A tunnel surrounding rock large deformation multi-source information fusion real-time early warning method and system

Abstract

This application provides a method and system for real-time early warning of large deformation of tunnel surrounding rock through multi-source information fusion. The method includes: acquiring a multi-source monitoring dataset of large deformation of surrounding rock, performing spatiotemporal registration, data self-correction, and mechanism-driven preprocessing to obtain a standard dataset for monitoring large deformation of surrounding rock; analyzing and processing the dataset using a physical information neural operator model to obtain the deformation rate and cumulative deformation; further analyzing and processing the dataset to obtain a fused feature vector and initial cognitive uncertainty measurement data; evaluating the early warning level of large deformation of tunnel surrounding rock through a multi-expert intelligent agent; and generating an early warning interpretation report. This application achieves real-time early warning of large deformation of tunnel surrounding rock based on multi-source information fusion by constructing multi-source three-dimensional perception, sensor data self-correction, physical information neural operator combined with mechanism-driven preprocessing, causal cross-linked fusion and Bayesian uncertainty quantification analysis, and cross-geological rapid adaptation combined with multi-expert intelligent agent collaborative decision-making.

Description

Technical Field

[0001] This application relates to the field of tunnel engineering surrounding rock disaster monitoring technology, specifically to a real-time early warning method and system for large deformation of tunnel surrounding rock through multi-source information fusion. Background Technology

[0002] During tunnel construction, large deformation of the surrounding rock under adverse geological conditions such as high-stress soft rock and fault fracture zones is a key hazard leading to support failure, lining cracking, and even collapse. Currently, the monitoring and early warning of large deformation of tunnel surrounding rock relies on multi-sensor joint acquisition, data fusion modeling, and graded threshold early warning; however, existing technologies have the following four problems: First, at the data acquisition and preprocessing level: the communication protocols, sampling frequencies, and spatial deployment scales of multi-source monitoring equipment are not uniform, resulting in a spatiotemporal disconnect between advanced geological forecast data and real-time monitoring data inside the tunnel; sensors are prone to zero-point drift and signal loss in the harsh environment of tunnels over long periods of time, and there is a lack of effective self-calibration and missing data completion mechanisms; heterogeneous data have large differences in dimensions, and existing normalization methods do not combine the surrounding rock mechanics mechanism for feature decoupling, resulting in noise and effective signal superposition, leading to low model input quality. Second, at the predictive model level: traditional fusion algorithms are highly subjective and cannot adapt to the nonlinear rheological characteristics of surrounding rock, easily failing under high-conflict data; deep learning models such as CNN and LSTM have weak interpretability, poor generalization ability for small sample geological conditions, fixed fusion weights, and cannot be dynamically updated; pure data-driven models do not embed the basic principles of rock mechanics, lack physical consistency, and their prediction accuracy drops sharply when crossing geological units. Third, at the early warning decision-making level: existing technologies mostly use fixed empirical thresholds, have poor reusability across geological and tunnel types, and mostly rely solely on deformation indicators without introducing uncertainty quantification mechanisms, resulting in high false alarm and false negative rates; early warning systems cannot provide prediction confidence, making it difficult for engineering decision-makers to judge the reliability of early warning signals. Fourth, at the data causality level: mechanically piling up multi-source data and ignoring the causal relationship structure of data leads to gradient conflicts and semantic gaps; the model has poor cross-geological generalization ability, requiring the collection of a large amount of labeled data for retraining when migrating to tunnels with different geological conditions, making it difficult to meet the needs of rapid deployment of early warning systems for rapid tunnel construction. For example, CN113408190A's model construction relies entirely on the temporal correlation of monitoring data, without incorporating the mechanical mechanisms of surrounding rock and geological parameter constraints. Under complex geological conditions such as fault fracture zones and soft rocks with high ground stress, its generalization ability is weak and the prediction error increases significantly. It only outputs single-point deformation prediction values, lacks the quantification of the uncertainty of the prediction results, cannot assess the reliability of the prediction, and is difficult to support reliable early warning decisions in high-risk areas. The model has insufficient interpretability, making it difficult to clarify the causal relationship between multi-source monitoring parameters and surrounding rock deformation, and cannot provide a clear mechanistic basis for adjusting on-site construction parameters.

[0003] Effective technical solutions are urgently needed to address the above problems. Summary of the Invention

[0004] The purpose of this application is to provide a real-time early warning method and system for large deformation of tunnel surrounding rock by multi-source information fusion. It can achieve accurate early warning based on multi-source information fusion by constructing multi-source three-dimensional perception, sensor data self-correction, physical information neural operator combined with mechanism-driven preprocessing, causal cross-linked fusion and Bayesian uncertainty quantification analysis, cross-geological rapid adaptation combined with multi-expert intelligent agent collaborative decision-making.

[0005] Firstly, this application provides a real-time early warning method for large deformation of tunnel surrounding rock by fusing multi-source information, including the following steps: A multi-source monitoring dataset of large deformation of surrounding rock was acquired, and spatiotemporal registration and data self-correction preprocessing were performed to obtain a corrected dataset of large deformation monitoring of surrounding rock. Mechanism-driven preprocessing was performed on the large deformation monitoring and correction dataset of surrounding rock to obtain the standard dataset of large deformation monitoring of surrounding rock. Based on the standard dataset for monitoring large deformation of surrounding rock and the correction dataset for monitoring large deformation of surrounding rock, the physical information neural operator model based on model-independent element learning (MAML) is used for analysis and processing to obtain the deformation rate and cumulative deformation of surrounding rock over a preset time period. Based on the monitoring and correction dataset of large deformation of surrounding rock, combined with the deformation rate and cumulative deformation, causal cross-linked multimodal fusion and Bayesian uncertainty quantification analysis were performed to obtain fused feature vectors and initial cognitive uncertainty measurement data. Based on the deformation rate and cumulative deformation of the surrounding rock, combined with the fused feature vector and the initial cognitive uncertainty measurement data, the analysis and processing are carried out by a multi-expert intelligent agent to obtain the early warning level of large deformation of the tunnel surrounding rock and the corresponding early warning interpretation report. Warning information is issued based on the warning level of large deformation of the surrounding rock in the tunnel, and the warning interpretation report is pushed to the management terminal for display.

[0006] Optionally, in the real-time early warning method for large deformation of tunnel surrounding rock described in this application, the step of acquiring a multi-source monitoring dataset for large deformation of surrounding rock and performing spatiotemporal registration and data self-correction preprocessing to obtain a large deformation monitoring correction dataset for surrounding rock includes: A multi-source three-dimensional sensing network was constructed, including advanced geological exploration, deep surrounding rock monitoring, support stress monitoring, environmental hydrological monitoring, and construction parameter acquisition, and a multi-source monitoring dataset of large deformation of surrounding rock was obtained. The multi-source monitoring dataset of large deformation of surrounding rock was time-series aligned and spatially registered, and the data self-correction preprocessing was performed by a preset data correction model based on DDPG reinforcement learning to obtain the monitoring correction dataset of large deformation of surrounding rock.

[0007] Optionally, in the real-time early warning method for large deformation of tunnel surrounding rock described in this application, the step of performing mechanism-driven preprocessing based on the large deformation monitoring and correction dataset of surrounding rock to obtain a standard dataset for large deformation monitoring of surrounding rock includes: Based on the large deformation monitoring correction dataset of surrounding rock, signal denoising is performed by wavelet packet hierarchical denoising and Kalman adaptive filtering to obtain the large deformation monitoring denoised dataset of surrounding rock. Based on the denoised dataset of large deformation monitoring of surrounding rock, missing data are filled in by spatial correlation of adjacent sensors, temporal grey GM model and neighborhood spatial interpolation to obtain the complete dataset of large deformation monitoring of surrounding rock. Based on the complete dataset of large deformation monitoring of surrounding rock, feature decoupling and normalization processing is performed through a preset normalization model of surrounding rock mechanical constraints, and feature encoding processing is performed to obtain the standard dataset of large deformation monitoring of surrounding rock.

[0008] Optionally, in the real-time early warning method for large deformation of tunnel surrounding rock described in this application, the step of analyzing and processing the large deformation monitoring standard dataset and the large deformation monitoring correction dataset of surrounding rock using a physical information neural operator model based on model-independent element learning (MAML) to obtain the predicted values ​​of the surrounding rock deformation rate and cumulative deformation over a preset time period includes: Based on the monitoring and correction dataset of large deformation of surrounding rock, parameters were extracted to obtain the number of microseismic events, microseismic dominant frequency offset rate, microseismic energy release rate, geostress change, geostress inversion value, surrounding rock category label, excavation disturbance data and support reaction force data, as well as the corresponding measured value of surrounding rock deformation rate and measured value of cumulative deformation. An initial model of the physical information neural operator was constructed, and the weighted sum of the data fitting loss and the partial differential equation physical constraint loss was used as the joint training objective function. The initial model of the physical information neural operator was trained based on the number of microseismic events, the microseismic dominant frequency offset rate, the microseismic energy release rate, the change in ground stress, the ground stress inversion value, the surrounding rock category label, the excavation disturbance data and the support reaction force data, as well as the corresponding measured values ​​of the surrounding rock deformation rate and the cumulative deformation, to obtain the physical information neural operator model. The standard dataset for monitoring large deformation of surrounding rock is input into the physical information neural operator model based on model-independent meta-learning (MAML) for analysis and processing to obtain the deformation rate and cumulative deformation value of surrounding rock over a preset time period.

