Longitudinal tearing recognition device and method for mining conveying belt

Abstract

The invention relates to the technical field of image processing, and discloses a device and a method for recognizing longitudinal tearing of a mining conveying belt, which are used for improving the accuracy of recognizing the longitudinal tearing of the mining conveying belt. The device comprises: an acquisition module performing structured light projection and image acquisition to obtain original image data; the extraction module performs preprocessing and contour line extraction on the original image data to obtain geometric feature information; the generation module performs feature point analysis and stress distribution analysis on the geometric feature information to generate dynamic change features and a stress distribution diagram; the fusion module carries out fusion analysis on the sensing data set to obtain conveyor belt tearing risk evaluation data; the analysis module performs tearing probability distribution analysis based on a multi-modal neural network on the conveyor belt tearing risk evaluation data to obtain tearing identification probability distribution data; and the early warning module performs early warning information analysis according to the tear identification probability distribution data to obtain target early warning information, and performs real-time early warning according to the target early warning information.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a device and method for identifying longitudinal tearing of a mining conveyor belt. Background Art

[0002] Mining conveyor belts are important transportation equipment in coal mines, mining and other industries. Their safe and stable operation directly affects production efficiency and safety. Traditional conveyor belt monitoring methods mainly rely on manual inspections or simple sensor monitoring, such as photoelectric switches, weight sensors, etc. Although these methods can detect obvious tears or breaks, it is difficult to detect early, small longitudinal tears in time.

[0003] However, these traditional methods have many shortcomings. Manual inspections are limited by inspection frequency and personnel experience, making it difficult to achieve all-weather, high-precision monitoring. Although simple sensor monitoring can achieve continuous monitoring, it can often only detect tearing when it develops to a more serious stage, and cannot achieve early warning. In addition, these methods are difficult to cope with complex working conditions, such as dust, vibration and other interference factors, and are prone to false alarms or missed alarms. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a device and method for identifying longitudinal tears in a mining conveyor belt, which are used to improve the accuracy of identifying longitudinal tears in a mining conveyor belt.

[0005] The present invention provides a device for identifying longitudinal tear of a mining conveyor belt, comprising: an acquisition module, used to perform structured light projection and image acquisition on the surface of a running conveyor belt, and obtain original image data of the conveyor belt surface; an extraction module, used to perform preprocessing and contour line extraction on the original image data, and obtain geometric feature information of the conveyor belt surface; a generation module, used to perform feature point analysis and force distribution analysis on the geometric feature information, and generate dynamic change characteristics and a force distribution diagram of the conveyor belt surface; a fusion module, used to perform fusion analysis on a set of sensor data of multiple sensors collected in advance, and obtain conveyor belt tear risk assessment data; an analysis module, used to perform tear probability distribution analysis on the conveyor belt tear risk assessment data based on a multimodal neural network, and obtain tear identification probability distribution data; an early warning module, used to perform early warning information analysis based on the tear identification probability distribution data, obtain target early warning information, and perform real-time early warning based on the target early warning information.

[0006] In the present invention, the acquisition module is used to: project multi-wavelength structured light on the surface of the conveyor belt to obtain multiple groups of structured light profiles; capture the multiple groups of structured light profiles with a high-speed camera to obtain an original multi-spectral image sequence; perform timestamp synchronization processing on the original multi-spectral image sequence to obtain a time-aligned image data set; perform adaptive exposure compensation on the time-aligned image data set to obtain image data with balanced illumination; perform denoising processing on the image data with balanced illumination by a Wiener filter algorithm to obtain denoised image data; perform geometric correction on the denoised image data to obtain a corrected image with distortion eliminated; perform feature matching on the corrected image by a SIFT algorithm to obtain a feature correspondence between consecutive frames; perform affine transformation calculation on the feature correspondence to obtain image registration parameters; perform spatial alignment on the corrected image according to the image registration parameters to obtain a registered image sequence; perform multi-frame super-resolution reconstruction on the registered image sequence to obtain the original image data of the conveyor belt surface.

[0007] In the present invention, the extraction module is used to: perform adaptive histogram equalization processing on the original image data to obtain a contrast enhanced image; perform edge preservation noise reduction processing on the contrast enhanced image through a bilateral filtering algorithm to obtain an optimized image; perform grayscale processing on the optimized image to obtain a grayscale image; perform adaptive threshold segmentation on the grayscale image through an Otsu algorithm to obtain a binary image; perform morphological opening operation processing on the binary image to obtain an image with small noise removed; perform edge detection on the image with small noise removed through an improved Canny algorithm to obtain an initial contour line image; perform Hough transform processing on the initial contour line image to obtain a set of main straight line segments; perform curve fitting on the set of main straight line segments through a RANSAC algorithm to obtain a smooth contour line; perform simplification processing on the smooth contour line through a Douglas-Peucker algorithm to obtain a set of key points; perform curve reconstruction on the set of key points through a B-spline interpolation algorithm to obtain geometric feature information of the conveyor belt surface.

[0008] In the present invention, a generation module is used to: extract feature points from the geometric feature information by the Harris corner detection algorithm to obtain a set of candidate feature points; perform non-maximum suppression processing on the candidate feature point set to obtain an optimized feature point set; generate feature descriptors from the optimized feature point set by the BRIEF algorithm to obtain a feature descriptor set; perform spatiotemporal correlation analysis on the feature descriptor set to obtain a feature point motion model; perform motion estimation on the feature point motion model by the optical flow method to obtain dynamic change characteristics of the conveyor belt surface; perform grid division processing on the geometric feature information to obtain multiple local areas; perform shape feature extraction on the multiple local areas by the curvature analysis algorithm to obtain local deformation data; perform tensor field construction on the local deformation data to obtain an initial force distribution map; perform stress propagation simulation on the initial force distribution map to obtain an optimized force distribution map; perform abnormal pattern recognition processing on the optimized force distribution map to obtain a force distribution map on the conveyor belt surface.

[0009] In the present invention, the fusion module is used to: perform piecewise constant approximation on the sensor data set to obtain a simplified data sequence; perform feature extraction on the simplified data sequence through an adaptive sliding window algorithm to obtain a local statistical feature set; perform multi-scale entropy analysis on the local statistical feature set to obtain a complexity index sequence; perform modal separation on the complexity index sequence through a variational mode decomposition algorithm to obtain a characteristic mode set; perform mutual information analysis on the characteristic mode set to obtain a sensor-to-sensor coupling strength matrix; perform sensor grouping on the coupling strength matrix through a spectral clustering algorithm to obtain a homogeneous sensor subset; perform data fusion on the homogeneous sensor subset to obtain a multi-source information fusion result; perform rule mapping on the multi-source information fusion result through a fuzzy inference system to obtain a preliminary risk assessment value; perform time series anomaly detection on the preliminary risk assessment value to obtain a risk mutation point sequence; perform trend analysis on the risk mutation point sequence through an exponentially weighted moving average model to obtain conveyor belt tearing risk assessment data.

