Coal conveyor belt deviation detection method and system

Abstract

The present invention provides a method and system for detecting deviation of a coal conveyor belt. By using the image fusion technology of visible light images and infrared thermal imaging images, edge contour information and temperature distribution data are organically combined, and a lightweight dual-branch neural network is used to respectively extract edge offset and temperature anomaly features, thereby effectively overcoming the defects that visible light images are easily affected by environmental interference and infrared thermal imaging images are difficult to associate with fault causes. At the same time, the structural equation model (SEM) and the dynamic time warping algorithm (DTW) are used to construct a dynamic association model through historical data and continuous frame analysis. This model can not only detect current faults in real time, but also predict potential risks based on the trend of feature changes, and identify in advance the causal chain between temperature rise caused by mechanical friction and deviation, so as to provide early fault warning, effectively avoid major accidents, and ensure production continuity.

Description

Technical Field

[0001] The present invention relates to the technical field of conveyor belt deviation detection, and in particular to a coal conveyor belt deviation detection method and a coal conveyor belt deviation detection system. Background Art

[0002] In modern industrial production systems, conveyor belt systems serve as the core carrier for the continuous transportation of bulk materials such as coal and ore. Their operational stability directly determines production efficiency and equipment safety. Especially in high-temperature, high-load industrial environments, the edge position and temperature status of the conveyor belt are key indicators for monitoring operational conditions. Traditional conveyor belt monitoring methods rely on a single sensor to detect edge deviation or temperature anomalies separately, making it difficult to comprehensively analyze the correlation between the two, resulting in insufficient accuracy in fault warnings. For example, edge detection systems based on a single visible light camera are easily affected by changes in lighting or dust obstructions, resulting in incorrect recognition. While infrared thermal imagers can capture temperature anomalies, they cannot directly correlate them with deviation issues. This limitation of single data leads to a lack of robustness in the warning model, making it difficult to adapt to dynamically changing industrial scenarios. Summary of the Invention

[0003] In order to solve the above technical problems, the present invention provides a method and system for detecting deviation of a coal conveyor belt.

[0004] The technical solution adopted in the present invention is as follows:

[0005] A method for detecting deviation of a coal conveyor belt comprises the following steps:

[0006] S1: using a multispectral imaging device to simultaneously collect multispectral image data of the conveyor belt during operation, the multispectral image data including visible light image data and infrared thermal imaging image data, preprocessing the collected original visible light image data and infrared thermal imaging image data, including denoising filtering and background correction operations to remove noise caused by environmental interference sources, and injecting physical interference simulation sample data to enhance robustness, thereby generating a first processed image set that is resistant to environmental interference;

[0007] S2: For the first processed image set, using an image fusion algorithm to perform spatial registration and feature extraction on the visible light image data and the infrared thermal imaging image data, integrating the conveyor belt edge position information and temperature distribution information, and generating a second fused image data set;

[0008] S3: Extracting the offset characteristics of the conveyor belt edge position and the distribution characteristics of the temperature anomaly points from the second fused image dataset using a lightweight dual-branch neural network, comprehensively analyzing the characteristic performance under the current operating state, and generating a first feature parameter set;

[0009] S4: Obtain edge offset records and temperature change records of the conveyor belt during a past period of time from the historical data stream, perform time series comparison with the first feature parameter set, analyze feature change trends, and generate a second associated feature data set;

[0010] S5: Analyze the dynamic causal relationship between edge deviation and temperature anomaly points using a time window-based structural equation model (SEM) based on the change trend in the second correlation feature data set, and output a third dynamic correlation map with causal direction markings;

[0011] S6: If a significant causal relationship is detected between the temperature anomaly point and the edge deviation in the third dynamic correlation map, a potential fault risk level R is comprehensively evaluated in combination with a preset mechanical friction heat threshold and a deviation tolerance threshold to generate a fourth risk assessment indicator;

[0012] S7: Mapping the spatial coupling relationship between the temperature anomaly point and the edge deviation position in the conveyor belt physical coordinate system based on the fourth risk assessment indicator to generate fifth risk distribution mapping data for subsequent determination of warning conditions;

[0013] S8: If the fifth risk distribution mapping data meets the preset deviation warning threshold, temperature anomaly threshold, and causal risk threshold, a graded alarm signal is triggered, the multispectral image data and historical data stream in the current state are recorded, and a sixth state storage repository is updated;

[0014] S9: For the updated data in the sixth state repository, a closed-loop optimization process is cyclically executed at fixed time intervals, high-value training sets are screened through causal-aware active learning, lightweight dual-branch neural network parameters and structural equation model (SEM) coefficients are updated, and a seventh optimization parameter set is generated to improve the accuracy of subsequent fault detection.

[0015] The preprocessing in S1 specifically includes: performing adaptive histogram equalization and bilateral filtering noise reduction operations on the visible light image data; performing non-uniformity correction and temperature drift compensation operations on the infrared thermal imaging image data; the physical interference simulation sample data includes dust occlusion data generated by a grayscale mask and heat wave distortion data generated by a random deformation field.

[0016] In S2, spatial registration realizes pixel-level alignment of visible light image and infrared thermal imaging image through scale-invariant feature transformation matching and random sampling consistency algorithm; feature extraction is performed by extracting visible light edge gradient map and infrared temperature saliency map, and performing adaptive weighted fusion operation according to preset weight ratio.

[0017] In S3, the lightweight dual-branch neural network includes: an edge offset branch, which uses a lightweight convolutional network to output an offset heat map and calculate the maximum offset; a temperature anomaly branch, which uses a network with a channel attention mechanism to locate the coordinates of temperature anomaly points and temperature extremes.

[0018] In S5, the dynamic causal relationship is determined by the following rules:

[0019] When the absolute value of the first path coefficient of the structural equation model (SEM) is greater than the preset coefficient threshold and the statistical significance is lower than the preset significance level, the causal direction of the deviation causing the temperature rise is marked;

[0020] When the absolute value of the second path coefficient is greater than the preset coefficient threshold and the statistical significance is lower than the preset significance level, the causal direction of the temperature rise aggravating the deviation is marked.

[0021] In S6, the calculation formula for comprehensively evaluating the potential failure risk level R is:

[0022]

[0023] Where, Indicates the conveyor belt edge offset detected in the current frame; Indicates the preset deviation tolerance threshold; represents the first weight factor; Indicates the temperature rise amplitude of the abnormal temperature point; Indicates the preset mechanical friction heat threshold; represents the second weight factor; Indicates the causal type assignment item. When the causal direction is deviation causing temperature rise, ; When the causal direction is that temperature rise aggravates deviation, ; represents the third weight factor, .

[0024] In S7, the conveyor belt physical coordinate system establishes a longitudinal coordinate axis and a transverse coordinate axis with the driving roller as the origin, maps the coordinates of the temperature anomaly point to the conveyor belt physical coordinate system, and performs a spatial superposition operation with the deviation displacement vector field.

[0025] In S8, the hierarchical alarm signal triggering rules include: triggering a first-level alarm when the fault risk level R reaches a first preset risk threshold and the offset exceeds a preset displacement limit; triggering a second-level alarm when the fault risk level R reaches a second preset risk threshold and the temperature extreme value exceeds a preset temperature baseline limit.