[0009] Optionally, in the real-time early warning method for large deformation of tunnel surrounding rock described in this application, the step of performing causal cross-linked multimodal fusion and Bayesian uncertainty quantification analysis based on the large deformation monitoring and correction dataset of surrounding rock combined with the deformation rate and cumulative deformation to obtain fused feature vectors and initial cognitive uncertainty measurement data includes: The monitoring and correction dataset of large deformation of surrounding rock is analyzed and processed through a preset dynamic Bayesian network causal graph to obtain data of the cause layer, data of the micro-damage layer, and data of the macro-response layer. Feature vectors are extracted from the data of the cause layer, the micro-damage layer, and the macro-response layer. These feature vectors are then analyzed and processed by a pre-defined causal decoupling network trained with the weighted sum of data reconstruction loss and counterfactual constraint loss as the objective function. Furthermore, cross-modal interaction is performed through a multi-head self-attention layer to obtain causal-enhanced fusion feature vectors. The causal enhancement fusion feature vector is fused with the surrounding rock deformation rate and cumulative deformation to obtain the fused feature vector; The initial cognitive uncertainty measurement data is obtained by processing the fused feature vectors using a pre-defined Bayesian uncertainty quantification analysis method.

[0010] Optionally, in the real-time early warning method for large deformation of tunnel surrounding rock described in this application, the step of analyzing and processing the deformation rate and cumulative deformation of the surrounding rock combined with the fused feature vector and initial cognitive uncertainty measurement data through a multi-expert intelligent agent to obtain the early warning level of large deformation of tunnel surrounding rock and the corresponding early warning interpretation report includes: Construct multi-expert intelligent agents, including physical expert intelligent agents, data expert intelligent agents, and engineering expert intelligent agents; Based on the surrounding rock deformation rate and cumulative deformation, combined with the fusion feature vector and initial cognitive uncertainty measurement data, as well as the historical engineering case database of large tunnel deformation, the analysis and processing are carried out by physical expert intelligent agent, data expert intelligent agent and engineering expert intelligent agent to obtain the surrounding rock deformation prediction value, cognitive uncertainty measurement data, surrounding rock deformation safety threshold and surrounding rock deformation treatment plan. Compare the predicted value of surrounding rock deformation with the safe threshold value of surrounding rock deformation; If the predicted value of surrounding rock deformation is less than the safe threshold for surrounding rock deformation, the warning level for large deformation of the surrounding rock in the tunnel is determined to be low risk. If the predicted value of surrounding rock deformation is greater than or equal to the safety threshold of surrounding rock deformation, the cognitive uncertainty measurement data will be compared with the preset cognitive uncertainty measurement benchmark threshold. If the cognitive uncertainty measurement data is less than or equal to the preset cognitive uncertainty measurement benchmark threshold, the early warning level of large deformation of the surrounding rock of the tunnel is determined to be high risk level, and a corresponding early warning interpretation report is generated based on the predicted value of the surrounding rock deformation and the surrounding rock deformation treatment plan. If the cognitive uncertainty measurement data is greater than the preset cognitive uncertainty measurement benchmark threshold, an encrypted monitoring reminder response will be output.

[0011] Secondly, this application provides a real-time early warning system for large deformation of tunnel surrounding rock through multi-source information fusion. The system includes a memory and a processor. The memory includes a program for a real-time early warning method for large deformation of tunnel surrounding rock through multi-source information fusion. When the program for the real-time early warning method for large deformation of tunnel surrounding rock through multi-source information fusion is executed by the processor, it implements the following steps: A multi-source monitoring dataset of large deformation of surrounding rock was acquired, and spatiotemporal registration and data self-correction preprocessing were performed to obtain a corrected dataset of large deformation monitoring of surrounding rock. Mechanism-driven preprocessing was performed on the large deformation monitoring and correction dataset of surrounding rock to obtain the standard dataset of large deformation monitoring of surrounding rock. Based on the standard dataset for monitoring large deformation of surrounding rock and the correction dataset for monitoring large deformation of surrounding rock, the physical information neural operator model based on model-independent element learning (MAML) is used for analysis and processing to obtain the deformation rate and cumulative deformation of surrounding rock over a preset time period. Based on the monitoring and correction dataset of large deformation of surrounding rock, combined with the deformation rate and cumulative deformation, causal cross-linked multimodal fusion and Bayesian uncertainty quantification analysis were performed to obtain fused feature vectors and initial cognitive uncertainty measurement data. Based on the deformation rate and cumulative deformation of the surrounding rock, combined with the fused feature vector and the initial cognitive uncertainty measurement data, the analysis and processing are carried out by a multi-expert intelligent agent to obtain the early warning level of large deformation of the tunnel surrounding rock and the corresponding early warning interpretation report. Warning information is issued based on the warning level of large deformation of the surrounding rock in the tunnel, and the warning interpretation report is pushed to the management terminal for display.

[0012] Optionally, in the real-time early warning system for large deformation of tunnel surrounding rock according to this application, the step of acquiring the multi-source monitoring dataset for large deformation of surrounding rock and performing spatiotemporal registration and data self-correction preprocessing to obtain the monitoring and correction dataset for large deformation of surrounding rock includes: A multi-source three-dimensional sensing network was constructed, including advanced geological exploration, deep surrounding rock monitoring, support stress monitoring, environmental hydrological monitoring, and construction parameter acquisition, and a multi-source monitoring dataset of large deformation of surrounding rock was obtained. The multi-source monitoring dataset of large deformation of surrounding rock was time-series aligned and spatially registered, and the data self-correction preprocessing was performed by a preset data correction model based on DDPG reinforcement learning to obtain the monitoring correction dataset of large deformation of surrounding rock.

[0013] Optionally, in the real-time early warning system for large deformation of tunnel surrounding rock according to the multi-source information fusion described in this application, the step of performing mechanism-driven preprocessing based on the large deformation monitoring and correction dataset of surrounding rock to obtain a standard dataset for large deformation monitoring of surrounding rock includes: Based on the large deformation monitoring correction dataset of surrounding rock, signal denoising is performed by wavelet packet hierarchical denoising and Kalman adaptive filtering to obtain the large deformation monitoring denoised dataset of surrounding rock. Based on the denoised dataset of large deformation monitoring of surrounding rock, missing data are filled in by spatial correlation of adjacent sensors, temporal grey GM model and neighborhood spatial interpolation to obtain the complete dataset of large deformation monitoring of surrounding rock. Based on the complete dataset of large deformation monitoring of surrounding rock, feature decoupling and normalization processing is performed through a preset normalization model of surrounding rock mechanical constraints, and feature encoding processing is performed to obtain the standard dataset of large deformation monitoring of surrounding rock.

[0014] Optionally, in the real-time early warning system for large deformation of tunnel surrounding rock described in this application, the step of analyzing and processing the standard dataset of large deformation monitoring of surrounding rock combined with the correction dataset of large deformation monitoring of surrounding rock through a physical information neural operator model based on model-independent element learning (MAML) to obtain the predicted values ​​of the deformation rate and cumulative deformation of surrounding rock for a preset time period includes: Based on the monitoring and correction dataset of large deformation of surrounding rock, parameters were extracted to obtain the number of microseismic events, microseismic dominant frequency offset rate, microseismic energy release rate, geostress change, geostress inversion value, surrounding rock category label, excavation disturbance data and support reaction force data, as well as the corresponding measured value of surrounding rock deformation rate and measured value of cumulative deformation. An initial model of the physical information neural operator was constructed, and the weighted sum of the data fitting loss and the partial differential equation physical constraint loss was used as the joint training objective function. The initial model of the physical information neural operator was trained based on the number of microseismic events, the microseismic dominant frequency offset rate, the microseismic energy release rate, the change in ground stress, the ground stress inversion value, the surrounding rock category label, the excavation disturbance data and the support reaction force data, as well as the corresponding measured values ​​of the surrounding rock deformation rate and the cumulative deformation, to obtain the physical information neural operator model. The standard dataset for monitoring large deformation of surrounding rock is input into the physical information neural operator model based on model-independent meta-learning (MAML) for analysis and processing to obtain the deformation rate and cumulative deformation value of surrounding rock over a preset time period.

[0015] As can be seen from the above, the real-time early warning method and system for large deformation of tunnel surrounding rock provided in this application achieves real-time early warning for large deformation of tunnel surrounding rock based on multi-source information fusion by constructing multi-source three-dimensional perception, sensor data self-correction, physical information neural operator combined mechanism driven preprocessing, causal cross-linked fusion and Bayesian uncertainty quantification analysis, cross-geological rapid adaptation combined with multi-expert intelligent agent collaborative decision-making.

