Dynamic visual flaw detection and online maintenance linkage method for mine conveying belt under complex working conditions

Abstract

This invention relates to the field of conveyor belt inspection technology, and particularly to a method for the linkage of dynamic visual flaw detection and online maintenance of mining conveyor belts under complex working conditions. The method includes associating and matching preliminary surface defect data with registered internal wire rope weak magnetic defect data to establish a mapping relationship between surface and internal defects, thus completing preliminary risk classification. This invention employs a surface defect identification scheme that fuses visible light texture features with laser depth three-dimensional features, simultaneously registering internal defect data from weak magnetic detection and establishing a mapping relationship between internal and external defects. Furthermore, it introduces a multi-cycle time-series cumulative verification mechanism based on the cyclic operation characteristics of the conveyor belt. This scheme enables integrated risk assessment of surface and internal defects, identifies high-risk defects in an active expansion state, and effectively filters out misjudgments caused by occasional environmental interference, improving the accuracy of defect identification and the reliability of system operation.

Description

Technical Field

[0001] This invention relates to the field of conveyor belt inspection technology, and in particular to a method for linking dynamic visual flaw detection and online maintenance of mining conveyor belts under complex working conditions. Background Technology

[0002] Mining belt conveyors are core equipment for underground coal transportation. The conveyor belt, as its load-bearing and traction component, operates under heavy loads and continuous conditions for extended periods, making it prone to surface wear, cracks, and internal defects such as broken wires and corrosion in the steel wire rope core. If these defects are not detected and addressed promptly, they can easily lead to safety accidents such as belt breakage and material spillage, seriously threatening mine production safety and the lives of personnel. Therefore, online non-destructive testing technology for conveyor belts is an important component of the mine safety assurance system.

[0003] Currently, conveyor belt flaw detection technologies mainly include visual inspection and weak magnetic field testing. Among them, machine vision inspection can efficiently identify surface defects of the belt, while weak magnetic field testing can detect hidden defects in the internal steel wire ropes. Both have been applied to some extent in engineering. However, under complex working conditions such as high dust, low illumination, large belt speed fluctuations, and frequent load changes underground, existing technologies still have many shortcomings: First, fixed image acquisition parameters and simple preprocessing methods make it difficult to adapt to image degradation caused by working condition fluctuations, resulting in a high rate of false detections and missed detections. Second, the fusion of multimodal detection data is shallow, only simple data stitching is achieved, the risk correlation between surface and internal defects is not established, and there is a lack of effective interference filtering mechanisms, making it susceptible to misjudgments due to occasional coal slag and dust interference. Third, the impact of conveyor belt deflection and tensile deformation on defect size detection under heavy load conditions is not considered, and there are systematic deviations in defect geometric parameters, which directly affect the accuracy of risk classification. Fourth, the use of a single-level detection mode with full precision and full area results in low computing power utilization, making it difficult to adapt to the detection needs of high-speed conveyor belts under limited edge computing power conditions underground. Fifth, the flaw detection system is disconnected from the maintenance process, only possessing basic alarm functions, and failing to form a closed-loop operation and maintenance system of detection, positioning, maintenance, and re-inspection, making it difficult to meet production needs in terms of operation and maintenance efficiency and hazard response speed.

[0004] Overall, existing flaw detection technologies for mining conveyor belts still have significant shortcomings in terms of adaptability to working conditions, reliability of detection, accuracy, computing power adaptability, and operation and maintenance coordination, and cannot fully meet the safety operation and maintenance needs under complex underground working conditions. Summary of the Invention

[0005] To overcome the problems mentioned in the background art, the present invention proposes a method for linking dynamic visual flaw detection and online maintenance of mining conveyor belts under complex working conditions.

[0006] The technical solution of this invention is: a method for linking dynamic visual flaw detection and online maintenance of mining conveyor belts under complex working conditions, comprising the following steps: S1: Simultaneously acquire visible light images of the conveyor belt surface, laser 3D contour data, and weak magnetic detection data of the internal steel wire rope. At the same time, acquire real-time belt speed, environmental dust concentration, and illuminance parameters. Based on the belt speed and environmental parameters, perform dynamic acquisition and control with belt speed-environment dual feedback to ensure stable data quality. S2: Perform three-level anti-interference preprocessing on the acquired surface image, including occlusion removal, dust defogging, and motion blur elimination, to output a high-quality surface image; S3: Based on laser 3D contour data and drive motor load data, calculate the transverse deflection and longitudinal elongation of the belt surface under the current load, and establish a load-belt deformation correction model. S4: Adopting a two-stage detection architecture of initial screening and fine inspection, it performs defect detection on the pre-processed surface image and three-dimensional contour data, and outputs the location, type and geometric parameters of preliminary surface defects; S5: Correlate and match the preliminary surface defect data with the registered internal wire rope weak magnetic defect data to establish a mapping relationship between surface and internal defects and complete the preliminary risk classification; S6: Call the load-strip deformation correction model to dynamically compensate and correct the geometric parameters of the initial defect in length, width and depth, and output the true defect size that is independent of the load; S7: Based on the cyclic operation characteristics of the conveyor belt, perform multi-operation cycle time-series cumulative verification, and only determine the features that are stably detected in multiple consecutive operating cycles as valid defects, and generate the final defect detection result and risk level. S8: Based on the risk level of the final defect, link the corresponding level of online maintenance process to complete the closed-loop operation and maintenance of defect location, maintenance execution and re-inspection verification.

