Tunnel boring machine surrounding rock convergence prediction and machine jam risk early warning method based on dynamic weighting fusion of SVR and MLP

Abstract

The application provides a tunnel boring machine surrounding rock convergence prediction and machine jam risk early warning method based on dynamic weighting fusion of SVR and MLP, belongs to the field of tunnel and underground engineering construction informatization and intelligent control technology, and comprises the following steps: data acquisition, data preprocessing, risk threshold setting, model training, double model prediction, weighted fusion, risk early warning and closed loop control. The application overcomes the defects of the existing surrounding rock convergence prediction method in tunnel boring machine construction, such as insufficient real-time performance, poor adaptability, low prediction accuracy, lack of model fusion complementary mechanism and lack of closed loop linkage with construction control.

Description

Technical Field

[0001] This invention relates to the field of information technology and intelligent control technology for tunnel and underground engineering construction, and in particular to a method for predicting the surrounding rock convergence of tunnel boring machines and providing early warning of machine jamming risk based on dynamic weighted fusion of SVR and MLP. Background Technology

[0002] Tunnel Boring Machines (TBMs) are widely used mechanized construction equipment in modern underground engineering. They can perform continuous operations such as tunneling, support, and muck removal under various geological conditions, offering advantages such as high tunneling speed, environmental friendliness, and high overall efficiency. However, when tunneling in weak surrounding rock, high-stress areas, or structurally complex sections, the surrounding rock may undergo significant convergent deformation due to stress release, creep, or structural damage. In severe cases, this can lead to the disappearance of the gap between the shield and the surrounding rock, a sharp increase in frictional resistance, and insufficient thrust of the TBM, resulting in jamming. With the rapid growth of TBM applications in railway, hydropower, transportation, mining, and municipal tunnel projects in China, long-distance, deep-buried, and geologically complex tunnels have emerged, greatly increasing the risk of TBM jamming. Jamming not only leads to construction interruptions and delays but can also cause equipment damage and substantial economic losses.

[0003] Existing methods for predicting surrounding rock convergence mainly include empirical formula methods, analytical theory methods, and numerical simulation methods. Empirical formula methods rely on historical engineering statistics and expert experience for prediction, and their applicability depends on similar geological conditions, making them difficult to adapt to complex and variable geological environments. Analytical theory methods are usually based on elastoplastic mechanical models or convergence-constraint methods to calculate surrounding rock deformation, and their underlying assumptions (such as homogeneity, isotropy, and instantaneous loading) often differ significantly from actual engineering conditions. Although numerical simulation methods have high accuracy, they are complex to model and computationally time-consuming, making it difficult to meet the requirements of real-time prediction and online early warning.

[0004] In recent years, with the improvement of information technology in shield tunnel engineering, the monitoring technology for tunnel boring equipment operation has become increasingly sophisticated. The recorded engineering data contains internal information about the tunnel boring equipment and its interaction with the external strata. Machine learning, due to its strong data analysis capabilities and the fact that it does not require prior theoretical formulas or expert knowledge, has a wider application scope compared to traditional modeling and statistical analysis methods. Artificial intelligence algorithms, especially machine learning models such as Support Vector Machine (SVM) and Artificial Neural Network (ANN), have been proven to be effective predictive tools when the relationship between independent and dependent variables is not clearly understood. Existing research shows that SVM and ANN can be effectively used to predict the surrounding rock convergence of TBM tunnel boring. However, the current application of artificial intelligence in TBM construction still faces the following problems: Disadvantages of existing technology: (1) Existing models are mostly single algorithms and lack a fusion and complementarity mechanism. Although existing studies have used SVM and ANN (including MLP) to predict the convergence of TBM tunnel surrounding rock, most of these studies only compare the predictive performance of the two models and do not organically integrate the two models to give full play to their respective advantages. A single model is prone to prediction fluctuations under specific geological conditions, making it difficult to guarantee the prediction stability during long-term construction.

[0005] (2) Lack of multi-source data fusion and feature optimization methods for surrounding rock convergence prediction. Existing methods do not fully utilize multi-source data such as borehole core tests, geological survey reports, laser section scanning and TBM field monitoring, and fail to effectively utilize the potential correlation of multi-source data, making it difficult to accurately predict the convergence trend of surrounding rock under complex nonlinear relationships.

[0006] (3) The model training and parameter optimization process is not dynamically updated in conjunction with real-time construction data. Most existing models are trained and used offline, and cannot adaptively update model parameters as geological conditions change gradually during construction, resulting in a gradual decrease in the prediction accuracy of the model during long-distance tunneling.

[0007] (4) The prediction results do not form a closed loop with risk warning and construction control. Most existing studies are limited to the prediction level and lack a complete closed-loop linkage mechanism from prediction to risk threshold determination and construction parameter adjustment suggestions. Even if a high-risk trend is predicted, it is impossible to take effective construction intervention measures in time before the risk occurs.

[0008] (5) Fixed risk warning thresholds and lack of dynamic adaptive capability. Existing risk warning methods for machine jamming mostly use fixed thresholds or are based on a single mechanical criterion. They fail to dynamically adjust the risk judgment criteria according to the distribution of historical construction data, changes in geological conditions and safety margins, resulting in a high false alarm rate or missed alarm rate.

[0009] (6) The model deployment lacks consideration for the real-time requirements of the construction site. Most existing methods rely on central servers or the cloud for calculation, resulting in large data transmission delays, which makes it difficult to meet the needs of construction sites for low-latency real-time prediction and early warning. Summary of the Invention

[0010] This invention provides a method for predicting the surrounding rock convergence of tunnel boring machines and for early warning of machine jamming risk based on dynamic weighted fusion of SVR and MLP, in order to overcome the shortcomings of existing methods for predicting the surrounding rock convergence of tunnel boring machines during construction, such as insufficient real-time performance, poor adaptability, low prediction accuracy, lack of model fusion and complementarity mechanism, and lack of closed-loop linkage with construction control.

[0011] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting rock convergence and predicting jamming risk of tunnel boring machines based on dynamic weighted fusion of SVR and MLP includes: The multi-source geological and mechanical parameters of the tunnel boring machine construction section and the measured value of the surrounding rock convergence are obtained. The multi-source geological and mechanical parameters include burial depth, geological strength index, rock mass quality index, uniaxial compressive strength, rock mass cohesion, internal friction angle, elastic modulus, uniaxial compressive strength and uniaxial tensile strength of rock mass. The multi-source geological and mechanical parameters are preprocessed to obtain preprocessed multi-source geological and mechanical parameters. A dynamic risk threshold is set based on the obtained measured values ​​of the surrounding rock convergence. A training dataset is constructed based on the preprocessed multi-source geological and mechanical parameters and the measured value of the surrounding rock convergence, and a support vector machine regression model and a multilayer perceptron neural network model are trained respectively. The multi-source geological and mechanical parameters collected in real time during construction are respectively input into the support vector machine regression model and the multilayer perceptron neural network model after training to obtain the first and second surrounding rock convergence prediction values. The first surrounding rock convergence prediction value and the second surrounding rock convergence prediction value are weighted and fused to obtain the final surrounding rock convergence prediction value; The final surrounding rock convergence prediction value is compared with the dynamic risk threshold. When the final surrounding rock convergence prediction value exceeds the dynamic risk threshold, a machine jam risk warning is triggered. Once the machine malfunction risk warning is triggered, a construction parameter adjustment instruction will be issued.

[0012] In this specification, the sources of the multi-source geological and mechanical parameters include borehole core test data, geological survey report data, and data from the tunnel boring machine on-site monitoring system. Time and space registration is performed on the data from different sources to ensure that each input data is accurately matched with the corresponding tunneling section, and conflicting or abnormal data is eliminated through consistency verification.

[0013] In this specification, the multi-source geological and mechanical parameters and the measured value of the surrounding rock convergence are used as construction data; the weight determination method for the weighted fusion is as follows: a validation set is divided from the construction data, which is the construction data corresponding to a set number of tunneling rings before the current time; the prediction accuracy scores of the support vector machine regression model and the multilayer perceptron neural network model on the validation set are calculated respectively, and the first fusion weight and the second fusion weight are dynamically allocated according to the ratio of the prediction accuracy scores.

[0014] In this specification, the final predicted value of the surrounding rock convergence is compared with the dynamic risk threshold. Based on the degree to which the final predicted value of the surrounding rock convergence exceeds the dynamic risk threshold, the risk level is divided into low risk level, medium risk level, and high risk level. The low risk level corresponds to prompting an increase in monitoring frequency, the medium risk level corresponds to pushing preventive construction adjustment suggestions, and the high risk level corresponds to automatically issuing construction parameter adjustment instructions.