[0016] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing embodiments of this application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a real-time early warning method for large deformation of tunnel surrounding rock based on multi-source information fusion, provided in this application embodiment; Figure 2 A flowchart illustrating the acquisition of a monitoring and correction dataset for large deformation of surrounding rock in a real-time early warning method based on multi-source information fusion for large deformation of surrounding rock in tunnels, provided in an embodiment of this application. Figure 3 A high-level flowchart of a real-time early warning method for large deformation of surrounding rock in tunnels, provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0020] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0021] Please refer to Figure 1 , Figure 1This is a flowchart illustrating a real-time early warning method for large deformation of tunnel surrounding rock based on multi-source information fusion, as described in some embodiments of this application. This method is used in terminal devices, such as computers and mobile phones. The method includes the following steps: S11. Obtain the multi-source monitoring dataset of large deformation of surrounding rock, and perform spatiotemporal registration and data self-correction preprocessing to obtain the monitoring correction dataset of large deformation of surrounding rock. S12. Based on the large deformation monitoring correction dataset of surrounding rock, perform mechanism-driven preprocessing to obtain the standard dataset of large deformation monitoring of surrounding rock. S13. Based on the standard dataset for monitoring large deformation of surrounding rock and the correction dataset for monitoring large deformation of surrounding rock, the data is analyzed and processed using a physical information neural operator model based on model-independent element learning (MAML) to obtain the deformation rate and cumulative deformation of surrounding rock over a preset time period. S14. Based on the monitoring and correction dataset of large deformation of surrounding rock, combined with the deformation rate and cumulative deformation, causal cross-linked multimodal fusion and Bayesian uncertainty quantification analysis were performed to obtain fused feature vectors and initial cognitive uncertainty measurement data. S15. Based on the deformation rate and cumulative deformation of the surrounding rock, combined with the fusion feature vector and the initial cognitive uncertainty measurement data, the data are analyzed and processed by a multi-expert intelligent agent to obtain the early warning level of large deformation of the tunnel surrounding rock and the corresponding early warning interpretation report. S16. Issue early warning information based on the early warning level of large deformation of the surrounding rock in the tunnel, and push the early warning interpretation report to the management terminal for display.

[0022] Further explanation is needed regarding the following steps: To achieve real-time early warning of large deformations in tunnel surrounding rock based on multi-source information fusion, the overall process involves: Step S11 collecting three-dimensional sensing data including advanced geological exploration, deep surrounding rock monitoring, support stress monitoring, environmental hydrological monitoring, and construction parameters. Data from different types of sensors are uniformly mapped onto the tunnel's longitudinal and circumferential standard spatiotemporal grids, and the sensor zero-drift and noise system errors are actively corrected based on the DDPG reinforcement learning model. Step S12 uses rock mechanics mechanisms as constraints to denoise, complete, and normalize the data, and performs feature encoding to input the data into the model. The process involves: providing standard format data; step S13 using a physical information neural operator combined with the MAML meta-learning framework to predict the deformation rate and cumulative deformation of the surrounding rock over a preset time period; step S14 establishing a causal map of the cause layer, micro-damage layer, and macro-response layer using a dynamic Bayesian network, separating causal factors and confounding factors through a causal decoupling network, and quantifying the prediction uncertainty using the Monte Carlo Dropout method; step S15 involving collaborative decision-making by a multi-expert intelligent agent, outputting the warning level based on a four-level warning threshold system; and step S16 triggering corresponding audible and visual alarms, pushing structured warning reports, and archiving historical cases.

[0023] Please refer to Figure 2 , Figure 2 This is a flowchart illustrating the process of obtaining a large-deformation monitoring and correction dataset for tunnel surrounding rock using a multi-source information fusion real-time early warning method according to some embodiments of this application. According to embodiments of the present invention, obtaining the multi-source monitoring dataset for large-deformation of surrounding rock, and performing spatiotemporal registration and data self-correction preprocessing to obtain the large-deformation monitoring and correction dataset includes: S21. Construct a multi-source three-dimensional sensing network that includes advanced geological exploration, deep surrounding rock monitoring, support stress monitoring, environmental hydrological monitoring, and construction parameter acquisition, and obtain a multi-source monitoring dataset of large deformation of surrounding rock. S22. Perform temporal alignment and spatial registration on the multi-source monitoring dataset of large deformation of surrounding rock, and perform data self-correction preprocessing through a preset data correction model based on DDPG reinforcement learning to obtain the monitoring correction dataset of large deformation of surrounding rock.

[0024] Further explanation is needed regarding the construction and data acquisition methods of the multi-source three-dimensional sensing network: Advanced geological exploration is conducted using the TSP seismic wave method and ground-penetrating radar to acquire data on the causal layers of stress changes and groundwater seepage conditions; deep surrounding rock monitoring is performed using FBG fiber optic grating sensors, BOTDA distributed fiber optic sensors, multi-point displacement gauges, and three-dimensional microseismic sensors to acquire data on the microscopic damage layer and macroscopic response layer, including deep strain, layered displacement, number of microseismic events, and energy release rate; support stress monitoring is performed using vibrating wire stress gauges and steel arch strain sensors to acquire supplementary macroscopic response layer data on anchor bolt axial force and initial support stress state; environmental hydrological monitoring is performed using pore water pressure gauges, temperature and humidity sensors, and blasting vibration sensors to acquire relevant data such as pore water pressure, surrounding rock temperature and humidity, and construction disturbance vibration velocity; the construction parameter acquisition module acquires data on excavation advance, support timing, and cycle advance. The specific data parameters mentioned above are adjusted and supplemented by those skilled in the art based on the specific conditions of tunnel construction.

[0025] In terms of spatiotemporal registration, a standard spatiotemporal grid is established along the longitudinal (mileage direction) and circumferential (section direction) axes of the tunnel with the tunnel entrance as the origin. All measuring points are mapped to unified grid nodes according to their coordinate positions, and the spatiotemporal alignment of the collected data is automatically performed using a unified sampling frequency and timestamp. In terms of data self-correction, after each tunneling cycle or when there is a sudden change in the environment, a mobile standard reference signal generator installed on the intelligent inspection device is used to periodically inject standard signals with known amplitudes into each sensor, record the deviation between the sensor output and the true value, and construct a DDPG (Deep Deterministic Policy Gradient) reinforcement learning correction model. In this model, the state space consists of the sensor's historical deviation feature vector, environmental parameters (temperature, humidity), and timestamp encoding; the action space consists of the Kalman filter process noise covariance adjustment coefficient and the sampling frequency adjustment amount; and the reward function is to minimize the root mean square error of the deviation between the corrected data and the reference true value. After training until the reward convergence, the model parameters are frozen for online self-correction. When the data missing rate of a certain sensor exceeds 5%, a spatiotemporal interpolation network is activated to complete the data of that sensor, ensuring data integrity.

[0026] According to an embodiment of the present invention, the step of performing mechanism-driven preprocessing based on the large deformation monitoring correction dataset of surrounding rock to obtain a standard dataset for large deformation monitoring of surrounding rock includes: Based on the large deformation monitoring correction dataset of surrounding rock, signal denoising is performed by wavelet packet hierarchical denoising and Kalman adaptive filtering to obtain the large deformation monitoring denoised dataset of surrounding rock. Based on the denoised dataset of large deformation monitoring of surrounding rock, missing data are filled in by spatial correlation of adjacent sensors, temporal grey GM model and neighborhood spatial interpolation to obtain the complete dataset of large deformation monitoring of surrounding rock. Based on the complete dataset of large deformation monitoring of surrounding rock, feature decoupling and normalization processing is performed through a preset normalization model of surrounding rock mechanical constraints, and feature encoding processing is performed to obtain the standard dataset of large deformation monitoring of surrounding rock.

[0027] Further explanation is needed regarding the specific implementation of signal denoising: Wavelet packet layered denoising decomposes the signal into different frequency band components through multi-layer decomposition of wavelet packets (4-6 layers), retaining the low-frequency components (cutoff frequency below 0.1Hz) containing the effective low-frequency trend of surrounding rock rheology, removing high-frequency noise caused by blasting disturbance (100-500Hz) and mechanical vibration (10-50Hz), and then reconstructing the denoised signal through inverse transformation; Kalman adaptive filtering establishes the deformation state equation of the tunnel surrounding rock, with the state variables being deformation displacement u and deformation rate v, and the state transition matrix is ​​established based on the assumption of uniform deformation. The optimal state is estimated through measurement update and prediction update cycles, and the process noise covariance Q is dynamically adjusted by the action output of the DDPG correction model to achieve the optimal linear estimation of the deformation signal; Three complementary strategies are adopted for missing data completion: (1) Spatial correlation completion of adjacent sensors, which is suitable for the situation where some data of multiple measuring points on the same monitoring section are missing, utilizing the spatial correlation of other sensors on the same section. (1) Relational regression estimation of missing values; (2) Time series gray GM(1,1) model completion, which is suitable for the case of missing data at several consecutive time points in the time series (no more than 10 sampling points are missing), and the missing data is extrapolated using the exponential prediction characteristics of the GM(1,1) model; (3) Neighborhood spatial interpolation completion, which is suitable for the case of sparse spatial data between adjacent monitoring sections, and spatial estimation is performed using the Kriging interpolation method; In terms of normalization and feature encoding, the Hawke-Brown strength criterion is used as the theoretical framework, and combined with the strength-stress ratio characteristics, the deformation rate, stress concentration coefficient and seepage coupling coefficient are decoupled and normalized to construct a preset surrounding rock mechanical constraint type normalization model, which eliminates the difference in dimensions while retaining the rock mass mechanical evolution characteristics (distinguishing the characteristics of elastic, plastic and fracture stages); Finally, one-hot encoding is used for discrete category features (surrounding rock category, etc.), and the above mechanical constraint type normalization is used for continuous numerical features to obtain the standard dataset for monitoring large deformation of surrounding rock, which is used as the input of the physical information neural operator.

