Tunnel deformation real-time early warning method and system based on LSTM-CNN model

Abstract

The invention discloses a tunnel deformation real-time early warning method and system based on an LSTM-CNN model, and belongs to the technical field of tunnel engineering safety monitoring. The method comprises four steps of data acquisition, spatio-temporal feature fusion processing, deformation prediction and risk assessment, and intelligent early warning and decision support: collecting multi-dimensional monitoring data through a distributed sensor network and recording spatio-temporal labels; after standardized noise reduction, extracting time trend and spatial distribution characteristics by using an LSTM-CNN fusion model, and constructing a time-space sequence data set; outputting a deformation prediction value based on the fusion features, and calculating a deviation degree in combination with a dynamic threshold model; and triggering multi-level early warning according to the deviation degree and performing visual display. The system comprises a data acquisition module, an edge calculation module, a cloud analysis module and an intelligent terminal module, and double-channel redundancy transmission is adopted to guarantee data continuity. Through spatio-temporal feature collaborative mining, dynamic threshold value adaptation and graded early warning, the deformation prediction precision and the risk assessment accuracy are improved, the transmission stability in an extreme environment is guaranteed, and the early warning decision efficiency is optimized.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering safety monitoring technology, and in particular to a real-time early warning method and system for tunnel deformation based on an LSTM-CNN model. Background Technology

[0002] As a critical infrastructure in transportation, water conservancy, and mining, tunnel engineering's structural safety directly impacts operational efficiency and the safety of personnel and property. With tunnel construction extending into deeper and more complex geological conditions, the demands for real-time and accurate deformation monitoring are increasing. Currently, the industry widely employs distributed sensor networks (such as fiber optic sensors and inclinometers) to collect tunnel structural parameters, covering multi-dimensional data such as crown settlement, horizontal convergence, surrounding rock stress, and environmental parameters. Simultaneously, machine learning technology is gradually being applied to deformation analysis, attempting to predict deformation trends through single time-series models (such as LSTM) or spatial feature models (such as CNN), providing data support and technological exploration for tunnel safety management.

[0003] However, existing technologies still have significant limitations in practical applications: First, traditional monitoring methods fail to adequately explore the spatiotemporal coupling characteristics of tunnel deformation. Single-time models struggle to capture spatial distribution patterns, and single-space models cannot depict long-term evolution trends, resulting in limited deformation prediction accuracy. Second, warning thresholds often use fixed values, making it difficult to adapt to dynamic changes in different geological conditions (such as soft rock and hard rock) and construction stages (excavation, support, and operation), easily leading to false alarms or missed alarms. Third, data transmission and processing lack redundancy, and in extreme environments, transmission interruptions or data noise can cause analysis delays. Fourth, warning response mechanisms are mostly single-level, making it difficult to match precise handling plans according to the degree of risk, affecting decision-making efficiency. These problems restrict the reliability and practicality of tunnel deformation early warning, necessitating an intelligent early warning solution that integrates spatiotemporal characteristics, dynamically adapts to working conditions, ensures stable transmission, and provides tiered responses. Summary of the Invention

[0004] The purpose of this invention is to address the problems in existing technologies, such as insufficient mining of spatiotemporal coupling features of tunnel deformation leading to limited prediction accuracy, poor adaptability of fixed thresholds resulting in false alarms and missed alarms, lack of data transmission redundancy leading to analysis lag, and a single early warning response mechanism affecting decision-making efficiency. This invention provides a real-time early warning method and system for tunnel deformation based on an LSTM-CNN model.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A real-time early warning method for tunnel deformation based on an LSTM-CNN model is provided, including the following steps:

[0007] S1. Data Acquisition: Real-time collection of multi-dimensional monitoring data of the tunnel structure through a distributed sensor network, including but not limited to crown settlement, horizontal convergence, surrounding rock stress and environmental temperature and humidity data, and simultaneous recording of spatiotemporal labels for each data point;

[0008] S2. Spatiotemporal feature fusion processing: Standardize and denoise the original data to construct a spatiotemporal sequence dataset;

[0009] An LSTM-CNN fusion model is used to extract features from spatiotemporal sequence data. The LSTM layer captures the deformation trend in the time dimension, while the CNN layer extracts the deformation distribution features in the spatial dimension through multi-scale convolutional kernels.

[0010] S3. Deformation Prediction and Risk Assessment: Based on the fused spatiotemporal characteristics, the model outputs the predicted deformation value within the future time window; a dynamic threshold model is constructed by combining historical safety data to calculate the deviation between the predicted value and the threshold.

[0011] S4. Intelligent Early Warning and Decision Support: Triggers multi-level early warning signals based on the deviation level, and displays the deformation trend and risk area distribution through a visual interface.

[0012] Furthermore, the standardized noise reduction process described in step S2 includes:

[0013] S2.1 Normalizes the original data using the following formula:

[0014]

[0015] Where, x i Here, μ represents the original data points, σ ​​represents the data mean, and σ represents the data standard deviation. This formula addresses the issue of inconsistent unit values ​​for different data types (e.g., settlement is measured in mm, stress in MPa) in tunnel monitoring data. By normalizing the data to the same order of magnitude (typically [-1, 1]), it avoids the LSTM-CNN model becoming overly sensitive to large numerical features due to differences in numerical ranges, thereby improving the stability and convergence speed of the model's feature learning.

[0016] S2.2 employs a sliding window method to interpolate and repair outlier data points, with the window length adaptively adjusted based on data fluctuation characteristics. The sliding window method, by adaptively adjusting the window length (e.g., a short window for high-frequency fluctuating data and a long window for low-frequency trend data), repairs outliers caused by sensor malfunctions or interference, ensuring data continuity and providing reliable input for subsequent feature extraction.

[0017] Furthermore, the LSTM-CNN fusion model training process described in step S2 includes:

[0018] S2.1 uses the cross-entropy loss function to optimize the model parameters, as shown in the formula:

[0019]

[0020] Among them, y i For real labels, Here, n represents the predicted probability, and n is the number of samples. This formula addresses the issue of insufficient prediction accuracy in LSTM-CNN models. By quantifying the difference between the predicted values ​​(deformation probability distribution) and the true values ​​(historical deformation labels), it guides iterative optimization of model parameters (such as the gating weights of LSTM and the convolutional kernel parameters of CNN), minimizing prediction errors and improving the model's ability to fit tunnel deformation trends.

