A tunnel deformation real-time monitoring and intelligent early warning management system

Abstract

The present application relates to the technical field of tunnel engineering monitoring, and discloses a tunnel deformation real-time monitoring and intelligent early warning management system. The distributed sensing network of the system is arranged on the surface of the tunnel lining, continuously captures three-dimensional deformation data of a plurality of monitoring points and encodes the three-dimensional deformation data into original deformation tensors; an adaptive learning engine receives the original deformation tensors, generates feature-enhanced deformation tensors; a dynamic control core stores a benchmark deformation specification, drives a hydraulic actuator to act, and measures a real-time deformation amount through a laser scanning unit; when the real-time deformation amount deviates from the benchmark specification by more than a dynamic tolerance, a recursive correction process is activated: a deformation residual tensor is calculated based on the feature-enhanced deformation tensor corresponding to the benchmark specification and the deformation tensor corresponding to the real-time deformation amount; the deformation residual tensor is optimized using a gradient descent algorithm to generate a deformation tensor; the optimized deformation tensor is iteratively updated to the benchmark specification, and the control instruction is cyclically executed until the deviation converges within the dynamic tolerance.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering monitoring technology, specifically to a real-time monitoring and intelligent early warning management system for tunnel deformation. Background Technology

[0002] Current tunnel deformation monitoring primarily relies on periodic manual measurements or fixed sensor threshold alarms. Existing technologies suffer from low data acquisition frequency, hindering continuous real-time monitoring. Data analysis methods are simplistic, often depending on single-point displacement exceedances, failing to capture the overall structural deformation characteristics. Control strategies based on fixed thresholds cannot adapt to the complex deformation behavior caused by the interaction between the tunnel's surrounding rock and support structure. The system lacks self-learning capabilities, making it difficult to adjust baseline parameters based on actual deformation. Existing methods require solutions to key technical challenges such as real-time continuous monitoring, overall deformation feature extraction, intelligent early warning, and adaptive control.

[0003] Traditional tunnel monitoring systems suffer from significant shortcomings in data processing and intelligent control. Sensor networks are sparsely deployed, resulting in insufficient spatial resolution of monitoring data. Deformation feature extraction methods are simplistic and fail to effectively identify deformation patterns and trends. Static and fixed benchmark specifications cannot reflect time-varying characteristics such as stress redistribution in surrounding rock and performance degradation of support structures. The rigid control command generation mechanism cannot dynamically adjust control strategies based on deformation development. Deviation correction processes lack optimization algorithms, leading to low adjustment efficiency. The inadequate recursive update mechanism hinders continuous performance optimization. Existing technologies necessitate the establishment of a fully adaptive management system encompassing data acquisition and intelligent control. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time monitoring and intelligent early warning management system for tunnel deformation, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a real-time monitoring and intelligent early warning management system for tunnel deformation, the system comprising:

[0006] Distributed sensor network, adaptive learning engine and dynamic control core;

[0007] The distributed sensor network is deployed on the surface of the tunnel lining to continuously capture three-dimensional deformation data from multiple monitoring points and encode them into raw deformation tensors.

[0008] The adaptive learning engine receives the original deformation tensor, extracts deformation feature maps through a convolutional neural network, and generates a feature-enhanced deformation tensor.

[0009] The dynamic control core stores a benchmark deformation specification. Based on the benchmark deformation specification and the feature-enhanced deformation tensor, it generates control commands to drive the hydraulic actuator. Simultaneously, it measures the real-time deformation of the tunnel using a laser scanning unit. When the deviation between the real-time deformation and the benchmark deformation specification exceeds the dynamic tolerance, the dynamic control core activates a recursive correction process: based on the feature-enhanced deformation tensor corresponding to the benchmark deformation specification and the deformation tensor corresponding to the real-time deformation, it calculates the deformation residual tensor; it optimizes the deformation residual tensor using a gradient descent algorithm to obtain the optimized deformation tensor; the system iteratively updates the optimized deformation tensor to the benchmark deformation specification and repeatedly executes the control commands until the deviation converges within the dynamic tolerance.

[0010] Preferably, the adaptive learning engine performs the following steps for each monitoring cycle: loading the original deformation tensor into a pre-trained deep belief network and outputting a deformation probability distribution; according to the deformation probability distribution, controlling the multispectral imaging device to acquire the thermal deformation spectrum of the tunnel, and resolving the actual deformation tensor from the thermal deformation spectrum;

[0011] Calculate the anisotropy index of the actual deformation tensor; when the anisotropy index is greater than or equal to a preset sensitivity threshold, call the deformation mapping function to convert the actual deformation tensor into a feature index tensor; the deformation mapping function is obtained through supervised learning training and associates the deformation with multi-dimensional feature indices.

