A method for monitoring tunnel wall deformation based on computer image recognition

Abstract

This invention belongs to the field of tunnel wall deformation monitoring, specifically relating to a tunnel wall deformation monitoring method based on computer image recognition. The method includes: collecting scalable data and integrating it into a subsystem; performing image stabilization at the edge and constructing a standardized preprocessing subsystem to process the data; constructing a multimodal fusion architecture of structural topology graph and spatiotemporal graph neural network; constructing a hierarchical warning mechanism based on uncertainty-driven risk scores and segmented adaptive thresholds, and linking BIM and work order flow to form a closed loop. This invention can achieve millimeter-level displacement resolution, significantly reduce false alarms and missed alarms, and support long-term online, traceable tunnel wall deformation monitoring and hierarchical early warning closed loop.

Description

Technical Field

[0001] This invention belongs to the field of tunnel wall deformation monitoring, and specifically relates to a tunnel wall deformation monitoring method based on computer image recognition. Background Technology

[0002] Chinese patent application number CN202411085345.1 discloses a method and system for monitoring tunnel wall deformation based on computer image recognition. The method includes acquiring images of the tunnel area and tunnel wall images at each acquisition point; performing deformation recognition and prediction on the tunnel wall images and each tunnel wall image to obtain deformation image recognition data and wall deformation prediction results; acquiring spatial location data corresponding to the deformation image recognition data and wall deformation prediction results, generating several deformation monitoring units in the tunnel's deformation monitoring area; using each deformation monitoring unit as a basic unit, acquiring sensor monitoring data and corresponding monitoring threshold intervals and comparing them to obtain sensor monitoring judgment results; and performing spatial matching analysis based on all sensor monitoring judgment results, corresponding deformation image recognition data, and wall deformation prediction results to obtain the corresponding tunnel wall deformation monitoring results.

[0003] Although this invention has proposed a relatively complete solution in the field of tunnel deformation monitoring technology, it lacks vibration-resistant image stabilization, mechanism decomposition, and uncertainty modeling, making it susceptible to scene noise. This leads to an increased proportion of tunnel wall deformation events that fail to be detected and trigger alarms, thus affecting long-term detection accuracy and alarm accuracy. Specifically, this is due to: equipment vibration caused by vehicle passage and image registration errors accumulating under backlight conditions, resulting in displacement estimation deviations; over-reliance on fixed thresholds weakens the generalization ability under different soil cover, lining, and temperature environments; and the lack of robust handling and online correction for obstruction, dust, condensation, and sensor zero drift, as well as the failure to utilize structural topology and spatiotemporal correlation to suppress local anomalies, can lead to single-point noise being misjudged as real deformation, resulting in unstable early warning and positioning drift. Therefore, a tunnel wall deformation monitoring method based on computer image recognition is urgently needed to solve the above problems. Summary of the Invention

[0004] This invention provides a method for monitoring tunnel wall deformation based on computer image recognition, aiming to solve the following problems: Existing technologies lack vibration-resistant image stabilization, mechanism decomposition, and uncertainty modeling, making them susceptible to scene noise. This leads to an increased proportion of tunnel wall deformation events exceeding limits going undetected and triggering alarms, thus affecting long-term detection accuracy and alarm precision. Specifically: vehicle traffic causes equipment vibration, and image registration errors accumulate under backlighting conditions, resulting in displacement estimation deviations; over-reliance on fixed thresholds weakens the generalization ability under different soil cover, lining, and temperature environments; the lack of robust handling and online correction for obstructions, dust, condensation, and sensor zero drift, along with the failure to utilize structural topology and spatiotemporal correlation to suppress local anomalies, leads to single-point noise being misjudged as real deformation, causing unstable early warning and positioning drift.

[0005] The technical solution adopted by this invention to solve the above problems is as follows: a tunnel wall deformation monitoring method based on computer image recognition, comprising: dividing the imaging device into a basic acquisition unit and an enhanced acquisition unit to obtain a progressive deployment path of initial monocular imaging followed by enhanced depth imaging capability; enabling pulse alignment time per second at the camera trigger and edge gateway, and writing Coordinated Universal Time timestamps to the acquired data; abstracting the data channel at the acquisition end and adopting QoS priority scheduling;

[0006] Phase-correlated full-frame registration and local optical flow of the pyramid are performed sequentially in the edge computing unit;

[0007] A standardized preprocessing subsystem is constructed to process the data;

[0008] The backbone identification link is performed on the image frame sequence after image stabilization and multi-source sensor data aligned to UTC in the cloud; a structural topology graph is constructed with circumferential joints, longitudinal joints and arch feet and edge weights are set with adjacency relationship and mechanical compatibility constraints; a topology-weighted edge weight function is constructed to calculate the edge weights of the structural topology graph;

[0009] Construct a hierarchical warning mechanism based on uncertainty-driven risk scores and segmented adaptive thresholds; set risk score thresholds to judge the calculated risk scores; and link BIM with work order flow to form a closed loop.

[0010] As a preferred implementation, the specific steps for abstracting the data channel at the acquisition end and employing QoS priority scheduling are as follows:

[0011] By abstracting the data channels at the acquisition end and adopting QoS priority scheduling, key frames and frames within the alarm time window are assigned to the high-priority queue Q1, ordinary video frames are assigned to the queue Q2, and background synchronization and historical backhaul are assigned to the queue Q3. When the uplink bandwidth is detected to be lower than the set threshold, only Q1 is guaranteed to upload at a rate of not less than 500kbps, while Q2 is sampled at 1 / 5 and Q3 is suspended from transmission.

