Belt tearing detection method, device and equipment based on visual identification

Abstract

The invention relates to the technical field of visual identification, and discloses a belt tearing detection method, device and equipment based on visual identification, and the method comprises the steps: carrying out the dual-channel image collection and preprocessing of the surface of a conveying belt, and obtaining a multi-channel preprocessing image; extracting a thermal difference feature and a texture structure feature of the multi-channel preprocessed image through a double-flow feature network, and inputting the thermal difference feature and the texture structure feature into a quaternion material deformation analysis model for deformation gradient tensor and invariant parameter calculation to obtain a tear feature description vector; performing hierarchical progressive identification and three-dimensional reconstruction analysis to obtain target tearing feature data; and risk assessment is carried out based on the target tear feature data to obtain a tear grade classification result, the interference of ambient temperature drift and non-uniform illumination is effectively eliminated, the calculation efficiency and accuracy of tear detection are improved, and false alarms and missing alarms are reduced.

Description

Technical Field

[0001] The present invention relates to the field of visual recognition technology, and in particular to a belt tear detection method, device and equipment based on visual recognition. Background Art

[0002] During long-term, high-load operation, conveyor belts are prone to tearing. Failure to detect and address these problems can lead to material leakage and reduced efficiency, or even to the shutdown of the entire production line or even a safety incident. Traditional conveyor belt monitoring methods rely primarily on manual inspections or single-sensor testing, which suffer from low accuracy, high false alarm rates, and poor environmental adaptability. This is particularly true under harsh operating conditions, such as dust pollution, high temperature, and high humidity, making it difficult for traditional detection technologies to achieve stable and reliable monitoring.

[0003] With the advancement of industrial automation and intelligentization, machine vision-based conveyor belt condition monitoring technology has gradually emerged. However, existing visual inspection systems primarily rely on a single information channel, making it difficult to simultaneously capture both the surface texture and internal thermodynamic changes of the conveyor belt. Furthermore, existing methods lack systematic theoretical support for information fusion, feature extraction, and deformation analysis, resulting in significant deficiencies in early detection of tears and prediction of their development trends. Especially for conveyor belts covered with material, surface obstructions can severely impact the detection performance of visible light images, while thermal imaging alone cannot provide accurate structural information. Summary of the Invention

[0004] The present invention provides a belt tear detection method, device and equipment based on visual recognition, which effectively eliminates the interference of ambient temperature drift and uneven lighting, improves the calculation efficiency and accuracy of tear detection, and reduces false alarms and missed alarms.

[0005] In a first aspect, the present invention provides a belt tear detection method based on visual recognition, the belt tear detection method based on visual recognition comprising: Performing dual-channel image acquisition on the conveyor belt surface to obtain original multi-channel image data, and preprocessing the original multi-channel image data to obtain a multi-channel preprocessed image; Extracting thermal difference features and texture structure features of the multi-channel preprocessed image through a dual-stream feature network, and inputting the thermal difference features and texture structure features into a quaternion material deformation analysis model to calculate deformation gradient tensors and invariant parameters to obtain a tear feature description vector; Performing hierarchical progressive recognition and stereo reconstruction analysis based on the tear feature description vector and the multi-channel pre-processed image to obtain target tear feature data; A multi-dimensional morphological feature vector is constructed based on the target tearing feature data, and a risk assessment is performed on the multi-dimensional morphological feature vector to obtain a tearing grade classification result.

[0006] In a second aspect, the present invention provides a belt tear detection device based on visual recognition, the belt tear detection device based on visual recognition comprising: An image acquisition module is used to perform dual-channel image acquisition on the conveyor belt surface to obtain original multi-channel image data, and pre-process the original multi-channel image data to obtain a multi-channel pre-processed image; a calculation module, configured to extract thermal difference features and texture structure features of the multi-channel preprocessed image through a dual-stream feature network, and input the thermal difference features and the texture structure features into a quaternion material deformation analysis model to calculate a deformation gradient tensor and invariant parameters to obtain a tear feature description vector; A reconstruction and analysis module, configured to perform hierarchical progressive recognition and stereo reconstruction analysis based on the tear feature description vector and the multi-channel pre-processed image to obtain target tear feature data; The risk assessment module is used to construct a multi-dimensional morphological feature vector based on the target tearing feature data, and perform risk assessment on the multi-dimensional morphological feature vector to obtain a tearing grade classification result.

[0007] A third aspect of the present invention provides a computer device comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned visual recognition-based belt tear detection method.

[0008] The technical solution provided by the present invention utilizes thermal imaging and visible light dual-channel acquisition technology to achieve comprehensive perception of the conveyor belt's surface temperature distribution and texture structure, overcoming the limitations of a single information source and improving the system's ability to detect tear damage in covered materials. Temperature calibration, illumination equalization, and image registration techniques effectively eliminate interference from ambient temperature drift and uneven lighting, improving image quality and feature extraction accuracy. An improved DenseNet and ResNet-50 dual-stream feature network enables parallel extraction and cross-attention fusion of thermal difference features and texture structure features, enhancing the system's ability to represent tear characteristics. By introducing a quaternion material deformation analysis model and calculating deformation gradient tensors and invariant parameters, the system accurately characterizes the belt's local deformation state and effectively distinguishes between elastic deformation and irreversible tears. A hierarchical recognition strategy combining coarse detection processing with fine segmentation improves the computational efficiency and accuracy of tear detection, reducing false positives and missed detections. Stereoscopic vision and structured light-assisted systems enable three-dimensional reconstruction and dynamic tracking of the tear area, capturing comprehensive information on tear morphology and development. Based on the morphological and dynamic characteristics of tearing, a comprehensive risk assessment model was constructed, enabling accurate classification of tear levels and prediction of development trends. Based on the tear level classification results, differentiated control instructions were generated through a multi-objective optimization algorithm, balancing the conflict between safety and production efficiency and improving the system's intervention effect on tearing conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0010] Figure 1 Schematic diagram of the steps of a belt tear detection method based on visual recognition in an embodiment of the present invention; Figure 2 Schematic diagram of the structure of a belt tear detection device based on visual recognition in an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0011] Embodiments of the present invention provide a method, device and apparatus for detecting belt tears based on visual recognition. The terms "first", "second", "third", "fourth", etc. (if any) in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way are interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products or apparatus.

[0012] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 In one embodiment of the present invention, a belt tear detection method based on visual recognition includes: Step S1, performing dual-channel image acquisition on the conveyor belt surface to obtain original multi-channel image data, and preprocessing the original multi-channel image data to obtain a multi-channel preprocessed image; It is understandable that the execution subject of the present invention can be a belt tear detection device based on visual recognition, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0013] Specifically, a thermal imaging camera and a visible light camera are mounted above the conveyor belt at a preset angle. Both cameras simultaneously cover the same inspection area and have a sufficient overlapping field of view. The cameras are mechanically arranged based on the conveyor belt width, camera mounting height, and camera viewing angle parameters. A calibration method for the overlapped field of view is used to perform geometric calibration to achieve a unified field of view. This process uses a calibration plate or a specific geometric pattern. By capturing images separately from both cameras and performing coordinate mapping to establish a transformation matrix, the thermal and visible light images are spatially aligned in subsequent steps. Parameters are configured for each imaging device. The thermal imaging camera sets the appropriate integration time, frame rate, and temperature range based on thermal sensitivity requirements and the site temperature range to form the thermal imaging acquisition parameters. The visible light camera sets the shutter time, gain, and frame rate based on the ambient light intensity, along with key parameters such as image resolution and dynamic range to form the visible light acquisition parameters. Furthermore, to synchronize the two-channel images, a unified time synchronization control system is constructed. A trigger signal is generated by a photoelectric encoder mounted on the conveyor belt. This signal simultaneously activates the image acquisition of both cameras and triggers a high-brightness LED flash system. This flash system utilizes pulsed drive technology, delivering light intensities up to 50,000 lux in a very short time. This allows for clear illumination of the conveyor belt surface even in harsh conditions such as heavy dust and flying materials, creating an ideal imaging environment for visible light cameras. Under the coordinated control of the aforementioned hardware system, dual-channel simultaneous imaging of the conveyor belt surface is performed based on pre-set thermal imaging and visible light acquisition parameters, synchronization trigger signals, and LED lighting status. This captures a set of highly temporally and spatially aligned image data—an image pair consisting of a first thermal image and a first visible light image—to form the raw multi-channel image data. This raw multi-channel image data undergoes temperature calibration, illumination equalization, and image registration. The thermal imaging image is temperature calibrated, and each frame is corrected in real time using a blackbody reference source. Thermal noise is suppressed using methods such as time-domain averaging filtering. Light equalization is then performed on the visible light image, automatically adjusting the LED intensity based on the on-site light intensity. An improved multi-scale Retinex algorithm is used to eliminate local overexposure and uneven brightness in shadow areas. A phase cross-correlation algorithm is used to achieve precise registration of the thermal imaging image and the visible light image, resulting in a multi-channel preprocessed image.

