A method and system for intelligent inspection of deformation of belt conveyor supports in trestle corridors

By employing a multimodal image preprocessing and multi-stage attention-based temporal prediction architecture, combined with a three-dimensional benchmark model, the accuracy and real-time performance issues of belt conveyor deformation detection within the trestle corridor were resolved. This enabled high-precision deformation detection and fault early warning, adapting to intelligent inspection under complex working conditions.

CN123066786APending Publication Date: 2026-11-10SHAANXI ZHENGYUAN SMART EQUIPMENT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611579388.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-13
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies cannot effectively solve the deformation detection of belt conveyor supports in the high dust, light fluctuation and strong vibration environment of the trestle corridor. They have problems such as insufficient detection accuracy, high false detection rate and inability to achieve all-weather real-time monitoring and early warning of faults.

Method used

By employing multimodal layered image preprocessing, weighted ICP two-level point cloud registration, multi-dimensional deformation quantization formula, and multi-stage attention temporal prediction architecture, combined with a three-dimensional benchmark model and a digital twin benchmark, a high-precision three-dimensional quantitative detection and deformation trend prediction system is achieved, thus constructing a closed-loop intelligent inspection system.

Benefits of technology

It improves the accuracy of support deformation detection under complex working conditions, significantly enhances environmental anti-interference ability, accurately distinguishes instantaneous vibration displacement from permanent structural deformation, reduces false detection rate, and realizes advanced fault prediction and all-weather real-time monitoring.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123066786A_ABST
    Figure CN123066786A_ABST
Patent Text Reader

Abstract

This invention discloses an intelligent inspection method and system for deformation of belt conveyor supports in trestle bridge corridors, relating to the field of machine vision processing technology. The method includes: utilizing a track-based inspection robot equipped with multiple sensors to collect images, point clouds, and attitude data; performing multi-modal preprocessing, including dust atomization removal, illumination equalization, and vibration correction, to eliminate interference from high dust levels, fluctuating illumination, and strong vibrations in the corridor; employing weighted ICP point cloud registration to achieve millimeter-level three-dimensional deformation-based detection; combining time-frequency analysis to distinguish between instantaneous vibration shifts and permanent structural deformations; constructing a MA-LSTM multi-stage attention time-series model, using two levels of attention to enhance the characteristics of key parts of the support and deformation acceleration periods, fitting the deformation evolution trend, and assessing fault risk; and establishing a closed-loop inspection mechanism of early warning, review, archiving, and iteration. This invention improves detection accuracy, enables early fault prediction, and is suitable for routine safety monitoring of belt conveyor supports in coal mines and ports.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of machine vision processing technology, and more specifically to an intelligent inspection method and system for deformation of belt supports in trestle corridors. Background Technology

[0002] The trestle corridor is the core load-bearing structure of the belt conveyor system. The internal belt supports serve as the main support structure, and defects such as bending, misalignment, loosening, corrosion, and deformation directly affect the operational safety of the belt conveyor system. If supports collapse or belt misalignment occurs, it can lead to major accidents such as production stoppages, equipment damage, and even casualties. Currently, the inspection of belt supports mainly relies on two methods: regular manual foot inspections and general visual inspection. Both have significant technical shortcomings, and there is no specific deformation inspection solution tailored to the special working conditions of the trestle corridor.

[0003] In the manual inspection mode, inspectors need to check section by section in a narrow, enclosed corridor with high dust and high noise. The working environment is harsh and the labor intensity is high. Furthermore, the inspection is limited by personnel experience, line of sight, and inspection frequency, resulting in a high rate of missed detection of early defects such as millimeter-level micro-bending and micro-displacement caused by loose bolts. At the same time, the inspection cycle is long, making it impossible to achieve all-weather real-time monitoring, making it difficult to detect sudden deformation failures in a timely manner, and posing safety hazards such as falls from heights and mechanical injuries.

[0004] Existing machine vision inspection solutions mostly use visible light cameras combined with convolutional neural networks for surface defect detection. Such solutions can achieve good results in open-air scenarios with stable lighting and clean environments, but they have limitations in adaptability to the special working conditions of trestle bridges and corridors. Insufficient environmental interference resistance: High dust concentration in the corridor can easily cause lens fogging and obstruction; large fluctuations in day and night lighting and uneven local lighting can easily cause overexposure / underexposure of images; vibration during equipment operation can easily cause image blurring and inter-frame misalignment; existing preprocessing algorithms such as single filtering and histogram equalization cannot eliminate multi-source interference at the same time, and there is a lack of quantitative image restoration calculation models, resulting in a high rate of loss of effective image features and a significant decrease in subsequent detection accuracy.

[0005] Insufficient detection dimensions and accuracy: Existing solutions are mostly based on feature comparison of two-dimensional images, which cannot accurately quantify the three-dimensional deformation of the support and is difficult to distinguish between instantaneous displacement and permanent structural deformation caused by vibration during normal operation of the support, resulting in a high false detection rate; some three-dimensional detection solutions using LiDAR have problems such as no weight optimization for point cloud registration and large matching errors, which make it difficult to meet the real-time and accuracy requirements of long-distance continuous inspection.

[0006] Insufficient intelligence and closed-loop capabilities: Existing technologies are mostly single-loop detection modes of acquisition-detection-alarm, without establishing a full life cycle deformation database for the support, unable to explore the deformation evolution law, and lack the ability to predict deformation trends and provide early warning of faults; at the same time, the inspection path is fixed, and it is impossible to dynamically adjust the inspection strategy according to the distribution of defects, resulting in low operation and maintenance efficiency and difficulty in achieving preventive operation and maintenance.

[0007] The time series prediction model is not targeted enough: existing deformation prediction mostly adopts general time series networks, which are not designed for the characteristics of large spatial differences in support deformation, nonlinear time evolution, and susceptibility to vibration and noise interference. They cannot simultaneously achieve focusing on key spatial parts and weighting of key time periods, resulting in insufficient accuracy in predicting the trend of early small deformations.

[0008] Therefore, how to propose an intelligent inspection method and system for the deformation of belt conveyor supports in the trestle corridor, and overcome the shortcomings of the existing technology, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0009] In view of this, the present invention provides an intelligent inspection method and system for the deformation of belt conveyor supports in trestle corridors, which is adapted to the special working conditions of trestle corridors with high dust, light fluctuations, strong vibrations, and narrow and enclosed spaces. Through a complete mathematical calculation model and a multi-stage attention-based temporal prediction architecture, it achieves high-precision three-dimensional quantitative detection of belt conveyor support deformation, advanced prediction of deformation trends, and fully closed-loop intelligent inspection. To achieve the above objectives, the present invention adopts the following technical solution: A method for intelligent inspection of deformation of belt conveyor supports in a trestle corridor includes: S1. Collect fault-free reference data of the entire belt support, and construct a three-dimensional reference model and digital twin reference body in combination with the original design parameters of the support, and calibrate the deformation threshold and vibration offset threshold of each part. S2. The inspection robot runs along the preset path and simultaneously collects visible light image sequences, three-dimensional point cloud data and attitude data of the belt support. S3. Preprocess the original image and output a standardized detection image; S4. Register and differentiate the three-dimensional reference model with the standardized detection image and real-time three-dimensional point cloud data respectively, quantify and calculate the deformation of each part of the support, and distinguish between instantaneous vibration displacement and permanent structural deformation by combining frequency characteristics, and identify the deformation type and location. S5. Based on the temporal deformation data of multi-cycle inspection, the probability and risk level of future failures are predicted by a multi-stage attention long short-term memory network model. S6. For deformation exceeding the threshold and high-risk faults triggering graded early warnings, the inspection path is adaptively adjusted to perform secondary verification scans, the inspection data is fully stored and the model is iteratively optimized to perform closed-loop intelligent inspection.