[0021] S2.2 uses the Particle Swarm Optimization (PSO) algorithm to dynamically adjust the forget gate weights of the LSTM layer and the number of convolutional kernels in the CNN layer to balance the model's generalization ability and computational efficiency. By dynamically optimizing the LSTM forget gate weights (affecting the strength of temporal feature retention) and the number of CNN convolutional kernels (affecting the spatial feature dimensionality), the PSO algorithm addresses the imbalance between model generalization ability and computational efficiency—avoiding overfitting (poor generalization ability) or parameter redundancy (low computational efficiency), enabling the model to operate stably in complex tunneling environments.

[0022] Furthermore, the dynamic threshold model construction method described in step S3 includes:

[0023] Based on the quantile distribution of historical monitoring data, and combined with a Bayesian network to dynamically update the threshold, the formula is:

[0024] T = α·Q 95％ +(1-α)·EMA(T prev );

[0025] Among them, Q 95％ The EMA(T) is the 95th percentile of historical data. prev The threshold value is the exponential moving average of the threshold value at the previous time step, and α is the adaptive weighting coefficient. This formula is used to solve the problem of "poor adaptability of fixed thresholds" in traditional early warning systems. Tunnel deformation is significantly affected by geological conditions (such as soft rock / hard rock) and construction stage (excavation / operation). Fixed thresholds are prone to false alarms (such as thresholds that are too lenient for hard rock tunnels) or false alarms (such as thresholds that are too strict for soft rock tunnels).

[0026] Among them, Q 95％ Ensure that thresholds are based on historical security data to avoid being detached from reality; EMA(T) prev To ensure that the threshold updates smoothly over time and avoid sudden changes, α is dynamically adjusted through a Bayesian network based on real-time working conditions (such as changes in the surrounding rock level) to make the threshold more consistent with the current tunnel condition. For example, α is increased in soft rock formations to improve the threshold sensitivity.

[0027] Furthermore, the multi-scale convolution kernel configuration described in step S2 is as follows:

[0028] Convolutional kernels of different sizes are deployed in parallel in the CNN layer to extract local, meso- and macro-level spatial features respectively, and multi-scale feature maps are generated by channel splicing.

[0029] This addresses the problem in traditional spatial feature extraction where "single-scale convolution is insufficient to cover the complex spatial structure of tunnels." Tunnel deformation exhibits multi-scale features (such as local abrupt changes at a single monitoring point, cooperative deformation of adjacent points, and overall trends across the entire cross-section). Multi-scale convolution kernels utilize:

[0030] Small cores (such as 3×3) capture local details (such as deformation around the vault);

[0031] Medium-sized kernels (e.g., 5×5) capture mesoscopic correlations (e.g., the coordinated deformation of sidewalls and arch waists);

[0032] Large-size cores (such as 7×7) capture macroscopic trends (such as the tendency of settlement throughout the tunnel);

[0033] Ultimately, by splicing and fusing multi-scale information through channels, the spatial distribution pattern of deformation is comprehensively depicted, providing spatial feature support for accurate prediction.

[0034] Furthermore, the deviation calculation method described in step S3 is as follows:

[0035]

[0036] in, Let T be the predicted deformation value for the j-th monitoring point. j is the dynamic threshold for the corresponding monitoring point, and m is the total number of monitoring points.

[0037] Furthermore, the multi-level early warning signals mentioned in step S4 include:

[0038] Level 1 Warning: When the deviation D is in the range of [0.8, 1.0), it indicates that the monitoring frequency needs to be increased.

[0039] Level II warning: When D is in the range [1.0, 1.2), the temporary support and reinforcement plan shall be activated;

[0040] Level 3 warning: When D≥1.2, an emergency evacuation order for personnel and a work stoppage order are triggered.

[0041] Furthermore, the method for constructing the spatiotemporal sequence dataset described in step S2 is as follows:

[0042] The time series data is segmented into fixed time steps and reconstructed into a three-dimensional tensor with the sensor data at the corresponding spatial location to form a spatiotemporally coupled input matrix.

[0043] A real-time early warning system for tunnel deformation based on an LSTM-CNN model includes:

[0044] Data acquisition module: Composed of distributed fiber optic grating sensors, inclinometers and temperature and humidity sensors, used to collect multi-dimensional monitoring data of the tunnel structure in real time;

[0045] Edge computing module: integrates FPGA chip to realize real-time noise reduction and spatiotemporal sequence reconstruction of data;

[0046] Cloud-based analytics platform: Deploys LSTM-CNN fusion models and dynamic thresholding algorithms to complete deformation prediction and risk assessment;

[0047] Intelligent terminal module: Receives early warning signals via mobile communication and displays the tunnel deformation status using AR augmented reality technology.

[0048] Furthermore, the data acquisition module and the edge computing module adopt a dual-channel redundant transmission architecture, wherein the main channel transmits real-time monitoring data through optical fiber, and the backup channel transmits key feature parameters synchronously through LoRa wireless technology; when the data transmission delay of the main channel exceeds a preset threshold, the edge computing module automatically switches to the backup channel and triggers a dynamic adjustment mechanism for the sensor sampling frequency to balance data integrity and transmission efficiency.

[0049] The beneficial effects of this invention are:

[0050] (1) The temporal trend and spatial distribution characteristics of tunnel deformation are extracted by LSTM-CNN fusion model, and local, meso and macro spatial features are captured by multi-scale convolution kernels, which realizes the deep characterization of deformation law and significantly improves the accuracy and reliability of tunnel deformation prediction.

[0051] (2) By combining the dynamic threshold model with historical data quantiles and Bayesian network to adaptively update the threshold, and with the deviation quantification evaluation mechanism, the problem of poor adaptability of fixed threshold under complex geological conditions is solved, effectively reducing false alarms and missed alarms in the early warning process and improving the accuracy of risk assessment.

[0052] (3) Through the collaborative work of the dual-channel redundant transmission architecture, edge computing module, and cloud analysis platform, the real-time acquisition, transmission and processing of monitoring data are realized, ensuring the continuity of data transmission in extreme environments. At the same time, combined with AR visualization and hierarchical early warning mechanism, the system response efficiency and decision support capabilities are improved. Attached Figure Description

[0053] Figure 1 A flowchart illustrating the specific steps of a real-time early warning method for tunnel deformation based on an LSTM-CNN model, provided in this embodiment of the invention.