[0015] In this specification, the training dataset is managed by a sliding time window. The sliding time window retains a set number of tunneling rings' construction data and measured values ​​of surrounding rock convergence as training samples in chronological order. Real-time collected construction parameters and measured values ​​of surrounding rock convergence are added to the training dataset, and historical data exceeding the range of the sliding time window are removed. The parameters of the support vector machine regression model and the multilayer perceptron neural network model are updated using an incremental learning algorithm.

[0016] In this specification, the support vector machine regression model and the multilayer perceptron neural network model are deployed on edge computing nodes near the tunnel boring machine construction site. The edge computing nodes have data caching and breakpoint resume functions, and can independently complete model inference and risk assessment when the communication link is interrupted. After the communication link is restored, the local prediction records are automatically synchronized with the central control platform.

[0017] In this specification, the construction parameter adjustment commands include commands for adjusting the advance speed, the cutterhead rotation speed, and the support installation rhythm; the adjusted actual construction parameters are fed back to the training dataset for subsequent model optimization training.

[0018] In this specification, the dynamic risk threshold is set as follows: the historical mean and standard deviation of the measured value of the surrounding rock convergence are calculated, and the product of the historical mean, the safety factor, and the standard deviation is used as the dynamic risk threshold; the safety factor is adaptively adjusted according to the construction stage and geological conditions.

[0019] In this specification, the support vector machine regression model uses a radial basis kernel function to achieve nonlinear mapping; the optimal combination of penalty coefficient, loss insensitive interval, and kernel width parameters of the support vector machine regression model is determined by a joint optimization strategy of cross-validation and hyperparameter search; the joint optimization strategy of hyperparameter search is to first use grid search to lock the approximate optimal interval of parameters in the global scope, and then use a genetic algorithm to perform a fine search within the approximate optimal interval.

[0020] In this specification, the multilayer perceptron neural network model adopts a multilayer fully connected structure, with the hidden layer using a nonlinear activation function and the output layer using a linear activation function. During training, the Levenberg-Marquardt optimization algorithm is used to optimize the parameters, and an early stopping mechanism is introduced to terminate training when the validation set error does not decrease for several consecutive rounds to prevent overfitting.

[0021] In summary, the present invention has at least the following beneficial effects: (1) Significantly improved prediction accuracy and stability. This invention constructs a dual-model prediction architecture of SVR and MLP and adopts a dynamic weighted fusion strategy based on real-time validation set performance, effectively combining the high accuracy advantage of SVR in small sample regression and the adaptability advantage of MLP in complex nonlinear pattern recognition. Compared with single-model prediction methods, the dynamic weighted fusion strategy can automatically adjust the fusion coefficients according to the recent model performance, so that the prediction results are always in the optimal state, significantly improving the accuracy and long-term stability of surrounding rock convergence prediction.

[0022] (2) The multi-source data fusion is sufficient and the feature representation is comprehensive. This invention integrates multi-source data such as borehole core test, geological exploration report, laser section scanning and TBM field monitoring. Through time series alignment, consistency verification and weighted fusion strategies, it ensures that the input features are accurately matched in time and space dimensions, and provides the model with comprehensive and reliable geological and construction parameter inputs.

[0023] (3) The dynamic threshold has strong adaptive capability and high early warning accuracy. This invention adopts a dynamic threshold setting method based on historical convergence distribution statistics and combines it with a safety factor for adaptive adjustment, overcoming the shortcomings of fixed thresholds having high false alarm or false alarm rates under complex geological conditions. The graded early warning mechanism (low, medium and high levels) can trigger construction adjustment strategies according to the differences in risk levels, effectively improving the accuracy and operability of risk early warning.

[0024] (4) A closed-loop linkage of prediction, early warning and control is realized. This invention establishes a real-time data channel with the TBM control system through industrial Ethernet or 5G private network, and automatically issues control commands to adjust the propulsion speed, cutterhead speed and support installation rhythm when the predicted convergence risk reaches a medium to high level. The integrated design of prediction, early warning and control enables the system to proactively intervene before the risk occurs, effectively reducing the probability of machine jamming accidents.

[0025] (5) The model has online adaptive update capability. This invention introduces an online incremental training mechanism based on a sliding time window, which can continuously incorporate the latest collected measured data and remove outdated data as construction progresses, so that the model always maintains the best adaptability to the current geological conditions and construction status, effectively solving the problem of the gradual decrease in prediction accuracy of offline models in long-distance tunneling.

[0026] (6) Low-latency real-time prediction meets the needs of construction sites. This invention deploys the prediction model on edge computing nodes close to the construction site, equipped with GPU acceleration and breakpoint resume function, which can run independently in the event of network interruption, greatly reducing data transmission latency and meeting the strict requirements of TBM construction for real-time prediction and rapid early warning.

[0027] (7) It has good scalability and promotional value. This invention reserves interfaces for accessing more data sources such as ground-penetrating radar, advanced geological forecasting, slag sample analysis, and gas monitoring. It supports access to more complex deep learning models such as CNN, LSTM, and Transformer, and can achieve model parameter sharing and collaborative optimization among multiple construction points through federated learning. This invention is not only applicable to various TBM constructions, but can also be extended to underground engineering fields such as mine roadway tunneling machines and underground hydropower station cavern excavation.

[0028] In summary, this invention achieves high-precision prediction of surrounding rock convergence and early warning of machine jamming risk during TBM construction through the synergistic effects of multi-source data fusion, dynamic weighted fusion prediction of SVR and MLP dual models, dynamic risk threshold setting, hierarchical early warning, and closed-loop construction control. It can maintain high accuracy and stability of prediction under various complex geological conditions, significantly improve construction safety and efficiency, and has outstanding engineering application value and broad promotion potential. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a schematic diagram of the tunnel boring machine surrounding rock convergence prediction and jamming risk warning method based on dynamic weighted fusion of SVR and MLP involved in this invention. Detailed Implementation

[0031] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0032] The following disclosure provides many different implementations or examples for carrying out different structures of the embodiments of the present invention. To simplify the disclosure of the embodiments of the present invention, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. Furthermore, reference numerals and / or reference letters may be repeated in different examples of the embodiments of the present invention; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various implementations and / or arrangements discussed.

[0033] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0034] like Figure 1 As shown, this embodiment provides a method for predicting the surrounding rock convergence of a tunnel boring machine and providing early warning of the machine jamming risk based on dynamic weighted fusion of SVR and MLP, including: The multi-source geological and mechanical parameters of the tunnel boring machine construction section and the measured value of the surrounding rock convergence are obtained. The multi-source geological and mechanical parameters include burial depth, geological strength index, rock mass quality index, uniaxial compressive strength, rock mass cohesion, internal friction angle, elastic modulus, uniaxial compressive strength and uniaxial tensile strength of rock mass. The multi-source geological and mechanical parameters are preprocessed to obtain preprocessed multi-source geological and mechanical parameters. A dynamic risk threshold is set based on the obtained measured values ​​of the surrounding rock convergence. A training dataset is constructed based on the preprocessed multi-source geological and mechanical parameters and the measured value of the surrounding rock convergence, and a support vector machine regression model and a multilayer perceptron neural network model are trained respectively. The multi-source geological and mechanical parameters collected in real time during construction are respectively input into the support vector machine regression model and the multilayer perceptron neural network model after training to obtain the first and second surrounding rock convergence prediction values. The first surrounding rock convergence prediction value and the second surrounding rock convergence prediction value are weighted and fused to obtain the final surrounding rock convergence prediction value; The final surrounding rock convergence prediction value is compared with the dynamic risk threshold. When the final surrounding rock convergence prediction value exceeds the dynamic risk threshold, a machine jam risk warning is triggered. Once the machine malfunction risk warning is triggered, a construction parameter adjustment instruction will be issued.

[0035] In some embodiments, the sources of the multi-source geological and mechanical parameters include borehole core test data, geological survey report data, and data from the tunnel boring machine on-site monitoring system. Time and space registration is performed on the data from different sources to ensure that each input data is accurately matched with the corresponding tunneling section, and conflicting or abnormal data is eliminated through consistency verification.

[0036] In some embodiments, the multi-source geological and mechanical parameters and the measured value of the surrounding rock convergence are used as construction data; the weights of the weighted fusion are determined as follows: a validation set is divided from the construction data, the validation set being the construction data corresponding to a set number of tunneling rings before the current time; the prediction accuracy scores of the support vector machine regression model and the multilayer perceptron neural network model on the validation set are calculated respectively, and the first fusion weight and the second fusion weight are dynamically allocated according to the ratio of the prediction accuracy scores.

