Intelligent monitoring system for coal conveying trestle belt

Abstract

The application discloses an intelligent monitoring system of a coal conveying trestle belt, comprising: a data acquisition module, which is used for acquiring a coal conveying belt image, measuring coal dust thickness distribution data on the surface of the belt in real time, obtaining a time sequence video stream of the coal conveying belt and a coal dust thickness matrix; a dynamic reference module, which is used for extracting a static belt structure according to the time sequence video stream, obtaining an initial reference feature map as a belt reference image without coal dust interference; a coal dust sensing and coding module, which is used for inhibiting high-shading area feature response according to the coal dust thickness matrix and coding a belt image of a current frame to obtain a current frame feature map; and a dynamic difference analysis module, which is used for performing difference analysis according to the current frame feature map and the belt reference image to obtain an abnormal area; the application can realize high-definition imaging of the surface of the belt and stable extraction of abnormal features under a complex environment of dynamic shading of coal dust and nonlinear attenuation of optical signals.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology, and more specifically to an intelligent monitoring system for a coal conveyor belt. Background Technology

[0002] Currently, safety monitoring systems for coal conveyor belts widely employ high-definition video surveillance technology. These systems deploy industrial cameras to capture real-time images of the belt's operation and utilize image processing algorithms (such as edge detection and template matching) to identify anomalies like belt tears and misalignment. Such solutions can meet basic detection needs in typical industrial scenarios. For example, in low-dust environments, cameras can clearly capture surface cracks or edge misalignments on the belt, and algorithms determine the location and severity of the anomaly through pixel-level analysis, thereby triggering alarm signals.

[0003] However, in the actual operating environment of coal conveyor trestle, the conveyor belt is exposed to high concentrations of coal dust (normally >50mg / m³) for extended periods. 3 In this environment, coal dust particles form a non-uniform adhesion layer on the camera lens and conveyor belt surface. This adhesion layer causes two major problems: the scattering effect of coal dust particles on visible and near-infrared light (the operating wavelength of conventional cameras) increases exponentially with the increase of the accumulation thickness; the image contrast of key features such as cracks and edge deformation on the conveyor belt surface will decrease significantly; traditional image enhancement algorithms (such as histogram equalization) can only linearly stretch the grayscale range and cannot recover the details obscured by coal dust. At the same time, during the coal transportation process, the vibration of the conveyor belt and the impact of the coal flow cause the coal dust adhesion layer to be constantly peeled off and re-accumulated. In the video stream captured by the camera, the texture of the conveyor belt surface and the distribution of coal dust are constantly changing dynamically. Traditional matching algorithms based on fixed templates (such as SSIM structural similarity comparison) lack dynamic benchmark references, resulting in an extremely high misjudgment rate.

[0004] Therefore, how to achieve high-definition imaging of the conveyor belt surface and stable extraction of abnormal features under complex environments such as dynamic coal dust shading and nonlinear attenuation of optical signals is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides an intelligent monitoring system for coal conveyor belts, which can achieve high-definition imaging of the belt surface and stable extraction of abnormal features under complex environments such as dynamic coal dust shading and nonlinear attenuation of optical signals.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] An intelligent monitoring system for a coal conveyor belt trestle includes:

[0008] The data acquisition module is used to acquire images of the coal conveyor belt and measure the coal dust thickness distribution data on the belt surface in real time, so as to obtain the time-series video stream of the coal conveyor belt and the coal dust thickness matrix.

[0009] The dynamic reference module is used to extract the static belt structure based on the time-series video stream and obtain an initial reference feature map as a belt reference image without coal dust interference.

[0010] The coal dust sensing and coding module is used to suppress the feature response of the high-obscurity area according to the coal dust thickness matrix and encode the belt image of the current frame to obtain the feature map of the current frame;

[0011] The dynamic difference analysis module is used to perform difference analysis based on the current frame feature map and the belt reference image to obtain abnormal regions.