[0054] Figure 2 This is a specific structural diagram of the CNN model in an LSTM-CNN fusion model provided in an embodiment of the present invention. Detailed Implementation

[0055] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0056] Example 1

[0057] participate Figure 1 This paper presents a real-time early warning method for tunnel deformation based on an LSTM-CNN model, which includes the following steps:

[0058] S1. Data Acquisition: Real-time collection of multi-dimensional monitoring data of the tunnel structure through a distributed sensor network, including but not limited to crown settlement, horizontal convergence, surrounding rock stress and environmental temperature and humidity data, and simultaneous recording of spatiotemporal labels for each data point;

[0059] S2. Spatiotemporal feature fusion processing: Standardize and denoise the original data to construct a spatiotemporal sequence dataset;

[0060] An LSTM-CNN fusion model is used to extract features from spatiotemporal sequence data. The LSTM layer captures the deformation trend in the time dimension, while the CNN layer extracts the deformation distribution features in the spatial dimension through multi-scale convolutional kernels.

[0061] S3. Deformation Prediction and Risk Assessment: Based on the fused spatiotemporal characteristics, the model outputs the predicted deformation value within the future time window; a dynamic threshold model is constructed by combining historical safety data to calculate the deviation between the predicted value and the threshold.

[0062] S4. Intelligent Early Warning and Decision Support: Triggers multi-level early warning signals based on the deviation level, and displays the deformation trend and risk area distribution through a visual interface.

[0063] The standardized noise reduction process described in step S2 includes:

[0064] S2.1 Normalizes the original data using the following formula:

[0065]

[0066] Where, x i Here are the original data points, μ is the data mean, and σ is the data standard deviation;

[0067] S2.2 uses a sliding window method to interpolate and repair abnormal data points, with the window length adaptively adjusted according to the data fluctuation characteristics.

[0068] The LSTM-CNN fusion model training process described in step S2 includes:

[0069] S2.1 uses the cross-entropy loss function to optimize the model parameters, as shown in the formula:

[0070]

[0071] Among them, y i For real labels, To predict the probability, n is the number of samples;

[0072] S2.2 uses the particle swarm optimization (PSO) algorithm to dynamically adjust the forget gate weights of the LSTM layer and the number of convolutional kernels of the CNN layer in order to balance the generalization ability and computational efficiency of the model.

[0073] The dynamic threshold model construction method described in step S3 includes:

[0074] Based on the quantile distribution of historical monitoring data, and combined with a Bayesian network to dynamically update the threshold, the formula is:

[0075] T = α·Q 95％ +(1-α)·EMA(T prev );

[0076] Among them, Q 95％ The EMA(T) is the 95th percentile of historical data. prev ) is the exponential moving average of the threshold at the previous time step, and α is the adaptive weighting coefficient.

[0077] The multi-scale convolution kernel configuration mentioned in step S2 is as follows:

[0078] Convolutional kernels of different sizes are deployed in parallel in the CNN layer to extract local, meso- and macro-level spatial features respectively, and multi-scale feature maps are generated by channel splicing.

[0079] The deviation calculation method described in step S3 is as follows:

[0080]

[0081] in, Let T be the predicted deformation value for the j-th monitoring point. j is the dynamic threshold for the corresponding monitoring point, and m is the total number of monitoring points.

[0082] The multi-level early warning signals mentioned in step S4 include:

[0083] Level 1 Warning: When the deviation D is in the range of [0.8, 1.0), it indicates that the monitoring frequency needs to be increased.

[0084] Level II warning: When D is in the range [1.0, 1.2), the temporary support and reinforcement plan shall be activated;

[0085] Level 3 warning: When D≥1.2, an emergency evacuation order for personnel and a work stoppage order are triggered.

[0086] The method for constructing the spatiotemporal sequence dataset described in step S2 is as follows:

[0087] The time series data is segmented into fixed time steps and reconstructed into a three-dimensional tensor with the sensor data at the corresponding spatial location to form a spatiotemporally coupled input matrix.

[0088] A real-time early warning system for tunnel deformation based on an LSTM-CNN model includes:

[0089] Data acquisition module: Composed of distributed fiber optic grating sensors, inclinometers and temperature and humidity sensors, used to collect multi-dimensional monitoring data of the tunnel structure in real time;

[0090] Edge computing module: integrates FPGA chip to realize real-time noise reduction and spatiotemporal sequence reconstruction of data;

[0091] Cloud-based analytics platform: Deploys LSTM-CNN fusion models and dynamic thresholding algorithms to complete deformation prediction and risk assessment;

[0092] Intelligent terminal module: Receives early warning signals via mobile communication and displays the tunnel deformation status using AR augmented reality technology.

[0093] The data acquisition module and the edge computing module adopt a dual-channel redundant transmission architecture. The main channel transmits real-time monitoring data through optical fiber, while the backup channel transmits key feature parameters synchronously through LoRa wireless technology. When the data transmission delay of the main channel exceeds a preset threshold, the edge computing module automatically switches to the backup channel and triggers a dynamic adjustment mechanism for the sensor sampling frequency to balance data integrity and transmission efficiency.

[0094] Example 2

[0095] A real-time early warning method for tunnel deformation based on the LSTM-CNN model and a corresponding system setup are provided. The specific implementation steps are as follows:

[0096] S1. Data Acquisition:

[0097] This step relies on the "data acquisition module" and related transmission architecture. Through a distributed sensor network and a dual-channel transmission mechanism, it enables real-time and reliable acquisition of multi-dimensional tunnel data, laying the foundation for subsequent analysis.

[0098] S1.1 Sensor Network Selection and Deployment: The selection of sensors for the data acquisition module must be adapted to the complex tunnel environment, based on the following criteria:

[0099] Fiber Bragg grating sensors are chosen for their electromagnetic interference resistance, corrosion resistance, and distributed measurement characteristics, making them suitable for environments inside tunnels that are humid, dusty, and subject to strong electromagnetic interference (such as tunnel boring machine operating areas). Their micron-level measurement accuracy can precisely capture minute deformations such as arch settlement and horizontal convergence, and a single optical fiber can connect multiple sensing points in series, enabling long-distance distributed monitoring.

[0100] The inclinometer employs a high-precision MEMS inclinometer, deployed on the surface of the support structure. By monitoring changes in the structure's tilt angle, it indirectly reflects the overall deformation trend of the tunnel cross-section. Its low power consumption makes it suitable for long-term burial, and its vibration resistance meets the dynamic monitoring requirements during construction.

