A tunnel construction deformation monitoring system based on machine vision and a monitoring method thereof

Abstract

This invention relates to the field of tunnel construction monitoring technology, and discloses a tunnel construction deformation monitoring system and method based on machine vision. The method acquires real-time images of the tunnel construction structure to obtain a sequence of tunnel surface images, while simultaneously monitoring the displacement and strain parameters of the tunnel structure. Based on the displacement and strain parameters, initial deformation prediction is performed to obtain predicted deformation parameters. Images of key areas inside the tunnel are acquired using machine vision equipment, and these images, combined with the predicted deformation parameters, are input into multiple deformation feature recognizers to identify actual deformation features. Finally, the predicted deformation parameters and actual deformation features are combined to perform a final deformation trend analysis, obtaining the tunnel deformation monitoring results. This method improves the targeting, comprehensiveness, and accuracy of deformation monitoring, and can more accurately reflect the deformation state of the tunnel structure, providing an effective technical means for tunnel construction safety monitoring.

Description

Technical Field

[0001] This invention relates to the field of tunnel construction monitoring technology, specifically to a tunnel construction deformation monitoring system and method based on machine vision. Background Technology

[0002] During tunnel construction, structural deformation monitoring is a crucial step in ensuring construction safety and project quality. Tunnel projects are situated in complex and variable geological environments, affected by factors such as ground pressure, groundwater, and construction disturbances. These factors make them highly susceptible to problems such as surrounding rock deformation, support structure cracking, and settlement. Failure to detect and warn of these issues in a timely manner can lead to serious safety accidents such as collapses, resulting in significant casualties and economic losses.

[0003] Traditional methods for monitoring tunnel deformation primarily rely on manual inspections and contact sensors. Manual inspections require personnel to enter the tunnel and measure data using instruments such as total stations and levels. This process is not only time-consuming and labor-intensive, but also suffers from low efficiency due to the harsh environment inside the tunnel (such as dust, insufficient lighting, and confined space). Data acquisition cycles are long, making real-time monitoring difficult and often resulting in monitoring delays, failing to capture the dynamic changes in deformation in a timely manner. Furthermore, the accuracy of manual measurements is easily affected by factors such as the operator's skill level and subjective judgment, leading to significant data errors.

[0004] Contact-based sensor monitoring methods, such as the deployment of strain gauges and displacement meters, while achieving a certain degree of automated monitoring, have significant limitations. Firstly, sensor deployment requires direct contact with the tunnel structure, potentially causing damage to the support structure and affecting its overall integrity. Secondly, the coverage area of ​​sensors is limited by the number and location of sensors, making it difficult to comprehensively reflect the deformation across the entire tunnel area, and potentially leaving blind spots for localized deformation in critical areas. Furthermore, factors such as vibration and electromagnetic interference in the tunnel construction environment can easily lead to sensor malfunctions or data drift, affecting the reliability of monitoring data. Additionally, the high maintenance and replacement costs of sensors increase the economic burden of engineering monitoring.

[0005] With the development of computer vision technology, machine vision monitoring methods are increasingly being applied to tunnel deformation monitoring. Existing machine vision methods primarily acquire images of the tunnel surface and use image processing algorithms for deformation analysis. However, these methods often rely solely on image information and lack consideration of the tunnel's structural mechanical parameters. Since tunnel deformation is the result of the combined effects of structural stress and geological conditions, single image feature analysis is insufficient to accurately reflect the intrinsic mechanism of deformation. It is easily affected by factors such as changes in lighting, image noise, and differences in surface texture, leading to low deformation recognition accuracy and failing to meet the engineering requirements for monitoring accuracy and reliability. Therefore, how to integrate multi-source monitoring data to improve the real-time performance, comprehensiveness, and accuracy of tunnel construction deformation monitoring has become an urgent problem to be solved in the field of tunnel engineering monitoring. Summary of the Invention

[0006] The purpose of this invention is to provide a machine vision-based method for monitoring tunnel construction deformation, in order to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a method for monitoring tunnel construction deformation based on machine vision, the method comprising:

[0008] Real-time image acquisition of the tunnel construction structure was performed to obtain a sequence of images of the tunnel surface, while simultaneously monitoring the displacement and strain parameters of the tunnel structure.

[0009] Based on the displacement and strain parameters, the initial deformation is predicted to obtain the predicted deformation parameters.

[0010] Using machine vision equipment, images of key areas inside the tunnel are acquired. Based on the predicted deformation parameters, the images of key areas are input into multiple deformation feature recognizers to identify and obtain the actual deformation features.

[0011] By combining the predicted deformation parameters and the actual deformation characteristics, the final deformation trend analysis is carried out to obtain the tunnel deformation monitoring results.

[0012] Preferably, real-time image acquisition is performed on the tunnel construction structure to obtain a sequence of tunnel surface images, while simultaneously monitoring the displacement and strain parameters of the tunnel structure, including:

[0013] Multiple cameras are deployed at key locations in the tunnel to continuously capture images of the tunnel surface, forming a sequence of tunnel surface images.

[0014] Sensor networks are used to measure the real-time displacement changes of the tunnel structure to obtain displacement parameters.

[0015] Strain data of the tunnel material is collected using strain gauges to obtain strain parameters;

[0016] The displacement and strain parameters are correlated with the timestamps of the tunnel surface image sequence.

[0017] Preferably, based on displacement parameters and strain parameters, initial deformation prediction is performed to obtain predicted deformation parameters, including:

[0018] Collect historical tunnel construction data, extract sample displacement parameter sets and sample strain parameter sets, and label sample deformation dimensions to form a sample deformation parameter set;

[0019] A deformation prediction model is constructed based on time series analysis;

[0020] The deformation prediction model is trained and tested using sample displacement parameter sets, sample strain parameter sets, and sample deformation parameter sets as training and testing data. The model is optimized after the error rate reaches the target.

[0021] Input the displacement parameters and strain parameters into the deformation prediction model, and obtain the predicted deformation parameters from the prediction output.