[0010] Optionally, step S1 includes: S1.1 controls the inspection robot to travel at a constant speed along the track at a calibrated speed, and simultaneously collects visible light reference video sequences, three-dimensional laser point cloud data and IMU attitude data of the belt support, and marks the key feature points of the support; S1.2 Based on the reference point cloud data, an improved iterative nearest point algorithm is used to register and stitch the point cloud to generate a three-dimensional point cloud reference model of the entire belt conveyor support; combined with the support design drawings and specifications, a parametric digital twin reference body is constructed to map the geometric dimensions and connection relationships of the support. S1.3 calibrate the allowable deformation thresholds for each part of the support, including bending deflection threshold, tilt angle threshold, horizontal displacement threshold, and bolt loosening displacement threshold; calibrate the vibration offset threshold to distinguish between normal vibration and structural deformation boundaries; and complete the registration of the multi-sensor spatial coordinate system.

[0011] Optionally, in step S2, a hardware-triggered synchronization mechanism is used to achieve microsecond-level synchronous acquisition of industrial cameras, lidar, and IMU.

[0012] Optionally, step S3 includes: S3.1 is based on an atmospheric scattering physics model. It automatically estimates dust concentration and atmospheric light value by combining a concentration sensing branch. It solves and corrects the transmittance through the dark channel prior. It combines a lightweight generative adversarial network to restore image details, remove dust fogging effect and retain key features of the support. S3.2 uses the multi-scale Retinex algorithm to decompose the incident and reflected components of the image, adaptively brightens dark areas and suppresses overexposed areas, and introduces an illumination smoothing constraint term to optimize the edge blur and halo of the support structure. S3.3 combines IMU attitude data to construct an affine transformation matrix, and combines ORB feature point matching and sparse optical flow method to calculate the global motion vector between frames, and performs pixel-level correction for image offset and misalignment caused by vibration.

[0013] Optionally, step S4 includes: S4.1 uses a lightweight semantic segmentation network to segment the support components of the preprocessed image, identify columns, beams, roller seats, and connecting bolts, and generate binary segmentation masks for each component to filter background interference. S4.2 Based on the previously calibrated camera intrinsic and extrinsic parameter matrices, establish the projection transformation relationship between 2D image pixels and the 3D world coordinate system, reverse project the binary segmentation mask of each component to the 3D space, filter out the point cloud subset of the corresponding component from the real-time 3D point cloud, and associate it with the corresponding component region in the 3D benchmark model. S4.3 adopts a two-level strategy of coarse registration and fine registration. For each component, its real-time point cloud subset is registered with the corresponding component region of the 3D benchmark model. First, coarse registration is completed by feature point matching, and then millimeter-level fine registration is completed by weighted iterative nearest point algorithm. The displacement of each point, beam deflection and column tilt angle are calculated by deformation transformation formula. S4.4 uses short-time Fourier transform to analyze the time-frequency characteristics of continuous frame deformation data, identifies high-frequency small-amplitude offsets as instantaneous vibration offsets, identifies low-frequency continuously increasing offsets as permanent structural deformations, and identifies deformation types such as bending, tilting, displacement, and loosening.

[0014] Optionally, the lightweight semantic segmentation network is an improved YOLO-Seg network with an inference speed of ≥30fps and a component segmentation accuracy of ≥95%; in the weighted iterative nearest point algorithm, the weight of the key feature point of the support is 1.5, and the weight of the ordinary structural point is 1.0.

[0015] Optionally, step S5 includes: S5.1 Organize historical inspection deformation data, construct a sample dataset according to time series, and expand the sample size by using time series interpolation enhancement for early small deformation samples. S5.2 Constructs a multi-stage attention long short-term memory network model. First, the first-level spatial attention mechanism is used to adaptively assign weights to the deformation features of multiple parts of the scaffold. Then, the LSTM network is used to extract the temporal evolution features. Subsequently, the second-level spatiotemporal fusion attention mechanism is used to achieve deep fusion of spatial and temporal features. Finally, the predicted features are output through residual splicing. S5.3 Input the latest inspection data, use the exponential growth model to fit the deformation evolution curve, calculate the probability of failure based on the Weibull distribution, and classify it into three risk levels: low, medium, and high.

[0016] Optionally, in step S5.2, the first-level spatial attention uses 1×1 convolution to generate query, key, and value vectors, and realizes spatial feature weighting through self-attention calculation; the second-level spatiotemporal fusion attention is based on the LSTM hidden state, fuses the features of the first-level spatial attention to calculate temporal weights, and weighted aggregates to obtain spatiotemporal features; the output adopts a structure of spatiotemporal aggregated features and original spatial feature residuals concatenated.

[0017] Optionally, step S6 includes: S6.1 triggers a first-level audible and visual warning and maintenance push for deformation exceeding the threshold, and triggers a second-level warning for high-risk predicted faults; S6.2 Based on deep reinforcement learning algorithm, dynamically adjust the inspection robot path and inspection frequency, and plan secondary scanning path for high-risk areas; S6.3 stores all inspection data throughout the entire process, forming an inspection log for the entire lifecycle of the stent, and regularly uses newly added data to iterate and optimize the detection and prediction models.

[0018] A smart inspection system for deformation of belt conveyor supports in a trestle corridor includes: a calibration module: used to collect fault-free reference data of the entire belt conveyor support, construct a three-dimensional reference model and a digital twin reference body in combination with the original design parameters of the support, and calibrate the deformation threshold and vibration offset threshold of each part; Acquisition module: Used for the inspection robot to run along a preset path and simultaneously acquire visible light image sequences, 3D point cloud data and attitude data of the belt support; Preprocessing module: Used to preprocess the original image and output a standardized detection image; Deformation recognition module: It is used to register and differentiate the three-dimensional reference model with the standardized detection image and real-time three-dimensional point cloud data, quantify and calculate the deformation of each part of the support, and distinguish between instantaneous vibration displacement and permanent structural deformation by combining frequency characteristics, and identify the deformation type and location. Prediction module: Based on time-series deformation data from multi-cycle inspections, it uses a multi-stage attention long short-term memory network model to predict the probability and risk level of future faults. Early warning and verification module: It is used to trigger graded early warnings for deformations exceeding the threshold and high-risk faults, adaptively adjust the inspection path to perform secondary verification scans, store all inspection data and iteratively optimize the model to perform closed-loop intelligent inspection.

