Industrial vision-based round pipe belt conveyor belt surface monitoring system

Abstract

The application discloses a round pipe belt conveyor belt surface monitoring system based on industrial vision, relates to the technical field of conveyor belt surface monitoring, and industrial vision acquisition module is used for acquiring multi-view images in the belt running process, and frame-level correction and feature alignment are carried out in combination with the belt movement rhythm; a non-linear mapping module constructs a non-rigid unwinding model of the belt surface, and realizes three-dimensional fitting reconstruction through tensor interpolation; a defect semantic extraction module identifies abnormal areas in the image based on a graph neural network and an attention mechanism; an abnormal trend analysis module constructs a space-time projection vector according to the defect development direction, periodic change and operation frequency, predicts a structural risk area and generates a maintenance path suggestion; a visualization module dynamically presents the analysis result in a three-dimensional model, realizes abnormal positioning, image backtracking and manual checking; the system has high-precision, predictable and interactive monitoring capabilities, and is suitable for intelligent evaluation and operation and maintenance management of the conveyor belt state under complex working conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of conveyor belt surface monitoring, in particular to a round pipe belt conveyor belt surface monitoring system based on industrial vision. BACKGROUND

[0002] Round pipe belt conveyors are widely used in mine, power, metallurgy, port and other bulk material conveying occasions due to their closed conveying, strong turning ability and excellent environmental protection performance. However, as the core component of such equipment, the surface state of the conveyor belt directly affects the conveying efficiency, safety and equipment operation life. The traditional conveyor belt surface monitoring method relies on manual inspection or fixed point image acquisition, which has limited detection range, poor real-time performance and is easily affected by human subjective factors, making it difficult to realize continuous, dynamic and accurate identification of the surface state of the round pipe belt.

[0003] Although some existing image recognition schemes can realize conveyor belt surface anomaly detection, most of them are limited to flat belt structures and cannot adapt to the image distortion and feature extraction problems caused by the curved shape, variable lighting and complex background of the round pipe belt conveyor. In addition, the existing methods lack intelligent fusion of multi-point image data and quantitative evaluation mechanism of abnormal change trend, and cannot provide targeted maintenance basis, limiting their practicality and reliability in high-load continuous operation scenarios.

[0004] Therefore, there is an urgent need for a round pipe belt conveyor belt surface monitoring system that integrates industrial vision perception technology, image processing algorithms and abnormal feature analysis models, which can automatically monitor the surface state of the belt with high precision, full coverage and real-time, to improve the intelligent level of equipment operation and maintenance and the accident warning capability. SUMMARY

[0005] The purpose of the present application is to provide a round pipe belt conveyor belt surface monitoring system based on industrial vision to solve the problems in the background art.

[0006] In order to achieve the above purpose, the present application provides the following technical scheme: a round pipe belt conveyor belt surface monitoring system based on industrial vision, comprising:

[0007] An industrial vision acquisition module is used to acquire multi-view image information under different curling states during belt operation, and to perform frame-level correction and feature alignment in combination with the belt motion rhythm;

[0008] A non-linear mapping module is used to construct a non-rigid unwinding model of the belt surface based on the curling deformation relationship between multi-view images, and to perform three-dimensional fitting reconstruction of the belt unwinding shape through tensor interpolation;

[0009] The defect semantic extraction module extracts features, semantic associations and spatial attributions of the abnormalities in the reconstructed image based on a graph neural network and an attention mechanism, and outputs a trend weight atlas in combination with historical defect evolution data;

[0010] The abnormal trend analysis module constructs a space-time projection vector based on the defect development direction, periodic intensity change and belt running frequency, predicts abnormal areas that may evolve into structural risks, and generates a trend suggestion path for maintenance scheduling;

[0011] The visualization module dynamically presents the recognition and prediction results on the three-dimensional belt model, supports the user to perform point selection image backtracking, trend evolution comparison and manual review intervention based on the abnormal areas.