[0010] In the present invention, the analysis module is used to: perform data enhancement processing on the conveyor belt tear risk assessment data to obtain an expanded data set; perform feature reconstruction on the expanded data set through an autoencoder to obtain a low-dimensional representation; perform multi-head attention mechanism processing on the low-dimensional representation to obtain key feature weights; perform spatial correlation modeling on the key feature weights through a graph convolutional network to obtain a feature map; perform gated recurrent unit processing on the feature map to obtain a time series feature sequence; perform uncertainty modeling on the time series feature sequence through a variational inference algorithm to obtain probability distribution parameters; perform Monte Carlo sampling on the probability distribution parameters to obtain a model integration sample; perform comprehensive decision-making on the model integration sample through a soft voting algorithm to obtain a preliminary tear probability estimate; perform Bayesian calibration on the preliminary tear probability estimate to obtain a calibrated probability distribution; and smooth the calibrated probability distribution through kernel density estimation to obtain tear recognition probability distribution data.

[0011] In the present invention, the early warning module is used to: perform kernel density estimation on the tear recognition probability distribution data to obtain a continuous probability density function; perform risk level classification on the continuous probability density function through an adaptive threshold algorithm to obtain a multi-level risk interval; perform Markov chain modeling on the multi-level risk interval to obtain a risk state transfer matrix; perform sequence prediction on the risk state transfer matrix through a forward algorithm to obtain future risk trends; perform change point detection on the future risk trends to obtain potential risk mutation points; perform credibility assessment on the potential risk mutation points through a fuzzy inference system to obtain credible risk events; perform decision tree classification on the credible risk events to obtain a warning level determination result; perform historical pattern matching on the warning level determination result through an association rule mining algorithm to obtain a similar case set; perform case reasoning on the similar case set to obtain a preliminary warning suggestion; perform a comprehensive assessment on the preliminary warning suggestion through a multi-criteria decision-making method to obtain target warning information, and trigger a corresponding warning mechanism based on the target warning information to perform real-time warning.

[0012] The present invention also provides a longitudinal tear identification system for a mining conveyor belt, comprising: performing structured light projection and image acquisition on the surface of a running conveyor belt to obtain original image data of the conveyor belt surface; preprocessing and contour line extraction on the original image data to obtain geometric feature information of the conveyor belt surface; performing feature point analysis and force distribution analysis on the geometric feature information to generate dynamic change characteristics and force distribution diagrams of the conveyor belt surface; performing fusion analysis on a set of sensor data from multiple sensors collected in advance to obtain conveyor belt tear risk assessment data; performing tear probability distribution analysis based on a multimodal neural network on the conveyor belt tear risk assessment data to obtain tear identification probability distribution data; performing warning information analysis based on the tear identification probability distribution data to obtain target warning information, and performing real-time warning based on the target warning information.

[0013] In the technical solution provided by the present invention, the acquisition module uses structured light projection technology to obtain high-quality original image data of the conveyor belt surface, providing a reliable data basis for subsequent analysis. The extraction module effectively extracts the geometric feature information of the conveyor belt surface through a series of image processing algorithms, such as adaptive histogram equalization, bilateral filtering, Otsu threshold segmentation, etc., and improves the accuracy and robustness of feature extraction. The generation module combines advanced technologies such as Harris corner detection, BRIEF algorithm, optical flow method, etc. to conduct in-depth analysis of geometric features, generate dynamic change characteristics and force distribution diagrams of the conveyor belt surface, and comprehensively capture the motion state and stress of the conveyor belt. The fusion module realizes the effective integration of multi-sensor data through a variety of data fusion technologies, such as variational mode decomposition, spectral clustering, fuzzy reasoning, etc., and improves the comprehensiveness and reliability of tear risk assessment. The analysis module adopts a method based on multimodal neural networks, including advanced algorithms such as autoencoders, graph convolutional networks, gated recurrent units, etc., to conduct in-depth analysis of tear risks, significantly improving the accuracy and generalization ability of tear recognition. The early warning module achieves accurate quantification and classification of tearing risks through methods such as kernel density estimation, Markov chain modeling, and decision tree classification, and generates targeted early warning information through multi-criteria decision-making methods to ensure the timeliness and effectiveness of early warnings. This multi-module, multi-algorithm collaborative method can not only achieve early identification of longitudinal tearing of conveyor belts, but also adapt to complex working conditions and environments, greatly reducing the probability of false alarms and missed alarms. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] Figure 1 Schematic diagram of a device for identifying longitudinal tearing of a conveyor belt for mining according to an embodiment of the present invention; Figure 2 The present invention is a flowchart of a method for identifying longitudinal tearing of a mining conveyor belt in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0018] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0019] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , Figure 1 It is a schematic diagram of a device for identifying longitudinal tear of a mining conveyor belt according to an embodiment of the present invention, comprising: an acquisition module 101, used for performing structured light projection and image acquisition on the surface of a running conveyor belt to obtain original image data of the conveyor belt surface; an extraction module 102, used for preprocessing and contour line extraction of the original image data to obtain geometric feature information of the conveyor belt surface; a generation module 103, used for performing feature point analysis and force distribution analysis on the geometric feature information to generate dynamic change characteristics and a force distribution diagram of the conveyor belt surface; a fusion module 104, used for performing fusion analysis on a set of sensor data of multiple sensors collected in advance to obtain conveyor belt tear risk assessment data; an analysis module 105, used for performing tear probability distribution analysis on the conveyor belt tear risk assessment data based on a multimodal neural network to obtain tear recognition probability distribution data; an early warning module 106, used for performing early warning information analysis based on the tear recognition probability distribution data to obtain target early warning information, and performing real-time early warning based on the target early warning information.

[0020] It should be noted that through the coordinated work of multiple functional modules, comprehensive and accurate identification and early warning of longitudinal tearing of the conveyor belt are achieved. The acquisition module 101 uses structured light projection technology to scan the surface of the running conveyor belt. Structured light is a special light source that can form a specific light pattern on the surface of an object. When these lights are projected onto the surface of the conveyor belt, they will deform due to the height change of the surface. The high-speed camera captures these deformation patterns to obtain three-dimensional information of the conveyor belt surface. This method can accurately capture subtle changes in the surface when the conveyor belt is running at high speed, providing high-quality original image data for subsequent analysis.

[0021] The extraction module 102 performs a series of preprocessing and contour extraction on the collected raw image data. The preprocessing includes operations such as image denoising and contrast enhancement to improve image quality. The contour extraction uses an improved Canny edge detection algorithm, which can effectively extract the contour of the conveyor belt surface through steps such as Gaussian filtering, non-maximum suppression and double threshold detection. The geometric feature information obtained after these processes lays the foundation for subsequent analysis.

[0022] The generation module 103 performs an in-depth analysis on the extracted geometric feature information. The feature point analysis uses the Harris corner point detection algorithm, which calculates the autocorrelation function of the local area of ​​the image to find the corner points in the image. These corner points usually represent the significant features of the conveyor belt surface. The force distribution analysis calculates the stress distribution by establishing a tensor field model of the conveyor belt surface. The combination of these two analyses generates the dynamic change characteristics and force distribution diagram of the conveyor belt surface, which fully reflects the motion state and stress condition of the conveyor belt.

[0023] The fusion module 104 processes data from multiple sensors. These sensors may include pressure sensors, temperature sensors, vibration sensors, etc. The module uses the variational mode decomposition (VMD) algorithm to decompose and reconstruct these multi-source data. The VMD algorithm can decompose complex signals into a series of intrinsic mode functions and effectively extract the features of different frequency components. Then, these features are grouped and fused through the spectral clustering algorithm to finally obtain comprehensive conveyor belt tear risk assessment data.