[0026] In S9, the causal-aware active learning is specifically as follows: screening samples in the structural equation model (SEM) whose absolute values ​​of path coefficients are greater than a preset screening threshold to form a high-value training set; fine-tuning the parameters of the lightweight two-branch neural network based on the high-value training set; and iteratively updating the path coefficients and intercept terms of the structural equation model (SEM) using the partial least squares algorithm.

[0027] A coal conveyor belt deviation detection system includes: a first generation module, which uses a multispectral imaging device to simultaneously collect multispectral image data of the conveyor belt when it is running, the multispectral image data including visible light image data and infrared thermal imaging image data, and pre-processes the collected original visible light image data and the infrared thermal imaging image data, including denoising filtering and background correction operations to remove noise caused by environmental interference sources, and injects physical interference simulation sample data to enhance robustness, thereby generating a first processed image set that is resistant to environmental interference; a second generation module, for the first processed image set, applies an image fusion algorithm to perform spatial registration and feature extraction on the visible light image data and the infrared thermal imaging image data, integrates the conveyor belt edge position information and temperature distribution information, and generates a second fused image set. a third generation module, which extracts the offset characteristics of the conveyor belt edge position and the distribution characteristics of the temperature anomaly points from the second fused image dataset through a lightweight dual-branch neural network, comprehensively analyzes the characteristic performance under the current operating state, and generates a first feature parameter set; a fourth generation module, which obtains the edge offset records and temperature change records of the conveyor belt operation in the past period of time from the historical data stream, performs a time series comparison with the first feature parameter set, analyzes the feature change trend, and generates a second associated feature dataset; an output module, which uses a time window-based structural equation model (SEM) to analyze the dynamic causal relationship between edge deviation and temperature anomaly points based on the change trend in the second associated feature dataset, and outputs a third dynamic association map with causal direction markings;

[0028] If a significant causal relationship is detected between the temperature anomaly point and edge deviation in the third dynamic correlation map, the fifth generation module comprehensively assesses the potential fault risk level R by combining the preset mechanical friction heat threshold and the deviation tolerance threshold to generate a fourth risk assessment indicator. Based on the fourth risk assessment indicator, the sixth generation module maps the spatial coupling relationship between the temperature anomaly point and the edge deviation position in the conveyor belt physical coordinate system to generate fifth risk distribution mapping data for subsequent warning condition determination. If the fifth risk distribution mapping data meets the preset deviation warning threshold, temperature anomaly threshold, and causal risk threshold, the triggering and updating module triggers a graded alarm signal, simultaneously records the multispectral image data and historical data stream in the current state, and updates the sixth state repository. The seventh generation module cyclically executes a closed-loop optimization process at fixed intervals based on the updated data in the sixth state repository. Using causal-aware active learning, it screens high-value models, updates the lightweight dual-branch neural network parameters and structural equation model (SEM) coefficients, and generates a seventh set of optimized parameters to improve the accuracy of subsequent fault detection.

[0029] Beneficial effects of the present invention:

[0030] (1) This invention organically combines edge profile information with temperature distribution data through image fusion technology that combines visible light images with infrared thermal imaging images. A lightweight dual-branch neural network is used to extract edge offset and temperature anomaly features, effectively overcoming the shortcomings of visible light images being susceptible to environmental interference and infrared thermal imaging images being difficult to correlate with fault causes. Compared with traditional single detection methods, this method improves the accuracy of edge recognition and the accuracy of determining the correlation between temperature anomalies and faults, while reducing the rates of false detection and missed detection.

[0031] (2) The present invention uses the structural equation model (SEM) and the dynamic time warping algorithm (DTW) to construct a dynamic correlation model through historical data and continuous frame analysis. It can not only detect current faults in real time, but also predict potential risks based on the trend of characteristic changes, and identify the causal chain between temperature rise and deviation caused by mechanical friction in advance, so as to provide early fault warning, effectively avoid major accidents, and ensure production continuity.

[0032] (3) The present invention establishes a causal-aware active learning closed-loop optimization mechanism. By screening high-value training sets, the lightweight dual-branch neural network and structural equation model (SEM) parameters are dynamically adjusted, thereby being able to adapt to harsh environments such as high temperature and high dust, as well as equipment aging and changes in working conditions, thereby improving the reliability of detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a main flow chart of a method for detecting deviation of a coal conveyor belt according to an embodiment of the present invention.

[0034] Figure 2 A flowchart of multispectral image acquisition and preprocessing according to an embodiment of the present invention;

[0035] Figure 3 A flowchart of feature extraction according to an embodiment of the present invention;

[0036] Figure 4 This is a flowchart of a special closed-loop optimization according to an embodiment of the present invention;

[0037] Figure 5 This is a main flow chart of a coal conveyor belt deviation detection system according to an embodiment of the present invention;

[0038] Figure 6 This is a closed-loop optimization sub-flowchart of the seventh generation module according to an embodiment of the present invention;

[0039] Figure 7 This is a decision flow chart of the trigger and update module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0041] like Figures 1-4 As shown, a method for detecting deviation of a coal conveyor belt according to an embodiment of the present invention includes the following steps:

[0042] S1: Multispectral imaging equipment is used to collect multispectral image data of the conveyor belt during operation. The multispectral image data includes visible light image data. and infrared thermal imaging image data The collected original visible light image data and infrared thermal imaging image data are preprocessed, including denoising filtering and background correction operations to remove the noise caused by environmental interference sources, and physical interference simulation sample data are injected to enhance robustness, generating a first processed image set that is resistant to environmental interference.

[0043] Visible light image data can be collected using a visible light camera, such as the Hikvision DS-2CD3T47FWDV2-IS. In a specific implementation of the present invention, the visible light camera can be installed 4 meters behind the conveyor belt drive roller, at a height of 2.8 meters, with a pitch angle of -3°. Visible light image data can be collected using an infrared thermal imager, such as the FLIR A615. During installation, the infrared thermal imager and visible light camera can be mounted side by side, with a horizontal offset of 0.2 meters (to avoid direct light source exposure). In addition, an ambient temperature sensor (PT100) can be installed near the infrared thermal imager to monitor the ambient temperature of the device in real time, which is used to correct the environmental parameters in the temperature drift compensation model and improve temperature measurement accuracy. When collecting image data, it can be set to synchronously capture 20 sets of images per second and transmit them to the edge computing server via a gigabit network.

[0044] By collecting multispectral image data, visible light image data reflects the visual information of the conveyor belt and its surrounding environment, and can present the shape, color, texture and other characteristics of the object; infrared thermal imaging image data focuses on displaying the temperature distribution on the surface of the object.

[0045] Among them, in step S1, the preprocessing specifically includes:

[0046] Adaptive histogram equalization and bilateral filtering noise reduction are performed on the visible light image data. Adaptive histogram equalization adjusts image contrast by calculating the grayscale distribution. This operation avoids the loss of detail that can occur with global equalization. It enhances the contrast between different image regions, making details like the conveyor belt's edges and surface texture clearer, facilitating subsequent extraction of key features such as edges.

[0047] Bilateral filtering considers both pixel spatial distance and grayscale differences, effectively removing image noise while preserving important information such as image edges. In coal transportation scenarios, the complex environment makes images susceptible to noise interference from dust, light fluctuations, and other factors. Bilateral filtering can remove noise from visible light images while maintaining the clarity of the conveyor belt edges, ensuring the accuracy of subsequent edge detection.