[0012] The feature index tensor is reduced in dimensionality using a tensor decomposition algorithm to obtain a low-rank feature tensor. The low-rank feature tensor is then fed back into a deep belief network for reconstruction, generating a reconstructed deformable tensor. This process is repeated until the anisotropy index is below a preset sensitivity threshold, and the final reconstructed deformable tensor is used as the feature-enhanced deformable tensor.

[0013] Preferably, the specific steps for applying the tensor decomposition algorithm to reduce the dimensionality of the feature index tensor include: performing Tucker decomposition on the feature index tensor to extract the core tensor and factor matrix; calculating the singular value sequence of the core tensor and retaining components with singular values ​​greater than the noise threshold; reconstructing the feature subspace using the factor matrix and the pruned core tensor; and projecting the feature subspace onto an orthogonal basis to obtain a low-rank feature tensor; wherein the noise threshold is dynamically set using an adaptive threshold algorithm and determined based on the energy distribution ratio of the feature index tensor.

[0014] Preferably, the steps for obtaining the optimized deformed tensor include: monitoring the residual gradient direction, residual magnitude change rate, and residual temporal fluctuation characteristics of each sampling point; fusing the residual gradient direction, residual magnitude change rate, and residual temporal fluctuation characteristics to generate a point state vector; inputting the point state vector to the reinforcement learning agent and outputting the adaptive weight value of each sampling point; aggregating the adaptive weight values ​​of all sampling points to form the optimized deformed tensor; the reinforcement learning agent is trained using the Q-learning algorithm, and the reward function is optimized based on historical convergence speed.

[0015] Preferably, the system further includes a pattern evolution module for constructing a deformation trend prediction model; the specific steps are as follows: collecting the original deformation tensor and the feature-enhanced deformation tensor within a continuous monitoring period, and calculating the correlation coefficient matrix between the tensors; applying a time series analysis algorithm to extract the principal components of the correlation coefficient matrix; inputting the principal components into a long short-term memory network, and outputting the predicted deformation tensor for future monitoring periods; based on the error between the predicted deformation tensor and the actual deformation tensor, using a backpropagation algorithm to update the weight parameters of the long short-term memory network.

[0016] Preferably, the step of adjusting the feature-enhanced deformation tensor based on the deformation residual tensor in the dynamic control core includes: calling the mode evolution module to obtain the predicted deformation tensor; calculating the difference tensor between the predicted deformation tensor and the benchmark deformation specification; weightedly fusing the difference tensor and the deformation residual tensor to generate a compensation tensor; solving the least squares solution of the compensation tensor using the conjugate gradient method to obtain a calibration vector; and superimposing the calibration vector onto the feature-enhanced deformation tensor to generate the optimized deformation tensor.

[0017] Preferably, the training process of the reinforcement learning agent includes: constructing a state space as a set of point state vectors and an action space as a weight adjustment policy; defining the immediate reward as a logarithmic function of the bias convergence rate; exploring the action space through an epsilon-greedy policy and updating the Q-value table; and fixing the policy network parameters when the training rounds reach a stable condition.

[0018] Preferably, the distributed sensor network integrates fiber optic grating sensors, piezoelectric accelerometers, and microwave radar.

[0019] Preferably, the adaptive learning engine uses a sensor fusion algorithm to align multi-source data to a unified spatiotemporal coordinate system to generate the original deformable tensor.

[0020] Preferably, the sensor fusion algorithm includes Kalman filtering and data registration steps.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] The adaptive learning engine extracts feature maps from the original deformation tensor using a convolutional neural network (CNN) to generate a feature-enhanced deformation tensor. The CNN employs a multi-layer convolutional kernel structure to progressively extract local features and global correlations from the deformation data. The feature mapping process captures the spatial distribution patterns and trends of the deformation field, identifying key deformation regions. The feature-enhanced deformation tensor integrates multi-scale feature information, highlighting deformation patterns that significantly impact structural safety. The tensor enhancement process improves the signal-to-noise ratio of deformation features, enhancing the accuracy of subsequent analysis. The dynamic control core generates control commands based on the comparison results between the benchmark deformation specification and the feature-enhanced deformation tensor. The benchmark deformation specification is scientifically formulated based on tunnel design parameters, geological conditions, and engineering experience. Comparative analysis calculates the degree of difference between the feature-enhanced tensor and the benchmark specification across multiple dimensions. The generation of control commands considers factors such as the magnitude, spatial distribution, and development speed of these differences. The command output drives the hydraulic actuator to make precise adjustments, controlling the development of tunnel deformation.