[0014] The first thermal image in the raw multi-channel image data undergoes temperature zero-point and gain calibration. This process uses a field-deployed reference blackbody source. By comparing the pixel grayscale values captured by the thermal imaging camera with the known blackbody temperature, a calibration coefficient is calculated for each pixel. This corrects temperature zero-point drift and linearizes the gain, generating temperature calibration data. This temperature calibration data is then fed into a time-domain averaging filter for thermal noise suppression. This filter uses a sliding window approach to weighted average successive frames, effectively eliminating high-frequency thermal fluctuations introduced by the infrared detector itself, and outputs a stable second thermal image. Simultaneously, the first visible light image in the raw multi-channel image data is analyzed using a brightness histogram to extract the overall brightness distribution of the image. The percentage of pixels in each grayscale interval is calculated to determine the current ambient light intensity. Based on this intensity distribution information, the pulse current of the LED flash system is dynamically adjusted, automatically adjusting the intensity of each flash to maintain an appropriate image brightness range under varying lighting conditions, ensuring good contrast and clarity in the captured second visible light image. To eliminate localized uneven illumination caused by factors such as scattered light and dust obstruction, a multi-scale Retinex process is performed on the second visible light image. The Retinex model decomposes the image into illumination and reflection components. Logarithmic domain operations and difference filtering are used to obtain two physically distinct image structures. A three-scale spatial domain filtering mechanism is used to reconstruct the illumination and reflection components. Low-scale layers preserve detailed textures, mid-scale layers address local contrast, and high-scale layers adjust overall brightness. The resulting reconstructed third visible light image is visually balanced and highlights local details. The second and third visible light images are fed into a phase cross-correlation calculation module. This module calculates the phase difference between the two images using a frequency-domain Fourier transform. It then locates extrema to determine the translation, rotation, and scale relationships between the two images, extracting the transformation parameters required for image registration. These registration transformation parameters are then applied to the second thermal image, and a spatial domain resampling interpolation algorithm is used to complete the spatial geometric transformation of the image. This interpolation process uses bilinear or cubic spline interpolation to ensure edge continuity and smoothness, resulting in a third thermal image that is spatially consistent with the visible light image. The third thermal imaging image is mapped to the coordinate system of the third visible light image to form a set of multi-channel pre-processed images with spatial alignment, clear texture and balanced illumination.

[0015] Step S2: extracting thermal difference features and texture structure features of the multi-channel preprocessed image through a dual-stream feature network, and inputting the thermal difference features and texture structure features into a quaternion material deformation analysis model to calculate deformation gradient tensors and invariant parameters to obtain a tear feature description vector; Specifically, the third thermal image from the multi-channel preprocessed image is input to the thermal feature extraction branch, while the third visible light image is simultaneously input to the texture feature extraction branch. Thermal and texture information in the images are processed in parallel along different paths. The thermal feature extraction branch employs an improved densely connected neural network architecture, consisting of a cascade of multiple convolutional modules, with each layer receiving feature maps from all previous layers as input. This allows the network to capture temperature variations on the belt surface at multiple scales, particularly responding to small temperature rises caused by factors such as friction and extrusion, extracting thermal difference features that characterize temperature anomalies. The texture feature extraction branch, on the other hand, utilizes a deep residual network architecture. This network utilizes skip connections to enhance the ability to retain detailed texture, edge contours, and structural distortions on the belt surface. This effectively extracts morphological changes caused by tears, cracks, and fatigue accumulation, and outputs high-dimensional texture structural features. A cross-attention fusion process is then performed on the thermal difference and texture structural features. This module analyzes the spatial correlation and complementarity of the two feature channels to adaptively generate fusion weights, prioritizing regions of thermal anomalies and regions of texture distortion to construct a fused feature map that incorporates both temperature anomalies and texture deformation information. The fused feature map is fed into a quaternion material deformation analysis model. By comparing the spatial deformation states of local regions in images between consecutive frames, the model describes the rotational deformation of the material block corresponding to each pixel on the belt surface in three-dimensional space using quaternions, and analyzes the trend and intensity of this deformation. By establishing a correspondence between image space and physical space, image changes are mapped to material deformation behavior, and local deformation parameters for each region on the belt surface are calculated. These deformation parameters are decomposed to extract key numerical features describing the principal directions and deformation intensity. These features effectively indicate whether a local region is in a normal elastic state or has entered an irreversible failure phase. Based on this combined information about principal deformation directions and intensity, a tear detection criterion is constructed. If a region exhibits a certain degree of deformation but can recover to its original state, it is considered elastic deformation and not a tear. If a region's deformation continues to intensify and fails to rebound over a long period of time, it is considered to have undergone irreversible tearing. This step generates a set of tear feature description vectors that integrate thermal sensing information, structural change information, and material deformation information.

[0016] The fused feature map is input into the quaternion material deformation analysis module. Within this module, an image feature point detection algorithm is used to extract a set of representative local feature point coordinates from the fused feature map. These feature points are located in areas of the image with significant thermal anomalies or texture distortion, effectively representing areas of active belt surface deformation. Successive image frames are analyzed over time, and a stable point-to-point mapping relationship is established between the two frames using an image matching strategy. This creates a feature point mapping table, which records the spatial position changes of the same physical point over time. Based on the feature point mapping table, a local coordinate system is constructed. This coordinate system is centered on the feature point and, combined with information from its neighboring pixels, a displacement field is constructed for the local area through interpolation or interpolation methods, thereby describing the continuous deformation behavior of the material at the microscopic scale. This displacement field reflects the relative displacement changes of each point in the two-dimensional image space. This displacement field is then mapped into three-dimensional Euclidean space to reconstruct the spatial deformation vector field, ensuring that each feature point not only contains planar displacement information but also describes depth changes. The displacement trajectory of each point in the spatial deformation vector field is converted into a quaternion representation, which more compactly and stably expresses the rotation, twist, and stretch changes of the object in three-dimensional space. Based on this, quaternion difference analysis is performed on consecutive time slices to calculate the rotation direction and angle changes of each local region between adjacent moments. The target quaternion is then extracted, which comprehensively expresses the rotational offset and stretch distortion of the local structure of the belt surface over a period of time. The target quaternion is serialized and analyzed on the time axis, and the intrinsic deformation data of the material during the evolution process is extracted through the quaternion difference method, effectively distinguishing between various states such as normal elastic recovery, plastic accumulation, and sudden rupture. Based on the intrinsic deformation data, the spatial partial derivatives of the corresponding coordinates of each feature point before and after deformation are inferred. The Jacobian matrix required for spatial mapping is constructed from this partial derivative information, and the deformation gradient tensor is calculated based on this information to describe the continuous deformation relationship of the material at the microscale.

[0017] Step S3: performing hierarchical progressive recognition and stereo reconstruction analysis based on the tear feature description vector and the multi-channel pre-processed image to obtain target tear feature data; Specifically, a preliminary rough inspection is performed based on the tear feature description vector and the corresponding multi-channel preprocessed image. During this process, a lightweight neural network model is used to mesh the image. Combined with the thermal deformation and texture anomaly information contained in the tear feature vector, the model quickly determines the presence of potential tearing within the delineated image cells, generating preliminary localization results for candidate tear regions. Candidate regions are then screened using a confidence threshold. Low-confidence regions are eliminated based on the classification probability or response strength output by the model, retaining only high-confidence bounding boxes as representative suspect regions. This significantly reduces the computational overhead of subsequent processing and improves overall recognition accuracy. The suspect region bounding boxes are mapped into the field of view of a binocular 3D depth camera. A structured light assistance system is simultaneously activated to project a grating pattern with coded features onto the belt surface. This coded grating enhances the surface texture, increasing the number of keypoints and localization accuracy during stereo matching, thereby enhancing the subsequent 3D reconstruction accuracy. Global stereo matching is then performed on the enhanced left and right view images, generating a disparity map containing pixel-level displacement information. This disparity map describes the parallax difference between the two cameras when observing the same object in space. Based on this disparity information, the system applies triangulation principles to convert the 2D disparity values of each valid pixel into corresponding 3D spatial coordinates, constructing an initial 3D structural model of the target tear region. This initial 3D structural model is then subjected to pixel-level tear segmentation, and a deep segmentation network is used to extract the tear boundary with high precision, ensuring the segmentation result has strong spatial contour clarity and integrity. Based on this segmentation result, a target appearance model of the tear region is constructed and mapped to the frequency domain. A corresponding response map is generated using Fourier transform techniques. Using a convolution operation, the peak similarity between this appearance model and subsequent frames is calculated in the frequency domain, enabling precise localization and continuous tracking of the tear region. As the tear region continues to move and deform during the belt cycle, the system updates the appearance model in real time and dynamically tracks the changes in the tear region's contour. Combined with spatial measurement data from each frame, the rates of change of the tear length, width, and maximum depth are calculated, reflecting the trend and rate of tear propagation. By tracking the tear location and dynamically monitoring its geometric parameters, a set of target tear feature data, including location, size, and morphological changes, is generated.