[0012] Preferably, it also includes an anomaly analysis module, which is used to extract spatiotemporal features based on the anomaly region and analyze the spatiotemporal evolution feature maps corresponding to different belt anomalies.

[0013] Preferably, the anomaly analysis includes:

[0014] Stack the binary images of anomaly regions from multiple consecutive frames into a three-dimensional tensor according to the time series.

[0015] By extracting anomalous evolution features at different time scales through parallel convolutional branches and performing pyramid pooling on the spatial dimension, local details and global morphological features are fused to obtain the final spatiotemporal features.

[0016] The correlation between abnormal regions in adjacent frames is analyzed by a self-attention mechanism to obtain a confidence matrix, and the spatiotemporal features are weighted accordingly.

[0017] Based on the weighted spatiotemporal features, a pre-defined anomaly pattern library is matched, and anomaly type labels and confidence scores are output.

[0018] Preferably, when the dynamic reference module separates the static conveyor belt structure from the dynamic coal dust interference through background modeling, it adopts an adaptive update strategy based on coal dust distribution: the coal dust-covered area is dynamically masked, and the background model is updated only for areas without coal dust or with low shading, thus preserving the integrity of the conveyor belt structure.

[0019] Preferably, the adaptive update strategy of the dynamic benchmark module includes the following steps:

[0020] A dynamic mask is generated based on the coal dust thickness matrix. Areas where the coal dust thickness exceeds a preset threshold are marked as high-occlusion areas, so that only the pixels in the non-masked low-occlusion areas are updated with background samples.

[0021] The geometric features of the static belt structure are periodically calibrated by combining belt load variation data.

[0022] Preferably, the adaptive update strategy of the dynamic reference module further includes: tracking the coal dust drift trajectory in the high-shading area using optical flow method; if the coal dust movement speed exceeds a preset threshold, then removing the dynamic interference signal of that area from the reference feature map.

[0023] Preferably, in the coal dust sensing and coding module, a dynamic weight map is generated through the coal dust thickness matrix to locally suppress the feature response of the high-occlusion area, while interpolating and compensating the features of the adjacent low-occlusion area to ensure the continuity of the belt texture.

[0024] Preferably, the feature suppression and compensation steps of the coal dust sensing coding module include:

[0025] The coal dust thickness matrix is ​​input into the convolutional layer to generate a dynamic weight map, and the weight values ​​are inversely proportional to the coal dust thickness.

[0026] The feature map output by the encoder is weighted to suppress the feature response in highly occluded regions;

[0027] Local interpolation is performed on the adjacent features of the suppressed region, and the texture continuity is restored by using the feature mean of the adjacent low-occlusion region.

[0028] By using skip connections, the original feature map is fused with the compensated feature map, preserving global contextual information.

[0029] Preferably, in the dynamic difference analysis module, the calculation of the difference matrix incorporates the coal dust thickness weight, dynamically attenuating the difference values ​​in highly obscured areas to avoid misjudgments caused by coal dust drift.

[0030] Preferably, the decay of the difference value in the dynamic difference analysis module is achieved through the following steps:

[0031] The coal dust thickness matrix is ​​normalized into an attenuation coefficient matrix; the greater the thickness, the higher the attenuation coefficient.

[0032] Attenuate the difference matrix pixel by pixel: Difference value = Original difference value × (1 - Attenuation coefficient);

[0033] Morphological filtering is applied to the attenuated difference matrix to preserve the continuous structural features of the belt edge and cracks;

[0034] Isolated noise points are removed by connected component analysis, and a binary map of the abnormal region is output.