[0024] The analysis module 105 uses a multimodal neural network to perform in-depth analysis on the fused risk assessment data. This network contains multiple subnetworks, such as a convolutional neural network (CNN) to process image data and a long short-term memory network (LSTM) to process time series data. Through the attention mechanism, the network can automatically learn the importance weights of different modal data. After training, the network outputs the probability distribution data of tear recognition, reflecting the tear risks at different locations and degrees.

[0025] The early warning module 106 performs early warning information analysis based on the obtained tear recognition probability distribution data. First, the probability distribution is smoothed using the kernel density estimation algorithm, and then the risk state is modeled and predicted using the Markov chain model. Based on the prediction results, the module uses a decision tree algorithm to classify the risk level, and finally generates targeted early warning information through a multi-criteria decision method and triggers the corresponding early warning mechanism.

[0026] For example, a mine uses the system to monitor a conveyor belt with a length of 1,000 meters. The acquisition module collects 50 frames of images per second, and the extraction module identifies about 1,000 feature points from them. The generation module analyzes the changes in these feature points and finds that a small 0.5mm bulge appears on the surface 800 meters away from the starting point, and the force distribution diagram shows that the stress is concentrated there. The fusion module combines the pressure sensor data and finds that the pressure there is slightly abnormal. The analysis module inputs this data into the neural network and concludes that the probability of early tearing is 75%. Based on this probability, combined with historical data and current operating conditions, the early warning module determines that it is a moderate risk and recommends detailed inspection and preventive maintenance within 24 hours. In this way, the system can detect and warn early before the tear develops to a serious stage, greatly improving the safety and service life of the conveyor belt.

[0027] By executing the above steps, the acquisition module uses structured light projection technology to obtain high-quality original image data of the conveyor belt surface, providing a reliable data basis for subsequent analysis. The extraction module effectively extracts the geometric feature information of the conveyor belt surface through a series of image processing algorithms, such as adaptive histogram equalization, bilateral filtering, Otsu threshold segmentation, etc., and improves the accuracy and robustness of feature extraction. The generation module combines advanced technologies such as Harris corner detection, BRIEF algorithm, and optical flow method to conduct in-depth analysis of geometric features, generate dynamic change characteristics and force distribution diagrams of the conveyor belt surface, and comprehensively capture the motion state and stress of the conveyor belt. The fusion module uses a variety of data fusion technologies, such as variational mode decomposition, spectral clustering, fuzzy reasoning, etc., to achieve effective integration of multi-sensor data and improve the comprehensiveness and reliability of tear risk assessment. The analysis module uses a method based on multimodal neural networks, including advanced algorithms such as autoencoders, graph convolutional networks, and gated recurrent units, to conduct in-depth analysis of tear risks, significantly improving the accuracy and generalization ability of tear recognition. The early warning module achieves accurate quantification and classification of tearing risks through methods such as kernel density estimation, Markov chain modeling, and decision tree classification, and generates targeted early warning information through multi-criteria decision-making methods to ensure the timeliness and effectiveness of early warnings. This multi-module, multi-algorithm collaborative method can not only achieve early identification of longitudinal tearing of conveyor belts, but also adapt to complex working conditions and environments, greatly reducing the probability of false alarms and missed alarms.

[0028] In a specific embodiment, the acquisition module 101 is specifically used to: (1) project multi-wavelength structured light onto the surface of the conveyor belt to obtain multiple groups of structured light profiles; (2) capture the multiple groups of structured light profiles with a high-speed camera to obtain an original multispectral image sequence; (3) perform time stamp synchronization processing on the original multispectral image sequence to obtain a time-aligned image data set; (4) perform adaptive exposure compensation on the time-aligned image data set to obtain image data with balanced illumination; (5) perform denoising processing on the image data with balanced illumination by using a Wiener filter algorithm to obtain denoised image data; (6) perform geometric correction on the denoised image data to obtain a corrected image with distortion eliminated; (7) perform feature matching on the corrected image by using a SIFT algorithm to obtain a feature correspondence relationship between consecutive frames; (8) perform affine transformation calculation on the feature correspondence relationship to obtain image registration parameters; (9) perform spatial alignment on the corrected image according to the image registration parameters to obtain a registered image sequence; (10) perform multi-frame super-resolution reconstruction on the registered image sequence to obtain the original image data of the conveyor belt surface.

[0029] Specifically, multi-wavelength structured light is projected onto the surface of the conveyor belt. Structured light is a special light source that can form a specific light pattern on the surface of an object. Using structured light of different wavelengths can obtain richer surface information, because light of different wavelengths reacts differently to surface materials. The multiple sets of structured light profiles obtained in this way are captured by a high-speed camera to form an original multispectral image sequence. The high-speed camera can capture clear images when the conveyor belt is running at high speed, and the multispectral image contains surface information at different wavelengths, which is conducive to subsequent analysis. The original multispectral image sequence is timestamped and synchronized to obtain a time-aligned image data set. This step ensures that images from different time points can correspond accurately, laying the foundation for subsequent dynamic analysis. Then, adaptive exposure compensation is performed on the time-aligned image data set to obtain image data with balanced illumination. The adaptive exposure compensation technology can dynamically adjust the exposure parameters according to different areas of the image, so that clear images can be obtained under different lighting conditions.

[0030] The Wiener filter algorithm is used to denoise the image data with balanced illumination to obtain denoised image data. The Wiener filter is a statistical filtering method that effectively removes noise while retaining image details by estimating the power spectrum of the image and the power spectrum of the noise. The denoised image data is geometrically corrected to obtain a corrected image with distortion eliminated. Geometric correction is mainly used to eliminate image deformation caused by camera angle or lens distortion. Then, the corrected image is feature matched by the SIFT (Scale Invariant Feature Transform) algorithm to obtain feature correspondences between consecutive frames. The SIFT algorithm can extract local features in the image, which are invariant to the scaling, rotation and brightness changes of the image, and are therefore particularly suitable for feature matching in dynamic scenes. The feature correspondence is calculated by affine transformation to obtain image registration parameters. Affine transformation can describe geometric relationships such as translation, rotation and scaling between images.

[0031] The corrected images are spatially aligned according to the calculated image registration parameters to obtain a registered image sequence. Spatial alignment ensures that images between different frames can overlap accurately, which is convenient for subsequent comparison and analysis. Finally, multi-frame super-resolution reconstruction is performed on the registered image sequence to obtain high-resolution original image data of the conveyor belt surface. Multi-frame super-resolution reconstruction technology generates high-resolution images by synthesizing multiple low-resolution images, which can provide more detailed information.

[0032] For example, in a mining conveyor belt monitoring system, three different wavelengths (red, green, and blue) of structured light sources are used, and high-speed cameras capture images at a speed of 100 frames per second. For a 1-meter-long conveyor belt, the system captures 30 sets of multispectral images in 0.1 seconds. These images are first aligned by time stamp to ensure that the images of the three wavelengths correspond accurately. Then, the adaptive exposure compensation algorithm analyzes the brightness distribution of each frame of the image and makes local adjustments to the areas that are too dark or too bright to make the brightness of the entire image balanced. The Wiener filter denoises the image based on the pre-estimated signal-to-noise ratio (for example, 20dB). The SIFT algorithm extracts about 1,000 feature points in each frame of the image, and establishes a correspondence between consecutive frames by comparing the descriptors of these feature points. Based on these correspondences, the affine transformation matrix between frames is calculated, and the typical matrix element value is between 0.98 and 1.02, indicating that the change between frames is small. Finally, the 10 registered images were reconstructed with super-resolution to obtain a high-definition image with a resolution four times higher, which clearly shows the tiny details on the conveyor belt surface, such as fine cracks as small as 0.1 mm wide.