[0022] The predicted deformation parameters are compared with the real-time displacement parameters to calibrate the model bias.

[0023] The parameters of the deformation prediction model are updated based on the calibration model bias.

[0024] Preferably, images of key areas inside the tunnel are acquired using machine vision equipment. Based on predicted deformation parameters, these key area images are input into multiple deformation feature recognizers to identify and obtain actual deformation features, including:

[0025] Machine vision equipment is used to focus on high-stress areas in the tunnel and collect images of key areas.

[0026] Based on the predicted deformation parameters, filter the range of matching deformation levels;

[0027] Select multiple deformation feature recognizers corresponding to the deformation level range, and each deformation feature recognizer includes multiple feature recognition paths based on image pattern matching;

[0028] The key region image is input into multiple deformation feature recognizers, and each deformation feature recognizer outputs a binary recognition result.

[0029] The proportion of binary recognition results that are "yes" is statistically analyzed, and the probability distribution of deformation level is calculated.

[0030] Select the deformation level with the highest probability as the actual deformation feature;

[0031] Based on the actual deformation characteristics, the acquisition angle of the key area image is adjusted accordingly.

[0032] Preferably, multiple deformation feature recognizers corresponding to the deformation level range are selected, and each deformation feature recognizer includes multiple feature recognition paths based on image pattern matching, including:

[0033] Multiple deformation feature recognizers corresponding to multiple deformation levels are pre-trained. The training data includes key region images of samples and binary results of sample deformation features.

[0034] The number of feature recognition paths is determined based on the error magnitude of the predicted deformation parameters.

[0035] Randomly activate feature recognition paths within multiple deformation feature recognizers;

[0036] Input the key region image into the activated feature recognition path and output a set of binary recognition results;

[0037] Aggregate the binary recognition result set and calculate the actual deformation features.

[0038] Preferably, by combining predicted deformation parameters and actual deformation characteristics, a final deformation trend analysis is performed to obtain tunnel deformation monitoring results, including:

[0039] Integrate predicted deformation parameters and actual deformation characteristics to form a deformation data vector;

[0040] Based on the moving window technique, time slicing is performed on the deformed data vector;

[0041] Analyze the deformation fluctuation characteristics within the time slice and calculate the average rate of change of the deformation fluctuation characteristics;

[0042] Construct deformation trend indicators, and generate tunnel deformation monitoring results based on these indicators;

[0043] The accuracy of the tunnel deformation monitoring results was verified by comparing them with historical data.

[0044] Preferably, based on the moving window technique, time slicing is performed on the deformed data vector, including: setting a fixed time window length and sliding step size, and dividing the deformed data vector according to the time sequence.

[0045] Preferably, the deformation trend index is constructed, including:

[0046] Analyze the direction of change in deformation data within the time slice, and calculate the upward and downward trends of deformation data.

[0047] Based on the upward and downward trend volumes, a modified trend indicator is obtained.

[0048] Preferably, the method further includes: formatting a report of the tunnel deformation monitoring results, transmitting the report to the central monitoring system, and triggering an alarm mechanism based on the tunnel deformation monitoring results.

[0049] Preferably, a machine vision-based tunnel construction deformation monitoring system is characterized in that it is used to implement the above-mentioned machine vision-based tunnel construction deformation monitoring method, the system comprising:

[0050] The image acquisition module is used to acquire real-time images of the tunnel construction structure, obtain image sequences of the tunnel surface, and simultaneously monitor the displacement and strain parameters of the tunnel structure.

[0051] The prediction module is used to predict the initial deformation based on the displacement and strain parameters, and obtain the predicted deformation parameters.

[0052] The feature recognition module is used to acquire images of key areas inside the tunnel using machine vision equipment. Based on the predicted deformation parameters, the key area images are input into multiple deformation feature recognizers to identify and obtain the actual deformation features.

[0053] The analysis module combines predicted deformation parameters and actual deformation characteristics to perform final deformation trend analysis and obtain tunnel deformation monitoring results.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] This machine vision-based method for monitoring tunnel construction deformation effectively overcomes the shortcomings of traditional monitoring methods through multi-dimensional data acquisition and fusion analysis. In the data acquisition phase, simultaneous acquisition of tunnel surface image sequences and monitoring of displacement and strain parameters breaks through the limitations of relying on a single data source. Image sequences can intuitively reflect the morphological changes of the tunnel surface, while displacement and strain parameters reveal the stress and deformation characteristics inside the structure from a mechanical perspective. The combination of these two provides richer basic data for deformation analysis, avoiding the one-sidedness of information caused by relying solely on images or a single mechanical parameter.

[0056] The initial deformation prediction stage, based on displacement and strain parameters, allows for a preliminary assessment of deformation trends from the perspective of structural mechanical response, providing clear guidance for subsequent image acquisition and analysis of key areas. This mechanical parameter-based prediction method makes image acquisition by machine vision equipment more targeted, avoiding the waste of computing power and information redundancy caused by indiscriminate image acquisition of the entire tunnel area, thus improving monitoring efficiency. By focusing on key areas, it is possible to more accurately capture parts that may undergo significant deformation, reducing the interference of invalid data on monitoring results.

[0057] In the actual deformation feature recognition process, the key area image is input into multiple deformation feature recognizers by combining predicted deformation parameters, fully leveraging the advantages of multi-recognizer collaborative analysis. Different deformation feature recognizers can perform specialized recognition for different manifestations of tunnel deformation (such as crack propagation, surface settlement, and support structure displacement), comprehensively capturing various deformation features in key areas. This segmented recognition approach avoids the omission or misjudgment of complex deformation features by a single recognizer, improving the completeness and accuracy of actual deformation feature extraction.