[0035] As can be seen from the above technical solution, compared with the prior art, this invention discloses an intelligent monitoring system for coal conveyor belts. The dynamic reference module separates the static belt structure from the time-series video stream and generates a reference image free from coal dust interference. By adaptively filtering out the instantaneous interference of dynamic coal dust, this module ensures the stable expression of key geometric features such as belt edges and supports, avoiding blurring or deformation of the reference image caused by coal dust movement. Even in areas covered by high concentrations of coal dust, the spatial topological relationship of the belt's basic structure is accurately preserved, providing a reliable reference for subsequent difference analysis. The coal dust perception coding module dynamically adjusts the feature response intensity based on the coal dust thickness matrix: it suppresses features in highly occluded areas (such as coal dust accumulation areas) to block the propagation of coal dust noise into the deep network; at the same time, it enhances the texture details (such as cracks and foreign objects) in low-occlusion areas through the coding process. This mechanism achieves preliminary separation of noise and effective signals in the feature extraction stage, so that the current frame feature map contains both local details under coal dust interference and maintains the semantic consistency of the overall belt structure. The dynamic difference analysis module directly focuses on the deviation of the conveyor belt surface from the static reference by comparing the current frame feature map with the reference image. Since the reference image eliminates dynamic interference from coal dust, and the current frame feature map suppresses noise in highly obscured areas through encoding, the difference matrix accurately reflects the difference in optical attenuation characteristics between real anomalies (such as tears or foreign objects) and coal dust-obscured areas. This module distinguishes between dynamic interference (instantaneous differences caused by coal dust movement) and static anomalies (persistent structural defects) from a mechanistic perspective, thus achieving robust detection of anomaly areas under complex optical attenuation environments. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0037] Figure 1 The attached figure is a schematic diagram of the intelligent monitoring system for a coal conveyor belt provided by the present invention.

[0038] Figure 2 This is a schematic diagram of the data processing flow of the dynamic benchmark module in an embodiment of the present invention.

[0039] Figure 3 This is a schematic diagram of the coal dust sensing and coding process in an embodiment of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] like Figure 1 This invention discloses an intelligent monitoring system for a coal conveyor belt, comprising:

[0042] The data acquisition module is used to collect real-time measurement data of coal dust thickness distribution on the surface of the coal conveyor belt, and obtain the time-series video stream and coal dust thickness matrix of the coal conveyor belt.

[0043] The dynamic reference module is used to extract the static belt structure from the time-series video stream and obtain an initial reference feature map as a belt reference image free from coal dust interference.

[0044] The coal dust sensing and coding module is used to suppress the feature response of high-obscurity areas based on the coal dust thickness matrix and encode the belt image of the current frame to obtain the feature map of the current frame.

[0045] The dynamic difference analysis module is used to perform difference analysis based on the current frame feature map and the belt reference image to obtain abnormal regions.

[0046] In this embodiment, the data acquisition module combines a polarized light-compensated camera with a laser scattering sensor to simultaneously acquire a high-definition video stream and a coal dust thickness matrix in a dusty environment. Near-infrared polarized light penetrates the coal dust layer and suppresses scattering noise, while laser scattering directly quantifies the degree of occlusion. The two complement each other to overcome the image blurring and feature loss caused by dust occlusion in traditional single-camera systems. The dynamic benchmark module extracts the static conveyor belt structure from the video stream to generate a benchmark image. It separates the conveyor belt body from the dynamic coal dust through background modeling and uses optical flow to track the movement trajectory of the coal dust to dynamically eliminate interference areas. At the same time, it introduces load data to calibrate the conveyor belt deformation, ensuring that the benchmark always reflects the real structure and avoiding misjudgments caused by coal dust movement or load changes in traditional static benchmarks. The coal dust perception encoding module generates dynamic weights based on the coal dust thickness matrix, suppresses the feature response of highly occluded areas, and uses neighborhood interpolation to compensate for the texture of the occluded conveyor belt. This eliminates coal dust movement noise while retaining key features such as cracks and edges, solving the problem of missed detections caused by the lack of features in occluded areas in traditional algorithms. The dynamic difference analysis module introduces an attenuation coefficient driven by coal dust thickness when calculating inter-frame differences, reducing the reliability of differences in highly obscured areas. It also removes isolated noise through morphological filtering and focuses on continuous abnormal areas (such as banded deviations and radial cracks), making the detection results both resistant to transient interference and maintaining high-precision positioning.