[0012] As a preferred embodiment, the specific steps of sequentially performing phase-correlated whole-frame registration and local optical flow of the pyramid in the edge computing unit are as follows:

[0013] Phase-correlated whole-frame registration is achieved by first performing a fast Fourier transform on two adjacent frames to obtain the spectrum, then normalizing the spectrum, and then performing an inverse fast Fourier transform on the normalized spectrum to obtain the positioning pulse peak coordinates on the spatial domain correlation surface as the whole-frame translation amount. The peak coordinates of the spatial domain correlation surface are then subjected to quadratic surface fitting interpolation to obtain the sub-pixel coordinates of the peak, and phase shift compensation is applied in the frequency domain according to the sub-pixel coordinates to reconstruct the aligned frame.

[0014] As a preferred embodiment, the step of sequentially performing phase-correlated whole-frame registration and pyramid local optical flow in the edge computing unit further includes:

[0015] First, a Gaussian pyramid of the target image is constructed. Then, the window gradient of each layer of the Gaussian pyramid after image stabilization is calculated from coarse to fine. The local optical flow normal equation of the pyramid is solved iteratively to obtain the incremental optical flow of that layer. Local resampling and registration of the target frame are then performed. The optical flow of that layer is then bilinearly upsampled and amplified by a scale factor as the initial displacement of the next layer. This process continues until the fine layer converges to the termination threshold, resulting in a local displacement vector field defined in the pixel coordinate system.

[0016] As a preferred embodiment, the specific steps for constructing a standardized preprocessing subsystem to process data are as follows:

[0017] By performing detrending, bandpass, and temperature compensation on acceleration and temperature time series and aligning them to the timestamp of each frame using UTC, sensor samples that match the image and frame one by one are obtained. By performing zero-drift online regression on the sensor samples at the edge and correcting the displacement estimation results in real time, a long-term stable deformation baseline is obtained. By defining a lightweight mode that only outputs stable keyframes and displacement candidate points, and defining a standard mode that outputs dense displacement fields and quality masks, output levels that adapt to different bandwidths and computing power are obtained. By caching evidence frame fragments in the alarm window and labeling the triggering reasons and threshold parameters, a traceable evidence chain is obtained. By enabling HDR multi-exposure acquisition in hotspot sections and using single-exposure acquisition in ordinary sections, computational and storage resources for hotspot sections can be prioritized.

[0018] As a preferred embodiment, the specific steps of performing the backbone identification link on the stabilized image frame sequence and UTC-aligned multi-source sensor data in the cloud are as follows:

[0019] Based on the signal-to-noise ratio threshold (SNR) and gradient magnitude threshold For each frame of the stabilized image frame sequence, pinhole distortion correction is first performed. ,and Valid pixels are used to generate a quality mask and determine the valid pixels participating in the calculation, thus obtaining the valid region, where SNR1 is the local signal-to-noise ratio. The local gradient is used; the dense two-dimensional displacement field is estimated by using normalized cross-correlation joint gradient constraint optical flow within the effective region, and outliers are eliminated by random sampling consistency, iterating until the proportion of inliers is not lower than a set lower limit; the normal information of each pixel is obtained by enabling binocular depth recovery and calculating local normal vectors at key cross sections; the displacement vector is decomposed into normal bulging displacement and in-plane slip displacement according to the normal direction to obtain physical quantities that can directly correspond to the bulging and shearing mechanisms.

[0020] As a preferred embodiment, the specific steps for calculating the edge weights of the structural topology graph using the topology-weighted edge weight function are as follows:

[0021] By constructing a topologically weighted edge weight function to calculate the edge weights of the structural topology graph, a weighted adjacency matrix is ​​obtained that is simultaneously influenced by geometric proximity, structural adjacency, historical correlation, and mechanical compatibility constraints. The formula for constructing the topologically weighted edge weight function is as follows:

[0022] ,

[0023] A represents the edge weights, exp represents the exponential function, B represents the spatial location vector of the detection unit, and i and j represent the x and y coordinates of the spatial location of the detection unit, respectively. It is the symbol for second normal form. For distance attenuation scale, The weights of adjacent indicators in the structure, For structural adjacency indicator functions, The weighting coefficients for historical correlation coefficients. Historical correlation coefficient The weighting coefficients for the mechanical compatibility term are... This is a mechanical compatibility term.

[0024] As a preferred embodiment, the specific steps for constructing the hierarchical alarm mechanism based on uncertainty-driven risk scores and segmented adaptive thresholds are as follows:

[0025] By constructing an uncertainty-driven risk score, the mean deformation estimate output by the spatiotemporal graph neural network is combined with the total uncertainty to form a risk score. Correction coefficients for overburden depth and lining type are introduced to obtain segmented adaptive threshold warning results. The formula for the uncertainty-driven risk score is as follows:

[0026] ,

[0027] Let G represent the deformation risk score of monitoring unit n at time m, and let G represent the deformation risk intensity. Let Q be the expected value of the deformation estimate for monitoring unit n at time m, where n is the monitoring unit and m is the time index. Let H be the variance of the estimated true deformation of monitoring unit n at time m, where H is the true deformation and Change is the variance of the estimated true deformation. It is the stability constant. Maintenance factor under operating conditions; The square root sign, This is scalar multiplication.