[0058] The final deformation trend analysis, by combining predicted deformation parameters and actual deformation characteristics, achieves a deep integration of mechanical prediction and visual observation. Predicted deformation parameters provide possible deformation trends from a theoretical perspective, while actual deformation characteristics verify and supplement the specific manifestations of deformation from an observational perspective. The two mutually corroborate and complement each other, providing a more comprehensive reflection of the deformation state of the tunnel structure. This fusion analysis approach not only considers the mechanical response laws of the structure but also incorporates observed surface morphology changes, making the deformation monitoring results more closely aligned with engineering realities and enabling a more accurate assessment of the tunnel structure's safety status, providing a more reliable basis for construction decisions. Simultaneously, the entire monitoring process achieves automated data acquisition and analysis, reducing manual intervention, improving the real-time nature and continuity of monitoring, and enabling the timely detection of potential deformation risks. Attached Figure Description

[0059] Figure 1 This is a schematic diagram illustrating the working principle of the machine vision-based tunnel construction deformation monitoring method described in this invention.

[0060] Figure 2 A detailed flowchart for initial deformation prediction;

[0061] Figure 3 A detailed flowchart for the deformation feature recognizer and path selection;

[0062] Figure 4 A detailed flowchart for the final deformation trend analysis. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] Please see Figure 1 This invention provides a machine vision-based tunnel construction deformation monitoring system and method, the method comprising:

[0065] A camera array deployed at key locations in the tunnel acquires real-time image sequences of the tunnel surface, while a displacement sensor network and strain gauges simultaneously collect structural displacement and material strain parameters. The image sequences and sensor data are spatiotemporally aligned using a unified timestamp. A deformation prediction model trained on historical construction data analyzes the current displacement and strain parameters, outputting predicted deformation parameters. Machine vision equipment dynamically adjusts the shooting angle based on the prediction results, focusing on high-stress areas to acquire images of key regions. Multiple parallel deformation feature recognizers perform multi-path feature analysis on the images, outputting actual deformation features. Predicted parameters and actual features are fused together by a data fusion module to generate a deformation data vector, which, after moving window analysis and trend index calculation, outputs the final monitoring results. The monitoring results automatically generate standardized reports and trigger a tiered early warning mechanism, achieving closed-loop processing from data acquisition to decision support.

[0066] Example 1: See Figure 2 The tunnel surface image acquisition system adopts a distributed deployment scheme, installing weather-resistant industrial cameras on the tunnel arch, sidewalls, and invert. The cameras are IP67 protected, have built-in anti-fog heating modules, and are equipped with 2-megapixel CMOS sensors and f / 1.8 large-aperture lenses. Installation locations are determined based on ground-penetrating radar scan results, focusing on fault zones, densely jointed areas, and joints in the support structure. The spacing between adjacent cameras is controlled within 3-5 meters to form overlapping field-of-view coverage. Image acquisition is performed at a rate of 10 frames per second, with a resolution of 1920×1080 pixels, and transmitted via a fiber optic ring network after H.265 encoding compression. The synchronously deployed displacement monitoring system uses Bragg grating sensing technology, pre-embedding 48-core sensing optical fibers within the initial support concrete, with a measurement point every 2 meters along the tunnel's longitudinal direction. The fiber optic demodulator acquires data at a 100Hz sampling rate, achieving three-dimensional displacement measurement through a phase demodulation algorithm, with an axial accuracy of 0.02mm and a radial accuracy of 0.05mm. Strain monitoring employs a full-bridge resistance strain gauge, with 120Ω foil strain gauges welded to key nodes of the steel arch frame to form a Wheatstone bridge circuit. The data acquisition module is equipped with a temperature compensation unit, acquiring micro-strain data at a frequency of 50Hz, with a measurement range of ±3000με and a resolution of 1με. All sensor data is transmitted via a 5G industrial router using the MQTT over TLS protocol, with each data packet appended with a GPS timestamp accurate to milliseconds. After receiving the data, the central processing system establishes a mapping relationship between image frames and sensor data using a time-series alignment algorithm, controlling the alignment error within 20 milliseconds.

[0067] The deformation prediction model was constructed through two stages: offline training and online prediction. Historical data collection covered at least three complete construction cycles of similar tunnel projects, with data samples including displacement time-series data, strain distribution data, and corresponding laser scanning deformation measurements. In the data preprocessing stage, cubic spline interpolation was used to fill in missing values, and Z-score standardization was used to eliminate dimensional differences. The model architecture employed a bidirectional LSTM neural network. The input layer had 12 neurons corresponding to the six displacement components (three each in the axial and radial directions) and the mean values ​​of the six strain zones. The hidden layer consisted of three layers: a first layer with 64 LSTM units, a second layer with 32 GRU units, and a third layer with 16 fully connected nodes. The output layer had three neurons, outputting the predicted axial deformation, radial convergence value, and crown settlement value. The training process used a sliding window strategy with a window length of 120 minutes and a sliding step size of 10 minutes. The Nadam algorithm was used as the optimizer, and the loss function was a weighted combination of smoothed L1 loss and cosine similarity. The model undergoes full training every 24 hours, retaining dynamic samples from the most recent 30 days for training data. During online prediction, real-time sensor data is input into the model after being filtered by the moving average, and a set of prediction parameters is generated every 30 seconds. The system compares the predicted displacement values ​​with the measured values ​​from the fiber optic sensors in real time. When the absolute deviation of five consecutive sampling points exceeds 0.15 mm, the model's incremental update mechanism is triggered. The update process employs an elastic weight fixation algorithm, adjusting the weights of the fully connected layers while preserving important parameters, with the update time controlled within 200 milliseconds.

[0068] The displacement parameter acquisition system employs distributed fiber optic acoustic sensing technology, laying single-mode sensing fibers along the tunnel's longitudinal direction. The demodulation equipment emits a 1550nm frequency-modulated continuous wave, analyzing the phase change of the Rayleigh scattering signal using coherent detection principles. Data processing utilizes the φ-OTDR algorithm, with a spatial resolution of 1 meter, covering a 2000-meter tunnel section. The system performs baseline calibration every 5 minutes to eliminate the effects of temperature drift. The strain monitoring network uses a hybrid network of vibrating wire strain gauges and resistance strain gauges, deploying vibrating wire sensors in areas of concrete stress concentration and installing resistance strain gauges at key points of steel reinforcement stress. The data acquisition terminal is equipped with a 6-channel 24-bit ADC, with a configurable sampling rate of 10-100Hz. The sensor network adopts a star topology, with regional aggregation nodes connected to the central server via an industrial Ethernet network. The spatiotemporal alignment system employs an improved dynamic time warping algorithm to establish a temporal correspondence between image sequences and sensor data. The alignment process first extracts the image timecode and sensor UTC timestamp, compensates for transmission delay through linear interpolation, and finally establishes a spatiotemporal index table with millisecond-level accuracy.