[0048] Furthermore, it also includes performing non-uniformity correction and temperature drift compensation on the infrared thermal imaging image data. Because the original detector of an infrared thermal imaging device is composed of numerous detection units, even within the same batch of products, there can be slight differences in the material properties and manufacturing processes of each detection unit. This difference causes different pixels to respond differently to infrared radiation at the same temperature, resulting in the voltage signal output by the original detector not truly reflecting the actual temperature of the object's surface. Non-uniformity correction allows for individual calibration of each pixel's characteristics, compensating for inconsistencies between pixels so that the corrected temperature value more accurately corresponds to the actual temperature distribution on the conveyor belt surface.

[0049] Temperature drift compensation is based on how temperature changes over time. Factors such as ambient temperature fluctuations and the device's own heat generation can cause the measured temperature to drift, deviating from the actual temperature. Using a temperature drift function, we analyze and model historical temperature data to predict temperature drift at different points in time. This compensation is then applied to the temperature value after non-uniformity correction, ensuring that the temperature reflected in the infrared thermal imaging image always represents the actual surface temperature of the conveyor belt, improving the reliability of temperature anomaly detection.

[0050] For the physical interference simulation sample data, it includes dust occlusion data generated by grayscale mask and heat wave distortion data generated by random deformation field.

[0051] During the coal transportation process, the conveyor belt will be faced with situations such as flying dust and heat wave interference caused by changes in ambient temperature. These interferences will affect the image quality and detection accuracy. By injecting these simulated data into the original image, it is possible to artificially create effects similar to actual interference, so that the analysis algorithms and models can "adapt" to these interferences in advance. For example, the dust occlusion data generated by the grayscale mask can simulate the effects of dust covering the conveyor belt on the blurring, grayscale changes, etc. on the visible light image; the heat wave distortion data generated by the random deformation field can simulate the distortion and deformation of the temperature distribution in the infrared thermal imaging image caused by the heat wave. The first processed image set resistant to environmental interference generated in this way In the subsequent detection process, the algorithm analyzes these images processed by interference simulation, which can effectively improve the robustness of conveyor belt deviation detection in actual complex environments and reduce the probability of false detection and missed detection.

[0052] S2: For the first processed image set, an image fusion algorithm is used to perform spatial registration and feature extraction on the visible light image data and infrared thermal imaging image data, and the conveyor belt edge position information and temperature distribution information are integrated to generate the second fused image data set. .

[0053] Visible light images can clearly present the appearance, edge contours and other visual features of the conveyor belt, while infrared thermal imaging images focus on reflecting the temperature distribution on the conveyor belt surface. Through image fusion, the advantages of these two types of images are complemented, and the edge position information of the conveyor belt and the temperature distribution information are integrated together. It can not only determine whether it is deviating based on the edge shape, but also combine temperature anomalies to assist in diagnosing the cause of the fault, thereby more accurately and comprehensively understanding the operating status of the conveyor belt and effectively improving the accuracy and reliability of fault detection.

[0054] Specifically, spatial registration achieves pixel-level alignment between visible light images and infrared thermal imaging images through scale-invariant feature transform (SIFT) matching and the random sample agreement consensus (RANSAC) algorithm. The SIFT algorithm extracts a large number of feature points that are scale-invariant, rotation-invariant, and illumination-invariant. Each feature point contains information such as position, scale, and orientation, and is described using a descriptor, which is a 128-dimensional vector. For example, in a conveyor belt image, the SIFT algorithm extracts feature points such as the belt edge corners and roller contours. These feature points are stable under different viewing angles and lighting conditions, providing a reliable basis for subsequent matching.

[0055] The RANSAC algorithm randomly selects sample points from two sets of feature points (feature points of visible light image and infrared thermal imaging image) and calculates the homography matrix by the least squares method. , where the homography matrix Describes the projective transformation relationship between two planes. In image registration, it is used to map points in a visible light image to corresponding positions in an infrared thermal imaging image. The formula is:

[0056]

[0057] Where, is the coordinate of the feature point in the visible light image, is the coordinate of the corresponding feature point in the infrared thermal imaging image.

[0058] In the embodiment of the present invention, the RANSAC algorithm calculates the homography matrix obtained by different sample points through multiple random samplings. , and count each homography matrix The number of corresponding inliers (correct matching points), select the homography matrix with the largest number of inliers As a final result, mismatched points are removed, and pixel-level alignment of the visible light image and the infrared thermal imaging image is achieved, ensuring that subsequent feature extraction and analysis are based on accurately aligned data.

[0059] Feature extraction is achieved by extracting the visible light edge gradient map Significant map with infrared temperature , and perform adaptive weighted fusion operation according to the preset weight ratio.

[0060] Specifically, edge information is obtained by calculating the gradient of each pixel in the visible light image. For example, the Sobel operator is used to convolve the image in the horizontal and vertical directions using two 3x3 convolution kernels to obtain the horizontal gradient. and vertical gradient The gradient amplitude is calculated using the following formula to form an edge gradient map. This map highlights the contour information such as the conveyor belt edge, providing a visual basis for determining whether the conveyor belt is deviating.

[0061]

[0062] Infrared temperature significance map By processing infrared thermal imaging images, areas with abnormal temperatures are highlighted. For example, methods such as threshold segmentation and local adaptive thresholding can be used to mark areas above or below the normal temperature range, forming a temperature saliency map. For example, a temperature threshold can be set and areas with temperatures above the threshold can be highlighted in the image, clearly showing temperature anomalies on the conveyor belt surface that may be caused by friction, uneven load, etc.

[0063] Extract visible light edge gradient map Significant map with infrared temperature Then, according to the preset weight ratio 、 Perform adaptive weighted fusion operation, where + = 1. Weight and According to the actual application scenario requirements, if you are more concerned about the edge deviation of the conveyor belt, you can increase it appropriately. If temperature anomaly detection is more critical, increase The value of .

[0064] Through this weighted fusion, the features of the two images are organically combined to generate a second fused image dataset. .in, The calculation formula is:

[0065]

[0066] S3: From the second fused image data set, a lightweight dual-branch neural network is used to extract the offset characteristics of the conveyor belt edge position and the distribution characteristics of the temperature anomaly points, and the characteristic performance under the current operating state is comprehensively analyzed to generate the first feature parameter set.

[0067] The lightweight dual-branch neural network includes an edge offset branch and a temperature anomaly branch. The edge offset branch uses a lightweight convolutional network to output an offset thermal map. , offset heat map The value of each pixel in the image represents the possible deviation of the conveyor belt edge at that location. The larger the value, the higher the possibility that the conveyor belt edge at that location deviates from the normal position. The coordinates of the maximum value point in the offset heat map are calculated. , and combined with the preset coordinate mapping relationship, the offset thermal map coordinates are converted into the coordinates of the actual conveyor belt physical coordinate system, and then the maximum offset is obtained. Among them, the maximum offset The expression is:

[0068]

[0069] The temperature anomaly branch uses a network with a channel attention mechanism to locate the coordinates of temperature anomaly points and temperature extremes. .