[0012] Preferably, the industrial vision acquisition module specifically comprises:

[0013] A plurality of industrial cameras are arranged on the operating path of the round pipe belt conveyor to obtain multi-view images of the belt in different crimping states;

[0014] According to the belt running speed information collected by the encoder, frame-level synchronization control is performed on the camera image acquisition;

[0015] Based on the corresponding relationship between the multi-view images, the image frames are time-aligned and spatial feature-corrected.

[0016] Preferably, the industrial vision acquisition module further comprises:

[0017] Extrinsic calibration parameters between the multi-view cameras are used to establish a three-dimensional point cloud re-projection model of the belt surface;

[0018] Fixed reference feature points are extracted from the multi-frame images, and frame-to-frame checking is performed in combination with the belt encoder displacement;

[0019] Through structural similarity analysis, abnormal interference frames caused by dust and jitter in the images are removed.

[0020] Preferably, the non-linear mapping module comprises:

[0021] A profile model of the belt running trajectory and crimping state is constructed to establish a non-rigid unwinding coordinate system;

[0022] The multi-view image projection points are mapped to the unwinding coordinate system, and an unwinding image tensor is generated;

[0023] Based on a tensor interpolation algorithm, the unwinding image is completed and smoothed to reconstruct the three-dimensional morphology of the belt surface.

[0024] Preferably, the tensor interpolation algorithm-based expansion image completion and smoothing process specifically includes: using a bidirectional B-spline interpolation method to interpolate and fill the tensor data in the length and width directions of the belt; giving a higher interpolation smoothness to the low-contrast area through a weight control function; and outputting a fitting quality score after reconstruction is completed.

[0025] Preferably, the defect semantic extraction module specifically includes:

[0026] A graph structure network model based on the expansion image tensor is constructed, and node image features are initialized.

[0027] Graph feature aggregation is performed through a graph neural network propagation mechanism.

[0028] The feature fusion result is clustered and classified to output a defect type, a spatial position, and a confidence score.

[0029] Preferably, the defect semantic extraction module further includes:

[0030] The historical defect evolution trajectory is encoded in a graph vector prototype manner to construct a semantic prior graph.

[0031] In the current graph node feature aggregation process, a semantic prototype attention channel is introduced.

[0032] Based on the similarity between the detection result and the historical evolution trend, a defect credibility level and a time window prediction reference value are output.

[0033] Preferably, the abnormal trend analysis module includes:

[0034] Time series data of the defect area are constructed, and area and gray dynamic features are extracted.

[0035] A predicted spatiotemporal projection vector is calculated in combination with the defect development direction and the belt running speed.

[0036] According to the trend intensity and the historical risk level, a structural risk prediction area and a maintenance suggestion path graph are generated.

[0037] Preferably, the abnormal trend analysis module further includes:

[0038] A belt surface risk thermodynamic distribution graph is calculated according to the defect projection vector trajectory, and a potential crack direction area is marked.

[0039] The running frequency of the belt is introduced as an adjustment factor of the spatiotemporal expansion function to correct the rate of the abnormal evolution path.

[0040] In combination with the trend evolution level, the running cycle, and the defect density, a maintenance task priority sorting list is output.

[0041] In the above technical solution, the application provides technical effects and advantages:

[0042] 1. The application provides a round pipe belt conveyor belt surface monitoring system based on industrial vision, which can realize image acquisition under multi-view, high frame rate and curling state during the operation of the conveyor belt, and reconstruct the real three-dimensional topography of the belt surface through nonlinear mapping and tensor interpolation technology. On this basis, the system introduces graph neural network and attention mechanism for defect semantic recognition, combines abnormal trend space projection algorithm, effectively identifies small cracks, tears, adhesives and other hidden dangers on the belt surface, and has defect development trend prediction and risk level evaluation capability, significantly improving the accuracy, real-time and intelligent level of monitoring.

[0043] 2. The application has the advantages of high automation, wide data coverage and strong model self-adaptation capability, can adapt to complex morphological changes and working condition disturbances in the round pipe belt conveying system, and reduce the false negative rate and false positive rate. At the same time, the system realizes three-dimensional defect positioning, image backtracking and manual intervention correction function through the visualization module, has good man-machine interaction and maintainability, and provides reliable technical support for the safe operation and maintenance of continuous conveying industry such as mine, power and metallurgy. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0045] Figure 1 The system module mind map of the present application. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application.