[0028] According to an embodiment of the present invention, the step of analyzing and processing the surrounding rock deformation rate and cumulative deformation prediction values ​​for a preset time period using a physical information neural operator model based on Model-Independent Element Learning (MAML) in conjunction with the surrounding rock large deformation monitoring standard dataset and the surrounding rock large deformation monitoring correction dataset includes: Based on the monitoring and correction dataset of large deformation of surrounding rock, parameters were extracted to obtain the number of microseismic events, microseismic dominant frequency offset rate, microseismic energy release rate, geostress change, geostress inversion value, surrounding rock category label, excavation disturbance data and support reaction force data, as well as the corresponding measured value of surrounding rock deformation rate and measured value of cumulative deformation. An initial model of the physical information neural operator was constructed, and the weighted sum of the data fitting loss and the partial differential equation physical constraint loss was used as the joint training objective function. The initial model of the physical information neural operator was trained based on the number of microseismic events, the microseismic dominant frequency offset rate, the microseismic energy release rate, the change in ground stress, the ground stress inversion value, the surrounding rock category label, the excavation disturbance data and the support reaction force data, as well as the corresponding measured values ​​of the surrounding rock deformation rate and the cumulative deformation, to obtain the physical information neural operator model. The standard dataset for monitoring large deformation of surrounding rock is input into the physical information neural operator model based on model-independent meta-learning (MAML) for analysis and processing to obtain the deformation rate and cumulative deformation value of surrounding rock over a preset time period.

[0029] Further explanation is needed regarding the following: the number of microseismic events originates from the three-dimensional microseismic sensor at the surrounding rock end, representing the frequency of microscopic fractures in the rock mass and serving as an indicator of damage accumulation. The microseismic dominant frequency offset rate also originates from the surrounding rock end microseismic sensor, representing the degree of microscopic damage to the rock mass; for example, a shift in the dominant frequency towards lower frequencies indicates a deterioration in the overall integrity of the rock mass. The microseismic energy release rate, also originating from the surrounding rock end microseismic sensor, represents the intensity of microscopic fractures in the rock mass and is a key input for predicting macroscopic deformation. The change in geostress originates from advanced geological exploration, representing the initial and evolving stress state of the rock mass and serving as the core input for the diffusion coefficient D(x,t). The geostress inversion value originates from advanced geological exploration and is used by the physical expert agent to calculate the interaction between the surrounding rock and the support. The surrounding rock category label originates from geological exploration data. The classification, input in one-hot encoded form, is used to distinguish the characteristics of different geological units. Excavation disturbance data is collected from the construction end and is used to represent the degree of disturbance to the surrounding rock during construction, serving as a component of the source term f(x,t). Support reaction force data is collected from the support end monitoring and is used to represent the constraint effect of the support structure on the surrounding rock, serving as another component of the source term f(x,t). The initial model of the physical information neural operator consists of a feature encoder (extracting high-dimensional feature vectors from preprocessed data), a diffusion coefficient network (a small-scale MLP, inputting microseismic energy release rate, inverted ground stress value, and surrounding rock category encoding, and outputting the spatiotemporal diffusion coefficient D(x,t)), and a PDE solution module (an operator layer based on the differentiable finite volume method). The governing equation is: Where u(x,t) is the surrounding rock deformation field (crown settlement or perimeter convergence value), D(x,t) is the spatiotemporal diffusion coefficient, f(x,t) is the excavation disturbance and support reaction source term, and ξ(x,t) is the preset error term; the total loss function of the training process is: L total =L data +λ PDE ×L PDE +λ reg ×L reg L data L represents the mean square error loss between the predicted and measured values. PDEL is the sum of squares of the PDE residuals at each grid point (physical constraint term). reg λ is the regularization term for the diffusion coefficient network weights. PDE The value range is 0.1~0.5 (determined through cross-validation). The Adam optimizer is used for training with a learning rate of 0.001 and 200 training rounds, so that the prediction results under any geological conditions can meet the basic laws of rock mechanics.

[0030] According to an embodiment of the present invention, the step of performing causal cross-linked multimodal fusion and Bayesian uncertainty quantification analysis based on the surrounding rock large deformation monitoring and correction dataset, combined with the surrounding rock deformation rate and cumulative deformation, to obtain fused feature vectors and initial cognitive uncertainty measurement data, includes: The monitoring and correction dataset of large deformation of surrounding rock is analyzed and processed through a preset dynamic Bayesian network causal graph to obtain data of the cause layer, data of the micro-damage layer, and data of the macro-response layer. Feature vectors are extracted from the data of the cause layer, the micro-damage layer, and the macro-response layer. These feature vectors are then analyzed and processed by a pre-defined causal decoupling network trained with the weighted sum of data reconstruction loss and counterfactual constraint loss as the objective function. Furthermore, cross-modal interaction is performed through a multi-head self-attention layer to obtain causal-enhanced fusion feature vectors. The causal enhancement fusion feature vector is fused with the surrounding rock deformation rate and cumulative deformation to obtain the fused feature vector; The initial cognitive uncertainty measurement data is obtained by processing the fused feature vectors using a pre-defined Bayesian uncertainty quantification analysis method.

[0031] Further explanation is needed regarding the pre-defined dynamic Bayesian network causal graph, which includes a causal layer, a micro-damage layer, and a macro-response layer. The causal layer (changes in geostress, groundwater seepage conditions) represents the fundamental driving force behind large deformations of the surrounding rock. The micro-damage layer (number of microseismic events, microseismic energy release rate, microseismic frequency shift rate) represents the micro-fracture evolution process of the rock mass. The macro-response layer (cumulative settlement of the arch, peripheral convergence rate, changes in axial force of the support structure) represents the observable results of the macro-deformation behavior of the surrounding rock. The causal relationship is that changes in geostress cause the accumulation of micro-fractures in the rock mass, which in turn leads to a macro-deformation response. The feature vector extraction and causal decoupling network uses independent one-dimensional CNN encoders (three convolutional layers with 16, 32, and 64 channels) to extract feature vectors f for the causal layer, micro-damage layer, and macro-response layer, respectively. cause f micro and f macro The concatenated features (192-dimensional) are input into a causal decoupling MLP (with a structure of 192, 128, and 64), and the output is a 32-dimensional causal factor z. causal (Eigencomponents with a real causal relationship to surrounding rock deformation) and 32-dimensional confusion factor zconfound (Spurious correlation characteristic components introduced by environmental interference and sensor noise), for z confound Apply a learnable adaptive additive noise perturbation, z causal This remains unchanged, enhancing robustness to changes in data distribution. The multi-head self-attention layer will de-perturb the z... causal Through a multi-head self-attention layer (head number = 4, hidden dimension 256), global interaction modeling is performed on causal factors from different data modalities. The long-range dependencies between the cause layer, micro-damage layer, and macro-response layer are captured for cross-modal interaction, and a causal enhancement fusion feature vector is output. The deformed prediction values ​​(deformation rate and cumulative deformation amount) are concatenated and spliced ​​to form the final fusion feature vector. Dropout activation (typical dropout rate 0.2) is maintained in both the training and inference stages of the physical information neural operator network. The same input is forward propagated T times (T≥50, 100 times in this embodiment) to statistically analyze the mean and variance of the prediction sample set. It is decomposed into initial cognitive uncertainty measurement data, which includes cognitive uncertainty measurement data (the model's cognitive boundary of the current data) and accidental uncertainty measurement data (the inherent noise level of the data). The cognitive uncertainty measurement data serves as the credibility basis for subsequent early warning decisions.

[0032] According to an embodiment of the present invention, the step of analyzing and processing the surrounding rock deformation rate and cumulative deformation amount by combining the fused feature vector and initial cognitive uncertainty measurement data through a multi-expert intelligent agent to obtain the early warning level of large deformation of the tunnel surrounding rock and the corresponding early warning interpretation report includes: Construct multi-expert intelligent agents, including physical expert intelligent agents, data expert intelligent agents, and engineering expert intelligent agents; Based on the surrounding rock deformation rate and cumulative deformation, combined with the fusion feature vector and initial cognitive uncertainty measurement data, as well as the historical engineering case database of large tunnel deformation, the analysis and processing are carried out by physical expert intelligent agent, data expert intelligent agent and engineering expert intelligent agent to obtain the surrounding rock deformation prediction value, cognitive uncertainty measurement data, surrounding rock deformation safety threshold and surrounding rock deformation treatment plan. Compare the predicted value of surrounding rock deformation with the safe threshold value of surrounding rock deformation; If the predicted value of surrounding rock deformation is less than the safe threshold for surrounding rock deformation, the warning level for large deformation of the surrounding rock in the tunnel is determined to be low risk. If the predicted value of surrounding rock deformation is greater than or equal to the safety threshold of surrounding rock deformation, the cognitive uncertainty measurement data will be compared with the preset cognitive uncertainty measurement benchmark threshold. If the cognitive uncertainty measurement data is less than or equal to the preset cognitive uncertainty measurement benchmark threshold, the early warning level of large deformation of the surrounding rock of the tunnel is determined to be high risk level, and a corresponding early warning interpretation report is generated based on the predicted value of the surrounding rock deformation and the surrounding rock deformation treatment plan. If the cognitive uncertainty measurement data is greater than the preset cognitive uncertainty measurement benchmark threshold, an encrypted monitoring reminder response will be output.