[0007] Preferably, the dynamic acquisition and control of the belt speed-environment dual feedback in step S1 specifically includes: The conveyor belt speed is obtained in real time by a rotary encoder, and the line frequency of the linear scan camera and the trigger frequency of the laser fill light are dynamically matched according to the target longitudinal sampling resolution to maintain a constant longitudinal sampling resolution within the belt speed fluctuation range of 0.5-6m / s. By collecting environmental parameters in real time through dust and light sensors, the laser supplementary light power is adaptively adjusted according to the dust concentration, and the camera exposure parameters are adaptively adjusted according to the ambient light intensity to compensate for the light intensity attenuation and brightness fluctuation caused by environmental interference.

[0008] Preferably, the three-level anti-interference preprocessing in step S2 specifically includes: The grayscale threshold and laser depth features are combined to identify areas covered by coal slag and gangue, which are marked as invalid detection areas and removed, thus completing the occlusion removal process. By combining real-time dust concentration parameters, an improved dark channel prior algorithm with adaptive intensity is used to eliminate the dust fogging effect in the image, improve image contrast, and complete the dust defogging process. Based on real-time belt speed data, a motion blur kernel is estimated, and a non-blind deconvolution algorithm is used to eliminate image motion blur caused by belt speed fluctuations, thus completing the motion blur elimination process.

[0009] Preferably, step S3 specifically includes: The cross-sectional profile of the conveyor belt is scanned in real time by a transversely arranged line structured light laser, the belt surface deflection curve is fitted, and the maximum transverse deflection and belt surface curvature parameters of the current belt surface are calculated. Collect the operating current and torque data of the drive motor, combine the conveyor belt tension mechanical model to calculate the real-time load and longitudinal tension, and calculate the longitudinal tensile rate based on the elastic modulus of the conveyor belt. Based on the transverse deflection parameter and the longitudinal tensile rate, compensation formulas for the transverse dimension, longitudinal dimension and depth dimension are constructed respectively, forming a load-strip deformation correction model.

[0010] Preferably, the two-stage detection architecture of initial screening and fine detection in step S4 specifically includes: First-level rapid initial screening: A lightweight detection model is used to quickly scan the low-resolution full-area image after downsampling, and output the coordinates and preliminary category of the suspected defect area. Normal and defect-free areas are skipped for further processing. Level 2 fine detection: This only targets the suspected defect areas output from the initial screening. It crops the corresponding high-resolution image and 3D contour data, inputs them into a high-precision segmentation and detection model, and completes the accurate classification, contour segmentation, and preliminary geometric parameter quantification of defects.

[0011] Preferably, step S4 also includes a speed-adaptive dynamic computing power scheduling step: The conveyor belt speed is acquired in real time. When the belt speed increases, the initial screening input resolution is reduced and the inference frame rate is increased to prioritize the detection coverage of the entire belt surface. When the belt speed decreases, the input resolution of the initial screening and fine inspection is increased to optimize the detection accuracy and make full use of idle computing resources.

[0012] Preferably, step S5 specifically includes: Based on encoder location stamps and a unified time reference, spatiotemporal dual registration is performed on three types of data: visible light, laser 3D, and weak magnetic field detection, to ensure accurate matching of multi-source data at the same physical location. Surface texture defect features, three-dimensional geometric defect features, and internal wire rope defect features were extracted separately. Establish a mapping relationship between internal and external defect locations. When the location of an internal defect is accompanied by surface degradation, the risk level of the defect is automatically increased. Defects with only internal defects and no surface characteristics are identified as early-stage hidden defects. Defects with only minor surface defects and no internal abnormalities are identified as ordinary surface damage.

[0013] Preferably, step S7 specifically includes: Using the entire circumference of the conveyor belt as a unit, precise matching of the circumferential position of defects is achieved based on encoder position data for each operating cycle; Record defect feature data for each operating cycle, and accumulate defect features at the same location across cycles; Only defects that are consistently detected over three or more consecutive operating cycles are considered valid defects, while occasional defects are classified as environmental interference and filtered out.

[0014] Preferably, the online maintenance process linked to the risk level described in step S8 specifically includes: The effective defects are divided into three risk levels. Among them, the first-level serious defects immediately trigger light and sound warnings and shutdown interlock signals, push emergency maintenance work orders, and force shutdown for maintenance. Level 2 critical defects generate high-priority repair work orders, and schedule planned downtime for repairs in the near future. Level 3 general defects are included in the routine maintenance log and handled uniformly during periodic maintenance.

[0015] Preferably, the closed-loop operation and maintenance described in step S8 specifically includes: Based on encoder position data, the defect is precisely located in the circumference. When the defect reaches the maintenance station, the conveyor belt is automatically controlled to decelerate or stop at a fixed point to guide the maintenance operation. After the maintenance is completed, the original defect location will be re-inspected in subsequent operation cycles to confirm the repair effect and update the defect status. Establish a defect lifecycle ledger to record data on the entire process of defect detection, development, repair, and re-inspection, and carry out preventive maintenance by combining trend prediction models.