[0037] In some embodiments, the final predicted value of surrounding rock convergence is compared with the dynamic risk threshold, and the risk level is divided into low risk level, medium risk level and high risk level according to the degree to which the final predicted value of surrounding rock convergence exceeds the dynamic risk threshold; the low risk level corresponds to prompting an increase in monitoring frequency, the medium risk level corresponds to pushing preventive construction adjustment suggestions, and the high risk level corresponds to automatically issuing construction parameter adjustment instructions.

[0038] In some embodiments, the training dataset is managed by a sliding time window. The sliding time window retains a set number of tunneling rings and measured values ​​of surrounding rock convergence as training samples in chronological order. Real-time collected construction parameters and measured values ​​of surrounding rock convergence are added to the training dataset, and historical data that exceeds the range of the sliding time window are removed. The parameters of the support vector machine regression model and the multilayer perceptron neural network model are updated using an incremental learning algorithm.

[0039] In some embodiments, the support vector machine regression model and the multilayer perceptron neural network model are deployed on edge computing nodes close to the tunnel boring machine construction site; the edge computing nodes have data caching and breakpoint resume functions, independently complete model inference and risk assessment when the communication link is interrupted, and automatically synchronize local prediction records with the central control platform after the communication link is restored.

[0040] In some embodiments, the construction parameter adjustment instructions include advance speed adjustment instructions, cutterhead rotation speed adjustment instructions, and support installation rhythm adjustment instructions; the adjusted actual construction parameters are fed back to the training dataset for subsequent model optimization training.

[0041] In some embodiments, the dynamic risk threshold is set by calculating the historical mean and standard deviation of the measured values ​​of the surrounding rock convergence, and using the product of the historical mean, the safety factor, and the standard deviation as the dynamic risk threshold; the safety factor is adaptively adjusted according to the construction stage and geological conditions.

[0042] In some embodiments, the support vector machine regression model uses a radial basis kernel function to achieve nonlinear mapping; the optimal combination of the penalty coefficient, loss insensitive interval, and kernel width parameters of the support vector machine regression model is determined by a joint optimization strategy of cross-validation and hyperparameter search; the joint optimization strategy of hyperparameter search is to first use grid search to lock the approximate optimal interval of parameters in the global scope, and then use a genetic algorithm to perform a fine search within the approximate optimal interval.

[0043] In some embodiments, the multilayer perceptron neural network model adopts a multilayer fully connected structure, with the hidden layer using a nonlinear activation function and the output layer using a linear activation function; during training, the Levenberg-Marquardt optimization algorithm is used to optimize the parameters, and an early stopping mechanism is introduced to terminate training when the validation set error does not decrease for several consecutive rounds to prevent overfitting.

[0044] In one embodiment, the hidden layer activation functions of the multilayer perceptron neural network model employ an alternating arrangement of hypercursive tangent sigmoid (TanSig) and logarithmic sigmoid (LogSig). Specifically, odd-numbered hidden layers use the TanSig activation function, while even-numbered hidden layers use the LogSig activation function. The TanSig function captures the non-linear variation trend of input features in the positive and negative intervals, while the LogSig function enhances the model's sensitivity to small changes in parameters. This alternating arrangement balances global fitting capability and local detail representation capability in feature learning at different levels.

[0045] In one embodiment, the incremental learning algorithm employs an incremental variant of stochastic gradient descent. Specifically, whenever a predetermined number of training samples containing tunneling loops are added, the current model parameters are used as initial values, and a predetermined number of gradient descent iterations are performed on the new data. The learning rate is set to a predetermined proportion of the initial learning rate (e.g., one-tenth) to avoid catastrophic forgetting. In another embodiment, the incremental learning algorithm employs an elastic weight consolidation algorithm. By calculating the importance weights of the model parameters, constraints are imposed on important parameters during updates to prevent the loss of key knowledge during incremental updates.

[0046] In one embodiment, the adaptive adjustment method of the safety factor k is: based on the safety factor baseline value... =1.8 is the initial value, which is dynamically adjusted according to the current construction stage and geological conditions. Construction stage correction factor. Determined based on TBM tunneling mileage: Initial tunneling stage (pre-set number of loops) Take 1.1, normal tunneling stage Take 1.0, the connection stage (final setting of quantity loop). Take 0.9. Geological condition correction factor. Based on the current Geological Strength Index (GSI): when GSI ≥ 60... Take 0.9, when 40≤GSI<60 Set to 1.0 when GSI < 40 Take 1.2. Final safety factor k = × × .

[0047] In one embodiment, each window covers the complete construction parameters and corresponding monitoring data of the nearest N rings, where N is the set number of tunneling rings, and the preferred value of N is 50 to 200 rings.

[0048] In one embodiment, the "set number of tunneling rings" of the verification set and the "set number of tunneling rings" of the sliding time window can be set independently, and the two can take the same or different values, preferably ranging from 30 to 200 rings; in a preferred embodiment, the verification set takes the 50 tunneling rings before the current time, and the sliding time window takes the most recent 100 tunneling rings.

[0049] In some embodiments, after obtaining the final surrounding rock convergence prediction value and before comparing the final surrounding rock convergence prediction value with the dynamic risk threshold, the present invention further includes a prediction confidence assessment step.

[0050] Since the support vector machine regression model and the multilayer perceptron neural network model can only learn the mapping relationship covered by the training data during the training process, when there are significant differences between the real-time collected construction parameters and the distribution of the training data (such as encountering special geological conditions not seen in the training samples), relying solely on the final surrounding rock convergence prediction value for risk assessment may lead to false alarms or missed alarms. Therefore, this invention introduces a prediction confidence assessment mechanism, attaching a confidence score to each weighted fusion output of the final surrounding rock convergence prediction value for comprehensive reference by the risk warning module during assessment.

[0051] The prediction confidence assessment step specifically includes the following sub-steps: Sub-step 1: Calculate the relative prediction residuals of each base model. For the real-time input data at the current moment, obtain the first surrounding rock convergence prediction value output by the support vector machine regression model and the second surrounding rock convergence prediction value output by the multilayer perceptron neural network model, and calculate the relative residuals of each base model relative to the final surrounding rock convergence prediction value. The formula for calculating the first relative residual is: ; The formula for calculating the second relative residual is: ; in, This is the predicted convergence value of the first surrounding rock. This is the second predicted value for the convergence of the surrounding rock. This is the final predicted value of the surrounding rock convergence. The preset small positive number (preferred value range is) to This is used to prevent the denominator from being zero. First relative residual Compared with the second relative residual The value range is The larger the value, the greater the deviation between the prediction results of the base model and the fusion results, indicating that the prediction behavior of the base model is abnormal under the current input conditions.

[0052] Sub-step 2: Calculate the Mahalanobis distance between the input data and the training dataset. The input vector is constructed from the multi-source geological and mechanical parameters collected in real-time at the current moment. Calculate the Mahalanobis distance of the input vector relative to the training dataset: ; in, Let be the mean vector of all input samples in the training dataset. Let covariance be the covariance matrix of all input samples in the training dataset. for Transpose of Mahalanobis distance. It measures the degree of deviation between the current input vector and the overall distribution of the training dataset. The larger the value, the greater the difference between the current input data and the training data distribution, and the lower the reliability of the model's prediction of that input.

[0053] Sub-step 3: Calculate the overall confidence score. The first relative residual... The second relative residual and the Mahalanobis distance Normalization and fusion are performed to obtain the comprehensive confidence score. : ; in, Let be the standard deviation of the Mahalanobis distance among all input samples in the training dataset. and The preset weighting coefficients (preferably within a certain range) to (Specific values ​​are determined through cross-validation). Overall confidence score. The range of values ​​is , The closer the value is The higher the confidence level of the current prediction result, The closer the value is The lower the confidence level of the current prediction result, the lower the confidence level.

[0054] Sub-step four: Grading processing based on confidence scores. The comprehensive confidence score... Compared with the preset low confidence threshold and high confidence threshold (Preferred) , (The specific value will be adjusted according to the project requirements) for comparison: when When the current prediction result is determined to be a high-confidence prediction, the final surrounding rock convergence prediction value is directly output to the risk warning step for threshold comparison. when When the current prediction result is determined to be a medium confidence prediction, the system outputs the final surrounding rock convergence prediction value to the risk warning step, and at the same time adds a prompt mark of "medium confidence in prediction, manual review is recommended" to the warning information. when If the current prediction result is determined to be a low-confidence prediction, the system will pause the output of the final surrounding rock convergence prediction value to the risk warning step and instead trigger the data quality review process. The data quality review process includes: checking whether there are sensor anomalies in the current input data, checking whether the current construction parameters exceed the coverage of the training data, and suggesting supplementary geological exploration or encrypted monitoring on site. The prediction output will be resumed after the input data returns to normal or is confirmed by human intervention.