[0018] Step S4: construct a multidimensional morphological feature vector based on the target tearing feature data, and perform risk assessment on the multidimensional morphological feature vector to obtain a tearing grade classification result.

[0019] Specifically, basic tear characteristic parameters, including the length, width, maximum depth, and total area of the tear region, are extracted from the target tear feature data. Based on the basic geometric features, the tear morphological complexity parameter is calculated. This parameter, defined by combining the nonlinear relationship between the tear boundary perimeter and area, characterizes whether the tear boundary exhibits significant irregular fluctuations and bifurcations, thereby reflecting the structural complexity of the tear failure. Furthermore, image second-order moment analysis is used to perform directional statistics on the tear region. The inclination angle of the tear region's principal axis—the region's primary extension direction in physical space relative to the belt's running direction—is calculated, yielding the tear's spatial directional variation trend, which is classified as a derivative feature of the tear morphology. A dynamic analysis of the tear's temporal evolution is conducted. Time-differentiation operations are performed on key parameters such as length, width, and depth in the target tear feature data, calculating the rate of change of these parameters per unit time. This constructs a dynamic tear variation feature, revealing whether the tear is stable or accelerating, and, to a certain extent, reflecting the tear's evolutionary path and potential risk trends. The aforementioned tear characteristics are integrated with their geometric, morphologically derived, and dynamic characteristics to construct a multidimensional morphological feature vector. This vector expresses the structural information of the tear region in multiple dimensions, including shape, location, direction, and evolution rate, in a high-dimensional numerical form. This multidimensional morphological feature vector is then fed into an XGBoost multi-classification model for feature combination judgment. The model utilizes an ensemble learning algorithm based on gradient boosting trees to perform nonlinear combination and judgment on the input multidimensional features, and automatically classifies the tear severity based on the combined relationships between the features. The model pre-defines four tear severity categories: first-degree tears (small, initial cracks); second-degree tears (slightly extended); third-degree tears (medium-sized and deep); and fourth-degree tears (significant structural damage and obvious danger). During training, the model learns the corresponding grade labels for different feature distributions using historical data. In actual operation, the model accurately classifies the tear region based on the multidimensional feature structure it presents, outputting a tear severity classification with clear physical meaning and risk level indication.

[0020] Based on the identified levels, different tear types are mapped to corresponding control modes. Level 1 tear, defined as the initial microcracking stage, is assigned to monitoring mode, where the system performs only real-time data recording and trend observation. Level 2 tear, corresponding to the mild damage stage, is assigned to load management mode, where system parameters are adjusted to slow the tear's progression. Level 3 tear, indicating a moderate tear, is mapped to standby switching mode, where a backup transmission unit is activated to take over the current system operation. Level 4 tear, indicating a severe tear requiring immediate safety shutdown, is assigned to safety shutdown mode, placing the system in a high-risk response state. These four modes collectively form the basis for defining the current tear control mode and determine the direction of subsequent system scheduling and execution strategies. Based on the current tear control mode, a built-in multi-objective optimization algorithm is used to solve the control parameters. This algorithm prioritizes minimizing tear risk, maximizing transmission efficiency, and minimizing equipment stress as parallel objectives. By adaptively adjusting the weighting coefficients of each objective, the algorithm determines the optimal control parameters that best meet the operational requirements for the current tear level. After obtaining the control parameters, a control model for rolling prediction is constructed, combining the current operating state of the belt with the dynamic characteristics of the equipment. Through simulation, the response behavior of the belt system and the tear development trajectory under different control conditions are simulated, generating a predictive control sequence containing control actions at multiple time nodes. This sequence is used to predict the comprehensive impact of control measures on the system and formulate a forward-looking dynamic response plan. Based on the control actions covered in the predictive control sequence, differentiated control instruction sets are generated for different tear levels. For first- and second-level tears, the system focuses on flexible adjustment, generating speed reduction instructions and feeding point offset instructions, respectively. By appropriately reducing the belt line speed and adjusting the material discharge position, the tension and impact load in the tear area are reduced, slowing the damage expansion. For third-level tears, a backup system switch instruction is generated, mobilizing the backup belt channel. After ensuring that the operating parameters of the two systems are aligned, the transfer is smoothly implemented. For the most severe fourth-level tears, a graded deceleration shutdown instruction is generated. This multi-stage deceleration process reduces system inertia and avoids structural impact caused by direct power outages, thereby achieving a safe and orderly shutdown process. The differentiated control instruction set generated above is input into the control instruction parsing module, and converted through the interface protocol of each actuator to generate precise control instructions in the format supported by field equipment such as frequency converters, tension adjustment devices and belt alignment mechanisms. These instructions are then sent to the corresponding control terminals through the industrial field bus system, so that each control operation can be executed accurately and efficiently at the device level, forming a closed-loop linkage mechanism between tear detection and control, significantly improving the safety, intelligence level and operational reliability of the belt system.

[0021] In an embodiment of the present invention, the system utilizes dual-channel acquisition technologies, including thermal imaging and visible light, to achieve comprehensive perception of the conveyor belt's surface temperature distribution and texture structure, overcoming the limitations of a single information source and improving the system's ability to detect tear damage in covered materials. Temperature calibration, illumination equalization, and image registration techniques effectively eliminate interference from ambient temperature drift and uneven lighting, improving image quality and feature extraction accuracy. An improved DenseNet and ResNet-50 dual-stream feature network enables parallel extraction and cross-attention fusion of thermal difference and texture features, enhancing the system's ability to characterize tear signatures. By introducing a quaternion material deformation analysis model and calculating deformation gradient tensors and invariant parameters, the system accurately characterizes the belt's local deformation state and effectively distinguishes between elastic deformation and irreversible tears. A hierarchical recognition strategy, combining coarse detection with fine segmentation, improves the computational efficiency and accuracy of tear detection, reducing false positives and missed detections. Stereo vision and structured light-assisted systems enable three-dimensional reconstruction and dynamic tracking of the tear region, capturing comprehensive information on tear morphology and development. Based on the morphological and dynamic characteristics of tearing, a comprehensive risk assessment model was constructed, enabling accurate classification of tear levels and prediction of development trends. Based on the tear level classification results, differentiated control instructions were generated through a multi-objective optimization algorithm, balancing the conflict between safety and production efficiency and improving the system's intervention effect on tearing conditions.

[0022] In a specific embodiment, the process of executing step S1 may specifically include the following steps: Install the thermal imaging camera and visible light camera above the conveyor belt at a preset angle, and calibrate the overlapping area of the field of view of the thermal imaging camera and visible light camera to obtain a unified acquisition field of view; According to the unified acquisition field of view, the thermal imaging camera parameters are configured to obtain thermal imaging acquisition parameters, and the visible light camera parameters are configured to obtain visible light acquisition parameters; The thermal imaging camera and visible light camera are time-synchronized and controlled to obtain a trigger signal. The trigger signal is then used to activate the high-brightness LED flash system, which illuminates the conveyor belt surface with instantaneous high-intensity illumination to obtain a visible light lighting environment. Based on thermal imaging acquisition parameters, visible light acquisition parameters, trigger signals, and visible light lighting environment, synchronous dual-channel image acquisition is performed on the conveyor belt surface, and the acquired first thermal imaging image and first visible light image are combined into an image pair to obtain original multi-channel image data; Temperature calibration, illumination equalization and image registration are performed on the original multi-channel image data to obtain a multi-channel preprocessed image.