[0019] As can be seen from the above technical solution, compared with the prior art, the present invention discloses an intelligent inspection method and system for deformation of belt conveyor supports in trestle corridors, which has the following beneficial effects: This invention addresses the unique working conditions of trestle bridges and corridors, characterized by high dust levels, fluctuating lighting, and strong vibrations. It constructs a multimodal layered image preprocessing system, employing a quantitative atmospheric scattering model, a multi-scale Retinex algorithm, and an affine correction model. This improves the effective feature retention rate of images, enhances the accuracy of support deformation detection under complex conditions, and significantly strengthens environmental anti-interference capabilities. Using weighted ICP two-level point cloud registration and multi-dimensional deformation quantification formulas, it can identify millimeter-level micro-deformations. Furthermore, through short-time Fourier transform frequency feature analysis, it accurately distinguishes between instantaneous vibration shifts and permanent structural deformations, reducing false detection rates. Based on a MA-LSTM multi-stage attention deformation trend prediction mechanism, it focuses on key stress areas through first-level spatial attention and on deformation acceleration periods through second-level spatiotemporal attention. Combined with residual structure to retain complete feature information, it improves prediction accuracy. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0021] Figure 1 This invention provides a schematic flowchart of an intelligent inspection method for deformation of belt conveyor supports in a trestle corridor. Detailed Implementation

[0022] 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.

[0023] This invention discloses an intelligent inspection method for deformation of belt conveyor supports in a trestle corridor, such as... Figure 1 As shown, it includes: S1. Collect fault-free reference data of the entire belt support, and construct a three-dimensional reference model and digital twin reference body in combination with the original design parameters of the support, and calibrate the deformation threshold and vibration offset threshold of each part. S2. The inspection robot runs along the preset path and simultaneously collects visible light image sequences, three-dimensional point cloud data and attitude data of the belt support. S3. Preprocess the original image and output a standardized detection image; S4. Register and differentiate the three-dimensional reference model with the standardized detection image and real-time three-dimensional point cloud data respectively, quantify and calculate the deformation of each part of the support, and distinguish between instantaneous vibration displacement and permanent structural deformation by combining frequency characteristics, and identify the deformation type and location. S5. Based on the temporal deformation data of multi-cycle inspection, the probability and risk level of future failures are predicted by a multi-stage attention long short-term memory network model. S6. For deformation exceeding the threshold and high-risk faults triggering graded early warnings, the inspection path is adaptively adjusted to perform secondary verification scans, the inspection data is fully stored and the model is iteratively optimized to perform closed-loop intelligent inspection.

[0024] In a specific implementation method, a smart inspection method for deformation of belt conveyor supports in a trestle corridor includes the following steps: S1 Preliminary Modeling and Parameter Calibration The track-mounted inspection robot, equipped with a high-definition industrial camera, a miniature 3D LiDAR, and an inertial measurement unit (IMU), completes data acquisition under fault-free conditions along the pre-set inspection track of the trestle corridor. Combined with the original design parameters of the support frame, a standardized 3D benchmark model and a digital twin mapping body are constructed to complete the system parameter calibration and threshold initialization.

[0025] S1.1 Full Reference Data Acquisition

[0026] The inspection robot is controlled to travel at a preset calibration speed along the track at a uniform speed, while simultaneously collecting visible light reference video sequences, three-dimensional laser point cloud data and IMU attitude data of the belt support. During the collection process, the position information of key feature points such as support connection nodes, welds, and bolts is marked.

[0027] (1) Definition of a unified coordinate system: World coordinate system O-XYZ: with the starting point of the corridor track end face as the origin, the X-axis along the direction of corridor extension, the Y-axis along the horizontal direction, and the Z-axis along the vertical direction; Image coordinate system o-uv: with the top left corner of the image as the origin, the u-axis is along the horizontal pixel direction, and the v-axis is along the vertical pixel direction; LiDAR coordinate system It is fixed to the radar body and converted to the world coordinate system through hand-eye calibration.

[0028] (2) Hand-eye calibration conversion relationship: ; in, Let be the homogeneous transformation matrix of the radar to the world coordinate system, including the rotation matrix. With translation vector The solution is obtained through the calibration plate calibration method.

[0029] S1.2 Construction of Three-Dimensional Benchmark Model

[0030] Based on the collected reference point cloud data, an improved Iterative Closest Point (ICP) algorithm is used to register and stitch the point cloud to generate a three-dimensional point cloud reference model of the entire belt conveyor support. At the same time, combined with the original engineering data such as the support's design drawings, specifications, and installation spacing, a parametric digital twin reference body of the support is constructed to map the geometric dimensions, connection relationships, and mechanical properties of each part of the support.

[0031] (1) Global transformation constraints for benchmark point cloud stitching: ; in, For the local point cloud of the i-th frame, , For the global rotation and translation transformation corresponding to the i-th frame, This is the stitched global baseline point cloud. This represents the total number of frames.

[0032] S1.3 Threshold and Coordinate System Calibration

[0033] Based on the support design specifications and operation and maintenance requirements, the allowable deformation threshold and vibration offset threshold for each part of the support are calibrated. All threshold parameters are used in subsequent testing and judgment processes to form a unified quantitative judgment standard. Allowable threshold for beam bending deflection: Allowable threshold for column tilt angle: Horizontal displacement allowable threshold: Bolt loosening displacement allowable threshold: Vibration offset amplitude threshold: Vibration frequency threshold: .

[0034] The first step in calibrating the multi-sensor spatial coordinate system was to: calibrate the camera's intrinsic parameters using a calibration board method to obtain the camera's intrinsic parameter matrix K; and calibrate the extrinsic parameter transformation matrices of the camera and lidar relative to the world coordinate system using a hand-eye calibration method. , A unified mapping relationship between the image pixel coordinate system, camera coordinate system, lidar coordinate system and world coordinate system is established to ensure that the spatial registration error of the multi-sensor system is ≤2mm, providing a calibration basis for subsequent 2D-3D spatial transformation.

[0035] S2 Real-time Image and Point Cloud Data Acquisition

[0036] The inspection robot runs along the track according to the preset inspection strategy. It acquires visible light image sequences, three-dimensional point cloud data and robot posture data of the belt support in the corridor in real time through a multi-sensor synchronous acquisition system. After the data is initially verified by the edge computing node, it is transmitted to the back-end processing terminal.

[0037] A hardware-triggered synchronization mechanism is employed to achieve microsecond-level synchronous data acquisition between industrial cameras, LiDAR, and IMU, minimizing synchronization time error. This ensures the spatiotemporal consistency of images and point cloud data; the dustproof lens cover of the inspection robot is equipped with an automatic blowing device to regularly remove dust from the lens surface and reduce the impact of dust obstruction; during the data acquisition process, the camera exposure parameters and supplementary light power are automatically adjusted according to the ambient light intensity of the corridor to adapt to the alternating light and dark conditions.

[0038] S3 Multimodal Image Adaptive Preprocessing

[0039] The acquired raw visible light images are subjected to three layers of adaptive processing: dust atomization removal, dynamic illumination equalization, and inter-frame vibration offset correction, in sequence to eliminate corridor environmental noise and vibration interference, and output high-definition, high-matching standardized detection images.