[0016] The beneficial effects of this invention are: 1. Compared to existing technologies that use fixed-parameter image acquisition schemes paired with a single image post-processing algorithm, these methods struggle to adapt to the complex working conditions of high dust levels, low illumination, and significant belt speed fluctuations in underground environments. This leads to various quality degradation issues such as image stretching, compression, motion blur, and fogging / occlusion, directly resulting in high false positive and false negative rates and poor system adaptability. This invention employs a dynamic acquisition and control mechanism with dual feedback from belt speed and environment, combined with a three-level preprocessing algorithm for occlusion removal, dust removal, fogging, and motion blur elimination. This forms a complete end-to-end anti-interference system, from dynamic adjustment of hardware acquisition parameters to interference elimination at the algorithm level. This solution can synchronously optimize the parameters of the acquisition and processing ends based on real-time operating status and environmental conditions, ensuring stable output of high-quality belt area images even under extremely complex conditions. It significantly reduces the negative impact of working condition fluctuations on detection results, and the system's environmental adaptability and data output stability far exceed conventional visual inspection schemes.

[0017] 2. Compared to existing technologies that often employ single visual inspection methods or simply stitch together multi-source inspection data, these technologies cannot identify hidden defects in the steel wire rope core inside the conveyor belt, nor can they distinguish between surface stains and slag and actual defects. Furthermore, they lack time-series verification of inspection results, making them susceptible to false positives due to occasional dust and debris interference, and they cannot determine the active expansion state of defects. This invention employs a surface defect identification scheme that fuses visible light texture features with laser depth 3D features. It simultaneously registers internal defect data from weak magnetic detection and establishes a mapping relationship between internal and external defects. It also leverages the cyclic operation characteristics of the conveyor belt to introduce a multi-cycle time-series cumulative verification mechanism. This scheme enables integrated risk assessment of surface and internal defects, accurately identifies high-risk defects in an active expansion state, effectively filters out false positives caused by occasional environmental interference, and significantly improves the accuracy of defect identification and the reliability of system operation.

[0018] 3. Compared to existing technologies that only correct image distortion caused by belt speed fluctuations and do not consider the effects of lateral deflection and longitudinal tensile deformation of the conveyor belt under heavy loads, resulting in systematic deviations in defect size and depth detection, and distortion of defect geometric parameters under heavy load conditions, this invention uses laser structured light to acquire the cross-sectional contour of the belt surface in real time. Combined with real-time load data calculated from drive motor parameters, a correction model coupling load and belt deformation is established to dynamically compensate and correct the detected defect geometric parameters. This solution effectively eliminates the problem of detection scale distortion caused by heavy load deformation, outputs the true defect geometric dimensions independent of the load across the entire load range, significantly improves the accuracy of risk classification, and avoids safety accidents caused by underestimation of defects under heavy load conditions.

[0019] 4. Compared to existing technologies that employ a single-level real-time detection mode with full bandwidth and full precision, where detection accuracy and resolution remain constant, insufficient frame rates at higher belt speeds can lead to missed detections, while idle computing resources are wasted at lower belt speeds. This makes it difficult to adapt to the limited computing power of downhole edge computing devices, resulting in high deployment costs for high-speed conveyor belts. This invention adopts a two-level detection architecture of initial screening followed by fine inspection. First, a lightweight model quickly scans the entire belt surface to screen for suspected areas. Then, high-resolution fine inspection is performed on these suspected areas. Simultaneously, the detection resolution and model inference priority are dynamically adjusted based on the real-time belt speed. This solution can significantly improve the detection and inference speed at the edge, adapting to the detection needs of higher belt speed conveyor belts under limited computing power conditions. It achieves dynamic optimal allocation of computing resources, effectively reducing the hardware cost and computing power threshold for downhole deployment.

[0020] 5. Compared to existing technologies that only focus on defect detection and alarms, the detection results are disconnected from on-site maintenance procedures. There is a lack of precise defect location methods and standardized closed-loop maintenance processes, resulting in slow response times, low operational efficiency, and difficulty in achieving full lifecycle tracking and preventative maintenance of defects. This invention employs a tiered defect response mechanism to match corresponding online maintenance procedures, combining encoder position data to achieve precise defect location and targeted shutdown. Simultaneously, it constructs a repair-repair and full lifecycle management mechanism for defects. This solution streamlines the entire process chain from defect detection to operational maintenance, enabling the matching of optimal maintenance strategies based on defect risk levels. This significantly shortens defect response time, improves maintenance efficiency, and promotes the upgrade of conveyor belt maintenance from passive fault repair to proactive predictive maintenance. Attached Figure Description

[0021] Figure 1 The flowchart shown is a method for linking dynamic visual flaw detection and online maintenance of mining conveyor belts under complex working conditions according to the present invention. Figure 2 The diagram shown is a flowchart of the closed-loop control of dynamic acquisition and regulation of belt speed and environment dual feedback in the linkage method of dynamic visual flaw detection and online maintenance of mining conveyor belt under complex working conditions of the present invention. Figure 3 The diagram illustrates the load-belt deformation correction and defect geometry compensation principle in the dynamic visual flaw detection and online maintenance linkage method for mining conveyor belts under complex working conditions according to the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] This embodiment is applied to a general-purpose steel wire rope core belt conveyor in underground coal mines. The rated width of the conveyor belt is 1200mm, the rated belt speed is 4m / s, and the actual operating belt speed fluctuates within a range of 0.5-6m / s. The detection station is deployed at the middle frame of the conveyor and is equipped with a mining explosion-proof and intrinsically safe edge computing host. All detection and control logic is executed locally at the underground edge end.