[0023] The development of benchmark deformation specifications requires four core steps: parameter acquisition, classification modeling, threshold calibration, and dynamic adaptation, to ensure the operability and engineering suitability of the specifications. The parameter acquisition stage involves comprehensively collecting structural parameters from the tunnel design phase (including lining thickness, reinforcement density, and concrete strength grade), surrounding rock classification data from the geological survey report (rock integrity coefficient, water content, and compressive strength), historical monitoring data from similar projects, and safety control indicators from current industry technical standards. The classification modeling stage involves classifying tunnels according to cross-sectional location (inlet, outlet, and middle sections), surrounding rock grade (I-V), and burial depth (shallow ≤50m, medium 50-100m, deep >100m), constructing a multi-dimensional benchmark model framework, with each category corresponding to an independent specification parameter system. The threshold calibration stage involves adjusting the initial threshold of the model based on initial deformation monitoring data during construction to ensure that the specifications meet both design expectations and actual engineering conditions; the calibration process should focus on the differences in deformation characteristics during different stages, such as the construction disturbance period and the surrounding rock stabilization period. Dynamic adaptation phase: Establish a benchmark specification update mechanism. When the tunnel experiences sudden changes in geological conditions or adjustments to the support structure, the specification parameters are iteratively optimized based on the newly added monitoring data to ensure the timeliness and accuracy of the specifications.

[0024] When the deviation between the real-time deformation and the benchmark specification exceeds the dynamic tolerance, the system activates a recursive correction process. The dynamic tolerance is dynamically adjusted according to tunnel safety requirements and monitoring accuracy to ensure timely and accurate early warning. The recursive correction process calculates the deformation residual tensor based on the feature-enhanced deformation tensor corresponding to the benchmark specification and the deformation tensor corresponding to the real-time deformation. The deformation residual tensor quantifies the difference between the current deformation state and the desired state, including information such as magnitude, direction, and distribution. The gradient descent algorithm is used to optimize the deformation residual tensor, finding the adjustment direction that minimizes the residual. The gradient descent algorithm determines the optimal step size and direction through iterative calculation to ensure convergence efficiency. The optimization process obtains the optimized deformation tensor, which represents the ideal deformation state after correction. The optimized deformation tensor is iteratively updated to the benchmark deformation specification to achieve dynamic optimization of the control objective. The benchmark specification update considers the actual response characteristics and long-term deformation patterns of the tunnel structure. The updated benchmark specification is more consistent with the current engineering state, improving the adaptability of the control strategy. The control commands are executed iteratively until the deviation converges within the dynamic tolerance, ensuring the system remains stable within a safe range. Through the synergistic effect of feature enhancement, intelligent comparison, and recursive optimization, precise control and long-term safety of tunnel deformation are achieved. Convolutional neural networks improve feature extraction capabilities, gradient descent algorithms ensure optimization efficiency, and recursive mechanisms enable continuous improvement. This intelligent management system significantly enhances the safety monitoring level of tunnel engineering. Attached Figure Description

[0025] Figure 1 This is a schematic diagram illustrating the working principle of the real-time tunnel deformation monitoring and intelligent early warning management system described in this invention.

[0026] Figure 2 A flowchart for feature enhancement in adaptive learning engines;

[0027] Figure 3 A flowchart for dimensionality reduction processing of feature index tensors;

[0028] Figure 4 To enhance the evolution of training metrics for learning agents;

[0029] Figure 5 This is a comparison chart of the Kalman filter effects. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figure 1This invention provides a real-time monitoring and intelligent early warning management system for tunnel deformation. The system includes a distributed sensor network, an adaptive learning engine, and a dynamic control core. The distributed sensor network is deployed on the surface of the tunnel lining, continuously capturing three-dimensional deformation data from multiple monitoring points and encoding them into raw deformation tensors. The adaptive learning engine receives the raw deformation tensors, extracts deformation feature maps through a convolutional neural network, and outputs feature-enhanced deformation tensors. The dynamic control core stores benchmark deformation specifications and generates control commands based on the benchmark deformation specifications and feature-enhanced deformation tensors to drive the hydraulic actuators. Simultaneously, it measures the real-time deformation of the tunnel through a laser scanning unit. When the deviation between the real-time deformation and the benchmark deformation specifications exceeds the dynamic tolerance, the dynamic control core activates a recursive correction process. Based on the feature-enhanced deformation tensors corresponding to the benchmark deformation specifications and the deformation tensors corresponding to the real-time deformation, it calculates the deformation residual tensor, optimizes the deformation residual tensor using a gradient descent algorithm, and obtains the optimized deformation tensor. The system iteratively updates the optimized deformation tensor to the benchmark deformation specifications and repeatedly executes control commands until the deviation converges within the dynamic tolerance.