[0033] In a specific embodiment, the extraction module 102 is specifically used to: (1) perform adaptive histogram equalization processing on the original image data to obtain a contrast enhanced image; (2) perform edge preservation noise reduction processing on the contrast enhanced image through a bilateral filtering algorithm to obtain an optimized image; (3) perform grayscale processing on the optimized image to obtain a grayscale image; (4) perform adaptive threshold segmentation on the grayscale image through the Otsu algorithm to obtain a binary image; (5) perform morphological opening operation processing on the binary image to obtain an image with small noise removed; (6) perform edge detection on the image with small noise removed through an improved Canny algorithm to obtain an initial contour line image; (7) perform Hough transform processing on the initial contour line image to obtain a set of main straight line segments; (8) perform curve fitting on the set of main straight line segments through the RANSAC algorithm to obtain a smooth contour line; (9) perform simplification processing on the smooth contour line through the Douglas-Peucker algorithm to obtain a set of key points; (10) perform curve reconstruction on the set of key points through the B-spline interpolation algorithm to obtain geometric feature information of the conveyor belt surface.

[0034] Specifically, the original image data is subjected to adaptive histogram equalization to obtain a contrast-enhanced image. Adaptive histogram equalization is an image enhancement technique that divides the image into multiple small regions, performs histogram equalization on each region separately, and then uses bilinear interpolation to eliminate artificial boundaries between regions, thereby improving the local contrast of the image. The contrast-enhanced image is subjected to edge-preserving denoising by a bilateral filtering algorithm to obtain an optimized image. Bilateral filtering is a nonlinear filtering method that reduces noise while preserving edges. The algorithm takes into account the spatial distance and grayscale value differences of pixels, and can keep edges sharp while smoothing the image. The optimized image is grayed to obtain a grayscale image. This step converts the color image into a single-channel grayscale image, simplifying subsequent processing.

[0035] The grayscale image is segmented by adaptive threshold using the Otsu algorithm to obtain a binary image. The Otsu algorithm is a method that automatically determines the optimal threshold. It finds the optimal segmentation threshold by maximizing the inter-class variance and divides the image into foreground and background. The binary image is processed by morphological opening operation to obtain an image with small noise removed. The opening operation is a process of corrosion followed by expansion, which can effectively remove small noise and protrusions while keeping the overall shape unchanged.

[0036] The improved Canny algorithm is used to detect the edge of the image after removing small noise points, and the initial contour line image is obtained. The Canny algorithm is a multi-stage edge detection algorithm, which includes Gaussian filtering, gradient calculation, non-maximum suppression and double threshold detection, and can effectively detect the edges in the image. The initial contour line image is processed by Hough transform to obtain the main straight line segment set. Hough transform is a method for detecting straight lines or other parameterized shapes from binary images, which is achieved by converting image space into parameter space. The RANSAC (Random Sample Consensus) algorithm is used to curve fit the main straight line segment set to obtain a smooth contour line. RANSAC is an iterative method for estimating the parameters of a mathematical model from a set of observation data containing outliers. It can find the best fitting curve in the presence of a large number of outliers. The Douglas-Peucker algorithm is used to simplify the smooth contour line to obtain a key point set. The Douglas-Peucker algorithm is a line simplification algorithm that reduces the number of points in a curve by recursively segmenting the line and removing unimportant points while maintaining the overall shape of the curve.

[0037] Finally, the curve is reconstructed by using the B-spline interpolation algorithm to obtain the geometric feature information of the conveyor belt surface. B-spline is a parametric curve that defines the curve shape through control points and can generate smooth and accurate curve representation.

[0038] For example, in a mining conveyor belt detection system, the original image size is 2048x1536 pixels. After adaptive histogram equalization processing, the local contrast of the image is significantly improved, making the subtle texture on the conveyor belt surface clearer. Bilateral filtering uses a spatial domain kernel radius of 5 pixels and a value range standard deviation of 50, which effectively removes noise in the image while retaining edge information. The optimal threshold automatically calculated by the Otsu algorithm is 127, which divides the image into two clear black and white areas. The improved Canny algorithm uses a low threshold of 30 and a high threshold of 90 to successfully detect the edge of the conveyor belt surface. The Hough transform detects 50 main straight line segments, and the RANSAC algorithm fits a smooth curve from these straight line segments, representing the contour of the conveyor belt. The Douglas-Peucker algorithm simplifies the contour that originally contained more than 1,000 points into 100 key points, reducing the amount of data while maintaining the shape characteristics. Finally, B-spline interpolation reconstructs these 100 key points into a smooth and continuous curve, accurately describing the geometric features of the conveyor belt surface.

[0039] In a specific embodiment, the generation module 103 is specifically used to: (1) extract feature points from geometric feature information using the Harris corner detection algorithm to obtain a set of candidate feature points; (2) perform non-maximum suppression processing on the candidate feature point set to obtain an optimized feature point set; (3) generate feature descriptors from the optimized feature point set using the BRIEF algorithm to obtain a feature descriptor set; (4) perform spatiotemporal correlation analysis on the feature descriptor set to obtain a feature point motion model; (5) perform motion estimation on the feature point motion model using the optical flow method to obtain dynamic change characteristics of the conveyor belt surface; (6) perform grid division processing on the geometric feature information to obtain multiple local areas; (7) perform shape feature extraction on multiple local areas using the curvature analysis algorithm to obtain local deformation data; (8) perform tensor field construction on the local deformation data to obtain an initial force distribution map; (9) perform stress propagation simulation on the initial force distribution map to obtain an optimized force distribution map; (10) perform abnormal pattern recognition processing on the optimized force distribution map to obtain a force distribution map on the conveyor belt surface.

[0040] Specifically, the Harris corner detection algorithm is used to extract feature points from geometric feature information. Harris corner detection is a corner detection method based on the local autocorrelation function of an image. It identifies corners by calculating the gradient changes around each pixel in the image. This step obtains a set of candidate feature points, which includes significant feature points on the conveyor belt surface.

[0041] The candidate feature point set is subjected to non-maximum suppression to obtain an optimized feature point set. Non-maximum suppression is a technique for screening the best local feature points. It compares the response values ​​of each candidate point with other points in its neighborhood and retains the local maximum point, thereby reducing redundant feature points. Then, the optimized feature point set is subjected to feature descriptor generation by the BRIEF (Binary Robust Independent Elementary Features) algorithm. The BRIEF algorithm is a binary descriptor that randomly selects pixel pairs around the feature point for brightness comparison to generate a binary string as a descriptor. This method has high computational efficiency and is robust to noise and illumination changes. The generated feature descriptor set is subjected to spatiotemporal correlation analysis to obtain a feature point motion model. This step compares the feature descriptors between different time frames to establish the correspondence between feature points in time and space, thereby constructing a feature point motion model. Subsequently, the feature point motion model is subjected to motion estimation by the optical flow method to obtain the dynamic change characteristics of the conveyor belt surface. The optical flow method is a technique for estimating pixel motion in an image. It estimates motion by calculating pixel displacement between adjacent frames based on the assumption of constant image brightness.