[0023] Specifically, a thermal imaging camera and a visible light camera are simultaneously mounted above the conveyor belt at a preset angle, based on the actual conveyor belt width, installation height, and site environment. This preset angle is calculated based on the optical axis angle, lens focal length, and field of view overlap requirements of the two cameras, ensuring that they cover the same working area during imaging. After physical installation, the fields of view of the thermal imaging camera and visible light camera are calibrated in the overlapping area using a calibration plate method. This involves placing a calibration pattern with known geometric features on the conveyor belt surface. The two cameras capture images of the same target area, extracting feature point locations using an image processing algorithm. The projected positions of the same calibration points in the different images are compared, and a spatial mapping transformation is calculated between the two. This establishes a geometric correspondence model between the thermal and visible light images. This step creates a unified captured field of view, ensuring that each subsequent image pair is highly consistent in spatial dimensions. Based on this unified field of view information, the parameters of the thermal imaging camera and visible light camera are configured to meet the imaging requirements of the specific industrial environment. For the thermal imaging camera, appropriate parameters such as the temperature measurement range, frame rate, integration time, and sensitivity threshold are configured based on the site temperature range, material thermal characteristics, and system resolution requirements. This allows the camera to accurately capture local temperature changes on the belt surface caused by friction, force, or strain, thereby constructing an early abnormal thermal distribution map. Furthermore, the visible light camera is configured with a sufficiently high resolution to capture texture details, and the shutter time, gain, and image dynamic range are appropriately set to accommodate the complex lighting conditions and rapid motion present in the working conditions. Because visible light image quality is significantly dependent on lighting conditions, it is integrated with the LED flash system for coordinated control to ensure sufficient visible light illumination and clear images during each image acquisition. To ensure time synchronization of the dual-channel image data, a unified time control mechanism is implemented. A high-precision photoelectric encoder installed on the conveyor belt drive acquires belt displacement information in real time, which serves as a synchronous trigger source to generate a unified trigger signal for the thermal imaging camera, visible light camera, and LED flash system. In this synchronization mechanism, a photoelectric encoder emits pulses in real time based on the belt's motion state. Data acquisition is triggered each time the belt moves a certain distance. At that moment, the system simultaneously initiates thermal and visible light image acquisition, and momentarily activates a high-brightness LED flash system to emit pulsed illumination. This provides short-term, intense illumination of the conveyor belt surface, thereby suppressing the effects of ambient light fluctuations, dust obstruction, and other factors on image quality and improving overall image contrast and edge clarity. Based on thermal and visible light acquisition parameters, trigger signals, and the visible light illumination environment, synchronized dual-channel image acquisition is performed on the conveyor belt surface. Each synchronized trigger generates both a thermal image and a visible light image. The system combines these two images into image pairs, indexed by time tags and spatial positions, to form raw multi-channel image data with temporal consistency and spatial alignment.Temperature calibration is performed on the raw multi-channel image data. Based on a standard blackbody temperature source or reference temperature model deployed on-site, the temperature value of each pixel in the image is zero-corrected and gain adjusted. This eliminates drift and nonlinear response issues caused by long-term infrared detector operation, ensuring that the thermal distribution in the image truly reflects physical temperature differences. After calibration, thermal noise suppression is performed. Time-domain filtering is used to average and smooth high-frequency interference signals in consecutive thermal image frames, thereby improving the stability of the image thermal signal. Simultaneously, illumination equalization is performed on the visible light image. The overall brightness distribution characteristics of the image are analyzed, and the current illumination intensity is estimated based on the image brightness histogram. Combined with the configuration parameters of the on-site LED system, a multi-scale Retinex algorithm is used to locally compensate and globally enhance the image brightness. This eliminates local overexposure, shadowing, and color imbalance, resulting in a visible light image with higher contrast and clearer details. To achieve pixel-level spatial registration of thermal and visible light images, a frequency-domain phase cross-correlation-based registration algorithm is employed. By calculating the cross-correlation function of the two images in the frequency domain and finding its maximum, the optimal spatial alignment parameters are determined. The thermal image is spatially resampled and transformed through an interpolation algorithm and mapped to the coordinate system of the visible light image to complete high-precision spatial registration of the two-channel images, ensuring that each pixel position in the two-channel images corresponds to the same position point in physical space, thereby obtaining a multi-channel preprocessed image.

[0024] In a specific embodiment, the process of performing temperature calibration, illumination equalization, and image registration on the original multi-channel image data to obtain a multi-channel pre-processed image may specifically include the following steps: Performing temperature zero point and gain calibration on the first thermal imaging image in the original multi-channel image data to obtain temperature calibration data, and inputting the temperature calibration data into a time domain averaging filter to perform thermal noise suppression processing to obtain a second thermal imaging image; Performing brightness histogram analysis on the first visible light image in the original multi-channel image data, calculating the ambient light intensity distribution, and dynamically adjusting the LED flash intensity according to the ambient light intensity distribution to obtain a second visible light image; Performing multi-scale Retinex processing on the second visible light image to obtain illumination and reflection components, and reconstructing the illumination and reflection components based on three-scale spatial domain filtering to obtain a third visible light image; performing phase cross-correlation calculation on the second thermal imaging image and the third visible light image to obtain image registration transformation parameters; The image registration transformation parameters are applied to the second thermal imaging image, and spatial domain resampling interpolation is performed to obtain a third thermal imaging image. The third thermal imaging image is mapped to a visible light image coordinate system corresponding to the third visible light image to obtain a multi-channel preprocessed image.

[0025] Specifically, temperature zero point and gain calibration are performed on the first thermal image in the raw multi-channel image data. Infrared thermal imaging equipment is prone to sensor drift and detector response nonlinearity after long-term operation, causing the imaged temperature to deviate from its true physical state. A reference temperature source or known temperature zone is introduced as a reference. The pixel grayscale values in the thermal image are compared with the reference temperature value, and a pixel-level calibration function is constructed. This function performs a global adjustment on the entire thermal image, correcting the temperature zero point offset and unifying the gain amplification factor, resulting in a temperature-calibrated image. The temperature-calibrated image is then input into a time-domain averaging filter. Leveraging the temporal continuity of the image sequence, a sliding weighted average is performed on multiple adjacent frames. This effectively suppresses high-frequency thermal noise components in the image while preserving the structural information of the true thermal distribution on the material surface. The result is a second thermal image with a stable thermal signal and clear boundaries. Simultaneously, a brightness histogram analysis is performed on the first visible light image in the raw image data. The number of pixels corresponding to each grayscale value in the entire image is counted, and a brightness distribution map is generated. This histogram reflects the overall illumination intensity of the current image within a spatial range. Based on the distribution density of each brightness region in the histogram, the image is judged to contain typical issues such as underexposure, local overbrightness, and shadows. Based on this information, the operating parameters of the LED flash system are dynamically adjusted to compensate for ambient light. This adjustment strategy utilizes a closed-loop feedback mechanism. By analyzing the difference between the current image brightness distribution and the preset ideal distribution, the LED flash current and pulse timing are reversely controlled to adjust the flash intensity and duration, ensuring a more balanced and clearer exposure for the next visible light image acquisition. After dynamic illumination control, a second visible light image is acquired. This second visible light image is then subjected to multi-scale Retinex processing. The Retinex algorithm is an illumination-reflection separation technique that simulates the human visual system's perception of light and texture, decomposing the image into two independent components: one representing local illumination intensity variations, and the other representing the reflective properties of the object's surface. In this process, the illumination and reflection components of the original image are calculated through logarithmic transformation and Gaussian difference calculation. These components are then processed at three different spatial scales: the low scale focuses on image edge details, the medium scale is used to identify medium-scale texture structures, and the high scale is used to correct overall brightness trends. Through scale weighting and inverse fusion, a third visible light image with uniform brightness and prominent texture is reconstructed. Phase cross-correlation is calculated on the second thermal imaging image and the third visible light image. By performing fast Fourier transforms on the two images and calculating their phase difference function in the frequency domain, the maximum correlation position is located, and the horizontal and vertical translation parameters, rotation angles, and scaling ratios of the two images are inferred to obtain the image registration transformation parameters.Based on the registration parameters, a spatial resampling and interpolation operation is performed on the second thermal image. Using bilinear interpolation, cubic spline interpolation, or other high-order interpolation methods, the image is geometrically transformed according to the specified rotation, scaling, and translation parameters. The grayscale value of each transformed pixel position is recalculated to generate a third thermal image. This third thermal image is directly mapped into the coordinate system of the third visible light image, achieving complete spatial overlap and pixel alignment of the two imaging data. At this point, the two images have a unified field of view, a matching pixel grid, and a corresponding texture structure. The two images are then combined to form the final multi-channel pre-processed image.

[0026] In a specific embodiment, the process of executing step S2 may specifically include the following steps: Sending the third thermal imaging image in the multi-channel preprocessed image to the thermal feature extraction branch, and sending the third visible light image in the multi-channel preprocessed image to the texture feature extraction branch; Multi-scale feature extraction is performed on the third thermal imaging image through the DenseNet structure of the thermal feature extraction branch to obtain thermal difference features; The ResNet-50 network structure of the texture feature extraction branch is used to extract deep features of the third visible light image to obtain texture structure features; The thermal difference features and texture structure features are fused with cross attention to obtain a fused feature map containing temperature anomaly and texture deformation information; The fused feature map is input into the quaternion material deformation analysis model. The local deformation state of the belt is represented by the quaternion transformation between consecutive frames and the spatial mapping relationship is calculated to obtain the deformation gradient tensor. The deformation gradient tensor is subjected to feature decomposition and invariant extraction to obtain the main deformation direction and intensity characteristic parameters. Based on the main deformation direction and intensity characteristic parameters, a discrimination criterion is constructed to distinguish elastic deformation from irreversible tearing, and the tearing feature description vector is obtained.

