A Deep Learning-Based Real-Time Monitoring and Control Method for Large Deformation in Deep Underground Engineering

Abstract

This invention discloses a method for real-time monitoring and control of large deformations in deep underground engineering based on deep learning. The method includes the deployment of a monitoring system, preprocessing of monitoring data, construction of a neural network model, training of the neural network model, output of prediction results, generation of control commands, and quantitative evaluation of control effects. Based on multi-source information such as surrounding rock strain, temperature, seepage pressure, and micro-seismic disturbance energy acquired synchronously, a deep learning model based on physical mechanism constraints is introduced, significantly improving the reliability and generalization ability of the prediction results. This enables intelligent identification of surrounding rock stability assessment and failure modes, automatic generation of control commands, and quantitative evaluation of implementation effects.

Description

Technical Field

[0001] This invention belongs to the field of underground engineering technology and relates to a method for real-time monitoring and control of large deformations in deep underground engineering based on deep learning. Background Technology

[0002] Deep underground engineering projects such as deep-buried tunnels and mine roadways are often located in complex geological environments characterized by high ground stress, high temperature, and high seepage pressure, making the surrounding rock prone to large deformations during construction. Such deformations can not only damage the support structure and cause the cavern to converge beyond permissible limits, but also lead to the expansion of loosened areas in the surrounding rock, further causing stress redistribution and exacerbating deformation, potentially even triggering major engineering accidents such as collapses. Furthermore, the development and connection of internal rock fissures during large deformations significantly weakens the impermeability of the surrounding rock, increases the risk of high-pressure water inrush, and also substantially increases support costs, delays construction, and adversely affects the long-term stability of the project. Therefore, achieving real-time monitoring and intelligent control of large deformations in the surrounding rock of deep engineering projects is a key technical challenge that urgently needs to be overcome in the field of geotechnical engineering.

[0003] Currently, the following prominent problems still exist in the monitoring and control of large deformations in deep underground engineering: 1) The causes of large deformations in surrounding rock are complex, involving the coupling effects of multiple fields such as stress field, seepage field, temperature field, and construction disturbance field. For example, high water pressure weakens the rock mass strength, stress concentration caused by blasting disturbance, and volume effect of rock mass caused by temperature changes. However, existing monitoring methods are mostly limited to the collection of single physical quantities, which makes it difficult to fully reflect the true state of the surrounding rock and easily leads to early warning delays or misjudgments; 2) Existing prediction models are mostly data-driven and lack the integration of physical mechanisms such as rock mechanics and seepage mechanics. In the context of complex deep geology and scarce monitoring data, such models are prone to overfitting, and the prediction results often deviate from the actual physical laws, resulting in insufficient reliability; 3) In actual engineering, the interpretation of monitoring data and subsequent control strategies rely heavily on engineering experience, lacking quantitative decision support and real-time feedback mechanisms, making it difficult to objectively evaluate the implementation effect of control measures.

[0004] In summary, current technologies have not yet achieved deep fusion of multi-physics information and effective embedding of mechanical mechanisms, nor do they possess closed-loop control capabilities based on real-time data. Therefore, it is currently necessary to develop an intelligent method that integrates real-time perception of multi-source information, physical mechanism-driven approaches, and dynamic closed-loop regulation to achieve accurate prediction and proactive control of large deformations in the surrounding rock of deep underground engineering projects. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a real-time monitoring and control method for large deformations in deep underground engineering based on deep learning. By combining deep learning methods and closed-loop control strategies, it achieves real-time, accurate, and intelligent perception and control of large deformation conditions of surrounding rock during the construction process of deep underground engineering, which is beneficial to improving construction safety, efficiency, quality, and intelligence.

[0006] A deep learning-based method for real-time monitoring and control of large deformations in deep underground engineering includes the following steps:

[0007] S1. Deployment of the monitoring system:

[0008] A distributed monitoring system is deployed in the surrounding rock area of ​​deep underground engineering projects, including a distributed fiber optic sensing module and a microseismic monitoring module. The distributed fiber optic sensing module uses a fiber optic grating array as the sensing element and is arranged in a grid pattern along the surface of the surrounding rock and deep boreholes to collect data on surrounding rock strain, temperature, and seepage pressure. The microseismic monitoring module consists of several three-component acceleration sensors arranged in the area affected by engineering excavation to monitor microseismic events induced by construction in real time and record the time of occurrence, disturbance energy density, and their three-dimensional spatial coordinates.

[0009] S2. Preprocessing of monitoring data:

[0010] The raw data acquired in step S1 is preprocessed, including noise reduction, normalization, spatiotemporal alignment, and coupling correction. Then, the surrounding rock strain is... e ( x , y , z , t ),temperature T ( x , y , z , t ), seepage pressure P ( x , y , z , t and perturbation energy density E ( x , y , z , t By combining them on a unified spatiotemporal grid, a spatiotemporally aligned multidimensional feature vector can be constructed.