[0024] The hardware system relied upon in this embodiment includes: one linear array industrial camera, two line structured light lasers, one set of weak magnetic field detection sensors, one incremental rotary encoder, one dust concentration sensor, one illuminance sensor, one edge computing host, one set of audible and visual early warning device, and a maintenance linkage control module. The rotary encoder is coaxially mounted on the driven roller shaft of the conveyor to collect real-time belt speed and position signals; the linear array camera and laser are installed inside a dustproof protective cover above the belt surface, facing the belt surface; the weak magnetic field detection sensor set is arranged laterally below the belt surface to collect the magnetic flux signal of the internal steel wire rope; the dust and illuminance sensors are installed next to the detection station; all devices are connected to the edge computing host, which uniformly executes data processing and control logic.

[0025] The complete implementation process of the present invention is described in detail below: S1: Simultaneously acquire multi-source detection data and execute dynamic acquisition and control based on both belt speed and environment feedback. This step is the initial stage of the detection process. It simultaneously acquires operating parameters, surface images, 3D contours, and internal magnetic signals. A dual feedback mechanism dynamically adjusts the acquisition parameters to ensure stable input data quality from the hardware perspective. Specifically, it includes: S11: Synchronous Acquisition of Multi-Source Data After the system is powered on, the pulse signal from the rotary encoder is used as a unified time reference to synchronously trigger data acquisition from the line scan camera, line structured light laser, and weak magnetic field detection sensor group. Simultaneously, real-time environmental data from the dust concentration sensor and illuminance sensor are read. All acquired data is tagged with a unified timestamp and encoder position label, providing a basis for subsequent registration.

[0026] S12: Acquisition synchronization control with speed feedback The encoder calculates the conveyor belt speed in real time. According to the preset target vertical sampling resolution The line frequency of the linear scan camera is dynamically matched with the laser trigger frequency. The formula for calculating the line frequency is: ; In this embodiment, the target longitudinal sampling resolution is set to 0.5 mm / line, meaning two lines of images are acquired per millimeter of belt length. When the belt speed is 0.5 m / s, the camera's line frequency is adjusted to 1000 Hz; when the belt speed is 6 m / s, the line frequency is synchronously increased to 12000 Hz. The line scan camera supports a line frequency adjustment range of 0-20 kHz, which can fully cover the belt speed fluctuation range of 0.5-6 m / s, eliminating image stretching and compression distortion caused by belt speed changes at the source of acquisition. The laser fill light trigger frequency is hard synchronized with the camera's line frequency at a 1:1 ratio, ensuring that each line of image corresponds to one laser fill light, avoiding uneven brightness between lines.

[0027] S13: Adaptive adjustment of light parameters based on environmental feedback Based on real-time collected dust concentration and illuminance data, the laser output power and camera exposure parameters are dynamically adjusted: Dust concentration adjustment: Dust concentration measurement range 0-1000mg / m³ 3 When the concentration is below 50 mg / m³ 3 At this time, the laser maintains 30% of its rated power; the concentration is 50-500 mg / m³. 3 Within the specified range, the power increases linearly with concentration up to 80% of the rated power; above 500 mg / m³... 3 At the same time, the power is maintained at 80% of the rated power to avoid overheating of the equipment due to excessive power and to compensate for light attenuation caused by dust; Illuminance adjustment: Illuminance measurement range is 0-1000 lux, with the goal of stabilizing the average gray value of the image in the range of 120-180. When the ambient light decreases, the camera exposure time and gain are linearly increased; when the light increases, the exposure parameters are reduced simultaneously to avoid overexposure and underexposure.

[0028] S2: Perform three-level anti-interference preprocessing on the area image. Based on hardware optimization, this step further eliminates operational interference through a three-level algorithm, outputting high-quality area images to provide reliable input for subsequent defect detection.

[0029] S21: Occlusion Removal Processing The method integrates grayscale thresholding and laser depth features to identify occluded areas: First, a reference surface is fitted to the laser 3D contour data to obtain a standard surface depth reference when there is no occlusion. For dark areas in the image with grayscale values ​​below 30, and corresponding raised areas with a depth data difference greater than 5mm from the reference surface, these are identified as occluders such as cinders and gangue. Pixels in these areas are marked as invalid detection areas and directly excluded in subsequent defect extraction. This avoids misjudging the edges of occluders as surface defects and eliminates the interference of occluded areas on size quantization.

[0030] S22: Adaptive Dust Demisting Process An improved dark channel prior algorithm is adopted, incorporating real-time dust concentration as a defogging intensity adjustment factor. Based on the standard dark channel prior's atmospheric light value and transmittance calculation logic, a defogging intensity coefficient is set. , The value is positively correlated with dust concentration: concentration below 50 mg / m³ 3 hour Perform mild demisting; concentration between 50-500 mg / m³ 3 hour Linear variation between 0.3 and 0.8; concentrations above 500 mg / m³ 3 hour It performs strong dehazing. By adaptively adjusting the intensity, it eliminates image fogging and contrast reduction caused by dust, while avoiding excessive dehazing that amplifies noise and loses defect details.