[0070] Specifically, the network with channel attention mechanism inputs the second fused image data After feature extraction and analysis, a feature map containing temperature information is output. By setting a preset temperature threshold , perform threshold screening on the feature map, and set the temperature value higher than The pixel points are marked as temperature anomaly points, and their coordinate sets That is the location of the temperature anomaly point. At the same time, among the marked temperature anomaly points, find the point with the largest temperature value, and its temperature value is the temperature extreme value. .

[0071] The calculation results of the edge offset branch and the temperature anomaly branch are combined to generate the first characteristic parameter set ,The first characteristic parameter set integrates the key quantitative ,information of conveyor belt edge offset and temperature anomaly, which is used for time series ,comparison, dynamic causal relationship analysis, and fault risk level assessment.

[0072] S4: From historical data stream Get the edge deviation record of the conveyor belt running in the past period of time and temperature change records , perform time series comparison with the first feature parameter set, analyze the feature change trend, and generate a second associated feature data set.

[0073] In a specific embodiment of the present invention, The value of can be set to the data of the past few hours to several days, such as the data of 8 hours of continuous production in a coal mine, to ensure that the historical data can reflect the characteristics of the normal operation and common fault conditions of the conveyor belt. These historical data are continuously recorded and stored during the daily operation of the conveyor belt through the detection process of steps S1-S3. In addition, the edge offset record and temperature change records in Indicates the number of data sampling points. The sampling frequency is determined according to the conveyor belt running speed and detection accuracy requirements. For example, it can be set to collect 1-10 data points per second to ensure that the data can accurately capture changes in the conveyor belt's operating status.

[0074] In an embodiment of the present invention, time series comparison is achieved by calculating the similarity between the current feature sequence and the historical feature sequence. Specifically, the similarity between the current feature sequence and the historical feature sequence is calculated using the dynamic time warping (DTW) algorithm, and the calculation formula is:

[0075]

[0076] Where, Represents the current feature sequence, which is the first feature parameter set generated by step S3 Arranged in chronological order, it reflects the current and recent operating characteristics of the conveyor belt.

[0077] Represents a historical feature sequence, i.e., an edge offset record obtained from a historical data stream and temperature change records .

[0078] is a weight coefficient used to adjust the importance of different dimensional features in similarity calculation. For example, if conveyor belt deviation has a greater impact on production safety, the weight of edge offset related features can be increased; if temperature anomalies are more likely to cause failures, the weight of temperature features can be increased. In a specific embodiment of the present invention, The value of needs to be optimized and adjusted according to specific working conditions and historical fault data.

[0079] The length of the matching path determines the length of the mapping relationship between corresponding points in the current sequence and the historical sequence when searching for the optimal match. The DTW algorithm constructs an m×n matrix (m is the length of the current sequence, n is the length of the historical sequence) and uses dynamic programming to calculate from the upper left corner of the matrix to the lower right corner, following a specific path search rule. The path that minimizes the distance metric in the formula is considered the optimal matching path between the two sequences.

[0080] and As matching indexes, they determine the specific data point locations that match each other in the current feature sequence and the historical feature sequence. The matching paths constructed through these indexes can align and compare two feature sequences with different time lengths and different change rhythms, thereby calculating their similarity.

[0081] If the current feature sequence is highly similar to the feature sequence under historical normal operating conditions, it indicates that the current operating state of the conveyor belt is stable; if the similarity is low and the similarity with the feature sequence before the historical failure gradually increases, it indicates that the conveyor belt may have a potential failure risk.

[0082] For example, if a trend in the edge offset sequence shows a gradual increase, and this trend is similar to historical records of belt deviations, combined with temperature fluctuations and the presence of an abnormally high temperature, a preliminary assessment of impending belt deviation can be made. Trend analysis can identify anomalies before a failure occurs, buying time to implement preventative measures.

[0083] The second correlation feature data set is generated by combining the comparison results of the current feature sequence with the historical feature sequence and the feature change trend analysis. The second correlation feature data set not only includes the current edge offset, temperature anomaly and other feature data, but also incorporates the similarity information between these features and historical data, change trend information, etc. For example, the second correlation feature data set may include the current maximum offset , current temperature extremes The similarity score with the historical normal state feature sequence, the slope of the edge deviation change, etc. These feature information can be used in step S5 to analyze the dynamic causal relationship between edge deviation and temperature anomaly using the structural equation model (SEM).

[0084] S5: Based on the change trend in the second correlation feature data set, the structural equation model (SEM) based on the time window is used to analyze the dynamic causal relationship between edge deviation and temperature anomaly points, and output a third dynamic correlation map with causal direction markings.

[0085] When detecting conveyor belt deviation, edge offset and abnormal temperature changes are the core research variables. Structural equation model (SEM) constructs a path model to analyze the dynamic causal relationship between them. The basic expression of structural equation model (SEM) is:

[0086]

[0087] Represents the temperature variable, which is used to describe the abnormal temperature of the conveyor belt surface; represents the offset variable, which reflects the degree to which the edge of the conveyor belt deviates from the normal position; Represents an offset variable; represents the temperature variable; and Represents the intercept term, which represents the starting value of the dependent variable when the independent variable is 0; 、 The error term is used to measure the part of the variable variation that the model cannot explain, including other influencing factors not included in the model and measurement errors; and They represent the first path coefficient and the second path coefficient of the structural equation model SEM respectively.

[0088] Among them, dynamic causal relationships can be determined by the following rules:

[0089] When the first path coefficient of the structural equation model SEM The absolute value is greater than the preset coefficient threshold And statistically significant Lower than the preset significance level , mark the causal direction of temperature rise caused by deviation.

[0090] Specifically, Measured offset variable For temperature variables degree of impact. The larger the absolute value of , the more significant the effect of the offset on temperature change. For example, =0.6 means that for every unit increase in the offset, the temperature variable will increase by 0.6 units accordingly, with other conditions remaining unchanged (the specific unit is determined by the quantization method of the actual variable). Determined based on a large amount of historical data and actual engineering experience. The absolute value is greater than the coefficient threshold , which shows that the influence of running deviation on temperature cannot be ignored in practical sense. To test the path coefficient Whether it is statistically significant reflects the probability of obtaining the current sample data results when there is no causal relationship between the assumed variables. Lower than the preset significance level When the value is 0.05 or 0.01, it means that the null hypothesis is rejected, that is, there is sufficient evidence to show that there is a real causal relationship between the offset and temperature, and it is not caused by chance factors.

[0091] When the second path coefficient The absolute value is greater than the preset coefficient threshold And statistically significant Lower than the preset significance level When the temperature rise is marked, the causal direction of the deviation is aggravated. The principle is similar to the determination of the temperature rise caused by the deviation. Measured temperature variables For offset variables The degree of impact, and As the criteria for judging the degree of influence and statistical significance respectively. For example, if And it meets the significance condition, which means that the temperature increase will cause the conveyor belt to deviate more. For every unit increase in temperature, the deviation will increase by 0.5 units accordingly.

[0092] In order to adapt to the dynamic changes in the operating status of the conveyor belt, the present invention adopts an analysis method based on a time window. The length of the time window is set according to the stability of the conveyor belt operation and the frequency of data changes. For example, data within the past 5 minutes or 10 minutes can be selected as a time window. In each time window, the second correlation feature data set is analyzed to calculate the parameters and causal relationships of the structural equation model SEM. As time goes by, the time window slides forward and the analysis results are continuously updated to capture the dynamic evolution of the causal relationship between conveyor belt deviation and temperature anomalies. In this way, the present invention can promptly detect changes in the causal relationship. For example, when the conveyor belt load changes, the transition of the causal relationship between deviation and temperature rise can be quickly detected, providing more timely and accurate information for fault warning.

