetration index On-site penetration index .

[0016] Preferably, in step S2, a robust normalization method is used to normalize the model input features, and the calculation formula is as follows:

[0017]

[0018] in, Input feature parameters to the model before normalization. This is the median of the input feature parameters. This is the difference between the upper and lower quartiles of the input feature parameter. These are the robustly normalized input feature parameters;

[0019] The unique thermal coding method is used to encode the surrounding rock classification label data into binary vectors.

[0020] Preferably, in step S3, the base classifier includes support vector machine, extreme gradient boosting tree, random forest, and gradient boosting decision tree, and the meta-classifier is gradient boosting decision tree; the source domain data sample set is divided into training set and test set according to a preset ratio, and the following is adopted: The TBM tunneling surrounding rock perception and prediction model was trained using cross-validation. The value is an integer greater than or equal to 3, and during the training process, the hyperparameters of the base classifier and the meta classifier are optimized using a grid search method or a metaheuristic algorithm.

[0021] Preferably, the feature alignment uses an invariant relation derived from the parameters of the TBM tunneling equipment to transform the six key general parameters from the source domain space to the target domain space. The calculation formula is as follows:

[0022]

[0023] in, , , , , , These are the propulsion speed in the source space, the cutterhead rotation speed, the total propulsion force, the cutterhead torque, the field penetration index, and the torque penetration index. , , , , , These are the propulsion speed, cutterhead rotation speed, total propulsion force, cutterhead torque, field penetration index, and torque penetration index when converted to the target domain space. The number of TBM tools in the source domain engineering. The number of TBM tools for the target domain project. For the TBM cutterhead diameter of the source domain engineering, The diameter of the TBM cutterhead for the target domain project.

[0024] Preferably, the feature alignment uses the CORAL correlation alignment method to transform the model input features in the source domain. The calculation formula is as follows:

[0025]

[0026]

[0027] in, The source domain characteristic matrix, The source domain feature matrix after CORAL correlation alignment transformation. Let be the transformation matrix. Let covariance be the feature of the source domain. Let covariance be the feature matrix of the target domain. This is the inverse matrix of the square root of the source domain covariance matrix, used for whitening features in the source domain. The square root of the target domain covariance matrix is ​​used to color the whitened source domain features to assign them the correlation and variance of the target domain features.

[0028] Preferably, the source domain includes TBM construction data from several different existing projects For integers greater than 1; in step S3, respectively using Training with source domain data sample sets from existing projects A pre-trained model for surrounding rock classification and prediction; the corresponding result obtained in step S4. An unsupervised transfer adaptation model or A supervised transfer adaptation model is used, employing majority voting or weighted voting methods to evaluate the transfer performance. The surrounding rock classification prediction results output by each migration adaptation model are fused to determine the final surrounding rock classification prediction result for the target domain; when using the weighted voting method, the first... Weights of the prediction results of each transfer adaptation model The calculation formula is as follows:

[0029]

[0030] in, For the first The F1 score of the migration adaptation model for source domain data on its own validation set. For the first The F1 score of the migration adaptation model for source domain data on its own validation set. and The values ​​are all from 1 to Integers.

[0031] Preferably, when iteratively training the unsupervised transfer adaptation model using known label data from the target domain and the TrAdaBoost algorithm, the first... In each iteration, the current classifier is calculated. Weighted error rate on known labeled samples in the target domain :

[0032]

[0033] in, The number of labeled samples is known for the target domain. For the target domain Input features of a known labeled sample, Classify and label its actual surrounding rock. For the first In the first iteration of the target domain The weights of a known labeled sample. For indicator functions, when The value is 1 if the condition is met, otherwise it is 0.

[0034] Update the target domain sample weight adjustment factor according to the following formula. Source domain sample weight decay factor :

[0035]

[0036] in, The number of source domain samples used in training. This represents the total number of iterations of the TrAdaBoost algorithm.

[0037] For source domain samples misclassified by the current classifier, multiply their weights by . Reduce the weight; for target domain samples misclassified by the current classifier, multiply their weights by . The output of the supervised transfer adaptation model is determined by the classifier ensemble obtained from the second half of the iterations after the iterations are completed.

[0038] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the TBM tunneling surrounding rock sensing methods based on feature alignment coupling TrAdaBoost.

[0039] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the TBM tunneling surrounding rock sensing methods based on feature alignment coupling TrAdaBoost.

[0040] The beneficial effects of this invention are:

[0041] (1) For cases where there are no known labeled data samples in the target domain, this invention proposes an unsupervised transfer learning method for TBM tunneling surrounding rock perception based on feature alignment: By using the CORAL correlation alignment method or the invariant relation derived from the parameters of the TBM tunneling equipment, feature alignment processing is performed on the data of different engineering source domains to achieve feature-level domain adaptation and reduce the domain offset between each source domain and the target domain; then, the source domain data after feature alignment is used to further iterate and train each pre-trained model to obtain unsupervised transfer adaptation models under different source domains, thereby improving the cross-domain generalization ability of the pre-trained models. In addition, when there are multiple source domains, the prediction results of different unsupervised transfer adaptation models are fused by majority voting or weighted voting to obtain more accurate target domain surrounding rock classification prediction results, reducing the randomness of prediction results using only a single unsupervised transfer adaptation model.

[0042] (2) For cases where there are a small number of known labeled data samples in the target domain, this invention proposes a supervised transfer learning method for TBM tunneling surrounding rock perception based on feature alignment coupled with TrAdaBoost: Firstly, feature alignment is performed on the source domain data of different engineering projects using the CORAL correlation alignment method or invariant relations derived from TBM tunneling equipment parameters, achieving feature-level domain adaptation, reducing the domain offset between the source and target domains, and further iteratively training each pre-trained model using the feature-aligned source domain data to obtain unsupervised transfer adaptation models under different source domains; Secondly, using a small number of known labeled data in the target domain and the TrAdaBoost algorithm, further iterative training and sample weight adjustment are performed on each unsupervised transfer adaptation model to achieve instance-level domain adaptation, obtaining supervised transfer adaptation models under different source domains, thus improving the cross-domain generalization ability of the pre-model from both feature and instance levels. Furthermore, by using majority voting (or weighted voting) on ​​the prediction results of different supervised transfer adaptation models, accurate target domain surrounding rock classification prediction results are obtained, reducing the randomness of prediction results obtained by using only a single supervised transfer adaptation model.

[0043] The method proposed in this invention can be effectively applied to new engineering application scenarios, enabling cross-domain high-precision surrounding rock classification and prediction of pre-trained models, thereby providing guidance for TBM tunneling parameter optimization and support decision-making, and ensuring safe and efficient TBM construction. Attached Figure Description

[0044] Figure 1 This is the overall flowchart of the present invention;

[0045] Figure 2 This invention provides a standardized preprocessing procedure for TBM construction data.

[0046] Figure 3 This is a violin diagram of the six key general features of the four TBM tunnel projects of this invention;

[0047] Figure 4 This is a schematic diagram illustrating the principle of the surrounding rock classification and prediction model based on Stacking ensemble learning of the present invention.