[0028] As a preferred embodiment, the specific steps for judging the calculated risk score by setting a risk score threshold are as follows:

[0029] After constructing the formula for the uncertainty-driven risk score, the calculated risk score is then processed according to segmented thresholds. The judgment is performed, and four threshold levels are set: Level 1 threshold DA, Level 2 threshold D2, Level 3 threshold D3, and Level 4 threshold D4. The specific judgment rules and their meanings are as follows: If green is selected as normal, it means that the unit is within the normal range and the short-term rate of change has not exceeded the stable rate, and no additional action is required.

[0030] If the value is yellow, it indicates that attention is needed, meaning that the unit has reached the lower warning limit or is showing a continuous upward trend, and it is necessary to increase the frequency of inspections and arrange on-site verification.

[0031] If the value is orange, an early warning will be issued, indicating that the unit has reached the warning threshold and exhibits clustered abnormalities in time or space. On-site inspections should be initiated and temporary control measures should be implemented.

[0032] If the alarm is triggered, it is classified as a red emergency, indicating that the unit has exceeded the danger alarm threshold and is accompanied by structural danger signs. Emergency measures should be taken immediately and traffic and construction activities should be suspended as appropriate.

[0033] As a preferred implementation method, the specific steps for forming a closed loop between the linked BIM and work order flow are as follows:

[0034] By mapping normal bulging, in-plane slip, crack indices and evidence frames to the BIM model and overlaying mileage and ring number, a visualization interface for three-dimensional spatial positioning and quantitative indicators is obtained. After handling the deformation hot zone and corresponding monitoring unit involved in the early warning event, the on-site cross-section of the tunnel monitored by the monitoring unit is re-measured, and the re-measurement results, threshold records and model version are sent back to the training set to obtain a traceable data loop.

[0035] The beneficial effects of this invention are as follows:

[0036] 1. By combining the mean deformation estimate and total uncertainty output by the spatiotemporal graph neural network into a risk score and introducing correction coefficients for soil cover depth and lining type, a segmented adaptive threshold alarm result is obtained, which can reduce false alarms and missed alarms under nighttime, dusty and backlight conditions.

[0037] 2. By mapping normal bulging, in-plane slip, crack indices and evidence frames to the BIM model and overlaying mileage and ring numbers, a visualization interface for three-dimensional spatial positioning and quantitative indicators is obtained, which can support on-site personnel to quickly locate and formulate maintenance plans.

[0038] 3. By performing a retest after the processing is completed and feeding the retest results, threshold records, and model version back into the training set, a traceable data loop is obtained, which supports fine-tuning with small samples and steadily improving model accuracy. Attached Figure Description

[0039] Figure 1 This is a flowchart of a method for monitoring tunnel wall deformation based on computer image recognition.

[0040] Figure 2 This is a technical rendering of a tunnel wall deformation monitoring method based on computer image recognition. Detailed Implementation

[0041] To make the technical means, creative features, and achieved objectives and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention. Unless otherwise specified, the experimental methods in the following embodiments are conventional methods, and the materials and reagents used in the following embodiments are commercially available unless otherwise specified.

[0042] Example 1 Figure 1A flowchart of a tunnel wall deformation monitoring method based on computer image recognition is presented, including: collecting scalable data and integrating it into a subsystem; performing image stabilization at the edge and constructing a standardized preprocessing subsystem to process the data; constructing a multimodal fusion architecture of structural topology graph and spatiotemporal graph neural network; constructing a hierarchical alarm mechanism based on uncertainty-driven risk score and segmented adaptive threshold, and linking BIM and work order flow to form a closed loop; the specific implementation steps are as follows:

[0043] Step 1: Collect scalable data and integrate it into the subsystem;