[0033] Further explanation is needed regarding the following: The physical expert agent, based on the characteristic curve method or convergence constraint method, takes into account the current ground stress, preset surrounding rock mechanical parameters (elastic modulus, Poisson's ratio, expansion coefficient, etc., indirectly obtained from surrounding rock type and geological survey), and support stiffness (indirectly obtained from support end monitoring in step 1). It calculates the interaction curve between the surrounding rock and the support, and outputs the surrounding rock deformation safety threshold (i.e., the ultimate deformation rate or cumulative deformation). The data expert agent uses a physical information neural operator model adapted from historical tunnel datasets through physical information element learning (MAML) to output the predicted value of surrounding rock deformation and cognitive uncertainty. The engineering expert agent constructs a database of historical engineering cases of large tunnel deformation and uses cosine similarity to search for similar cases in the database, outputting the surrounding rock deformation treatment plan for the most similar case. Finally, the warning level is determined by comparing the threshold of the predicted value of surrounding rock deformation and the cognitive uncertainty measurement data. Low-risk levels can be monitored normally, while high-risk levels require triggering a warning. If the cognitive uncertainty measurement data is greater than the preset cognitive uncertainty measurement benchmark threshold, the output information is insufficient, and it is recommended to encrypt the monitoring reminder and temporarily not trigger the warning to prevent false alarms.

[0034] It is worth mentioning that, according to an embodiment of the present invention, the step of processing the fused feature vector using a preset Bayesian uncertainty quantification analysis method to obtain initial cognitive uncertainty measurement data includes: The fused feature vectors are sampled T times using the Monte Carlo random inactivation (MC Dropout) method; The mean and variance of the T sampling results are calculated to obtain the initial cognitive uncertainty measurement data.

[0035] Further explanation is needed. For the same input, forward propagation is performed T times (T≥50, 100 times in this embodiment) to statistically predict the mean and variance of the sample set. This is decomposed into initial cognitive uncertainty measurement data, which includes cognitive uncertainty measurement data (the model's cognitive boundary of the current data) and random uncertainty measurement data (the inherent noise level of the data).

[0036] Please refer to Figure 3 , Figure 3 This is a high-level flowchart of a real-time early warning method for large deformation of surrounding rock in tunnels, which is based on multi-source information fusion, according to some embodiments of this application.

[0037] This invention also discloses a real-time early warning system for large deformation of tunnel surrounding rock through multi-source information fusion, comprising a memory and a processor. The memory includes a program for a real-time early warning method for large deformation of tunnel surrounding rock through multi-source information fusion. When the processor executes the program for a real-time early warning method for large deformation of tunnel surrounding rock through multi-source information fusion, the following steps are implemented: A multi-source monitoring dataset of large deformation of surrounding rock was acquired, and spatiotemporal registration and data self-correction preprocessing were performed to obtain a corrected dataset of large deformation monitoring of surrounding rock. Mechanism-driven preprocessing was performed on the large deformation monitoring and correction dataset of surrounding rock to obtain the standard dataset of large deformation monitoring of surrounding rock. Based on the standard dataset for monitoring large deformation of surrounding rock and the correction dataset for monitoring large deformation of surrounding rock, the physical information neural operator model based on model-independent element learning (MAML) is used for analysis and processing to obtain the deformation rate and cumulative deformation of surrounding rock over a preset time period. Based on the monitoring and correction dataset of large deformation of surrounding rock, combined with the deformation rate and cumulative deformation, causal cross-linked multimodal fusion and Bayesian uncertainty quantification analysis were performed to obtain fused feature vectors and initial cognitive uncertainty measurement data. Based on the deformation rate and cumulative deformation of the surrounding rock, combined with the fused feature vector and the initial cognitive uncertainty measurement data, the analysis and processing are carried out by a multi-expert intelligent agent to obtain the early warning level of large deformation of the tunnel surrounding rock and the corresponding early warning interpretation report. Warning information is issued based on the warning level of large deformation of the surrounding rock in the tunnel, and the warning interpretation report is pushed to the management terminal for display.

[0038] Further explanation is needed regarding the following steps: To achieve real-time early warning of large deformations in tunnel surrounding rock based on multi-source information fusion, the overall process involves: Step S11 collecting three-dimensional sensing data including advanced geological exploration, deep surrounding rock monitoring, support stress monitoring, environmental hydrological monitoring, and construction parameters. Data from different types of sensors are uniformly mapped onto the tunnel's longitudinal and circumferential standard spatiotemporal grids, and the sensor zero-drift and noise system errors are actively corrected based on the DDPG reinforcement learning model. Step S12 uses rock mechanics mechanisms as constraints to denoise, complete, and normalize the data, and performs feature encoding to input the data into the model. The process involves: providing standard format data; step S13 using a physical information neural operator combined with the MAML meta-learning framework to predict the deformation rate and cumulative deformation of the surrounding rock over a preset time period; step S14 establishing a causal map of the cause layer, micro-damage layer, and macro-response layer using a dynamic Bayesian network, separating causal factors and confounding factors through a causal decoupling network, and quantifying the prediction uncertainty using the Monte Carlo Dropout method; step S15 involving collaborative decision-making by a multi-expert intelligent agent, outputting the warning level based on a four-level warning threshold system; and step S16 triggering corresponding audible and visual alarms, pushing structured warning reports, and archiving historical cases.

[0039] According to an embodiment of the present invention, the step of acquiring a multi-source monitoring dataset for large deformation of surrounding rock and performing spatiotemporal registration and data self-correction preprocessing to obtain a corrected monitoring dataset for large deformation of surrounding rock includes: A multi-source three-dimensional sensing network was constructed, including advanced geological exploration, deep surrounding rock monitoring, support stress monitoring, environmental hydrological monitoring, and construction parameter acquisition, and a multi-source monitoring dataset of large deformation of surrounding rock was obtained. The multi-source monitoring dataset of large deformation of surrounding rock was time-series aligned and spatially registered, and the data self-correction preprocessing was performed by a preset data correction model based on DDPG reinforcement learning to obtain the monitoring correction dataset of large deformation of surrounding rock.

[0040] Further explanation is needed regarding the construction and data acquisition methods of the multi-source three-dimensional sensing network: Advanced geological exploration is conducted using the TSP seismic wave method and ground-penetrating radar to acquire data on the causal layers of stress changes and groundwater seepage conditions; deep surrounding rock monitoring is performed using FBG fiber optic grating sensors, BOTDA distributed fiber optic sensors, multi-point displacement gauges, and three-dimensional microseismic sensors to acquire data on the microscopic damage layer and macroscopic response layer, including deep strain, layered displacement, number of microseismic events, and energy release rate; support stress monitoring is performed using vibrating wire stress gauges and steel arch strain sensors to acquire supplementary macroscopic response layer data on anchor bolt axial force and initial support stress state; environmental hydrological monitoring is performed using pore water pressure gauges, temperature and humidity sensors, and blasting vibration sensors to acquire relevant data such as pore water pressure, surrounding rock temperature and humidity, and construction disturbance vibration velocity; the construction parameter acquisition module acquires data on excavation advance, support timing, and cycle advance. The specific data parameters mentioned above are adjusted and supplemented by those skilled in the art based on the specific conditions of tunnel construction.

[0041] In terms of spatiotemporal registration, a standard spatiotemporal grid is established along the longitudinal (mileage direction) and circumferential (section direction) axes of the tunnel with the tunnel entrance as the origin. All measuring points are mapped to unified grid nodes according to their coordinate positions, and the spatiotemporal alignment of the collected data is automatically performed using a unified sampling frequency and timestamp. In terms of data self-correction, after each tunneling cycle or when there is a sudden change in the environment, a mobile standard reference signal generator installed on the intelligent inspection device is used to periodically inject standard signals with known amplitudes into each sensor, record the deviation between the sensor output and the true value, and construct a DDPG (Deep Deterministic Policy Gradient) reinforcement learning correction model. In this model, the state space consists of the sensor's historical deviation feature vector, environmental parameters (temperature, humidity), and timestamp encoding; the action space consists of the Kalman filter process noise covariance adjustment coefficient and the sampling frequency adjustment amount; and the reward function is to minimize the root mean square error of the deviation between the corrected data and the reference true value. After training until the reward convergence, the model parameters are frozen for online self-correction. When the data missing rate of a certain sensor exceeds 5%, a spatiotemporal interpolation network is activated to complete the data of that sensor, ensuring data integrity.