[0048] Figure 5 This is the row-normalized confusion matrix of the source domain pre-trained model of the present invention, which directly predicts engineering results across the target domain;

[0049] Figure 6 This is the row-normalized confusion matrix of the prediction results of the source domain and target domain pre-trained models of this invention on their own validation set;

[0050] Figure 7 This is the row-normalized confusion matrix of the cross-engineering prediction results of the unsupervised transfer adaptation model based on deduced invariant relations for feature-level domain adaptation in this invention.

[0051] Figure 8 The present invention provides a row-normalized confusion matrix for cross-engineering prediction results of a supervised transfer adaptation model based on CORAL feature alignment coupled with TrAdaBoost for feature-level and instance-level domain adaptation. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] The first aspect of this invention provides a method for sensing surrounding rock in TBM tunneling based on feature alignment coupling TrAdaBoost, which includes the following steps:

[0054] Step S1: Obtain TBM construction data from the source and target domains. The source domain consists of TBM construction data from existing projects, and the target domain consists of TBM construction data from new projects. Perform data standardization preprocessing on the TBM construction data from the source and target domains to obtain effective rock-breaking time-series data. Calculate the statistical characteristic values ​​of each parameter in the stable rock-breaking stage as representative values, and select key common parameters from the source and target domains as model input features. Unify the surrounding rock classification standards for the source and target domains, and use the unified surrounding rock classification as the model output feature. Further, in step S1, the data standardization preprocessing includes sequentially performing tunneling state discrimination, data change point identification, and outlier removal; the statistical characteristic value is the mean value of each parameter in the stable rock-breaking stage; the model input feature includes the advance speed. Cutter head speed Total propulsion Cutter head torque Torque penetration index and on-site penetration index There are a total of 6 key general parameters; among them, penetration per revolution Torque penetration index On-site penetration index Furthermore, in step S1, the classification standard for the surrounding rock of both the source domain and the target domain uniformly adopts the engineering geological classification method for surrounding rock in the "Code for Geological Investigation of Water Conservancy and Hydropower Projects" (GB50487-2008), and the surrounding rock conditions are divided into five categories from good to bad: I, II, III, IV, and V, which are used as the output features of the model.

[0055] Step S2 involves normalizing the model input features, performing feature encoding on the surrounding rock classification label data, and establishing source domain data sample sets and target domain data sample sets for TBM tunneling surrounding rock sensing. Further, in step S2, a robust normalization method is used to normalize the model input features, and a one-hot encoding method is used to encode the surrounding rock classification label data into binary vectors. The robust normalization calculation formula is as follows:

[0056]

[0057] in, Input feature parameters to the model before normalization. This is the median of the input feature parameters. This is the difference between the upper and lower quartiles of the input feature parameter (i.e., the interquartile range). These are the robustly normalized input feature parameters.

[0058] Step S3: Select a base classifier and a meta-classifier. Establish a TBM (Tunnel Boring Machine) surrounding rock perception and prediction model based on Stacking ensemble learning technology. Train and optimize the model using the source domain data sample set to obtain a pre-trained model for surrounding rock classification and prediction. Further, in step S3, the base classifier includes support vector machines, extreme gradient boosting trees, random forests, and gradient boosting decision trees, and the meta-classifier is a gradient boosting decision tree. Divide the source domain data sample set into a training set and a test set according to a preset ratio, and use... ( The model is trained using cross-validation, and the hyperparameters of the base classifier and meta-classifier are optimized during the training process using grid search or meta-heuristic algorithms.

[0059] Step S4: Perform transfer learning based on whether there is known surrounding rock classification label data in the target domain, and output the surrounding rock classification prediction result for the target domain, including both unsupervised and supervised transfer learning:

[0060] Scenario 1 (Unsupervised Transfer Learning, Step S4-1): If there is no known rock classification label data in the target domain, unsupervised transfer learning is carried out based on the feature-level domain adaptation algorithm. Specifically, it includes: Step S4-1-1, processing the model input features of the source domain using feature alignment to achieve feature-level domain adaptation between the source and target domains; Step S4-1-2, fixing the hyperparameters of the rock classification prediction pre-trained model in Step S3, and using the source domain data after feature alignment in Step S4-1-1 to further iteratively train the pre-trained model to obtain an unsupervised transfer adaptation model with higher prediction accuracy; Step S4-1-3, inputting the model input features of the target domain into the unsupervised transfer adaptation model obtained in Step S4-1-2, and outputting the rock classification prediction results of the target domain in real time.

[0061] Scenario 2 (Supervised Transfer Learning, Step S4-2): If the target domain has a small amount of known rock classification label data, supervised transfer learning is carried out based on the feature-level + instance-level domain adaptation algorithm. Specifically, this includes: Step S4-2-1, processing the model input features of the source domain using feature alignment to achieve feature-level domain adaptation between the source and target domains; Step S4-2-2, fixing the hyperparameters of the rock classification prediction pre-trained model from Step S3, and using the source domain data processed by feature alignment in Step S4-2-1 to further iteratively train the pre-trained model, obtaining an unsupervised transfer adaptation model with higher prediction accuracy; Step S4-2-3, using a small amount of known label data from the target domain and the TrAdaBoost algorithm to perform supervised transfer learning on the unsupervised transfer adaptation model obtained in Step S4-2-2. The supervised transfer adaptation model undergoes further iterative training. During training, the sample weights of the source and target domains are iteratively adjusted: the weights of source domain samples similar in distribution to the target domain are increased, while the weights of source domain samples unrelated to the target domain distribution are decreased. This achieves instance-level domain adaptation between the source and target domains, resulting in a supervised transfer adaptation model with further improved prediction accuracy. In step S4-2-4, the model input features of samples in the target domain other than those with known labels are input into the supervised transfer adaptation model obtained in step S4-2-3, and the surrounding rock classification prediction results for the target domain are output in real time. Further, the feature alignment uses the CORAL correlation alignment method, or uses invariant relations derived from TBM tunneling equipment parameters, to transform (align) the model input features of the source domain from the source domain space to the target domain space. Further, when using invariant relations derived from TBM tunneling equipment parameters for feature alignment, the calculation formulas for transforming the six key general parameters from the source domain space to the target domain space are as follows:

[0062]

[0063] in, , , , , , These are the propulsion speed in the source space, the cutterhead rotation speed, the total propulsion force, the cutterhead torque, the field penetration index, and the torque penetration index. , , , , , These are the propulsion speed, cutterhead rotation speed, total propulsion force, cutterhead torque, field penetration index, and torque penetration index when converted to the target domain space. The number of TBM tools in the source domain engineering. The number of TBM tools for the target domain project. For the TBM cutterhead diameter of the source domain engineering, Let be the diameter of the TBM cutterhead in the target domain. Furthermore, when using the CORAL correlation alignment method for feature alignment, the calculation formula for transforming the model input features of the source domain is as follows:

[0064]

[0065]