[0027] Specifically, the multi-channel images are effectively decoupled and task-divided. The third thermal image is fed into the thermal feature extraction branch, while the third visible light image is fed into the texture feature extraction branch, leveraging the complementary information between the two image types. Thermal images, with their high sensitivity to temperature gradients, reveal subtle thermal anomalies on the belt surface caused by accumulated stresses such as friction, stretching, and fatigue. Visible light images, at high resolution, retain rich information such as physical texture, boundary morphology, and structural continuity. Their ability to resolve cracks, tears, or deformed edges far exceeds that of infrared images. Therefore, the dual-branch architecture enables the system to simultaneously mine key features of both thermal anomalies and geometric distortion. In the thermal feature extraction path, an improved DenseNet network architecture is employed to perform deep multi-scale feature extraction on the third thermal image. DenseNet utilizes a dense connection mechanism, where each layer directly transmits its feature map to all subsequent layers, enabling cross-layer feature reuse, thereby improving feature representation and gradient transfer efficiency. The network consists of multiple dense blocks, each integrating convolution, normalization, and activation operations. Connected layers are used to reduce the number of channels and downsample features. The resulting thermal difference feature map contains thermal gradient distributions at all scales and preserves local anomaly patterns caused by weak thermal signals generated in the early stages of a tear. The texture feature extraction path is based on the ResNet-50 network, which utilizes a residual connection architecture to prevent gradient vanishing and maintain the continuous transmission of detailed information even with deep network layers. In this channel, the third visible light image is processed through multiple layers of convolution to gradually extract a multi-level feature representation, from edges to textures to higher-level structures. This channel exhibits particularly high sensitivity to discontinuous texture features such as tensile deformation, crack edges, and material damage on the belt surface. The resulting texture structure feature map from the ResNet-50 network preserves the macroscopic trends of the belt surface material structure while accurately capturing details of local anomalies. A cross-attention fusion process is performed on the thermal difference and texture structure features to construct a fused feature map. The cross-attention mechanism maps the two feature maps into a query vector, a key vector, and a value vector, respectively. An inner product operation is used to calculate a similarity weight matrix between the channels. Based on these weights, the feature maps are dynamically weighted and combined. This allows the fused feature map to highlight the strong response characteristics of areas with temperature anomalies while also locating areas of possible tearing in the geometric structure. The fused feature map is then input into a quaternion material deformation analysis model, which is used to model and identify the local deformation state of the belt surface material. This analysis process extracts stable feature points from the image by processing the fused feature maps between consecutive frames, tracks their positional changes, and calculates the displacement trajectory of these points in three-dimensional space.To represent the rotational and tensile behavior induced by these displacements, a quaternion transformation is employed for modeling. This transforms the spatial deformation state of each local feature point into a four-dimensional vector. This vector compactly encodes the rotation angle and axis in space, avoiding the singularities common in Euler angle or rotation matrix representations. Quaternion difference analysis is performed between adjacent frames to extract local rotation increments and construct a deformation mapping. Based on the quaternion trajectory analysis results, the spatial coordinate changes between feature points before and after deformation are derived, and the deformation gradient tensor is calculated from this information. This tensor numerically represents the mapping of any point in the microscopic deformation field, describing the strain distribution, degree of stretching, and shear direction of the region during continuous deformation. To reveal the physical mechanism of the tearing behavior, the deformation tensor is eigendecomposed and a series of physically meaningful invariants are extracted, such as deformation strength, local stretching ratio, and deformation symmetry. These invariants are used to directly assess whether a particular region of the material is within the elastic deformation range or has undergone irreversible damage. A deformation state discrimination criterion is constructed by combining the principal deformation direction and strength characteristic parameters. By comparing the temporal evolution trend of the invariant with a preset threshold, it is possible to distinguish which regions are in a normal elastic expansion state and which are in a potential tearing state or experiencing irreversible fracture. All regions with clear abnormal characteristics are encoded as tear feature description vectors.

[0028] In a specific embodiment, the step of inputting the fused feature map into the quaternion material deformation analysis model, representing the local deformation state of the belt by quaternion transformation between consecutive frames, and calculating the spatial mapping relationship to obtain the deformation gradient tensor may specifically include the following steps: The fused feature map is input into the quaternion material deformation analysis model to perform feature point detection, and a local feature point coordinate set is obtained. The corresponding relationship between consecutive frames is established based on the local feature point coordinate set to obtain a feature point mapping table; Construct a local coordinate system based on the feature point mapping table, and calculate the displacement field of the area around each feature point according to the local coordinate system; Mapping the displacement field to three-dimensional Euclidean space to obtain a spatial deformation vector field, and converting the spatial deformation vector field into a quaternion representation; Quaternion transformation is performed through quaternion representation to obtain target quaternions that represent local rotation and stretching, and time series analysis and quaternion difference calculation are performed on the target quaternions to obtain material intrinsic deformation data; The coordinate partial derivatives of the corresponding points before and after deformation are calculated according to the intrinsic deformation data of the material, and the spatial mapping Jacobian matrix is constructed based on the coordinate partial derivatives to obtain the deformation gradient tensor.

[0029] Specifically, the fused feature map is input into the quaternion deformation analysis module, which performs a series of deep geometric calculations based on image processing and motion modeling. Initially, the system performs feature point detection on the fused feature map. Using a feature point extraction algorithm based on the Harris corner enhancement mechanism or the FAST weighted response mechanism, the system automatically extracts stable points from the image, including areas of thermal anomalies, texture distortion, and structural boundary intersections. These feature points exhibit high positioning accuracy and temporal consistency and are selected as core data elements for subsequent deformation tracking. The two-dimensional coordinates of all identified feature points in the image plane are stored in a local feature point coordinate table, forming a spatial distribution map of high-response areas on the fused image. To establish temporal correspondence during the continuous deformation process of the material, the local feature points in the current frame are matched point by point with the corresponding points in the adjacent previous frame. High-precision matching is achieved through multiple mechanisms, including feature descriptor alignment, local graph structure analysis, and image pyramid registration. This generates a feature point mapping table between consecutive frames. This mapping table contains the spatial coordinate position of each feature point in the two previous frames, as well as the displacement vector between them. After obtaining the point correspondences in the time series, a local coordinate system is constructed with each feature point as the center. This coordinate system uses an equidistant neighborhood model or triangulation network in the two-dimensional image grid to define the area surrounding the feature point, ensuring that the neighborhood coverage contains a sufficient number of information points while maintaining the integrity of the deformation trend of a single region. Based on this, the microscopic displacement field of the region is calculated by analyzing the displacement differences between each neighborhood point, reflecting the local deformation characteristics at the pixel level. The displacement field is mapped to three-dimensional Euclidean space. Using binocular vision principles, combined with the spatial registration relationship between thermal and visible light images and stereo imaging geometry models, the two-dimensional displacement vector of each feature point in the image is converted to its displacement vector in the three-dimensional coordinate system, forming a spatial deformation vector field. This vector field is a point-based, vector-represented continuous spatial description that records the local deformation, orientation change, and scale adjustment of each feature point in space. To more compactly represent this deformation information and avoid the gimbal lock and rotational singularities common in Euler angle representation, the spatial displacement vector is converted into a quaternion representation. During the conversion process, quaternion data is constructed based on the rotation axis vector and rotation angle of each point, completely and efficiently encoding spatial rotation behavior into a four-dimensional vector structure. Quaternion transformations are performed on the target area using quaternion representations, and temporal changes between images are represented as continuous quaternion transformation trajectories. By comparing the quaternion differences between previous and next frames, the rotation increment and stretching behavior of each local area are determined. Sequential analysis is then performed on each target quaternion, tracking its evolution over time to determine the persistence, directionality, and cumulative trend of the deformation in that area.To capture the intrinsic physical response of the material, the quaternion sequence is differentiated to extract the implicit intrinsic deformation data. This data reflects the material's true internal deformation process under stress or fatigue, providing a crucial basis for distinguishing elastic recovery from irreversible failure. The partial derivatives of the coordinates of corresponding points before and after deformation are calculated based on the intrinsic deformation data. The three-dimensional coordinates of the same feature point in both the original and deformed states are obtained, and the rates of change of the first-order derivatives along each axis are calculated. These derivatives form the fundamental variables describing the spatial differential mapping. The partial derivatives of all points are combined to construct the Jacobian matrix of the spatial mapping. This matrix represents a local linear transformation from the original space to the deformed space, describing the distortion, stretching, and compression behavior of the coordinate system in multidimensional space. Based on this constructed Jacobian matrix, the deformation gradient tensor is calculated. This tensor is a high-order tensor structure that incorporates multidimensional coordinate, directional, and scale transformations, mathematically characterizing the continuous deformation mechanism of the belt under stress, friction, or tearing. By analyzing the eigenvalues of the tensor and decomposing the principal directions, key deformation features such as principal strain direction, local expansion rate, and shear ratio are extracted.