[0055] Technical effects of this embodiment: First, a two-dimensional confidence assessment framework. Existing TBM surrounding rock convergence prediction methods are all single-point predictions, only outputting predicted values ​​without providing confidence information. This invention is the first in the field of TBM surrounding rock convergence prediction to construct a confidence assessment system from two independent dimensions: "model prediction consistency" (characterized by the relative residuals of the two models) and "adaptability of input data distribution" (characterized by Mahalanobis distance). The two dimensions complement and corroborate each other, and an anomaly in either dimension can trigger a confidence downgrade.

[0056] Second, the design links relative residuals with fusion results. Existing model fusion methods only focus on the accuracy of the fused predictions, failing to utilize the dispersion between the base model predictions as an indicator of prediction reliability. This invention creatively incorporates the relative residuals of the base models into the confidence assessment—when the predictions of the SVR and MLP models are highly consistent, the relative residuals are small, and the confidence is high; when the predictions of the two models diverge, the relative residuals are large, and the confidence is low. This design utilizes the "dual-model divergence" as a natural indicator of prediction reliability.

[0057] Third, the correlation between Mahalanobis distance and the distribution of training data. Existing methods cannot identify the applicable boundaries of their predictions when encountering unseen geological conditions after model deployment. This invention, by calculating the Mahalanobis distance between the current input and the training dataset, quantitatively characterizes whether the current prediction falls within the model's knowledge coverage, enabling the system to "know what it doesn't know," and proactively downgrade rather than blindly outputting when the input data exceeds the training distribution range.

[0058] Fourth, a confidence-driven tiered processing mechanism. Unlike existing technologies that use confidence assessment only as a post-analysis tool, this invention directly embeds confidence scores into the prediction → early warning workflow. Differentiated processing strategies are adopted based on the confidence level (high / medium / low)—high confidence level triggers a normal early warning, medium confidence level triggers an additional review prompt, and low confidence level suspends output and triggers a data quality review. This achieves a paradigm shift from "passively accepting prediction results" to "actively managing prediction reliability."

[0059] The technical concept of this invention is as follows: This invention first uses multiple types of sensors deployed on the tunnel boring machine and construction section to collect multi-dimensional geological and mechanical parameters, including burial depth, geological strength index, rock mass quality indicators, uniaxial compressive strength, rock mass cohesion, internal friction angle, elastic modulus, and uniaxial compressive and tensile strength of the rock mass. The collected data encompasses real-time information from borehole core experiments, geological survey reports, and the TBM construction monitoring system, thus forming a comprehensive input data foundation.

[0060] In the data processing stage, this invention performs missing value imputation, outlier removal, and normalization on the original data, mapping parameters of different physical dimensions to the [0,1] interval. When the feature dimension is high or redundant information exists, principal component analysis and other methods are used for feature dimensionality reduction, which reduces computational complexity and improves the stability and generalization ability of model training.

[0061] This invention constructs two artificial intelligence prediction models: a support vector machine (SVM) regression model and a multilayer perceptron neural network (MLPN) model. The SVM regression model uses a radial basis function (RBF) kernel to achieve nonlinear mapping and optimizes key hyperparameters such as the penalty coefficient, loss-insensitive interval, and kernel width through cross-validation. The MPN model adopts a multilayer fully connected structure, with the hidden layer activation function using a combination of hypercursive tangent sigmoid and logarithmic sigmoid, and the output layer using a linear activation function to adapt to continuous prediction tasks.

[0062] In the prediction phase, the present invention inputs the preprocessed real-time construction parameters into the trained support vector machine and multilayer perceptron models, and fuses the prediction results of the two by weighted averaging. Combining the high accuracy advantage of support vector machine in small sample regression and the adaptability advantage of multilayer perceptron in complex pattern recognition, a more stable and robust prediction output is obtained.

[0063] This invention establishes an adaptively adjustable risk threshold. This threshold is dynamically adjusted based on the distribution of historical construction data and safety margins to adapt to different geological conditions and construction stages. When the predicted surrounding rock convergence exceeds this threshold, the system immediately triggers a risk warning.

[0064] The risk warning module compares the prediction results with the risk threshold and generates construction adjustment suggestions when the threshold is reached or exceeded. These suggestions include optimization schemes for parameters such as advance speed, cutterhead speed, and support installation rhythm. The suggestions are then executed through the communication interface with the TBM control system, achieving closed-loop linkage between prediction, warning, and control.

[0065] To ensure the model maintains high-accuracy prediction capabilities even when geological conditions change, this invention introduces an online incremental training mechanism. This mechanism updates the training dataset based on a sliding time window, incorporating the latest collected measured data into the training samples and discarding outdated data, thereby maintaining the model's adaptability to current construction conditions.

[0066] The predictive model of this invention can be deployed on edge computing nodes close to the construction site, including industrial computers or portable servers with GPU acceleration capabilities, to achieve low-latency prediction and local early warning. In the event of a network outage, the edge computing nodes can continue to operate independently and cache the prediction results, synchronizing with the central monitoring platform once the network is restored.

[0067] In the data processing process, this invention not only performs conventional normalization and feature dimensionality reduction, but also adopts a weighted fusion strategy for data from different sources and with different levels of precision. This makes high-precision data (such as laboratory test results) have a higher weight in training, while low-precision or noisy data (such as some online monitoring data) automatically have a lower weight in model training, thereby reducing the impact of uncertainty on the prediction results.

[0068] This invention employs an optimization strategy combining grid search and genetic algorithm in the hyperparameter search of the support vector machine model. First, grid search is used to quickly lock the approximate optimal range of parameters over a large range. Then, the genetic algorithm is used to perform a fine search within a local range to obtain a better combination of penalty coefficient C, loss-insensitive range ε, and kernel parameter γ, thereby improving the robustness of the model under different geological conditions.

[0069] This invention employs the Levenberg-Marquardt optimization algorithm instead of the traditional gradient descent algorithm in the training of multilayer perceptron models, improving the convergence speed and accuracy of the model in nonlinear regression problems. Simultaneously, an early stopping mechanism is introduced to terminate training if the validation set error fails to decrease for several consecutive rounds, thus preventing overfitting.

[0070] The multi-model weighted fusion strategy of this invention can not only use static weights, but also dynamically adjust the weights based on the performance of the real-time validation set. This dynamic adjustment module can continuously evaluate the accuracy of SVM and MLP models during the construction process and allocate new fusion coefficients according to the performance over a recent period, so that the fusion prediction results are always in the optimal state.

[0071] The risk threshold calculation module of this invention supports adaptive adjustment based on statistical methods. Specifically, the threshold T = μ + kσ is determined by calculating the historical convergence mean μ and standard deviation σ, combined with a safety factor k. In special geological sections, out-of-distribution detection methods can be introduced to identify abnormal deformation trends, and the threshold can be tightened or relaxed accordingly.

[0072] The risk warning mechanism of this invention is not only a simple over-limit alert, but also includes a graded warning mechanism that divides the risk level into three levels: low, medium, and high. For different risk levels, this system generates differentiated construction adjustment strategies. For example, for low risk, only monitoring is prompted; for medium risk, it is recommended to set up support in advance; and for high risk, instructions to reduce the advance speed and adjust the cutterhead operation are directly issued.

[0073] The closed-loop construction control interface of this invention uses industrial Ethernet or wireless industrial protocols (such as Wi-Fi 6 industrial version, 5G private network) to achieve high-speed data interaction with the TBM control system, ensuring that the transmission delay of prediction and adjustment commands does not exceed 1 second, thereby meeting the real-time requirements.

[0074] The online incremental training module of this invention uses a sliding window mechanism to manage the training dataset. Each update uses only the construction and monitoring data of the most recent N rings and discards the oldest data to ensure that the model reflects the latest features of the current geological and construction status. This module supports execution locally on edge computing nodes, avoiding the delay caused by the backhaul of a large amount of data.

[0075] This invention supports a hot update mechanism for models, which loads a model version trained and optimized with new data without stopping the prediction service, thus ensuring uninterrupted construction. The hot update process includes parallel operation of the old and new models, performance comparison, and seamless switching, with a switching latency of no more than 100 milliseconds.

[0076] When deployed at edge computing nodes, this invention is equipped with data caching and breakpoint resume functions, which can still perform predictions normally even in the event of network interruption, and automatically synchronize local prediction records with the central control platform after the network is restored, ensuring data integrity and continuity of construction logs.

[0077] This invention can be extended to access more data sources, such as ground-penetrating radar, advanced geological forecasting, slag sample analysis, and gas monitoring, to further enhance the comprehensiveness of model input and improve the ability to predict and prevent special geological disasters (such as water inrush, mudslides, and gas).