[0066] in, The source domain characteristic matrix, The source domain feature matrix after CORAL correlation alignment transformation. Let be the transformation matrix. Let covariance be the feature of the source domain. Let be the covariance matrix of the target domain features; It is the inverse matrix of the square root of the source domain covariance matrix, used to perform whitening operations on the source domain features; The square root of the target domain covariance matrix is ​​used to color the whitened source domain features, assigning them the relevance and variance of the target domain features. Further, when iteratively training the unsupervised transfer learning model using known label data of the target domain and the TrAdaBoost algorithm, the... In each iteration, the current classifier is calculated. Weighted error rate on known labeled samples in the target domain Based on this, the target domain sample weight adjustment factor is calculated. Source domain sample weight decay factor ,in The number of source domain samples used in training. This represents the total number of iterations; for source domain samples misclassified by the current classifier, their weights are multiplied by [the number of iterations]. To reduce the weight, for target domain samples misclassified by the current classifier, multiply their weights by [the appropriate factor]. The size is increased; after iteration, the output of the supervised transfer adaptation model is determined by the ensemble of classifiers obtained from the latter half of the iterations. Further, the source domain includes... ( TBM construction data from several different existing projects; when In step S3, respectively, the following methods are used: Training with source domain data sample sets from existing projects A pre-trained model for surrounding rock classification and prediction is obtained accordingly in step S4. An unsupervised transfer adaptation model or A supervised transfer adaptation model is used, employing majority voting or weighted voting methods to evaluate the transfer performance. The surrounding rock classification prediction results output by each migration adaptation model are fused to determine the final surrounding rock classification prediction result for the target domain; when using the weighted voting method, the first... Weights of the prediction results of each transfer adaptation model The calculation formula is as follows:

[0067]

[0068] in, For the first The F1 score of the migration adaptation model for source domain data on its own validation set. For the first The F1 score of the migration adaptation model for source domain data on its own validation set. and The values ​​are all from 1 to Integers.

[0069] A second aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the TBM tunneling surrounding rock sensing method based on feature alignment coupling TrAdaBoost as described in the first aspect of the present invention.

[0070] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the TBM tunneling surrounding rock sensing method based on feature alignment coupling TrAdaBoost as described in the first aspect of the present invention.

[0071] like Figure 1 As shown, this embodiment of the invention provides a TBM tunneling surrounding rock sensing method based on feature alignment coupling TrAdaBoost, including steps S1 to S4, which are described in detail below.

[0072] Step S1 involves performing data standardization preprocessing on the source and target domains to obtain effective rock-breaking time-series data. Statistical characteristic values ​​of each parameter in the stable rock-breaking stage are calculated as representative values, and key common parameters in both the source and target domains are selected as input features for the model. The surrounding rock classification standards in the source and target domains are unified, and the surrounding rock classification is used as the model's output feature. Step S1 specifically includes steps S1-1 to S1-3.

[0073] Step S1-1: The TBM construction data of the source domain and the target domain are sampled in a single tunneling cycle (i.e., one tunneling process of the TBM from start to stop). Each tunneling cycle includes useless data and effective rock breaking data. Data standardization preprocessing is performed on the source domain and the target domain. For the raw data, the following processes are used in sequence: tunneling status (whether tunneling is underway) discrimination, data change point (inflection point) identification, and outlier removal to obtain effective rock breaking data.

[0074] Step S1-2: Calculate the mean values ​​of each parameter in the stable rock-breaking stage as representative values, and select six key general parameters from the source and target domains as model input features, including: propulsion velocity. Cutter head speed Total propulsion Cutter head torque Torque penetration index and on-site penetration index .

[0075] Of the six key general parameters mentioned above, propulsion speed is... Cutter head speed Total propulsion Cutter head torque Torque penetration index can be directly acquired by the sensing system on the TBM. and on-site penetration index The composite parameters reflecting the rock-machine interaction mechanism are calculated from directly acquired parameters. The penetration depth per revolution of the TBM is defined. for:

[0076]

[0077] in, Per revolution penetration, that is, the distance the TBM advances along the tunnel axis per revolution of the cutterhead. To accelerate, This represents the rotational speed of the cutter head.

[0078] Torque penetration index and on-site penetration index They are defined as follows:

[0079]

[0080] in, The torque penetration index represents the cutterhead torque required per unit penetration depth. The penetration index represents the total thrust required per unit penetration depth. The torque of the cutter head. As the overall driving force, For each revolution penetration degree. and It reflects the excavability of the surrounding rock, is closely related to the surrounding rock conditions, and is a key feature characterizing the nature of rock-machine interaction.

[0081] Steps S1-3: The surrounding rock classification standards for both the source and target domains are uniformly adopted using the engineering geological classification method for surrounding rock in the "Code for Geological Investigation of Water Conservancy and Hydropower Projects" (GB 50487-2008). The surrounding rock conditions are divided into five categories—I, II, III, IV, and V—from best to worst, and these categories are used as output features of the model. It should be noted that when the original surrounding rock classification standards used in the source and target domains are inconsistent, they can be converted and unified to the same classification standard based on the grading criteria of each classification standard (such as rock strength, rock mass integrity, structural surface condition, groundwater condition, etc.) to eliminate differences in the output label space.

[0082] Step S2 involves normalizing the model input features from step S1, performing feature encoding on the surrounding rock classification label data, and establishing source domain data sample sets and target domain data sample sets for TBM tunneling surrounding rock sensing. Step S2 specifically includes steps S2-1 and S2-2.

[0083] Step S2-1: The six input features selected—feed speed, cutterhead rotation speed, total feed force, cutterhead torque, torque penetration index, and field penetration index—are normalized using a robust normalization method. The calculation formula is as follows:

[0084]

[0085] in, These are the input feature parameters before normalization. This is the median of the input feature parameters. This is the difference between the upper and lower quartiles of the input feature parameter (i.e., the interquartile range). These are the input feature parameters after robust normalization. Compared to standardization methods based on mean and standard deviation, robust normalization scales based on median and interquartile range, making it more robust to outliers in TBM construction data that are difficult to completely remove.

[0086] The label data of surrounding rock classification I to V are processed into 5-dimensional binary vectors using the one-hot encoding method. That is, surrounding rock classification I, II, III, IV and V correspond to (1, 0, 0, 0, 0), (0, 1, 0, 0, 0), (0, 0, 1, 0, 0), (0, 0, 0, 1, 0) and (0, 0, 0, 0, 1) respectively.

[0087] Step S2-2: Establish source domain data sample sets and target domain data sample sets for TBM tunneling surrounding rock sensing, respectively. The source domain data sample set includes... ( The target domain data sample set contains known surrounding rock classification label data of 1 existing project; the target domain data sample set contains all unknown label data of 1 new project (corresponding to unsupervised transfer learning), or a large amount of unknown label data and a small amount of known label data of 1 new project (corresponding to supervised transfer learning).

[0088] Step S3 involves selecting a base classifier and a meta-classifier, establishing a TBM tunneling surrounding rock perception and prediction model based on Stacking ensemble learning technology, and training and optimizing the model using the source domain data sample set obtained in step S2 to obtain a high-precision surrounding rock classification and prediction pre-trained model. Step S3 specifically includes steps S3-1 and S3-2.

[0089] Step S3-1: Establish a rock classification prediction model based on Stacking ensemble learning technology. The Stacking ensemble learning model consists of two layers: the first layer is the base classifier layer, composed of multiple heterogeneous base classifiers in parallel. Each base classifier learns from the input features and outputs its own category prediction probability. The second layer is the meta-classifier layer, which uses the outputs of each base classifier as input features for secondary learning, and outputs the final rock classification prediction result after fusing the prediction information of each base classifier. In this embodiment, the selected base classifiers include Support Vector Machine (SVM), Extreme Gradient Boosting Tree (XGBoost), Random Forest (RF), and Gradient Boosting Decision Tree (GBDT), and the selected meta-classifier is Gradient Boosting Decision Tree (GBDT). By using Stacking ensemble learning to complement the advantages of multiple heterogeneous classifiers, higher prediction accuracy and stronger stability can be obtained than with a single classifier.