[0047] To further implement the above technical solution, the anomaly analysis module is used to extract spatiotemporal features based on the anomaly region and analyze the spatiotemporal evolution feature maps corresponding to different belt anomalies. The steps include: stacking the binary images of the anomaly region in multiple consecutive frames into a three-dimensional tensor according to the time series; extracting the anomaly evolution features at different time scales through parallel convolution branches, and performing pyramid pooling on the spatial dimension to fuse local details and global morphological features to obtain the final spatiotemporal features; analyzing the correlation between anomaly regions in adjacent frames through a self-attention mechanism to obtain a confidence matrix, and weighting the spatiotemporal features; matching the weighted spatiotemporal features with a preset anomaly pattern library, and outputting anomaly type labels and confidence scores.

[0048] Specifically, firstly, the binary images of the abnormal region from multiple consecutive frames are stacked into a three-dimensional tensor according to the time series, constructing a spatiotemporal data carrier. This tensor retains the dynamic change information of the abnormal region in the time dimension (such as deviation trend and crack propagation direction), and maps the location and morphological distribution of the abnormal region in the spatial dimension (such as band-like or radial patterns). Next, the abnormal evolution features at different time scales are extracted through parallel convolutional branches—the short-time branch captures inter-frame abrupt changes (such as the instantaneous appearance of foreign objects), the medium-time branch models trend-based shifts (such as the gradual spread of deviation), and the long-time branch analyzes cumulative effects (such as the slow extension of cracks). At the same time, pyramid pooling is performed on the spatial dimension to fuse local details (crack edge sharpness) and global morphology (overall deformation trend of the belt), ensuring that the feature representation takes into account both microscopic and macroscopic information.

[0049] Furthermore, a self-attention mechanism is used to enhance temporal correlation: a query vector, a key vector, and a value vector are generated for each anomalous region in the 3D tensor, and the similarity weights of anomalous regions at the same location in adjacent frames are calculated. If a region is continuously activated across multiple consecutive frames (e.g., ≥5 frames), its confidence weight is increased; if it only appears in a single frame, it is judged as transient noise and its weight is reduced. This mechanism effectively distinguishes between genuine anomalies (continuous evolution) and transient interference (random noise).

[0050] Finally, the weighted spatiotemporal features are matched against a pre-defined anomaly pattern library: the deviation pattern corresponds to continuous linear offset in the edge region, the tearing pattern matches the radial diffusion trajectory of cracks, and the foreign object intrusion pattern identifies random jumps in isolated areas. Based on the matching results, anomaly type labels and confidence scores are output, triggering tiered alarm commands. This process, through spatiotemporal modeling and physically driven pattern matching, solves the problems of traditional threshold methods failing to distinguish anomaly types and being susceptible to transient interference, thus improving the interpretability and reliability of detection.

[0051] The method for constructing the exception mode library is as follows:

[0052] Misalignment mode: Extract the progressive offset features of the belt edge over multiple consecutive frames. Spatially, it exhibits a band-like diffusion, and temporally, the offset increases linearly.

[0053] Tear mode: Detects the extension direction and speed of cracks on the belt surface. Spatially, the cracks are radially distributed, and over time, the crack length increases exponentially.

[0054] Foreign object intrusion mode: Identifies isolated abnormal areas outside the belt structure, with abrupt spatial outlines and random jumps in regional location over time.

[0055] like Figure 2 To further implement the above technical solution, the dynamic benchmark module adopts an adaptive update strategy based on coal dust distribution when separating the static belt structure from the dynamic coal dust interference through background modeling: the coal dust-covered area is dynamically masked, and the background model is updated only for areas without coal dust or with low shading, thus preserving the integrity of the belt structure.

[0056] The adaptive update strategy of the dynamic reference module includes the following steps: generating a dynamic mask based on the coal dust thickness matrix, marking areas where the coal dust thickness exceeds a preset threshold as high-occlusion areas, so that only the pixels in the low-occlusion areas without the mask are updated with background samples; tracking the coal dust drift trajectory in the high-occlusion areas using optical flow method, and removing the dynamic interference signal in the area if the coal dust movement speed exceeds a preset threshold; and periodically calibrating the geometric features of the static belt structure in combination with belt load change data.