[0032] Example 1: See Figure 2 The adaptive learning engine loads the original deformation tensor into a pre-trained deep belief network for each monitoring cycle and outputs a deformation probability distribution. In some embodiments, the multispectral imaging device is controlled to acquire thermal deformation maps of the tunnel based on the deformation probability distribution, and the actual deformation tensor is parsed from the thermal deformation maps. The anisotropy index of the actual deformation tensor is calculated. In specific implementations, the anisotropy index is calculated using the following formula:

[0033]

[0034] in: Represents the anisotropy index. , , These represent three eigenvalues ​​of the actual deformable tensor. When the anisotropy index is greater than or equal to a preset sensitivity threshold, the deformable mapping function is invoked to convert the actual deformable tensor into a feature index tensor. The deformable mapping function is obtained through supervised learning and associates the deformable quantity with multi-dimensional feature indices. A tensor decomposition algorithm is applied to reduce the dimensionality of the feature index tensor, resulting in a low-rank feature tensor. This low-rank feature tensor is then fed back into a deep belief network for reconstruction, generating a reconstructed deformable tensor. This process is repeated until the anisotropy index falls below the preset sensitivity threshold, and the final reconstructed deformable tensor is used as the feature-enhanced deformable tensor. Optionally, the preset sensitivity threshold is set based on the tunnel structure characteristics.

[0035] Supervised learning training data preparation: Collect tunnel deformation sample data under different geological conditions and construction conditions. The samples need to cover multiple scenarios such as normal deformation, slight over-limit deformation, and moderate risk deformation. Each sample contains input data (actual deformation tensor) and label data (corresponding multi-dimensional feature indicators). The feature indicators include key dimensions such as deformation rate, deformation uniformity, and concentration of dangerous areas.

[0036] Supervised model architecture: A multilayer perceptron is selected as the basic model for supervised learning. The model includes an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is consistent with the dimension of the actual deformation tensor. There are two hidden layers. The number of neurons in the first layer is 1.5 times that of the input layer, and the number of neurons in the second layer is 0.8 times that of the first layer. The number of neurons in the output layer matches the dimension of the multidimensional feature index.

[0037] Supervised learning training process implementation: During data preprocessing, the collected sample data is normalized to eliminate dimensional differences between different dimensions and ensure training stability. The sample set is randomly divided into training, validation, and test sets in a 7:2:1 ratio. During model training, mean squared error is used as the loss function, and stochastic gradient descent is employed to optimize model parameters. The loss value of the validation set is monitored in real-time during training; if the loss value of the validation set does not decrease for five consecutive training epochs, the learning rate is adjusted and training continues. During model optimization, the generalization ability of the model is evaluated using validation set data. If overfitting is found, Dropout regularization is used to suppress it. Finally, the model with the highest accuracy on the test set is retained as the core model for the deformation mapping function. During model deployment and updates, the trained model is embedded into the adaptive learning engine. During system operation, incremental training is performed on the model every 100 new valid sample data sets accumulated to continuously optimize mapping accuracy.

[0038] In some embodiments, the multispectral imaging device acquires thermal deformation maps in the infrared band. In specific implementations, the deep belief network adopts a stacked structure of restricted Boltzmann machines. Optionally, the deformation mapping function is trained using supervised learning based on a multilayer perceptron. It can be understood that the tensor decomposition algorithm uses Tucker decomposition. In specific implementations, when the tensor decomposition algorithm uses Tucker decomposition, the feature index tensor is decomposed using Tucker decomposition to extract the core tensor and factor matrix; the singular value sequence of the core tensor is calculated, and components with singular values ​​greater than the noise threshold are retained; the feature subspace is reconstructed using the factor matrix and the pruned core tensor; the feature subspace is projected onto an orthogonal basis to obtain a low-rank feature tensor.