[0042] According to an embodiment of the present invention, the step of performing mechanism-driven preprocessing based on the large deformation monitoring correction dataset of surrounding rock to obtain a standard dataset for large deformation monitoring of surrounding rock includes: Based on the large deformation monitoring correction dataset of surrounding rock, signal denoising is performed by wavelet packet hierarchical denoising and Kalman adaptive filtering to obtain the large deformation monitoring denoised dataset of surrounding rock. Based on the denoised dataset of large deformation monitoring of surrounding rock, missing data are filled in by spatial correlation of adjacent sensors, temporal grey GM model and neighborhood spatial interpolation to obtain the complete dataset of large deformation monitoring of surrounding rock. Based on the complete dataset of large deformation monitoring of surrounding rock, feature decoupling and normalization processing is performed through a preset normalization model of surrounding rock mechanical constraints, and feature encoding processing is performed to obtain the standard dataset of large deformation monitoring of surrounding rock.

[0043] Further explanation is needed regarding the specific implementation of signal denoising: Wavelet packet layered denoising decomposes the signal into different frequency band components through multi-layer decomposition of wavelet packets (4-6 layers), retaining the low-frequency components (cutoff frequency below 0.1Hz) containing the effective low-frequency trend of surrounding rock rheology, removing high-frequency noise caused by blasting disturbance (100-500Hz) and mechanical vibration (10-50Hz), and then reconstructing the denoised signal through inverse transformation; Kalman adaptive filtering establishes the deformation state equation of the tunnel surrounding rock, with the state variables being deformation displacement u and deformation rate v, and the state transition matrix is ​​established based on the assumption of uniform deformation. The optimal state is estimated through measurement update and prediction update cycles, and the process noise covariance Q is dynamically adjusted by the action output of the DDPG correction model to achieve the optimal linear estimation of the deformation signal; Three complementary strategies are adopted for missing data completion: (1) Spatial correlation completion of adjacent sensors, which is suitable for the situation where some data of multiple measuring points on the same monitoring section are missing, utilizing the spatial correlation of other sensors on the same section. (1) Relational regression estimation of missing values; (2) Time series gray GM(1,1) model completion, which is suitable for the case of missing data at several consecutive time points in the time series (no more than 10 sampling points are missing), and the missing data is extrapolated using the exponential prediction characteristics of the GM(1,1) model; (3) Neighborhood spatial interpolation completion, which is suitable for the case of sparse spatial data between adjacent monitoring sections, and spatial estimation is performed using the Kriging interpolation method; In terms of normalization and feature encoding, the Hawke-Brown strength criterion is used as the theoretical framework, and combined with the strength-stress ratio characteristics, the deformation rate, stress concentration coefficient and seepage coupling coefficient are decoupled and normalized to construct a preset surrounding rock mechanical constraint type normalization model, which eliminates the difference in dimensions while retaining the rock mass mechanical evolution characteristics (distinguishing the characteristics of elastic, plastic and fracture stages); Finally, one-hot encoding is used for discrete category features (surrounding rock category, etc.), and the above mechanical constraint type normalization is used for continuous numerical features to obtain the standard dataset for monitoring large deformation of surrounding rock, which is used as the input of the physical information neural operator.

[0044] According to an embodiment of the present invention, the step of analyzing and processing the surrounding rock deformation rate and cumulative deformation prediction values ​​for a preset time period using a physical information neural operator model based on Model-Independent Element Learning (MAML) in conjunction with the surrounding rock large deformation monitoring standard dataset and the surrounding rock large deformation monitoring correction dataset includes: Based on the monitoring and correction dataset of large deformation of surrounding rock, parameters were extracted to obtain the number of microseismic events, microseismic dominant frequency offset rate, microseismic energy release rate, geostress change, geostress inversion value, surrounding rock category label, excavation disturbance data and support reaction force data, as well as the corresponding measured value of surrounding rock deformation rate and measured value of cumulative deformation. An initial model of the physical information neural operator was constructed, and the weighted sum of the data fitting loss and the partial differential equation physical constraint loss was used as the joint training objective function. The initial model of the physical information neural operator was trained based on the number of microseismic events, the microseismic dominant frequency offset rate, the microseismic energy release rate, the change in ground stress, the ground stress inversion value, the surrounding rock category label, the excavation disturbance data and the support reaction force data, as well as the corresponding measured values ​​of the surrounding rock deformation rate and the cumulative deformation, to obtain the physical information neural operator model. The standard dataset for monitoring large deformation of surrounding rock is input into the physical information neural operator model based on model-independent meta-learning (MAML) for analysis and processing to obtain the deformation rate and cumulative deformation value of surrounding rock over a preset time period.

[0045] Further explanation is needed regarding the following: the number of microseismic events originates from the three-dimensional microseismic sensor at the surrounding rock end, representing the frequency of microscopic fractures in the rock mass and serving as an indicator of damage accumulation. The microseismic dominant frequency offset rate also originates from the surrounding rock end microseismic sensor, representing the degree of microscopic damage to the rock mass; for example, a shift in the dominant frequency towards lower frequencies indicates a deterioration in the overall integrity of the rock mass. The microseismic energy release rate, also originating from the surrounding rock end microseismic sensor, represents the intensity of microscopic fractures in the rock mass and is a key input for predicting macroscopic deformation. The change in geostress originates from advanced geological exploration, representing the initial and evolving stress state of the rock mass and serving as the core input for the diffusion coefficient D(x,t). The geostress inversion value originates from advanced geological exploration and is used by the physical expert agent to calculate the interaction between the surrounding rock and the support. The surrounding rock category label originates from geological exploration data. The classification, input in one-hot encoded form, is used to distinguish the characteristics of different geological units. Excavation disturbance data is collected from the construction end and is used to represent the degree of disturbance to the surrounding rock during construction, serving as a component of the source term f(x,t). Support reaction force data is collected from the support end monitoring and is used to represent the constraint effect of the support structure on the surrounding rock, serving as another component of the source term f(x,t). The initial model of the physical information neural operator consists of a feature encoder (extracting high-dimensional feature vectors from preprocessed data), a diffusion coefficient network (a small-scale MLP, inputting microseismic energy release rate, inverted ground stress value, and surrounding rock category encoding, and outputting the spatiotemporal diffusion coefficient D(x,t)), and a PDE solution module (an operator layer based on the differentiable finite volume method). The governing equation is: Where u(x,t) is the surrounding rock deformation field (crown settlement or perimeter convergence value), D(x,t) is the spatiotemporal diffusion coefficient, f(x,t) is the excavation disturbance and support reaction source term, and ξ(x,t) is the preset error term; the total loss function of the training process is: L total =L data +λ PDE ×L PDE +λ reg ×L reg L data L represents the mean square error loss between the predicted and measured values. PDE L is the sum of squares of the PDE residuals at each grid point (physical constraint term). reg λ is the regularization term for the diffusion coefficient network weights. PDE The value range is 0.1~0.5 (determined through cross-validation). The Adam optimizer is used for training with a learning rate of 0.001 and 200 training rounds, so that the prediction results under any geological conditions can meet the basic laws of rock mechanics.

[0046] According to an embodiment of the present invention, the step of performing causal cross-linked multimodal fusion and Bayesian uncertainty quantification analysis based on the surrounding rock large deformation monitoring and correction dataset, combined with the surrounding rock deformation rate and cumulative deformation, to obtain fused feature vectors and initial cognitive uncertainty measurement data, includes: The monitoring and correction dataset of large deformation of surrounding rock is analyzed and processed through a preset dynamic Bayesian network causal graph to obtain data of the cause layer, data of the micro-damage layer, and data of the macro-response layer. Feature vectors are extracted from the data of the cause layer, the micro-damage layer, and the macro-response layer. These feature vectors are then analyzed and processed by a pre-defined causal decoupling network trained with the weighted sum of data reconstruction loss and counterfactual constraint loss as the objective function. Furthermore, cross-modal interaction is performed through a multi-head self-attention layer to obtain causal-enhanced fusion feature vectors. The causal enhancement fusion feature vector is fused with the surrounding rock deformation rate and cumulative deformation to obtain the fused feature vector; The initial cognitive uncertainty measurement data is obtained by processing the fused feature vectors using a pre-defined Bayesian uncertainty quantification analysis method.

[0047] Further explanation is needed regarding the pre-defined dynamic Bayesian network causal graph, which includes a causal layer, a micro-damage layer, and a macro-response layer. The causal layer (changes in geostress, groundwater seepage conditions) represents the fundamental driving force behind large deformations of the surrounding rock. The micro-damage layer (number of microseismic events, microseismic energy release rate, microseismic frequency shift rate) represents the micro-fracture evolution process of the rock mass. The macro-response layer (cumulative settlement of the arch, peripheral convergence rate, changes in axial force of the support structure) represents the observable results of the macro-deformation behavior of the surrounding rock. The causal relationship is that changes in geostress cause the accumulation of micro-fractures in the rock mass, which in turn leads to a macro-deformation response. The feature vector extraction and causal decoupling network uses independent one-dimensional CNN encoders (three convolutional layers with 16, 32, and 64 channels) to extract feature vectors f for the causal layer, micro-damage layer, and macro-response layer, respectively. cause f micro and f macro The concatenated features (192-dimensional) are input into a causal decoupling MLP (with a structure of 192, 128, and 64), and the output is a 32-dimensional causal factor z. causal (Eigencomponents with a real causal relationship to surrounding rock deformation) and 32-dimensional confusion factor z confound (Spurious correlation characteristic components introduced by environmental interference and sensor noise), for z confound Apply a learnable adaptive additive noise perturbation, z causal This remains unchanged, enhancing robustness to changes in data distribution. The multi-head self-attention layer will de-perturb the z... causal Through a multi-head self-attention layer (head number = 4, hidden dimension 256), global interaction modeling is performed on causal factors from different data modalities. The long-range dependencies between the cause layer, micro-damage layer, and macro-response layer are captured for cross-modal interaction, and a causal enhancement fusion feature vector is output. The deformed prediction values ​​(deformation rate and cumulative deformation amount) are concatenated and spliced ​​to form the final fusion feature vector. Dropout activation (typical dropout rate 0.2) is maintained in both the training and inference stages of the physical information neural operator network. The same input is forward propagated T times (T≥50, 100 times in this embodiment) to statistically analyze the mean and variance of the prediction sample set. It is decomposed into initial cognitive uncertainty measurement data, which includes cognitive uncertainty measurement data (the model's cognitive boundary of the current data) and accidental uncertainty measurement data (the inherent noise level of the data). The cognitive uncertainty measurement data serves as the credibility basis for subsequent early warning decisions.