[0031] S23: Motion blur removal processing Based on real-time belt speed With camera exposure time Calculate the length of the motion fuzzy kernel The blur direction is the longitudinal direction of the conveyor belt movement, generating a one-dimensional uniform motion blur kernel. The Richardson-Lucy non-blind deconvolution algorithm is used to restore the image based on the known blur kernel. Since the blur kernel is directly calculated from the belt speed and exposure parameters, blind estimation is not required. Compared with traditional blind deblurring algorithms, the operation speed is improved by more than 40%, which can effectively adapt to motion blur scenarios with frequent belt speed fluctuations and start-stop phases.

[0032] S3: Establish a load-strip deformation correction model This step addresses the detection distortion caused by conveyor belt deflection and stretching under heavy loads. A correction model is built based on real-time profile and load data to provide a basis for subsequent defect size correction.

[0033] S31: Calculation of Cross-sectional Profile and Deflection Parameters of a Surface A horizontally arranged line structured light laser projects a laser stripe covering the entire bandwidth onto the strip surface. A line scan camera simultaneously acquires images of the stripe, and the depth value of each pixel is calculated using triangulation principles to obtain the depth profile curve of the strip surface cross-section. A quadratic polynomial fitting is then performed on the profile curve to obtain the strip surface deflection curve. ,in, The horizontal axis is... The depth coordinates are used; the perpendicular distance between the line connecting the two endpoints of the curve and the lowest point of the curve is calculated, which represents the maximum lateral deflection. In this embodiment, the unloaded deflection is less than 5mm, while the maximum deflection under full load can reach 30-50mm.

[0034] S32: Load and elongation calculation based on motor parameters Collect the stator operating current of the drive motor With torque The current-load correspondence is established in advance through offline calibration: the no-load current is recorded. With full load current The corresponding load ranges from 0 to the rated load. The real-time load calculation formula is: ; Based on the conveyor belt tension mechanics model, longitudinal tension With load Positive correlation; based on the comprehensive elastic modulus of the conveyor belt With cross-sectional area Calculate the longitudinal elongation ratio: ; The comprehensive elastic modulus and cross-sectional area can be obtained from the conveyor belt's factory parameters or through offline calibration.

[0035] S33: Establishment of a Multidimensional Deformation Correction Model Based on the transverse deflection parameter and the longitudinal elongation rate, three-dimensional dimensional compensation formulas are constructed to form a complete load-strip deformation correction model: Lateral dimension correction: the lateral straight-line distance of the defect obtained by projecting the image. Converting to the actual lateral dimension corresponding to the arc length of the surface, the formula is: ,in, , The coordinates of the defect's lateral boundary; Longitudinal dimension correction: This involves correcting the longitudinal length of the detected defect. Corrected to the actual length without load. The formula is This eliminates the error of excessive length caused by stretching; Depth dimension correction: Using the real-time deflected surface profile as the reference plane, the relative depth of the defect is recalculated. The formula is ,in, This is the original detection depth. The reference depth of deflection at the defect location. This is used to determine the average reference depth of the strip surface, eliminating the depth reference offset caused by the overall deflection of the strip surface.

[0036] S4: Defect detection is performed using a two-tiered architecture of initial screening and fine inspection. This step addresses the issue of limited edge computing power by balancing speed and accuracy through a tiered detection architecture, while also combining dynamic scheduling of computing power across the entire bandwidth to ensure detection coverage.

[0037] S41: First stage, lightweight and rapid initial screening A lightweight target detection model (such as an improved YOLOv8n network) is used to perform fast full-area scanning on the downsampled low-resolution band image. In this embodiment, the original image has a horizontal resolution of 4096 pixels, which is downsampled to 1024 pixels in the initial screening stage (downsampling ratio 1:4). The model only outputs the bounding box coordinates, category confidence, and preliminary category (crack, wear, hole, etc.) of suspected defects. The confidence threshold is set to 0.3: areas with a confidence level higher than 0.3 are identified as suspected areas and output to the next level; normal areas with a confidence level lower than 0.3 are skipped directly for further processing. The single-frame inference time of this level is less than 10ms, which can meet the real-time scanning requirements of the entire band at a band speed of 6m / s.

[0038] S42: Level 2, High-resolution fine detection For suspected defect areas only identified in the initial screening, the Region of Interest (ROI) is cropped from the original high-resolution image and corresponding 3D depth data. This ROI is then input into a high-precision instance segmentation model (such as Mask R-CNN) for fine-grained inference. The model outputs a precise contour mask, subdivision category, pixel-level size, and depth information for the defect, achieving accurate defect classification and preliminary geometric parameter quantization. Because only suspected areas, representing less than 5% of the total area, are processed, the overall computational cost is significantly reduced.