[0057] For example, the laser scanner generates a coal dust distribution matrix on the conveyor belt surface in real time. It detects that the coal dust accumulation on the right half of the belt reaches a thickness of 5-8 mm (a preset threshold of 3 mm), while the coal dust on the left half is thinner (0-2 mm). At this point, the system automatically marks the right half as a "high-occlusion area," generates a dynamic mask to cover this area, and only allows the low-occlusion area on the left half to participate in the background model update. This strategy avoids noise pollution of the coal dust accumulation area and preserves the original features of static structures such as the belt edges and supports.

[0058] In the monitoring footage, coal dust in the high-obscurity area was moving due to airflow. The system detected a clump of coal dust moving upwards and to the right at a speed of 15 pixels per frame (the preset speed threshold is 10 pixels per frame) using optical flow. The system immediately marked the pixel area covered by this motion trajectory as a dynamic interference signal and removed the pixel data from these locations from the feature map of the current frame. Simultaneously, a fixed foreign object (such as metal fragment) was detected on the conveyor belt surface in the low-obscurity area. Since the background model of this area was not contaminated, the system accurately identified the location of the foreign object and triggered an alarm.

[0059] When the belt load increases from 500 tons to 800 tons, the load cells synchronously transmit load data. Based on a preset belt deformation model, the system predicts that the belt's mid-section sag will increase by 12% under the current load, and then dynamically compensates and calibrates the belt's geometric parameters (such as edge curvature and support spacing) in the reference model. During calibration, the system prioritizes using undeformed structural features in low-obscuration areas as reference points to ensure that the corrected background model matches the actual belt shape. This process is executed periodically every 30 seconds to continuously adapt to the slow deformation of the belt caused by load changes.

[0060] In this embodiment, in the coal mine conveyor belt monitoring system, the present invention dynamically adjusts the monitoring system's calibration of the belt's physical shape (such as curvature, edge position, etc.) based on changes in the actual weight of coal carried by the belt. This avoids misjudging belt deformation caused by load changes as abnormal (such as misalignment or breakage). Here are a few specific examples:

[0061] Dynamic compensation for belt sag: When unloaded, the belt is straight, but when fully loaded, the middle section sags due to weight, forming an arc (e.g., sag increases by 15cm). Calibration logic: The load cell detects a current load of 800 tons (preset full load is 1000 tons); the system calculates the expected sag based on the deformation model (e.g., for every 100 tons increase in load, the sag increases by 3cm); the baseline in the middle of the belt is adjusted from a straight line to an arc offset downwards by 9cm; in subsequent detection, the system uses the adjusted arc as a reference to determine whether abnormal sag has occurred.

[0062] Lateral offset correction at the belt edge: When the load increases, the belt may slightly expand to both sides due to gravity (e.g., 2cm on each side). Calibration logic: Load data triggers a calibration cycle. The system obtains the current actual edge position (e.g., the left edge shifts 1.8cm to the right, and the right edge shifts 2.1cm to the left); the "standard edge line" for visual detection is adjusted from a fixed coordinate to a dynamic range (e.g., the allowable position range of the left edge changes from X = 100 ± 1cm to X = 101.8 ± 1cm); if the edge is detected to be outside the corrected range, it is considered misalignment.

[0063] Deformation compensation for belt support spacing: Under excessive load, the belt between two fixed supports may stretch, resulting in a visually larger "support spacing" (e.g., the actual support spacing is 5 meters, but it appears as 5.2 meters in the image after stretching). Calibration logic: Calculate the belt stretching ratio based on load data (e.g., 4% stretching rate under 800 tons load); dynamically correct the original support spacing benchmark value of 5 meters to 5.2 meters; if the spacing is detected to exceed ±5% of the correction value (e.g., >5.46 meters), trigger an alarm for support loosening or breakage.

