A belt tearing detection method and system based on full-form perception and multi-scale reconstruction

Abstract

The application discloses a kind of based on full form perception and multi-scale reconstruction belt tear detection method and system, it is related to computer vision and industrial defect detection field.The method constructs full form tear dataset using double space expansion enhancement, embeds multi-branch anisotropic feature reconstruction operator in the feature pyramid fusion stage of improved model based on YOLO26 architecture, and is trained using full form perception geometric loss;Inference stage will real-time video stream be decoded as continuous image frame or directly input single frame image, output detection class, confidence and position coordinate, and combined confidence, crack length and continuous frame count execute joint early warning.The scheme is used for the real-time detection and alarm of conveyor belt tear, under the premise of guaranteeing very low power consumption, different scale and complex form belt tear is realized to high-precision, strong robustness detection.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and industrial defect detection technology, and more specifically, to a belt tear detection method and system based on full-morphology perception and multi-scale reconstruction. Background Technology

[0002] As a crucial conveying equipment in industrial production, the surface health of belt conveyors directly affects the safety and efficiency of the entire production line. Due to prolonged high-load operation, impact from sharp materials, and equipment aging, belt surfaces are highly susceptible to tearing. If not detected and addressed promptly, tiny cracks can rapidly expand into long, penetrating tears, leading to significant economic losses or even safety accidents.

[0003] Traditional contact sensors, such as leak switches and coil detectors, are susceptible to mechanical wear and dust interference, resulting in high false alarm and missed alarm rates. They typically only trigger physical alarms after the belt has been completely torn or broken, completely lacking early warning capabilities. Meanwhile, early traditional machine vision methods, such as edge detection and threshold segmentation, rely heavily on manually designed features. In harsh industrial environments such as coal mines and metallurgy, where there is high dust, drastic changes in lighting, and extremely complex belt surface textures, the algorithms have very poor robustness and cannot meet the requirements of continuous and stable production.

[0004] In recent years, deep learning-based target detection algorithms, especially the YOLO series, have gradually become the mainstream research direction for belt tear detection due to their powerful automatic feature extraction capabilities and inference speed. However, real belt surface defects have extremely high morphological diversity, encompassing extremely small cracks, complex and tortuous irregular cracks, and long, through-the-belt tears with high aspect ratios. Faced with such complex industrial scenarios, existing deep learning detection methods and published related patents generally suffer from the following structural limitations: 1. Most existing improved methods rely on standard square convolutional kernels or conventional attention mechanisms such as CBAM and CA for feature fusion. These operators are "isotropic" in terms of spatial receptive field expansion (i.e., they expand uniformly outward in the length and width directions), which cannot effectively fit the slender tear features with extreme aspect ratios; 2. Existing object detection patents usually use IoU, CIoU, or DIoU as regression loss functions for bounding boxes. These traditional loss functions are also isotropic (i.e., the width and height errors of the bounding box contribute the same weight to the gradient), which easily leads to missed detection of small targets and cannot provide directional gradient excitation, resulting in slow convergence and inaccurate localization of slender bounding boxes; 3. The model is susceptible to parameter collapse caused by environmental noise and augmented data interference, resulting in poor model generalization ability. Therefore, the industry urgently needs a visual inspection method for belt tears that can break through the isotropic limitations of existing convolution operators and loss functions, achieve accurate perception of the entire shape, and perfectly balance early minute defects and long-distance tear detection while ensuring extremely low computing power consumption and high noise resistance. Summary of the Invention

[0005] In view of this, and addressing the shortcomings of existing technologies such as mismatched receptive fields for defects with extreme aspect ratios in complex industrial environments, easy omission of early-stage small targets, and lack of geometric perception and noise resistance in the loss function, the purpose of this invention is to provide a belt tear detection method and system based on full-morphology perception and multi-scale reconstruction. This method can achieve high-precision detection of belt tears of different scales and shapes while ensuring extremely low computational power consumption. Full-morphology perception refers to the ability of the detection model to simultaneously perform feature extraction, boundary localization, and category recognition for belt tear defects of different scales, aspect ratios, and textures, including obvious short tears, fine short tears, obvious long tears, fine long tears, and irregular tears; this capability is mainly achieved through a multi-branch anisotropic feature reconstruction operator and a full-morphology perception geometric loss.