[0101] The temperature and humidity sensor is an explosion-proof type, which is suitable for high temperature or high humidity environments that may exist in the tunnel (such as when passing through water-rich strata), ensuring the stability of environmental parameter acquisition and providing data support for analyzing the impact of environmental factors on deformation.

[0102] The density and location of the sensors are dynamically adjusted according to the mechanical properties of the tunnel structure.

[0103] In high-risk areas such as fault fracture zones and weak interlayers, the density of sensor deployment is increased (e.g., one set every 5 meters) to cover key parts such as the arch, side walls, and invert arch, forming a "three-dimensional monitoring network".

[0104] In homogeneous hard rock sections, the deployment density can be appropriately reduced (e.g., one set every 10-15 meters), with a focus on monitoring the stress concentration areas of the arch and sidewalls, taking into account both monitoring costs and risk control.

[0105] All sensors are fixed to the surface of the tunnel structure by anti-corrosion brackets. The brackets are flexibly connected to the structure to avoid positional displacement caused by vibration (such as vehicle traffic or construction disturbance) and ensure the accuracy of the spatial tags.

[0106] S1.2 Real-time acquisition of multidimensional data:

[0107] The sensor network collects data in real time at a dynamic sampling frequency (adjusted according to the risk level), specifically including:

[0108] Structural deformation data, such as crown settlement, horizontal convergence value, and clearance change, reflect changes in the tunnel's geometry.

[0109] Mechanical characteristic data: surrounding rock stress, internal forces of support structure, axial force of anchor bolts, etc., characterizing the stress state of the structure;

[0110] Environmental impact data: Temperature, humidity, wind speed, etc. inside the tunnel, used to analyze the potential impact of environmental factors on deformation;

[0111] During the data acquisition process, the data acquisition terminal performs preliminary verification on the raw data (such as data format and transmission integrity verification), removes obviously invalid data (such as negative settlement values), and ensures that the data entering subsequent processing has basic reliability.

[0112] S1.3 Spatiotemporal Tag Synchronization Record:

[0113] Each piece of collected data is synchronously bound to a spatiotemporal tag, providing a coordinate reference for subsequent spatiotemporal feature analysis:

[0114] Time stamping: Time synchronization is achieved through GPS timing or BeiDou positioning base stations deployed in the tunnel. The timestamps are accurate to the millisecond level, ensuring the time consistency of data from different sensors, with the error controlled within 10 milliseconds.

[0115] Spatial Labels: Based on the three-dimensional spatial positioning information of tunnel mileage station and cross-section coordinates (such as "K12+345.6-ring number 23-arch top"), the location of the monitoring point corresponding to the data is identified and associated with the tunnel BIM model to provide a spatial reference for AR visualization;

[0116] Spatiotemporal tags and monitoring data are bound together using a hash algorithm and stored in the local cache of the edge computing module to ensure data traceability and provide reliable historical samples for model iteration.

[0117] S1.4 Dual-channel redundant transmission: The data acquisition module and the edge computing module adopt a dual-channel redundant transmission architecture to ensure the continuity of data transmission.

[0118] Main channel (fiber optic transmission): Single-mode fiber is used as the main transmission link, with a transmission rate of no less than 1Gbps, supporting real-time transmission of all raw data (including deformation, stress, environmental data, and spatiotemporal tags). The fiber optic route is laid along the tunnel sidewall to avoid areas affected by construction disturbance, and fiber optic connector protection boxes are installed every 100 meters to reduce the risk of physical damage.

[0119] Backup Channel (LoRa Wireless Transmission): Utilizing the LoRaWAN protocol as a backup link, operating in the ISM unlicensed frequency band, with a transmission distance of several kilometers, it supports low-power transmission of key characteristic parameters (such as peak distortion, average distortion, and anomaly markers). A relay node is deployed every 200 meters within the tunnel to ensure complete signal coverage and prevent transmission interruptions due to obstruction.

[0120] Transmission priority management: The main channel prioritizes the transmission of all data, while the backup channel transmits key characteristic parameters synchronously, forming a redundant transmission mode of "full data + core data", which ensures data integrity while reducing transmission pressure.

[0121] S1.5 Transmission Adaptive Switching Mechanism: The edge computing module monitors the status of the main channel in real time to achieve intelligent switching of the transmission link.

[0122] Delay monitoring: The main channel transmission delay (the difference between the data sending time and the receiving acknowledgment time) is calculated through a "send-response" mechanism. When the delay exceeds the preset threshold for three consecutive times, it is determined that the main channel is faulty.

[0123] Automatic switching: When the main channel fails, the edge computing module immediately switches to the backup channel and triggers the compression and retransmission of locally cached data (preserving key features) to ensure the data continuity of the cloud analysis platform;

[0124] Recovery Switching: After the main channel is restored, it automatically switches back to fiber optic transmission, raises the sampling frequency to the default value, and restores full data transmission;

[0125] Dynamic adjustment of sampling frequency: When switching to the backup channel, the sampling frequency of non-critical data (such as environmental data) is automatically reduced according to the channel bandwidth, giving priority to the transmission of data such as deformation and stress, and balancing data integrity and transmission efficiency.

[0126] S2. Spatiotemporal feature fusion processing:

[0127] This step is achieved through the collaboration of the "edge computing module" and the "cloud analysis platform," including data preprocessing, spatiotemporal sequence construction, LSTM-CNN model feature extraction and training optimization, in order to mine the spatiotemporal correlation features of tunnel deformation.

[0128] S2.1 Data Standardization and Noise Reduction: The raw data needs to be standardized and denoised to improve its quality and provide a reliable foundation for model input.

[0129] S2.1.1 Data cleaning and anomaly identification: Before normalization, valid data is screened through multi-dimensional verification;

[0130] Based on the mechanical properties of the tunnel structure, reasonable ranges for each parameter are set (e.g., the daily change in arch settlement should not exceed 1 / 5 of the allowable value of the structural design), and data exceeding the range are marked as suspicious values.

[0131] Compare the deformation and stress data at the same monitoring point (e.g., stress should increase synchronously when settlement increases), and mark any contradictory trends as outliers;

[0132] Examine the gradient of data changes between adjacent timestamps, and mark a mutation value when the rate of change exceeds three times the historical average gradient;

[0133] Abnormal data is temporarily stored in the abnormal database of the edge computing module, pending subsequent interpolation repair or manual review.