[0042] The geometric feature information is meshed to obtain multiple local areas. Meshing divides the conveyor belt surface into several small areas for local analysis. The shape features of these local areas are extracted through the curvature analysis algorithm to obtain local deformation data. Curvature analysis calculates the curvature of the surface at each point, which can effectively capture the local deformation of the surface. A tensor field is constructed for the local deformation data to obtain an initial force distribution diagram. A tensor field is a mathematical tool used to describe the stress state of an object in all directions. By converting local deformation data into a tensor field, the stress distribution on the conveyor belt surface can be intuitively represented. Next, a stress propagation simulation is performed on the initial force distribution diagram to obtain an optimized force distribution diagram. The stress propagation simulation takes into account the physical properties of the material and the law of stress transfer, and can more accurately reflect the actual stress distribution.

[0043] The optimized force distribution map is processed by abnormal pattern recognition to obtain the force distribution map of the conveyor belt surface. Abnormal pattern recognition compares the current force distribution with the distribution pattern under normal conditions to identify potential abnormal areas, thereby accurately locating the location where there may be a risk of tearing.

[0044] For example, in an actual conveyor belt monitoring system, a 10-meter-long conveyor belt was analyzed. The Harris corner detection algorithm initially identified 5,000 candidate feature points, and 1,000 optimized feature points were retained after non-maximum suppression. The BRIEF algorithm generates a 256-bit binary descriptor for each feature point. Through spatiotemporal correlation analysis, the movement of these feature points in 30 consecutive frames of images was tracked, and a feature point motion model was established. The optical flow method analysis showed that at 7.5 meters from the starting point, a group of feature points moved 20% slower than the surrounding points, indicating that there may be anomalies there. The meshing divided the conveyor belt surface into 500 small areas of 20 cm × 20 cm. Curvature analysis found that the surface curvature value near the abnormal movement area was 30% higher than that of the normal area. The tensor field construction showed that the principal stress direction in this area was at a 45-degree angle to the conveyor belt movement direction, and the stress intensity was 1.5 times the normal value. Stress propagation simulation showed that this abnormal stress distribution could extend to a range of 1 meter around it within 48 hours. The abnormal pattern recognition ultimately determined that there was a moderate tear risk in the area and recommended a detailed inspection at the next routine maintenance. This sophisticated analysis enables the system to identify potential tear risks at an early stage, greatly improving the safety and service life of the conveyor belt.

[0045] In a specific embodiment, the fusion module 104 is specifically used to: (1) perform piecewise constant approximation on the sensor data set to obtain a simplified data sequence; (2) perform feature extraction on the simplified data sequence through an adaptive sliding window algorithm to obtain a local statistical feature set; (3) perform multi-scale entropy analysis on the local statistical feature set to obtain a complexity index sequence; (4) perform modal separation on the complexity index sequence through a variational mode decomposition algorithm to obtain a characteristic mode set; (5) perform mutual information analysis on the characteristic mode set to obtain a sensor coupling strength matrix; (6) perform sensor grouping on the coupling strength matrix through a spectral clustering algorithm to obtain a homogeneous sensor subset; (7) perform data fusion on the homogeneous sensor subset to obtain a multi-source information fusion result; (8) perform rule mapping on the multi-source information fusion result through a fuzzy inference system to obtain a preliminary risk assessment value; (9) perform time series anomaly detection on the preliminary risk assessment value to obtain a risk mutation point sequence; (10) perform trend analysis on the risk mutation point sequence through an exponentially weighted moving average model to obtain conveyor belt tear risk assessment data.

[0046] Specifically, the sensor data set is piecewise-constant approximated to obtain a simplified data sequence. Piecewise constant approximation is a data compression technology that divides a continuously changing data sequence into multiple segments, each of which is represented by a constant value, thereby significantly reducing the amount of data while retaining key features. The simplified data sequence is feature extracted using an adaptive sliding window algorithm to obtain a local statistical feature set. The adaptive sliding window algorithm can dynamically adjust the window size according to the local characteristics of the data, thereby improving computational efficiency while ensuring feature extraction accuracy. Multiscale entropy analysis is performed on the local statistical feature set to obtain a complexity index sequence. Multiscale entropy is a method to measure the complexity of a time series, which reflects the inherent complexity of the data by calculating sample entropy at different time scales.

[0047] Then, the complexity index sequence is modally separated by the variational mode decomposition (VMD) algorithm to obtain the characteristic mode set. VMD is a non-recursive signal decomposition method that can decompose complex signals into a series of intrinsic mode functions, each of which represents a specific frequency component of the signal. The mutual information analysis is performed on the characteristic mode set to obtain the sensor coupling strength matrix. Mutual information analysis measures the degree of mutual dependence between different sensor data and helps to identify the correlation between sensors. The coupling strength matrix is ​​grouped by the spectral clustering algorithm to obtain a homogeneous sensor subset. Spectral clustering is a clustering method based on graph theory. It uses the similarity between data points to construct a graph and then clusters them by the spectrum (eigenvalue) of the graph. Data fusion is performed on the homogeneous sensor subset to obtain the multi-source information fusion result. Data fusion integrates the information of multiple sensors and improves the reliability and integrity of the data.

[0048] The fuzzy inference system is used to map the multi-source information fusion results to obtain a preliminary risk assessment value. The fuzzy inference system constructs a series of fuzzy rules based on expert knowledge and can handle uncertainty and ambiguity in the data. The preliminary risk assessment value is subjected to time series anomaly detection to obtain a sequence of risk mutation points. Time series anomaly detection identifies abnormal changes in risk assessment values ​​that may indicate potential tearing risks. Finally, the risk mutation point sequence is subjected to trend analysis through the exponentially weighted moving average (EWMA) model to obtain the conveyor belt tearing risk assessment data. The EWMA model is a time series prediction method that gives higher weight to recent data and can effectively capture changes in risk trends.

[0049] For example, in an actual conveyor belt monitoring system, there are 10 different types of sensors, including pressure sensors, vibration sensors, and temperature sensors, which collect 100 data points per second. Through piecewise constant approximation, every 10 data points are compressed into 1 representative value to obtain a simplified data sequence. The adaptive sliding window algorithm uses an initial window size of 50 data points, dynamically adjusts the window size between 30 and 70 according to data changes, and extracts 20 local statistical features including mean, standard deviation, kurtosis, etc. Multiscale entropy analysis calculates sample entropy at 5 different scales to obtain a 5-dimensional complexity index sequence. The VMD algorithm decomposes this sequence into 3 main modes, representing high-frequency, medium-frequency, and low-frequency components. Mutual information analysis calculates the mutual information between these 3 modes and constructs a 10×10 sensor coupling strength matrix. Spectral clustering divides the 10 sensors into 3 homogeneous subsets.

[0050] The data fusion stage uses the Dempster-Shafer evidence theory to synthesize the information of each subset. The fuzzy inference system contains 20 expert-defined rules, such as "If the pressure is high and the vibration is strong, the risk is high." Time series anomaly detection uses the local anomaly factor (LOF) algorithm to identify 5 risk mutation points. The EWMA model uses an attenuation factor of 0.2 to perform trend analysis on these mutation points.