[0090] Step S3-2: Randomly sample according to a preset ratio (e.g., 9:1) ( The source domain data sample sets of each of the different projects were divided into training and test sets, and the following methods were adopted: ( The aforementioned surrounding rock classification and prediction model was trained using cross-validation. During training, grid search or meta-heuristic algorithms (such as genetic algorithms, particle swarm optimization, etc.) were employed to optimize the hyperparameters of the base classifier and meta-classifier in the model. Through this process, a model targeting... Data from different engineering source domains Find the optimal combination of hyperparameters, and the corresponding A high-precision pre-trained model for classifying and predicting surrounding rock.

[0091] Step S4 establishes the transfer learning framework and algorithm implementation for TBM tunneling surrounding rock perception, including both unsupervised and supervised transfer learning. Feature alignment is the core operation for achieving feature-level domain adaptation. This invention provides two flexibly selectable feature alignment methods: one is an invariant relationship derived from TBM tunneling equipment parameters, and the other is the CORAL correlation alignment method. The two feature alignment methods are explained first, followed by the specific implementation processes of unsupervised and supervised transfer learning.

[0092] (I) Invariant Relationships Derived Based on TBM Tunneling Equipment Parameters

[0093] The differences in equipment parameters such as cutterhead diameter and number of cutters (rollers) used in different projects lead to systematic distribution shifts in collected parameters, including feed speed, cutterhead rotation speed, total feed force, and cutterhead torque, as well as their composite parameters, between the source and target domains. This invention, starting from the rock-machine interaction mechanism, derives the characteristic transformation invariance relationship from the source domain space to the target domain space based on the following physical assumptions:

[0094] (a) Under the same surrounding rock conditions, the net tunneling speed of a TBM is mainly determined by the excavability of the surrounding rock and is independent of the equipment size, i.e., the advance speed. (a) To keep the cutting speed and wear level of the tool constant between the source and target domains; (b) To make the tool cutting speed and wear level comparable, the edge speed of tool discs with different diameters (and) (Proportional) to remain consistent, where (c) The diameter of the cutter head; the average thrust borne by a single hob ( This reflects the essence of the interaction between the cutterhead and the rock mass, and remains unchanged under the same surrounding rock conditions. (d) Number of cutting tools; (d) Turret torque It is approximately equal to the sum of the products of the rolling force of each hob and its installation radius, assuming the hobs are approximately uniformly arranged radially along the cutter head. and Proportional, of which The average rolling force of a single cutter remains constant under the same surrounding rock conditions.

[0095] Based on the above assumptions (a) and (b), we can obtain , Based on hypothesis (c), we can obtain Based on assumption (d), we can obtain ; and then from each penetration degree achievable , and then combine and Based on the definition, the complete calculation formula for transforming the six input features from the source domain space to the target domain space can be derived as follows:

[0096]

[0097] in, It is the speed of propagation of the source domain space. It is the propulsion speed for transitioning to the target domain space. It is the rotational speed of the cutter head in the source domain space. It is the tool shroud rotation speed converted to the target domain space. It is the total propulsion force of the source domain space. It is the total propulsion force for the transformation to the target domain space. It is the cutter head torque in the source domain space. It is the cutter head torque converted to the target domain space. It is the field penetration index of the source domain space. It is the field penetration index transformed into the target domain space. It is the torque penetration index of the source domain space. It is the torque penetration index transformed to the target domain space. This refers to the number of TBM tools used in source domain engineering. It is the number of TBM tools in the target domain project. It is the diameter of the TBM cutterhead in the source domain engineering. It is the diameter of the TBM cutterhead in the target domain project.

[0098] The aforementioned invariant relation can complete the feature space transformation from the source domain to the target domain using only two easily obtainable equipment parameters: the diameter of the cutter head and the number of cutters. It does not require any labeled data in the target domain, has clear physical meaning, and is easy to calculate. It is particularly suitable for scenarios where data is scarce at the beginning of a new project.

[0099] (ii) CORAL-related alignment methods

[0100] The CORAL (CORrelation Alignment) method reduces the distributional differences between the source and target domains by aligning the second-order statistics (covariance) of source and target domain features. The formula for calculating CORAL correlation alignment of source domain features is as follows:

[0101]

[0102]

[0103] in, The source domain characteristic matrix, The source domain feature matrix after CORAL correlation alignment transformation. Let be the transformation matrix. Let covariance be the feature of the source domain. Let be the covariance matrix of the target domain features; It is the inverse matrix of the square root of the source domain covariance matrix. Its function is to perform whitening operation on the source domain features, converting the data into an isotropic distribution. It is the square root of the covariance matrix of the target domain. Its function is to perform a coloring operation on the whitened source domain features, assigning them the correlation and variance of the target domain features.

[0104] To ensure the invertibility of the covariance matrix and the stability of numerical calculations, regularization can be applied to the covariance matrix in practical calculations:

[0105]

[0106] in, This represents the operation of calculating the covariance matrix. The feature matrix of the target domain, It is the identity matrix. The regularization coefficient can be taken as... The CORAL-related alignment method aligns the feature space based on the statistical characteristics of the data, without requiring labeled data in the target domain. It is suitable for scenarios where a certain amount of unlabeled data has been accumulated in the target domain.

[0107] (III) Unsupervised transfer learning (Step S4-1, Case 1)

[0108] If there is no known surrounding rock classification label data in the target domain, unsupervised transfer learning is carried out based on the feature-level domain adaptation algorithm, specifically including steps S4-1-1 to S4-1-3.

[0109] Step S4-1-1: Using the CORAL correlation alignment method or the invariant relation derived from the TBM tunneling equipment parameters, respectively... Alignment processing is performed on the source domain features of different projects to achieve feature-level domain adaptation, resulting in feature-aligned features. Source domain data.

[0110] Step S4-1-2: Fix the condition in step S3 The hyperparameters of the base classifier and meta classifier in the pre-trained model for surrounding rock classification and prediction are used, after feature alignment processing in step S4-1-1. Data from different engineering source domains, using ( Cross-validation method Further iterative training is performed on each pre-trained model to obtain... An unsupervised transfer adaptation model with higher prediction accuracy.

[0111] Step S4-1-3: Input the input features of the target domain into the data obtained in step S4-1-2. In an unsupervised transfer adaptation model, the target domain is output in real time. The results of the surrounding rock classification prediction. When When the output prediction result is the final surrounding rock classification prediction result for the target domain, then... At that time, the final surrounding rock classification prediction result of the target domain is determined by majority voting or weighted voting.

[0112] When using weighted voting, the first Weights of the prediction results of each unsupervised transfer adaptation model The calculation formula is as follows:

[0113]

[0114] in, For the first The weights of the prediction results of an unsupervised transfer adaptation model. For the first The F1 score of an unsupervised transfer adaptation model on its own validation set for source domain data. For the first The F1 score of an unsupervised transfer adaptation model on its own validation set for source domain data. ( () represents the number of source domains for different projects.