[0044] Specifically, the steps for acquiring data and integrating it into the subsystem are as follows: By dividing the imaging equipment into basic acquisition units and enhanced acquisition units, a progressive deployment path is obtained, from the initial deployment of monocular imaging to the subsequent enhancement of depth imaging capabilities. This reduces initial procurement costs and allows for segmented expansion and functional upgrades as needed. The basic acquisition unit consists of a low-light camera and near-infrared supplementary lighting, used to acquire high signal-to-noise ratio visible light and near-infrared image frames of the tunnel wall under conditions of backlighting, dust attenuation, and wet reflection, and outputs a video stream with a unified timestamp, providing clear texture input for image stabilization registration and 2D displacement estimation. The enhanced acquisition unit consists of a binocular camera, used to acquire parallax and depth information and recover the local normal vector and 3D point cloud of the wall, thereby supporting normal bulging and millimeter-level 3D deformation calculations. By setting a driver adaptation layer on the device side and ensuring compatibility with the Real-Time Streaming Protocol (RTSP) and Industrial Serial Communication Protocol (IPC), a more comprehensive and efficient system is established. Protocol (ISCP) enables plug-and-play access to existing cameras and sensors, thereby protecting existing assets and shortening the time-to-live phase. Existing assets refer to reusable hardware and network resources already deployed in the field, such as industrial cameras, near-infrared illuminators, sensors, edge gateways, and switches. The time-to-live phase refers to the overall acceptance period from device access, time synchronization and calibration to the generation of stable alarms. This is achieved by enabling pulse-per-second (PPS) alignment time on camera triggers and edge gateways, and writing the acquired data to Coordinated Universal Time (UTC). Time (UTC) timestamps are used to achieve millisecond-level timing consistency across devices; by installing reference components on fixed brackets and vehicle platforms and calibrating the positions of coded patches, reproducible initial values ​​of extrinsic parameters and field of view coverage are obtained, thereby reducing power-on calibration time and replacement and debugging costs; by abstracting the data channels at the acquisition end and adopting QoS priority scheduling, key frames and frames within the alarm time window are assigned to high-priority queue Q1, ordinary video frames to queue Q2, and background synchronization and historical backhaul to queue Q3. When the uplink bandwidth is detected to be lower than a set threshold, only Q1 is guaranteed to be at a lower priority. Uploading at rates below 500kbps, while sampling Q2 at 1 / 5 and pausing transmission for Q3, allows for priority uploading of key frames and maintenance of the integrity of alarm periods even under bandwidth-constrained conditions, improving availability in weak network conditions. The tunnel is divided into monitoring units according to mileage markers and ring numbers, for example, every 10 meters of the tunnel is a unit, and each unit is registered with fields such as soil cover depth, lining type (e.g., secondary lining), shotcrete thickness, start and end mileage, and center coordinates, thereby obtaining a collection index containing structural semantics, providing engineering priors for subsequent threshold partitioning and segmented capacity expansion.By applying a hydrophilic coating to the protective window and configuring a passive air duct, an anti-condensation imaging window is obtained, thereby reducing the frequency of manual wiping and downtime maintenance. By initially installing and activating only basic acquisition units and covering key sections, such as the arch foot and adjacent sections of the track bed, and areas prone to leakage, a small operational closed loop is achieved, enabling the rapid formation of a usable system within budget constraints.

[0045] Step 2: Perform image stabilization at the edge and build a standardized preprocessing subsystem to process the data;

[0046] Specifically, the steps for image stabilization at the edge are as follows: Phase-correlated whole-frame registration and Pyramidal Lucas-Kanade Local Optical Flow (Pyramidal LK) are sequentially performed in the edge computing unit to obtain a stable image field of no more than 0.06 pixels, thereby establishing a stable reference coordinate system under vibration and jitter conditions. The edge computing unit refers to computing devices such as CPUs, memory, storage, and network interfaces deployed on-site and responsible for preprocessing and preliminary inference. The phase-correlated whole-frame registration involves performing a Fast Fourier Transform on two adjacent frames to obtain the spectrum, then normalizing the spectrum, and then performing an Inverse Fast Fourier Transform on the normalized spectrum to obtain the positioning pulse peak coordinates on the spatial domain correlation surface as the whole-frame translation. Quadratic surface fitting interpolation is then performed on the peak coordinates of the spatial domain correlation surface to obtain the sub-pixel coordinates of the peak. Phase-shift compensation is then applied in the frequency domain according to the sub-pixel coordinates to reconstruct the aligned frame, thus completing the whole-frame registration and providing a stable reference coordinate system for subsequent local optical flow and displacement calculations. The Pyramidal LK... LK first constructs a Gaussian pyramid of the target image, calculates the window gradient of each layer of the Gaussian pyramid after image stabilization from coarse to fine, iteratively solves the local optical flow normal equation of the pyramid to obtain the incremental optical flow of that layer, performs local resampling on the target frame, and performs local resampling registration on the target frame. The optical flow of this layer is then bilinearly upsampled and amplified by a scale factor as the initial displacement of the next layer until the fine layer converges to the termination threshold, thus obtaining a local displacement vector field defined in the pixel coordinate system. By inversely compensating the image with the stabilized field and performing multi-scale Retinex illumination normalization on the compensated image, an edge-normalized frame sequence is obtained, which can reduce the impact of low light changes on recognition quality. Pluggable operators are configured on the edge side for management. The system loads the stabilized field inverse compensation image in a containerized manner, thereby achieving an on-demand start / stop and load-based scaling operation mechanism, which reduces hardware / software coupling and upgrade risks. The stabilized field inverse compensation image is obtained by negating the stabilized field and geometrically resampling the target frame to obtain a compensation image aligned with the reference frame. The pluggable operator manager manages image stabilization, denoising, decomposition, alarm detection, and other processing operators during runtime by providing standardized loading and upgrade interfaces, thereby achieving on-demand start / stop and version rollback operator orchestration capabilities. The containerization method packages each functional component and its dependencies into independent containers and deploys and runs them in isolation using images, thereby obtaining a portable, elastically scalable, and quickly rollback online operating environment.