[0011] The dimension of this vector is equal to the sum of the number of all physical field parameters contained in it at a specific spatiotemporal point, and it is directly used as the input of the neural network model in step S3;

[0012] Preferably, step S2 involves preprocessing the raw data, specifically including the following steps:

[0013] S201. Data Denoising and Quality Improvement: The original data is filtered and denoised to eliminate high-frequency noise introduced by construction vibration and electromagnetic interference, thereby improving the signal-to-noise ratio. At the same time, abnormal data is identified and replaced based on statistical principles to remove outliers that significantly deviate from the data distribution pattern. Interpolation algorithms are used to fill data gaps to ensure the continuity of the time series.

[0014] S202, Data Normalization: Mapping the values ​​of the original data to a unified [0,1] interval, thereby eliminating scale differences caused by different physical dimensions and ensuring that all monitoring parameters have comparable weights in the deep learning model; the normalized parameter values X norm The calculation expression is:

[0015]

[0016] in, X These are the original parameter values. X max and X min These are the global maximum and minimum values ​​of the parameter recorded in the historical engineering database and the current monitoring period, respectively.

[0017] S203, Spatiotemporal Alignment and Data Fusion: After noise reduction, restoration, and normalization, the monitoring data are aligned according to their precise spatiotemporal coordinates. x , y , z , t Spatial registration and temporal synchronization are performed to ensure that data from different sensors correspond to each other at the same spatial location and the same point in time, forming a spatiotemporally consistent multidimensional dataset.

[0018] S204, Data Coupling Correction:

[0019] The strain of the surrounding rock after steps S201 to S203 e 0 ( x , y , z , t The surrounding rock strain is then corrected. e ( x , y , z , t It satisfies the following expression:

[0020]

[0021] in, β 1 ,β 2 and β 3 These are coupling coefficients characterizing the contribution of temperature, seepage pressure, and disturbance energy density to the strain of the surrounding rock. These coefficients need to be calibrated through indoor physical and mechanical tests on representative rock samples.

[0022] S205. Construction of multidimensional feature vector: The data processed by steps S201 to S204 are combined on a unified spatiotemporal grid point to form the multidimensional feature vector.

[0023] S3. Construction of Neural Network Model:

[0024] A neural network model based on deep learning and physical constraints is constructed to fuse multi-physics field monitoring data and predict the surrounding rock condition and generate control commands; the neural network model includes an input layer, a feature extraction layer, a physical constraint layer and an output layer connected in sequence.

[0025] The input layer is configured to receive the multidimensional feature vector generated in step S2, and adopts a fully connected structure. The number of neurons is consistent with the dimension of the multidimensional feature vector, which is used to map the input data to a high-dimensional feature space.

[0026] The feature extraction layer is configured to automatically learn and fuse the complex spatial and temporal correlation features of the surrounding rock state from the multidimensional feature vectors. This layer adopts a hybrid neural network architecture, including a multimodal spatial feature extractor, a temporal evolution feature extractor, and a feature fusion module. The multimodal spatial feature extractor is implemented through a three-dimensional convolutional neural network and is used to extract spatial correlation features and coupling patterns between monitoring points at different spatial locations, as well as between multiple physical field parameters such as strain, temperature, seepage pressure, and disturbance energy. The temporal evolution feature extractor is implemented through a long short-term memory network and is used to capture the dynamic evolution law and historical dependence of the multidimensional feature vectors in the time series. The feature fusion module is implemented through an attention mechanism and is used to adaptively weight and fuse the spatial features output by the multimodal spatial feature extractor and the temporal features output by the temporal evolution feature extractor, and enhance the attention weight of feature information in high-risk areas.

[0027] The physical constraint layer is configured to embed the basic principles of geotechnical mechanics into the model's learning process, ensuring that the model's prediction results conform to physical laws and enhancing generalization. This layer is located between the feature extraction layer and the output layer, and physical constraints are achieved by introducing constraint terms based on physical equations into the model training loss function. The physical equation constraint terms include: constraints based on the effective stress principle, ensuring that the predicted surrounding rock stress state and pore water pressure conform to the effective stress relationship; constraints based on the rock constitutive relationship, ensuring that the predicted stress-strain path conforms to the basic mechanical behavior of the rock mass; and constraints based on the energy conservation principle, ensuring that the energy released by microseismic events is consistent with the energy consumption trend of surrounding rock deformation and failure.

[0028] The output layer is configured to generate the final prediction results and control parameters based on the fusion features extracted by the feature extraction layer and the requirements of the physical constraint layer. This layer uses a linear activation function, and its neuron output values ​​correspond to: the predicted maximum deformation of key locations in the surrounding rock within the next 24 hours; a stability assessment index characterizing the overall stability state of the current surrounding rock; a classification probability indicating the potential failure mode of the current surrounding rock, including compression type, shear type, and seepage-induced type; and quantitative control parameters used to guide subsequent control measures.

[0029] S4. Training the neural network model:

[0030] The total loss function of the neural network model L total Data loss L data With physical constraint loss L phys It consists of, that is, satisfying the following expression:

[0031]

[0032] in, L data The mean squared error between the model's predicted values ​​and the actual monitored values ​​is used to ensure the model's ability to fit the monitoring data. L phys The weighted sum of the physical constraint error terms is constructed by introducing the effective stress principle, rock constitutive relation and energy conservation principle, and is incorporated into the total loss function in the form of a weighted sum to ensure that the model output conforms to physical common sense. m 1 and m 2 These are the weighting coefficients, and m 1 + m 2 The condition is satisfied with 1, and the specific value is determined through cross-validation.