[0051] Finally, it was concluded that within the 8 hours of monitoring, the conveyor belt showed a continuously increasing tear risk at 150 meters from the starting point, and the risk value increased from the initial 0.2 to 0.7 (full score 1). This result prompted the maintenance personnel to conduct a timely inspection and found an early longitudinal tear, which was repaired in time to avoid a possible major accident. This example shows how the method can achieve early and accurate identification of conveyor belt tear risks through complex data processing and analysis technology.

[0052] In a specific embodiment, the analysis module 105 is specifically used to: (1) perform data enhancement processing on the conveyor belt tear risk assessment data to obtain an expanded data set; (2) perform feature reconstruction on the expanded data set through an autoencoder to obtain a low-dimensional representation; (3) perform multi-head attention mechanism processing on the low-dimensional representation to obtain key feature weights; (4) perform spatial correlation modeling on the key feature weights through a graph convolutional network to obtain a feature map; (5) perform gated recurrent unit processing on the feature map to obtain a time series feature sequence; (6) perform uncertainty modeling on the time series feature sequence through a variational inference algorithm to obtain probability distribution parameters; (7) perform Monte Carlo sampling on the probability distribution parameters to obtain a model integration sample; (8) perform comprehensive decision-making on the model integration sample through a soft voting algorithm to obtain a preliminary tear probability estimate; (9) perform Bayesian calibration on the preliminary tear probability estimate to obtain a calibrated probability distribution; (10) perform smoothing on the calibrated probability distribution through kernel density estimation to obtain tear recognition probability distribution data.

[0053] Specifically, the conveyor belt tear risk assessment data was augmented to obtain an expanded data set. Data augmentation is a technique to improve the generalization ability of a model, including adding noise, rotation, scaling, and other operations to generate more diverse training data. The features of the augmented data set were reconstructed through an autoencoder to obtain a low-dimensional representation. An autoencoder is an unsupervised learning algorithm that learns an efficient representation of data by compressing the input data into a low-dimensional latent space and then reconstructing the original input. This step effectively reduces the dimensionality of the data while retaining key information.

[0054] The low-dimensional representation is processed by a multi-head attention mechanism to obtain the key feature weights. The multi-head attention mechanism is a technique that allows the model to focus on different parts of the input at the same time. It uses multiple attention heads to capture feature correlations at different levels. This step identifies the most important features and their weights for tear risk assessment. The spatial correlation of the key feature weights is modeled through a graph convolutional network to obtain a feature map. A graph convolutional network is a deep learning method for processing non-Euclidean structured data, which can effectively capture the spatial relationship between features. This step establishes the topological relationship between features and forms a comprehensive feature map.

[0055] The feature map is processed by a gated recurrent unit (GRU) to obtain a time series feature sequence. GRU is a recursive neural network that can effectively handle long-term dependency problems, thereby capturing the dynamic characteristics of features changing over time. This step converts the static feature map into a dynamic time series feature sequence. The uncertainty of the time series feature sequence is modeled by a variational inference algorithm to obtain the probability distribution parameters. Variational inference is a method of approximate Bayesian inference that estimates complex posterior distributions by optimizing a variational objective. This step quantifies the uncertainty of the model prediction and provides a confidence interval for risk assessment.

[0056] Monte Carlo sampling is performed on the probability distribution parameters to obtain model ensemble samples. Monte Carlo sampling is a method to approximate complex probability distributions through random sampling. This step generates multiple possible model predictions, which provides a basis for subsequent ensemble decisions. A comprehensive decision is made on the model ensemble samples through the soft voting algorithm to obtain a preliminary tearing probability estimate. Soft voting is an ensemble learning method that makes a final decision by combining the prediction results of multiple models. This step integrates the predictions of multiple sampling models to improve the stability and accuracy of the prediction.

[0057] A Bayesian calibration is performed on the preliminary tear probability estimate to obtain a calibrated probability distribution. Bayesian calibration is a method of adjusting the model output probability to make it closer to the true posterior probability. This step improves the reliability of the probability estimate and makes the risk assessment more accurate.

[0058] Finally, the calibration probability distribution is smoothed by kernel density estimation to obtain the tear recognition probability distribution data. Kernel density estimation is a non-parametric method used to estimate the probability density function of a random variable. This step smoothes the probability distribution and reduces the noise effect that may be caused by discrete sampling.

[0059] For example, in an actual conveyor belt monitoring system, the original risk assessment data contains 100 samples, each with 20 features. Through data enhancement, the sample size is expanded to 1,000. The autoencoder compresses the 20-dimensional features into an 8-dimensional latent space. The multi-head attention mechanism uses 4 attention heads, each focusing on 2 dimensions to obtain the key feature weights. The graph convolutional network uses 3 layers of convolution, with 16 convolution kernels per layer, to construct the feature map. The GRU network contains 64 hidden units and processes 24 hours of time series data. The variational inference algorithm estimates the mean and variance parameters at each time point. Monte Carlo sampling generates 100 model samples. The soft voting algorithm synthesizes the predictions of these 100 samples. Bayesian calibration uses historical data as a priori to adjust the probability distribution. The kernel density estimation uses a Gaussian kernel with a bandwidth of 0.1 to obtain the final tear recognition probability distribution.

[0060] This complex processing flow ultimately outputs a continuous probability distribution function, showing the probability of the conveyor belt tearing at different locations within the next 24 hours. For example, at 500 meters from the starting point, the probability of tearing within the next 6 hours is 0.02, 12 hours is 0.05, and rises to 0.15 within 24 hours. This sophisticated probability distribution provides a key reference for maintenance decisions, allowing staff to arrange maintenance in a targeted manner, greatly improving the safety and efficiency of conveyor belt operations.

[0061] In a specific embodiment, the early warning module 106 is specifically used to: (1) perform kernel density estimation on the tear recognition probability distribution data to obtain a continuous probability density function; (2) perform risk level classification on the continuous probability density function through an adaptive threshold algorithm to obtain a multi-level risk interval; (3) perform Markov chain modeling on the multi-level risk interval to obtain a risk state transfer matrix; (4) perform sequence prediction on the risk state transfer matrix through a forward algorithm to obtain a future risk trend; (5) perform change point detection on the future risk trend to obtain a potential risk mutation point; (6) perform credibility assessment on the potential risk mutation point through a fuzzy reasoning system to obtain a credible risk event; (7) perform decision tree classification on the credible risk event to obtain a warning level determination result; (8) perform historical pattern matching on the warning level determination result through an association rule mining algorithm to obtain a similar case set; (9) perform case reasoning on the similar case set to obtain a preliminary warning suggestion; (10) perform a comprehensive assessment on the preliminary warning suggestion through a multi-criteria decision method to obtain target warning information, and trigger a corresponding warning mechanism based on the target warning information to perform real-time warning.

[0062] It should be noted that the kernel density estimation is performed on the tear identification probability distribution data to obtain a continuous probability density function. Kernel density estimation is a non-parametric method that places a kernel function (usually a Gaussian kernel) on each data point and then adds these kernel functions to obtain a smooth probability density estimate. This step converts discrete probability data into a continuous density function, which is helpful for subsequent analysis. The continuous probability density function is divided into risk levels through an adaptive threshold algorithm to obtain multi-level risk intervals. The adaptive threshold algorithm dynamically adjusts the threshold according to the distribution characteristics of the probability density and divides the risk into multiple levels such as low, medium, and high. This method can better adapt to the risk distribution under different working conditions.