[0040] S3.1 Dust atomization removal

[0041] Based on the improved dark channel prior and atmospheric scattering model, the fogged image caused by dust occlusion is restored. In view of the characteristics of dust concentration fluctuation in the corridor, a concentration perception branch is introduced to automatically estimate the atmospheric light value and transmittance. Combined with the detail repair module of lightweight generative adversarial network (GAN), the key details such as the support edge and bolt node are preserved while removing the dust fogging effect.

[0042] Atmospheric scattering physics model: ; in: The original dust-containing image was captured in pixels. The grayscale value at that location; The grayscale value of the restored clear image; : pixel The transmittance of the medium at a certain point reflects the ability of dust to transmit light, and its value ranges from [0,1]. : Global atmospheric light value, representing the background light intensity of the dusty environment within the corridor.

[0043] Dark channel value calculation: ; in, For pixels A 15x15 neighborhood window, The image grayscale value for the c color channel.

[0044] Concentration-sensing transmittance correction formula: ; in, The local dust concentration is estimated and obtained by normalizing the brightness of the dark channel. This is the concentration correction factor, with a value ranging from 0.1 to 0.3, and 0.25 for high dust conditions in the corridor.

[0045] S3.2 Dynamic Illumination Equalization

[0046] An improved multi-scale Retinex algorithm is used to enhance images with uneven illumination. By decomposing the incident and reflected components of the image, local overexposure is suppressed and dark areas are brightened, while the edge contrast of the support structure is preserved. For scenes with alternating strong light and shadow in the corridor, illumination smoothing constraints are introduced to avoid local halo and color distortion problems after processing.

[0047] Single-scale Retinex output: ; in, Let i be a Gaussian wrapping function at scale i. For convolution operations, the Gaussian function expression is: ; Multi-scale Retinex (MSR) weighted fusion output: ; The Gaussian standard deviations of the three scales are as follows: Corresponding weights .

[0048] Lighting smoothing constraint: ; in, For the incident light component, The smoothing weighting coefficient is set to 0.01, which is used to suppress halos caused by sudden changes in illumination.

[0049] S3.3 Inter-frame vibration offset correction

[0050] By combining IMU attitude data with image ORB feature point matching results, the sparse optical flow method is used to calculate the global motion vector between frames to correct image offset and blur caused by device vibration. By matching corner features and geometric structure features on the support, pixel-level alignment between consecutive frames is achieved, eliminating inter-frame misalignment caused by vibration.

[0051] Image affine transformation model based on IMU pose: ; in, These are the pixel coordinates of the original image. Here are the pixel coordinates of the corrected image; H is the affine transformation matrix, and the pitch angle is acquired by the IMU. Roll angle Yaw angle The calculation yielded: ; in, , This represents the translation amount of the image plane, obtained from the linear displacement data of the IMU through intrinsic parameter conversion.

[0052] Fundamental constraint equations for sparse optical flow: ; in, , This represents the gradient of the image grayscale in the x and y directions. For time gradient, , The optical flow velocity is the speed at which a pixel is viewed.

[0053] S4 Precise Deformation Detection and Type Differentiation

[0054] The preprocessed image data and the real-time acquired 3D point cloud data are compared with the 3D benchmark model in multiple dimensions. The deformation is accurately obtained through quantization calculation formula, and the deformation type is distinguished by combining frequency characteristics.

[0055] S4.1 Segment Semantic Segmentation and Region Localization

[0056] A lightweight semantic segmentation network (improved YOLO-Seg) is used to extract the support region from the preprocessed image, identify key components such as the support column, beam, roller seat, and connecting bolt, segment the support structure region to be detected, generate a binary segmentation mask for each type of component, and filter out background interference such as belts, materials, and corridor walls. The component segmentation accuracy is ≥95%, and the inference speed is ≥30fps.

[0057] S4.2 2D-3D Spatial Mapping and Component-Level Point Cloud Association

[0058] Based on the previously calibrated camera intrinsic and extrinsic parameter matrices, the projection transformation relationship between 2D image pixels and the 3D world coordinate system is established. The 2D component mask regions obtained by semantic segmentation are mapped to the 3D point cloud space. The real-time point cloud subsets corresponding to each component are selected and associated one by one with the corresponding component regions in the 3D benchmark model, providing a component-level matching basis for subsequent accurate registration.

[0059] (1) 2D-3D projection transformation model

[0060] Based on the imaging principle of a pinhole camera, a forward projection relationship from 3D world coordinates to 2D image pixel coordinates is established: ; Where (u,v) are the pixel coordinates in the image pixel coordinate system; The depth value of the spatial point in the camera coordinate system is provided by the synchronously acquired lidar point cloud; This is the camera intrinsic parameter matrix, which includes internal parameters such as focal length and principal point coordinates. This is the homogeneous extrinsic parameter matrix from the world coordinate system to the camera coordinate system, which includes the rotation matrix and the translation vector; The coordinates of the point in space are the three-dimensional coordinates in the world coordinate system.

[0061] Correspondingly, by combining the depth information synchronously acquired by the lidar, the above formula can be used to achieve the reverse projection from 2D pixel coordinates to 3D world coordinates, mapping the pixel region in the image to a set of 3D spatial points.

[0062] (2) Component-level point cloud screening and benchmark association

[0063] For the binary mask regions of the four types of components—columns, beams, roller seats, and connecting bolts—obtained by S4.1 semantic segmentation, the component-level point clouds are associated with the baseline model one-to-one according to the following steps: ① Traverse all valid pixels within the mask of each type of component, combine them with the synchronously acquired 3D point cloud depth information, and map the pixels to the world coordinate system through the above-mentioned back projection transformation to obtain the 2D-3D matching point set corresponding to the component. ② Using the matching point set as the center, set a spatial distance threshold of 5mm, and filter out the real-time point cloud subset of the corresponding component from the real-time acquired global 3D point cloud. (k is the component category number), further removing background point cloud interference; ③ Reference areas of the same type of component pre-divided in the three-dimensional reference model A hierarchical mapping relationship is established between 2D segmentation mask, real-time component point cloud, and reference component region to achieve one-to-one matching at the component level.

[0064] S4.3 3D Point Cloud Registration and Difference Calculation

[0065] A two-stage registration strategy of coarse registration + fine registration is adopted. For each type of component, its real-time point cloud subset is registered with the corresponding component region in the 3D reference model: coarse registration is first completed by feature point matching, and then millimeter-level fine registration is completed by the improved weighted ICP algorithm. After registration, the 3D coordinate difference of each corresponding point is calculated to obtain quantitative parameters such as deformation offset, tilt angle, and displacement distance of each part of the support.

[0066] Weighted ICP registration objective function: ; in, For the i-th corresponding point in the current component's real-time point cloud subset, For the corresponding point in the component area of ​​the baseline model; For rotation matrix, It is a translation vector; As for point weights, the weight of key feature points (nodes, bolts) of the support is 1.5, and the weight of ordinary structural points is 1.0. Using the Euclidean distance operator, the optimal solution is obtained by iteratively minimizing the objective function. and .