[0048] According to an embodiment of the present invention, the step of analyzing and processing the surrounding rock deformation rate and cumulative deformation amount by combining the fused feature vector and initial cognitive uncertainty measurement data through a multi-expert intelligent agent to obtain the early warning level of large deformation of the tunnel surrounding rock and the corresponding early warning interpretation report includes: Construct multi-expert intelligent agents, including physical expert intelligent agents, data expert intelligent agents, and engineering expert intelligent agents; Based on the surrounding rock deformation rate and cumulative deformation, combined with the fusion feature vector and initial cognitive uncertainty measurement data, as well as the historical engineering case database of large tunnel deformation, the analysis and processing are carried out by physical expert intelligent agent, data expert intelligent agent and engineering expert intelligent agent to obtain the surrounding rock deformation prediction value, cognitive uncertainty measurement data, surrounding rock deformation safety threshold and surrounding rock deformation treatment plan. Compare the predicted value of surrounding rock deformation with the safe threshold value of surrounding rock deformation; If the predicted value of surrounding rock deformation is less than the safe threshold for surrounding rock deformation, the warning level for large deformation of the surrounding rock in the tunnel is determined to be low risk. If the predicted value of surrounding rock deformation is greater than or equal to the safety threshold of surrounding rock deformation, the cognitive uncertainty measurement data will be compared with the preset cognitive uncertainty measurement benchmark threshold. If the cognitive uncertainty measurement data is less than or equal to the preset cognitive uncertainty measurement benchmark threshold, the early warning level of large deformation of the surrounding rock of the tunnel is determined to be high risk level, and a corresponding early warning interpretation report is generated based on the predicted value of the surrounding rock deformation and the surrounding rock deformation treatment plan. If the cognitive uncertainty measurement data is greater than the preset cognitive uncertainty measurement benchmark threshold, an encrypted monitoring reminder response will be output.

[0049] Further explanation is needed regarding the following: The physical expert agent, based on the characteristic curve method or convergence constraint method, takes into account the current ground stress, preset surrounding rock mechanical parameters (elastic modulus, Poisson's ratio, expansion coefficient, etc., indirectly obtained from surrounding rock type and geological survey), and support stiffness (indirectly obtained from support end monitoring in step 1). It calculates the interaction curve between the surrounding rock and the support, and outputs the surrounding rock deformation safety threshold (i.e., the ultimate deformation rate or cumulative deformation). The data expert agent uses a physical information neural operator model adapted from historical tunnel datasets through physical information element learning (MAML) to output the predicted value of surrounding rock deformation and cognitive uncertainty. The engineering expert agent constructs a database of historical engineering cases of large tunnel deformation and uses cosine similarity to search for similar cases in the database, outputting the surrounding rock deformation treatment plan for the most similar case. Finally, the warning level is determined by comparing the threshold of the predicted value of surrounding rock deformation and the cognitive uncertainty measurement data. Low-risk levels can be monitored normally, while high-risk levels require triggering a warning. If the cognitive uncertainty measurement data is greater than the preset cognitive uncertainty measurement benchmark threshold, the output information is insufficient, and it is recommended to encrypt the monitoring reminder and temporarily not trigger the warning to prevent false alarms.

[0050] It is worth mentioning that, according to an embodiment of the present invention, the step of processing the fused feature vector using a preset Bayesian uncertainty quantification analysis method to obtain initial cognitive uncertainty measurement data includes: The fused feature vectors are sampled T times using the Monte Carlo random inactivation (MC Dropout) method; The mean and variance of the T sampling results are calculated to obtain the initial cognitive uncertainty measurement data.

[0051] Further explanation is needed. For the same input, forward propagation is performed T times (T≥50, 100 times in this embodiment) to statistically predict the mean and variance of the sample set. This is decomposed into initial cognitive uncertainty measurement data, which includes cognitive uncertainty measurement data (the model's cognitive boundary of the current data) and random uncertainty measurement data (the inherent noise level of the data).

[0052] This invention discloses a real-time early warning method and system for large deformation of tunnel surrounding rock based on multi-source information fusion. It achieves real-time early warning of large deformation of tunnel surrounding rock based on multi-source information fusion by constructing multi-source three-dimensional perception, sensor data self-correction, physical information neural operator combined mechanism driven preprocessing, causal cross-linked fusion and Bayesian uncertainty quantification analysis, cross-geological rapid adaptation combined with multi-expert intelligent agent collaborative decision-making.

[0053] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0054] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0055] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0056] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0057] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

Claims

1. A real-time early warning method for large deformation of surrounding rock in tunnels, characterized in that, Includes the following steps: A multi-source monitoring dataset of large deformation of surrounding rock was acquired, and spatiotemporal registration and data self-correction preprocessing were performed to obtain a corrected dataset of large deformation monitoring of surrounding rock. Mechanism-driven preprocessing was performed on the large deformation monitoring and correction dataset of surrounding rock to obtain the standard dataset of large deformation monitoring of surrounding rock. Based on the standard dataset for monitoring large deformation of surrounding rock and the correction dataset for monitoring large deformation of surrounding rock, the physical information neural operator model based on model-independent element learning (MAML) is used for analysis and processing to obtain the deformation rate and cumulative deformation of surrounding rock over a preset time period. Based on the monitoring and correction dataset of large deformation of surrounding rock, combined with the deformation rate and cumulative deformation, causal cross-linked multimodal fusion and Bayesian uncertainty quantification analysis were performed to obtain fused feature vectors and initial cognitive uncertainty measurement data. Based on the deformation rate and cumulative deformation of the surrounding rock, combined with the fused feature vector and the initial cognitive uncertainty measurement data, the analysis and processing are carried out by a multi-expert intelligent agent to obtain the early warning level of large deformation of the tunnel surrounding rock and the corresponding early warning interpretation report. Warning information is issued based on the warning level of large deformation of the surrounding rock in the tunnel, and the warning interpretation report is pushed to the management terminal for display.

2. The method for real-time early warning of large deformation of tunnel surrounding rock by multi-source information fusion according to claim 1, characterized in that, The process of acquiring a multi-source monitoring dataset for large deformation of surrounding rock, and performing spatiotemporal registration and data self-correction preprocessing to obtain a corrected monitoring dataset for large deformation of surrounding rock includes: A multi-source three-dimensional sensing network was constructed, including advanced geological exploration, deep surrounding rock monitoring, support stress monitoring, environmental hydrological monitoring, and construction parameter acquisition, and a multi-source monitoring dataset of large deformation of surrounding rock was obtained. The multi-source monitoring dataset of large deformation of surrounding rock was time-series aligned and spatially registered, and the data self-correction preprocessing was performed by a preset data correction model based on DDPG reinforcement learning to obtain the monitoring correction dataset of large deformation of surrounding rock.

3. The method for real-time early warning of large deformation of tunnel surrounding rock by multi-source information fusion according to claim 2, characterized in that, The step involves performing mechanism-driven preprocessing based on the large deformation monitoring and correction dataset of surrounding rock to obtain a standard dataset for large deformation monitoring of surrounding rock, including: Based on the large deformation monitoring correction dataset of surrounding rock, signal denoising is performed by wavelet packet hierarchical denoising and Kalman adaptive filtering to obtain the large deformation monitoring denoised dataset of surrounding rock. Based on the denoised dataset of large deformation monitoring of surrounding rock, missing data are filled in by spatial correlation of adjacent sensors, temporal grey GM model and neighborhood spatial interpolation to obtain the complete dataset of large deformation monitoring of surrounding rock. Based on the complete dataset of large deformation monitoring of surrounding rock, feature decoupling and normalization processing is performed through a preset normalization model of surrounding rock mechanical constraints, and feature encoding processing is performed to obtain the standard dataset of large deformation monitoring of surrounding rock.