[0078] This invention is not only applicable to the prediction of surrounding rock convergence and the early warning of machine jamming risk during the construction of tunnel boring machines, but can also be extended to the deformation prediction and risk management of other underground engineering equipment, such as mine roadway tunnel boring machines, underground hydropower station cavern excavation and large-section underground space engineering, etc., and has high versatility and promotion value.

[0079] This invention achieves high-precision real-time prediction of the surrounding rock convergence trend and prevention and control of machine jamming risk in TBM construction through the synergistic effect of innovative technologies such as multi-source data fusion, artificial intelligence modeling, multi-model fusion, dynamic risk threshold, adaptive construction adjustment, online incremental training and edge deployment, effectively improving construction safety, automation level and engineering economy.

[0080] This invention introduces a data quality assessment mechanism throughout the model training and prediction process. By setting three categories of indicators—data integrity, timeliness, and consistency—the real-time collected data is scored. When the score is lower than a set threshold, data re-collection or removal is automatically triggered, thereby preventing low-quality data from adversely affecting prediction accuracy.

[0081] The multi-source data fusion module of this invention has a time series alignment function, which can align data with different sampling frequencies and different acquisition times to a unified time base, and use linear interpolation or spline interpolation to fill in the missing data points, ensuring the consistency of input features in the time dimension and improving the model's utilization of time series information.

[0082] This invention introduces a confidence assessment mechanism in the prediction output stage of support vector machine and multilayer perceptron models. By analyzing information such as model residual distribution and predicted value variance, a confidence score is added to each prediction result, which is then comprehensively considered by the risk warning module when making judgments, so as to avoid false alarms or missed alarms due to a single abnormal prediction.

[0083] The dynamic weight adjustment algorithm of this invention can automatically change the weight ratio of SVM and MLP in the fusion prediction according to the changes in the construction stage. For example, in the early stage when the geological conditions are stable and the number of samples is small, the weight of SVM is increased first; in the middle and later stages when the amount of data gradually increases and the geological changes are large, the weight of MLP is increased, so as to give full play to the complementary advantages of the two types of models.

[0084] The risk classification strategy of this invention can be combined with the task scheduling of the construction management system. It can not only adjust construction parameters under high-risk conditions, but also automatically postpone key processes in the plan and prepare emergency support plans and construction materials in advance, thereby organically combining risk management with construction plan optimization.

[0085] The online incremental training mechanism of this invention supports distributed execution, that is, local model updates are performed on multiple edge computing nodes, and the updates of each node are periodically merged into a global model through parameter aggregation algorithms (such as FedAvg), so as to realize knowledge sharing among multiple construction points and ensure data privacy and security.

[0086] The edge computing node of this invention deploys a lightweight model inference engine, which can quickly run SVM and MLP models under limited computing resources, and supports batch prediction mode, which can process data from multiple time windows in one inference, improving real-time performance and throughput.

[0087] The system architecture provided by this invention reserves open interfaces, allowing for the subsequent integration of deep learning models (such as Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), Transformers, etc.) as additional prediction models, and incorporating them into a multi-model fusion framework to achieve iterative upgrades in prediction capabilities.

[0088] The early warning information output of this invention can be presented in various forms, including on-site audible and visual alarms, control room large screen display, and mobile terminal push, to adapt to different construction organization and management models and ensure that the early warning information is transmitted to relevant personnel in the shortest possible time.

[0089] In the implementation of this invention, all prediction, early warning and control commands are recorded in a log. The log includes data source, model version, predicted value, risk level, adjustment command and execution feedback, which facilitates subsequent construction analysis, model optimization and accident tracing.

[0090] The technical effects of this invention are as follows: by fully integrating multi-source geological and construction data, and combining artificial intelligence algorithm modeling and multi-model fusion prediction, the accuracy and stability of surrounding rock convergence prediction are significantly improved; by introducing dynamic risk thresholds and closed-loop construction control, prediction, early warning and construction adjustment are integrated; and by supporting online model optimization and edge computing deployment, the real-time construction needs under complex and variable geological conditions can be met.

[0091] In summary, this invention achieves high-precision prediction of surrounding rock convergence and early warning of machine jamming risk during TBM construction through multi-source data fusion, artificial intelligence modeling, multi-model fusion prediction, dynamic risk threshold setting, hierarchical early warning, and closed-loop construction control. It can maintain high accuracy and stability of prediction under various complex geological conditions, significantly improve construction safety and efficiency, and reduce the incidence of sudden shutdowns and accidents. It has outstanding engineering application value and broad promotion potential.

[0092] In the specific implementation of this invention, some steps can be appropriately adjusted according to different geological conditions, construction equipment models and construction organization designs, so as to achieve the optimal performance of prediction and early warning functions under different working conditions.

[0093] like Figure 1 As shown, this invention presents an artificial intelligence-based method for predicting surrounding rock convergence and providing early warning of machine jamming risks for tunnel boring machines (TBMs). The method comprises a data acquisition module, a data preprocessing module, an AI prediction model module, a multi-model fusion module, a risk threshold calculation module, a risk early warning module, and a construction control closed-loop module. These modules interact bidirectionally via high-speed data links and control signals, forming a closed-loop system for real-time prediction and construction adjustments. This system can be deployed at edge computing nodes or a central dispatch and control center on the construction site, and achieves low-latency communication with the TBM main control system via industrial Ethernet or a 5G private network.

[0094] The data acquisition module is used to obtain multi-dimensional input parameters related to surrounding rock deformation prediction during construction. These include: burial depth H, geological strength index GSI, rock mass quality index RQD, uniaxial compressive strength UCS, rock mass cohesion c, internal friction angle φ, elastic modulus E, uniaxial compressive strength scm, and uniaxial tensile strength stm. These parameters are obtained through various methods: ① Drill core tests are used to obtain the mechanical properties and integrity indicators of rocks; ② The geological survey report provides information on the macroscopic geological distribution, fault and fracture zone distribution; ③The TBM field monitoring system collects real-time data on propulsion force, cutterhead speed, slag density, and support installation. ④ A laser profile scanner collects tunnel cross-sectional deformation information; ⑤ The convergence meter is used to measure the radial displacement change of a critical section.

[0095] All data is centrally aggregated through the data acquisition gateway and synchronized to the data preprocessing module.

[0096] In a preferred embodiment, the geological parameters such as cohesion c, internal friction angle φ, elastic modulus E, uniaxial compressive strength scm, and uniaxial tensile strength stm can be calculated using the Hoek-Brown criterion and the Mohr-Coulomb criterion. Taking the Hoek-Brown criterion as an example, its basic expression is: ; in, This represents the maximum principal stress at which the rock mass fails. For the minimum principal stress, For the uniaxial compressive strength of rock, It is the Hoek–Brown constant. The rock mass integrity coefficient. These are empirical indices. These coefficients can be fitted using regression analysis based on GSI and rock test data to obtain mechanical characteristic parameters that can be used as input for AI models.

[0097] The data preprocessing module is used to normalize the original input parameters to eliminate the influence of different units and numerical ranges on model training. The normalization formula used in this invention is as follows: ; in, For the original data in the i-th group and j-th column, and These are the maximum and minimum values ​​of the j-th column parameter, respectively. This method can map all input features to... The interval ensures the comparability of the contributions of different features to the model prediction and helps to accelerate the convergence speed of gradient descent.

[0098] In the data preprocessing process, to further improve the reliability of the input data, the present invention performs the following processing steps: ① Missing value imputation: Interpolation using the mean of time series data from adjacent processes is preferred. If the missing value rate exceeds the threshold, multiple imputation (MICE) is used for imputation. ② Outlier Removal: Outliers are identified using the interquartile range (IQR) method. Outliers are removed when data falls within a certain range. Outliers outside the mean are considered abnormal and removed, or the 3σ principle is used to remove outliers exceeding the mean ± 3 times the standard deviation. ③ Feature dimensionality reduction: Principal component analysis (PCA) is used to reduce redundant feature dimensions, retaining only principal components with a cumulative contribution rate of over 95%, in order to reduce the computational load of the model and improve generalization performance.

[0099] The AI ​​prediction model module includes a Support Vector Machine Regression (SVR) model and a Multilayer Perceptron (MLP) neural network model. The SVR model uses a Radial Basis Function (RBF): ; in, The kernel width parameter determines the distribution density of support vectors in the feature space. This model avoids the overfitting problem of traditional regression models in predicting nonlinear complex relationships by minimizing structural risk.