[0047] Embodiment 1, please refer to Figure 1 The round pipe belt conveyor belt surface monitoring system based on industrial vision described in this embodiment comprises:

[0048] Industrial vision acquisition module, used for acquiring multi-view image information under different curling states during belt operation, and combining belt movement rhythm for frame-level correction and feature alignment;

[0049] a non-linear mapping module configured to construct a non-rigid unwrapping model of the belt surface based on the curl deformation relationship between the multi-view images, and to perform three-dimensional fitting reconstruction on the unwrapped form of the belt by means of tensor interpolation;

[0050] a defect semantic extraction module configured to extract features, associate semantics and attribute spaces of abnormalities existing in the reconstructed images based on a graph neural network and an attention mechanism, and output a trend weight atlas in combination with historical defect evolution data;

[0051] an abnormal trend analysis module configured to construct a space-time projection vector based on a defect development direction, a periodic intensity change and a belt running frequency, predict an abnormal region that may evolve into a structural risk, and generate a trend suggestion path for maintenance scheduling;

[0052] a visualization module configured to dynamically present the recognition and prediction results on the three-dimensional belt model, and support a user to perform point selection image backtracking, trend evolution comparison and manual review intervention based on the abnormal region.

[0053] In order to make the objects, technical solutions and advantages of the present application clearer, the industrial vision acquisition module in the "industrial vision-based round pipe belt conveyor belt surface monitoring system" is described in detail below in combination with specific embodiments.

[0054] It should be understood that the following embodiments are only used to illustrate the present application, but not to limit the protection scope of the present application. Any equivalent replacement or modification made within the spirit and principle of the present application should be covered within the protection scope of the present application.

[0055] In the present embodiment, the industrial vision acquisition module is deployed at a key conveying path section of the round pipe belt conveyor, and includes a plurality of industrial camera units, a synchronous trigger controller, an image cache device and an ambient light adaptation assembly, etc.

[0056] The industrial camera units are arranged on the outer side of the belt ascending section or descending section along the conveying direction, and a plurality of pairs of opposite cameras are arranged on both sides of the belt according to the belt running state, and a multi-view image acquisition array is formed by the angle difference. Each camera has a sampling frame rate of not less than 200 fps, a resolution of not less than 1280x1024, supports global shutter, and is suitable for motion target capture under high-speed pipe winding state.

[0057] The synchronous trigger controller is configured to send acquisition signal pulses according to the belt running rhythm, so as to ensure that all cameras capture different angle images on the same physical section at the same time point, and realize frame-level synchronous acquisition.

[0058] The circular pipe belt conveyor is in a continuous crimping state during operation, and has the characteristics of nonlinear change in spatial form. In order to accurately capture the image data of the belt surface, the industrial vision acquisition module needs to perform the following crimping adaptive processing steps:

[0059] The system acquires belt displacement data through the encoder installed at the driving wheel or the driven wheel, combines the spatial calibration relationship between the camera array, and constructs the spatial spline curve of the belt running path. Further combining the acceleration and angular velocity information measured by the visual inertial unit, the attitude change amount (pitch angle, yaw angle, roll angle) of the belt at each acquisition time is calculated.

[0060] Based on the calculated attitude parameters and crimping trajectory, the reverse projection method is used for spatial registration of images collected at different angles, and the belt segments in the field of view of each camera are mapped into a unified intermediate view coordinate system. This process uses tensor remapping technology to eliminate the perspective distortion caused by bending and twisting.

[0061] In order to ensure the time consistency and spatial continuity of the collected data, the following frame-level synchronization mechanism and feature alignment strategy are used in this embodiment:

[0062] All image frames are bound with encoder displacement values when collected, and the displacement is used as a time sequence index to realize the correspondence between images at different time periods.

[0063] Considering the camera response delay and wiring asynchronicity, the system sets a frame time tolerance threshold (for example, ±5ms) in the central processing unit, and selects the image frame with the highest matching degree with the target displacement in the image buffer for alignment.

[0064] Using fixed reference patterns on the belt surface (such as weld marks and regularly attached coded markers) as visual anchor points, mixed features are used for cross-frame registration to eliminate dynamic background interference and achieve accurate alignment.