[0039] Example 2: See Figure 3When applying tensor decomposition algorithms to reduce the dimensionality of feature index tensors, Tucker decomposition is performed on the feature index tensors to extract the core tensor and factor matrix. The singular value sequence of the core tensor is calculated, and components with singular values ​​greater than the noise threshold are retained. The noise threshold is dynamically set using an adaptive threshold algorithm, determined based on the energy distribution ratio of the feature index tensor. Therefore, the formula for calculating the noise threshold is:

[0040]

[0041] in: Indicates the noise threshold. This represents the energy retention ratio coefficient. This represents the total number of singular values. Let k represent the k-th singular value. The eigenspace is reconstructed using the factor matrix and the pruned core tensor. The eigenspace is then projected onto an orthogonal basis to obtain a low-rank eigenspace tensor. In practice, the orthogonal basis is constructed through the Gram-Schmidt orthogonalization process.

[0042] When obtaining the optimized deformed tensor, the residual gradient direction, residual magnitude change rate, and residual temporal fluctuation characteristics of each sampling point are monitored. These residual gradient direction, residual magnitude change rate, and residual temporal fluctuation characteristics are fused to generate a point state vector. The point state vector is input to the reinforcement learning agent, which outputs an adaptive weight value for each sampling point. The adaptive weight values ​​of all sampling points are aggregated to form the optimized deformed tensor. In some embodiments, the reinforcement learning agent is trained using the Q-learning algorithm, and the reward function is optimized based on historical convergence speed. Optionally, the residual gradient direction is calculated using the central difference method, and the residual magnitude change rate is represented using a logarithmic ratio. It can be understood that the residual temporal fluctuation characteristics are characterized by autoregressive model coefficients. In specific implementations, when the residual temporal fluctuation characteristics are characterized by autoregressive model coefficients, an autoregressive model is fitted to the residual time series data of each sampling point. These coefficients serve as components of the point state vector to describe the dynamic behavior of the residuals. Optionally, the order of the autoregressive model is determined by an information criterion, and the coefficient vector is normalized to eliminate the influence of dimensions. It can be understood that the fitting process uses a sliding window update to adapt to changes in the time series data. In practice, the point state vector is normalized, the Q-value table of the reinforcement learning agent is stored in tabular form, and in some embodiments, the reward function is designed as the logarithmic function of the bias convergence rate. Optionally, the output of the adaptive weight value is implemented using the Softmax function to realize the probability distribution.

[0043] Example 3: The system also includes a pattern evolution module for constructing a deformation trend prediction model. This module collects the original deformation tensor and the feature-enhanced deformation tensor within continuous monitoring periods and calculates the correlation coefficient matrix between the tensors. A time series analysis algorithm is applied to extract the principal components of the correlation coefficient matrix. These principal components are input into a long short-term memory (LSM) network, which outputs the predicted deformation tensor for future monitoring periods. Based on the error between the predicted and actual deformation tensors, a backpropagation algorithm is used to update the weight parameters of the LSM network. In some embodiments, the time series analysis algorithm employs an autoregressive integral moving average model. It can be understood that the LSM network includes a forget gate, an input gate, and an output gate structure. Optionally, the correlation coefficient matrix is ​​calculated using the Pearson correlation coefficient method.

[0044] When the dynamic control core enhances the deformation tensor based on the deformation residual tensor, it calls the mode evolution module to obtain the predicted deformation tensor and calculates the difference tensor between the predicted deformation tensor and the benchmark deformation specification. The difference tensor and the deformation residual tensor are then weighted and fused to generate a compensation tensor. In specific implementation, the weighted fusion process is represented as follows:

[0045]

[0046] in: Represents the compensation tensor. Represents the difference tensor. Represents the deformed residual tensor. and This represents the preset fusion coefficient. The least-squares solution of the compensation tensor is obtained using the conjugate gradient method to obtain the calibration vector. This calibration vector is then superimposed on the feature-enhanced deformation tensor to generate the optimized deformation tensor. In some embodiments, the conjugate gradient method has a maximum number of iterations and a convergence tolerance. It can be understood that the least-squares solution is obtained through a system of normal equations. Optionally, the calibration vector superposition uses tensor addition. In a specific implementation, the tensor addition operation is carried out as follows: the calibration vector and the feature-enhanced deformation tensor are added element-wise along their corresponding dimensions to obtain the optimized deformation tensor. In some embodiments, the shape consistency between the calibration vector and the feature-enhanced deformation tensor is checked before addition to ensure dimensional matching before the addition operation can be performed. It can be understood that element-wise addition is implemented through array operations, with each element calculated independently. Optionally, if the dimensions do not match, a broadcast mechanism is used to expand the calibration vector to the same dimension as the feature-enhanced deformation tensor, and the addition result directly overwrites or updates the target tensor storage area. In practice, the fusion coefficients α and β are adjusted based on historical data, and the difference tensor is calculated using an element-wise difference method. The dimension of the compensation tensor is kept consistent with that of the feature enhancement deformation tensor.