[0115] (iv) Supervised transfer learning (step S4-2, case two)

[0116] If the target domain has a small amount of known surrounding rock classification label data, then supervised transfer learning is carried out based on the feature-level + instance-level domain adaptation algorithm, specifically including steps S4-2-1 to S4-2-4.

[0117] Step S4-2-1: Using the CORAL correlation alignment method or the invariant relation derived from the TBM tunneling equipment parameters, respectively... Alignment processing is performed on the source domain features of different projects to achieve feature-level domain adaptation, resulting in feature-aligned features. Source domain data.

[0118] Step S4-2-2: Fix the condition in step S3 The hyperparameters of the base classifier and meta classifier in the pre-trained model for surrounding rock classification and prediction are used, after feature alignment processing in step S4-2-1. Data from different engineering source domains, using ( Cross-validation method Further iterative training is performed on each pre-trained model to obtain... An unsupervised transfer adaptation model with higher prediction accuracy.

[0119] Step S4-2-3: Utilize a small amount of known label data from the target domain and the TrAdaBoost algorithm to refine the results obtained in Step S4-2-2. Each unsupervised transfer adaptation model undergoes further iterative training. During training, the sample weights of the source and target domains are iteratively adjusted (i.e., increasing the weights of source domain samples with distributions similar to the target domain and decreasing the weights of source domain samples with distributions unrelated to the target domain), achieving instance-level domain adaptation between the source and target domains. A supervised transfer adaptation model with further improved prediction accuracy.

[0120] A specific implementation of the TrAdaBoost algorithm for iterative training and sample weight adjustment is as follows. The source domain samples with known labels (number of samples) after feature alignment are... ) and a small number of known labeled samples in the target domain (number of samples) The samples are merged into a training set, and the weights of each sample are initialized; in the first... ( In each iteration, the classifier is trained using the current sample weights. and calculate Weighted error rate on known labeled samples in the target domain :

[0121]

[0122] in, The number of labeled samples is known for the target domain. For the target domain Input features of a known labeled sample, This is the classification label for the actual surrounding rock of the sample. For classifier The predicted surrounding rock classification results for this sample, For the first In the first iteration of the target domain The weights of a known labeled sample. This is an indicator function that takes the value 1 when the condition within the parentheses is true (i.e., the prediction is wrong), and 0 otherwise.

[0123] Calculate the target domain sample weight adjustment factor Source domain sample weight decay factor :

[0124]

[0125] in, For the first Classifier in rounds of iteration Weighted error rate on known labeled samples in the target domain The number of labeled samples in the source domain used for training is known. This represents the total number of iterations of the TrAdaBoost algorithm.

[0126] Then update the sample weights according to the following formula:

[0127]

[0128] in, and The first Wheel and the first In the first iteration The weights of each sample, Given a set of known labeled samples from the source domain. The target domain is a set of known labeled samples. From the above formula, it can be seen that for a source domain sample that is misclassified, its weight is multiplied by... ( The weight of a source domain sample is reduced by multiplying its weight by the target domain weight, thus assuming the source domain sample is unrelated to the target domain distribution and gradually reducing its impact on training. For misclassified target domain samples, their weights are multiplied by the weight of the source domain sample. (when hour This increases the learning capacity of the model for difficult-to-distinguish samples in the target domain. After the iteration is completed, the second half of the iterations (the second half) will be used to further improve the model's learning capacity for difficult-to-distinguish samples in the target domain. Turn number The classifier ensemble obtained from the round determines the output of the supervised transfer adaptation model, thereby achieving instance-level domain adaptation between the source and target domains.

[0129] Step S4-2-4: Input the input features of samples in the target domain other than the known label data into the data obtained in step S4-2-3. In a supervised transfer adaptation model, the target domain is output in real time. The results of the surrounding rock classification prediction. When When the output prediction result is the final surrounding rock classification prediction result for the target domain, then... At that time, the final surrounding rock classification prediction result of the target domain is determined by majority voting or weighted voting.

[0130] When using weighted voting, the first Weights of prediction results from supervised transfer adaptation models The calculation formula is as follows:

[0131]

[0132] in, For the first The weights of the prediction results of a supervised transfer adaptation model. For the first The F1 score of a supervised transfer adaptation model on its own validation set for each source domain data. For the first The F1 score of a supervised transfer adaptation model on its own validation set for each source domain data. ( () represents the number of source domains for different projects.

[0133] The target domain surrounding rock classification prediction results output by step S4 above can provide guidance for the optimization and adjustment of TBM tunneling parameters and support scheme decisions for new projects in the target domain, ensuring safe and efficient TBM construction.

[0134] Specific examples

[0135] The following section uses four TBM tunnel projects in my country—the Yinchuo-Jiliao Project, the Yinsong Project, the Dianzhong Water Diversion Project, and Project A—as engineering backgrounds to provide a detailed explanation of the implementation process and effects of the method of this invention.

[0136] (1) Execution process of step S1

[0137] Step S1-1: Collect TBM construction data for four tunnel projects: Yinchuo-Jiliao, Yinsong, Dianzhong Water Diversion, and Project A. Yinchuo-Jiliao is the source domain, and Yinsong, Dianzhong Water Diversion, and Project A are the three target domains. TBM equipment information for the source and target domains is shown in Table 1. A single tunneling cycle from start-up to shutdown of the TBM forms a regular tunneling loop. TBM construction data for both the source and target domains are sampled using a single tunneling cycle. Data analysis shows that each tunneling cycle can be divided into five working stages: start-up, empty-push, ascending, stabilizing, and shutdown. The ascending and stabilizing stages are the effective rock-breaking stages.

[0138] Table 1. TBM equipment information for source and target domain projects.

[0139] parameter Yinchuo Jiliao Pine Planting Project Central Yunnan Water Diversion Project Project A Cutter head diameter (m) 5.20 7.93 9.83 4.75 Number of hobbing cutters 34 56 65 28 Number of data items collected 401 199 381 200 Maximum propulsion displacement (m) 1.8 1.8 2.0 1.8 Rated cutterhead thrust (kN) 11340 23260 31526 13608 Rated cutter head torque (kN·m) 3340 8410 15719 2870 Maximum propulsion speed (mm / min) 120 120 120 120 Maximum cutter head speed (rev / min) 11.45 7.6 7.4 11.3 TBM type Open Open Open Open

[0140] The TBM construction data standardization preprocessing operation involves performing data standardization preprocessing on both the source and target domains. The process is as follows: Figure 2 As shown, firstly, a binary state discrimination function is constructed to determine the tunneling state (whether tunneling is underway) based on the raw data, thereby identifying the data in the tunneling state. Next, a data change point (inflection point) identification algorithm is used to obtain two change points between the data of the unsupported section, the rising section, and the stable section, thus obtaining the effective rock-breaking data for the rising and stable sections. Finally, a box plot method is used to process the stable section data, removing outlier data that exceeds the upper and lower bounds.