[0093] Through the above-mentioned causal relationship determination, a third dynamic association map with causal direction markings is output. In the third dynamic association map, the causal relationship between edge deviation and temperature anomaly points is graphically displayed. The causal direction can be represented by arrows, with the arrows pointing from the cause variable to the result variable. For example, if it is determined that the deviation caused the temperature rise, an arrow is drawn from the node representing the offset to the node representing the temperature. At the same time, the thickness or color of the arrow can be coded according to the size of the path coefficient to intuitively display the strength of the impact. In addition, relevant parameter information, such as the path coefficient value and the statistical significance level, can also be annotated in the third dynamic association map.

[0094] S6: If a significant causal relationship is detected between the temperature anomaly point and the edge deviation in the third dynamic correlation map, the potential fault risk level R is comprehensively evaluated in combination with the preset mechanical friction heat threshold and the deviation tolerance threshold to generate a fourth risk assessment indicator.

[0095] Specifically, the present invention quantifies and integrates the conveyor belt edge offset, the temperature rise at the abnormal temperature point, and the causal relationship between the two to produce a numerical indicator that intuitively reflects the degree of potential failure risk, namely the potential failure risk level R. A higher value indicates a greater likelihood of conveyor belt failure, thus helping personnel take timely measures to reduce the probability of accidents.

[0096] Specifically, the calculation formula for the comprehensive evaluation of potential failure risk level R is:

[0097]

[0098] Where, The conveyor belt edge offset detected in the current frame, in millimeters. This offset is calculated by the edge offset branch of the lightweight two-branch neural network in step S3 and represents the actual distance the conveyor belt edge deviates from its normal position at the current moment. For example, at a certain detection moment, the calculated value d = 15 mm indicates that the conveyor belt edge is offset by 15 mm from its normal position.

[0099] Indicates the preset deviation tolerance threshold, i.e. the maximum safe deviation allowed for the conveyor belt, in millimeters. The preset deviation tolerance threshold is determined based on the design standards, operating conditions, and safety regulations of the conveyor belt equipment. For example, for a certain type of coal conveyor belt, after engineering testing and safety assessment, the preset deviation tolerance threshold is set to =50 mm, indicating that the risk of failure increases significantly when the conveyor belt deflection exceeds this value.

[0100] Represents the first weight factor, that is, the contribution weight of the offset in risk assessment, with a value range of 0-1; The size of reflects the impact of the offset on the overall fault risk. Its value is adjusted based on historical data and experience of faults caused by deviation in actual working conditions. If deviation often leads to serious faults in actual operation, it can be appropriately increased. to highlight the importance of offset in risk assessment.

[0101] Represents the normalized index of the deviation, which compares the actual deviation with the deviation tolerance threshold to obtain a relative value, which is used to measure the deviation ratio of the deviation degree relative to the safe range. =20 mm, =50 mm, , indicating that the current offset has reached 40% of the safety tolerance.

[0102] Indicates the temperature rise of the temperature anomaly point, that is, the difference between the current temperature and the baseline temperature, in degrees Celsius; it is detected and calculated by the temperature anomaly branch in step S3. The baseline temperature refers to the average temperature of the conveyor belt during normal operation, which can be obtained through long-term monitoring data statistics. For example, if the current temperature of a temperature anomaly point is 60 degrees Celsius and the baseline temperature is 40 degrees Celsius, then =60℃-40℃=20℃.

[0103] Indicates the preset mechanical friction heat threshold, that is, the critical temperature rise value that triggers the fault warning, in degrees Celsius; the preset mechanical friction heat threshold is determined based on the heat resistance of the conveyor belt material, the friction characteristics of the mechanical components, and historical fault data. For example, for a specific conveyor belt, after testing and analysis, the threshold is set. =30℃. When the temperature rise of the abnormal temperature point exceeds this value, it indicates that there may be hidden dangers of failure caused by excessive mechanical friction or other reasons.

[0104] Represents the second weight factor, that is, the contribution weight of the temperature rise in the risk assessment; the value range is between 0 and 1, and ; The size of reflects the impact of temperature anomaly on the overall failure risk, and is also adjusted according to the actual operating conditions in which temperature anomaly causes failure. If temperature anomaly is more likely to cause serious failure, increase The value of .

[0105] Represents the normalized index of temperature rise, which compares the actual temperature rise with the critical temperature rise value to obtain a relative value reflecting the degree of temperature anomaly. , hour, , indicating that the temperature rise at the current temperature anomaly point has reached about 67% of the critical temperature rise value.

[0106] Indicates the causal type assignment item, which is assigned according to the causal direction determined by S5; when the causal direction is that the deviation causes the temperature rise, ; When the causal direction is that temperature rise aggravates deviation, The impact of two different causal relationships on failure risk is distinguished by assigning items.

[0107] Indicates the third weight factor, that is, the additional weight of the causal type in the risk assessment; the value range is between 0-1, which is used to measure the additional impact of the causal relationship on the overall failure risk. If a specific causal relationship is more likely to cause serious failures, it can be appropriately increased. The value of .

[0108] In a specific embodiment of the present invention, when calculating the potential fault risk level R, first obtain the conveyor belt edge offset d and the temperature rise amplitude of the temperature abnormal point in the current frame from step S3. , get the causal type assignment item from step S5 , and call the pre-set 、 as well as 、 、 Then these parameters are substituted into the calculation formula of potential fault risk level R to calculate the potential fault risk level R, that is, to generate the fourth risk assessment index. 、 、 , =20 mm, =50 mm, =20℃, , and the causal relationship is that the deviation causes the temperature rise, that is Substituting these values ​​into the formula yields: .

[0109] S7: Based on the fourth risk assessment indicator, the spatial coupling relationship between the temperature anomaly point and the edge deviation position is mapped in the physical coordinate system of the conveyor belt to generate fifth risk distribution mapping data for subsequent determination of warning conditions.

[0110] In this method, the conveyor belt's physical coordinate system establishes a longitudinal coordinate axis (y) and a transverse coordinate axis (x) with the drive roller as the origin. The coordinates of the temperature anomaly point are mapped to the conveyor belt's physical coordinate system and spatially superimposed with the deviation displacement vector field. The longitudinal coordinate axis (y) aligns with the conveyor belt's running direction and represents the length of the conveyor belt. The transverse coordinate axis (x) is perpendicular to the conveyor belt's running direction and represents the width of the conveyor belt.

[0111] By establishing a coordinate system with the drive roller as the origin, the present invention can closely link the mechanical structure and operating status of the conveyor belt, thereby accurately mapping the detected risk information to the actual physical location. For example, on a coal conveyor belt with a length of 100 meters and a width of 1.2 meters, this coordinate system can accurately represent the position coordinates of any point on the conveyor belt. For example, the coordinates of a point 20 meters from the starting point of the drive roller and located in the middle of the conveyor belt can be expressed as (0, 20) (assuming the horizontal middle position is x = 0).