[0069] The model training dataset contains three sets of data: displacement parameters, strain parameters, and corresponding total station measured deformation. During data augmentation, time warp and Gaussian noise injection strategies were employed to expand the sample size to five times that of the original data. Network training adopted a phased strategy: the first phase froze the LSTM layer weights, training only the fully connected layers; the second phase unfroze all parameters for fine-tuning. Regularization measures included a weight decay coefficient set to 1e-4 and the application of Dropout with a 50% probability to the hidden layers. Model deployment utilized the TensorRT optimized inference engine, running on a Jetson AGXXavier edge computing device. The prediction service was exposed via a gRPC interface, supporting 100 concurrent calls per second. The online calibration module set two levels of deviation thresholds: when the instantaneous deviation exceeded 0.3 mm or the average deviation after 10 consecutive calibrations exceeded 0.1 mm, a model retraining task was automatically generated. The retraining process employed a transfer learning strategy, retaining the convolutional layer weights and updating only the fully connected layer parameters, with training time controlled within 15 minutes. The calibrated model is gradually replaced with the online service using a canary deployment strategy to ensure uninterrupted operation of the prediction service.

[0070] The system implements multi-source data fusion processing, with the image acquisition unit and sensor network achieving microsecond-level time synchronization via the PTPv2 protocol. The displacement data processing flow includes three steps: outlier filtering (based on the 3σ criterion), trend term removal (EMD decomposition), and downsampling (Kalman filtering). Strain data is used to generate a two-dimensional strain contour map using spatial interpolation, and key feature values ​​are extracted through contour lines. The prediction model input vector is constructed using a sliding window method with a window length of 60 minutes, containing 720 displacement samples and 3600 strain samples. During feature engineering, 12 statistical features are extracted: mean, variance, range, and autocorrelation coefficient of displacement data; kurtosis, skewness, energy entropy, and power spectral density of strain data. Post-processing of the model output uses exponentially weighted moving average smoothing with a smoothing coefficient α set to 0.2. The deviation detection module uses the CUSUM control chart algorithm, setting a cumulative deviation threshold of 0.8 mm. When a model update is triggered, the system automatically creates a training task queue, prioritizing data from periods with high deviation. After training, A / B testing is used to verify the model improvement effect. The entire data processing flow runs in a containerized environment, with Kubernetes enabling elastic scheduling of computing resources.

[0071] Example 2: See Figure 3The key area image acquisition utilizes a high-precision gimbal camera system equipped with a three-axis servo motor drive, enabling ±180° horizontal rotation and +90° / -30° pitch adjustment. The gimbal features a 20-megapixel global shutter CMOS sensor, paired with an f / 2.8-16 motorized zoom lens, achieving a resolution of 50μm / pixel at a minimum focusing distance of 0.3 meters. After receiving the predicted deformation parameters, the system determines the center point coordinates of the target area using a spatial coordinate transformation algorithm. The positioning process employs an iterative nearest-point algorithm to match the predicted deformation position with the actual 3D point cloud model, controlling the positioning error within 2 cm. When the predicted value enters the Level I (0-3mm), Level II (3-6mm), or Level III (>6mm) warning range, the gimbal automatically adjusts to the preset observation pose. For Level III warning areas, the system activates macro mode, automatically switching the lens focal length to 150mm, and providing shadowless illumination with a ring LED fill light group.

[0072] The deformation feature recognition system comprises nine independent recognizers, with three recognizers corresponding to each deformation level. The recognizers employ a containerized deployment architecture, with each recognizer containing three feature processing paths: the first path uses an improved U-Net convolutional neural network, extracting multi-scale features from the input image through five downsampling and five upsampling operations; the second path applies a multi-directional Sobel operator for edge enhancement, combined with Hough transform to detect linear features; and the third path uses a local binary mode and gray-level co-occurrence matrix fusion algorithm to extract texture features. Each path outputs a binary recognition result, which is then integrated and judged through a weighted voting mechanism. The weight allocation is dynamically adjusted based on the path's historical accuracy, with an initial weight set to 0.4:0.3:0.3.

[0073] The recognizer training employs a two-stage transfer learning strategy. The base model uses a ResNet-101 architecture pre-trained on the COCO dataset, with the input layer adjusted to 640×480 pixels to fit the tunnel image scale. Fine-tuning training uses an annotated dataset containing typical deformation features such as cracks, spalling, and leakage, with annotation information including defect type, geometric size, and spatial distribution. Data augmentation employs random rotation (±15°), brightness adjustment (±20%), and local occlusion strategies. The training process uses a hierarchical learning rate setting: 1e-5 for convolutional layers and 1e-4 for fully connected layers. The focus loss function addresses the sample imbalance problem, and the RAdam algorithm is selected as the optimizer.

[0074] During the online identification phase, the system activates the corresponding recognizer cluster based on the predicted deformation level. Each recognizer employs a random path selection mechanism, randomly activating two feature paths for each processing iteration. The outputs of the six activated paths are input to the decision fusion module, which calculates the crack identification confidence score and the estimated deformation. When the confidence score falls below the 85% threshold, a multi-view re-acquisition mechanism is triggered. The re-acquisition strategy generates three supplementary observation angles: a main viewpoint ±15° horizontal offset and a +10° pitch viewpoint. New images are input into the recognition system for cross-validation, and the final result is the median of the multi-view identification results.