[0067] Deformation parameter calculation: 3D displacement at any point: ; Beam bending deflection (maximum normal displacement at the midpoint of the beam): ;in, The normal unit vector of the reference beam; Column tilt angle: ;in, The normal vector of the column in the baseline model. The normal vector of the column is obtained by fitting the real-time point cloud.

[0068] S4.3 Intelligent Distinction of Deformation Types

[0069] By combining short-time Fourier transform (STFT) analysis with frequency characteristics of continuous frame deformation data, high-frequency, small-amplitude offsets are identified as instantaneous vibration offsets, while low-frequency, continuously increasing offsets are identified as permanent structural deformations. At the same time, based on the comparison between the deformation amount and a preset threshold, specific deformation types such as bending, tilting, displacement, and loosening are identified, and the deformation location is accurately located.

[0070] Short-time Fourier transform of deformed time-series signals: ; in, It is a deformation displacement time sequence signal for consecutive frames. This is the Hanning window function, with a window length of 32 frames.

[0071] Deformation type determination criteria: For different types of deformation, the results are compared with the corresponding allowable thresholds, and a determination is made based on frequency characteristics. If the deformation magnitude And the main frequency of the signal It is determined to be an instantaneous vibration offset and is not included in the structural deformation; If the deformation continues to increase, the signal frequency... Furthermore, the deformation exceeds the allowable threshold for the corresponding part (beam deflection exceeds [w], column tilt exceeds [w]). If the displacement exceeds [d] or the bolt displacement exceeds [δ], it is considered a permanent structural deformation. For frequency The corresponding amplitude, This is the main frequency of the signal.

[0072] S5 Time-Series Deformation Trend Prediction

[0073] Based on continuous multi-cycle inspection deformation data, a multi-stage attention long short-term memory network (MA-LSTM) prediction model is constructed to explore the evolution law of stent deformation and achieve advanced prediction of failure risk.

[0074] S5.1 Time Series Dataset Construction

[0075] Organize the deformation data of each part of the historical inspection support output in step S4, construct a sample dataset according to time series, and label the deformation development stage and fault label; use time series interpolation to expand the sample size for small sample early deformation data, and improve the model's ability to identify early defects.

[0076] Each sample input is a sequence of deformation features of multiple parts of the stent over T consecutive time steps. The feature dimension F corresponds one-to-one with the deformation variables output by S4, including six spatial features: column tilt, beam deflection, bolt displacement, node displacement, horizontal offset, and vertical settlement. The sample labels are the deformation variables and fault levels corresponding to future time points. The dataset is strictly divided into training, validation, and test sets in chronological order. Test set data is not used in model training to ensure the objectivity of the evaluation.

[0077] Training of S5.2 Multi-stage Attention Temporal Prediction Model

[0078] A multi-stage attention-based long short-term memory network (MA-LSTM) prediction model was constructed. The model employs a hierarchical architecture of spatial attention encoding, temporal feature extraction, spatiotemporal attention fusion, and residual output. Addressing the spatial variability and nonlinear temporal evolution of scaffold deformation data, a two-stage attention mechanism is used to allocate weights to key spatial components and enhance weights for key temporal periods. Residual connections are introduced to preserve original feature information, accurately uncovering the temporal evolution patterns of scaffold deformation. The model training process is as follows: (1) First-level spatial attention feature encoding The dataset contains spatial feature sequences of support deformation from multiple parts, including columns, beams, connecting bolts, and welds. A 1×1 convolutional transformation is used to generate a query vector Q, a key vector K, and a value vector V. Self-attention calculation is employed to adaptively assign weights to spatial features from different parts, strengthening the feature representation of key nodes and weak points in the support structure while suppressing interference from unstructured background features. The output is a spatial attention-weighted feature sequence. The specific calculation formula is as follows: ; ; ; ; in: The input is the spatial feature sequence of the support deformation. The time step length, For spatial feature dimensions; These are 1×1 convolutional mappings that generate queries, keys, and values, respectively. The dimension of the key vector is used to scale the dot product result to avoid gradient vanishing. The operation is performed along the feature dimension to ensure that the sum of the output weights is 1. This is the feature sequence of the first-level spatial attention weighted output.

[0079] (2) LSTM temporal feature extraction

[0080] The weighted feature sequence output from the first-level spatial attention is sequentially input into a two-layer LSTM network step by step to capture the long-term temporal dependence and short-term evolution trend of the scaffold deformation at each time step. Effective temporal information is filtered through a gating mechanism involving forget gates, input gates, unit states, and output gates to remove high-frequency noise caused by instantaneous vibration interference, retain the long-term evolution characteristics of structural deformation, and output the hidden state feature sequence at each time step. The LSTM unit calculation process at time t is as follows: ① Forget gate: Controls the degree to which the cell state information from the previous moment is retained, filtering out outdated vibration interference information. ; ② Input gate: Controls the degree of update of the input features at the current time step, and writes in the newly added structural deformation information: ; ; ③ Cell state update: Combine the outputs of the forget gate and the input gate to update the cell state at the current time step: ; ④ Output gate: Controls the output level of the hidden state at the current time step, and outputs the temporal feature vector: ; ; in: The spatial attention-weighted deformation feature input at time t; These are the hidden layer output and cell state at the previous time step, respectively; This is the weight matrix corresponding to the gating. For the corresponding bias vector; Here, is the Sigmoid activation function, and tanh is the hyperbolic tangent activation function; For Hadamard products (element-wise multiplication); Let t be the hidden state feature vector output by the LSTM at time t.

[0081] (3) Second-level spatiotemporal fusion attention mechanism

[0082] Based on the hidden state sequences at each time step of the LSTM output, this method integrates the global spatial features from the first-level spatial attention output, calculates the attention weights at different time steps, and focuses on key periods of accelerated deformation growth. This achieves deep fusion of spatial features and temporal evolution characteristics, addressing the issues of single-time-series attention being susceptible to the randomness of scaffold operation and insufficient hidden state quality. The specific calculation process is as follows: ① Temporal query vector generation: Based on the previous time-series hidden state output by the LSTM, a temporal query vector with the same dimension as the spatial features is generated. ; in: The time-series query vector at time t is consistent with the dimension of the single-time spatial features; Let H be the weight matrix and bias vector of the attention network, and H be the dimension of the LSTM hidden state.

[0083] ② Calculation of spatiotemporal attention weights: The attention weights at each time step are obtained by performing an inner product between the temporal query vector and the spatial attention feature vector at each time step, followed by scaling and normalization. ; ; in: Let i be the attention score at time i. =F is the spatial feature dimension, used to scale the dot product to avoid gradient vanishing; Let be the normalized attention weight vector from time 1 to time t.

[0084] ③ Weighted Feature Aggregation: Based on the attention weights, the spatial features at each time step are weighted and summed to obtain a context feature vector that integrates spatiotemporal information. ; in: The attention weights at time i are calculated at time t. This is the context feature vector that has been weighted and aggregated by spatiotemporal attention.