[0033] The training sample set consists of historical engineering measured data and virtual samples based on physical mechanism simulation to enhance the model's generalization ability under extreme conditions. The model training adopts the Adam optimization algorithm, with the initial learning rate set to 0.001, and a learning rate decay strategy is introduced to suppress overfitting.

[0034] S5. Output of prediction results:

[0035] The neural network model trained in step S4 performs forward calculation on the multidimensional feature vector processed in step S2, and outputs the predicted deformation value, stability assessment index and surrounding rock failure mode.

[0036] The predicted deformation value is the predicted maximum deformation value at key locations of the surrounding rock within the next 24 hours. U max ;

[0037] The stability assessment index I s ( x , y , z , t The quantitative index reflecting the stability state of the surrounding rock satisfies the following expression:

[0038]

[0039] in, e p This represents the peak strain of the rock mass. P p The critical seepage pressure;

[0040] The surrounding rock failure modes are output through a Softmax classifier deployed in the output layer, which outputs the probability distributions of three failure modes: compression failure, shear failure, and seepage-induced failure, respectively. P sq , P sh and P se Correspondingly, the probability of the highest probability destruction mode is... P max =max{ P sq , P sh , P se};

[0041] S6. Generation of control commands:

[0042] when U max ≥[ U max]、 I s ( x , y , z , t )≤[ I s ]or P max ≥[ P When [], the generation of control commands is triggered; among which, [ U max [This is a preset critical deformation threshold, set according to the engineering geological report;] I s [This refers to the stability safety threshold, determined according to the safety requirements of deep underground engineering design;] P [This represents the probability threshold for the destructive mode;]

[0043] Preferably, in step S6, when P max = P sq At that time, it reflects that the dominant type of surrounding rock failure mode is compression failure, and the neural network model generates control parameters that are beneficial to improving the radial constraint force of the surrounding rock; when P max = P sh At that time, it reflects that the dominant type of surrounding rock failure is shear failure, and the neural network model generates control parameters that are beneficial to suppressing structural plane slip; when P max = P se At that time, it was reflected that the dominant type of surrounding rock failure mode was seepage-induced failure. The neural network model generated control parameters that are conducive to reducing pore water pressure. The control parameters that are conducive to improving the radial constraint force of the surrounding rock include the increase in anchor bolt preload, the increase in grouting pressure, and the increase in support structure layout density. The control parameters that are conducive to suppressing structural surface slippage include the increase in anchor cable locking force and the increase in the thickness of the sprayed layer on the surrounding rock surface. The control parameters that are conducive to reducing pore water pressure include the increase in drainage hole flow rate and the increase in water-stopping grouting pressure.

[0044] S7. Quantitative evaluation of control effectiveness:

[0045] After implementing control measures according to the control instructions in step S6, the surrounding rock strain, temperature, seepage pressure, and microseismic energy density data are re-acquired and pre-processed according to steps S1 and S2 to construct a control effect index. I t The calculation expression is shown below:

[0046]

[0047] Where, Δ e Δ represents the change in strain after the implementation of control measures relative to the strain before the implementation of control measures. E The value represents the change in microseismic energy density, indicating the change in the average microseismic energy density within 2 hours after the implementation of control measures relative to the average microseismic energy density within 2 hours before the implementation of control measures; Δ P The seepage pressure change value represents the change in seepage pressure after the implementation of control measures relative to the seepage pressure before the implementation of control measures. e pre , E pre , P pre These are the surrounding rock strain, microseismic energy density, and seepage pressure before the implementation of control measures; α 1 , α 2 and α 3 These are weighting coefficients, and their specific values ​​are predetermined based on engineering geological conditions, the dominant type of surrounding rock failure mode, and engineering practice experience using the analytic hierarchy process (AHP) or expert scoring method. α 1 + α 2 + α 3 =1.

[0048] Preferably, in step S7, when the control effect index I t When the value is less than -0.5, it is determined that the current surrounding rock control measures have achieved the expected effect. At this time, the multi-source monitoring data and the executed control commands within this period are integrated into sample pairs and stored in the model training database for subsequent incremental learning. During the incremental learning process, in order to maintain the stability of the existing features of the model, the parameters of the convolutional neural network and long short-term memory network layers are kept unchanged, and only the fully connected weights of the output layer are fine-tuned and iteratively updated.

[0049] In summary, compared with existing technologies, the beneficial effects of this invention are as follows: Addressing the problems of low accuracy in interpreting monitoring data, prediction results that do not conform to physical laws, and subjective estimation of control parameters lacking a closed loop in the accurate prediction and active control technologies for large deformation of surrounding rock in deep underground engineering, this invention proposes a real-time monitoring and control method for large deformation in deep underground engineering based on deep learning. This method includes the deployment of a monitoring system, preprocessing of monitoring data, construction of a neural network model, training of the neural network model, output of prediction results, generation of control commands, and quantitative evaluation of control effects, achieving the following breakthrough improvements:

[0050] 1) Deep integration of multi-physics field monitoring data has been achieved: By deploying distributed optical fiber sensing and microseismic monitoring modules, multi-source information such as surrounding rock strain, temperature, seepage pressure and disturbance energy density is collected simultaneously. After preprocessing and coupling correction, a spatiotemporally unified multi-dimensional feature vector is constructed, which overcomes the limitations of traditional single physical field monitoring and provides a comprehensive and reliable data foundation for surrounding rock condition identification and deformation prediction.