[0039] S43: Speed-Adaptive Dynamic Computing Power Scheduling By combining real-time belt speed data, the parameters and computing power allocation of the two-stage detection are dynamically adjusted to achieve high-speed coverage and low-speed accuracy improvement. When belt speed When in high-speed mode: the initial screening downsampling ratio is adjusted to 1:5 to increase the inference frame rate and prioritize the detection coverage of the entire area; the fine detection uses half-precision inference to compress the inference time of a single area and avoid backlog of suspected areas and missed detection. when At this time, it enters normal speed mode: the initial screening maintains a 1:4 downsampling ratio, and the fine detection adopts full-precision inference to balance detection speed and accuracy; When belt speed When the system enters low-speed mode, the initial screening downsampling ratio is adjusted to 1:2, and multi-scale detection is enabled for fine inspection to improve the accuracy of identifying minute defects and make full use of idle computing power.

[0040] S5: Multimodal data registration and risk classification of internal and external defect correlation This step integrates surface vision, 3D contour, and internal weak magnetic data to establish a mapping relationship between internal and external defects, enabling a refined assessment of defect risks.

[0041] S51: Spatiotemporal Dual Registration of Multi-Source Data Using encoder pulse signals as a unified reference, dual time and space registration is achieved, with time synchronization error controlled within 1ms. Time registration: A hard-triggered synchronization method is adopted, in which each encoder pulse synchronously triggers the camera line acquisition, laser flash and weak magnetic data sampling, ensuring that the acquisition time of the three types of data strictly corresponds; Spatial registration: Each encoder pulse corresponds to a 0.1mm advance of the conveyor belt. With the fixed position of the conveyor as the zero point, all detection data are marked with the corresponding pulse count value as a position label to ensure that the same pulse count corresponds to the same physical position of the conveyor belt, and the spatial positioning error is less than 5mm.

[0042] S52: Multi-dimensional Defect Feature Extraction Extract the corresponding defect features from the three types of data respectively: Visible light images: extract texture, grayscale, and edge features to identify texture defects such as surface cracks, scratches, and holes; Laser 3D data: Extract depth, height, and contour features to identify geometric defects such as depressions, protrusions, and thickness wear, and can effectively distinguish between protruding coal slag and real depressions; Weak magnetic field detection data: Extracting abrupt changes in magnetic flux and waveform distortion features to identify hidden defects such as broken wires, corrosion, and joint slippage in internal steel wire ropes.

[0043] S53: Risk Assessment of Internal and External Defects Establish a mapping relationship between internal and external defect locations, using a location tolerance of ±50mm as the matching threshold. When the location difference between an internal defect and a surface defect is within the tolerance range, they are identified as related defects at the same location, and the risk level is adjusted according to the following rules: If the location of an internal defect is accompanied by surface cracks, dents or other deterioration features, the defect is determined to be in an active expansion state and is automatically upgraded by one level based on the original internal defect risk level. If there are only internal defects and no corresponding surface deterioration, they are judged as early hidden defects, with a risk level of level two, and are included in the key observation sequence, with the detection frequency doubled. If there are only minor surface defects and no internal abnormalities, it is judged as ordinary surface damage, with a risk level of three, and is managed as a routine defect.

[0044] S6: Call the correction model to complete dynamic compensation of defect geometric parameters This step calls the load-strip deformation correction model established in step S3, and substitutes the preliminary defect geometric parameters (lateral width, longitudinal length, and depth) output in step S4 into the corresponding correction formula one by one to complete the dynamic compensation of the defect size and output the real defect geometric size that is independent of the load.

[0045] For example, under full load conditions, the initial transverse width of a defect is 95mm, but after arc length correction, the actual width is 100mm, reducing the error from 5% to 0. The initial longitudinal length is 103mm, but after tensile rate correction, the actual length is 100mm, reducing the error from 3% to 0. The initial depth is 8mm, but after deflection datum correction, the actual depth is 5mm, eliminating the artificial depth caused by surface deflection. Through this correction, the defect size detection error can be stably controlled within 5% across the entire load range, providing an accurate quantitative basis for risk classification.

[0046] S7: Perform multi-cycle timing cumulative verification This step relies on the cyclical operation of the conveyor belt to filter out occasional environmental interference through cross-cycle verification, and outputs the final valid defect results.

[0047] S71: Circumferential defect location matching The total number of encoder pulses N corresponding to a full circumference of the conveyor belt is pre-calibrated, meaning the encoder outputs N pulses per circumference of the conveyor belt. Within each inspection cycle, all defects are marked with a corresponding pulse count value n (0 ≤ n < N). Defects in different operating cycles are matched in position using the pulse count values, with a matching tolerance of ±10 pulses (corresponding to ±1mm position deviation), eliminating minor positional shifts caused by belt speed fluctuations.

[0048] S72: Accumulation of Cross-Period Characteristics and Validity Determination Establish a defect time sequence ledger. After each running cycle of detection is completed, match the defect features of the current cycle with the historical ledger: the cumulative number of defects that are successfully matched is incremented by 1; new features that are not successfully matched are added to the ledger and the initial number of detections is 1.

[0049] A defect is considered a valid defect and included in the final test results only if it is consistently detected for three or more consecutive operating cycles. If a defect is detected less than three times in total, or is not detected for one consecutive cycle, it is considered an occasional interference (such as temporarily attached coal slag, flying dust particles, etc.) and is removed from the valid defects.

[0050] S8: Implement a tiered, interconnected online maintenance process and complete closed-loop operation and maintenance. This step connects the testing and maintenance processes, implements tiered responses based on defect risk levels, and achieves closed-loop management of the entire process of defect location, repair, and re-inspection.