[0141] Step S1-2: Calculate the mean values ​​of each parameter in the stable section (stable rock-breaking stage) as representative values. The TBM used in the source domain Yinchuo-Jiliao Project has 401 collected parameters, while the TBMs used in the target domains Yinsong Project, Dianzhong Water Diversion Project, and Project A have 199, 381, and 200 collected parameters, respectively. By analyzing the parameters of the source domain and the three target domains, six key common parameters shared by all domains are selected as model input features, including: propulsion speed. Cutter head speed Total propulsion Cutter head torque Torque penetration index and on-site penetration index A violin diagram of six key common features of four TBM tunnel projects is shown below. Figure 3 As shown, due to differences in engineering geological conditions and equipment parameters, there is a significant distribution shift of various features among different projects.

[0142] Step S1-3: The classification standard for surrounding rocks in both the source and target domains is uniformly adopted using the engineering geological classification method for surrounding rocks in the "Code for Geological Investigation of Water Conservancy and Hydropower Projects" (GB 50487-2008). The surrounding rock conditions are divided into five categories from best to worst: I, II, III, IV, and V, which are used as the output features of the model.

[0143] (2) Execution process of step S2

[0144] Step S2-1: Normalize the 6 selected input features using the aforementioned robust normalization method; process the surrounding rock classification I to V label data into 5-dimensional binary vectors using one-hot encoding.

[0145] Step S2-2: Establish source domain data sample sets and target domain data sample sets for TBM tunneling surrounding rock perception, each containing 6 input features and 1 output variable. The number of tunneling cycle samples for the source and target domain projects is shown in Table 2. The source domain data sample set includes known surrounding rock classification label data from one existing project, the Yinchuo-Jiliao Water Diversion Project (i.e.,...). To more effectively illustrate the versatility of the method proposed in this invention, the embodiments used three target domain sample sets: the Yinsong Project, the Dianzhong Water Diversion Project, and Project A. Each target domain sample set was configured in two ways: ① all unknown surrounding rock classification label data; ② a large amount of unknown surrounding rock classification label data and a small amount of known surrounding rock classification label data.

[0146] Table 2 Number of tunneling cycle samples for source and target domain projects

[0147] Classification of surrounding rock Yinchuo Jiliao Pine Planting Project Central Yunnan Water Diversion Project Project A Ⅰ 0 0 0 0 Ⅱ 1120 389 0 0 Ⅲ 3498 4998 1392 1471 Ⅳ 1011 1286 627 434 Ⅴ 189 111 2901 90 total 5818 7538 4920 1995

[0148] (3) Execution process of step S3

[0149] Step S3-1: Establish a rock classification and prediction model based on Stacking ensemble learning technology. The selected base classifiers include Support Vector Machine, Extreme Gradient Boosting Tree, Random Forest, and Gradient Boosting Decision Tree, and the selected meta-classifier is Gradient Boosting Decision Tree. A schematic diagram of the principle of the rock classification and prediction model based on Stacking ensemble learning is shown below. Figure 4 As shown.

[0150] Step S3-2: The source domain data sample set is randomly divided into a training set and a test set at a ratio of 9:1, with 5236 and 582 samples in the training and test sets, respectively. The model training process employs 10-fold cross-validation for performance evaluation and parameter optimization. A genetic algorithm is used to globally optimize the key hyperparameters of the base classifier and meta-classifier in the Stacking framework, while other non-key hyperparameters are set to the default values ​​from the Scikit-learn library. Through the above process, the optimal hyperparameter combination for the source domain data and the corresponding high-precision surrounding rock classification and prediction pre-trained model are obtained.

[0151] The source domain pre-trained model was directly used to predict engineering data in three target domains to evaluate its cross-engineering prediction performance. The row-normalized confusion matrix of the cross-engineering prediction results of the source domain pre-trained model in the target domain is as follows: Figure 5 As shown in Table 3, the statistical index values ​​of the prediction results of the source domain pre-trained model directly across engineering applications are shown in Table 3.

[0152] Table 3. Statistical values ​​of prediction results of source domain pre-trained models directly across engineering applications.

[0153] Target domain engineering Surrounding rock category Accuracy Precision Recall F1 score Number of samples Pine Planting Project Ⅱ — 0.045 0.167 0.070 30 Pine Planting Project Ⅲ — 0.788 0.820 0.804 557 Pine Planting Project Ⅳ — 0.306 0.124 0.177 153 Pine Planting Project Ⅴ — 0 0 0 14 Pine Planting Project overall 63.8% 0.285 0.278 0.263 754 Central Yunnan Water Diversion Project Ⅱ — 0 0 0 0 Central Yunnan Water Diversion Project Ⅲ — 0.158 0.279 0.202 147 Central Yunnan Water Diversion Project Ⅳ — 0 0 0 54 Central Yunnan Water Diversion Project Ⅴ — 1.000 0.010 0.020 291 Central Yunnan Water Diversion Project overall 8.9% 0.290 0.072 0.056 492 Project A Ⅱ — 0 0 0 0 Project A Ⅲ — 0.805 0.232 0.361 142 Project A Ⅳ — 0.227 0.458 0.303 48 Project A Ⅴ — 0.300 0.600 0.400 10 Project A overall 30.5% 0.333 0.323 0.266 200

[0154] Depend on Figure 5 As shown in Table 3, when the source domain pre-trained model is directly applied across engineering projects, the overall prediction accuracy in the three target domains is only 63.8%, 8.9%, and 30.5%, respectively. The prediction performance is significantly degraded, indicating that there is a clear domain shift between the source domain and the target domain. Directly applying the source domain pre-trained model is difficult to meet engineering requirements.

[0155] Furthermore, to better analyze the cross-engineering prediction performance of the target domains, based on the same process described above, the optimal hyperparameter combinations for the three target domain data and the corresponding three high-precision surrounding rock classification and prediction pre-trained models were obtained. The row-normalized confusion matrices of the prediction results on their own validation sets for the source and target domain pre-trained models are shown below. Figure 6 As shown, it can achieve high prediction accuracy on its own validation set, further confirming that the performance degradation when applied across engineering projects is due to domain shift rather than the model itself.

[0156] (4) Execution process of step S4

[0157] Step S4-1 (Case 1, Unsupervised Transfer Learning): Assuming that there is no known surrounding rock classification label data for any of the three target domains, unsupervised transfer learning is carried out based on the feature-level domain adaptation algorithm.

[0158] Step S4-1-1: For each of the three different target domains, based on the invariant relationship derived from the TBM tunneling equipment parameters (the cutterhead diameter and the number of cutters are taken from Table 1), the six input features of the source domain of the Yinchuo-Jiliao Project are aligned to achieve feature-level domain adaptation, resulting in three sets of source domain data after feature alignment for the three target domains.

[0159] Step S4-1-2: Fix the hyperparameters of the base classifier and meta classifier in the pre-trained model for surrounding rock classification and prediction in step S3. Using the three sets of source domain data after feature alignment processing in step S4-1-1, further iterative training of the pre-trained model is carried out using 10-fold cross-validation to obtain three unsupervised transfer adaptation models with higher prediction accuracy for the three target domains.

[0160] Step S4-1-3: Input the input features of the three target domains into the three corresponding unsupervised transfer adaptation models obtained in step S4-1-2, and output the surrounding rock classification prediction results of the three target domains in real time. The row-normalized confusion matrix of the cross-engineering prediction results of the unsupervised transfer adaptation model based on the derived invariant relation for feature-level domain adaptation is as follows. Figure 7 As shown in Table 4, the statistical index values ​​of the cross-project prediction results of the unsupervised transfer adaptation model are presented.