[0064] like Figure 3To further implement the above technical solution, in the coal dust sensing and coding module, a dynamic weight map is generated through the coal dust thickness matrix to locally suppress the feature response of the high-shading area, while interpolating and compensating the features of the adjacent low-shading area to ensure the continuity of the belt texture.

[0065] The feature suppression and compensation steps of the coal dust sensing coding module include: inputting the coal dust thickness matrix into the convolutional layer to generate a dynamic weight map, with the weight value inversely proportional to the coal dust thickness; weighting the feature map output by the encoder to suppress the feature response of high-occlusion areas; performing local interpolation on the adjacent features of the suppressed areas and restoring texture continuity using the feature mean of adjacent low-occlusion areas; and fusing the original feature map and the compensated feature map through skip connections to preserve global context information.

[0066] For example,

[0067] The system first uses a laser scanner to obtain the coal dust thickness distribution matrix on the conveyor belt surface. For example, the coal dust accumulation thickness in the middle section of the belt (coordinate range X = 100-200, Y = 50-150) reaches 5-8 mm, while the thickness in the two edge areas is only 0.2-1 mm. This thickness data is input into a single-channel 1×1 convolutional layer. The convolution kernel parameters are designed as negative correlation functions; for example, the weights are calculated using the formula... (where d is the coal dust thickness), and the weight values ​​are compressed to between 0 and 1 using the Sigmoid function. For areas with a thickness of 8 mm, the weight value is reduced to about 0.15, while the weight value for areas with a thickness of 0.5 mm remains at 0.89. Finally, a dynamic weight map inversely proportional to the coal dust thickness is generated, in which high-shading areas are dark (low weight) and low-shading areas are light (high weight).

[0068] Next, in the original feature map extracted by the encoder (such as ResNet-34), the texture features in the middle section of the conveyor belt are blurred due to coal dust coverage. For example, the activation value of a key feature channel representing the belt mesh is 0.9 in the normal region, but is suppressed to 0.3 in the middle section. The system multiplies the dynamic weight map with the original feature map channel by channel, further reducing the activation value in the highly occluded middle section—for example, when the weight of a pixel is 0.2, the feature values ​​of all channels at that point are scaled to one-fifth of their original values, and the aforementioned key feature value drops from 0.3 to 0.06. This operation significantly weakens the noise signal in the coal dust interference area, but also causes local loss of texture information.

[0069] To address this issue, the system employs local interpolation compensation for each suppressed pixel (e.g., regions with a weight less than 0.3): A 5×5 neighborhood window is defined centered on a highly occluded point (X=120, Y=80), and pixels with a weight greater than 0.7 (e.g., the eight adjacent low-occluded points) are selected. The mean feature value of these points is then calculated. For example, if the activation value of this highly occluded point in a certain channel was originally suppressed to 0.06, by taking the mean of the surrounding eight points (assuming their feature values ​​are 0.55–0.65), the feature value of this point is restored to 0.58 after interpolation, approaching the 0.65 of the normal region. This process allows for the reconstruction of a coherent belt mesh texture in the middle section covered by coal dust; for example, the originally broken longitudinal stripes are aligned with the unoccluded areas on both sides after interpolation.

[0070] Finally, the system fuses the compensated feature map with the original encoder output via skip connections. Specifically, the compensated feature map (256×256×64 channels) and the original feature map (also 256×256×64 channels) are concatenated along the channel dimension to form a fused feature map of 256×256×128, which is then compressed back to 64 channels via a 1×1 convolution. During this process, the convolutional kernel automatically learns the weight allocation for the two types of features—for example, in low-occlusion areas, the original features retain 70% of the weight to preserve details (such as a high response value of 0.88 for a 0.5 mm crack on the right edge), while in high-occlusion areas, the compensated features occupy 85% of the weight to enhance the repaired structure (such as a response value of 0.72 for a longitudinal tear in the middle section). The fused feature map retains both the subtle anomalies unaffected by coal dust and repairs the continuous texture in occluded areas. When finally transmitted to the downstream detection module, it can successfully identify the longitudinal tear in the middle section covered by coal dust (feature response exceeding the threshold of 0.6) and the undisturbed transverse cracks on both sides.