[0006] A belt tear detection method and system based on full-morphology perception and multi-scale reconstruction includes the following steps:

[0007] Step S1: Acquire real-time video of the belt surface and original image of the torn belt surface during the operation of the belt conveyor; classify the belt surface tears into five types of defects based on the spatial geometric characteristics of the defects: obvious short tears, fine short tears, obvious long tears, fine long tears and irregular tears.

[0008] Step S2: The original images acquired in Step S1 are preprocessed and augmented using a dual-space expansion enhancement strategy. The original images are combined to construct a belt tear defect dataset covering all types of cracks. The dataset is then uniformly labeled with bounding boxes using a labeling tool and divided into a training set and a validation set.

[0009] Step S3: Construct a full-morphological perception target detection model based on the improved YOLO26 architecture; the model sequentially includes: a backbone network for multi-scale feature extraction, a multi-branch anisotropic feature reconstruction operator embedded in the feature pyramid fusion stage for deep feature morphological decoupling and reconstruction, and a detection head for outputting defect category probability, confidence level, and predicted bounding box position coordinates;

[0010] Step S4: Input the training dataset described in step S2 into the full-morphology perception target detection model described in step S3, use full-morphology perception geometric loss to calculate the comprehensive error between the predicted bounding box and the labeled bounding box in step S2, and perform backpropagation and parameter update until the model converges to obtain the trained detection model.

[0011] Step S5: Decode the real-time acquired video stream into continuous image frames, or directly input the images into the trained detection model for forward inference, output the detection category, confidence level and predicted bounding box position coordinates of the belt tear, and trigger an alarm signal in combination with multi-dimensional joint safety warning logic.

[0012] Furthermore, in step S2, the dual-space expansion enhancement strategy specifically includes:

[0013] (1) Pixel space enhancement: Explicit geometric and photometric transformations are performed at the image pixel level, including random rotation, flipping, adding Gaussian noise, and blurring and brightness contrast adjustment to simulate an industrial environment;

[0014] (2) Latent space enhancement: Using the original image as a guiding condition, a diffusion model is used to generate belt crack variants with unknown texture changes, rare lighting conditions and different coal dust covering materials in the latent space to achieve semantic-level augmentation.

[0015] Furthermore, in step S3, the specific structure and operating logic of the improved full-morphology perception target detection model based on the YOLO26 architecture are as follows:

[0016] (1) Backbone network for multi-scale feature extraction: A lightweight convolutional network architecture is adopted to perform downsampling and feature extraction on the preprocessed image layer by layer from bottom to top, and output multi-scale initial feature maps containing shallow surface texture details and deep strong semantic information in sequence.

[0017] (2) Omni_G_Block, a multi-branch anisotropic feature reconstruction operator embedded in the feature pyramid fusion stage for deep feature morphological decoupling and reconstruction: receives the multi-scale initial feature map and performs cross-layer cascading and feature fusion; wherein, the multi-branch anisotropic feature reconstruction operator is embedded in the deep feature fusion stage.

[0018] When constructing this operator, an anisotropic phantom reconstruction mechanism, a first of its kind in this invention, is introduced: traditional lightweight networks such as GhostNet usually use symmetric isotropic convolution to generate redundant phantom features; while the anisotropic phantom reconstruction mechanism abandons blind generation. It first extracts highly condensed intrinsic semantics, and then, taking into account the highly directional nature of belt tearing, it innovatively uses long-distance strip probes with extremely low parameter cost to regenerate anisotropic phantom features with strong direction awareness along orthogonal directions.