[0075] The image processing workflow comprises three stages: preprocessing, feature extraction, and decision fusion. In the preprocessing stage, lens distortion correction and perspective transformation are performed to convert the tilted image into an orthographic projection view. In the feature extraction stage, a convolutional neural network path outputs a 1024-dimensional feature vector, an edge detection path generates an edge density distribution map, and a texture analysis path calculates texture response values ​​in eight directions. The decision fusion module employs a fuzzy inference system, with input parameters including crack continuity index, edge density peak value, and texture disorder. The output results include deformation level classification and crack width estimation with millimeter-level accuracy.

[0076] The system employs a dynamic feedback mechanism, adjusting acquisition parameters based on the recognition results. Upon identifying a Level III deformation feature, the sampling rate is automatically increased to 30 frames per second, and structured light projection mode is activated to acquire the 3D shape. If the recognition confidence level remains below the threshold, the system automatically switches to a higher resolution mode and increases the supplementary lighting intensity. All recognition results are associated with spatial coordinate information and displayed in real-time in the tunnel BIM model. The data processing pipeline adopts an asynchronous architecture, with image acquisition, feature extraction, and result fusion distributed across three computing nodes, executing in parallel, with single-frame processing latency controlled within 300 milliseconds.

[0077] The deformation feature database employs a hierarchical storage structure, retaining original images for 30 days, feature vectors for 180 days, and recognition results permanently. The query interface supports searches by station range, deformation level, and time interval. The system performs self-diagnosis every 24 hours, verifying the recognizer's accuracy using standard test images; if the deviation exceeds 5%, a model retraining process is automatically triggered. The retraining process uses an online learning strategy, incrementally updating the fully connected layer parameters while maintaining the convolutional layer weights unchanged. The updated model is validated through shadow mode; once confirmed to be error-free, it is switched to the production environment.

[0078] The gimbal control system employs a closed-loop feedback mechanism, with position accuracy verified by both a rotary encoder and a visual odometry. The motion control algorithm uses adaptive PID control, dynamically adjusting the proportional coefficient based on the target distance. The image stabilization mechanism combines electronic image stabilization with mechanical vibration damping to maintain image clarity even in construction vibration environments. The lighting system features eight independently controllable LED arrays with an adjustable color temperature within the 3000K-6500K range, automatically adjusting output power based on ambient light intensity. All equipment parameters are mapped in real-time through a digital twin system, allowing operators to remotely adjust the shooting parameters of any camera from the control center.

[0079] Example 3: See Figure 4 The deformation data fusion processing adopts a hierarchical and progressive architecture, integrating predicted deformation parameters with actual deformation features into a unified representation. The data vector construction process first performs principal component analysis on the displacement parameters, extracting the first three principal components as displacement components; after generating a two-dimensional distribution field for the strain parameters through Kriging interpolation, the gradient magnitudes in eight directions are calculated as strain components; the visual features are compressed into a 256-dimensional feature vector by a convolutional neural network, and then dimensionality reduction using t-SNE is applied to obtain three visual components. The final 14-dimensional data vector is represented as V=[d1,d2,d3,s1,...,s8,v1,v2,v3], where d represents the principal displacement component, s represents the strain gradient, and v corresponds to the visual features. The vector elements are scaled to the [0,1] interval using min-max normalization to eliminate dimensional differences.

[0080] The time window processing employs a dual buffering mechanism, setting a fixed 60-second main window and a 10-second sliding sub-window. The main window buffer uses a circular queue structure with a capacity of 360 sampling points (corresponding to a 6Hz sampling rate). As the window slides, new data overwrites the oldest data, maintaining temporal continuity. The sub-window processing unit performs discrete wavelet transform on the data within the window, using the db4 wavelet basis function for 5-level decomposition, extracting approximation coefficients and detail coefficients as time-frequency features. The fluctuation characteristics of the displacement components are obtained by calculating the mean of the absolute values ​​of the first-order differences, expressed by the formula:

[0081]

[0082] in: Indicates the intensity of displacement fluctuation. This represents the number of sampling points within the window. This represents the principal component value of the displacement at the i-th sampling point. The strain component fluctuation analysis employs the directional gradient method to calculate the Euclidean distance of the strain gradient between adjacent sampling points. The visual component uses an optical flow algorithm to extract the motion vectors of feature points, and the histogram entropy value of the vector amplitude is used as a fluctuation index.

[0083] The deformation trend analysis module employs a multi-scale feature fusion strategy. Within a 60-second window, the linear regression slope and intercept of each component are calculated; at a 10-minute macroscopic scale, the trend and periodic terms are decomposed using Hodrick-Prescott filtering. The trend index construction process first performs a Hilbert transform on the displacement component to solve for the instantaneous frequency and phase; the strain component obtains the energy change rate by calculating the time derivative of the strain energy density; and the visual component uses a three-dimensional LBP operator to extract spatiotemporal texture features. All features are weighted and fused using an attention mechanism, with the attention weights adaptively adjusted based on the feature variance.

[0084] The risk scoring model employs a three-layer feedforward neural network. The input layer receives 14 feature parameters, the hidden layer has 32 neurons, and the output layer generates a risk value from 0 to 100. The network training uses Wasserstein distance as the loss function, and the optimization process applies a gradient penalty strategy. The model is automatically updated every 24 hours, and the training data retains labeled samples from the most recent 7 days. After scoring, the results are filtered using a moving average, with a window width of 5 sampling points and a smoothing coefficient of 0.3.

[0085] The anomaly detection subsystem employs the isolation forest algorithm, constructing 100 isolation trees in the feature space. Each tree is built by randomly selecting features and segmentation values, and the anomaly score is calculated as the average path length required to isolate a data point. When the anomaly score exceeds a dynamic threshold, a detailed diagnostic process is triggered. The diagnostic process includes three steps: displacement-strain cross-validation, multi-time period data comparison, and spatial correlation analysis. The validation results generate an anomaly report, indicating the possible cause types and impact ranges.