[0071] To further implement the above technical solution, the calculation of the difference matrix incorporates coal dust thickness weights to dynamically attenuate the difference values ​​in highly obscured areas, avoiding misjudgments caused by coal dust drift. The difference value attenuation is achieved through the following steps: normalizing the coal dust thickness matrix into an attenuation coefficient matrix, with a higher attenuation coefficient for larger thicknesses; performing pixel-by-pixel attenuation on the difference matrix: difference value = original difference value × (1 - attenuation coefficient); performing morphological filtering on the attenuated difference matrix to preserve the continuous structural features of belt edges and cracks; and eliminating isolated noise points through connected component analysis to output a binary map of the abnormal area.

[0072] In this embodiment, the system first acquires coal dust thickness distribution data on the current conveyor belt surface using a laser scanner, generating a coal dust thickness matrix. In this matrix, the thickness values ​​in the middle section of the conveyor belt are significantly higher than those in the side edge areas due to thicker coal dust accumulation. To convert the thickness data into attenuation coefficients, the system normalizes the thickness matrix—mapping the maximum thickness value to an attenuation coefficient of 1.0, the minimum thickness value to 0, and calculating other values ​​linearly. For example, if the coal dust thickness in a certain area is 80% of the maximum value, its attenuation coefficient is 0.8. In the normalized attenuation coefficient matrix, high-shading areas exhibit high attenuation values ​​close to 1, while low-shading areas are close to 0.

[0073] Next, the system performs a pixel-level comparison between the current frame image and the background model to generate an original difference matrix. In the original difference matrix, the coal dust drifting area exhibits scattered high difference values ​​due to dynamic interference, while the real crack area presents a continuous high difference band. At this point, the system performs a pixel-by-pixel attenuation operation on the difference matrix: the difference value of each pixel is multiplied by (1 - the attenuation coefficient at the corresponding position). For example, if the original difference value of the high-occlusion area is 0.9, it may be reduced to 0.18 (0.9 × (1 - 0.8)) after attenuation, while the difference value of the low-occlusion area is only slightly reduced from 0.9 to 0.81 (0.9 × (1 - 0.1)). This operation significantly reduces the dynamic interference of coal dust, but may cause fractures in the real crack area (e.g., some pixels are over-attenuated when the crack passes through the high-occlusion area).

[0074] To repair the fractured structure, the system performs a morphological closing operation on the attenuated difference matrix: a 3×3 rectangular kernel is used to dilate the matrix followed by erosion. This operation connects adjacent high-difference regions; for example, a crack divided into two segments by a highly occluded region is filled at the fracture point after the closing operation, forming a continuous region. Subsequently, an opening operation (erosion followed by dilation) is used to eliminate isolated noise points—for example, random noise points have their difference values ​​reduced to zero after erosion and cannot be recovered after dilation, while true cracks are preserved due to their strong continuity.

[0075] Finally, the system performs connected component analysis on the processed difference matrix. All high-difference regions are marked using the 8-neighborhood connectivity rule, and the area and average difference value of each connected region are calculated. An area threshold (e.g., minimum 20 pixels) is set to filter out minor noise, retaining only regions exceeding the threshold as anomalies. For example, sporadic points caused by coal dust are discarded due to insufficient area, while long, narrow connected regions formed by genuine cracks are fully preserved. The output results are displayed as a binary map to mark the anomaly locations, directly guiding on-site personnel in maintenance.