[0047] The specific steps for constructing a standardized preprocessing subsystem to process data are as follows: By performing detrending, bandpass, and temperature compensation on acceleration and temperature time series data and aligning them to the timestamp of each frame using UTC, sensor samples that match the image and frame one-to-one are obtained; by performing zero-drift online regression on the sensor samples at the edge and correcting the displacement estimation results in real time, a long-term stable deformation baseline is obtained, thereby reducing periodic recalibration and lowering the cost of manual inspection; by defining a lightweight mode that only outputs stabilized keyframes and displacement candidate points, and defining a standard mode that outputs dense displacement fields and quality masks, output levels adapted to different bandwidths and computing power are obtained, thereby ensuring unified scheduling of multiple lines; the keyframes refer to image frames that, after stabilization and distortion correction, are used as reference registration and alarm evidence; the displacement candidate points refer to the set of pixel coordinates selected on the keyframes based on texture intensity thresholds and corner scoring thresholds, used as the precise input for subsequent local optical flow or sub-pixel matching; the quality mask is a probability map of the same size as the image, used to label each... The validity of each pixel is comprehensively determined based on the signal-to-noise ratio threshold, gradient magnitude threshold, stable in-pixel marking, and occlusion detection results. Regions with a value of 1 participate in displacement calculation, while regions with a value of 0 are not included in the calculation and are omitted during uplink. By caching evidence frame fragments in the alarm window and labeling the triggering cause and threshold parameters, a traceable evidence chain is obtained, which can support data link-level review and audit evidence collection. By enabling HDR multi-exposure acquisition in hotspot sections and using single-exposure acquisition in ordinary sections, computing and storage resources for hotspot sections can be prioritized, thus ensuring stable recognition of key sections under the condition of limited computing power and storage scale in the early stage. The aforementioned limited computing power and storage scale conditions refer to the objective environment of limited edge node and link resources in the early stage of deployment. For example, edge nodes have no independent GPU, the CPU is quad-core, memory ≤8GB, local available storage ≤256GB, and the average uplink bandwidth is ≤1Mbps with occasional jitter. Under this resource constraint, priority allocation of acquisition, computing, and storage strategies is required.

[0048] Step 3: Construct a multimodal fusion architecture of structural topology graph and spatiotemporal graph neural network;

[0049] Specifically, the steps for constructing a multimodal fusion architecture of structural topology graph and spatiotemporal graph neural network are as follows: A backbone recognition link is executed on the stabilized image frame sequence and multi-source sensor data aligned to UTC in the cloud. Specifically, this involves: based on the signal-to-noise ratio (SNR) threshold and the gradient magnitude threshold... For each frame of the stabilized image frame sequence, pinhole distortion correction is first performed. ,and A quality mask is generated by determining whether a pixel is valid, and the valid pixels participating in the calculation are identified, thereby obtaining the effective region. Here, SNR1 is the local signal-to-noise ratio. For local gradient estimation, a dense two-dimensional displacement field is estimated by using normalized cross-correlation joint gradient constraint optical flow within the effective region. Outliers are eliminated using Random Sample Consensus (RANSAC), iterating until the proportion of inliers is not lower than a set lower limit, thus ensuring the registration residual is no greater than 0.05 pixels. By enabling binocular depth recovery and calculating local normal vectors at key cross-sections, normal information for each pixel is obtained, supporting mechanism decomposition. By decomposing the displacement vector into normal bulging displacement and in-plane slip displacement, physical quantities directly corresponding to bulging and shearing mechanisms are obtained, supporting the consistency comparison of laser cross-section sampling results. A structural topology diagram is constructed using circumferential seams, longitudinal seams, and arch feet, with edge weights set based on adjacency relationships and mechanical compatibility constraints. For example, if two monitoring units are located in the same circumferential seam and adjacent ring blocks, the edge weight is set to 0.1; if they are located in the same longitudinal seam... If the edges are adjacent to each other, the edge weight is set to 0.6; if it is the arch foot and its adjacent arch waist, the edge weight is set to 0.8; otherwise, the edge weight is set to 0. Adjacent edges are corrected according to mechanical compatibility constraints. For example, when the displacement gradient difference is large, the edge weight is increased by +0.1; otherwise, it is decreased by -0.1. This yields a graph structure representing structural coupling, providing engineering priors for spatiotemporal modeling. By constructing a topologically weighted edge weight function, the edge weights of the structural topology graph are calculated, resulting in a weighted adjacency matrix simultaneously influenced by geometric proximity, structural adjacency, historical correlation, and mechanical compatibility constraints. This allows for the suppression of single-point anomaly propagation and improved spatiotemporal fusion robustness under conditions of image occlusion and local noise. The formula for constructing the topologically weighted edge weight function is as follows:

[0050] ,

[0051] A represents the edge weight, exp represents the exponential function, B represents the spatial location vector of the detection unit, and i and j represent the horizontal and vertical coordinates of the spatial location of the detection unit, respectively. It is the symbol for second normal form. For distance attenuation scale, The weights of adjacent indicators in the structure, For structural adjacency indicator functions, The weighting coefficients for historical correlation coefficients. Historical correlation coefficient The weighting coefficients for the mechanical compatibility term are... For mechanical compatibility terms;

[0052] After constructing the topologically weighted edge weight function, the normal bulging, in-plane slip, crack width, temperature, acceleration, overburden depth, and lining type are organized into node feature vectors according to monitoring units and time steps. A weighted adjacency matrix is ​​used as the spatial connection input to the spatiotemporal graph neural network, thereby obtaining the mean deformation estimate and total uncertainty of each monitoring unit at each time step. This ensures robust output even under occlusion and short-term data loss. The spatiotemporal graph neural network uses a weighted adjacency matrix to represent spatial connections and performs graph convolution and temporal modeling on the node temporal features for joint encoding, thus outputting a supervised learning model of the mean deformation estimate and uncertainty of each monitoring unit at each time step. Thin plate splines are used on the parametric plane of the tunnel wall. By incorporating topological preservation regularity into the objective function for solving the thin plate spline coefficients and interpolating the deformation field between nodes, a continuous surface displacement distribution without non-physical tearing is obtained, which can provide regional-level evidence for hot zone location and trend analysis. The hot zone location refers to threshold segmentation of indicators such as normal bulging displacement, in-plane slip displacement, crack width, and risk score, and merging of connected components to output the center coordinates, boundary polygons, and area of ​​each continuous over-limit area, which is used to identify high-risk locations on the tunnel wall. The trend analysis calculates the linear slope, growth rate, and change points of the monitoring unit's time series after seasonal and noise reduction, and judges the state as rising deterioration, basically stable, or improving based on fixed criteria, thus assessing the direction of deformation evolution.