[0161] Table 4 Statistical values ​​of cross-project prediction results of unsupervised transfer adaptation model

[0162] Target domain engineering Surrounding rock category Accuracy Precision Recall F1 score Number of samples Pine Planting Project Ⅱ — 0.089 0.133 0.107 30 Pine Planting Project Ⅲ — 0.802 0.837 0.819 557 Pine Planting Project Ⅳ — 0.496 0.412 0.450 153 Pine Planting Project Ⅴ — 1.000 0.071 0.133 14 Pine Planting Project overall 70.8% 0.597 0.363 0.377 754 Central Yunnan Water Diversion Project Ⅱ — 0 0 0 0 Central Yunnan Water Diversion Project Ⅲ — 0.787 0.476 0.593 147 Central Yunnan Water Diversion Project Ⅳ — 0.184 0.167 0.175 54 Central Yunnan Water Diversion Project Ⅴ — 0.804 0.450 0.577 291 Central Yunnan Water Diversion Project overall 42.7% 0.443 0.273 0.336 492 Project A Ⅱ — 0 0 0 0 Project A Ⅲ — 0.887 0.662 0.758 142 Project A Ⅳ — 0.439 0.521 0.476 48 Project A Ⅴ — 0.467 0.700 0.560 10 Project A overall 63.0% 0.448 0.471 0.449 200

[0163] Depend on Figure 7 As shown in Table 4, after the above unsupervised transfer learning, the overall prediction accuracy of the model for surrounding rock classification in the three target domains increased from 63.8%, 8.9%, and 30.5% to 70.8%, 42.7%, and 63.0%, respectively, all of which were improved to a certain extent, verifying the effectiveness of feature-level domain adaptation based on the derived invariant relations.

[0164] Step S4-2 (Case 2, Supervised Transfer Learning): Assuming that each of the three target domains has a small amount of known surrounding rock classification label data, supervised transfer learning is carried out based on the feature-level + instance-level domain adaptation algorithm.

[0165] Step S4-2-1: For each of the three different target domains, the CORAL correlation alignment method is used to align the six input features of the source domain of the Yinchuo-Jiliao Project to achieve feature-level domain adaptation, resulting in three sets of source domain data after feature alignment for the three target domains.

[0166] Step S4-2-2: Fix the hyperparameters of the base classifier and meta classifier in the pre-trained model for surrounding rock classification prediction in step S3. Using the three sets of source domain data after feature alignment processing in step S4-2-1, further iterative training of the pre-trained model is carried out using 10-fold cross-validation to obtain three unsupervised transfer adaptation models with higher prediction accuracy for the three target domains.

[0167] Step S4-2-3: For each of the three different target domains, using a small amount of known label data for each target domain (in this embodiment, 10% of the engineering samples of each target domain are uniformly used as known label data) and the TrAdaBoost algorithm, further iterative training is carried out on the three unsupervised transfer adaptation models obtained in step S4-2-2. By iteratively adjusting the sample weights of the source domain and the target domain according to the aforementioned weight update formula during the training process, instance-level domain adaptation between the source domain and the target domain is achieved, resulting in three supervised transfer adaptation models with further improved prediction accuracy for the three target domains.

[0168] Step S4-2-4: Input the input features of samples from the three target domains, excluding the known label data, into the three corresponding supervised transfer adaptation models obtained in step S4-2-3, and output the surrounding rock classification prediction results for the three target domains in real time. The row-normalized confusion matrix of the cross-engineering prediction results of the supervised transfer adaptation model based on CORAL feature alignment coupled with TrAdaBoost for feature-level + instance-level domain adaptation is shown below. Figure 8 As shown in Table 5, the statistical index values ​​of the cross-engineering prediction results of the supervised transfer adaptation model are presented.

[0169] Table 5 Statistical values ​​of cross-project prediction results of supervised transfer adaptation model

[0170] Target domain engineering Surrounding rock category Accuracy Precision Recall F1 score Number of samples Pine Planting Project Ⅱ — 0.125 0.233 0.163 30 Pine Planting Project Ⅲ — 0.854 0.785 0.818 557 Pine Planting Project Ⅳ — 0.530 0.641 0.580 153 Pine Planting Project Ⅴ — 1.000 0.071 0.133 14 Pine Planting Project overall 72.0% 0.627 0.432 0.423 754 Central Yunnan Water Diversion Project Ⅲ — 0.954 0.707 0.812 147 Central Yunnan Water Diversion Project Ⅳ — 0.442 0.352 0.392 54 Central Yunnan Water Diversion Project Ⅴ — 0.826 0.966 0.891 291 Central Yunnan Water Diversion Project overall 82.1% 0.741 0.675 0.698 492 Project A Ⅱ — 0 0 0 0 Project A Ⅲ — 0.953 0.866 0.907 142 Project A Ⅳ — 0.769 0.833 0.800 48 Project A Ⅴ — 0.833 1.000 0.909 10 Project A overall 86.5% 0.639 0.675 0.654 200

[0171] Depend on Figure 8 As shown in Table 5, after the above-mentioned supervised transfer learning, the overall prediction accuracy of the model for surrounding rock classification in the three target domains was further improved to 72.0%, 82.1%, and 86.5%, respectively. This is a significant improvement compared to direct cross-engineering applications (63.8%, 8.9%, and 30.5%) and unsupervised transfer learning (70.8%, 42.7%, and 63.0%). Moreover, the prediction performance improvement of supervised transfer learning is better than that of unsupervised transfer learning, which verifies the effectiveness and superiority of the feature-level + instance-level dual-layer domain adaptation strategy.

[0172] This invention also provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the various steps of the TBM tunneling surrounding rock perception method based on feature alignment coupling TrAdaBoost. The processor can be a central processing unit (CPU), a graphics processing unit (GPU), or other general-purpose or special-purpose processor with data processing capabilities. The memory can include non-transitory computer-readable storage media such as read-only memory, random access memory, disk storage, and flash memory. The electronic device can be deployed at the TBM construction site as an industrial control computer, edge computing device, or remote server, receiving real-time tunneling parameter data collected by the TBM sensing system and outputting surrounding rock classification and prediction results.