[0030] In a specific embodiment, the process of executing step S3 may specifically include the following steps: Perform rough inspection on the tear feature description vector and multi-channel pre-processed image to obtain preliminary positioning results of candidate tear areas; Perform confidence threshold screening on the preliminary positioning results of the candidate tearing area to obtain the suspicious area bounding box; Map the suspicious area bounding box to the binocular 3D depth camera field of view, activate the structured light auxiliary system to project the coded grating to enhance the texture features, and obtain the enhanced stereo vision input image; Perform global stereo matching calculation on the enhanced stereo vision input image to obtain a disparity map, and convert the two-dimensional disparity of the disparity map into a three-dimensional point cloud through the principle of triangulation to obtain the initial three-dimensional structural model of the torn area; Perform pixel-level tear segmentation on the initial 3D structure model to obtain accurately segmented tear regions. Build a target appearance model based on the accurately segmented tear regions, and calculate the response map of the target appearance model in the Fourier domain to locate the tear position. Based on the response graph, the tearing position is located and the dynamic changes of the tearing area are tracked. At the same time, the change rates of the tearing length, width and depth are calculated to obtain the target tearing feature data.

[0031] Specifically, the tear feature description vector and the corresponding frame's multi-channel image are input to the coarse inspection module. This module utilizes a lightweight deep neural network, such as the MobileNetV3 architecture, combined with an image segmentation strategy to divide the entire image into fixed-size grid regions. A convolutional feature extractor then rapidly classifies the local image information within each grid, determining whether the region contains potential tear anomalies. To improve detection rates, the classification threshold in the coarse inspection phase is conservatively set to minimize missed detections of potential tears. During this process, geometric scale information, thermal anomaly intensity, and texture complexity metrics from the tear feature description vector are incorporated as reference factors, forming a feature fusion input with the convolutional features to enhance the coarse inspection module's recognition capabilities and region screening accuracy. The coarse inspection outputs preliminary localization information for several candidate tear regions, but these regions contain varying degrees of uncertainty. Therefore, the system performs confidence threshold screening based on this information. By analyzing the response score of each candidate region, regions with a confidence score above a preset threshold are selected, thereby constructing a set of bounding boxes for suspicious regions. The bounding box of the suspicious area is spatially mapped and accurately projected into the imaging field of view of the binocular 3D depth camera through geometric transformation. To improve the accuracy of depth information acquisition and enhance the contrast of texture features, a structured light-assisted projection system is simultaneously activated to project a pre-designed coded grating pattern onto the surface of the torn area. This coded pattern enhances image features in areas with insufficient texture or smooth surfaces by introducing an artificial texture stripe structure. This improves the binocular stereo matching algorithm's ability to match image points, reduces matching failures and disparity errors, and forms an enhanced stereo input image. A global stereo matching algorithm is used to calculate disparity on this enhanced stereo input image. Using SGBM (semi-global block matching) or a multi-path dynamic programming method based on graph optimization, the pixel displacements of the same target points in the left and right images are matched. The disparity value for each pair of matching points is obtained, generating a complete disparity map. The disparity map is an intermediate product that reflects the 3D depth information of each pixel in the image. Smaller disparity values indicate farther targets, while larger disparity values indicate closer targets. Applying the principle of triangulation based on the disparity map, the 2D disparity information of each pixel is combined with the camera baseline distance and focal length parameters to calculate its true position coordinates in 3D space. Point cloud data is generated point by point, and an initial 3D structural model of the tear area is reconstructed. This model presents the true 3D topography of the tear area in the form of a high-density spatial lattice, including features such as surface ridges, tear cracks, and sunken areas. This initial 3D structural model is input into the deep segmentation module to achieve pixel-level tear boundary segmentation. This module uses a U-Net++ architecture, leveraging its multi-scale nested skip connections to fuse deep semantic and shallow texture features for high-precision extraction of tear edges.The precise boundary contours of the tear region are extracted from the segmentation results and used as the basis for constructing an appearance model of the tear target. This model captures the spatial distribution, edge features, and crack morphology of the tear and can be mapped to the frequency domain. To improve the robustness and speed of the target tracking process, the appearance model is converted to a frequency domain representation and a Fourier domain response map is calculated. Kernel correlation filtering or phase correlation techniques are used to quickly search for peak locations in subsequent frames that closely match the original model, accurately locating the tear region. Based on the tear region location information provided by the Fourier domain response map, the tear target is continuously tracked and its morphological evolution during belt operation is dynamically monitored. The length, width, and maximum depth of the tear region are spatially measured in each frame, and their rates of change per unit time are calculated through time series analysis. The tear growth rate, width expansion trend, and depth increase are calculated to determine the tear's propagation rate and directionality. This information is then used to update the dynamic evolution dimension of the tear feature description vector, outputting a set of target tear feature data.

[0032] In a specific embodiment, the process of executing step S4 may specifically include the following steps: Extracting basic tearing feature parameters from target tearing feature data; The shape complexity parameters are calculated based on the basic characteristic parameters of the tearing, and the main axis direction of the tearing area is calculated by the second-order moment to obtain the derived characteristics of the tearing morphology; Perform time difference calculation on target tearing feature data to obtain tearing dynamic change characteristics; The basic characteristic parameters of tearing, the derived characteristics of tearing morphology and the dynamic change characteristics of tearing are combined into a multi-dimensional morphological feature vector; The multidimensional morphological feature vector is input into the XGBoost multi-classification model for feature combination judgment, and the tear is divided into tear grade classification results including first-grade tear, second-grade tear, third-grade tear and fourth-grade tear.

[0033] Specifically, the basic geometric parameters of the tear are extracted from the target tear feature data, including the length, width, maximum depth, overall area and boundary perimeter of the tear area, which together constitute the basic morphological information of the tear. Among them, the tear length reflects the extension scale of the tear along the running direction of the belt, the width describes its lateral damage range, the maximum depth reveals the degree of penetration of the crack into the thickness of the belt, and the area, as a reflection of the overall damaged area, helps to assess the macroscopic damage degree of the tear. These parameters are obtained through pixel projection and voxel resampling of the three-dimensional reconstructed model, with high physical accuracy and time continuity. Based on the basic geometric description of the tear, its structural complexity is quantitatively modeled, and the shape complexity parameter is calculated by the known tear area and boundary perimeter. This parameter is a dimensionless ratio that reflects the degree of regularity of the tear boundary. The higher the value, the more complex and irregular the tear boundary is, which is associated with high-risk tearing morphologies such as strong material fracture, edge tearing, and multi-point extension. To analyze the structural extension direction of the tear region in space, second-order moments are calculated based on the region's 2D contour or 3D point cloud projection. Starting from the centroid of the tear image, the distribution offset of all edge pixels along each principal axis is statistically analyzed to determine the principal axis direction of the tear. This direction is expressed as an angle and its spatial inclination relative to the belt's direction of motion. The extracted principal axis direction reflects the structural orientation of the tear region and helps determine whether asymmetric tear evolution is caused by interference between equipment components or unilateral force. By combining complexity parameters with principal axis directions, morphological derivative modeling is achieved from basic geometric quantities to higher-order structural features. The time dimension is introduced to perform dynamic analysis of the target tear feature data. The tear geometric parameters of the current frame and several previous frames are time-differentiated to generate dynamic characteristics, including tear length growth rate, width expansion rate, and maximum depth increase rate. These time-differential indices capture the propagation rate and damage trend during tear evolution by modeling the numerical changes in tear characteristics per unit time. Methods such as sliding window averaging and time series fitting are introduced to smooth short-term fluctuations, improve the stability and predictability of dynamic features, and obtain the dynamic change characteristics of tearing. The basic characteristic parameters of tearing, the derived characteristics of tearing morphology, and the dynamic change characteristics of tearing are combined into a multi-dimensional morphological feature vector. The multi-dimensional vector is input into the XGBoost multi-classification model based on the gradient boosting tree. During the training phase, the model has learned the correspondence between various feature patterns and tear levels based on a large number of labeled tearing samples. Within the model, each decision tree divides the discrimination path for some features, and the integrated result of multiple trees represents the final classification output. Compared with traditional linear classifiers, XGBoost has the ability to model nonlinear feature relationships and interactions between features, and can maintain high classification accuracy and generalization capabilities under complex working conditions.During the prediction stage, the XGBoost model automatically determines the level category of the tear based on the input feature vector. The classification system is divided into four levels according to the scale and severity of the tear: Level 1 tear represents early microcracks with small geometric scale, simple shape, dynamic changes approaching zero, and low risk; Level 2 tear manifests as obvious cracks and is in a state of slight expansion, with a certain degree of morphological complexity and slight dynamic evolution; Level 3 tear refers to a tear with a long length and a medium depth, a complex crack morphology and a rapid evolution trend, and there is a risk of further deterioration; Level 4 tear is the most serious type of structural damage, with characteristics such as large-area cracking, deep penetration, irregular crack expansion, and drastic dynamic changes. It is an important signal that the system should immediately issue an early warning and take measures.