[0053] Step 4: Construct a hierarchical warning mechanism that uses uncertainty-driven risk scores and segmented adaptive thresholds, and link it with BIM and work order flow to form a closed loop;

[0054] Specifically, the steps for constructing a hierarchical alarm mechanism based on uncertainty-driven risk scores and segmented adaptive thresholds are as follows: An uncertainty-driven risk score is constructed by combining the mean deformation estimate output by the spatiotemporal graph neural network with the total uncertainty to form a risk score. Correction coefficients for overburden depth and lining type are introduced to obtain segmented adaptive threshold alarm results. This reduces false alarms and missed alarms under nighttime, dusty, and backlighting conditions. The formula for the uncertainty-driven risk score is:

[0055] ,

[0056] Let G represent the deformation risk score of monitoring unit n at time m, and let G represent the deformation risk intensity. Let Q be the expected value of the deformation estimate for monitoring unit n at time m, where n is the monitoring unit and m is the time index. Let H be the variance of the estimated true deformation of monitoring unit n at time m, where H is the true deformation and Change is the variance of the estimated true deformation. It is the stability constant. Maintenance factor under operating conditions; The square root sign, This is scalar multiplication;

[0057] After constructing the formula for the uncertainty-driven risk score, the calculated risk score is then processed according to segmented thresholds. The judgment is performed, and four threshold levels are set: Level 1 threshold DA, Level 2 threshold D2, Level 3 threshold D3, and Level 4 threshold D4. The specific judgment rules and their meanings are as follows: If green is selected as normal, it means that the unit is within the normal range and the short-term rate of change has not exceeded the stable rate, and no additional action is required.

[0058] If the value is yellow, it indicates that attention is needed, meaning that the unit has reached the lower warning limit or is showing a continuous upward trend, and it is necessary to increase the frequency of inspections and arrange on-site verification.

[0059] If the value is orange, an early warning will be issued, indicating that the unit has reached the warning threshold and exhibits clustered abnormalities in time or space. On-site inspections should be initiated and temporary control measures should be implemented.

[0060] If the alarm is triggered, it is classified as a red emergency, indicating that the unit has exceeded the danger alarm threshold and is accompanied by structural danger signs. Emergency measures should be taken immediately and traffic and construction activities should be suspended as appropriate.

[0061] By setting four threshold levels to judge the risk score, the alarm-handling action can be linked in real time, thereby shortening the response time from alarm to handling and solidifying the handling process.

[0062] The specific steps for forming a closed loop between BIM and work order flow are as follows: By mapping normal bulging, in-plane slip, crack indicators, and evidence frames to the BIM model and overlaying mileage and ring numbers, a visual interface for three-dimensional spatial positioning and quantitative indicators is obtained. This supports on-site personnel in quickly locating and formulating maintenance plans for the deformation hot zones, crack locations, and their quantitative parameters, such as length, average width, and growth rate, of the tunnel wall. After handling the deformation hot zones and corresponding monitoring units involved in the early warning event, the on-site cross-section of the tunnel monitored by the monitoring unit is re-measured, and the re-measurement results, threshold records, and model version are sent back to the training set, thereby obtaining a traceable data closed loop. This enables small-sample fine-tuning and steadily improving the accuracy of the BIM model.

[0063] like Figure 2The figure shows the technical effect of a tunnel wall deformation monitoring method based on computer image recognition. The black lines represent the technical effect of the present invention, while the gray lines represent the technical effect achieved by the prior art. As can be seen from the figure, the technical effect of the present invention is superior to that of the prior art.

[0064] Example 2, based on Example 1, describes a method for monitoring tunnel wall deformation using computer image recognition. The specific solution is as follows:

[0065] Step 1: Collect data on-site and align the time.

[0066] Specifically, the steps for collecting data and aligning time on-site are as follows: By installing low-light industrial cameras and near-infrared supplementary lighting at the arch crown, arch waist, and arch foot, and enabling PPS time synchronization and writing UTC timestamps to the camera triggers and edge gateways, a consistent image stream and sensor stream with a time sequence of no more than 1ms across devices is obtained, thus providing a unified time reference for subsequent fusion; by dividing monitoring units according to mileage and ring number and registering the soil cover depth and lining type, a structural semantic index of unit ID—mileage—ring number—soil cover—lining is obtained, which can support threshold partitioning based on segment elements such as mileage—ring number, soil cover depth, and lining type, and subsequent spatiotemporal graph neural network modeling of structural topology and time sequence features with monitoring units as nodes.

[0067] Step two: Perform jitter correction and standardization.

[0068] Specifically, the steps for jitter correction and standardization are as follows: Phase-correlated whole-frame registration and Pyramidal LK local optical flow are used to obtain a stable image field with a registration residual of no more than 0.05 pixels, thus enabling the establishment of a stable reference coordinate system under train vibration and equipment jitter conditions; The target frame is inversely compensated using the stable image field, and Retinex illumination normalization and pinhole model distortion correction are performed to obtain a standardized frame sequence with controlled noise and unified coordinates, thereby improving the stability of displacement estimation; Zero-drift online regression is performed on sensor samples, and displacement estimation is corrected in real time to obtain a long-term stable deformation baseline, thereby reducing the cost of periodic recalibration and manual inspection.