[0134] S2.1.2 Data Normalization and Interpolation Repair:

[0135] Normalization: Normalize the original data that has passed the verification to eliminate the impact of differences in units on model training and map the data to the [-1,1] interval, which facilitates cross-feature learning of LSTM and CNN.

[0136] Outlier interpolation repair: A sliding window method is used to repair outlier data points, with the window length adaptively adjusted according to the data type.

[0137] High-frequency fluctuation data (such as temperature and humidity) are captured using a short window (5-10 sampling points) and instantaneous changes are preserved through linear interpolation.

[0138] Low-frequency trend data (such as arch settlement) are processed using a long window (20-30 sampling points) and noise is smoothed by cubic spline interpolation.

[0139] The repaired data needs to pass a consistency check again to ensure it conforms to physical laws and avoids introducing new errors.

[0140] S2.2 Spatiotemporal Sequence Dataset Construction: The preprocessed data needs to be converted into a spatiotemporally coupled input format for use by the LSTM-CN N model.

[0141] S2.2.1 Time dimension segmentation, the selection of time step is based on the time scale characteristics of tunnel deformation:

[0142] Short-term forecast window: used for early warning within 12 hours, with a time step of 1 hour. Each window contains 24 historical data points (from the previous 24 hours) to capture short-term fluctuations.

[0143] Mid-term forecast window: used for 1-7 day trend analysis, with a time step of 6 hours. Each window contains 28 historical data points (the previous 7 days), balancing details and trends.

[0144] An overlapping sliding method is used for segmentation (overlap rate 50%) to ensure the continuity of the time series and avoid information loss.

[0145] S2.2.2 Spatial Dimension Correlation and Tensor Reconstruction:

[0146] The monitoring point data of the same cross section are sorted according to the circumferential distribution (arch top → left arch waist → left side wall → inverted arch → right side wall → right arch waist) to form a cross section circumferential feature vector; the monitoring point data of the same type of adjacent 5 rings are sorted according to mileage to form a longitudinal trend vector.

[0147] Based on the dynamic selection of tunnel stages (stress and settlement channels are retained during construction, and vibration and temperature and humidity channels are added during operation), redundant features are reduced.

[0148] By integrating time window, spatial distribution and feature channel data, a three-dimensional tensor of "time step × spatial dimension × feature channel" (such as "24 steps × 32 monitoring points × 5 features") is formed to achieve coupled storage of spatiotemporal information.

[0149] S2.3 LSTM-CNN Fusion Model Feature Extraction: The LSTM-CNN model deployed on the cloud analysis platform achieves a deep characterization of deformation patterns by collaboratively extracting spatiotemporal features.

[0150] S2.3.1 The LSTM layer extracts temporal features. The LSTM layer captures the deformation trend of the time dimension through a gating mechanism:

[0151] Forget Gate: Set a relatively large initial weight (0.8-1.0) to prioritize the preservation of long-term trends (such as the creep characteristics of soft rock tunnels) and avoid interference from short-term fluctuations;

[0152] Input gate: Assign high weights to abrupt change data (such as stress surge) to enhance feature learning at key time points;

[0153] Output gate: Integrates historical and current features, and outputs a deformation trend vector in the time dimension (such as the change in settlement rate in the next 24 hours);

[0154] Configure a dropout layer (dropout rate 0.2-0.3) to prevent overfitting, and use a truncated backpropagation algorithm to avoid gradient vanishing and improve training stability.

[0155] S2.3.2 CNN Layer Extraction of Spatial Features: The CNN layer extracts spatial features in parallel through multi-scale convolutional kernels, achieving feature abstraction from local to macroscopic levels.

[0156] First convolutional layer: Deploy 3×3 and 5×5 small-sized convolutional kernels (32 of each) to extract local details (such as deformation abrupt changes at a single monitoring point);

[0157] The second convolutional layer deploys 7×7 and 9×9 medium-sized convolutional kernels (64 of each) to capture mesoscopic correlations (such as the cooperative deformation of 3-5 adjacent monitoring points);

[0158] The third convolutional layer: deploys 11×11 large-size convolutional kernels (128 kernels) to extract macroscopic trends (such as the overall settlement tendency of the entire tunnel);

[0159] Each convolutional layer is followed by a 2×2 max pooling layer to reduce the feature map size through downsampling; finally, multi-scale features are fused by channel concatenation and then compressed by 1×1 convolution to improve model efficiency.

[0160] S2.4LSTM-CNN Model Training Optimization: Model training improves prediction accuracy and generalization ability through loss function optimization and parameter tuning.

[0161] S2.4.1 Training data partitioning and loss function optimization include:

[0162] A time-series splitting method is adopted (70% training set, 20% validation set, and 10% test set) to avoid "future information leakage"; the dataset is dynamically updated every month, incorporating new data and removing old data to ensure that the model adapts to the latest deformation features;

[0163] The cross-entropy loss function is used to measure the deviation between the predicted and the true values. The loss is minimized by the Adam gradient descent algorithm (learning rate 0.001-0.0001, decaying with iteration), and the model weights are updated iteratively.

[0164] S2.4.2 Particle Swarm Optimization (PSO) Parameter Tuning: Key parameters are dynamically optimized using PSO to balance model performance and efficiency.

[0165] LSTM forget gate weights (affect the strength of temporal feature retention), CNN convolutional kernel number (affects the spatial feature dimension);

[0166] Initialize the particle swarm (each group of particles represents a set of parameter combinations), and find the optimal parameters by iteratively updating the particle positions and velocities, with the goal of minimizing the loss function;

[0167] This enables the model to maintain stable feature extraction capabilities under complex geological conditions (such as fault zones) while reducing cloud computing load.

[0168] S3. Deformation Prediction and Risk Assessment:

[0169] This step, based on the prediction results output by the fusion model, combines the deviation calculated by the dynamic threshold model to achieve a quantitative assessment of risk and provide a basis for early warning decisions.

[0170] S3.1 Deformation Prediction Output: The model outputs multi-dimensional prediction results to meet the needs of different decision-making scenarios.

[0171] S3.1.1 Prediction Window and Content Design:

[0172] Real-time forecast: Outputs deformation values ​​for the next hour for emergency risk assessment;

[0173] Short-term forecast: Outputs the deformation trend for the next 12-24 hours, including settlement, convergence value, deformation rate, etc.