[0019] Let the deep feature map input to this operator be... ,in, Indicates the number of channels. and These represent the height and width of the feature map, respectively. This operator first processes the input feature map... Channel compression and intrinsic semantic extraction are performed, followed by morphological decoupling and reconstruction through strip convolution branches in both horizontal and vertical directions. The specific mathematical expression is as follows:

[0020] First, using Convolution extracts intrinsic semantic features To achieve channel dimensionality reduction and cross-channel information integration:

[0021] In the formula, express Convolution operation, For batch normalization, It is a non-linear activation function.

[0022] Subsequently, the intrinsic features Parallel inputs are fed into two mutually orthogonal long-range strip probe branches for polymorphic mapping:

[0023] a. Horizontal phantom extraction branch: using... Strip convolution captures long-range dependencies along the horizontal direction, generating horizontal phantom features. To ensure feature diversity and significantly reduce the number of parameters, grouped convolution is used, expressed as:

[0024] In the formula, The SiLU activation function is used. For batch normalization. express Grouped convolution operation, where The span of the strip convolution is set according to the specific objective; The number of groups is determined by the greatest common divisor of the number of intrinsic channels and the number of horizontal phantom channels;

[0025] b. Vertical Phantom Extraction Branch: (Using...) Strip convolution captures long-range dependencies along the vertical direction, generating vertical phantom features. The expression is:

[0026] In the formula, The SiLU activation function is used. For batch normalization. express Grouped convolution operation, where The span of the strip convolution is set according to the specific objective; The number of groups is determined by the greatest common divisor of the number of intrinsic channels and the number of vertical phantom channels;

[0027] Finally, channel stitching and residual reconstruction: integrating intrinsic features The phantom features from the two directions mentioned above are directly spliced ​​and fused along the channel dimension to form the full morphological receptive field. Cross-channel information is then converged through a transitional convolutional layer, and residual connections are introduced to form the final output of the operator. :

[0028] in, This represents the output feature map after reconstruction by the multi-branch anisotropic feature reconstruction operator Omni_G_Block, which is used for subsequent feature pyramid fusion and detector prediction.

[0029] (3) Detection head used to output predicted boundary and category probability: Receives the reconstructed multi-scale fusion features and outputs in parallel the category probability of five types of defects: obvious short tear, fine short tear, obvious long tear, fine long tear and irregular tear, as well as the position coordinates of the predicted bounding box.

[0030] Furthermore, in step S4, the specific calculation logic for the full-morphological perception geometric loss is as follows:

[0031] First, when calculating the bounding box regression loss, a custom empty shell network module is introduced to truncate and block the gradient backpropagation of the original distributed focus loss, so that it always returns a zero tensor consistent with the target shape, thus ensuring the purity of the full morphological perception geometric loss during backpropagation.

[0032] Subsequently, let the center coordinates and width and height of the model's predicted bounding box be... The center coordinates and width and height of the annotation bounding box marked in step S2 are: The square of the diagonal distance between the two smallest bounding rectangles is The square of the Euclidean distance between the center points is Based on the area of ​​the bounding box. Set a threshold for the area of ​​small targets. The loss is calculated adaptively in two branches:

[0033] (1) Regarding area For regular and slender targets, calculate the anisotropic phantom penalty term. With dynamic focus factor regression loss :

[0034] To enable the loss function to apply additional directional gradient penalties based on the specific direction of the belt crack's extension, the width and height of the predicted bounding box and the labeled bounding box are extracted, and the horizontal sensing weights are calculated accordingly. and vertical perception weights :

[0035]

[0036] Simultaneously, the adaptive adjustment coefficient of the aspect ratio of the labeled bounding box is calculated. And perform an upper limit truncation operation on it, with the maximum value limited to 1. ,

[0037] Based on this, an anisotropic phantom penalty item is constructed. :

[0038] Using outlier measure

[0039]

[0040] Finally, the bounding box regression loss for this branch is calculated as follows:

[0041]

[0042] In the formula, To predict the intersection-union ratio (IU) between the bounding box and the labeled bounding box, ϵ is a smoothed minimum value to prevent division by zero. The parameters for dynamic focusing are set according to the specific target.