4. The real-time early warning method for large deformation of tunnel surrounding rock by multi-source information fusion according to claim 3, characterized in that, The process involves analyzing and processing the standard dataset for large deformation monitoring of surrounding rock combined with the correction dataset using a physical information neural operator model based on Model-Independent Element Learning (MAML) to obtain predicted values ​​of the deformation rate and cumulative deformation over a preset time period. This includes: Based on the monitoring and correction dataset of large deformation of surrounding rock, parameters were extracted to obtain the number of microseismic events, microseismic dominant frequency offset rate, microseismic energy release rate, geostress change, geostress inversion value, surrounding rock category label, excavation disturbance data and support reaction force data, as well as the corresponding measured value of surrounding rock deformation rate and measured value of cumulative deformation. An initial model of the physical information neural operator was constructed, and the weighted sum of the data fitting loss and the partial differential equation physical constraint loss was used as the joint training objective function. The initial model of the physical information neural operator was trained based on the number of microseismic events, the microseismic dominant frequency offset rate, the microseismic energy release rate, the change in ground stress, the ground stress inversion value, the surrounding rock category label, the excavation disturbance data and the support reaction force data, as well as the corresponding measured values ​​of the surrounding rock deformation rate and the cumulative deformation, to obtain the physical information neural operator model. The standard dataset for monitoring large deformation of surrounding rock is input into the physical information neural operator model based on model-independent meta-learning (MAML) for analysis and processing to obtain the deformation rate and cumulative deformation value of surrounding rock over a preset time period.

5. The real-time early warning method for large deformation of tunnel surrounding rock according to claim 4, characterized in that, The process involves using the surrounding rock large deformation monitoring and correction dataset, combined with the surrounding rock deformation rate and cumulative deformation, to perform causal cross-linked multimodal fusion and Bayesian uncertainty quantification analysis, resulting in fused feature vectors and initial cognitive uncertainty measurement data, including: The monitoring and correction dataset of large deformation of surrounding rock is analyzed and processed through a preset dynamic Bayesian network causal graph to obtain data of cause layer, micro-damage layer and macro-response layer. Feature vectors are extracted from the data of the cause layer, the micro-damage layer, and the macro-response layer. These feature vectors are then analyzed and processed by a pre-defined causal decoupling network trained with the weighted sum of data reconstruction loss and counterfactual constraint loss as the objective function. Furthermore, cross-modal interaction is performed through a multi-head self-attention layer to obtain causal-enhanced fusion feature vectors. The causal enhancement fusion feature vector is fused with the surrounding rock deformation rate and cumulative deformation to obtain the fused feature vector; The initial cognitive uncertainty measurement data are obtained by processing the fused feature vectors using a pre-defined Bayesian uncertainty quantification analysis method.

6. The method for real-time early warning of large deformation of tunnel surrounding rock by multi-source information fusion according to claim 5, characterized in that, The process involves analyzing and processing the surrounding rock deformation rate and cumulative deformation using a fusion feature vector and initial cognitive uncertainty measurement data through a multi-expert intelligent agent to obtain the tunnel surrounding rock large deformation early warning level and corresponding early warning interpretation report, including: Construct multi-expert intelligent agents, including physical expert intelligent agents, data expert intelligent agents, and engineering expert intelligent agents; Based on the surrounding rock deformation rate and cumulative deformation, combined with the fusion feature vector and initial cognitive uncertainty measurement data, as well as the historical engineering case database of large tunnel deformation, the analysis and processing are carried out by physical expert intelligent agent, data expert intelligent agent and engineering expert intelligent agent to obtain the surrounding rock deformation prediction value, cognitive uncertainty measurement data, surrounding rock deformation safety threshold and surrounding rock deformation treatment plan. Compare the predicted value of surrounding rock deformation with the safe threshold value of surrounding rock deformation; If the predicted value of surrounding rock deformation is less than the safe threshold for surrounding rock deformation, the warning level for large deformation of the surrounding rock in the tunnel is determined to be low risk. If the predicted value of surrounding rock deformation is greater than or equal to the safety threshold of surrounding rock deformation, the cognitive uncertainty measurement data will be compared with the preset cognitive uncertainty measurement benchmark threshold. If the cognitive uncertainty measurement data is less than or equal to the preset cognitive uncertainty measurement benchmark threshold, the early warning level of large deformation of the surrounding rock of the tunnel is determined to be high risk level, and a corresponding early warning interpretation report is generated based on the predicted value of the surrounding rock deformation and the surrounding rock deformation treatment plan. If the cognitive uncertainty measurement data is greater than the preset cognitive uncertainty measurement benchmark threshold, an encrypted monitoring reminder response will be output.

7. A real-time early warning system for large deformation of surrounding rock in tunnels, characterized in that, The system includes a memory and a processor. The memory contains a program for a real-time early warning method based on multi-source information fusion for large deformation of tunnel surrounding rock. When the program for the real-time early warning method based on multi-source information fusion for large deformation of tunnel surrounding rock is executed by the processor, it performs the following steps: A multi-source monitoring dataset of large deformation of surrounding rock was acquired, and spatiotemporal registration and data self-correction preprocessing were performed to obtain a corrected dataset of large deformation monitoring of surrounding rock. Mechanism-driven preprocessing was performed on the large deformation monitoring and correction dataset of surrounding rock to obtain the standard dataset of large deformation monitoring of surrounding rock. Based on the standard dataset for monitoring large deformation of surrounding rock and the correction dataset for monitoring large deformation of surrounding rock, the physical information neural operator model based on model-independent element learning (MAML) is used for analysis and processing to obtain the deformation rate and cumulative deformation of surrounding rock over a preset time period. Based on the monitoring and correction dataset of large deformation of surrounding rock, combined with the deformation rate and cumulative deformation, causal cross-linked multimodal fusion and Bayesian uncertainty quantification analysis were performed to obtain fused feature vectors and initial cognitive uncertainty measurement data. Based on the deformation rate and cumulative deformation of the surrounding rock, combined with the fused feature vector and the initial cognitive uncertainty measurement data, the analysis and processing are carried out by a multi-expert intelligent agent to obtain the early warning level of large deformation of the tunnel surrounding rock and the corresponding early warning interpretation report. Warning information is issued based on the warning level of large deformation of the surrounding rock in the tunnel, and the warning interpretation report is pushed to the management terminal for display.

8. The real-time early warning system for large deformation of tunnel surrounding rock according to claim 7, characterized in that, The process of acquiring a multi-source monitoring dataset for large deformation of surrounding rock, and performing spatiotemporal registration and data self-correction preprocessing to obtain a corrected monitoring dataset for large deformation of surrounding rock includes: A multi-source three-dimensional sensing network was constructed, including advanced geological exploration, deep surrounding rock monitoring, support stress monitoring, environmental hydrological monitoring, and construction parameter acquisition, and a multi-source monitoring dataset of large deformation of surrounding rock was obtained. The multi-source monitoring dataset of large deformation of surrounding rock was time-series aligned and spatially registered, and the data self-correction preprocessing was performed by a preset data correction model based on DDPG reinforcement learning to obtain the monitoring correction dataset of large deformation of surrounding rock.

9. The real-time early warning system for large deformation of tunnel surrounding rock according to claim 8, characterized in that, The step involves performing mechanism-driven preprocessing based on the large deformation monitoring and correction dataset of surrounding rock to obtain a standard dataset for large deformation monitoring of surrounding rock, including: Based on the large deformation monitoring correction dataset of surrounding rock, signal denoising is performed by wavelet packet hierarchical denoising and Kalman adaptive filtering to obtain the large deformation monitoring denoised dataset of surrounding rock. Based on the denoised dataset of large deformation monitoring of surrounding rock, missing data are filled in by spatial correlation of adjacent sensors, temporal grey GM model and neighborhood spatial interpolation to obtain the complete dataset of large deformation monitoring of surrounding rock. Based on the complete dataset of large deformation monitoring of surrounding rock, feature decoupling and normalization processing is performed through a preset normalization model of surrounding rock mechanical constraints, and feature encoding processing is performed to obtain the standard dataset of large deformation monitoring of surrounding rock.

10. The real-time early warning system for large deformation of tunnel surrounding rock according to claim 9, characterized in that, The process involves analyzing and processing the standard dataset for large deformation monitoring of surrounding rock combined with the correction dataset using a physical information neural operator model based on Model-Independent Element Learning (MAML) to obtain predicted values ​​of the deformation rate and cumulative deformation over a preset time period. This includes: Based on the monitoring and correction dataset of large deformation of surrounding rock, parameters were extracted to obtain the number of microseismic events, microseismic dominant frequency offset rate, microseismic energy release rate, geostress change, geostress inversion value, surrounding rock category label, excavation disturbance data and support reaction force data, as well as the corresponding measured value of surrounding rock deformation rate and measured value of cumulative deformation. An initial model of the physical information neural operator was constructed, and the weighted sum of the data fitting loss and the partial differential equation physical constraint loss was used as the joint training objective function. The initial model of the physical information neural operator was trained based on the number of microseismic events, the microseismic dominant frequency offset rate, the microseismic energy release rate, the change in ground stress, the ground stress inversion value, the surrounding rock category label, the excavation disturbance data and the support reaction force data, as well as the corresponding measured values ​​of the surrounding rock deformation rate and the cumulative deformation, to obtain the physical information neural operator model. The standard dataset for monitoring large deformation of surrounding rock is input into the physical information neural operator model based on model-independent meta-learning (MAML) for analysis and processing to obtain the deformation rate and cumulative deformation value of surrounding rock over a preset time period.

Images