[0100] The multilayer perceptron model consists of an input layer, several hidden layers, and an output layer. The hidden layer activation functions combine the hypercursive tangent sigmoid (TanSig) and logarithmic sigmoid (LogSig) to simultaneously ensure the smoothness of the nonlinear feature map and the gradient stability within the saturation region. The output layer uses a linear activation function to directly regress the continuous values ​​of the maximum convergence of the surrounding rock. During training, the Levenberg-Marquardt (LM) optimization algorithm is employed. This algorithm combines the advantages of gradient descent and Gauss-Newton methods, offering fast convergence and high stability, and incorporates an early stopping mechanism to prevent overfitting.

[0101] In a preferred embodiment, both the SVR and MLP models employ K-fold cross-validation (preferably K=10) for hyperparameter optimization. For SVR, the main optimizations are the penalty parameter C (controlling the balance between model complexity and error), the loss-insensitive interval ε (determining regression tolerance), and the kernel width γ (affecting the smoothness of feature mapping). For MLP, the optimizations are the number of hidden layers, the number of neurons per layer, the learning rate, and the regularization coefficient to ensure the model's generalization ability under different geological conditions.

[0102] The multi-model fusion module combines the prediction results of SVR and MLP in a weighted average manner to offset the bias of a single model under specific operating conditions. The fusion formula is as follows: ; in, The weights can be dynamically adjusted based on the prediction error of the historical validation set. For example, when the SVR performs better under recent operating conditions, the system can automatically increase the weights. The proportion, and vice versa, increases. The weight.

[0103] The risk threshold calculation module dynamically sets the early warning threshold based on the distribution of the maximum convergence amount in historical construction sections. The calculation formula is: ; in, This is the average of the historical maximum convergence. Standard deviation, The safety factor is set at (preferred range 1.5~2.0, determined according to the engineering safety level). This dynamic threshold method can adapt to the deformation characteristics of different geological sections, avoiding excessive false alarms or missed alarms caused by static thresholds under high-risk or low-risk conditions.

[0104] The risk warning module categorizes risk levels into three levels: low, medium, and high, based on the ratio of the predicted result to the risk threshold. Specifically, when the predicted surrounding rock convergence exceeds the dynamic threshold but does not exceed 1.2 times that threshold, the system automatically classifies it as low risk; when the predicted value exceeds 1.2 times the threshold but does not exceed 1.5 times the threshold, it is classified as medium risk; and when the predicted value exceeds 1.5 times the threshold, it is directly classified as high risk. In actual operation, this classification strategy can be adjusted according to the safety level of different geological sections to adapt to engineering needs with different risk tolerances, and it supports real-time modification of risk classification parameters on the system interface without stopping the prediction service.

[0105] In low-risk situations, the system prompts construction personnel to increase the monitoring frequency of the current section through multiple terminal interfaces (including on-site control console, dispatch center, and mobile devices), such as shortening the convergence measurement interval and increasing the number of slag sample tests. In medium-risk situations, the system generates and pushes preventive construction adjustment suggestions, including pre-laying or reinforcing supports, optimizing the ratio of advance speed to cutterhead rotation speed, and adjusting the slag amendment ratio. In high-risk situations, the risk warning module directly calls the construction control closed-loop module, sending real-time adjustment commands through a high-speed data link with the TBM main control system. This automatically executes measures including reducing advance speed, adjusting cutterhead torque, accelerating support installation speed, and increasing advanced geological surveys to ensure effective intervention before deformation exceeds limits.

[0106] The construction control closed-loop module establishes a real-time data channel with the TBM control system using an industrial Ethernet or 5G private network, achieving millisecond-level response and ensuring that the end-to-end latency from risk prediction to construction parameter adjustment does not exceed 1 second. In a preferred embodiment, the closed-loop module is also equipped with a command execution status feedback function, which verifies the accuracy of the control response by comparing the deviation between the actual execution parameters and the issued commands, and triggers a secondary control compensation mechanism when the deviation exceeds a set range.

[0107] The online incremental training module utilizes a sliding time window mechanism to update the model training data. Each window covers the complete construction parameters and corresponding monitoring data of the most recent N rings (preferably 50 to 200 rings). As the window slides, the system automatically removes historical data that exceeds the window range to reduce the interference of old data on the model's predictive performance under current geological conditions. This mechanism enables the model weights to be dynamically updated as construction conditions gradually change, improving adaptability to gradual or abrupt geological changes.

[0108] Model training and prediction are both deployed on edge computing nodes close to the construction site. These nodes are equipped with high-performance GPUs for accelerating deep learning computations and have offline caching capabilities. When communication with the central control platform is interrupted, the edge nodes can independently complete model inference, risk assessment, and control command issuance, and cache all running data and prediction results locally. Once the network is restored, the system automatically performs data retransmission and result synchronization to ensure the integrity and continuity of construction records.

[0109] In one embodiment, the edge computing node integrates a lightweight inference engine (such as an AI inference framework optimized based on TensorRT or ONNXRuntime), which controls the computation time of a single prediction to within 500 milliseconds and supports batch processing mode, enabling parallel inference of data from multiple time windows simultaneously, thereby maintaining the real-time performance and stability of predictions under complex operating conditions.

[0110] During operation, the system records all predicted values, risk levels, control commands, and their execution feedback information in a timestamp-ordered log database. The log data includes not only numerical information but also metadata such as risk assessment criteria, model version number, and training data time range. This facilitates subsequent engineering analysis, model performance evaluation, and accountability, and supports multi-condition queries by time, construction section, risk level, and other criteria.

[0111] In a real-world TBM construction case, the method of this invention significantly improved the accuracy of convergence prediction, and the coefficient of determination R on the test set increased. 2 The accuracy was stabilized above 0.96, and the mean square error (MSE) was reduced to 0.12. Compared with traditional empirical prediction methods, potential machine jamming risks were detected an average of 3 to 5 rings earlier. Sudden downtime events during construction were reduced by about 30%, significantly improving construction efficiency and equipment utilization.

[0112] In another embodiment, the system of the present invention incorporates data obtained from ground-penetrating radar and advanced drilling. This data is input as additional features into the AI ​​prediction model, resulting in a reduction of convergence prediction error in complex geological sections such as fault fracture zones by approximately 20% compared to when this data is not integrated. In long-distance tunnel construction under variable working conditions, this solution effectively reduces prediction misjudgments caused by geological abrupt changes.

[0113] The method of this invention is not only applicable to single-shield TBM construction, but also to double-shield and hybrid shield tunneling machine construction. In areas with weak surrounding rock, high ground stress, and abundant water, the method maintains stable predictive performance, especially in high ground stress sections where it can identify potential rockburst risks in advance and issue early warnings of water inrush in water-rich sections.

[0114] When the system of this invention is deployed simultaneously at multiple construction sites, federated learning can be used to share model parameter updates among different construction sites without transmitting the original data. This mechanism not only protects the data privacy and security of each construction project, but also improves the generalization ability of the overall model under different geological conditions through cross-project model knowledge fusion.

[0115] The system interface supports cross-platform and multi-terminal access, including LED screens at construction sites, monitoring terminals in dispatch rooms, and mobile applications supporting iOS and Android. Each terminal can display different levels of information based on user permissions. For example, construction personnel can view the risk level and adjustment suggestions for the current construction section, while managers can view the overall risk trend and model operation status, enabling information synchronization and collaborative decision-making among all personnel.

[0116] Upon triggering a high-risk event, the system automatically archives the relevant data, execution instructions, and feedback results, and pushes them to the construction safety management platform, automatically generating an accident prevention analysis report. This report includes an analysis of the cause of the risk, an assessment of its potential impact, the countermeasures already taken, and follow-up recommendations, facilitating the construction unit to develop more scientific prevention and control strategies in subsequent construction.

[0117] The system is designed with interfaces reserved for deep learning frameworks, supporting the future integration of more complex model architectures, such as convolutional neural networks (CNN) for extracting image-based geological features, long short-term memory networks (LSTM) for capturing the temporal dependence of construction parameters, and the Transformer architecture for enhancing multivariate long-term time-series prediction capabilities. This will enable more refined and adaptive convergence prediction and risk warning functions in future construction scenarios.

[0118] The risk threshold calculation module not only relies on historical data statistics and dynamic calculation results, but also supports external expert intervention. Under special working conditions or emergencies (such as geological disasters, major equipment failures, sudden water inrushes, etc.), on-site engineers or geological experts can input manually set risk thresholds in real time through the system interface, overriding the automatically calculated results, thus ensuring a high degree of consistency between the system's risk assessment and emergency construction strategies. In manual adjustment mode, the system automatically records the adjustment reason, adjustment time, operator identity, and the original automatically calculated threshold, ensuring subsequent traceability and analysis.