[0051] 2) Improve the predictive reliability of neural network models: By introducing deep learning models based on physical mechanism constraints, the reliability and generalization ability of prediction results are significantly improved; by embedding the basic principles of geotechnical mechanics into the neural network, the model not only relies on data-driven approaches but also conforms to physical laws, effectively avoiding overfitting or predictive outputs that violate common sense due to data scarcity under extreme working conditions.

[0052] 3) It realizes the intelligent identification of surrounding rock stability assessment and failure mode: the model can output the future key deformation, stability index and typical failure mode classification probability, providing quantitative and multi-dimensional judgment basis for control decision, overcoming the shortcomings of traditional methods that rely on human experience and lack quantitative support.

[0053] 4) Achieved precise matching of control measures: A closed-loop intelligent control mechanism was constructed, which can automatically generate control commands and quantitatively evaluate the implementation effect based on the prediction results; for different dominant damage modes, it outputs quantitative control parameters with clear physical meaning and engineering guidance value, and objectively quantitatively evaluates the control effect;

[0054] 5) The model has online learning and optimization capabilities: By introducing the control effect index and incremental learning mechanism, the system can accumulate field control samples and continuously optimize the output layer weights, thereby adapting to changes in geological conditions and construction dynamics, and enhancing the long-term applicability and stability of the system. Attached Figure Description

[0055] Figure 1 This is a flowchart of the deep learning-based real-time monitoring and control method for large deformations in deep underground engineering, as described in this invention. Detailed Implementation

[0056] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments given herein are for illustration and explanation only and are not intended to limit the present invention.

[0057] This application discloses, as follows: Figure 1 The deep learning-based real-time monitoring and control method for large deformations in deep underground engineering, as shown, includes the following steps:

[0058] S1. Deployment of the monitoring system: A distributed monitoring system is deployed in the surrounding rock area of ​​the deep underground engineering project, including a distributed fiber optic sensing module and a microseismic monitoring module. The distributed fiber optic sensing module uses a corrosion-resistant and electromagnetic interference-resistant fiber optic grating array as the sensing element, and is arranged in a grid pattern along the surface of the surrounding rock and deep boreholes to collect data on surrounding rock strain, temperature, and seepage pressure. The microseismic monitoring module consists of several three-component acceleration sensors arranged in the area affected by the engineering excavation, used to monitor microseismic events induced by construction in real time and record the time of occurrence, disturbance energy density, and their three-dimensional spatial coordinates. The monitoring range of the distributed monitoring system covers the key deformation areas of the surrounding rock of the deep underground engineering project, the area affected by the engineering excavation, and the potential risk sections identified in the geological survey report, including but not limited to the arch, sidewalls, the area one diameter ahead of the tunnel face, and potential slip sections of the surrounding rock.

[0059] S2. Preprocessing of monitoring data: The raw data collected in step S1 is preprocessed, including noise reduction, normalization, spatiotemporal alignment, and coupling correction. Then, the surrounding rock strain is... e ( x , y , z , t ),temperature T ( x , y , z , t ), seepage pressure P ( x , y , z , t and perturbation energy density E ( x , y , z , t The data is combined at a unified spatiotemporal grid point to construct a spatiotemporally aligned multidimensional feature vector. The dimension of this vector is equal to the sum of the number of all physical field parameters contained at a specific spatiotemporal point, and it is directly used as the input to the neural network model in step S3. In specific implementation, step S2 preprocesses the original data, including the following steps:

[0060] S201. Data Denoising and Quality Improvement: The original data is filtered and denoised to eliminate high-frequency noise introduced by construction vibrations and electromagnetic interference, improving the signal-to-noise ratio. Simultaneously, outlier identification and replacement are performed based on statistical principles to remove outliers that significantly deviate from the data distribution pattern. Interpolation algorithms are used to fill data gaps, ensuring the continuity of the time series. Specifically, according to 3... sThe criteria remove data that differ from the data mean by more than three times the standard deviation. Then, time series linear interpolation is used to repair the missing data, that is, the average value of the data adjacent to the missing data in the time series is used to fill the missing data.

[0061] S202, Data Normalization: Mapping the values ​​of the original data to a unified [0,1] interval, thereby eliminating scale differences caused by different physical dimensions and ensuring that all monitoring parameters have comparable weights in the deep learning model; the normalized parameter values X norm The calculation expression is:

[0062] (1)

[0063] in, X These are the original parameter values. X max and X min These are the global maximum and minimum values ​​of the parameter recorded in the historical engineering database and the current monitoring period, respectively.

[0064] S203, Spatiotemporal Alignment and Data Fusion: After noise reduction, restoration, and normalization, the monitoring data are aligned according to their precise spatiotemporal coordinates. x , y , z , t Spatial registration and temporal synchronization are performed to ensure that data from different sensors correspond to each other at the same spatial location and time point, forming a spatiotemporally consistent multidimensional dataset. Specifically, the coordinates of all sensors are unified to the construction coordinate system of deep underground engineering, and the data from each sensor are interpolated to a unified grid point; at the same time, the data from each sensor are unified to the same sampling rate based on the clock of the data acquisition server.