[0063] Markov chain modeling is performed on multi-level risk intervals to obtain the risk state transition matrix. Markov chain is a random process model that describes state transition. By analyzing historical data, the transition probability between different risk levels is calculated to construct the state transition matrix. This step captures the dynamic characteristics of risk levels changing over time. The forward algorithm is used to perform sequence prediction on the risk state transition matrix to obtain future risk trends. The forward algorithm is a dynamic programming method used to calculate the probability of an observation sequence in a hidden Markov model. Here, it is used to predict the changing trend of risk levels over a period of time in the future.

[0064] Perform change point detection on future risk trends to obtain potential risk mutation points. Change point detection aims to identify mutations or abnormal points in time series, which often represent drastic changes in risk status. Use fuzzy inference system to evaluate the credibility of potential risk mutation points and obtain credible risk events. Fuzzy inference system builds a series of fuzzy rules based on expert knowledge, which can handle uncertainty and ambiguity in risk assessment. Perform decision tree classification on credible risk events to obtain warning level determination results. Decision tree is an intuitive classification method that classifies risk events into different warning levels through a series of conditional judgments. Use association rule mining algorithm to match historical patterns on warning level determination results to obtain a set of similar cases. Association rule mining can discover implicit relationships between data items and help find historical cases similar to the current situation.

[0065] Case reasoning is performed on a set of similar cases to obtain preliminary warning suggestions. Case reasoning is a method of solving current problems based on past experience. By analyzing the handling methods of similar cases, reference suggestions are provided for the current situation. Finally, the preliminary warning suggestions are comprehensively evaluated through the multi-criteria decision-making method to obtain the target warning information, and the corresponding warning mechanism is triggered according to the target warning information to perform real-time warning. The multi-criteria decision-making method considers multiple evaluation indicators, such as risk level, economic impact, safety impact, etc., to comprehensively evaluate the necessity and urgency of the warning.

[0066] For example, in an actual conveyor belt monitoring system, a conveyor belt with a length of 1000 meters is analyzed. The kernel density estimation uses a Gaussian kernel with a bandwidth of 0.05 to obtain a continuous risk probability density function. The adaptive threshold algorithm divides the risk into three levels: low (0-0.3), medium (0.3-0.7), and high (0.7-1.0). The Markov chain model calculates the state transition matrix based on the data of the past 30 days, such as the probability of going from low risk to medium risk is 0.1, and the probability of going from medium risk to high risk is 0.2.

[0067] The forward algorithm predicts the risk trend for the next 24 hours, and the change point detection identifies that a sudden change in risk may occur in the 18th hour. The fuzzy reasoning system evaluates the credibility of this mutation point as 0.8. The decision tree classifies this event as "emergency warning level". Association rule mining finds 5 similar cases from the historical database, and case reasoning leads to the recommendation: "Carry out a comprehensive inspection within 12 hours, focusing on the area 600-700 meters from the starting point."

[0068] The multi-criteria decision-making considers the risk level (weight 0.4), economic impact (weight 0.3) and safety impact (weight 0.3), with a comprehensive score of 0.85 (full score 1), and determines that an early warning is needed immediately. The system immediately triggers the early warning mechanism, sends an emergency notice to the maintenance personnel, and reduces the conveyor belt speed by 30%. The maintenance personnel completed the inspection within 4 hours after receiving the notice and found an initial longitudinal tear of 2 cm in length 650 meters from the starting point. It was repaired in time to avoid a possible major accident. This example shows how this method can achieve early and accurate early warning of the risk of conveyor belt tearing through complex data processing and analysis technology, effectively ensuring production safety.

[0069] The embodiment of the present invention also provides a method for identifying longitudinal tearing of a mining conveyor belt. Figure 2 As shown, the flowchart of the method for identifying longitudinal tears of a mining conveyor belt specifically includes: S201, performing structured light projection and image acquisition on the surface of a running conveyor belt to obtain original image data of the conveyor belt surface; S202, preprocessing and contour line extraction on the original image data to obtain geometric feature information of the conveyor belt surface; S203, performing feature point analysis and force distribution analysis on the geometric feature information to generate dynamic change characteristics and force distribution diagrams of the conveyor belt surface; S204, performing fusion analysis on the sensor data sets of multiple sensors collected in advance to obtain conveyor belt tear risk assessment data; S205, performing tear probability distribution analysis on the conveyor belt tear risk assessment data based on a multimodal neural network to obtain tear recognition probability distribution data; S206, performing warning information analysis based on the tear recognition probability distribution data to obtain target warning information, and performing real-time warning based on the target warning information.

[0070] In the embodiment of the present application, structured light projection technology is used to obtain high-quality original image data of the conveyor belt surface, providing a reliable data basis for subsequent analysis. Through a series of image processing algorithms, such as adaptive histogram equalization, bilateral filtering, Otsu threshold segmentation, etc., the geometric feature information of the conveyor belt surface is effectively extracted, and the accuracy and robustness of feature extraction are improved. Combined with advanced technologies such as Harris corner detection, BRIEF algorithm, and optical flow method, the geometric features are deeply analyzed to generate dynamic change characteristics and force distribution diagrams of the conveyor belt surface, and the motion state and stress of the conveyor belt are fully captured. Through a variety of data fusion technologies, such as variational mode decomposition, spectral clustering, fuzzy reasoning, etc., the effective integration of multi-sensor data is achieved, and the comprehensiveness and reliability of tear risk assessment are improved. A method based on multimodal neural networks, including advanced algorithms such as autoencoders, graph convolutional networks, and gated recurrent units, is used to conduct in-depth analysis of tear risks, significantly improving the accuracy and generalization ability of tear recognition. Through methods such as kernel density estimation, Markov chain modeling, and decision tree classification, accurate quantification and classification of tearing risks are achieved, and targeted early warning information is generated through multi-criteria decision-making methods, ensuring the timeliness and effectiveness of the early warning. It can not only achieve early identification of longitudinal tears in the conveyor belt, but also adapt to complex working conditions and environments, greatly reducing the probability of false alarms and missed alarms.

[0071] The above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the embodiments, a person skilled in the art should understand that the specific implementation modes of the present invention can still be modified or replaced by equivalents, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A device for identifying longitudinal tearing of a conveyor belt for mining, characterized in that: include: An acquisition module is used to perform structured light projection and image acquisition on the surface of the conveyor belt in operation to obtain original image data of the conveyor belt surface; An extraction module, used for preprocessing the original image data and extracting contour lines to obtain geometric feature information of the conveyor belt surface; A generation module, used to perform feature point analysis and force distribution analysis on the geometric feature information, and generate dynamic change characteristics and force distribution diagram of the conveyor belt surface; A fusion module is used to perform fusion analysis on the sensor data set of multiple sensors collected in advance to obtain conveyor belt tear risk assessment data; An analysis module, used for performing tear probability distribution analysis on the conveyor belt tear risk assessment data based on a multimodal neural network to obtain tear recognition probability distribution data; The early warning module is used to analyze early warning information according to the tear recognition probability distribution data, obtain target early warning information, and make real-time early warning according to the target early warning information.