[0174] Mid-term forecast: Outputs the cumulative deformation over the next 7 days for adjustment of support scheme;

[0175] Each predicted value is accompanied by a 95% confidence interval to reflect uncertainty (the confidence interval is wider in geologically complex areas) to help determine the reliability of the risk.

[0176] S3.1.2 Prediction result verification mechanism:

[0177] The model is used to make "pseudo-predictions" on known historical data. The deviation between the predicted value and the actual value is compared. When the deviation exceeds a preset threshold, the model is retrained.

[0178] Based on the structural bearing capacity limit, determine whether the predicted value exceeds the reasonable range (e.g., settlement exceeds the design allowable value). If it does, mark it as high risk.

[0179] Check the consistency between the predicted trend and the historical trend. When abrupt changes occur (such as a sudden change from subsidence to uplift), investigate the cause in conjunction with geological data to avoid model misjudgment.

[0180] S3.2 Dynamic Threshold Model Construction: The dynamic threshold model combines historical data with real-time operating conditions for adaptive updates, ensuring the rationality of the threshold.

[0181] S3.2.1 Calculation of quantiles for historical data:

[0182] The 95th percentile of historical data was calculated using a 3-month sliding window to reflect recent deformation characteristics;

[0183] Give high weight to recent data (e.g., 0.7 for data from the past month and 0.3 for data from the past 1-3 months) to avoid outdated data affecting the threshold.

[0184] Outliers caused by sensor malfunctions or construction disturbances are removed to ensure the representativeness of the quantiles.

[0185] S3.2.2 Bayesian network threshold optimization: Bayesian network dynamically adjusts the threshold by fusing multi-source information.

[0186] Network structure: The network uses "geological conditions (surrounding rock grade, water content)", "construction stage" and "environmental factors" as input nodes and "threshold adjustment coefficient α" as output nodes.

[0187] Parameter update: For each new batch of data, the network parameters are updated by maximum likelihood estimation to make α fit the current working conditions (e.g., α increases when the surrounding rock grade decreases, thus improving threshold sensitivity).

[0188] Emergency calibration: In the event of special conditions such as sudden water flow or earthquake, the emergency mode is automatically triggered, temporarily increasing the α value to make the threshold more conservative and ensure safety.

[0189] S3.3 Deviation Calculation and Spatial Aggregation: Deviation measures the degree of deviation between the predicted deformation value and the threshold, providing a basis for the classification of early warning levels.

[0190] S3.3.1 Single Monitoring Point Deviation Calculation: For different monitoring points with varying structural importance (e.g., the crown has a higher weight than the sidewalls), the dynamic threshold T is adjusted accordingly.j Multiply by an importance coefficient; recent forecasts are given higher weight (e.g., the weight of the last 3 hours is 1.5 times that of the earlier ones) to highlight the impact of short-term trends.

[0191] S3.3.2 Regional Deviation Aggregation Analysis: Kriging interpolation is used to generate a full tunnel heat map based on the deviation of monitoring points, filling the gaps between monitoring points; regions are divided into 50-meter intervals, and the regional average and maximum deviations are calculated to identify high-risk areas; combined with the structural mechanics transmission path, the potential impact of high-risk areas on adjacent areas (such as whether arch deformation will trigger a chain reaction in the sidewalls) is analyzed, providing a comprehensive basis for early warning response.

[0192] S4. Intelligent Early Warning and Decision Support:

[0193] This step utilizes the "intelligent terminal module" to achieve multi-level early warning and visual decision support, ensuring timely and accurate risk management.

[0194] S4.1 Multi-level Early Warning Signal and Response Mechanism: Triggers a three-level early warning based on the numerical range of deviation D, with a corresponding progressive response process:

[0195] S4.1.1 Level 1 Warning (D∈[0.8,1.0), Level of Attention Risk):

[0196] The instruction data acquisition module increases the sampling frequency of the corresponding area (e.g., from 1 time / minute to 1 time / 10 seconds) and starts video monitoring to observe the scene in real time;

[0197] The edge computing module generates a deformation trend report every hour, including "recent rate", "comparison with history", and "analysis of influencing factors", which is then pushed to the administrator's terminal.

[0198] The system prompts monitoring personnel to check the sensor status on-site, eliminate false alarms caused by equipment malfunctions, and input the verification results into the system for model optimization.

[0199] S4.1.2 Level II Early Warning (D∈[1.0,1.2), Intervention-level Risk):

[0200] The cloud platform automatically retrieves the "temporary support plan library" and recommends suitable solutions (such as anchor bolt type, spacing, and shotcrete strength), and marks "applicable geology", "construction period" and "resource requirements";

[0201] Connect to the resource management platform to query available materials, equipment, and personnel information, and generate scheduling suggestions;

[0202] Real-time monitoring of deformation data after reinforcement measures are implemented, evaluation of effectiveness, and triggering of contingency plan upgrades if unsatisfactory.

[0203] S4.1.3 Level 3 Early Warning (D≥1.2, Emergency Level Risk):

[0204] Evacuation instructions are issued through tunnel sound and light alarms and broadcasts, and smart terminals display the optimal route (avoiding high-risk areas) and count the number of people evacuated in real time.

[0205] Push early warning information to emergency response teams, property owners, and construction units, and share data, videos, and contingency plans;

[0206] Activate drone inspections and portable monitoring devices to supplement fixed sensor data and fully grasp the deformation dynamics;

[0207] Once the danger has passed, the structural safety must be assessed by experts, and the warning can only be lifted if the conditions are met.

[0208] S4.2 Visualization and Decision Support: Smart terminals use AR technology to visualize data and assist in efficient decision-making.

[0209] S4.2.1AR 3D Model Dynamic Display:

[0210] The tunnel BIM model is integrated with real-time deformation data, and the AR interface displays the amount of deformation with color gradients (non-warning colors) (blue for slight deformation, red for severe deformation).

[0211] Simulate the deformation and evolution process over the next 24 hours to visually demonstrate the trend of risk spread;

[0212] It supports arbitrary cross-sectional cutting to display deformation distribution and stress state, helping to identify weak points in the structure.