[0047] Example 4: The training process of the reinforcement learning agent includes constructing a state space as a set of point state vectors, an action space as the weight adjustment policy, and defining the immediate reward as the logarithmic function of the bias convergence rate; the formula for calculating the immediate reward is:

[0048]

[0049] in: Indicates an immediate reward. The deviation convergence rate is calculated by dividing the change in deviation between the current monitoring period and the previous monitoring period by the time interval. An epsilon-greedy strategy is used to explore the action space and update the Q-value table. In specific implementations, the epsilon-greedy strategy sets the exploration probability to decay linearly from the initial value to the minimum value. The Q-value table uses a two-dimensional array structure to store state-action values, with the state space dimension consistent with the point state vector dimension. The action space includes three basic strategies: weight increase, weight decrease, and weight maintenance. In some embodiments, the point state vector includes three components: residual gradient direction, residual magnitude change rate, and residual temporal fluctuation characteristics. Optionally, the output of the weight adjustment strategy is mapped to a continuous value domain. When the training rounds reach a stable condition, the network parameters are fixed. The stable condition is defined as the change in Q-value over multiple consecutive training rounds being lower than a preset threshold. In specific implementations, the number of training rounds is set to a fixed number of iterations, Q-value updates use a temporal difference learning method, and the deviation convergence rate is calculated based on a sliding window of historical monitoring data.

[0050] Table 1: Training Parameters for Reinforcement Learning Agents

[0051] Parameter name Parameter value describe Initial value of exploration probability 0.9 epsilon-greedy strategy starting value Exploring the minimum probability 0.1 lower bound of the epsilon-greedy strategy Learning rate 0.01 Q value update rate Discount factor 0.95 Future reward decay coefficient Stability threshold 0.001 Q-value change threshold

[0052] In practical implementation, the point state vector is normalized to a uniform dimension, and the action space policy output is converted into a probability distribution using the Softmax function. The Q-value table is initialized with zero-padding. In some embodiments, the bias convergence rate is calculated using an exponentially weighted moving average method; optionally, the training epoch interval is set to a fixed time period. The state space set is indexed using a hash function, and the action space policy is implemented using a lookup table. The reward value is pruned to a reasonable range to prevent numerical overflow.

[0053] See Figure 4The figure illustrates the evolution of key metrics for reinforcement learning agents during training. Four metrics are represented: bias convergence rate (blue solid line), immediate reward (red solid line), Q-value change (green solid line), and exploration probability (gray dashed line). The horizontal axis represents the number of training epochs, while the vertical axes on the left and right sides represent the changes in metric values ​​and exploration probability, respectively. In the early stages of training, the exploration probability maintains a high initial value (0.9), prompting the agent to explore extensively in the action space. The bias convergence rate and immediate reward increase rapidly, while the Q-value change gradually decreases. As training progresses, the exploration probability linearly decays to a minimum value (0.1), and the agent gradually utilizes the learned policy. The bias convergence rate continues to increase and stabilizes, the immediate reward also steadily increases, and the Q-value change approaches 0, indicating that the policy is stabilizing. This intuitively reflects the transformation of reinforcement learning agents from exploration to exploitation during training, as well as the synergistic optimization of bias convergence, reward acquisition, and Q-value stability. It is perfectly consistent with the training mechanism of reinforcement learning agents in the project, which is based on the Q-learning algorithm and uses the logarithm of the bias convergence rate as the reward function. It demonstrates the dynamic balance between exploration and exploitation and the training convergence characteristics under the epsilon-greedy strategy.

[0054] Example 5: A distributed sensor network integrates a fiber optic grating sensor, a piezoelectric accelerometer, and a microwave radar. The fiber optic grating sensor measures the strain distribution of the structure, the piezoelectric accelerometer captures vibration acceleration signals, and the microwave radar acquires displacement change information. The adaptive learning engine uses a sensor fusion algorithm to align multi-source data to a unified spatiotemporal coordinate system, generating the original deformation tensor. The sensor fusion algorithm includes Kalman filtering and data registration steps. In specific implementation, the state equation of the Kalman filtering process is expressed as:

[0055]

[0056] in: Let k represent the system state vector at time k. Represents the state transition matrix. This represents the system state vector at time k-1. This represents the process noise vector. The data registration step achieves spatiotemporal alignment of multi-source data through coordinate transformation. In specific implementations, the sampling frequency of the fiber optic grating sensor is set to 1000Hz, the measurement range of the piezoelectric accelerometer is configured to ±50g, and the resolution of the microwave radar is set to 0.1mm. In some embodiments, the unified spatiotemporal coordinate system adopts the tunnel design coordinate system. It can be understood that the process noise covariance matrix of the Kalman filter is configured according to the sensor characteristics. The data registration step includes timestamp synchronization and spatial coordinate transformation operations. The sensor fusion algorithm outputs an original deformable tensor with dimensions of m×n×3, where m represents the number of monitoring points, n represents the time series length, and 3 represents the three-dimensional spatial coordinate components.