[0076] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0077] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent monitoring system for a coal conveyor belt trestle, characterized in that, include: The data acquisition module is used to acquire images of the coal conveyor belt and measure the coal dust thickness distribution data on the belt surface in real time, so as to obtain the time-series video stream of the coal conveyor belt and the coal dust thickness matrix. The dynamic reference module is used to extract the static belt structure based on the time-series video stream and obtain an initial reference feature map as a belt reference image without coal dust interference. When the dynamic benchmark module separates the static conveyor belt structure from the dynamic coal dust interference through background modeling, it adopts an adaptive update strategy based on coal dust distribution: it performs dynamic masking on the coal dust-covered area and updates the background model only for areas without coal dust or with low shading, thus preserving the integrity of the conveyor belt structure. The coal dust sensing and coding module is used to suppress the feature response of high-occlusion areas according to the coal dust thickness matrix and encode the belt image of the current frame to obtain the feature map of the current frame. In the coal dust sensing and coding module, a dynamic weight map is generated through the coal dust thickness matrix to locally suppress the feature response of high-occlusion areas, while interpolating and compensating the features of adjacent low-occlusion areas to ensure the continuity of belt texture. The dynamic difference analysis module is used to perform difference analysis based on the current frame feature map and the belt reference image to obtain abnormal regions; In the dynamic difference analysis module, the calculation of the difference matrix incorporates coal dust thickness weighting, dynamically attenuating the difference values ​​in highly obscured areas to avoid misjudgments caused by coal dust drift. This difference value attenuation in the dynamic difference analysis module is achieved through the following steps: The coal dust thickness matrix is ​​normalized into an attenuation coefficient matrix; the greater the thickness, the higher the attenuation coefficient. Attenuate the difference matrix pixel by pixel: Difference value = Original difference value × (1 - Attenuation coefficient); Morphological filtering is applied to the attenuated difference matrix to preserve the continuous structural features of the belt edge and cracks; Isolated noise points are removed by connected component analysis, and a binary map of the abnormal region is output.

2. The intelligent monitoring system for a coal conveyor belt trestle according to claim 1, characterized in that, It also includes an anomaly analysis module, which is used to extract spatiotemporal features based on the anomaly region and analyze the spatiotemporal evolution feature maps corresponding to different belt anomalies.

3. The intelligent monitoring system for a coal conveyor belt trestle according to claim 2, characterized in that, The anomaly analysis includes: Stack the binary images of anomaly regions from multiple consecutive frames into a three-dimensional tensor according to the time series. By extracting anomalous evolution features at different time scales through parallel convolutional branches and performing pyramid pooling on the spatial dimension, local details and global morphological features are fused to obtain the final spatiotemporal features. The correlation between abnormal regions in adjacent frames is analyzed by a self-attention mechanism to obtain a confidence matrix, and the spatiotemporal features are weighted accordingly. Based on the weighted spatiotemporal features, a pre-defined anomaly pattern library is matched, and anomaly type labels and confidence scores are output.

4. The intelligent monitoring system for a coal conveyor belt trestle according to claim 1, characterized in that, The adaptive update strategy of the dynamic benchmark module includes the following steps: A dynamic mask is generated based on the coal dust thickness matrix. Areas where the coal dust thickness exceeds a preset threshold are marked as high-occlusion areas, so that only the pixels in the non-masked low-occlusion areas are updated with background samples. The geometric features of the static belt structure are periodically calibrated by combining belt load variation data.

5. The intelligent monitoring system for a coal conveyor belt trestle according to claim 4, characterized in that, The adaptive update strategy of the dynamic reference module also includes: tracking the coal dust drift trajectory in the high-shading area using optical flow method; if the coal dust movement speed exceeds a preset threshold, then removing the dynamic interference signal of that area from the reference feature map.

6. The intelligent monitoring system for a coal conveyor belt trestle according to claim 1, characterized in that, The feature suppression and compensation steps of the coal dust sensing coding module include: The coal dust thickness matrix is ​​input into the convolutional layer to generate a dynamic weight map, and the weight values ​​are inversely proportional to the coal dust thickness. The feature map output by the encoder is weighted to suppress the feature response in highly occluded regions; Local interpolation is performed on the adjacent features of the suppressed region, and the texture continuity is restored by using the feature mean of the adjacent low-occlusion region. By using skip connections, the original feature map is fused with the compensated feature map, preserving global contextual information.

Images