[0086] The data visualization system maps the analysis results into a 3D heatmap, which is then overlaid on the tunnel BIM model. The heatmap uses the HSL color space for color coding; hue represents deformation type, saturation reflects the rate of change, and brightness corresponds to the risk level. The view supports timeline dragging, allowing users to revisit the deformation state at any given moment. Interactive features include profile analysis tools, contour line generators, and trend comparison views. All visualization elements are rendered using WebGL acceleration, supporting real-time synchronous viewing across multiple devices.

[0087] In terms of system implementation, the data processing pipeline adopts a microservice architecture, comprising five service units: data access, window computation, feature extraction, trend analysis, and result output. Services communicate via the gRPC protocol, and message serialization uses the Protocol Buffers format. Computationally intensive tasks are deployed on GPU nodes, and a Redis cluster is used to store real-time data in the in-memory database. The task scheduler is built on Kubernetes, supporting elastic scaling of computing resources. The quality monitoring module continuously tracks the processing latency of each service, automatically triggering horizontal scaling when the P99 latency exceeds 500 milliseconds.

[0088] The deformation feature database employs a time-partitioned storage strategy, with daily data stored independently in a columnar database. The Zstandard compression algorithm strikes a balance between compression ratio and decompression speed. The query interface supports SQL-like syntax, providing multi-dimensional filtering by spatial region, time range, and feature value. Data export formats include CSV, JSON, and Parquet to meet the needs of different analysis tools. All data operations are logged in an audit log, which is secured by blockchain technology to ensure log immutability.

[0089] System integration testing employed fault injection methods to simulate abnormal scenarios such as network latency, data loss, and compute node failure. The recovery mechanism included three strategies: data retransmission, computation rollback, and service degradation. Performance testing showed that under an input pressure of 1000 data points per second, the end-to-end processing latency remained stable within 800 milliseconds. Resource monitoring showed that under normal operating conditions, CPU utilization remained between 40-60%, and memory usage did not exceed 70% of the allocated amount. The security protection system comprised three lines of defense: transport layer encryption, access control lists, and behavior auditing. Penetration testing was conducted regularly to update the protection rules.

[0090] Example 4: Deformation trend analysis employs a dual-time window processing mechanism, with the main window width set to 300 seconds and the sub-window width to 60 seconds. The system slides the sub-window every 10 seconds and refreshes the main window data every 300 seconds. In actual monitoring of a certain tunnel section, the system recorded data changes over 30 consecutive minutes. The displacement component processing flow first performs linear fitting on the displacement sequence within the main window, calculating the slope angle of the fitted line. When continuous deformation occurred at monitoring station K25+380, the system captured the process of the displacement slope gradually increasing from -0.5° to +2.3°. Positive slope angle values ​​indicate an upward trend, negative values ​​indicate a downward trend, and the absolute value reflects the intensity of the change. Simultaneously, the coefficient of determination is calculated to assess the reliability of the linear fitting; when the coefficient is below 0.6, nonlinear trend analysis is initiated.

[0091] Strain component analysis employs an energy density calculation method. The system integrates the strain gradient in eight directions within each sub-window to obtain the strain energy density value. Comparison of energy density values ​​between adjacent sub-windows is calculated using a relative rate of change, calculated as the absolute value of the difference between the current and previous values ​​divided by the previous value. When the energy density in the K25+380 segment increases from 1.8 J / m³ to 2.7 J / m³ between 10:15 and 10:20, the system calculates a 50% rate of change. A warning flag is triggered when this rate of change exceeds a 25% threshold.

[0092] Visual texture feature extraction employs an improved LBP-TOP algorithm. The system treats the image sequence within a 60-second sub-window as a spatiotemporal cube, calculating local binary patterns in the XY plane and dynamic texture changes in the XT and YT planes. Each sub-window outputs a texture disorder index, ranging from 0 to 1, with higher values ​​indicating more drastic surface texture changes. At 10:15 in the K25+380 segment, the system detected a jump in the texture disorder of the arch concrete surface from 0.35 to 0.68, indicating abnormal surface deformation.

[0093] The trend indicators are represented by a three-dimensional vector: [displacement trend angle, strain rate of change, texture disorder]. The changes in trend indicators for five consecutive time points in this segment are shown in Table 1.

[0094] Table 1: The trend indicators of this section at 5 consecutive time points are as follows.

[0095]

[0096] Risk value mapping is implemented using a radial basis function network. The network input layer receives three trend indicators, and the hidden layer uses 24 Gaussian kernel functions, with the center points determined based on historical data clustering. The output layer generates risk values ​​from 0 to 100 through linear combination. Based on the data in the table above, the system calculated a risk value of 73.5 at 10:15:20, exceeding the orange alert threshold of 60. Network parameters are automatically updated weekly, and training data uses labeled samples from the most recent three months.

[0097] The upward trend was calculated using a modified Mann-Kendall test. The system assessed the significance of the trend in the displacement sequence by statistically analyzing the proportion of sampling points that met the following criteria: three consecutive points showing a monotonically increasing trend with the change exceeding the measurement error. In the K25+380 case, 83% of the sampling points detected between 10:15 and 10:18 met the upward trend criteria. The downward trend was determined through strain energy density analysis, calculating the average negative gradient of the energy density curve. A significant downward trend was identified when this value remained below -0.15 J / m³·s for three consecutive minutes.

[0098] In terms of system implementation, the trend analysis engine is deployed on a dedicated computing node, equipped with dual Xeon processors and 128GB of memory. The data processing pipeline consists of four stages: the data standardization stage scales the raw parameters to a uniform dimension; the feature extraction stage computes three trend indicators in parallel; the fusion stage constructs the input vector; and the evaluation stage outputs a risk value. The entire process adopts a pipelined parallel design, with a single analysis latency controlled within 150 milliseconds.

[0099] The visualization system maps trend indicators to dynamic radar charts, with three axes representing displacement, strain, and texture trends, respectively. In the K25+380 case, the radar chart at 10:15:20 shows significant outward expansion of the strain and texture axes, while the displacement axis exhibits a sharp angle. Simultaneously, a heatmap is displayed showing the longitudinal risk distribution of the tunnel, with high-risk sections marked in dark red. The user interface allows users to click on any data point to view a detailed trend breakdown chart.