[0043] (2) Regarding area For extremely small targets, calculate the loss based on normalized Wasserstein distance. :

[0044] For extremely small targets To address the problem of gradient vanishing due to its tendency to approach zero, the bounding box is modeled as a two-dimensional Gaussian distribution, and the square of the Wasserstein distance between the predicted and labeled bounding boxes is calculated. :

[0045]

[0046] Normalize it using an exponential function to obtain the NWD similarity measure, and then calculate the regression loss for small targets:

[0047]

[0048] The model is set according to specific objectives; the model is based on... and The mask determination automatically summarizes the corresponding losses to complete the backpropagation.

[0049] Furthermore, in step S5, the multi-dimensional joint security early warning logic specifically includes:

[0050] The system issues a shutdown or maintenance warning signal when any or a combination of the following conditions are met: (1) The confidence level of a certain torn bounding box detected in a single frame exceeds the preset danger threshold; (2) The pixel length of the detected elongated torn bounding box in its main extension direction exceeds the safety threshold; (3) In the historical cache queue of multiple consecutive frames, based on spatiotemporal coordinate tracking, the cumulative number of torn defects detected at the same spatial location exceeds the set fault tolerance threshold.

[0051] Compared with the prior art, the present invention has the following advantages:

[0052] 1. This invention introduces an anisotropic phantom reconstruction mechanism for belt tear target detection by constructing a multi-branch anisotropic feature reconstruction operator. It specifically optimizes the structure for the high aspect ratio specificity of industrial belt tears, abandoning redundant traditional isotropic operators, and extracts the complete geometric skeleton of slender tears with extremely low parameter cost, effectively solving the structural omission problem of traditional algorithms in long-range dependencies.

[0053] 2. This invention overcomes the isotropic limitation of traditional loss functions such as CIoU / DIoU. The built-in anisotropic phantom penalty term enables the network to obtain adaptive directional gradient excitation when regressing slender targets, resulting in more accurate localization. Gradient loss due to early microcracks is compensated by normalized Wasserstein distance. At the same time, the specially designed native distributed focus loss blocking mechanism and dynamic focusing factor ensure that the model maintains extremely high robustness when facing severe environmental noise and enhanced dirty labels generated by diffusion models.

[0054] 3. When dealing with multi-scale datasets containing five complex tearing patterns, this invention has a relatively small number of parameters, only about 1.75 M, lower than the benchmark model YOLO26n, and a computational load of only 7.1 GFLOPS. On an NVIDIA RTX 4080, it achieves a pure inference speed of 764.4 FPS and an end-to-end real-time inference speed of 308.1 FPS, significantly reducing the computing power and storage requirements for edge computing devices in industrial settings. While improving the average detection accuracy mAP@0.5 to 96.7%, it also meets the ultra-real-time end-to-end early warning requirements of high-speed belt conveyors, demonstrating promising prospects for industrial applications. Attached Figure Description

[0055] Figure 1 A flowchart of a belt tear detection method based on full-morphology perception and multi-scale reconstruction provided in an embodiment of the present invention;

[0056] Figure 2 This is a schematic diagram of the overall structure of the full-morphology perception target detection model based on the improved YOLO26 architecture provided in an embodiment of the present invention.