[0065] S204, Data Coupling Correction: Considering the cross-interference of temperature, seepage pressure, and disturbance energy density on the surrounding rock strain, the surrounding rock strain after processing in steps S201 to S203 is corrected. e 0 ( x , y , z , t Physical field coupling correction is performed, and the corrected surrounding rock strain is... e ( x , y , z , t It satisfies the following expression:

[0066] (2)

[0067] in, β1 , β 2 and β 3 These are coupling coefficients characterizing the contributions of temperature, seepage pressure, and disturbance energy density to the surrounding rock strain. These coefficients need to be calibrated through laboratory physical and mechanical tests on representative rock samples. In practice, after calibration, the following coefficients are obtained: β 1 =0.05, β 2 =0.12, β 3 =0.08.

[0068] S205. Construction of the multidimensional feature vector: The data processed in steps S201 to S204 are combined on a unified spatiotemporal grid to form the multidimensional feature vector. The dimension of this vector is equal to the sum of the number of all physical field parameters contained in it at a specific spatiotemporal point, and it is directly used as the input to the neural network model in step S3.

[0069] S3. Construction of Neural Network Model: Construct a neural network model based on deep learning and physical constraints to fuse multi-physics field monitoring data and predict the surrounding rock condition and generate control commands; the neural network model includes an input layer, a feature extraction layer, a physical constraint layer and an output layer connected in sequence.

[0070] In specific implementation, the input layer is configured to receive the multidimensional feature vector generated in step S2, adopting a fully connected structure with the number of neurons matching the dimension of the multidimensional feature vector, and is used to map the input data to a high-dimensional feature space. The feature extraction layer is configured to automatically learn and fuse the complex correlation features of the surrounding rock state in the spatial and temporal dimensions from the multidimensional feature vector; this layer adopts a hybrid neural network architecture, including a multimodal spatial feature extractor, a temporal evolution feature extractor, and a feature fusion module; the multimodal spatial feature extractor is implemented through a three-dimensional convolutional neural network, used to extract spatial correlation features and coupling patterns between monitoring points at different spatial locations, as well as between multiple physical field parameters such as strain, temperature, seepage pressure, and disturbance energy; the temporal evolution feature extractor is implemented through a long short-term memory network, used to capture the dynamic evolution law and historical dependence of the multidimensional feature vector in the time series; the feature fusion module is implemented through an attention mechanism, used to adaptively weight and fuse the spatial features output by the multimodal spatial feature extractor and the temporal features output by the temporal evolution feature extractor, and to enhance the attention weight of feature information in high-risk areas. The physical constraint layer is configured to embed the fundamental principles of geotechnical mechanics into the model's learning process, ensuring that the model's predictions conform to physical laws and enhancing generalization. This layer is located between the feature extraction layer and the output layer, and implements physical constraints by introducing constraint terms based on physical equations into the model training loss function. These physical equation constraint terms include constraints based on the effective stress principle. L 1 This ensures that the predicted surrounding rock stress state and pore water pressure conform to the effective stress relationship; based on constraints from the rock constitutive relation. L 2 This ensures that the predicted stress-strain path conforms to the basic mechanical behavior of the rock mass; constraints based on the principle of energy conservation. L 3 This ensures that the energy released by microseismic events is consistent with the energy consumption trend of surrounding rock deformation and failure.

[0071] The output layer is configured to generate the final prediction results and control parameters based on the fusion features extracted by the feature extraction layer and the requirements of the physical constraint layer. This layer uses a linear activation function, and its neuron output values ​​correspond to: the predicted maximum deformation of key locations in the surrounding rock within the next 24 hours; a stability assessment index characterizing the overall stability of the current surrounding rock; a classification probability indicating the potential failure mode of the current surrounding rock, including compression type, shear type, and seepage-induced type; and quantitative control parameters used to guide subsequent control measures.

[0072] S4. Training of the neural network model: The total loss function of the neural network model. L total Data loss L dataWith physical constraint loss L phys It consists of, that is, satisfying the following expression:

[0073] (3)

[0074] in, L data The mean squared error between the model's predicted values ​​and the actual monitored values ​​is used to ensure the model's ability to fit the monitoring data. L phys The weighted sum of the physical constraint error terms is constructed by introducing the effective stress principle, rock constitutive relation and energy conservation principle, and is incorporated into the total loss function in the form of a weighted sum to ensure that the model output conforms to physical common sense. m 1 and m 2 These are the weighting coefficients, and m 1 + m 2 The condition is satisfied with 1, and the specific value is determined through cross-validation; in practice, L phys = d 1 · L 1 + d 2 · L 2 + d 3 · L 3 ,in d 1 , d 2 and d 3 All are weighting coefficients, and satisfy the following conditions: d 1 + d 2 + d 3 =1.

[0075] The training sample set consists of historical engineering measured data and virtual samples based on physical mechanism simulation to enhance the model's generalization ability under extreme conditions. The model training adopts the Adam optimization algorithm, with the initial learning rate set to 0.001, and a learning rate decay strategy is introduced to suppress overfitting.