[0065] In the case of local image blur or occlusion, the system determines the resampling window according to the belt running speed and image blur gradient, dynamically adjusts the camera exposure time and gain, re-collects high-quality frames and covers the distorted area.

[0066] In order to improve the stability and robustness of the industrial vision acquisition module under complex working conditions, this embodiment introduces a redundant view design and an image fault tolerance strategy:

[0067] Each belt area is covered by at least two cameras, and the texture consistency (based on the structural similarity index SSIM) of the same area in different images is calculated to dynamically evaluate the data quality.

[0068] System integrates ambient light sensor, adjusts camera exposure parameters in real time according to light intensity, ensures stable image contrast, and is suitable for environments with dramatic changes in light such as open air and tunnels.

[0069] If the texture entropy of a camera image is lower than the preset threshold due to dust obstruction or light reflection, the system automatically discards the frame and compensates it with redundant cameras, ensuring the continuity and reliability of the entire belt image.

[0070] The processed image sequence is marked as a multi-view homogeneous image in the unwound state, which can be directly transmitted to the shape reconstruction module and defect feature extraction module, realizing seamless connection from visual acquisition to defect identification. In addition, all image data, position index, timestamp, and attitude parameters are packaged into standardized data packets for parallel processing and multi-source fusion in downstream modules.

[0071] Based on the curl deformation relationship between multi-view images, a non-rigid unwinding model of the belt surface is constructed, which includes:

[0072] The extrinsic parameters (rotation matrix R and translation vector T) of each industrial camera deployed around the belt are calibrated to obtain the spatial pose of each view relative to the unified reference coordinate system.

[0073] At the same time, the intrinsic parameters (focal length f, principal point c, and distortion coefficient k) are calibrated to ensure that the image geometric information can be accurately restored.

[0074] According to the design parameters of the circular pipe belt conveyor, a parametric model of the belt cross-sectional profile is established, denoted as a closed curve C(t), t∈[0,2π].

[0075] Combined with the running angle of the belt, the cross-sectional envelope of the belt at a certain time frame is constructed in three-dimensional space, represented as B(x,y,z).

[0076] For each image, the three-dimensional space coordinates corresponding to the image points are projected using the camera intrinsic and extrinsic parameters, obtaining the pixel projection mapping function from the image to the physical surface: ; where p is the pixel coordinate, K is the intrinsic matrix, R and T are the camera extrinsic parameters.

[0077] In three-dimensional space, a plurality of projection point sets (P1, P2,..., P n ) are extracted from the belt surface, n is the total number of projection points; the trajectory curve S(u) along the center line of the belt surface is fitted, where u represents the belt length direction coordinate.

[0078] Based on the tangent vector and normal vector of the curve S(u), a local unwinding plane is constructed to generate the unwinding coordinate system (u,v), where v is the coordinate along the width direction of the belt.

[0079] The spatial point cloud obtained by projecting the multi-view images is remapped to the (u, v) coordinate system to form a non-rigid unwrapped surface image projection plane.

[0080] Each image point is mapped to a pixel position (uᵢ, vᵢ) on the unwrapped plane and stored in an unwrapped mapping table M(u, v).

[0081] The unwrapped form of the belt is three-dimensionally fitted and reconstructed by tensor interpolation, specifically including:

[0082] All unwrapped image data points are organized into a three-dimensional tensor T∈ , where U is the number of sampling points in the length direction, V is the number of sampling points in the width direction, and C is the number of channels (such as grayscale or RGB).

[0083] For occluded and reflective areas in the multi-camera image stitching process, a neighborhood weight average + least squares fitting method is used for hole filling to construct tensor density.

[0084] In the unwrapped coordinate system, a third-order B-spline interpolation is applied to smooth fill the tensor data to ensure spatial continuity.

[0085] The interpolation formula is: , where is the B-spline basis function, is the control point intensity value, and m is the number of control points.

[0086] After interpolation, a smooth and continuous image tensor T′ is formed as the basis for the unwrapped belt surface data.