[0085] ④ Residual concatenation and output: The spatiotemporal attention aggregation features are concatenated with the original spatial attention features at the current time step using residual concatenation. The final temporal prediction feature output is obtained through linear mapping, which preserves the original feature information while enhancing the weights of key features. ; in: B is the output layer weight matrix, and B is the output layer bias vector. For feature splicing operations, This is the deformation prediction feature vector output by the model at time t.

[0086] (4) Model training and parameter optimization

[0087] The model is trained using a mini-batch stochastic gradient descent algorithm combined with the Adam optimizer, with mean squared error (MSE) as the loss function and minimizing the prediction error on the validation set as the convergence criterion. The network weights and bias parameters are updated layer by layer using backpropagation. An early stopping strategy is employed to prevent overfitting; training is terminated when the validation set loss no longer decreases after 20 consecutive epochs, and the optimal model parameters are saved. The loss function is calculated as follows: ; Where: m is the batch sample size. Let i be the actual deformability of the i-th sample. These are the deformation variables predicted by the model.

[0088] S5.3 Deformation Trend and Fault Prediction

[0089] The latest inspection data is input into the trained model to fit the evolution curve of stent deformation, predict the deformation in different future cycles, assess the probability and expected time of failures such as stent collapse and belt misalignment, and classify them into three levels: low risk, medium risk, and high risk.

[0090] Deformation evolution exponential growth model: ; in, Let be the deformation at time t. is the initial deformation, and k is the deformation growth rate coefficient, which is obtained by fitting historical data.

[0091] Failure probability calculation (Weibull distribution model): ; Where m is the shape parameter, taken as m=2; Risk level classification: Low risk: P < 10%; Medium risk: ; High risk: P>50%.

[0092] S6 Closed-Loop Early Warning and Review Inspection

[0093] Based on deformation detection results and trend prediction results, a closed-loop inspection mechanism of early warning, review, archiving and optimization is constructed to achieve a complete closed loop of intelligent operation and maintenance.

[0094] S6.1 Tiered Early Warning Trigger

[0095] For detected permanent deformations exceeding the threshold and predicted medium- to high-risk faults, a tiered audible and visual early warning system is triggered, along with notifications from the operation and maintenance platform. High-risk faults are simultaneously linked to the emergency stop protection logic of the belt conveyor system to prevent accidents from occurring.

[0096] S6.2 Adaptive Verification Inspection

[0097] The inspection robot's inspection path and frequency are dynamically adjusted based on the Deep Q-Network (DQN) reinforcement learning algorithm. The state space is based on the bracket risk level, deformation location, and inspection accuracy requirements, while the action space is based on the robot's movement speed, acquisition resolution, and inspection path points. The reward function is the weighted sum of detection accuracy and inspection efficiency. The robot automatically plans a secondary refined scanning path for high-risk areas, reduces the robot's movement speed, and increases the acquisition resolution. It performs multi-view, high-precision verification and inspection of deformed areas to verify the authenticity and severity of the deformation.

[0098] S6.3 Data Source Tracing and Model Iteration

[0099] All raw data, test results, verification records, and fault handling information from each inspection are stored in the stent deformation database to form a full lifecycle inspection log. At the same time, the time series prediction model and detection model are regularly iterated and optimized with newly added data to continuously improve the accuracy of detection and prediction.

[0100] In a specific embodiment, the deformation inspection of the belt conveyor supports in the main transport trestle corridor of a coal mine is taken as an example. The trestle corridor is 1200m long, 2.2m wide, and 2.5m high. The belt conveyor supports are arranged at 3m intervals. The supports are steel structures and mainly have deformation risks such as bending, tilting, and loose bolts. The corridor has high dust concentration, poor lighting conditions, and strong equipment vibration, making traditional manual inspection difficult.

[0101] This embodiment uses a track-mounted inspection robot as the data collection platform. The robot is equipped with a 2-megapixel high-definition industrial camera (30fps), a 16-line miniature LiDAR (range accuracy ±5mm), and a six-axis IMU (attitude accuracy ±0.1°). It also features an automatic dust-proof lens cover and an adaptive lighting module. The inspection track is laid along the top of the corridor, covering the entire conveyor belt support. The specific inspection implementation steps are as follows: Step 1: Preliminary Modeling and Parameter Calibration The inspection robot was controlled to travel at a constant speed of 0.5 m / s along the track, simultaneously collecting visible light reference video, 3D point cloud, and IMU attitude data for all 400 support frames along the entire section. During the data acquisition process, 12 key feature points, including the bottom of the support columns, beam nodes, and bolt connection points, were marked. The transformation matrix of the radar to the world coordinate system was obtained through hand-eye calibration. The spatial registration error after calibration is 1.8 mm.

[0102] After denoising and downsampling preprocessing of the collected reference point cloud data, an improved weighted ICP algorithm is used to register and stitch the point cloud segment by segment to generate a three-dimensional point cloud reference model of the entire belt conveyor support with a point cloud resolution of 5mm. At the same time, combined with the support design drawings (the columns are made of No. 10 channel steel, the crossbeams are made of No. 8 channel steel, and the installation height is 1.8m), a parametric digital twin reference body is constructed to map the geometric dimensions and connection relationships of each part of the support.

[0103] According to the "Maintenance Specifications for Belt Conveyor Equipment in Coal Mines", the allowable deformation thresholds for each part are specified: allowable bending deflection threshold for the crossbeam. The allowable threshold for the column tilt angle Horizontal displacement allowable threshold Bolt loosening displacement allowable threshold Simultaneously calibrate the vibration offset threshold: vibration amplitude threshold. Vibration frequency threshold .

[0104] Step 2: Real-time image and point cloud data acquisition

[0105] During routine inspections, the inspection robot patrols along its track at a speed of 1 m / s every 4 hours according to a preset cycle. A hardware-triggered synchronization mechanism enables microsecond-level synchronous data acquisition from the camera, LiDAR, and IMU, with a synchronization time error of ≤0.8 ms. This allows for real-time acquisition of visible light images, 3D point clouds, and attitude data of the support structure. During data acquisition, a light sensor automatically adjusts the camera's exposure time and supplemental lighting power to adapt to changes in lighting within the corridor. A lens cleaning device activates every 30 minutes to remove dust from the lens surface. The collected data is transmitted to edge computing nodes via the 5G network within the corridor, where it undergoes validity verification before being uploaded to the backend processing terminal.

[0106] Step 3: Adaptive preprocessing of multimodal images

[0107] The background process performs three layers of processing on the original image sequentially: Dust atomization removal: First, the dust concentration and atmospheric light value of the image are estimated by the concentration sensing branch. The image is initially restored based on the atmospheric scattering model. The transmittance correction coefficient α is set to 0.25. Then, a lightweight GAN network is used to repair the details of the restored image, supplement the texture details of the support edge and bolt node, and improve the image clarity.

[0108] Dynamic illumination equalization: A three-scale Retinex algorithm is used with Gaussian standard deviations of 15, 80, and 250, corresponding to weights of 0.5, 0.3, and 0.2, respectively. This decomposes the incident and reflected light components of the image, adaptively brightens dark areas, and suppresses overexposed areas. At the same time, an edge preservation constraint (smoothing weight λ=0.01) is introduced to avoid blurring of the support structure edges, resulting in improved image contrast after processing.