[0076] S5. Output of prediction results: The neural network model trained in step S4 performs forward calculation on the multidimensional feature vector processed in step S2, and outputs the predicted value of deformation, stability assessment index and surrounding rock failure mode.

[0077] The predicted deformation value is the predicted maximum deformation value at key locations of the surrounding rock within the next 24 hours. U max ;

[0078] The stability assessment index I s ( x , y , z , t The quantitative index reflecting the stability state of the surrounding rock satisfies the following expression:

[0079] (4)

[0080] in, e p This represents the peak strain of the rock mass. P p The critical seepage pressure; I s The value range is [0,1]. The closer the value is to 1, the higher the stability of the surrounding rock; the closer the value is to 0, the worse the stability of the surrounding rock.

[0081] The surrounding rock failure modes are output through a Softmax classifier deployed in the output layer, which outputs the probability distributions of three failure modes: compression failure, shear failure, and seepage-induced failure, respectively. P sq , P sh and P se Correspondingly, the probability of the highest probability destruction mode is... P max =max{ P sq , P sh , P se}

[0082] S6. Generation of control commands: When U max ≥[ U max ]、 I s ( x , y , z , t )≤[ I s ]or P max ≥[ P When [ ], control commands are generated; among them, [U max [This is a preset critical deformation threshold, set according to the engineering geological report;] I s [This refers to the stability safety threshold, determined according to the safety requirements of deep underground engineering design;] P [] represents the probability threshold for the destruction mode.

[0083] when P max = P sq At that time, it reflects that the dominant type of surrounding rock failure mode is compression failure, and the neural network model generates control parameters that are beneficial to improving the radial constraint force of the surrounding rock; when P max = P sh At that time, it reflects that the dominant type of surrounding rock failure is shear failure, and the neural network model generates control parameters that are beneficial to suppressing structural plane slip; when P max = P se At that time, it was reflected that the dominant type of surrounding rock failure mode was seepage-induced failure, and the control parameters generated by the neural network model were beneficial to reducing pore water pressure.

[0084] In specific implementation, in step S6, the control parameters that are beneficial to improving the radial constraint force of the surrounding rock include the increase in anchor bolt preload, the increase in grouting pressure, and the increase in the density of the support structure layout; the control parameters that are beneficial to suppressing structural surface slippage include the increase in anchor cable locking force and the increase in the thickness of the sprayed layer on the surrounding rock surface; and the control parameters that are beneficial to reducing pore water pressure include the increase in drainage hole flow rate and the increase in water-stopping grouting pressure.

[0085] S7. Quantitative Evaluation of Control Effectiveness: After implementing control measures according to the control instructions in step S6, the surrounding rock strain, temperature, seepage pressure, and microseismic energy density data are re-acquired and pre-processed according to steps S1 and S2 to construct a control effectiveness index. I t The calculation expression is shown below:

[0086] (5)

[0087] Where, Δ e Δ represents the change in strain after the implementation of control measures relative to the strain before the implementation of control measures. E The value represents the change in microseismic energy density, indicating the change in the average microseismic energy density within 2 hours after the implementation of control measures relative to the average microseismic energy density within 2 hours before the implementation of control measures; Δ PThe seepage pressure change value represents the change in seepage pressure after the implementation of control measures relative to the seepage pressure before the implementation of control measures. e pre , E pre , P pre These are the surrounding rock strain, microseismic energy density, and seepage pressure before the implementation of control measures; α 1 , α 2 and α 3 These are weighting coefficients, and their specific values ​​are predetermined based on engineering geological conditions, the dominant type of surrounding rock failure mode, and engineering practice experience using the analytic hierarchy process (AHP) or expert scoring method. α 1 + α 2 + α 3 =1.

[0088] In specific implementation, when the control effect index I t When the value is less than -0.5, it is determined that the current surrounding rock control measures have achieved the expected effect. At this time, the multi-source monitoring data and the executed control commands within this period are integrated into sample pairs and stored in the model training database for subsequent incremental learning. During the incremental learning process, in order to maintain the stability of the existing features of the model, the parameters of the convolutional neural network and long short-term memory network layers are kept unchanged, and only the fully connected weights of the output layer are fine-tuned and iteratively updated.