[0087] An image tensor T′ is used to generate a belt surface texture map.

[0088] Meanwhile, a three-dimensional model is fitted through the following parameterization equation by combining the profile function and the belt center trajectory S(u): , where C(u) is the belt center trajectory, R(u) is the local rotation matrix, and N(v) is the profile point vector (such as a circular or elliptical cross-section). The model constructs the unwrapped belt surface in three dimensions at each u cross-section through C(u) + N(v).

[0089] The image tensor T′ is mapped to the three-dimensional reconstructed surface S(u, v) to achieve high-precision restoration of the belt surface form in the physical space.

[0090] The obtained three-dimensional model can be used for defect positioning, subsequent surface anomaly identification, dynamic trend tracking, visual simulation display, etc.

[0091] The embodiment provides a defect semantic extraction module suitable for a belt surface monitoring system of a round pipe belt conveyor, which is based on a graph neural network and a multi-scale attention mechanism, performs defect identification, semantic clustering and spatial position attribution on an unfolded and reconstructed belt surface image data, and generates a trend weight atlas in combination with historical defect data, so as to support defect risk prediction and intelligent maintenance decision.

[0092] The main input of the defect semantic extraction module is a three-dimensional reconstructed unfolded image tensor output by the front-end nonlinear mapping module, and specifically includes:

[0093] The belt unfolded image is T'∈ , U is the sampling number in the length direction of the belt, V is the sampling number in the width direction of the belt, and C is the number of image channels.

[0094] The reconstructed surface coordinate mapping matrix is M(u, v);

[0095] The encoded displacement information and the timestamp are used for defect development tracking.

[0096] The historical defect database is Dh, and contains defect type, morphology, position, confidence and evolution track information collected in the past.

[0097] The unfolded image tensor T' is mapped into a graph structure data G=(V, E), wherein:

[0098] Each pixel point (uᵢ, vᵢ) is mapped into a graph node vᵢ∈V;

[0099] A plurality of adjacent pixels are connected to form an edge eᵢⱼ∈E, and the weight wᵢⱼ of the edge is jointly defined by the gray scale / texture similarity and the spatial distance: ; wherein sim is an image feature similarity function, and a and β are weighting coefficients.

[0100] Each node vᵢ is assigned an initial feature vector fᵢ, including but not limited to pixel value (RGB or gray scale), texture local binary pattern (LBP), Sobel gradient amplitude and HOG histogram, and matching degree of similar defect features in the historical defect database (initialized to 0).

[0101] A graph convolution network (GCN) or a dynamic graph attention network (GAT) is used to update the context feature information of each node: ; wherein represents the feature of node i at the lth layer, represents the attention weight, which measures the dependence degree of node i on neighbor j; represents a learnable weight matrix; and σ represents an activation function (such as ReLU); represents the adjacent node set of node i.

[0102] Introduce multi-scale region windows (such as 3x3, 5x5, 7x7) to extract feature information of different receptive fields, and apply attention weighting to different scale features of each node to obtain a final fusion vector ; wherein is the attention coefficient of different scales s, which is obtained by self-learning from data through the Softmax function, and S is the total number of scales.

[0103] The fused graph node features are sent to a clustering module (such as K-Means, DBSCAN, or a graph clustering algorithm) to obtain several high-similarity region clusters, each of which represents a potential abnormal region.

[0104] The "semantic prototype vector" of historical defect samples can be introduced as a guide center in the clustering process to improve the class discriminability of clustering.

[0105] Each region cluster is sent to a trained defect classification model (such as a multi-layer perceptron MLP or a lightweight CNN) to output a defect type label (such as crack, delamination, adherent, tear, etc.) and a confidence score.

[0106] The classification model is pre-trained using multi-class labeled images and combined with transfer learning to improve the generalization ability under small samples.

[0107] Using the mapping matrix M(u,v) and the three-dimensional expansion model, the (u,v) coordinates of each defect region can be mapped to the three-dimensional physical coordinates (X,Y,Z) of the actual belt surface, realizing the physical positioning of the defect.