[0051] S81: Defect Classification and Response Mechanism Based on the type, size, and associated status of valid defects, they are divided into three risk levels, corresponding to different maintenance strategies: Level 1 (Severe Defect): Determined if any of the following conditions are met: The area of ​​broken wires in the internal wire rope accounts for 10% or more of the total area of ​​the wire rope at that location; the conveyor belt joint slippage exceeds the standard; the diameter of a penetrating hole in the belt surface is ≥50mm; or there is a transverse crack with a depth exceeding 1 / 2 of the belt thickness. A Level 1 defect immediately triggers an audible and visual alarm, sends a shutdown interlock signal to the conveyor control system, forces a deceleration stop, and pushes an emergency maintenance work order to the ground monitoring platform. Level 2 (Critical Defect): Internal broken wire area accounts for 3%-10%; surface crack depth exceeds 1 / 3 of the strip thickness; local wear exceeds 20% of the strip thickness. Level 2 defects generate high-priority planned maintenance work orders, which are pushed to the maintenance terminal and scheduled for maintenance within the nearest planned downtime window; Level 3 (General Defects): Minor surface scratches, wear less than 10% of the belt thickness, shallow surface dents without internal abnormalities, etc. Level 3 defects are included in the routine maintenance log and handled uniformly during monthly scheduled maintenance.

[0052] S82: Precise Defect Location and Repair Execution Defect location is achieved based on encoder position data: the system records the pulse position value corresponding to the defect. When the defect is 10m away from the maintenance station, the system sends a deceleration signal to control the conveyor belt to run at low speed. When the defect position is precisely aligned with the maintenance station, a stop signal is sent, achieving defect-point stopping. Maintenance personnel can start work directly without manual searching. For systems equipped with automatic repair devices, they can directly link with grinding, glue application, and other actuators to complete the automated repair of surface defects.

[0053] S83: Inspection, Repair, and Full Lifecycle Management After the overhaul is completed, the re-inspection process is initiated. In the following three consecutive operating cycles, the original defect location is inspected in detail to confirm whether the defect has been eliminated and whether the repair quality meets the standards. If the re-inspection is qualified, it is marked as repaired and the defect log is updated. If the re-inspection is unqualified, a new overhaul work order is generated and a second overhaul is arranged.

[0054] The system establishes a full lifecycle ledger of defects, recording the first detection time of each defect, dimensional changes in each cycle, risk level changes, maintenance records, re-inspection results, and other full-process data. Based on historical data, a grey prediction model is used to predict the defect expansion trend and remaining lifespan, and preventive maintenance is arranged in advance, realizing the transformation from "fault maintenance" to "predictive maintenance".

[0055] Those skilled in the art should understand that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art can modify or make equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for linking dynamic visual flaw detection and online maintenance of mining conveyor belts under complex working conditions, characterized in that, Includes the following steps: S1: Simultaneously acquire visible light images of the conveyor belt surface, laser 3D contour data, and weak magnetic detection data of the internal steel wire rope. At the same time, acquire real-time belt speed, environmental dust concentration, and illuminance parameters. Based on the belt speed and environmental parameters, perform dynamic acquisition and control with belt speed-environment dual feedback to ensure stable data quality. S2: Perform three-level anti-interference preprocessing on the acquired surface image, including occlusion removal, dust defogging, and motion blur elimination, to output a high-quality surface image; S3: Based on laser 3D contour data and drive motor load data, calculate the transverse deflection and longitudinal elongation of the belt surface under the current load, and establish a load-belt deformation correction model. S4: Adopting a two-stage detection architecture of initial screening and fine inspection, it performs defect detection on the pre-processed surface image and three-dimensional contour data, and outputs the location, type and geometric parameters of preliminary surface defects; S5: Correlate and match the preliminary surface defect data with the registered internal wire rope weak magnetic defect data to establish a mapping relationship between surface and internal defects and complete the preliminary risk classification; S6: Call the load-strip deformation correction model to dynamically compensate and correct the geometric parameters of the initial defect in length, width and depth, and output the true defect size that is independent of the load; S7: Based on the cyclic operation characteristics of the conveyor belt, perform multi-operation cycle time-series cumulative verification, and only determine the features that are stably detected in multiple consecutive operating cycles as valid defects, and generate the final defect detection result and risk level. S8: Based on the risk level of the final defect, link the corresponding level of online maintenance process to complete the closed-loop operation and maintenance of defect location, maintenance execution and re-inspection verification.

2. The method for linking dynamic visual flaw detection and online maintenance of mining conveyor belts under complex working conditions as described in claim 1, characterized in that, The dynamic acquisition and control of belt speed-environment dual feedback mentioned in step S1 specifically includes: The conveyor belt speed is obtained in real time by a rotary encoder, and the line frequency of the linear scan camera and the trigger frequency of the laser fill light are dynamically matched according to the target longitudinal sampling resolution to maintain a constant longitudinal sampling resolution within the belt speed fluctuation range of 0.5-6m / s. By collecting environmental parameters in real time through dust and light sensors, the laser supplementary light power is adaptively adjusted according to the dust concentration, and the camera exposure parameters are adaptively adjusted according to the ambient light intensity to compensate for the light intensity attenuation and brightness fluctuation caused by environmental interference.