[0089] The above describes one or more embodiments of the present invention in a relatively specific and detailed manner, but it should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A method for real-time monitoring and control of large deformations in deep underground engineering based on deep learning, characterized in that, Includes the following steps: S1. Deployment of the monitoring system: Deploy a distributed monitoring system in the surrounding rock area of ​​the deep underground engineering to collect data on surrounding rock strain, temperature and seepage pressure, and monitor microseismic events induced by construction in real time and record the time of occurrence, disturbance energy density and its three-dimensional spatial coordinates. S2. Preprocessing of monitoring data: The raw data collected in step S1 is preprocessed, including noise reduction, normalization, spatiotemporal alignment, and coupling correction. Then, the surrounding rock strain is... ε ( x , y , z , t ),temperature T ( x , y , z , t ), seepage pressure P ( x , y , z , t and perturbation energy density E ( x , y , z , t By combining them on a unified spatiotemporal grid, a spatiotemporally aligned multidimensional feature vector can be constructed. S3. Construction of Neural Network Model: Construct a neural network model based on deep learning and physical constraints to fuse multi-physics field monitoring data and predict the surrounding rock condition and generate control commands; S4. Training the Neural Network Model: Train the neural network model, and the total loss function... L total Data loss L data With physical constraint loss L phys Composition; physical constraints are achieved by introducing constraint terms based on physical equations into the model training loss function; the physical equation constraint terms include: constraints based on the effective stress principle to ensure that the predicted stress state of the surrounding rock conforms to the effective stress relationship with the pore water pressure; constraints based on the rock constitutive relationship to ensure that the predicted stress-strain path conforms to the basic mechanical behavior of the rock mass; and constraints based on the energy conservation principle to ensure that the energy released by microseismic events is consistent with the energy consumption trend of surrounding rock deformation and failure. S5. Output of Prediction Results: The neural network model trained in step S4 performs forward calculations on the multidimensional feature vectors processed in step S2, outputting the predicted deformation value, stability assessment index, and surrounding rock failure mode; the predicted deformation value is the predicted maximum deformation value at key locations of the surrounding rock within the next 24 hours. U max ; The stability assessment index I s ( x , y , z , t The quantitative index reflecting the stability state of the surrounding rock satisfies the following expression: in, ε p This represents the peak strain of the rock mass. P p The critical seepage pressure; The surrounding rock failure modes are output through a Softmax classifier deployed in the output layer, which outputs the probability distributions of three failure modes: compression failure, shear failure, and seepage-induced failure, respectively. P sq , P sh and P se Correspondingly, the probability of the highest probability destruction mode is... P max =max{ P sq , P sh , P se }; S6. Generation of control commands: When U max ≥[ U max ]、 I s ( x , y , z , t )≤[ I s ]or P max ≥[ P When [], the generation of control commands is triggered; among which, [ U max [This is a preset critical deformation threshold, set according to the engineering geological report;] I s [This refers to the stability safety threshold, determined according to the safety requirements of deep underground engineering design;] P [This represents the probability threshold for the destructive mode;] S7. Quantitative Evaluation of Control Effectiveness: After implementing control measures according to the control instructions in step S6, the surrounding rock strain, temperature, seepage pressure, and microseismic energy density data are re-acquired and pre-processed according to steps S1 and S2 to construct a control effectiveness index. I t .

2. The method for real-time monitoring and control of large deformations in deep underground engineering based on deep learning according to claim 1, characterized in that, In step S1, the distributed monitoring system includes a distributed optical fiber sensing module and a microseismic monitoring module. The distributed optical fiber sensing module uses a fiber optic grating array as the sensing element and is deployed in a grid pattern along the surface of the surrounding rock and deep boreholes to collect data on surrounding rock strain, temperature, and seepage pressure. The microseismic monitoring module consists of several three-component acceleration sensors arranged within the area affected by the engineering excavation. It is used to monitor microseismic events induced by construction in real time and record the time of occurrence, disturbance energy density, and their three-dimensional spatial coordinates.

3. The method for real-time monitoring and control of large deformations in deep underground engineering based on deep learning according to claim 1, characterized in that, Step S2 involves preprocessing the raw data, specifically including the following steps: S201. Data Denoising and Quality Improvement: The original data is filtered and denoised to eliminate high-frequency noise introduced by construction vibration and electromagnetic interference, thereby improving the data signal-to-noise ratio. At the same time, based on statistical principles, abnormal data is identified and replaced to remove outliers that significantly deviate from the data distribution pattern, and interpolation algorithms are used to fill in data gaps. S202, Data Normalization: Mapping the values ​​of the original data to a unified [0,1] interval, thereby eliminating scale differences caused by different physical dimensions and ensuring that all monitoring parameters have comparable weights in the deep learning model; the normalized parameter values X norm The calculation expression is: in, X These are the original parameter values. X max and X min These are the global maximum and minimum values ​​of the parameter recorded in the historical engineering database and the current monitoring period, respectively. S203, Spatiotemporal Alignment and Data Fusion: After noise reduction, restoration, and normalization, the monitoring data are aligned according to their precise spatiotemporal coordinates. x , y , z , t Spatial registration and temporal synchronization are performed to ensure that data from different sensors correspond to each other at the same spatial location and the same point in time, forming a spatiotemporally consistent multidimensional dataset. S204, Data Coupling Correction: The strain of the surrounding rock after steps S201 to S203 ε 0 ( x , y , z , t The surrounding rock strain is then corrected. ε ( x , y , z , t It satisfies the following expression: in, β 1 , β 2 and β 3 These are coupling coefficients characterizing the contribution of temperature, seepage pressure, and disturbance energy density to the strain of the surrounding rock. These coefficients need to be calibrated through indoor physical and mechanical tests on representative rock samples. S205. Construction of multidimensional feature vector: The data processed by steps S201 to S204 are combined on a unified spatiotemporal grid point to form the multidimensional feature vector.