[0108] Evolution trajectory data similar to the current detected defect is called from the historical database Dh, and a trend evolution function f(t) is constructed according to parameters such as defect category, area growth rate, and gray change gradient, and a trend weight atlas W(u,v,t) is generated to indicate the key areas and possible degree of future evolution.

[0109] The trend weight corresponding to each point (u,v) in the atlas can be represented as: ; wherein represents the historical growth rate; represents the current time series intensity fluctuation; represents the defect morphology change factor; represents the weighted coefficient. This atlas provides a region weighted scoring reference for operation and maintenance, which can be used for maintenance scheduling or device self-diagnosis triggering.

[0110] The embodiment relates to an abnormal trend analysis module for a round pipe belt conveyor belt surface monitoring system, which is used for further fusing defect development direction, periodic intensity change and belt running frequency information on the basis of acquiring defect semantic features and spatial attribution information, constructing a time-space projection vector model, predicting an abnormal area which may evolve into a structural risk, and generating a trend suggestion path diagram, so as to provide an intelligent auxiliary decision basis for equipment maintenance scheduling.

[0111] The abnormal trend analysis module depends on the following multi-source input data.

[0112] A defect set output by a defect semantic extraction module:

[0113] Including the type, position coordinates (X, Y, Z), shape features and recognition confidence of each defect.

[0114] Defect time sequence information:

[0115] Observation images, gray values, areas and edge gradients of each defect point at multiple time nodes.

[0116] Belt running frequency and rhythm information:

[0117] Output by a driving system encoder, and the unit is m / s.

[0118] For each detected defect area, image frames at multiple time points are collected on a time axis, and a feature time sequence F(t) = [A(t), G(t), E(t)] is constructed corresponding to the area A(t) of each defect, the gray mean G(t) and the edge complexity E(t). The sequence reflects the development trend of the defect and is the basis for subsequent projection analysis.

[0119] A fast Fourier transform (FFT) or wavelet analysis method is used to perform frequency domain conversion on the time sequence F(t), extract the frequency spectrum features of the periodic change signal, and judge whether the abnormality has periodic enhancement (such as intermittent expansion of some types of mechanical impact caused cracks).

[0120] The periodic index P is defined as: ; wherein is the frequency domain transformation; if there is a stable main frequency component, record the period T as a trend modeling parameter.

[0121] The centroid position of the defect area at continuous time points is calculated using a vector difference method to calculate the displacement direction of the defect area: ; T is the total number of time points; the direction vector represents the spatial trend of the possible evolution of the defect.

[0122] In order to reflect the development rate of the defect at different time stages, the evolution rate S(t) of each time point is defined: ; and the growth rate is normalized , to strengthen the role of high growth in the projection.

[0123] Comprehensive defect development direction , growth rate and belt running speed , generate the spatio-temporal projection vector field V(x, y, z, t), that is, model the potential expansion path of the current defect in the future time window ΔM: ; all vectors are synthesized to obtain the abnormal spatio-temporal path prediction area Ω.

[0124] In the Ω area, the structure risk factor R is calculated, which is used to measure whether the defect is likely to evolve into a high-risk event that causes damage to the device structure: ; wherein: represents the average growth rate; represents the spatial expansion intensity; represents the risk coefficient based on the defect type (such as a crack weight higher than an adherend); represents the weighting factor obtained by experience or model tuning.

[0125] If R> set threshold θ, it is marked as a "potential structure risk point".

[0126] Connect all defect areas with risk levels through their vector fields V to generate a trend suggestion path graph Tpath(u, v), which represents the belt segment area that needs to be focused on inspection and maintenance in the future ΔM period.

[0127] Tpath is composed of the following information:

[0128] The starting point is the center position of the current defect; the direction is the main development direction of the defect; the distance is the predicted expansion range; and the risk level (high, medium, and low) and recommended inspection priority are attached. The final result is output as a graphical visualization trend graph and a structured maintenance suggestion table.

[0129] The embodiment provides a visualization module applied to a belt surface monitoring system of a round pipe belt conveyor, which is used to dynamically map various identification, prediction and evaluation results output by a defect semantic extraction module and an abnormal trend analysis module on a high-fidelity three-dimensional belt model. The module also supports user point selection image backtracking, trend evolution comparative analysis and artificial review interaction on abnormal areas, thereby improving system interpretability and engineering usability.