[0112] The coordinates of the temperature anomaly points obtained from step S3 , it is necessary to convert from the image coordinate system to the conveyor belt physical coordinate system. Since there is a certain geometric relationship between the image coordinate system and the actual physical coordinate system during the image acquisition process, it is necessary to convert it through pre-calibrated parameters. Specifically, during the equipment installation and commissioning phase, a calibration plate of known size will be used to calibrate the multispectral imaging equipment to obtain the conversion matrix between the image coordinate system and the physical coordinate system. Through this conversion matrix, the coordinates of the temperature anomaly point in the image are converted to Converted to the actual position coordinates in the conveyor belt physical coordinate system For example, after conversion, a temperature anomaly originally located at (100, 200) in the image coordinate system might correspond to a location in the conveyor belt's physical coordinate system of (0.2, 30), 30 meters from the starting point of the drive roller and 0.2 meters to the right of the conveyor belt. This allows the temperature anomaly to be accurately located at its actual physical location on the conveyor belt.

[0113] The belt edge offset calculated in step S3 is used to construct the deviation vector field. This deviation vector field describes the deviation direction and magnitude at different belt locations. For each point on the belt, there is a corresponding deviation vector. The direction of this vector indicates the deviation direction of the belt edge (e.g., left or right), and the length of the vector indicates the magnitude of the deviation. For example, if the deviation vector at a certain belt location indicates a 5 mm rightward offset, then the deviation vector in the deviation vector field will be horizontally pointing to the right, and its length will correspond to a physical distance of 5 mm.

[0114] The mapped temperature anomaly coordinates are spatially superimposed with the deviation displacement vector field. This means the temperature anomaly is marked with a specific marker (e.g., a red dot) in the conveyor belt's physical coordinate system. The deviation displacement vector field is simultaneously plotted as an arrow in the same coordinate system to generate the fifth risk distribution map. This allows operators to intuitively identify the specific location of the temperature anomaly on the conveyor belt and the deviation of the conveyor belt edge at that location.

[0115] Therefore, in the embodiment of the present invention, by observing the fifth risk distribution mapping data, the staff can intuitively understand the current risk status of the conveyor belt, including the risk location, risk type (deviation or temperature abnormality) and the relationship between the two.

[0116] S8: If the fifth risk distribution mapping data meets the preset deviation warning threshold, temperature anomaly threshold and causal risk threshold, a graded alarm signal is triggered, and the multispectral image data and historical data stream in the current state are recorded, and the sixth state repository is updated.

[0117] In the steps of the present invention, the fifth risk distribution mapping data is first compared with the preset deviation warning threshold, temperature abnormality threshold and causal risk threshold to determine whether there is a risk in the current conveyor belt operation state. Only when the risk distribution mapping data meets these basic risk threshold conditions, indicating that there is a certain degree of risk hidden danger, will further analysis be conducted based on the fault risk level R, offset d, temperature extreme value, etc. And other specific indicators, according to the set trigger rules to determine whether to trigger a level one alarm or a level two alarm.

[0118] Specifically, the hierarchical alarm signal triggering rules include:

[0119] When the fault risk level R reaches the first preset risk threshold And the offset d exceeds the preset displacement limit It should be noted that the first level alarm indicates that the equipment is in a state close to failure and requires the operator to pay immediate attention and take corresponding measures, such as adjusting the operating parameters of the conveyor belt, checking mechanical components, etc., to avoid the occurrence of failure. For example, , If R=0.75 and d=35 mm are calculated in a certain test, the triggering conditions of the first level alarm are met and the system will issue a first level alarm signal to remind the staff that the risk of conveyor belt deviation is high.

[0120] When the fault risk level R reaches the second preset risk threshold And the temperature extremes Exceeding the preset temperature baseline limit It should be noted that the second level alarm indicates that not only is the overall risk very high, but the temperature anomaly is also very serious, and the equipment is very likely to fail in a short period of time. After the second level alarm is issued, it is necessary to immediately take measures such as emergency shutdown and conduct a comprehensive inspection and maintenance of the equipment to prevent major accidents. For example, when , , if a certain test R=0.9, , a secondary alarm is triggered, and the system will alert the staff in a stronger way (such as a high-decibel alarm, flashing warning lights, etc.).

[0121] When a graded alarm signal is triggered, the system records the current multispectral image data and historical data streams, and updates the sixth state repository for use in fault analysis and tracing, model optimization, and training. For example, the recorded multispectral image data includes information such as the conveyor belt's appearance and temperature distribution at the moment of the alarm. By analyzing these images, technicians can intuitively understand the conveyor belt's condition before the fault occurred. Combined with records of offsets and temperature changes in the historical data stream, they can trace the cause and development of the fault.

[0122] It should be noted that the above preset deviation warning threshold and displacement limit The preset temperature anomaly threshold and temperature baseline limit are determined based on the mechanical design parameters, operation safety standards and historical fault data of the conveyor belt equipment. The causal risk threshold is primarily determined based on factors such as the material properties of the conveyor belt, the frictional heating patterns of mechanical components, and the influence of ambient temperature. The causal risk threshold is derived from an in-depth analysis of the causal relationship between conveyor belt deviation and temperature anomalies, as well as a summary of actual failure cases. The dynamic causal relationship between the two has been clarified in step S5, and corresponding causal risk thresholds are set based on the severity and probability of failures caused by different causal relationships. For example, if the causal relationship is determined to be one of temperature rise exacerbating deviation, a lower causal risk threshold can be set to trigger an alarm more promptly, as this situation could lead to rapid deterioration of the fault.

[0123] S9: For the updated data in the sixth state repository, the closed-loop optimization process is executed cyclically at fixed time intervals. High-value training sets are screened through causal-aware active learning, and the lightweight two-branch neural network parameters and structural equation model (SEM) coefficients are updated to generate the seventh optimized parameter set for improving the accuracy of subsequent fault detection.

[0124] Due to the complex and ever-changing operating environment of coal conveyor belts, factors such as equipment wear and operating condition adjustments can cause the detection system's accuracy to gradually decline. This paper utilizes the continuously updated multispectral image data, historical data streams, and fault detection results stored in the sixth state repository to construct a feedback system. By cyclically executing the optimization process, a lightweight two-branch neural network and structural equation model (SEM) can adapt to the dynamic changes in the conveyor belt's operating state, improving the detection accuracy of deviations and related faults, reducing the probability of false detections and missed detections, and extending the system's effective service life.

[0125] Specifically, causal-aware active learning is as follows:

[0126] The absolute value of the path coefficient in the screening structural equation model (SEM) is greater than the preset screening threshold The samples constitute a high-value training set; among them, the preset screening threshold Determined based on a large amount of historical data and actual engineering experience. and the second path coefficient ) whose absolute value is greater than the screening threshold , indicating that the causal relationship between conveyor belt deviation and temperature anomaly in the corresponding sample is significant, and contains more effective information that helps model learning. For example, in the detection data of a certain period of time, the path coefficients of multiple samples are obtained through structural equation model SEM analysis. If If the absolute value of the path coefficient is set to 0.3, samples with a path coefficient greater than 0.3 are selected. These samples reflect typical operating conditions with strong causal relationships, such as "deviation causing temperature rise" or "temperature rise exacerbating deviation," and constitute a high-value training set. This screening method avoids the interference of large amounts of redundant data on model training, focusing on samples that are helpful for model optimization, thereby improving training efficiency and quality.