[0119] The prediction process of this invention can be seamlessly integrated with a BIM (Building Information Modeling) platform. Through dynamic association with 3D geological and structural models, the predicted convergence values ​​are mapped in real time to the BIM visualization environment, forming a 3D risk distribution map of the construction section. This function can intuitively present potential deformation areas, risk levels, and changing trends, facilitating construction managers to quickly identify high-risk points in 3D space and formulate targeted construction measures. It also supports VR / AR device access, enabling immersive risk visualization.

[0120] The hardware required for system deployment includes high-performance computers (IPCs) with industrial-grade protection capabilities, GPU accelerator cards (for deep learning inference and training), multi-source data acquisition terminals (including laser profile scanners, convergence meters, pressure sensors, ground-penetrating radar receivers, etc.), and high-speed communication modules (industrial Ethernet or 5G private network). All hardware must meet the requirements of the high humidity, high dust, and strong vibration conditions at tunnel construction sites and possess modular replacement capabilities to reduce maintenance downtime.

[0121] This invention can be deeply integrated with a construction log management system to form a full lifecycle digital management chain covering data collection, predictive calculation, risk warning, construction adjustment, and result feedback. When recording construction logs, the system automatically associates the predicted values, risk levels, adjustment measures, and implementation effects of each stage of tunneling, thereby providing data support for subsequent construction strategy optimization and experience accumulation.

[0122] The system software adopts a modular architecture design, with each functional module (data acquisition, preprocessing, model prediction, multi-model fusion, risk calculation, closed-loop control, etc.) running as an independent service, and communicating with each other through API interfaces. This architecture not only facilitates functional expansion and version iteration, but also enables rapid restart or hot-swap to restore functionality in the event of a module failure, improving system maintainability and reliability.

[0123] Regarding energy conservation and optimization, this invention can dynamically adjust the data acquisition frequency and model operation frequency according to the real-time risk level. In low-risk conditions, the system automatically reduces the frequency of data acquisition and inference calls to decrease sensor power consumption and computational load; in high-risk conditions, the system automatically switches to a high-frequency monitoring mode to ensure timely detection and rapid response to risk changes. This mechanism can effectively reduce energy consumption and hardware wear during long-term construction.

[0124] The model training phase supports multi-threaded parallel computing and can utilize the multi-core architecture of GPUs for batch matrix operations to accelerate training, significantly reducing training time. For historical datasets containing millions of samples, the system, with GPU acceleration and parallel optimization, can reduce training time from several hours to several minutes, thereby improving the timeliness of model updates.

[0125] The system can be configured with a periodic self-check function, which automatically checks the connectivity of the data acquisition link, the working status of sensors, the health of model operation, and the usage of storage space at a set cycle. When problems such as data interruption, sensor drift, or abnormal model prediction are detected, the system will automatically generate alarms and push them to the operation and maintenance personnel, while also recording them in the system operation and maintenance log to ensure that problems can be fixed in an early stage.

[0126] When the system detects a persistent, systematic deviation between predicted and measured values ​​over a long period (e.g., the mean deviation consistently exceeds a set threshold), it will automatically trigger a model retraining process. This process will optimize the model parameters based on the latest collected field data, update the prediction weights, and automatically deploy the new model to edge computing nodes after it passes the validation set accuracy test, ensuring the long-term stability of prediction performance.

[0127] Construction parameter adjustment suggestions are generated collaboratively by a built-in expert knowledge base and an AI decision engine. The expert knowledge base stores empirical construction adjustment strategies for various geological conditions, while the AI ​​decision engine dynamically generates the optimal adjustment plan by combining real-time prediction results and historical construction data. This dual-engine mechanism can improve the relevance and effectiveness of suggestions by taking into account expert experience while leveraging the adaptive capabilities of machine learning.

[0128] The system supports a multilingual user interface, including Chinese, English, Spanish, and other languages, and can automatically switch the interface language based on the user's login region. This feature facilitates the system's application in construction projects across different countries and regions, and is particularly suitable for international joint-contracting projects.

[0129] The data communication process employs the TLS (Transport Layer Security) encryption protocol to ensure that data transmitted between the data collection terminal, edge nodes, and central server is protected from man-in-the-middle attacks and data theft during network transmission. It also supports a two-way authentication mechanism based on digital certificates to further enhance communication security.

[0130] The system-generated logs can be integrated with third-party security audit platforms to achieve full compliance tracking of construction data, forecast results, and adjustment instructions, meeting the security audit and compliance inspection requirements of engineering supervision departments or owners. Log storage supports hierarchical encryption and access control to prevent unauthorized access and tampering.

[0131] The software algorithm code of this invention can be deployed by encapsulating it in Docker containers. The container image can run across different operating system platforms (such as Linux and Windows Server) and supports rapid migration and elastic scaling. This containerization solution can significantly simplify the system deployment process and reduce the environment configuration costs when migrating between different construction sites.

[0132] The system can achieve fault-tolerant operation by deploying redundant hardware and software instances on critical edge computing nodes. When the primary node fails, the backup node can take over the task within seconds, ensuring that the prediction and risk warning functions are uninterrupted, thereby improving the system's reliability during high-risk construction phases.

[0133] In long-distance tunnel construction, this invention supports the segmented deployment of multiple prediction units. Each unit independently collects data for its segment and performs local predictions, which are then managed and coordinated by a central scheduling platform. This distributed architecture can reduce data transmission latency while ensuring prediction timeliness, and ensures that the normal operation of other units is not affected when one unit malfunctions.

[0134] The model version management module can record and track parameter settings, training dataset range, and performance metrics (such as R) for each model training and deployment. 2 This allows for the creation of a traceable model lifecycle archive, including information such as model type, MSE, and MAE, along with deployment time. This feature not only aids in engineering quality management and technology iteration evaluation but also enables rapid rollback to a historically stable version when model performance fluctuates.

[0135] The system supports two-way data linkage with the construction progress management system. By receiving real-time information such as construction task plans, ring number progress, and equipment team arrangements, it dynamically adjusts the construction pace and resource allocation based on current risk prediction results. For example, when predicting construction in high-risk sections, the system can suggest slowing down the progress, increasing the number of support teams, or allocating additional monitoring equipment to minimize construction risks; while in low-risk sections, it can suggest appropriately accelerating the construction pace to improve overall project efficiency.

[0136] Upon triggering a high-risk warning, the system automatically executes a full data protection mechanism, storing all raw monitoring data, processed characteristic data, prediction results, and control commands within a set time window (preferably 30 minutes to 2 hours) before and after the risk trigger in a dedicated accident investigation database. This data is archived chronologically and includes a data integrity check code to ensure the accuracy and traceability of subsequent accident investigations and analyses.

[0137] The system can automatically generate predictive trend charts, risk level change curves, and risk statistics charts based on historical records throughout the entire construction cycle, including the temporal distribution of high, medium, and low risks, risk trigger frequency, and the probability of risk occurrence in different geological sections. These visualization results can be displayed in the dispatch center, project management platform, and mobile devices, providing management with quantitative decision-making basis and assisting in the formulation of subsequent construction plans and risk control strategies.

[0138] System maintenance personnel can upgrade and optimize model parameters, system configuration files, and software versions through a secure and encrypted remote interface (supporting SSH, VPN, or dedicated tunnel protocols). In remote maintenance mode, the system automatically generates a complete backup before the upgrade, ensuring a quick rollback to a stable version in case of upgrade anomalies, thereby reducing maintenance risks.

[0139] The hardware and software architecture of this invention adopts a modular replacement design. The hardware modules include computing nodes, communication modules, sensor terminals, etc., while the software modules include data acquisition, preprocessing, predictive modeling, risk analysis, and control command generation. This modular design not only reduces the difficulty of repairing a single module failure but also allows different construction projects to flexibly combine system functions according to their own needs, thereby effectively reducing overall maintenance costs.

[0140] In low-risk construction areas, the system can automatically switch to energy-saving mode, reducing data acquisition frequency, minimizing model calls, and limiting the proportion of background computing resources used, thereby reducing GPU and CPU power consumption and extending hardware lifespan. This mode is particularly suitable for scenarios involving long-term construction under stable geological conditions, significantly reducing system energy consumption and operating costs.

[0141] Bayesian optimization algorithms can be introduced during the model optimization phase to perform a global optimization search on the hyperparameters (such as C, ε, γ, hidden layer size, learning rate, etc.) of models like SVR and MLP. Compared to traditional grid search and random search, Bayesian optimization can find better parameter combinations with fewer iterations, thereby reducing computational overhead while maintaining prediction accuracy.

[0142] In the anomaly detection phase, the system can introduce the Isolation Forest algorithm to identify unsupervised outliers in the input data. This algorithm identifies observations that significantly differ from historical data distributions by recursively partitioning the feature space, effectively eliminating anomalous data caused by sensor malfunctions, communication errors, or extreme geological conditions, thereby improving the stability and accuracy of model training and prediction.