[0034] In a specific embodiment, the belt tear detection method based on visual recognition further includes the following steps: Based on the tear level classification results, the first-level tear is assigned to the monitoring mode, the second-level tear is assigned to the load management mode, the third-level tear is assigned to the standby switching mode, and the fourth-level tear is assigned to the safety shutdown mode to obtain the current tear control mode; Solve the control parameters of the current tear control mode through multi-objective optimization to obtain the optimized control parameters. Based on the optimized control parameters, predict the belt system response and tear development trajectory to obtain a predictive control sequence. Based on the predictive control sequence, speed reduction and feeding point offset instructions are generated for the first and second level tearing, backup system switching instructions are generated for the third level tearing, and graded deceleration and shutdown instructions are generated for the fourth level tearing, thus obtaining a differentiated control instruction set. The differentiated control instruction set is converted into precise control instructions, and the precise control instructions are sent to the frequency converter, tension adjustment device and alignment mechanism through the field bus.

[0035] Specifically, a control mode matching mechanism is implemented based on the tear grade classification results. Within this mechanism, the lowest-grade tear, level one, is assigned to monitoring mode. This manifests as early-stage microcracks or fine surface fatigue textures, posing no direct damage risk, and therefore requires only continuous observation and dynamic recording. Level two tear, characterized by clear geometric cracking and a potential for expansion, is assigned to load management mode, where crack evolution is mitigated by actively adjusting the operating load. When the tear reaches level three, characterized by moderate tear depth or rapid growth, the system triggers backup system switching mode, preparing to implement redundant switching of the conveyor path to prevent damage to the main belt. When level four tear, characterized by severe damage or through-breaks, is identified, the system immediately switches to safety shutdown mode, prioritizing equipment and personnel safety. Through level-based pattern mapping, the current tear control mode is updated in real time. A multi-objective optimization solution is then applied to determine the control parameters for the current tear control mode. A multi-objective optimization problem is designed to address the three objectives of minimizing risk control, maximizing production efficiency, and minimizing mechanical stress, and an adaptive particle swarm optimization algorithm is employed to solve it. In this algorithm, each particle represents a set of candidate control parameters, such as belt speed, load distribution ratio, and tension coefficient. Through a dynamic weighted combination of objective functions, each parameter set is evaluated for its suitability and overall effectiveness for the current control mode. For first- and second-degree tears, maintaining production efficiency and reducing stress are prioritized, while for third- and fourth-degree tears, minimizing tear risk and mechanical impact is prioritized. Through population iteration, local search, and a global feedback mechanism, the algorithm rapidly converges to an optimal solution and outputs a set of optimized control parameters. These optimized control parameters are input into the system's internal digital twin simulation platform to predict the belt system's response and tear trajectory. In the prediction phase, a high-fidelity model of the belt conveyor system is constructed, and dynamic simulations are performed over a period of time based on the optimized control parameter operating conditions. By coupling material flow, equipment load, structural stress, and tear evolution models, the system's expected response over the next tens to hours, if implemented, is predicted. This includes indicators such as whether the tear zone will expand, whether the system tension is overloaded, and whether the energy consumption change is acceptable. The predicted control sequence is then output. This prediction sequence, based on a timeline, indicates the system state adjustment actions that should be taken at each control time step, and forms a tearing risk curve, providing early warning of potential risks caused by unreasonable control. Based on this predictive control sequence, tearing control is transformed from a static classification process to a dynamic instruction set generation process.In a first-level tear scenario, speed deceleration and feed point offset commands are automatically generated. The former reduces the belt line speed by incremental steps, slowing crack propagation. The latter controls the three-dimensional feeder to adjust the feed direction, shifting the heavy load area away from the crack to the intact area, reducing local stress concentration. In a second-level tear scenario, the system retains these commands but enhances their control frequency and response thresholds, such as increasing the deceleration frequency and shortening the offset response time. When the tear enters the third-level tear state, the system activates the backup system switch command. A status assessment system checks whether the backup belt path is in standby mode. It then calculates the optimal switch window, prioritizing periods of low material load. This triggers a series of pre-set processes, including backup belt preheating, power connection, and conveyor path switching, to ensure a smooth and seamless switchover. When the tear enters the fourth-level tear state, the system automatically executes a graded deceleration shutdown command. This involves running at a medium speed for several seconds, followed by a speed reduction buffer, and finally a safe shutdown with tension released, preventing secondary damage to the belt structure caused by sudden braking. All of these control actions are integrated into a differentiated control instruction set. The differentiated control instruction set is converted into precise control instructions. This conversion process, based on a control mapping template, encodes abstract control behavior parameters into a digital command format through a model converter, annotating the target controller address, execution time, parameter range, and response strategy. All precise control instructions are sent via the fieldbus to the inverter, tension adjustment device, material alignment mechanism, and feeding equipment in the control system, ensuring that each physical execution node responds to the command according to a unified timing, achieving a closed-loop linkage from information perception to execution feedback.

[0036] The above describes the belt tear detection method based on visual recognition in the embodiment of the present invention. The following describes the belt tear detection device based on visual recognition in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a belt tear detection device based on visual recognition includes: An image acquisition module is used to acquire dual-channel images of the conveyor belt surface to obtain original multi-channel image data, and pre-process the original multi-channel image data to obtain a multi-channel pre-processed image; A calculation module is used to extract thermal difference features and texture structure features of multi-channel preprocessed images through a dual-stream feature network, and input the thermal difference features and texture structure features into a quaternion material deformation analysis model to calculate the deformation gradient tensor and invariant parameters to obtain a tear feature description vector; The reconstruction analysis module is used to perform hierarchical progressive recognition and stereo reconstruction analysis based on the tear feature description vector and multi-channel pre-processed images to obtain target tear feature data; The risk assessment module is used to construct a multidimensional morphological feature vector based on the target tearing feature data, and perform risk assessment on the multidimensional morphological feature vector to obtain a tearing grade classification result.

[0037] By integrating the aforementioned components, the present invention achieves comprehensive perception of the conveyor belt's surface temperature distribution and texture structure through dual-channel acquisition technology using thermal imaging and visible light. This overcomes the limitations of a single information source and enhances the system's ability to detect tear damage in covered materials. Temperature calibration, illumination equalization, and image registration techniques effectively eliminate interference from ambient temperature drift and uneven lighting, improving image quality and feature extraction accuracy. An improved DenseNet and ResNet-50 dual-stream feature network enables parallel extraction and cross-attention fusion of thermal and texture features, enhancing the system's ability to characterize tear signatures. By introducing a quaternion material deformation analysis model and calculating deformation gradient tensors and invariant parameters, the system accurately characterizes the belt's local deformation state and effectively distinguishes between elastic deformation and irreversible tears. A hierarchical recognition strategy combining coarse detection processing with fine segmentation improves the computational efficiency and accuracy of tear detection, reducing false positives and missed detections. Stereo vision and structured light-assisted systems enable three-dimensional reconstruction and dynamic tracking of the tear region, capturing comprehensive information on tear morphology and development. Based on the morphological and dynamic characteristics of tearing, a comprehensive risk assessment model was constructed, enabling accurate classification of tear levels and prediction of development trends. Based on the tear level classification results, differentiated control instructions were generated through a multi-objective optimization algorithm, balancing the conflict between safety and production efficiency and improving the system's intervention effect on tearing conditions.

[0038] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, memory, display screen, input device, network interface and database connected via a system bus. The processor of the computer design is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the above method is implemented.

[0039] Those skilled in the art will understand that Figure 3The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0040] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0041] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0042] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A belt tear detection method based on visual recognition, characterized in that: include: Performing dual-channel image acquisition on the conveyor belt surface to obtain original multi-channel image data, and preprocessing the original multi-channel image data to obtain a multi-channel preprocessed image; Extracting thermal difference features and texture structure features of the multi-channel preprocessed image through a dual-stream feature network, and inputting the thermal difference features and texture structure features into a quaternion material deformation analysis model to calculate deformation gradient tensors and invariant parameters to obtain a tear feature description vector; Performing hierarchical progressive recognition and stereo reconstruction analysis based on the tear feature description vector and the multi-channel pre-processed image to obtain target tear feature data; A multi-dimensional morphological feature vector is constructed based on the target tearing feature data, and a risk assessment is performed on the multi-dimensional morphological feature vector to obtain a tearing grade classification result.

2. The belt tear detection method based on visual recognition according to claim 1 is characterized in that: The dual-channel image acquisition of the conveyor belt surface to obtain original multi-channel image data, and preprocessing the original multi-channel image data to obtain a multi-channel preprocessed image, includes: Installing a thermal imaging camera and a visible light camera above the conveyor belt at a preset angle, and calibrating the overlapping areas of the fields of view of the thermal imaging camera and the visible light camera to obtain a unified acquisition field of view; According to the unified acquisition field of view, the thermal imaging camera is configured with parameters to obtain thermal imaging acquisition parameters, and the visible light camera is configured with parameters to obtain visible light acquisition parameters; Performing time synchronization control on the thermal imaging camera and the visible light camera to obtain a trigger signal, and activating a high-brightness LED flash system through the trigger signal to perform instantaneous high-intensity illumination on the conveyor belt surface to obtain a visible light illumination environment; Based on the thermal imaging acquisition parameters, the visible light acquisition parameters, the trigger signal, and the visible light illumination environment, synchronously capturing dual-channel images of the conveyor belt surface, and combining the captured first thermal imaging image and first visible light image into an image pair to obtain original multi-channel image data; Temperature calibration, illumination equalization and image registration are performed on the original multi-channel image data to obtain a multi-channel pre-processed image.