[0069] Step 3: Calculate the mechanistic quantities and integrate the graphical model;

[0070] Specifically, the steps for calculating mechanistic quantities and fusing the graph model are as follows: By calculating dense two-dimensional displacements on standardized frames and fusing binocular parallax or structured light depth, pixel-level displacement vectors and wall normals are obtained, which can provide accurate input for mechanism decomposition; by orthogonally decomposing the dense displacement vectors according to the local wall normals, normal bulging displacements and in-plane slip displacements are obtained, which can distinguish between bulging mechanisms and shear mechanisms under the same coordinates and dimensions, and provide comparable physical quantities for subsequent threshold warning and structural assessment; by constructing a structural topology graph with monitoring units as nodes and calculating edge weights according to geometric proximity, structural adjacency, historical correlation and mechanical compatibility, normal bulging, in-plane slip, crack parameters, temperature, acceleration, soil cover depth and lining type are used as node features to input spatiotemporal graph neural network modeling, thereby obtaining the mean deformation estimate and total uncertainty of each node, which can maintain robust output under occlusion and data loss.

[0071] Step 4: Build a risk scoring and early warning system and close the loop with BIM work orders;

[0072] Specifically, the steps for constructing a risk scoring and early warning system and using BIM work orders in a closed loop are as follows: A risk score is calculated by constructing an uncertainty-driven risk score, and segmented thresholds are set to implement green, yellow, orange, and red graded alarms, thereby obtaining alarm results with low false alarms and low false negatives, improving long-term online stability. The segmented thresholds for green, yellow, orange, and red graded alarms are as follows: green indicates that the monitoring indicator is within the normal range and the short-term change rate has not exceeded the preset stable rate, requiring no additional action; yellow indicates that the monitoring indicator has exceeded the lower warning limit but has not yet reached the engineering handling threshold, indicating that increased monitoring frequency and on-site verification are needed; orange indicates that the monitoring indicator has exceeded the early warning threshold and shows clustered anomalies in space or time, for example: multiple [indicators] in the same segment [area not specified]. Points exceeding limits indicate the need to initiate on-site inspections and implement temporary control measures; red indicates that the monitored indicators have exceeded the maximum alarm threshold and are accompanied by structural hazards, such as rapidly increasing normal bulging and significant tunnel crack expansion, indicating the need for immediate emergency measures and possible traffic interruption; by mapping normal bulging, in-plane slip, and crack indicators to the BIM model and automatically generating work orders, a closed-loop process of alarm—work order—on-site arrival—retest—receipt is obtained, which can shorten the response time from alarm to action and solidify the action process; after the action is completed, retesting is performed through the corresponding monitoring unit and the retest results, trigger thresholds, and model version are fed back into the training dataset, thus obtaining a traceable model update closed loop, which can continuously improve the accuracy of alarm judgment.

[0073] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring tunnel wall deformation based on computer image recognition, characterized in that, include: The tunnel wall image stream and sensor stream are time-aligned, and monitoring units are divided according to mileage and ring number. The soil cover depth and lining type of each monitoring unit are registered to obtain a structural semantic index. The tunnel wall image stream is stabilized and quality masked to obtain the effective region corresponding to the structural semantic index. Dense two-dimensional displacement field calculation is performed on the effective area, and the displacement vector is decomposed by combining the wall normal obtained by binocular depth recovery to obtain the normal bulging displacement and in-plane slip displacement of each monitoring unit. Using the monitoring unit as a node, and constructing a structural topology graph with circumferential joints, longitudinal joints, and arch feet, the edge weights of the structural topology graph are calculated based on geometric proximity, structural adjacency, historical correlation coefficient, and mechanical compatibility term to obtain a weighted adjacency matrix. The normal bulging displacement, the in-plane slip displacement, the temperature and acceleration data in the sensor stream, the soil cover depth, the lining type, and the weighted adjacency matrix are input into the spatiotemporal graph neural network to obtain the mean deformation estimate and total uncertainty of each monitoring unit. The mean value of the deformation estimate and the total uncertainty are combined to form a risk score. Correction coefficients for the soil cover depth and lining type are introduced. The risk score is classified and judged according to the segmented threshold to obtain the segmented classification warning result. The tunnel wall image stream is stabilized and quality masked to obtain the effective region corresponding to the structural semantic index, including: Phase-correlated whole-frame registration is performed on adjacent image frames in the tunnel wall image stream to obtain the whole-frame translation amount; The target frame is compensated based on the frame translation amount to obtain a frame-aligned image. Perform pyramid local optical flow calculations on the entire frame aligned image to obtain a local image-stabilized field; Based on the local image stabilization field, reverse compensation is performed on the target frame to obtain a stabilized image frame; A quality mask is generated based on the signal-to-noise ratio, gradient magnitude, in-frame markers, and occlusion detection results of each pixel in the stabilized image frame. The effective region is obtained by associating the pixel regions marked as valid in the quality mask with the monitoring unit numbers in the structural semantic index; A dense two-dimensional displacement field calculation is performed on the effective region, and the displacement vector is decomposed by combining the wall normal obtained from binocular depth recovery to obtain the normal bulging displacement and in-plane slip displacement of each monitoring unit, including: Within the effective region, normalized cross-correlation joint gradient-constrained optical flow is used to estimate the dense two-dimensional displacement field, and outliers are eliminated using random sampling consistency. The local wall normal at the corresponding pixel position is calculated by combining the binocular depth recovery results, and the displacement vector in the dense two-dimensional displacement field is associated with the local wall normal. The projection of the displacement vector onto the local normal direction of the wall surface is determined as the normal bulging displacement; The component of the displacement vector after removing the normal projection is determined as the in-plane sliding displacement. Using the monitoring unit as a node, and constructing a structural topology graph with circumferential joints, longitudinal joints, and arch foot, the edge weights of the structural topology graph are calculated based on geometric proximity, structural adjacency, historical correlation coefficient, and mechanical compatibility terms to obtain a weighted adjacency matrix, including: The monitoring unit is used as a node; The structural adjacency relationship between nodes is determined based on the circumferential joint, longitudinal joint, and arch foot. Geometric proximity terms are determined based on the spatial relationship between the monitoring units; Historical correlation terms are determined based on the historical correlation coefficients between monitoring units; The mechanical compatibility terms are determined based on the normal bulging displacement and in-plane slip displacement of adjacent monitoring units; The edge weights of the structural topology graph are calculated based on the geometric proximity term, the structural adjacency relationship, the historical correlation term, and the mechanical compatibility term to obtain the weighted adjacency matrix; The structural adjacency relationships between nodes are determined based on the circumferential joints, longitudinal joints, and arch foot, including: When two monitoring units are located in the same annular gap and are adjacent ring blocks, the two monitoring units are determined to be adjacent to each other in the annular gap, and the edge weight used for the structural adjacent relationship is set to 0.