[0057] Figure 3 A schematic diagram of the internal structure of the multi-branch anisotropic feature reconstruction operator Omni_G_Block and the anisotropic phantom reconstruction module provided in the embodiments of the present invention;

[0058] Figure 4 A structural block diagram of a belt tear detection system based on full-morphology perception and multi-scale reconstruction provided in an embodiment of the present invention;

[0059] Figure 5 The diagram shows a comparison of the detection accuracy of various types of defects and the overall accuracy before and after model improvement, as provided in the embodiments of the present invention. Detailed Implementation

[0060] To make the intended purpose, technical means, and effects of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0061] It should be noted that, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as those in the technical field to which this invention pertains. Methods in the following embodiments that do not specifically specify experimental conditions employ conventional techniques within that technical field. It should be understood that the described embodiments are only a part of the examples of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on these embodiments of this invention without creative effort are within the protection scope of this invention.

[0062] Example 1

[0063] To address the limitations of existing deep learning object detection algorithms in industrial conveyor belt tear detection, such as the inability to effectively extract the geometric skeleton of defects with extreme aspect ratios due to the isotropic limitations of feature extraction operators and loss functions, the high likelihood of missing small targets, and the susceptibility of models to reduced generalization ability due to harsh environmental noise and enhanced dirty labels, this invention proposes a conveyor belt tear detection method and system based on full-morphology perception and multi-scale reconstruction. This method innovatively introduces an anisotropic phantom reconstruction mechanism for morphological decoupling of low-level features and designs a full-morphology perception geometric loss to adaptively provide directional gradient excitation and noise resistance, thereby achieving high-precision and robust detection of full-morphology conveyor belt tears with extremely low computational power. Figure 1 As shown, the belt tear detection method based on full-morphology perception and multi-scale reconstruction provided in this embodiment runs on a deep learning workstation or edge computing device equipped with a high-performance CPU or GPU. The specific steps are as follows:

[0064] Step S1: During the operation of the belt conveyor, the original images and videos of the belt surface are collected in real time by industrial cameras installed at key nodes of the belt; based on the spatial geometric characteristics of the tear, the belt surface tears are classified into five types of defects: obvious short tears, fine short tears, obvious long tears, fine long tears, and irregular tears.

[0065] Step S1.1: Using industrial cameras deployed at key nodes of the belt conveyor, acquire raw images of the torn belt surface during the operation of the belt conveyor in real time, and process them to a uniform resolution.

[0066] Step S1.2: Based on the spatial geometric characteristics of the tear, the surface tear of the belt is strictly divided into five types of defects: obvious short tear, fine short tear, obvious long tear, fine long tear and irregular tear.

[0067] Step S2: The original images acquired in Step S1 are preprocessed and augmented using a dual-space expansion enhancement strategy. These images are then combined with the original images to construct a dataset covering all types of belt tear defects. Bounding boxes are annotated into the dataset using a labeling tool, and the dataset is divided into training and validation sets.

[0068] Step S2.1: Perform pixel space enhancement. Explicit geometric and photometric transformations are performed at the image pixel level, specifically including random rotation, flipping, adding Gaussian noise, and simulating blurring and brightness / contrast adjustments to mimic industrial environment conditions, in order to simulate the physical shake and sensor noise of industrial cameras.

[0069] Step S2.2: Perform latent space augmentation. Using the original image as a guiding condition, a diffusion model is used to generate variants in the latent space to achieve semantic-level augmentation. In this embodiment, the Stable Diffusion model is used, with a denoising intensity of 0.25, to generate belt crack variants with unknown texture variations, rare lighting conditions, and different coal dust covering materials.

[0070] Step S2.3: Merge the original and enhanced images of the belt tear to construct a dataset, and use annotation tools to annotate the bounding boxes of the dataset. Considering the long-tailed distribution of the belt tear data, to ensure the model is adequately trained, the final dataset is divided into training and validation sets in a 9:1 ratio.

[0071] Step S3: Construct a full-morphological perception target detection model based on the improved YOLO26 architecture; the model sequentially includes: a backbone network for multi-scale feature extraction, a multi-branch anisotropic feature reconstruction operator embedded in the feature pyramid fusion stage for deep feature morphological decoupling and reconstruction, and a detection head for outputting defect category probability, confidence level and predicted bounding box position coordinates.