[0130] The module receives various data inputs from other modules of the system, including but not limited to:

[0131] Three-dimensional unwound belt model data: three-dimensional mesh model or point cloud model (format such as OBJ, PLY or VTK) of the unwound state of the belt generated by the nonlinear mapping module;

[0132] Defect detection and classification results: physical space coordinates of the defect area, type label, confidence level, area boundary and texture image segment;

[0133] Trend prediction vector atlas: vector projection path, weight intensity atlas and prediction risk level generated by the abnormal trend analysis module;

[0134] Original image and processed image: including source view image, unwound image, crimped state image and abnormal feature image layer;

[0135] Timeline index data: time correlation model for establishing image backtracking and trend comparison.

[0136] Based on the (u, v) coordinate system and physical space mapping matrix M(u, v) provided by the nonlinear unwinding module, a three-dimensional parameterized surface model of the belt is constructed;

[0137] The grid node point is generated by the interpolation of the belt center trajectory S(u) and the transverse profile function C(v), forming a texture coordinate grid model.

[0138] Map the unwound image tensor T' to the model texture map, and generate UV map coordinates on the belt surface;

[0139] Map all defect areas and their contour boundaries to the three-dimensional model surface through M(u, v) to construct multiple abnormal layers, such as:

[0140] Current defect contour;

[0141] Defect type label (icon + color coding);

[0142] Trend path vector arrow;

[0143] Evolution thermodynamic map, etc.

[0144] The layers have adjustable display switch, transparency adjustment, abnormal level filtering and other properties.

[0145] The user can click on any abnormal area on the three-dimensional model surface through the mouse or touch control method, and the system will locate the corresponding image block according to the (u, v) value of the point in the unwound coordinate system;

[0146] The system automatically pops up an "image backtracking window" to display the following content:

[0147] The original acquisition image (multi-view) corresponding to this position in each time period;

[0148] Unfold the historical texture changes of the region in the image;

[0149] Defect growth trend curve (area change, gray gradient evolution, etc.);

[0150] Abnormal semantic feature changes (such as shape, edge complexity).

[0151] Users can select two or more time nodes, and the system supports "side-by-side image comparison" or "dynamic layer superposition" to display the defect evolution process;

[0152] Optional features include: texture change map, edge distribution map, heat distribution map, risk level trajectory map;

[0153] Trend change map supports zooming, dragging, timeline sliding, and layer transparency adjustment.

[0154] If the user believes that the system recognition result has errors, the "rectification" button can be clicked to enter the manual review mode;

[0155] Users can perform the following operations:

[0156] Region contour fine-tuning: drag the boundary node to modify the defect boundary;

[0157] Label replacement: modify the defect type label;

[0158] Missed detection supplement: manually frame new abnormal areas and label them;

[0159] Confidence score: adjust the confidence score of the recognition result.

[0160] The system records each modification operation and generates a feedback log for subsequent model training dataset self-evolution.

[0161] All defect points marked as "medium-high risk" will be automatically summarized into a "maintenance recommendation list";

[0162] The system is sorted according to the following rules:

[0163] Risk level (R value);

[0164] Trend speed (growth rate);

[0165] Distance from the driving end position;

[0166] Same region historical fault records, etc.

[0167] Support export to operation and maintenance work order or API interface transmission to enterprise MES / CMMS system.

[0168] The visualization module interface can be implemented using WebGL+Vue front-end architecture or Qt / C++ native application architecture, and the core interface components include:

[0169] Main view window (3D view rotation, zoom, section);

[0170] Abnormal layer control area (layer switch, filter);

[0171] Timeline slider (support dynamic preview and comparison);

[0172] Abnormal list and trend path panel;

[0173] Artificial interaction toolbar (label, confirm, correct, note);

[0174] System settings and interface menu (import / export, synchronization control, etc.).

[0175] The module can be deployed on a local industrial control station, or can be connected to an upper computer through a network to realize linkage between a field and a background operation and maintenance system.

[0176] The above is merely a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which shall be encompassed within the protection scope of the present application.