[0109] Inter-frame vibration offset correction: Combining the attitude vibration data output by the IMU, ORB feature point matching combined with sparse optical flow method is used to calculate the global motion vector between consecutive frames. The image is then translated and rotated using an affine transformation matrix to eliminate inter-frame misalignment caused by device vibration. After correction, the inter-frame registration error is ≤0.8 pixels.

[0110] Step 4: Precise Deformation Detection and Type Differentiation

[0111] A lightweight semantic segmentation network is used to segment the support components of the preprocessed image, identifying four key components: columns, beams, roller seats, and bolts, while filtering out background interference such as conveyor belts, raw coal, and corridor walls.

[0112] 2D-3D mapping is performed based on previously calibrated camera intrinsic and extrinsic parameters: the camera intrinsic parameter calibration result is the focal length. Pixels The pixel coordinates are (512, 384). After camera extrinsic calibration, the translation error is ≤1.5mm and the rotation error is ≤0.1°. The segmentation masks of the four types of parts are back-projected into 3D space, and a 5mm distance threshold is set to filter the part point cloud subset from the global point cloud. The part point cloud filtering accuracy reaches 96.5%, and it matches the corresponding part regions of the benchmark model one by one.

[0113] Based on the segmentation results, the point cloud region corresponding to the support is located. First, coarse registration between the real-time point cloud and the reference point cloud is achieved through feature point matching. Then, a weighted ICP algorithm is used for fine registration, with key feature points weighted at 1.5 and ordinary points at 1.0, achieving a registration accuracy of 0.5mm. After registration, the displacement, beam deflection, and column tilt angle at each point are calculated using deformation reduction formulas. For example, the calculated displacement at the midpoint of a support beam is 8.2mm, and the normal deflection... It is within the permissible range.

[0114] A short-time Fourier transform was performed on 10 consecutive frames of deformation data, with a window length of 32 frames, to analyze its frequency characteristics: if the offset amplitude is ≤3mm and the dominant frequency is ≥5Hz, it is determined to be an instantaneous vibration offset; if the offset amplitude continues to increase and the dominant frequency is <1Hz, it is determined to be a permanent structural deformation. In this embodiment, a horizontal displacement of 9.2mm was detected at the bottom of a column, with a dominant frequency of 0.7Hz, which was determined to be a permanent displacement deformation, exceeding the allowable threshold.

[0115] Step 5: Predicting the Temporal Deformation Trend

[0116] We compiled the deformation data from the past 6 months of inspections and constructed a time series sample set for each support and each part. The feature dimensions included 6 spatial features such as column tilt, beam deflection, and bolt displacement. The time step was set to 24 steps. For early small deformation samples, we used temporal interpolation to expand the sample size. A total of 12,000 training samples were constructed and divided into training set, validation set and test set in an 8:1:1 ratio.

[0117] The MA-LSTM multi-stage attention prediction model was trained as follows: The first-stage spatial attention used a single 1×1 convolutional layer to generate Q / K / V values; the LSTM part used a two-layer network with 64 hidden nodes; the second-stage spatiotemporal fusion attention fused spatial features and temporal hidden states, and the output adopted a residual concatenation structure. The Adam optimizer was used for training, with an initial learning rate of 0.002, a batch size of 64, and a maximum training epoch of 200. An early stopping strategy was adopted, terminating training if the validation set loss did not decrease for 20 consecutive epochs. After training, the average error in predicting variables for the next 7 days was ≤7.8%.

[0118] The latest inspection data is input into the model, and an exponential growth model is used to fit the deformation evolution curve. The probability of failure is calculated using the Weibull distribution. For example, for the aforementioned displacement deformation support, the fitting yields a deformation growth rate k = 0.012 / day, predicting that the deformation will reach 12.5mm after 15 days, with a failure probability P = 61%, which is classified as high risk.

[0119] To verify the performance of the proposed model, comparative experiments were conducted with general LSTM and CNN-LSTM models on the same dataset. The mean absolute percentage error (MAPE) was used as the evaluation metric, and the results are as follows: The general LSTM model has an average error of 16.2% in predicting deformation over the next 7 days. The CNN-LSTM model has an average error of 12.7% in predicting deformation over the next 7 days. The model of this invention has an average deformation prediction error of 7.8% over the next 7 days, which is 51.9% higher than the general LSTM and 38.6% higher than the CNN-LSTM. The prediction error over the next 30 days is more than 42% higher than the general LSTM, which verifies the effectiveness of the multi-stage attention architecture.

[0120] Tests showed that after three layers of preprocessing, the image retained 92.3% of the support edge features, which is 85.7% higher than the 49.7% of the traditional single filtering algorithm.

[0121] Step 6: Closed-loop early warning and review inspection

[0122] Tiered early warning: For detected permanent deformation exceeding the threshold, a first-level audible and visual early warning is immediately triggered, and the fault location, deformation type, and deformation amount are pushed to the maintenance personnel; for supports that are judged to be high-risk, a second-level early warning is triggered, and the maintenance personnel are advised to conduct on-site verification as soon as possible, and the belt conveyor system is simultaneously reduced in speed.

[0123] Adaptive review: For high-risk areas, the review and inspection path is automatically planned based on deep reinforcement learning algorithm, the inspection robot is controlled to reduce its speed to 0.3m / s, the acquisition resolution of camera and lidar is improved, and the faulty support is scanned from multiple perspectives to verify the authenticity and severity of the deformation. The review results are updated to the operation and maintenance platform in a synchronous manner.

[0124] Data archiving and model iteration: All raw data, test results, review records and handling information of each inspection are stored in full to form a full life cycle inspection log for the stent; every half month, the detection model and prediction model are fine-tuned and iterated using newly added inspection data to continuously improve model accuracy.

[0125] The results show that the method has a 98.7% accuracy rate in detecting support deformation in the coal mine trestle corridor, can identify a minimum deformation of 0.8 mm, has a false detection rate of only 1.2%, and can successfully predict the failure of a support beam bending aggravation 21 days in advance, thus saving time for equipment maintenance. The inspection efficiency is 22 times higher than that of manual inspection, and significant application results have been achieved.

[0126] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0127] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent inspection of deformation of belt conveyor supports in a trestle bridge corridor, characterized in that, include: S1. Collect fault-free reference data of the entire belt support, and construct a three-dimensional reference model and digital twin reference body in combination with the original design parameters of the support, and calibrate the deformation threshold and vibration offset threshold of each part. S2. The inspection robot runs along the preset path and simultaneously collects visible light image sequences, three-dimensional point cloud data and attitude data of the belt support. S3. Preprocess the original image and output a standardized detection image; S4. Register and differentiate the three-dimensional reference model with the standardized detection image and real-time three-dimensional point cloud data respectively, quantify and calculate the deformation of each part of the support, and distinguish between instantaneous vibration displacement and permanent structural deformation by combining frequency characteristics, and identify the deformation type and location. S5. Based on the temporal deformation data of multi-cycle inspection, the probability and risk level of future failures are predicted by a multi-stage attention long short-term memory network model. S6. For deformation exceeding the threshold and high-risk faults triggering graded early warnings, the inspection path is adaptively adjusted to perform secondary verification scans, the inspection data is fully stored and the model is iteratively optimized to perform closed-loop intelligent inspection.