[0072] Step S3.1: Using a lightweight convolutional network architecture, the preprocessed image is downsampled and feature extracted layer by layer from bottom to top, and multi-scale initial feature maps containing shallow surface texture details and deep strong semantic information are output sequentially.

[0073] Step S3.2: Embed the multi-branch anisotropic feature reconstruction operator Omni_G_Block and extract intrinsic features. Receive the multi-scale initial feature maps and perform cross-layer cascading and feature fusion; wherein, in the deep feature fusion stage, the multi-branch anisotropic feature reconstruction operator is embedded. This operator introduces the anisotropic phantom reconstruction mechanism pioneered in this invention, abandoning the traditional lightweight network approach of blindly generating redundant phantom features using symmetric isotropic convolution.

[0074] Let the deep feature map input to this operator be... First, using Convolution extracts intrinsic semantic features To achieve channel dimensionality reduction and cross-channel information integration:

[0075]

[0076] In the formula, express Convolution operation, For batch normalization, It is a non-linear activation function.

[0077] Step S3.3: Input the intrinsic features in parallel into two mutually orthogonal long-range strip probe branches for polymorphic mapping.

[0078] use Strip convolution captures long-range dependencies along the horizontal direction, generating horizontal phantom features. To ensure feature diversity and significantly reduce the number of parameters, grouped convolution is used, expressed as:

[0079]

[0080] In the formula, The SiLU activation function is used. The strip convolution span is preferably set to 7 in this embodiment; The number of groups is determined by the greatest common divisor of the number of intrinsic channels and the number of horizontal phantom channels;

[0081] use Strip convolution captures long-range dependencies along the vertical direction, generating vertical phantom features. The expression is:

[0082]

[0083] In the formula, The SiLU activation function is used. The strip convolution span is preferably set to 7 in this embodiment; The number of groups is determined by the greatest common divisor of the number of intrinsic channels and the number of vertical phantom channels;

[0084] Step S3.4: Intrinsic features The phantom features from the two directions mentioned above are directly spliced ​​and fused along the channel dimension to form the full morphological receptive field. Cross-channel information is then converged through a transitional convolutional layer, and residual connections are introduced to form the final output of the operator. :

[0085]

[0086] Step S4: Input the training dataset described in step S2 into the full-morphology perception target detection model, use full-morphology perception geometric loss to calculate the comprehensive error between the predicted bounding box and the labeled bounding box, and perform backpropagation and parameter update until the model converges to obtain the trained detection model.

[0087] Step S4.1: When calculating the bounding box regression loss, a custom empty shell network module is introduced to truncate and block the gradient backpropagation of the original distributed focus loss, so that it always returns a zero tensor consistent with the target shape, thus ensuring the purity of the full morphological perception geometric loss during backpropagation.

[0088] Step S4.2: Set the center coordinates and width / height of the predicted bounding box as follows: The center coordinates and width and height of the bounding box are marked as follows: The square of the diagonal distance between the two smallest bounding rectangles is The square of the Euclidean distance between the center points is Based on the area of ​​the bounding box. Set a threshold for the area of ​​small targets. In this embodiment, it is preferably set to For each pixel, the loss is calculated adaptively in two branches.

[0089] Step S4.3: Regarding the area For regular and slender targets, calculate the anisotropic phantom penalty term. With dynamic focus factor regression loss .

[0090] Extract the width and height of the predicted bounding box and the labeled bounding box, and calculate the horizontal perception weights for each. and vertical perception weights :

[0091]

[0092] Simultaneously, the adaptive adjustment coefficient of the aspect ratio of the labeled bounding box is calculated. And perform an upper limit truncation operation on it. In this embodiment, it is preferable to set a maximum value limit. The value is set to 4.0 to prevent gradient explosion caused by extreme aspect ratios.