[0100] The anomaly handling mechanism includes a trend conflict detection function. When the displacement trend and strain trend change in opposite directions (e.g., displacement increases while strain decreases), the system automatically triggers a review process. The review process includes checking the sensor calibration status, re-acquiring image data, and manual confirmation. In the K25+380 case, the system detected a mismatch between displacement and texture trends at 10:15:25. Review revealed that this was due to texture analysis deviation caused by dirt on the camera lens.

[0101] Data storage employs a tiered archiving strategy. Raw trend data is retained for 30 days, feature vectors for 180 days, and risk assessment results are stored permanently. The query interface supports searches by time range, spatial location, and risk level. The system generates a trend analysis report weekly, including change graphs and statistical summaries for each indicator. All analysis results are signed with digital certificates to ensure data integrity and traceability.

[0102] The quality control module performs standard tests periodically. The test dataset contains trend features of 12 typical deformation scenarios, and the system runs the test process every 24 hours. When the detection accuracy deviation exceeds 5%, the model retraining process is automatically triggered. The training process adopts an incremental learning approach, incorporating new samples while retaining existing knowledge, and the training time is controlled within 30 minutes. After the updated model is verified through shadow runs, it is switched to the production environment.

[0103] Example 5: The monitoring result report adopts a structured data encapsulation format, defining a data pattern containing six core fields. The tunnel station number field records the mileage location of the monitoring point, using a kilometer + meter format accurate to the centimeter level. The timestamp field records UTC time with added time zone information, achieving millisecond-level accuracy. The deformation parameter field stores 12 indicators, including three-dimensional displacement, principal strain direction, and maximum strain value. The feature recognition result field includes deformation level classification, crack width estimation, and defect distribution coordinates. The trend index field stores three floating-point values: displacement trend angle, strain change rate, and texture disorder. The risk level field uses an enumeration type to label four levels: safe, caution, warning, and danger. The report generation module performs data packaging every 10 seconds, using JSON-LD format for semantic description, compressing the file size to less than 5KB.

[0104] The central monitoring system employs a distributed message processing architecture, deploying an Apache Kafka cluster as the data bus. The cluster is configured with three Broker nodes, each with a dedicated topic for report data, divided into three partitions. The producer API is integrated into the monitoring terminal, pushing report data via persistent TCP connections. The consumer group comprises three instances: an alarm service, a storage service, and a visualization service, using a collaborative consumption model to achieve load balancing. The message queue retains data for the most recent 24 hours, and SSD arrays ensure high throughput. The system's peak processing capacity reaches 2000 reports per second, with an average end-to-end latency controlled within 50 milliseconds.

[0105] The alarm triggering mechanism implements a tiered response strategy. When the risk level field enters the alert state (corresponding to a value of 30-60), a yellow alert procedure is triggered: the control center console illuminates a yellow indicator light, the electronic signboard marks the affected chainage range, and a lightweight notification is pushed to mobile devices. When the risk value rises to the 60-80 range, an orange alert is triggered: the audible and visual alarm group is activated, voice prompts are played repeatedly, a response plan including support and reinforcement suggestions is automatically generated, and an alarm message is sent to the responsible engineer via SMS. When the risk value exceeds 80, a red alert state is entered: the construction power to the affected area is cut off, an emergency evacuation broadcast is triggered, and the tunnel structure safety assessment model is simultaneously activated. All alert events are recorded with a response timeline accurate to milliseconds.

[0106] The data persistence system employs a hybrid storage architecture. Raw monitoring reports are written to an InfluxDB time-series database, stored in time-partitioned locations with spatial indexes. Analysis results are persisted to a PostgreSQL relational database, containing schema-based reports and associated metadata. A Redis cluster is deployed in the caching layer to store the most recent 15 minutes of hot data. The archiving system performs monthly cold data migrations, transferring historical data to an object storage system with a permanent retention policy. The query interface supports both ODBC and JDBC protocols, providing multi-dimensional retrieval capabilities based on time range, spatial location, and risk level.

[0107] The system interface services are exposed through an API gateway, deploying the OAuth 2.0 authentication protocol. External access requires an access token, which is valid for 2 hours. The data query interface supports both JSON and XML formats, with a promised response time of 99% of requests completed within 300 milliseconds. The management interface provides device status monitoring, alarm history query, and system configuration update functions; critical operations require dual authentication. Interface traffic is rate-limited, with a maximum of 100 requests per second per client. All transmitted data is encrypted using TLS 1.3, with critical fields additionally encrypted using AES-256.

[0108] The blockchain-based evidence storage system constructs a private blockchain network containing four consensus nodes. Each time a monitoring report is generated, a SHA-256 hash value is calculated and written to the blockchain as a digital fingerprint. When an alert event is triggered, complete event details, including response actions, operators, and execution time, are generated and uploaded to the blockchain as the Merkle root value. The audit trail interface supports inputting time ranges and operation types, returning immutable operation logs. Evidence storage data undergoes off-chain backup every 24 hours, with backup files stored in physical isolation.

[0109] The mobile interaction system develops cross-platform applications, supporting both iOS and Android systems. Alarm message push integrates APNs and FCM dual channels to ensure message delivery. The application interface is displayed in three sections: a real-time monitoring view shows a heatmap of tunnel longitudinal profile risks; a historical analysis view provides data trend playback for any time period; and an alarm management view lists unprocessed alarms. Offline mode caches data from the last 72 hours, automatically synchronizing operation records after network recovery. Device binding uses a two-way authentication mechanism, allowing remote data erasure in case of device loss.

[0110] The system maintenance module enables full lifecycle management. The configuration center supports dynamic parameter updates, with changes taking effect immediately without a restart. The log system collects runtime logs from all components and performs centralized analysis via the ELK stack. The monitoring panel displays 50 metrics in real time, including CPU load, memory usage, and network traffic. The self-diagnostic program performs hardware checks, data consistency verification, and performance benchmark tests every morning. The maintenance terminal provides a command-line interface to support remote troubleshooting and system repair.