[0213] S4.2.2 Decision Support Knowledge Base Linkage:

[0214] Search the historical case database, match cases with similar risk characteristics, and push disposal plans and effect evaluations;

[0215] Built-in tools quickly calculate support parameters (such as anchor bolt bearing capacity and concrete thickness) to provide a reference for scheme design;

[0216] Automatically record the entire process of early warning response (instruction time, measures, and responsible person) to form a traceable log for post-event review and process optimization.

[0217] This embodiment comprehensively covers the relevant technical features, and through the synergy of methods and systems, forms a complete real-time early warning solution for tunnel deformation. Its advantages are reflected in:

[0218] LSTM and CNN work together to capture the temporal trend and spatial correlation of deformation, significantly improving prediction accuracy compared to traditional models; combining historical data with real-time operating conditions to adjust thresholds solves the adaptability problem of fixed thresholds under complex geological conditions, reducing false alarms and missed alarms; dual-channel architecture and adaptive switching mechanism ensure uninterrupted data transmission in extreme environments, achieving 99.9% system availability; progressive response from awareness to emergency response, combined with AR visualization and knowledge base support, shortens risk handling time by more than 30% and makes resource allocation more rational.

[0219] Example 3

[0220] See Figure 2 This paper presents a specific structure of the CNN model in an LSTM-CNN fusion model. This model is applied in a real-time early warning system for tunnel deformation, employing multi-scale convolution and feature fusion. Its specific structure is as follows:

[0221] Input layer:

[0222] The input layer receives a preprocessed three-dimensional spatiotemporal feature tensor, the dimension of which is deeply bound to the spatial attributes of the tunnel monitoring data. The tensor dimension is "H×W×C", where:

[0223] H represents the spatial height dimension, corresponding to the distribution pattern of longitudinal monitoring points in the tunnel. For example, when a monitoring section is set up every 5 meters along the tunnel mileage direction, the value of H is determined by the total number of monitoring sections, ensuring longitudinal coverage of the entire tunnel length;

[0224] W represents the spatial width dimension, corresponding to the distribution of lateral monitoring points within a single monitoring section. Taking a standard circular tunnel as an example, W can be divided in a circumferential sequence of "arch crown → left arch waist → left side wall → invert arch → right side wall → right arch waist," accurately reflecting the deformation differences at different locations within the section.

[0225] C represents the feature channel dimension, integrating multiple types of monitoring data. In practical applications, C can include channels such as crown settlement, horizontal convergence, surrounding rock stress, temperature and humidity, etc. Each channel corresponds to the time series data of a type of physical quantity, realizing the spatial correlation of multi-source information.

[0226] The input layer transforms scattered monitoring data into structured spatial tensors, providing a unified input format for subsequent convolutional operations. During implementation, the tensor needs to be reconstructed in real-time via an edge computing module to ensure that the data dimensions match the convolutional layer parameters, avoiding feature extraction biases caused by dimensionality incompatibility.

[0227] Multi-scale convolutional branch layers:

[0228] The multi-scale convolutional branch layer employs a three-parallel branch design, using convolutional kernels of different sizes to achieve hierarchical extraction of spatial features from local to macroscopic levels, adapting to the feature differences of tunnel deformation at different spatial scales:

[0229] Local Feature Extraction Branch: This branch focuses on subtle deformation features around a single monitoring point, using a convolutional kernel of size a×a (e.g., 3×3) with a quantity of N1. In tunnel monitoring, local features correspond to the deformation gradient around a single sensor, such as the correlation between a monitoring point on the arch and the stress change within a 50cm radius. The convolutional layer iterates through the input tensor using a sliding window, calculating the feature response within the local region to capture the spatial distribution patterns of subtle deformations. Subsequent activation layers use the ReLU function to enhance feature discriminative power and filter out invalid noise through nonlinear transformation. The pooling layer uses 2×2 max pooling to retain the maximum feature value within the local region, strengthening key deformation signals while reducing data volume, and outputting a local feature map with dimensions H1×W1×N1.

[0230] Mesoscopic Feature Extraction Branch: This branch targets the collaborative deformation features of adjacent monitoring point groups, employing a b×b convolutional kernel (b>a, e.g., 5×5), with the number of kernels set to N². Mesoscopic features correspond to the deformation correlation within a specific region of the tunnel cross-section, such as the synchronous convergence trend of the arch waist and sidewall in the left half of the cross-section. The convolutional layer expands the receptive field to cover multiple adjacent monitoring points, calculating the feature correlation within the region. The activation layer also uses the ReLU function to enhance the nonlinear feature representation. The pooling layer compresses the data dimension through 2×2 max pooling, outputting a mesoscopic feature map of H²×W²×N², focusing on preserving the collaborative deformation patterns within the region.

[0231] Macroscopic Feature Extraction Branch: This branch focuses on the overall deformation trend of the entire tunnel or multiple combined sections, using c×c convolutional kernels (c > b, e.g., 7×7) with a quantity of N3. Macroscopic features correspond to the deformation tendencies of the entire tunnel cross-section or multiple mileage sections, such as the difference in settlement rates between the inlet and outlet sections. The convolutional layers cover a large spatial area through the maximum receptive field, capturing the deformation transmission patterns across regions; the configuration of the activation and pooling layers is consistent with the previous two branches, outputting a macroscopic feature map of H3×W3×N3, reflecting the overall spatial deformation trend of the tunnel.

[0232] The parallel design of the three branches ensures full-scale spatial feature coverage from millimeter-level local deformation to hundred-meter-level macroscopic trends, solving the problem that single-scale convolution is difficult to adapt to the complex spatial structure of tunnels.

[0233] Feature splicing layer:

[0234] The feature concatenation layer, located downstream of the three convolutional branches, is responsible for merging local, meso-level, and macro-level feature maps along the channel dimension. Specifically, the local feature map (H1×W1×N1), meso-level feature map (H2×W2×N2), and macro-level feature map (H3×W3×N3) are aligned according to their spatial dimensions and then concatenated along the feature channel direction to form a fused multi-scale feature map (Hf×Wf×(N1+N2+N3)). Here, Hf and Wf represent the spatial dimensions of the concatenated map, and an automatic alignment mechanism ensures that the spatial coordinates of different branches match.

[0235] This process addresses the limitations of single-scale features: for example, when a sudden increase in local deformation occurs in a region, local features can accurately pinpoint the abrupt change, meso-scale features can reflect the impact of this change on adjacent monitoring points, and macro-scale features can determine whether there is a risk of cascading deformation across the entire tunnel. The fusion of multi-scale features enables the model to identify both subtle deformation signals and grasp the overall trend, providing comprehensive spatial evidence for subsequent risk assessment.