[0093]

[0094] Based on this, an anisotropic phantom penalty item is constructed. :

[0095]

[0096] Using outlier measure

[0097]

[0098] Finally, the bounding box regression loss for this branch is calculated as follows:

[0099]

[0100] In the formula, To predict the intersection-union ratio (IU) between the bounding box and the labeled bounding box, ϵ is a smoothed minimum value to prevent division by zero. For dynamic focusing adjustment parameters, it is preferably set to 1.9 in this embodiment.

[0101] Step S4.4: Regarding the area For extremely small targets, calculate the loss based on normalized Wasserstein distance. .

[0102] For extremely small targets To address the problem of gradient vanishing due to its tendency to approach zero, the bounding box is modeled as a two-dimensional Gaussian distribution, and the square of the Wasserstein distance between the predicted and labeled bounding boxes is calculated. :

[0103]

[0104] Normalize it using an exponential function to obtain the NWD similarity measure, and then calculate the regression loss for small targets:

[0105]

[0106] In this embodiment, the value is preferably set to 12.8; the model is based on... and The mask determination automatically summarizes the corresponding losses to complete the backpropagation.

[0107] Step S4.5: Model parameter configuration and training iteration. In this embodiment, the input image size is set to 640×640, the batch size is 64, an adaptive optimizer and cosine annealing learning rate are used, and the total number of training iterations is set to 200. The model updates the weights through backpropagation using the fully morphologically perceptual geometric loss set above, until training converges, and the optimal model weight file is saved.

[0108] Step S5: Decode the video stream of the torn belt surface acquired in real time in step S1 into continuous image frames or directly input a single frame image, input it into the trained detection model for forward inference, output the detection category, confidence level and position coordinates of the predicted bounding box of the belt tear, and trigger an alarm signal in combination with the multi-dimensional joint safety early warning logic.

[0109] Step S5.1: Decode the real-time video stream data collected on site into continuous image frames or directly input a single frame image, input it into the belt tear detection model deployed on the edge computing device for forward inference, and the model outputs the belt tear category, confidence level and coordinates of the predicted bounding box of the current frame.

[0110] Step S5.2: Set specific multi-dimensional joint security warning parameters. In this embodiment, the single-frame extremely high-risk confidence threshold is set to 0.85; the long-side pixel security threshold for slender cracks is 50 pixels; and the continuous detection frame tolerance threshold is 5 frames.

[0111] Step S5.3: When any or a combination of the following conditions are met, the system issues a shutdown or maintenance warning signal: (1) The confidence of a certain torn bounding box detected in a single frame exceeds the preset danger threshold; (2) The pixel length of the detected elongated torn bounding box in its main extension direction exceeds the safety threshold; (3) In the historical cache queue of multiple consecutive frames, based on spatiotemporal coordinate tracking, the cumulative number of torn defects detected at the same spatial location exceeds the set fault tolerance threshold.

[0112] Example 2

[0113] This embodiment provides a belt tear detection system based on full-morphology perception and multi-scale reconstruction. This system is a virtual device for implementing the belt tear detection method described in Embodiment 1, and includes:

[0114] 1. Image acquisition and classification module: used to acquire original images of the torn belt surface and real-time video of the belt conveyor, and classify the belt surface tears into five defect types based on spatial geometric features: obvious short tears, fine short tears, obvious long tears, fine long tears and irregular tears;

[0115] 2. Data Augmentation Module: This module is used to preprocess and expand the pixel space and latent space of the acquired original images using a dual-space expansion augmentation strategy, and to construct a full-morphology belt tear defect dataset together with the original images, and to annotate the belt tears in the dataset with bounding boxes.

[0116] 3. Full-morphology target detection module: It has a built-in full-morphology perception target detection model based on the improved YOLO26 architecture, which includes a backbone network for multi-scale feature extraction, a multi-branch anisotropic feature reconstruction operator embedded in the feature pyramid fusion stage for deep feature morphology decoupling and reconstruction, and a detection head for outputting predicted bounding boxes, detection categories and confidence scores.