[0127] Fine-tune the parameters of a lightweight two-branch neural network based on a high-value training set. Specifically, the fine-tuning process can use a backpropagation algorithm to update the weights and bias parameters in the network with the goal of minimizing the loss function. The loss function can be the mean squared error (MSE) or cross-entropy loss function, which measures the difference between the network's predictions and the actual labels.

[0128] For example, in the edge offset branch, the network performs forward propagation based on the conveyor belt edge offset data in the high-value training set to calculate the predicted offset heatmap. A loss function is then used to calculate the error between the predicted value and the true offset. The backpropagation algorithm then propagates the error from the output layer back through each layer of the network. The weights and bias parameters of each layer are adjusted based on the error gradient, enabling the network to more accurately output conveyor belt edge offsets in subsequent predictions. The parameter fine-tuning process for the temperature anomaly branch is similar, with continuous optimization of network parameters improving the accuracy of locating temperature anomalies.

[0129] After fine-tuning its parameters, the lightweight dual-branch neural network can better adapt to new features and changing patterns that emerge during conveyor belt operation. For example, when a conveyor belt experiences localized wear and tear due to long-term use, resulting in subtle changes in edge features, the fine-tuned network can more accurately capture these changes, improving the accuracy of edge offset detection. For temperature anomaly detection, it can more sensitively identify temperature anomalies caused by equipment aging, sudden load changes, and other factors, enhancing the system's adaptability to complex operating conditions.

[0130] The partial least squares algorithm is used to iteratively update the path coefficients and intercept terms of the structural equation model (SEM). During the update process, the offset, temperature and other variables in the high-value training set are used as inputs, and the causal relationship determination results are used as the goal. The path coefficients in the structural equation model (SEM) are continuously adjusted through the PLS algorithm. 、 and the intercept term 、 .

[0131] The updated Structural Equation Model (SEM) can more accurately analyze the causal relationship between conveyor belt deviation and temperature anomalies, providing a more reliable basis for potential failure risk assessment. For example, when changes in conveyor belt operating conditions, such as increased load or fluctuating ambient temperature, cause the causal relationship between deviation and temperature rise to shift, the updated SEM can promptly capture these changes, making the consideration of causal relationships in subsequent risk assessments more accurate, thereby improving the timeliness and accuracy of fault warnings.

[0132] Through the above-mentioned fine-tuning of the parameters of the lightweight two-branch neural network and the updating of the coefficients of the structural equation model (SEM), the seventh optimized parameter set is generated. Among them, the seventh optimized parameter set includes the weights and bias parameters of the optimized lightweight two-branch neural network and the path coefficients, intercept terms and other key parameters of the structural equation model (SEM). In the subsequent conveyor belt deviation detection process, the parameters in the seventh optimized parameter set will be used for calculation and analysis. Specifically, the lightweight two-branch neural network extracts features from the newly collected multispectral image data based on the optimized parameters, and the structural equation model (SEM) uses the updated coefficients to analyze the causal relationship between deviation and temperature anomalies, thereby achieving more accurate fault detection and risk assessment. At the same time, as new data continues to accumulate, the closed-loop optimization process will continue to run and continuously update the optimization parameter set.

[0133] Corresponding to the method for detecting deviation of a coal conveyor belt in the above embodiment, the present invention further provides a system for detecting deviation of a coal conveyor belt.

[0134] like Figure 5-Figure 7 As shown, a coal conveyor belt deviation detection system according to an embodiment of the present invention includes a first generation module, a second generation module, a third generation module, a fourth generation module, an output module, a fifth generation module, a sixth generation module, a trigger and update module and a seventh generation module.

[0135] Among them, the first generation module uses a multispectral imaging device to simultaneously collect multispectral image data of the conveyor belt when it is running. The multispectral image data includes visible light image data and infrared thermal imaging image data. The collected original visible light image data and infrared thermal imaging image data are preprocessed, including denoising filtering and background correction operations to remove noise caused by environmental interference sources, and physical interference simulation sample data is injected to enhance robustness, generating a first processed image set that is resistant to environmental interference.

[0136] The second generation module applies an image fusion algorithm to the first processed image set, performs spatial registration and feature extraction on the visible light image data and infrared thermal imaging image data, integrates the conveyor belt edge position information and temperature distribution information, and generates a second fused image data set.

[0137] The third generation module extracts the offset characteristics of the conveyor belt edge position and the distribution characteristics of the temperature anomaly points from the second fused image data set through a lightweight dual-branch neural network, comprehensively analyzes the feature performance under the current operating state, and generates a first feature parameter set.

[0138] The fourth generation module obtains the edge offset records and temperature change records of the conveyor belt operation in the past period from the historical data stream, compares them with the first feature parameter set in time series, analyzes the feature change trend, and generates a second associated feature data set.

[0139] The output module uses the structural equation model (SEM) based on the time window to analyze the dynamic causal relationship between edge deviation and temperature anomaly points based on the change trend in the second correlation feature data set, and outputs a third dynamic correlation map with causal direction markings.

[0140] If a significant causal relationship is detected between the temperature anomaly point and the edge deviation in the third dynamic correlation map, the fifth generation module comprehensively evaluates the potential fault risk level R based on the preset mechanical friction heat threshold and the deviation tolerance threshold to generate a fourth risk assessment indicator.

[0141] The sixth generation module maps the spatial coupling relationship between the temperature anomaly point and the edge deviation position in the physical coordinate system of the conveyor belt based on the fourth risk assessment indicator, and generates fifth risk distribution mapping data for subsequent judgment of warning conditions.

[0142] If the fifth risk distribution mapping data meets the preset deviation warning threshold, temperature anomaly threshold and causal risk threshold, the trigger and update module triggers a graded alarm signal, records the multispectral image data and historical data stream in the current state, and updates the sixth state repository.

[0143] The seventh generation module cyclically executes the closed-loop optimization process at fixed time intervals for the updated data in the sixth state repository, screens high-value templates through causal-aware active learning, updates the lightweight two-branch neural network parameters and structural equation model (SEM) coefficients, and generates the seventh optimization parameter set to improve the accuracy of subsequent fault detection.