1. When two monitoring units are located in the same longitudinal joint and are adjacent blocks, the two monitoring units are determined to be in a longitudinal joint adjacent relationship, and the edge weight used for the structural adjacent relationship is set to 0.

6. When two monitoring units are located at the arch foot and the adjacent arch waist respectively, the relationship between the two monitoring units is determined to be an arch foot constraint relationship, and the edge weight used for the structural adjacency relationship is set to 0.8; When two monitoring units do not satisfy the circumferential joint adjacency relationship, the longitudinal joint adjacency relationship, and the arch foot constraint relationship, the two monitoring units are determined to be non-structural adjacency relationships, and the edge weight used for structural adjacency relationships is set to 0. After obtaining the segmented and graded alarm results, the following is also included: The normal bulging displacement, in-plane slip displacement and evidence frame corresponding to the segmented and graded alarm results are mapped to the corresponding mileage and ring number positions in the building information model to obtain the deformation evidence location results. A corresponding processing work order is generated based on the location results of the deformed evidence. After the on-site retest corresponding to the disposal work order is completed, the retest results, trigger thresholds and model versions are associated with the corresponding monitoring unit number and sent back to the training set to obtain traceable retest records.

2. The tunnel wall deformation monitoring method according to claim 1, characterized in that, The tunnel wall image stream and sensor stream are time-aligned, and monitoring units are divided according to mileage and ring number. The soil cover depth and lining type of each monitoring unit are registered to obtain a structural semantic index, including: Based on the pulse per second time pair between the camera trigger and the edge gateway, Coordinated Universal Time timestamps are written for the image frames in the tunnel wall image stream and the sensor samples in the sensor stream; The image frames and sensor samples are aligned to the same time base according to the Coordinated Universal Time timestamp. The monitoring unit number is determined by the tunnel mileage and lining ring number. The soil cover depth, lining type, start and end mileage and center coordinates of each monitoring unit are associated with the monitoring unit number to obtain the structural semantic index.

3. The tunnel wall deformation monitoring method according to claim 2, characterized in that, The normal bulging displacement, the in-plane slip displacement, the temperature and acceleration data in the sensor stream, the overburden depth, the lining type, and the weighted adjacency matrix are input into a spatiotemporal graph neural network to obtain the mean deformation estimate and total uncertainty of each monitoring unit, including: Based on the monitoring unit number and time step, the normal bulging displacement, the in-plane sliding displacement, the temperature data, the acceleration data, the soil cover depth, and the lining type are organized into node feature vectors; The weighted adjacency matrix is ​​used as the spatial connection relationship between nodes; The node feature vectors and the weighted adjacency matrix are input into the spatiotemporal graph neural network to obtain the mean deformation estimate and total uncertainty of each monitoring unit at the corresponding time step.

4. The tunnel wall deformation monitoring method according to claim 3, characterized in that, The mean value of the deformation estimate and the total uncertainty are combined to form a risk score. Correction coefficients for overburden depth and lining type are introduced. The risk score is then graded according to a segmented threshold to obtain a segmented graded warning result, including: Read the mean deformation estimate, total uncertainty, soil cover depth, and lining type of the corresponding monitoring unit; The mean of the deformation estimate is combined with the total uncertainty to form a risk score; By introducing correction coefficients for the soil cover depth and the lining type, the risk score is segmented to obtain a segmented risk score. The segmented risk score is compared with the segmented threshold group, and one of the following is output according to the threshold range in which the segmented risk score is located: green normal, yellow attention, orange warning or red emergency, to obtain the segmented graded alarm result.

Images