[0236] Output layer:

[0237] The output layer receives the stitched multi-scale feature map and transforms it into a one-dimensional feature vector through a flattening operation. Specifically, the three-dimensional feature map (Hf×Wf×(N1+N2+N3)) is flattened into a vector of length Hf×Wf×(N1+N2+N3) according to the row-major principle. This vector contains full-scale spatial feature information of tunnel deformation.

[0238] In the entire LSTM-CNN fusion model, this one-dimensional vector is concatenated with the temporal feature vector extracted by the LSTM model to form a spatiotemporal fusion feature. For example, the "sudden increase in local deformation of the arch" signal in the spatial feature vector will be combined with the "accelerated settlement rate in the past 3 hours" signal in the temporal feature vector to jointly support deformation prediction and risk assessment, ultimately improving the accuracy and timeliness of tunnel deformation early warning.

[0239] Through the hierarchical design of the above model, the CNN model can achieve the following effects in tunnel deformation monitoring: First, multi-scale convolution solves the problem of feature capture in different spatial ranges of the tunnel; second, feature stitching enables collaborative analysis of local and global features; and third, fusion with LSTM provides a foundation for spatiotemporal correlation modeling. In practical applications, this model can effectively identify different types of spatial features such as concentrated local deformation around fault zones and uniform settlement across the entire cross-section, providing refined spatial dimension support for tunnel safety monitoring and significantly reducing the risk of false alarms or missed alarms due to the omission of spatial features.

[0240] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A real-time early warning method for tunnel deformation based on an LSTM-CNN model, characterized in that, Includes the following steps: S1. Data Acquisition: Real-time collection of multi-dimensional monitoring data of the tunnel structure through a distributed sensor network, including but not limited to crown settlement, horizontal convergence, surrounding rock stress and environmental temperature and humidity data, and simultaneous recording of spatiotemporal labels for each data point; S2. Spatiotemporal feature fusion processing: Standardize and denoise the original data to construct a spatiotemporal sequence dataset; An LSTM-CNN fusion model is used to extract features from spatiotemporal sequence data. The LSTM layer captures the deformation trend in the time dimension, while the CNN layer extracts the deformation distribution features in the spatial dimension through multi-scale convolutional kernels. S3. Deformation Prediction and Risk Assessment: Based on the fused spatiotemporal characteristics, the model outputs the predicted deformation values ​​within the future time window; A dynamic threshold model is constructed by combining historical security data to calculate the deviation between the predicted value and the threshold. S4. Intelligent Early Warning and Decision Support: Triggers multi-level early warning signals based on the deviation level, and displays the deformation trend and risk area distribution through a visual interface.

2. The method according to claim 1, characterized in that, The standardized noise reduction process described in step S2 includes: S2.1 Normalizes the original data using the following formula: Where, x i Here are the original data points, μ is the data mean, and σ is the data standard deviation; S2.2 uses a sliding window method to interpolate and repair abnormal data points, with the window length adaptively adjusted according to the data fluctuation characteristics.

3. The method according to claim 1, characterized in that, The LSTM-CNN fusion model training process described in step S2 includes: S2.1 uses the cross-entropy loss function to optimize the model parameters, as shown in the formula: Among them, y i For real labels, To predict the probability, n is the number of samples; S2.2 uses the particle swarm optimization (PSO) algorithm to dynamically adjust the forget gate weights of the LSTM layer and the number of convolutional kernels of the CNN layer in order to balance the generalization ability and computational efficiency of the model.

4. The method according to claim 1, characterized in that, The dynamic threshold model construction method described in step S3 includes: Based on the quantile distribution of historical monitoring data, and combined with a Bayesian network to dynamically update the threshold, the formula is: T=α·Q 95％ +(1-α)·EMA(T prev ); Among them, Q 95％ The EMA(T) is the 95th percentile of historical data. prev ) is the exponential moving average of the threshold at the previous time step, and α is the adaptive weighting coefficient.

5. The method according to claim 1, characterized in that, The multi-scale convolution kernel configuration mentioned in step S2 is as follows: Convolutional kernels of different sizes are deployed in parallel in the CNN layer to extract local, meso- and macro-level spatial features respectively, and multi-scale feature maps are generated by channel splicing.

6. The method according to claim 1, characterized in that, The deviation calculation method described in step S3 is as follows: in, Let T be the predicted deformation value for the j-th monitoring point. j is the dynamic threshold for the corresponding monitoring point, and m is the total number of monitoring points.

7. The method according to claim 1, characterized in that, The multi-level early warning signals mentioned in step S4 include: Level 1 Warning: When the deviation D is in the range of [0.8, 1.0), it indicates that the monitoring frequency needs to be increased. Level II warning: When D is in the range [1.0, 1.2), the temporary support and reinforcement plan shall be activated; Level 3 warning: When D≥1.2, an emergency evacuation order for personnel and a work stoppage order are triggered.

8. The method according to claim 1, characterized in that, The method for constructing the spatiotemporal sequence dataset described in step S2 is as follows: The time series data is segmented into fixed time steps and reconstructed into a three-dimensional tensor with the sensor data at the corresponding spatial location to form a spatiotemporally coupled input matrix.

9. A real-time early warning system for tunnel deformation based on an LSTM-CNN model, characterized in that, include: Data acquisition module: Composed of distributed fiber optic grating sensors, inclinometers and temperature and humidity sensors, used to collect multi-dimensional monitoring data of the tunnel structure in real time; Edge computing module: integrates FPGA chip to realize real-time noise reduction and spatiotemporal sequence reconstruction of data; Cloud-based analytics platform: Deploys LSTM-CNN fusion models and dynamic thresholding algorithms to complete deformation prediction and risk assessment; Smart terminal module: Receives early warning signals via mobile communication and displays the tunnel deformation status using AR augmented reality technology.

10. The system according to claim 9, characterized in that, The data acquisition module and the edge computing module adopt a dual-channel redundant transmission architecture, wherein the main channel transmits real-time monitoring data through optical fiber, and the backup channel transmits key feature parameters synchronously through LoRa wireless technology. When the data transmission delay of the main channel exceeds a preset threshold, the edge computing module automatically switches to the backup channel and triggers a dynamic adjustment mechanism for the sensor sampling frequency to balance data integrity and transmission efficiency.

Images