3. The method for linking dynamic visual flaw detection and online maintenance of mining conveyor belts under complex working conditions as described in claim 1, characterized in that, The three-level anti-interference preprocessing described in step S2 specifically includes: The grayscale threshold and laser depth features are combined to identify areas covered by coal slag and gangue, which are marked as invalid detection areas and removed, thus completing the occlusion removal process. By combining real-time dust concentration parameters, an improved dark channel prior algorithm with adaptive intensity is used to eliminate the dust fogging effect in the image, improve image contrast, and complete the dust defogging process. Based on real-time belt speed data, a motion blur kernel is estimated, and a non-blind deconvolution algorithm is used to eliminate image motion blur caused by belt speed fluctuations, thus completing the motion blur elimination process.

4. The method for linking dynamic visual flaw detection and online maintenance of mining conveyor belts under complex working conditions as described in claim 1, characterized in that, Step S3 specifically includes: The cross-sectional profile of the conveyor belt is scanned in real time by a transversely arranged line structured light laser, the belt surface deflection curve is fitted, and the maximum transverse deflection and belt surface curvature parameters of the current belt surface are calculated. Collect the operating current and torque data of the drive motor, combine the conveyor belt tension mechanical model to calculate the real-time load and longitudinal tension, and calculate the longitudinal tensile rate based on the elastic modulus of the conveyor belt. Based on the transverse deflection parameter and the longitudinal tensile rate, compensation formulas for the transverse dimension, longitudinal dimension and depth dimension are constructed respectively, forming a load-strip deformation correction model.

5. The method for linking dynamic visual flaw detection and online maintenance of mining conveyor belts under complex working conditions as described in claim 1, characterized in that, The two-stage detection architecture of initial screening and fine-tuning described in step S4 specifically includes: First-level rapid initial screening: A lightweight detection model is used to quickly scan the low-resolution full-area image after downsampling, and output the coordinates and preliminary category of the suspected defect area. Normal and defect-free areas are skipped for further processing. Level 2 fine detection: This only targets the suspected defect areas output from the initial screening. It crops the corresponding high-resolution image and 3D contour data, inputs them into a high-precision segmentation and detection model, and completes the accurate classification, contour segmentation, and preliminary geometric parameter quantification of defects.

6. The method for linking dynamic visual flaw detection and online maintenance of mining conveyor belts under complex working conditions as described in claim 5, characterized in that, Step S4 also includes a speed-adaptive dynamic computing power scheduling step: The conveyor belt speed is acquired in real time. When the belt speed increases, the initial screening input resolution is reduced and the inference frame rate is increased to prioritize the detection coverage of the entire belt surface. When the belt speed decreases, the input resolution of the initial screening and fine inspection is increased to optimize the detection accuracy and make full use of idle computing resources.

7. The method for linking dynamic visual flaw detection and online maintenance of mining conveyor belts under complex working conditions as described in claim 1, characterized in that, Step S5 specifically includes: Based on encoder location stamps and a unified time reference, spatiotemporal dual registration is performed on three types of data: visible light, laser 3D, and weak magnetic field detection, to ensure accurate matching of multi-source data at the same physical location. Surface texture defect features, three-dimensional geometric defect features, and internal wire rope defect features were extracted separately. Establish a mapping relationship between internal and external defect locations. When the location of an internal defect is accompanied by surface degradation, the risk level of the defect is automatically increased. Defects with only internal defects and no surface characteristics are identified as early-stage hidden defects. Defects with only minor surface defects and no internal abnormalities are identified as ordinary surface damage.

8. The method for linking dynamic visual flaw detection and online maintenance of mining conveyor belts under complex working conditions as described in claim 1, characterized in that, Step S7 specifically includes: Using the entire circumference of the conveyor belt as a unit, precise matching of the circumferential position of defects is achieved based on encoder position data for each operating cycle; Record defect feature data for each operating cycle, and accumulate defect features at the same location across cycles; Only defects that are consistently detected over three or more consecutive operating cycles are considered valid defects, while occasional defects are classified as environmental interference and filtered out.

9. The method for linking dynamic visual flaw detection and online maintenance of mining conveyor belts under complex working conditions as described in claim 1, characterized in that, Step S8, which describes the online maintenance process linked to risk levels, specifically includes: The effective defects are divided into three risk levels. Among them, the first-level serious defects immediately trigger light and sound warnings and shutdown interlock signals, push emergency maintenance work orders, and force shutdown for maintenance. Level 2 critical defects generate high-priority repair work orders, and schedule planned downtime for repairs in the near future. Level 3 general defects are included in the routine maintenance log and handled uniformly during periodic maintenance.

10. The method for linking dynamic visual flaw detection and online maintenance of mining conveyor belts under complex working conditions as described in claim 1, characterized in that, The closed-loop operation and maintenance described in step S8 specifically includes: Based on encoder position data, the defect is precisely located in the circumference. When the defect reaches the maintenance station, the conveyor belt is automatically controlled to decelerate or stop at a fixed point to guide the maintenance operation. After the maintenance is completed, the original defect location will be re-inspected in subsequent operation cycles to confirm the repair effect and update the defect status. Establish a defect lifecycle ledger to record data on the entire process of defect detection, development, repair, and re-inspection, and carry out preventive maintenance by combining trend prediction models.

Images