2. The device for identifying longitudinal tearing of a conveyor belt for mining use according to claim 1, characterized in that: The acquisition module is used to: project multi-wavelength structured light onto the surface of the conveyor belt to obtain multiple groups of structured light profiles; capture the multiple groups of structured light profiles with a high-speed camera to obtain an original multi-spectral image sequence; and perform time stamp synchronization processing on the original multi-spectral image sequence to obtain a time-aligned image data set; Performing adaptive exposure compensation on the time-aligned image data set to obtain image data with balanced illumination; The illumination-balanced image data is denoised by a Wiener filter algorithm to obtain denoised image data; the denoised image data is geometrically corrected to obtain a corrected image with distortion eliminated; feature matching is performed on the corrected image by a SIFT algorithm to obtain a feature correspondence relationship between consecutive frames; affine transformation is performed on the feature correspondence relationship to obtain image registration parameters; the corrected image is spatially aligned according to the image registration parameters to obtain a registered image sequence; and multi-frame super-resolution reconstruction is performed on the registered image sequence to obtain original image data of the conveyor belt surface.

3. The device for identifying longitudinal tear of a mining conveyor belt according to claim 1, characterized in that: The extraction module is used to: perform adaptive histogram equalization processing on the original image data to obtain a contrast-enhanced image; perform edge-preserving noise reduction processing on the contrast-enhanced image through a bilateral filtering algorithm to obtain an optimized image; Performing grayscale processing on the optimized image to obtain a grayscale image; performing adaptive threshold segmentation on the grayscale image using the Otsu algorithm to obtain a binary image; The binary image is subjected to morphological opening operation to obtain an image with small noise removed; the image with small noise removed is subjected to edge detection by improved Canny algorithm to obtain an initial contour line image; the initial contour line image is subjected to Hough transform processing to obtain a set of main straight line segments; the main straight line segment set is subjected to curve fitting by RANSAC algorithm to obtain a smooth contour line; the smooth contour line is simplified by Douglas-Peucker algorithm to obtain a set of key points; the key point set is subjected to curve reconstruction by B-spline interpolation algorithm to obtain geometric feature information of the conveyor belt surface.

4. The device for identifying longitudinal tear of a mining conveyor belt according to claim 1, characterized in that: A generation module is used to: extract feature points from the geometric feature information using the Harris corner detection algorithm to obtain a candidate feature point set; perform non-maximum suppression processing on the candidate feature point set to obtain an optimized feature point set; generate feature descriptors from the optimized feature point set using the BRIEF algorithm to obtain a feature descriptor set; perform spatiotemporal correlation analysis on the feature descriptor set to obtain a feature point motion model; perform motion estimation on the feature point motion model using the optical flow method to obtain dynamic change characteristics of the conveyor belt surface; The geometric feature information is meshed to obtain multiple local areas; shape features are extracted from the multiple local areas through a curvature analysis algorithm to obtain local deformation data; a tensor field is constructed for the local deformation data to obtain an initial force distribution map; stress propagation simulation is performed on the initial force distribution map to obtain an optimized force distribution map; abnormal pattern recognition is performed on the optimized force distribution map to obtain a force distribution map on the conveyor belt surface.

5. The device for identifying longitudinal tear of a mining conveyor belt according to claim 1, characterized in that: A fusion module is used to: perform piecewise constant approximation on the sensor data set to obtain a simplified data sequence; perform feature extraction on the simplified data sequence through an adaptive sliding window algorithm to obtain a local statistical feature set; perform multi-scale entropy analysis on the local statistical feature set to obtain a complexity index sequence; perform modal separation on the complexity index sequence through a variational mode decomposition algorithm to obtain a characteristic mode set; perform mutual information analysis on the characteristic mode set to obtain a coupling strength matrix between sensors; perform sensor grouping on the coupling strength matrix through a spectral clustering algorithm to obtain a homogeneous sensor subset; Performing data fusion on the homogeneous sensor subset to obtain a multi-source information fusion result; Perform rule mapping on the multi-source information fusion result through a fuzzy inference system to obtain a preliminary risk assessment value; The preliminary risk assessment value is subjected to time series anomaly detection to obtain a risk mutation point sequence; the risk mutation point sequence is subjected to trend analysis through an exponentially weighted moving average model to obtain conveyor belt tear risk assessment data.

6. The device for identifying longitudinal tear of a conveyor belt for mining use according to claim 1, characterized in that: The analysis module is used to: perform data enhancement processing on the conveyor belt tear risk assessment data to obtain an expanded data set; perform feature reconstruction on the expanded data set through an autoencoder to obtain a low-dimensional representation; and perform multi-head attention mechanism processing on the low-dimensional representation to obtain key feature weights; The spatial correlation of the key feature weights is modeled through a graph convolutional network to obtain a feature map; the feature map is processed by a gated recurrent unit to obtain a time series feature sequence; the time series feature sequence is modeled by uncertainty through a variational inference algorithm to obtain probability distribution parameters; Monte Carlo sampling is performed on the probability distribution parameters to obtain a model integration sample; a comprehensive decision is made on the model integration sample through a soft voting algorithm to obtain a preliminary tearing probability estimate; the preliminary tearing probability estimate is Bayesian calibrated to obtain a calibrated probability distribution; the calibrated probability distribution is smoothed by kernel density estimation to obtain tearing recognition probability distribution data.

7. The device for identifying longitudinal tear of a conveyor belt for mining use according to claim 6, characterized in that: An early warning module is used to: perform kernel density estimation on the tear recognition probability distribution data to obtain a continuous probability density function; perform risk level classification on the continuous probability density function through an adaptive threshold algorithm to obtain a multi-level risk interval; Performing Markov chain modeling on the multi-level risk intervals to obtain a risk state transfer matrix; performing sequence prediction on the risk state transfer matrix through a forward algorithm to obtain future risk trends; Perform change point detection on the future risk trend to obtain potential risk mutation points; Conduct credibility assessment on the potential risk mutation point through a fuzzy inference system to obtain a credible risk event; Performing decision tree classification on the credible risk events to obtain a warning level determination result; Perform historical pattern matching on the warning level determination results through an association rule mining algorithm to obtain a similar case set; Conduct case reasoning on the similar case set to obtain preliminary warning suggestions; The preliminary warning suggestions are comprehensively evaluated through a multi-criteria decision-making method to obtain target warning information, and the corresponding warning mechanism is triggered according to the target warning information to perform real-time warning.

8. A method for identifying longitudinal tearing of a mining conveyor belt, characterized in that: include: Performing structured light projection and image acquisition on the surface of the conveyor belt in operation to obtain original image data of the conveyor belt surface; performing preprocessing and contour line extraction on the original image data to obtain geometric feature information of the conveyor belt surface; Performing feature point analysis and force distribution analysis on the geometric feature information to generate dynamic change characteristics and force distribution diagram of the conveyor belt surface; A fusion analysis is performed on the sensor data set of multiple sensors collected in advance to obtain conveyor belt tear risk assessment data; a tear probability distribution analysis based on a multimodal neural network is performed on the conveyor belt tear risk assessment data to obtain tear identification probability distribution data; a warning information analysis is performed based on the tear identification probability distribution data to obtain target warning information, and a real-time warning is performed based on the target warning information.

Images