[0144] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications and improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for detecting deviation of a coal conveyor belt, characterized in that: The following steps are involved: S1: using a multispectral imaging device to simultaneously collect multispectral image data of the conveyor belt during operation, the multispectral image data including visible light image data and infrared thermal imaging image data, preprocessing the collected original visible light image data and infrared thermal imaging image data, including denoising filtering and background correction operations to remove noise caused by environmental interference sources, and injecting physical interference simulation sample data to enhance robustness, thereby generating a first processed image set that is resistant to environmental interference; S2: For the first processed image set, using an image fusion algorithm to perform spatial registration and feature extraction on the visible light image data and the infrared thermal imaging image data, integrating the conveyor belt edge position information and temperature distribution information, and generating a second fused image data set; S3: Extracting the offset characteristics of the conveyor belt edge position and the distribution characteristics of the temperature anomaly points from the second fused image dataset using a lightweight dual-branch neural network, comprehensively analyzing the characteristic performance under the current operating state, and generating a first feature parameter set; S4: Obtain edge offset records and temperature change records of the conveyor belt during a past period of time from the historical data stream, perform time series comparison with the first feature parameter set, analyze feature change trends, and generate a second associated feature data set; S5: Analyze the dynamic causal relationship between edge deviation and temperature anomaly points using a time window-based structural equation model (SEM) based on the change trend in the second correlation feature data set, and output a third dynamic correlation map with causal direction markings; S6: If a significant causal relationship is detected between the temperature anomaly point and the edge deviation in the third dynamic correlation map, a potential fault risk level R is comprehensively evaluated in combination with a preset mechanical friction heat threshold and a deviation tolerance threshold to generate a fourth risk assessment indicator; S7: Mapping the spatial coupling relationship between the temperature anomaly point and the edge deviation position in the conveyor belt physical coordinate system based on the fourth risk assessment indicator to generate fifth risk distribution mapping data for subsequent determination of warning conditions; S8: If the fifth risk distribution mapping data meets the preset deviation warning threshold, temperature anomaly threshold, and causal risk threshold, a graded alarm signal is triggered, the multispectral image data and historical data stream in the current state are recorded, and a sixth state storage repository is updated; S9: For the updated data in the sixth state repository, the closed-loop optimization process is cyclically executed at fixed time intervals, high-value training sets are screened through causal-aware active learning, the lightweight two-branch neural network parameters and the structural equation model SEM coefficients are updated, and the seventh optimization parameter set is generated.

2. The method for detecting deviation of a coal conveyor belt according to claim 1, characterized in that: The pre-processing in S1 includes: performing adaptive histogram equalization and bilateral filtering noise reduction operations on the visible light image data; performing non-uniformity correction and temperature drift compensation operations on the infrared thermal imaging image data; The physical interference simulation sample data includes dust occlusion data generated by a grayscale mask and heat wave distortion data generated by a random deformation field.

3. The method for detecting deviation of a coal conveyor belt according to claim 2, characterized in that: In S2, spatial registration realizes pixel-level alignment of visible light image and infrared thermal imaging image through scale-invariant feature transformation matching and random sampling consistency algorithm; feature extraction is performed by extracting visible light edge gradient map and infrared temperature saliency map, and performing adaptive weighted fusion operation according to preset weight ratio.

4. The method for detecting deviation of a coal conveyor belt according to claim 3, characterized in that: In S3, the lightweight dual-branch neural network includes: The edge offset branch uses a lightweight convolutional network to output the offset heat map and calculate the maximum offset; The temperature anomaly branch uses a network with a channel attention mechanism to locate the coordinates of temperature anomaly points and temperature extremes.

5. The method for detecting deviation of a coal conveyor belt according to claim 4, characterized in that: In S5, the dynamic causal relationship is determined by the following rules: When the absolute value of the first path coefficient of the structural equation model (SEM) is greater than the preset coefficient threshold and the statistical significance is lower than the preset significance level, the causal direction of the deviation causing the temperature rise is marked; When the absolute value of the second path coefficient is greater than the preset coefficient threshold and the statistical significance is lower than the preset significance level, the causal direction of the temperature rise aggravating the deviation is marked.

6. The method for detecting deviation of a coal conveyor belt according to claim 5, characterized in that: In S6, the calculation formula for comprehensively evaluating the potential failure risk level R is: Where, Indicates the conveyor belt edge offset detected in the current frame; Indicates the preset deviation tolerance threshold; represents the first weight factor; Indicates the temperature rise amplitude of the abnormal temperature point; Indicates the preset mechanical friction heat threshold; represents the second weight factor; Indicates the causal type assignment item. When the causal direction is deviation causing temperature rise, ; When the causal direction is that the temperature rise aggravates the deviation, ; represents the third weight factor, .

7. The method for detecting deviation of a coal conveyor belt according to claim 6, characterized in that: In S7, the conveyor belt physical coordinate system establishes a longitudinal coordinate axis and a transverse coordinate axis with the driving roller as the origin, maps the coordinates of the temperature anomaly point to the conveyor belt physical coordinate system, and performs a spatial superposition operation with the deviation displacement vector field.

8. The method for detecting deviation of a coal conveyor belt according to claim 7, characterized in that: In S8, the hierarchical alarm signal triggering rules include: When the fault risk level R reaches the first preset risk threshold and the offset exceeds the preset displacement limit, a level 1 alarm is triggered; A level 2 alarm is triggered when the fault risk level R reaches a second preset risk threshold and the temperature extreme value exceeds a preset temperature baseline limit.

9. The method for detecting deviation of a coal conveyor belt according to claim 8, characterized in that: In S9, the active learning of causal perception is specifically as follows: Screening the samples whose absolute values ​​of path coefficients in the structural equation model (SEM) are greater than the preset screening threshold to form a high-value training set; Fine-tuning parameters of a lightweight two-branch neural network based on the high-value training set; The partial least squares algorithm was used to iteratively update the path coefficients and intercept terms of the structural equation model (SEM).

10. A coal conveyor belt deviation detection system, characterized in that: include: A first generation module is configured to simultaneously collect multispectral image data of the conveyor belt during operation using a multispectral imaging device, the multispectral image data including visible light image data and infrared thermal imaging image data, preprocess the collected original visible light image data and infrared thermal imaging image data, including denoising filtering and background correction operations to remove noise caused by environmental interference sources, and inject physical interference simulation sample data to enhance robustness, thereby generating a first processed image set that is resistant to environmental interference; a second generation module, applying an image fusion algorithm to the first processed image set, performing spatial registration and feature extraction on the visible light image data and the infrared thermal imaging image data, integrating the conveyor belt edge position information and temperature distribution information, and generating a second fused image data set; A third generation module extracts the offset characteristics of the conveyor belt edge position and the distribution characteristics of the temperature anomaly points from the second fused image data set through a lightweight two-branch neural network, comprehensively analyzes the characteristic performance under the current operating state, and generates a first feature parameter set; A fourth generation module obtains edge offset records and temperature change records of the conveyor belt operation over a period of time from the historical data stream, compares them with the first feature parameter set in time series, analyzes feature change trends, and generates a second associated feature data set; an output module, which uses a time-window-based structural equation model (SEM) to analyze the dynamic causal relationship between edge deviation and temperature anomaly points based on the change trend in the second correlation feature data set, and outputs a third dynamic correlation map with causal direction markings; a fifth generating module, which, if a significant causal relationship is detected between the temperature anomaly point and the edge deviation in the third dynamic correlation map, comprehensively evaluates the potential fault risk level R by combining a preset mechanical friction heat threshold and a deviation tolerance threshold to generate a fourth risk assessment indicator; a sixth generation module, mapping the spatial coupling relationship between the temperature anomaly point and the edge deviation position in the physical coordinate system of the conveyor belt based on the fourth risk assessment indicator, and generating fifth risk distribution mapping data for subsequent determination of warning conditions; a triggering and updating module, which triggers a graded alarm signal if the fifth risk distribution mapping data satisfies a preset deviation warning threshold, temperature anomaly threshold, and causal risk threshold, simultaneously records the multispectral image data and historical data stream in the current state, and updates a sixth state repository; The seventh generation module cyclically executes the closed-loop optimization process at fixed time intervals for the updated data in the sixth state repository, screens high-value templates through causal-aware active learning, updates the lightweight two-branch neural network parameters and structural equation model (SEM) coefficients, and generates a seventh optimization parameter set for improving the accuracy of subsequent fault detection.

Images