Claims

1. A belt surface monitoring system for circular tube belt conveyors based on industrial vision, characterized in that: include: The industrial vision acquisition module is used to acquire multi-view image information of the belt under different curling states during the belt operation, and to perform frame-level correction and feature alignment in combination with the belt movement rhythm; The industrial vision acquisition module specifically includes: deploying multiple industrial cameras along the running path of the circular tube belt conveyor to acquire multi-view images of the belt in different curling states; performing frame-level synchronization control on the camera image acquisition based on the belt running speed information acquired by the encoder; and performing time alignment and spatial feature correction on the image frames based on the correspondence between the multi-view images. The industrial vision acquisition module further includes: establishing a three-dimensional point cloud reprojection model of the belt surface using the extrinsic calibration parameters between multi-view cameras; extracting fixed reference feature points in multiple frames of images and performing inter-frame verification in combination with belt encoder displacement; and removing abnormal interference frames caused by dust and shaking in the images through structural similarity analysis. The nonlinear mapping module is used to construct a non-rigid unfolding model of the belt surface based on the curling deformation relationship between multi-view images, and to perform three-dimensional fitting and reconstruction of the unfolding shape of the belt through tensor interpolation. The nonlinear mapping module includes: constructing a profile model of the belt running trajectory and the curled state, establishing a non-rigid unfolding coordinate system; mapping the projection points of multi-view images to the unfolding coordinate system and generating unfolding image tensors; and performing completion and smoothing processing on the unfolded image based on tensor interpolation algorithms to reconstruct the three-dimensional morphology of the belt surface. The tensor interpolation algorithm-based image completion and smoothing process specifically includes: using bidirectional B-spline interpolation to interpolate and fill the tensor data in both the length and width directions of the belt; assigning higher interpolation smoothness to low-contrast regions through a weight control function; and outputting a fitting quality score after reconstruction. The defect semantic extraction module, based on graph neural networks and attention mechanisms, performs feature extraction, semantic association, and spatial attribution on anomalies in reconstructed images, and outputs a trend weight map by combining historical defect evolution data. The abnormal trend analysis module constructs a spatiotemporal projection vector based on the defect development direction, periodic intensity changes, and belt operation frequency, predicts abnormal areas that may evolve into structural risks, and generates trend suggestion paths for maintenance scheduling. The visualization module dynamically presents the identification and prediction results on the 3D belt model, supporting users to perform point-to-point image backtracking, trend evolution comparison, and manual verification based on abnormal areas.

2. The industrial vision-based circular tube belt conveyor belt surface monitoring system according to claim 1, characterized in that: The defect semantic extraction module specifically includes: Construct a graph structure network model based on unfolded image tensors and initialize node image features; Graph feature aggregation is performed using a graph neural network propagation mechanism. The feature fusion results are clustered and classified to output the defect type, spatial location, and confidence score.

3. The industrial vision-based belt surface monitoring system for circular tube conveyors according to claim 2, characterized in that: The defect semantic extraction module further includes: Encode the historical defect evolution trajectory using graph vector prototypes to construct a semantic prior graph; In the current graph node feature aggregation process, a semantic prototype attention channel is introduced; Based on the similarity between the detection results and historical evolution trends, the defect confidence level and time window prediction reference value are output.

4. The industrial vision-based belt surface monitoring system for circular tube conveyors according to claim 1, characterized in that: The abnormal trend analysis module includes: Construct time-series data of the defect area and extract its area and grayscale dynamic features; Based on the defect development direction and belt speed, the predicted spatiotemporal projection vector is calculated; Based on the trend intensity and historical risk level, a structural risk prediction area and maintenance recommendation path map are generated.

5. The industrial vision-based belt surface monitoring system for circular tube conveyors according to claim 4, characterized in that: The abnormal trend analysis module further includes: Calculate the risk thermal distribution map of the belt surface based on the defect projection vector trajectory, and mark the potential crack orientation areas; The belt running frequency is introduced as an adjustment factor for the spatiotemporal spread function to correct the rate of abnormal evolution paths; Based on the trend evolution level, operation cycle, and defect density, output a priority ranking list of maintenance tasks.

Images