[0111] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0112] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring tunnel construction deformation based on machine vision, characterized in that, The method includes: Real-time image acquisition of the tunnel construction structure was performed to obtain a sequence of images of the tunnel surface, while simultaneously monitoring the displacement and strain parameters of the tunnel structure. Based on the displacement and strain parameters, the initial deformation is predicted to obtain the predicted deformation parameters. Using machine vision equipment, images of key areas inside the tunnel are acquired. Based on the predicted deformation parameters, the images of key areas are input into multiple deformation feature recognizers to identify and obtain the actual deformation features. By combining the predicted deformation parameters and the actual deformation characteristics, the final deformation trend analysis is carried out to obtain the tunnel deformation monitoring results; Machine vision equipment is used to acquire images of key areas inside the tunnel. Based on predicted deformation parameters, these images are input into multiple deformation feature recognizers to identify and obtain the actual deformation features, including: Machine vision equipment is used to focus on high-stress areas in the tunnel and collect images of key areas. Based on the predicted deformation parameters, filter the range of matching deformation levels; Select multiple deformation feature recognizers corresponding to the deformation level range, and each deformation feature recognizer includes multiple feature recognition paths based on image pattern matching; The key region image is input into multiple deformation feature recognizers, and each deformation feature recognizer outputs a binary recognition result. The proportion of binary recognition results that are "yes" is statistically analyzed, and the probability distribution of deformation level is calculated. Select the deformation level with the highest probability as the actual deformation feature; Based on the actual deformation characteristics, the acquisition angle of the key area image is adjusted accordingly.

2. The tunnel construction deformation monitoring method based on machine vision according to claim 1, characterized in that, Real-time image acquisition of the tunnel construction structure yields a sequence of images of the tunnel surface, while simultaneously monitoring the displacement and strain parameters of the tunnel structure, including: Multiple cameras are deployed at key locations in the tunnel to continuously capture images of the tunnel surface, forming a sequence of tunnel surface images. Sensor networks are used to measure the real-time displacement changes of the tunnel structure to obtain displacement parameters. Strain data of the tunnel material is collected using strain gauges to obtain strain parameters; The displacement and strain parameters are correlated with the timestamps of the tunnel surface image sequence.

3. The method for monitoring tunnel construction deformation based on machine vision according to claim 1, characterized in that, Based on the displacement and strain parameters, initial deformation prediction is performed to obtain the predicted deformation parameters, including: Collect historical tunnel construction data, extract sample displacement parameter sets and sample strain parameter sets, and label sample deformation dimensions to form a sample deformation parameter set; A deformation prediction model is constructed based on time series analysis; The deformation prediction model is trained and tested using sample displacement parameter sets, sample strain parameter sets, and sample deformation parameter sets as training and testing data. The model is optimized after the error rate reaches the target. Input the displacement parameters and strain parameters into the deformation prediction model, and obtain the predicted deformation parameters from the prediction output. The predicted deformation parameters are compared with the real-time displacement parameters to calibrate the model bias. The parameters of the deformation prediction model are updated based on the calibration model bias.

4. The tunnel construction deformation monitoring method based on machine vision according to claim 1, characterized in that, Select multiple deformation feature recognizers corresponding to the deformation level range. Each deformation feature recognizer includes multiple feature recognition paths based on image pattern matching, including: Multiple deformation feature recognizers corresponding to multiple deformation levels are pre-trained. The training data includes key region images of samples and binary results of sample deformation features. The number of feature recognition paths is determined based on the error magnitude of the predicted deformation parameters. Randomly activate feature recognition paths within multiple deformation feature recognizers; Input the key region image into the activated feature recognition path and output a set of binary recognition results; Aggregate the binary recognition result set and calculate the actual deformation features.

5. The tunnel construction deformation monitoring method based on machine vision according to claim 1, characterized in that, By combining predicted deformation parameters and actual deformation characteristics, a final deformation trend analysis is conducted to obtain tunnel deformation monitoring results, including: Integrate predicted deformation parameters and actual deformation characteristics to form a deformation data vector; Based on the moving window technique, time slicing is performed on the deformed data vector; Analyze the deformation fluctuation characteristics within the time slice and calculate the average rate of change of the deformation fluctuation characteristics; Construct deformation trend indicators, and generate tunnel deformation monitoring results based on these indicators; The accuracy of the tunnel deformation monitoring results was verified by comparing them with historical data.

6. The tunnel construction deformation monitoring method based on machine vision according to claim 5, characterized in that, Based on the moving window technique, time slicing is performed on the deformed data vector, including setting a fixed time window length and sliding step size, and dividing the deformed data vector according to the time series.

7. The tunnel construction deformation monitoring method based on machine vision according to claim 5, characterized in that, Construct deformation trend indicators, including: Analyze the direction of change in deformation data within the time slice, and calculate the upward and downward trends of deformation data. Based on the upward and downward trend volumes, a modified trend indicator is obtained.

8. The tunnel construction deformation monitoring method based on machine vision according to claim 1, characterized in that, Also includes: The tunnel deformation monitoring results are formatted into a report, which is then transmitted to the central monitoring system. An alarm mechanism is triggered based on the tunnel deformation monitoring results.

9. A tunnel construction deformation monitoring system based on machine vision, characterized in that, For implementing the machine vision-based tunnel construction deformation monitoring method according to any one of claims 1 to 8, the system comprises: The image acquisition module is used to acquire real-time images of the tunnel construction structure, obtain image sequences of the tunnel surface, and simultaneously monitor the displacement and strain parameters of the tunnel structure. The prediction module is used to predict the initial deformation based on the displacement and strain parameters, and obtain the predicted deformation parameters. The feature recognition module is used to acquire images of key areas inside the tunnel using machine vision equipment. Based on the predicted deformation parameters, the key area images are input into multiple deformation feature recognizers to identify and obtain the actual deformation features. The analysis module is used to combine predicted deformation parameters and actual deformation characteristics to perform final deformation trend analysis and obtain tunnel deformation monitoring results.

Images