[0173] This invention also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the various steps of the aforementioned TBM tunneling surrounding rock sensing method based on feature alignment coupling TrAdaBoost. The computer-readable storage medium includes, but is not limited to, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0174] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A TBM tunneling surrounding rock sensing method based on feature alignment coupling TrAdaBoost, characterized in that, Includes the following steps: Step S1: Obtain TBM construction data from the source domain and the target domain. The source domain is the TBM construction data of an existing project, and the target domain is the TBM construction data of a new project. Perform data standardization preprocessing on the TBM construction data from the source domain and the target domain to obtain effective rock breaking time series data. Statistical characteristic values ​​of each parameter in the stable rock breaking stage are calculated as representative values, and key general parameters in the source and target domains are selected as model input features. The classification standards for surrounding rocks in the source and target domains are unified, and the unified surrounding rock classification is used as the output feature of the model. Step S2: Normalize the input features of the model, encode the features of the surrounding rock classification label data, and establish source domain data sample set and target domain data sample set for TBM tunneling surrounding rock perception respectively. Step S3: Select the base classifier and meta classifier, establish a TBM tunneling surrounding rock perception and prediction model based on Stacking ensemble learning technology, and use the source domain data sample set to train the model and optimize the hyperparameters to obtain a pre-trained model for surrounding rock classification and prediction. Step S4: Perform transfer learning based on whether there is known surrounding rock classification label data in the target domain, and output the surrounding rock classification prediction result for the target domain: If there is no known rock classification label data in the target domain, feature alignment is used to process the model input features of the source domain to achieve feature-level domain adaptation between the source and target domains. The hyperparameters of the rock classification prediction pre-training model are fixed, and the rock classification prediction pre-training model is iteratively trained using the source domain data after feature alignment to obtain an unsupervised transfer adaptation model. The model input features of the target domain are input into the unsupervised transfer adaptation model, and the rock classification prediction result of the target domain is output. If the target domain has a small amount of known rock classification label data, feature alignment is used to process the model input features of the source domain to achieve feature-level domain adaptation between the source and target domains. The hyperparameters of the pre-trained rock classification prediction model are fixed, and the pre-trained model is iteratively trained using the feature-aligned source domain data to obtain an unsupervised transfer adaptation model. Then, the unsupervised transfer adaptation model is iteratively trained using the known label data of the target domain and the TrAdaBoost algorithm. During training, the weights of source and target domain samples are iteratively adjusted to increase the weight of source domain samples with distributions similar to the target domain and decrease the weight of source domain samples with distributions unrelated to the target domain, achieving instance-level domain adaptation between the source and target domains to obtain a supervised transfer adaptation model. The model input features of samples in the target domain other than the known label data are input into the supervised transfer adaptation model, and the rock classification prediction result for the target domain is output.

2. The TBM tunneling surrounding rock sensing method based on feature alignment coupling TrAdaBoost according to claim 1, characterized in that, In step S1, the data standardization preprocessing operation includes sequentially performing tunneling state discrimination, data change point identification, and outlier removal; the statistical feature value is the mean of each parameter in the stable rock breaking stage; the model input feature includes the propulsion speed. Cutter head speed Total propulsion Cutter head torque Torque penetration index and on-site penetration index There are a total of 6 key general parameters; among them, penetration per revolution Torque penetration index On-site penetration index .

3. The TBM tunneling surrounding rock sensing method based on feature alignment coupling TrAdaBoost according to claim 2, characterized in that, In step S2, a robust normalization method is used to normalize the model input features. The calculation formula is as follows: ； in, Input feature parameters to the model before normalization. This is the median of the input feature parameters. This is the difference between the upper and lower quartiles of the input feature parameter. These are the robustly normalized input feature parameters; The unique thermal coding method is used to encode the surrounding rock classification label data into binary vectors.

4. The TBM tunneling surrounding rock sensing method based on feature alignment coupling TrAdaBoost according to claim 1, characterized in that, In step S3, the base classifier includes support vector machine, extreme gradient boosting tree, random forest, and gradient boosting decision tree, and the meta-classifier is gradient boosting decision tree; the source domain data sample set is divided into training set and test set according to a preset ratio, and then... The TBM tunneling surrounding rock perception and prediction model was trained using cross-validation. The value is an integer greater than or equal to 3, and during the training process, the hyperparameters of the base classifier and the meta classifier are optimized using a grid search method or a metaheuristic algorithm.

5. The TBM tunneling surrounding rock sensing method based on feature alignment coupling TrAdaBoost according to claim 2, characterized in that, The feature alignment employs an invariant relation derived from TBM tunneling equipment parameters to transform the six key general parameters from the source domain space to the target domain space. The calculation formula is as follows: ； in, , , , , , These are the propulsion speed in the source space, the cutterhead rotation speed, the total propulsion force, the cutterhead torque, the field penetration index, and the torque penetration index. , , , , , These are the propulsion speed, cutterhead rotation speed, total propulsion force, cutterhead torque, field penetration index, and torque penetration index when converted to the target domain space. The number of TBM tools in the source domain engineering. The number of TBM tools for the target domain project. For the TBM cutterhead diameter of the source domain engineering, The diameter of the TBM cutterhead for the target domain project.

6. The TBM tunneling surrounding rock sensing method based on feature alignment coupling TrAdaBoost according to claim 1, characterized in that, The feature alignment uses the CORAL correlation alignment method to transform the model input features in the source domain. The calculation formula is as follows: ； ； in, The source domain characteristic matrix, The source domain feature matrix after CORAL correlation alignment transformation. The transformation matrix is... Let covariance be the feature of the source domain. Let covariance be the feature matrix of the target domain. This is the inverse matrix of the square root of the source domain covariance matrix, used for whitening features in the source domain. The square root of the target domain covariance matrix is ​​used to color the whitened source domain features to assign them the correlation and variance of the target domain features.

7. The TBM tunneling surrounding rock sensing method based on feature alignment coupling TrAdaBoost according to claim 1, characterized in that, The source domain includes TBM construction data from several different existing projects For integers greater than or equal to 1; in step S3, respectively using Training with source domain data sample sets from existing projects A pre-trained model for classifying and predicting surrounding rocks; The corresponding result obtained in step S4 An unsupervised transfer adaptation model or A supervised transfer adaptation model, when When the value equals 1, the surrounding rock classification prediction result output by the transfer adaptation model is the final surrounding rock classification prediction result for the target domain, while when... When the value is greater than 1, the majority voting method or weighted voting method shall be used. The surrounding rock classification prediction results output by each migration adaptation model are fused to determine the final surrounding rock classification prediction result for the target domain; when using the weighted voting method, the first... Weights of the prediction results of each transfer adaptation model The calculation formula is as follows: ； in, For the first The F1 score of the migration adaptation model for source domain data on its own validation set. For the first The F1 score of the migration adaptation model for source domain data on its own validation set. and The values ​​are all from 1 to Integers.

8. The TBM tunneling surrounding rock sensing method based on feature alignment coupling TrAdaBoost according to claim 1, characterized in that, When iteratively training the unsupervised transfer adaptation model using known label data in the target domain and the TrAdaBoost algorithm, the first... In each iteration, the current classifier is calculated. Weighted error rate on known labeled samples in the target domain : ； in, The number of labeled samples is known for the target domain. For the target domain Input features of a known labeled sample, Classify and label its actual surrounding rock. For the first In the first iteration of the target domain The weights of a known labeled sample. For indicator functions, when The value is 1 if the condition is met, otherwise it is 0. Update the target domain sample weight adjustment factor according to the following formula. Source domain sample weight decay factor : ； in, The number of source domain samples used in training. This represents the total number of iterations of the TrAdaBoost algorithm. For source domain samples misclassified by the current classifier, multiply their weights by . Reduce the weight; for target domain samples misclassified by the current classifier, multiply their weights by . The output of the supervised transfer adaptation model is determined by the classifier ensemble obtained from the second half of the iterations after the iterations are completed.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the TBM tunneling surrounding rock sensing method based on feature alignment coupling TrAdaBoost as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the TBM tunneling surrounding rock sensing method based on feature alignment coupling TrAdaBoost as described in any one of claims 1 to 8.

Images