4. The method for real-time monitoring and control of large deformations in deep underground engineering based on deep learning according to claim 1, characterized in that, In step S3, the neural network model includes an input layer, a feature extraction layer, a physical constraint layer, and an output layer connected in sequence; the input layer is configured to receive the multidimensional feature vector generated in step S2, adopts a fully connected structure, and the number of neurons is consistent with the dimension of the multidimensional feature vector, which is used to map the input data to a high-dimensional feature space; The feature extraction layer is configured to automatically learn and fuse the complex correlation features of the surrounding rock state in the spatial and temporal dimensions from the multi-dimensional feature vectors. This layer adopts a hybrid neural network architecture, including a multimodal spatial feature extractor, a temporal evolution feature extractor, and a feature fusion module. The multimodal spatial feature extractor is implemented through a three-dimensional convolutional neural network and is used to extract the spatial correlation features and coupling patterns between monitoring points at different spatial locations, as well as between multi-physical field parameters such as strain, temperature, seepage pressure, and disturbance energy. The temporal evolution feature extractor is implemented through a long short-term memory network and is used to capture the dynamic evolution law and historical dependence of the multi-dimensional feature vectors in the time series. The feature fusion module is implemented through an attention mechanism and is used to adaptively weight and fuse the spatial features output by the multimodal spatial feature extractor and the temporal features output by the temporal evolution feature extractor, and enhance the attention weight of feature information in high-risk areas. The physical constraint layer is configured to embed the basic principles of geotechnical mechanics into the model's learning process, ensuring that the model's prediction results conform to physical laws and enhancing generalization. This layer is located between the feature extraction layer and the output layer. The output layer is configured to generate the final prediction results and control parameters based on the fusion features extracted by the feature extraction layer and the requirements of the physical constraint layer. This layer uses a linear activation function, and its neuron output values ​​correspond to: the predicted maximum deformation of key locations in the surrounding rock within the next 24 hours; a stability assessment index characterizing the overall stability of the current surrounding rock; a classification probability indicating the potential failure mode of the current surrounding rock, including compression type, shear type, and seepage-induced type; and quantitative control parameters used to guide subsequent control measures.

5. The method for real-time monitoring and control of large deformations in deep underground engineering based on deep learning according to claim 1, characterized in that, In step S4, the total loss function L total Satisfy the following expression: in, L data The mean squared error between the model's predicted values ​​and the actual monitored values ​​is used to ensure the model's ability to fit the monitoring data. L phys The weighted sum of the physical constraint error terms is constructed by introducing the effective stress principle, rock constitutive relation and energy conservation principle; μ 1 and μ 2 These are the weighting coefficients, and μ 1 + μ 2 The learning rate is set to 1, with the specific value determined through cross-validation. The training sample set consists of historical engineering measured data and virtual samples based on physical mechanism simulation to enhance the model's generalization ability under extreme conditions. The model training uses the Adam optimization algorithm, with the initial learning rate set to 0.001, and a learning rate decay strategy is introduced to suppress overfitting.

6. The method for real-time monitoring and control of large deformations in deep underground engineering based on deep learning according to claim 1, characterized in that, In step S6, when P max = P sq At that time, it was reflected that the dominant type of surrounding rock failure mode was compression failure, and the neural network model generated control parameters that were beneficial to improving the radial constraint force of the surrounding rock. when P max = P sh At that time, it was reflected that the dominant type of surrounding rock failure mode was shear failure, and the control parameters generated by the neural network model were beneficial to suppressing the slip of the structural plane. when P max = P se At that time, it was reflected that the dominant type of surrounding rock failure mode was seepage-induced failure. The neural network model generated control parameters that are conducive to reducing pore water pressure. The control parameters that are conducive to improving the radial constraint force of the surrounding rock include the increase in anchor bolt preload, the increase in grouting pressure, and the increase in support structure layout density. The control parameters that are conducive to suppressing structural surface slippage include the increase in anchor cable locking force and the increase in the thickness of the sprayed layer on the surrounding rock surface. The control parameters that are conducive to reducing pore water pressure include the increase in drainage hole flow rate and the increase in water-stopping grouting pressure.

7. The method for real-time monitoring and control of large deformations in deep underground engineering based on deep learning according to claim 1, characterized in that, In step S7, the control effect index I t The calculation expression is shown in the following formula: Where, Δ ε Δ represents the change in strain after the implementation of control measures relative to the strain before the implementation of control measures. E The value represents the change in microseismic energy density, indicating the change in the average microseismic energy density within 2 hours after the implementation of control measures relative to the average microseismic energy density within 2 hours before the implementation of control measures; Δ P The seepage pressure change value represents the change in seepage pressure after the implementation of control measures relative to the seepage pressure before the implementation of control measures. ε pre , E pre , P pre These are the surrounding rock strain, microseismic energy density, and seepage pressure before the implementation of control measures; α 1 , α 2 and α 3 These are weighting coefficients, and their specific values ​​are predetermined based on engineering geological conditions, the dominant type of surrounding rock failure mode, and engineering practice experience using the analytic hierarchy process (AHP) or expert scoring method. α 1 + α 2 + α 3 =1; when the control effect index I t When the value is less than -0.5, it is determined that the current surrounding rock control measures have achieved the expected effect. At this time, the multi-source monitoring data and the executed control commands within this period are integrated into sample pairs and stored in the model training database for subsequent incremental learning. During the incremental learning process, in order to maintain the stability of the existing features of the model, the parameters of the convolutional neural network and long short-term memory network layers are kept unchanged, and only the fully connected weights of the output layer are fine-tuned and iteratively updated.

Images