[0143] The risk warning function can be synchronized with the tunnel safety broadcast system. When the system detects a medium- or high-risk level, it automatically triggers an internal tunnel broadcast or voice prompt to remind construction personnel to take timely protective measures. Furthermore, it can be linked with on-site warning lights, variable message signs, and other hardware to form a multi-channel, multi-form safety reminder mechanism, ensuring that information is promptly delivered to all relevant personnel.

[0144] The system's operating parameters can be customized according to the management standards of different construction companies, including risk level classification standards, data collection frequency, model update cycle, early warning trigger delay, and log retention period. This function ensures that the system meets the corresponding safety, quality, and management requirements in projects of different companies, countries, or regions.

[0145] The overall system of this invention can be deployed on a cloud-based centralized management platform to achieve data aggregation, model sharing, and unified scheduling across multiple construction projects. It also supports a local independent operation mode under conditions of unstable network connectivity, meeting the real-time prediction and early warning needs of construction sites in remote areas. This dual-mode architecture provides flexible deployment options for projects of different sizes and conditions.

[0146] In summary, this invention achieves high-precision prediction of surrounding rock convergence and early warning of machine jamming risk during TBM construction through multi-source data fusion, artificial intelligence modeling, multi-model fusion prediction, dynamic risk threshold setting, hierarchical early warning, and closed-loop construction control. It can maintain high accuracy and stability of prediction under various complex geological conditions, significantly improve construction safety and efficiency, and reduce the incidence of sudden shutdowns and accidents. It has outstanding engineering application value and broad promotion potential.

[0147] The embodiments described above are for illustrative purposes only and are not intended to limit the invention. Therefore, any changes in numerical values ​​or substitutions of equivalent elements should still fall within the scope of this invention.

[0148] The above detailed description will enable those skilled in the art to understand that the present invention can indeed achieve the aforementioned objectives and has complied with the provisions of the Patent Law.

[0149] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention. The above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

[0150] It should be noted that the above description of the process is for illustrative purposes only and does not limit the scope of this specification. Those skilled in the art can make various modifications and changes to the process under the guidance of this specification. However, these modifications and changes remain within the scope of this specification.

[0151] The basic concepts have been described above. Obviously, for those skilled in the art who have read this application, the above disclosure is merely illustrative and does not constitute a limitation of this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, and therefore, such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this application.

[0152] Furthermore, this application uses specific terms to describe its embodiments. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different positions in this specification do not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application can be appropriately combined.

Claims

1. A method for predicting surrounding rock convergence and predicting jamming risk of tunnel boring machines based on dynamic weighted fusion of SVR and MLP, characterized in that, include: The multi-source geological and mechanical parameters of the tunnel boring machine construction section and the measured value of the surrounding rock convergence are obtained. The multi-source geological and mechanical parameters include burial depth, geological strength index, rock mass quality index, uniaxial compressive strength, rock mass cohesion, internal friction angle, elastic modulus, uniaxial compressive strength and uniaxial tensile strength of rock mass. The multi-source geological and mechanical parameters are preprocessed to obtain preprocessed multi-source geological and mechanical parameters. A dynamic risk threshold is set based on the obtained measured values ​​of the surrounding rock convergence. A training dataset is constructed based on the preprocessed multi-source geological and mechanical parameters and the measured value of the surrounding rock convergence, and a support vector machine regression model and a multilayer perceptron neural network model are trained respectively. The multi-source geological and mechanical parameters collected in real time during construction are respectively input into the support vector machine regression model and the multilayer perceptron neural network model after training to obtain the first and second surrounding rock convergence prediction values. The first surrounding rock convergence prediction value and the second surrounding rock convergence prediction value are weighted and fused to obtain the final surrounding rock convergence prediction value; The final surrounding rock convergence prediction value is compared with the dynamic risk threshold. When the final surrounding rock convergence prediction value exceeds the dynamic risk threshold, a machine jam risk warning is triggered. Once the machine malfunction risk warning is triggered, a construction parameter adjustment instruction will be issued.

2. The method for predicting surrounding rock convergence and predicting jamming risk of tunnel boring machines based on dynamic weighted fusion of SVR and MLP as described in claim 1, characterized in that, The sources of the multi-source geological and mechanical parameters include borehole core test data, geological survey report data, and data from the tunnel boring machine on-site monitoring system. Time and space registration is performed on the data from different sources to ensure that each input data is accurately matched with the corresponding tunneling section, and conflicting or abnormal data is eliminated through consistency verification.

3. The method for predicting surrounding rock convergence and predicting jamming risk of tunnel boring machines based on dynamic weighted fusion of SVR and MLP as described in claim 1, characterized in that, The multi-source geological and mechanical parameters and the measured value of the surrounding rock convergence are used as construction data; the weights of the weighted fusion are determined as follows: a validation set is divided from the construction data, which is the construction data corresponding to a set number of tunneling rings before the current time; the prediction accuracy scores of the support vector machine regression model and the multilayer perceptron neural network model on the validation set are calculated respectively, and the first fusion weight and the second fusion weight are dynamically allocated according to the ratio of the prediction accuracy scores.

4. The method for predicting surrounding rock convergence and predicting jamming risk of tunnel boring machines based on dynamic weighted fusion of SVR and MLP as described in claim 1, characterized in that, The final surrounding rock convergence prediction value is compared with the dynamic risk threshold. Based on the degree to which the final surrounding rock convergence prediction value exceeds the dynamic risk threshold, the risk level is divided into low risk level, medium risk level and high risk level. The low-risk level corresponds to prompting increased monitoring frequency, the medium-risk level corresponds to pushing preventative construction adjustment suggestions, and the high-risk level corresponds to automatically issuing construction parameter adjustment instructions.

5. The method for predicting surrounding rock convergence and predicting jamming risk of tunnel boring machines based on dynamic weighted fusion of SVR and MLP as described in claim 1, characterized in that, The training dataset is managed by a sliding time window. The sliding time window retains a set number of tunneling rings of construction data and measured values ​​of surrounding rock convergence as training samples in chronological order. Real-time collected construction parameters and measured values ​​of surrounding rock convergence are added to the training dataset, and historical data that exceeds the range of the sliding time window are removed. The parameters of the support vector machine regression model and the multilayer perceptron neural network model are updated using an incremental learning algorithm.

6. The method for predicting surrounding rock convergence and predicting jamming risk of tunnel boring machines based on dynamic weighted fusion of SVR and MLP as described in claim 1, characterized in that, The support vector machine regression model and the multilayer perceptron neural network model are deployed on edge computing nodes near the tunnel boring machine construction site. The edge computing nodes have data caching and breakpoint resume functions, and can independently complete model inference and risk assessment when the communication link is interrupted. After the communication link is restored, the local prediction records are automatically synchronized with the central control platform.

7. The method for predicting surrounding rock convergence and predicting jamming risk of tunnel boring machines based on dynamic weighted fusion of SVR and MLP as described in claim 1, characterized in that, The construction parameter adjustment commands include commands to adjust the advance speed, cutterhead rotation speed, and support installation rhythm; the adjusted actual construction parameters are fed back to the training dataset for subsequent model optimization training.

8. The method for predicting surrounding rock convergence and predicting jamming risk of tunnel boring machines based on dynamic weighted fusion of SVR and MLP as described in claim 1, characterized in that, The dynamic risk threshold is set by calculating the historical mean and standard deviation of the measured value of the surrounding rock convergence, and using the product of the historical mean, the safety factor, and the standard deviation as the dynamic risk threshold. The safety factor is adaptively adjusted according to the construction stage and geological conditions.

9. The method for predicting surrounding rock convergence and predicting jamming risk of tunnel boring machines based on dynamic weighted fusion of SVR and MLP as described in claim 1, characterized in that, The support vector machine regression model uses a radial basis kernel function to achieve nonlinear mapping; the optimal combination of penalty coefficient, loss insensitive interval and kernel width parameters of the support vector machine regression model is determined by a joint optimization strategy of cross-validation and hyperparameter search; the joint optimization strategy of hyperparameter search is to first use grid search to lock the approximate optimal interval of parameters in the global range, and then use a genetic algorithm to perform a fine search within the approximate optimal interval.

10. The method for predicting surrounding rock convergence and predicting jamming risk of tunnel boring machines based on dynamic weighted fusion of SVR and MLP as described in claim 1, characterized in that, The multilayer perceptron neural network model adopts a multilayer fully connected structure, with nonlinear activation functions in the hidden layers and linear activation functions in the output layers. During training, the Levenberg-Marquardt optimization algorithm is used to optimize the parameters, and an early stopping mechanism is introduced to terminate training when the validation set error does not decrease for several consecutive rounds to prevent overfitting.

Images