3. The belt tear detection method based on visual recognition according to claim 2, characterized in that: The performing temperature calibration, illumination equalization, and image registration on the original multi-channel image data to obtain a multi-channel pre-processed image includes: performing temperature zero point and gain calibration on a first thermal imaging image in the original multi-channel image data to obtain temperature calibration data, and inputting the temperature calibration data into a time domain averaging filter to perform thermal noise suppression processing to obtain a second thermal imaging image; Performing a brightness histogram analysis on the first visible light image in the original multi-channel image data, calculating an ambient light intensity distribution, and dynamically adjusting the LED flash intensity according to the ambient light intensity distribution to obtain a second visible light image; performing multi-scale Retinex processing on the second visible light image to obtain illumination and reflection components, and reconstructing the illumination and reflection components based on three-scale spatial domain filtering to obtain a third visible light image; performing phase cross-correlation calculation on the second thermal imaging image and the third visible light image to obtain image registration transformation parameters; The image registration transformation parameters are applied to the second thermal imaging image, spatial domain resampling interpolation is performed to obtain a third thermal imaging image, and the third thermal imaging image is mapped to a visible light image coordinate system corresponding to the third visible light image to obtain a multi-channel preprocessed image.

4. The belt tear detection method based on visual recognition according to claim 1, characterized in that: The method extracts thermal difference features and texture structure features of the multi-channel preprocessed image through a dual-stream feature network, and inputs the thermal difference features and texture structure features into a quaternion material deformation analysis model to calculate deformation gradient tensors and invariant parameters to obtain a tear feature description vector, including: Sending the third thermal imaging image in the multi-channel pre-processed image to the thermal feature extraction branch, and sending the third visible light image in the multi-channel pre-processed image to the texture feature extraction branch; Performing multi-scale feature extraction on the third thermal imaging image using the DenseNet structure of the thermal feature extraction branch to obtain thermal difference features; Performing depth feature extraction on the third visible light image using the ResNet-50 network structure of the texture feature extraction branch to obtain texture structure features; Performing cross-attention fusion processing on the thermal difference feature and the texture structure feature to obtain a fusion feature map containing temperature anomaly and texture deformation information; The fused feature map is input into the quaternion material deformation analysis model, the local deformation state of the belt is represented by the quaternion transformation between consecutive frames, and the spatial mapping relationship is calculated to obtain the deformation gradient tensor; The deformation gradient tensor is subjected to feature decomposition and invariant extraction to obtain main deformation direction and intensity characteristic parameters, and a discrimination criterion is constructed based on the main deformation direction and the intensity characteristic parameters to distinguish elastic deformation from irreversible tearing, thereby obtaining a tearing feature description vector.

5. The belt tear detection method based on visual recognition according to claim 4 is characterized in that: The fusion feature map is input into the quaternion material deformation analysis model, the local deformation state of the belt is represented by the quaternion transformation between consecutive frames, and the spatial mapping relationship is calculated to obtain the deformation gradient tensor, including: Inputting the fused feature map into a quaternion material deformation analysis model to perform feature point detection to obtain a local feature point coordinate set, and establishing a correspondence between consecutive frames based on the local feature point coordinate set to obtain a feature point mapping table; Constructing a local coordinate system based on the feature point mapping table, and calculating a displacement field for an area around each feature point according to the local coordinate system; Mapping the displacement field to a three-dimensional Euclidean space to obtain a spatial deformation vector field, and converting the spatial deformation vector field into a quaternion representation; Performing quaternion transformation using the quaternion representation to obtain a target quaternion representing local rotation and stretching, and performing time series analysis and quaternion difference calculation on the target quaternion to obtain material intrinsic deformation data; The coordinate partial derivatives of the corresponding points before and after the deformation are calculated according to the intrinsic deformation data of the material, and the spatial mapping Jacobian matrix is constructed according to the coordinate partial derivatives to obtain the deformation gradient tensor.

6. The belt tear detection method based on visual recognition according to claim 1, characterized in that: The step of performing hierarchical progressive recognition and stereo reconstruction analysis based on the tear feature description vector and the multi-channel pre-processed image to obtain target tear feature data includes: Performing rough inspection on the tear feature description vector and the multi-channel pre-processed image to obtain a preliminary positioning result of a candidate tear region; Performing confidence threshold screening on the preliminary positioning results of the candidate tearing region to obtain a suspicious region bounding box; Mapping the suspicious area bounding box to the field of view of the binocular 3D depth camera, activating the structured light auxiliary system to project the coded grating to enhance the texture features, and obtaining an enhanced stereo vision input image; Performing a global stereo matching calculation on the enhanced stereo vision input image to obtain a disparity map, and converting the two-dimensional disparity of the disparity map into a three-dimensional point cloud through the principle of triangulation to obtain an initial three-dimensional structural model of the torn area; Performing pixel-level tear segmentation on the initial three-dimensional structure model to obtain accurately segmented tear regions, establishing a target appearance model based on the accurately segmented tear regions, and calculating a response map of the target appearance model in the Fourier domain to locate the tear position; Based on the response graph, the tearing position is located and the dynamic change of the tearing area is tracked. At the same time, the change rate of the tearing length, width and depth is calculated to obtain the target tearing feature data.

7. The belt tear detection method based on visual recognition according to claim 1, characterized in that: The multi-dimensional morphological feature vector is constructed based on the target tearing feature data, and risk assessment is performed on the multi-dimensional morphological feature vector to obtain a tearing grade classification result, including: Extracting tearing basic characteristic parameters from the target tearing characteristic data; Calculating shape complexity parameters based on the basic tearing characteristic parameters, and calculating the principal axis direction of the tearing area by the second-order moment to obtain the tearing morphology derivative characteristics; Performing time difference calculation on the target tearing characteristic data to obtain tearing dynamic change characteristics; Combining the tearing basic characteristic parameters, the tearing morphology derivative characteristics and the tearing dynamic change characteristics into a multi-dimensional morphological feature vector; The multidimensional morphological feature vector is input into the XGBoost multi-classification model for feature combination judgment, and the tear is divided into tear level classification results including first-level tear, second-level tear, third-level tear and fourth-level tear.

8. The belt tear detection method based on visual recognition according to claim 7, characterized in that: The belt tear detection method based on visual recognition also includes: Based on the tear level classification result, the first tear is assigned to the monitoring mode, the second tear is assigned to the load management mode, the third tear is assigned to the standby switching mode, and the fourth tear is assigned to the safety shutdown mode, thereby obtaining the current tear control mode; Solving the control parameters of the current tear control mode through multi-objective optimization to obtain optimized control parameters, and predicting the belt system response and tear development trajectory based on the optimized control parameters to obtain a predicted control sequence; Based on the predictive control sequence, speed reduction and feeding point offset instructions are generated for the first and second level tearing, backup system switching instructions are generated for the third level tearing, and graded deceleration and shutdown instructions are generated for the fourth level tearing, thereby obtaining a differentiated control instruction set; The differentiated control instruction set is converted into a precise control instruction, and the precise control instruction is sent to the frequency converter, the tension adjustment device and the alignment mechanism via a field bus.

9. A belt tear detection device based on visual recognition, characterized in that: Used to perform the belt tear detection method based on visual recognition according to any one of claims 1 to 8, the belt tear detection device based on visual recognition comprises: An image acquisition module is used to perform dual-channel image acquisition on the conveyor belt surface to obtain original multi-channel image data, and pre-process the original multi-channel image data to obtain a multi-channel pre-processed image; a calculation module, configured to extract thermal difference features and texture structure features of the multi-channel preprocessed image through a dual-stream feature network, and input the thermal difference features and the texture structure features into a quaternion material deformation analysis model to calculate a deformation gradient tensor and invariant parameters to obtain a tear feature description vector; A reconstruction and analysis module, configured to perform hierarchical progressive recognition and stereo reconstruction analysis based on the tear feature description vector and the multi-channel pre-processed image to obtain target tear feature data; The risk assessment module is used to construct a multi-dimensional morphological feature vector based on the target tearing feature data, and perform risk assessment on the multi-dimensional morphological feature vector to obtain a tearing grade classification result.

10. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and is characterized in that when the processor executes the computer program, the belt tear detection method based on visual recognition described in any one of claims 1 to 8 is implemented.

Images