2. The intelligent inspection method for deformation of belt conveyor supports in a trestle corridor according to claim 1, characterized in that, Step S1 includes: S1.1 controls the inspection robot to travel at a constant speed along the track at a calibrated speed, and simultaneously collects visible light reference video sequences, three-dimensional laser point cloud data and IMU attitude data of the belt support, and marks the key feature points of the support; S1.2 Based on the reference point cloud data, an improved iterative nearest point algorithm is used to register and stitch the point cloud to generate a three-dimensional point cloud reference model of the entire belt conveyor support; combined with the support design drawings and specifications, a parametric digital twin reference body is constructed to map the geometric dimensions and connection relationships of the support. S1.3 calibrate the allowable deformation thresholds for each part of the support, including bending deflection threshold, tilt angle threshold, horizontal displacement threshold, and bolt loosening displacement threshold; calibrate the vibration offset threshold to distinguish between normal vibration and structural deformation boundaries; and complete the registration of the multi-sensor spatial coordinate system.

3. The intelligent inspection method for deformation of belt conveyor supports in a trestle corridor according to claim 1, characterized in that, In step S2, a hardware-triggered synchronization mechanism is used to achieve microsecond-level synchronous acquisition of industrial cameras, lidar, and IMU.

4. The intelligent inspection method for deformation of belt conveyor supports in a trestle corridor according to claim 1, characterized in that, Step S3 includes: S3.1 is based on an atmospheric scattering physics model. It automatically estimates dust concentration and atmospheric light value by combining a concentration sensing branch. It solves and corrects the transmittance through the dark channel prior. It combines a lightweight generative adversarial network to restore image details, remove dust fogging effect and retain key features of the support. S3.2 uses the multi-scale Retinex algorithm to decompose the incident and reflected components of the image, adaptively brightens dark areas and suppresses overexposed areas, and introduces an illumination smoothing constraint term to optimize the edge blur and halo of the support structure. S3.3 combines IMU attitude data to construct an affine transformation matrix, and combines ORB feature point matching and sparse optical flow method to calculate the global motion vector between frames, and performs pixel-level correction for image offset and misalignment caused by vibration.

5. The intelligent inspection method for deformation of belt conveyor supports in a trestle corridor according to claim 1, characterized in that, Step S4 includes: S4.1 uses a lightweight semantic segmentation network to segment the support components of the preprocessed image, identify columns, beams, roller seats, and connecting bolts, and generate binary segmentation masks for each component to filter background interference. S4.2 Based on the previously calibrated camera intrinsic and extrinsic parameter matrices, establish the projection transformation relationship between 2D image pixels and the 3D world coordinate system, reverse project the binary segmentation mask of each component to the 3D space, filter out the point cloud subset of the corresponding component from the real-time 3D point cloud, and associate it with the corresponding component region in the 3D benchmark model. S4.3 adopts a two-level strategy of coarse registration and fine registration. For each component, its real-time point cloud subset is registered with the corresponding component region of the 3D benchmark model. First, coarse registration is completed by feature point matching, and then millimeter-level fine registration is completed by weighted iterative nearest point algorithm. The displacement of each point, beam deflection and column tilt angle are calculated by deformation transformation formula. S4.4 uses short-time Fourier transform to analyze the time-frequency characteristics of continuous frame deformation data, identifies high-frequency small-amplitude offsets as instantaneous vibration offsets, identifies low-frequency continuously increasing offsets as permanent structural deformations, and identifies deformation types such as bending, tilting, displacement, and loosening.

6. The intelligent inspection method for deformation of belt conveyor supports in a trestle corridor according to claim 1, characterized in that, Step S5 includes: S5.1 Organize historical inspection deformation data, construct a sample dataset according to time series, and expand the sample size by using time series interpolation enhancement for early small deformation samples. S5.2 Constructs a multi-stage attention long short-term memory network model. First, the first-level spatial attention mechanism is used to adaptively assign weights to the deformation features of multiple parts of the scaffold. Then, the LSTM network is used to extract the temporal evolution features. Subsequently, the second-level spatiotemporal fusion attention mechanism is used to achieve deep fusion of spatial and temporal features. Finally, the predicted features are output through residual splicing. S5.3 Input the latest inspection data, use the exponential growth model to fit the deformation evolution curve, calculate the probability of failure based on the Weibull distribution, and classify it into three risk levels: low, medium, and high.

7. The intelligent inspection method for deformation of belt conveyor supports in a trestle corridor according to claim 6, characterized in that, In step S5.2, the first-level spatial attention uses 1×1 convolution to generate query, key, and value vectors, and realizes spatial feature weighting through self-attention calculation; the second-level spatiotemporal fusion attention is based on the LSTM hidden state, fuses the features of the first-level spatial attention to calculate temporal weights, and weighted aggregation is used to obtain spatiotemporal features. The output end adopts a structure that splices spatiotemporal aggregated features with the residuals of the original spatial features.

8. The intelligent inspection method for deformation of belt conveyor supports in a trestle corridor according to claim 1, characterized in that, Step S6 includes: S6.1 triggers a first-level audible and visual warning and maintenance push for deformation exceeding the threshold, and triggers a second-level warning for high-risk predicted faults; S6.2 Based on deep reinforcement learning algorithm, dynamically adjust the inspection robot path and inspection frequency, and plan secondary scanning path for high-risk areas; S6.3 stores all inspection data throughout the entire process, forming an inspection log for the entire lifecycle of the stent, and regularly uses newly added data to iterate and optimize the detection and prediction models.

9. An intelligent inspection system for deformation of belt conveyor supports in a trestle bridge corridor, characterized in that, Includes: Calibration module: used to collect fault-free reference data of the entire belt support, combine the original design parameters of the support to construct a three-dimensional reference model and a digital twin reference body, and calibrate the deformation threshold and vibration offset threshold of each part; Acquisition module: Used for the inspection robot to run along a preset path and simultaneously acquire visible light image sequences, 3D point cloud data and attitude data of the belt support; Preprocessing module: Used to preprocess the original image and output a standardized detection image; Deformation recognition module: It is used to register and differentiate the three-dimensional reference model with the standardized detection image and real-time three-dimensional point cloud data, quantify and calculate the deformation of each part of the support, and distinguish between instantaneous vibration displacement and permanent structural deformation by combining frequency characteristics, and identify the deformation type and location. Prediction module: Based on time-series deformation data from multi-cycle inspections, it uses a multi-stage attention long short-term memory network model to predict the probability and risk level of future faults. Early warning and verification module: It is used to trigger graded early warnings for deformations exceeding the threshold and high-risk faults, adaptively adjust the inspection path to perform secondary verification scans, store all inspection data and iteratively optimize the model to perform closed-loop intelligent inspection.