[0057] Fiber Bragg grating sensors are evenly spaced along the tunnel lining surface, piezoelectric accelerometers are installed at key structural locations, and microwave radars are deployed at the tunnel arch and sidewalls. In practice, time synchronization of multi-source data is achieved through a GPS clock module. The Kalman filter in the sensor fusion algorithm uses an extended Kalman filter to process the nonlinear observation model. The data registration step includes coordinate system unification and unit standardization. In some embodiments, the conversion coefficient between the wavelength drift and strain values ​​of the fiber Bragg grating sensors is determined through calibration experiments. It can be understood that the output signal of the piezoelectric accelerometers is low-pass filtered, and the ranging data from the microwave radar is obtained by averaging multiple measurements. The origin of the unified spatiotemporal coordinate system is set at the center of the tunnel entrance section, with the x-axis pointing in the tunnel excavation direction, the y-axis pointing horizontally in the tunnel cross-section, and the z-axis pointing vertically upwards. In practice, the generation process of the original deformation tensor includes a data validity check step, and invalid data is compensated for using interpolation methods. The output frequency of the sensor fusion algorithm is set to 100Hz to meet the requirements of real-time monitoring. The temperature measurement function of the fiber optic grating sensor is used for ambient temperature compensation. The frequency response characteristics of the piezoelectric accelerometer cover the range of 0.5-2000Hz. The sampling interval of the microwave radar is configured to 10 milliseconds.

[0058] See Figure 5 In the sensor fusion stage of the real-time tunnel deformation monitoring system, Kalman filtering technology is used to achieve noise reduction and fusion of multi-source sensor data. The blue curve in the figure represents the raw deformation measurement data captured by the distributed sensor network (integrating fiber optic grating sensors, piezoelectric accelerometers, and microwave radar). This raw data exhibits significant fluctuations and contains substantial noise, reflecting the original state of the multi-source data during acquisition due to environmental interference and equipment errors. The orange curve represents the data after Kalman filtering, which shows significantly improved smoothness and more accurately reflects the true trend of tunnel deformation. In practice, Kalman filtering optimizes the original deformation tensor by iteratively updating the state equation and observation equation, combined with the statistical characteristics of process and observation noise. This process effectively filters out high-frequency noise and abnormal fluctuations in the raw data, making the processed data more closely resemble the actual physical laws of tunnel deformation. This provides highly reliable basic data support for subsequent adaptive learning engines to extract deformation features and for dynamic control cores to perform deformation regulation and early warning, demonstrating the crucial role of Kalman filtering in multi-source sensor data fusion and noise suppression.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A tunnel deformation real-time monitoring and intelligent early warning management system, characterized in that, The system comprises a distributed sensing network, an adaptive learning engine and a dynamic control core; The distributed sensing network is deployed on the surface of the tunnel lining for continuously capturing three-dimensional deformation data of multiple monitoring points and encoding the data into original deformation tensors; The adaptive learning engine receives the original deformation tensors, extracts deformation feature maps through a convolutional neural network, and generates feature-enhanced deformation tensors; The dynamic control core stores a baseline deformation specification, generates control instructions according to the baseline deformation specification and the feature-enhanced deformation tensors, drives the hydraulic actuator to act, and simultaneously measures the real-time deformation of the tunnel through a laser scanning unit; When the deviation between the real-time deformation and the baseline deformation specification exceeds the dynamic tolerance, the dynamic control core activates a recursive correction process: based on the feature-enhanced deformation tensor corresponding to the baseline deformation specification and the deformation tensor corresponding to the real-time deformation, a deformation residual tensor is calculated; The gradient descent algorithm is used to optimize the deformation residual tensor to obtain an optimized deformation tensor; The system iteratively updates the optimized deformation tensor to the baseline deformation specification and repeatedly executes the control instructions until the deviation converges within the dynamic tolerance; For each monitoring period, the adaptive learning engine performs the following steps: loads the original deformation tensors into a pre-trained deep belief network, and outputs a deformation probability distribution; According to the deformation probability distribution, a multi-spectral imaging device is controlled to collect a thermal deformation map of the tunnel, and an actual deformation tensor is resolved from the thermal deformation map; An anisotropy index of the actual deformation tensor is calculated; When the anisotropy index is greater than or equal to a preset sensitivity threshold, a deformation mapping function is called to convert the actual deformation tensor into a feature index tensor; The deformation mapping function is trained through supervised learning and is associated with deformation and multi-dimensional feature indexes; A tensor decomposition algorithm is applied to reduce the dimension of the feature index tensor to obtain a low-rank feature tensor; the low-rank feature tensor is fed back to the deep belief network for reconstruction to generate a reconstructed deformation tensor; the above process is repeated until the anisotropy index is lower than the preset sensitivity threshold, and the final reconstructed deformation tensor is taken as the feature-enhanced deformation tensor.

2. The tunnel deformation real-time monitoring and intelligent early warning management system according to claim 1, characterized in that, The specific steps of applying the tensor decomposition algorithm to reduce the dimension of the feature index tensor include: performing Tucker decomposition on the feature index tensor to extract a core tensor and a factor matrix; calculating the singular value sequence of the core tensor and retaining components with singular values greater than a noise threshold; reconstructing a feature subspace using the factor matrix and the pruned core tensor; projecting the feature subspace onto an orthogonal basis to obtain a low-rank feature tensor; wherein the noise threshold is dynamically set through an adaptive threshold algorithm and is determined based on the energy distribution ratio of the feature index tensor.

3. The tunnel deformation real-time monitoring and intelligent early warning management system according to claim 2, characterized in that, The step of obtaining the optimized deformation tensor comprises: monitoring the residual error gradient direction, residual error amplitude change rate and residual error time sequence fluctuation characteristics of each sampling point; fusing the residual error gradient direction, residual error amplitude change rate and residual error time sequence fluctuation characteristics to generate a point state vector; inputting the point state vector into a reinforcement learning agent to output an adaptive weight value of each sampling point; aggregating the adaptive weight values of all sampling points to form the optimized deformation tensor; and training the reinforcement learning agent through a Q-learning algorithm, and optimizing a reward function based on a historical convergence speed.

4. The tunnel deformation real-time monitoring and intelligent early warning management system according to claim 1, characterized in that, The system further comprises a mode evolution module for constructing a deformation trend prediction model; the specific steps are: collecting original deformation tensors and feature-enhanced deformation tensors in a continuous monitoring period, calculating a correlation coefficient matrix between the tensors; applying a time series analysis algorithm to extract principal components of the correlation coefficient matrix; inputting the principal components into a long short-term memory network to output a predicted deformation tensor in a future monitoring period; and updating weight parameters of the long short-term memory network using a back propagation algorithm based on an error between the predicted deformation tensor and an actual deformation tensor.

5. The tunnel deformation real-time monitoring and intelligent early warning management system according to claim 4, characterized in that, The step of adjusting the feature-enhanced deformation tensor based on the deformation residual error tensor comprises: calling the mode evolution module to obtain a predicted deformation tensor; calculating a difference tensor between the predicted deformation tensor and a reference deformation specification; weighting and fusing the difference tensor and the deformation residual error tensor to generate a compensation tensor; solving a least square solution of the compensation tensor using a conjugate gradient method to obtain a calibration vector; and superimposing the calibration vector on the feature-enhanced deformation tensor to generate an optimized deformation tensor.

6. The tunnel deformation real-time monitoring and intelligent early warning management system according to claim 3, characterized in that, The training process of the reinforcement learning agent comprises: constructing a state space as a set of point state vectors, and an action space as a weight adjustment strategy; defining an immediate reward as a logarithmic function of a bias convergence rate; exploring the action space through an epsilon-greedy strategy, and updating a Q value table; and fixing strategy network parameters when a training round reaches a stable condition.

7. The tunnel deformation real-time monitoring and intelligent early warning management system according to claim 1, characterized in that, The distributed sensing network integrates fiber Bragg grating sensors, piezoelectric accelerometers and microwave radars.

8. The tunnel deformation real-time monitoring and intelligent early warning management system according to claim 7, characterized in that, The adaptive learning engine aligns multi-source data to a unified space-time coordinate system using a sensor fusion algorithm to generate an original deformation tensor.

9. The tunnel deformation real-time monitoring and intelligent early warning management system according to claim 8, characterized in that, The sensor fusion algorithm comprises Kalman filtering and data registration steps.

Images