[0117] 4. Model Training and Loss Calculation Module: This module inputs the training dataset into the full-morphology object detection module, uses full-morphology perception geometric loss to calculate the comprehensive error between the predicted bounding box and the labeled bounding box, and performs backpropagation and parameter updates to obtain the trained detection model.

[0118] 5. Inference and Joint Early Warning Module: This module decodes the real-time video stream into continuous image frames or directly inputs the images into the trained detection model for forward inference, outputs the detection category, confidence level, and location coordinates of belt tearing, and triggers alarm signals in conjunction with multi-dimensional joint safety early warning logic.

[0119] The above modules work together. The image acquisition and classification module transfers the collected data to the data augmentation module. The processed data is fed into the full-morphology target detection module to extract features. During the training phase, the model training and loss calculation module optimizes the parameters. Finally, during the inference phase, the inference and joint early warning module outputs the results and executes the alarm action.

[0120] Those skilled in the art should understand that the above division of system modules is merely for clarity of description. In practical applications, the functions of the above modules can be merged into one or more hardware entities, or the functions of a certain module can be further broken down into multiple sub-modules to complete the task.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0122] in conclusion

[0123] In summary, this invention proposes a belt tear detection method and system based on full-morphological perception and multi-scale reconstruction. The detection network of this method possesses extremely lightweight and high-precision perception characteristics. The basic process involves constructing a robust dataset based on a dual-space expansion enhancement strategy, performing full-morphological feature decoupling using an improved YOLO26 model based on an embedded anisotropic phantom reconstruction mechanism, adaptive training based on full-morphological perception geometric loss, and security early warning based on multi-dimensional joint logic. This effectively avoids the structural missed detections and inaccurate localization problems caused by traditional isotropic operators and loss functions.

[0124] First, this invention constructs a multi-branch anisotropic feature reconstruction operator, Omni_G_Block, and introduces an anisotropic phantom reconstruction mechanism. Utilizing long-range strip probes to orthogonally regenerate features, it not only extracts the complete geometric skeleton of slender tears with extremely low parameter cost but also overcomes the structural limitations of traditional networks in long-range dependencies, enabling the model to accurately perceive defects with extreme aspect ratios. Second, this invention constructs a full-morphology-aware geometric loss. By introducing anisotropic phantom penalty terms, it provides adaptive directional gradient excitation for slender targets. Simultaneously, it effectively compensates for gradient loss in early, extremely small cracks through normalized Wasserstein distance, achieving accurate regression of defects across all scales. Third, this invention specifically incorporates a blocking mechanism and a dynamic focusing factor to effectively suppress negative gradients caused by noise from harsh industrial environments and edge-blurred dirty labels generated by diffusion model enhancement, significantly improving the model's robustness and generalization ability in complex contexts. Finally, this invention constructs a multi-dimensional joint safety early warning logic, which combines single-frame high-risk confidence, long-side pixel safety threshold and spatiotemporal coordinate continuous tracking, effectively filtering out occasional false alarms caused by dust and reflection, and comprehensively improving the reliability of the alarm system.

[0125] The proposed belt tear detection method, trained and tested on a self-built dataset, demonstrates that it significantly improves the detection accuracy for five complex morphological defects: obvious short tears, minute short tears, obvious long tears, minute long tears, and irregular tears. The model has a low parameter count of approximately 1.75 M and a computational cost of only 7.1 GFLOPS. On an NVIDIA RTX 4080, it achieves a pure inference speed of 764.4 FPS and an end-to-end real-time inference speed of 308.1 FPS, greatly reducing the computational power threshold and memory consumption in industrial settings. Simultaneously, the average detection accuracy (mAP@0.5) reaches 96.7%, meeting the ultra-real-time end-to-end early warning requirements for high-speed belt conveyors and demonstrating promising industrial application prospects.

[0126] It will be readily understood by those skilled in the art that, without conflict, the above-mentioned preferred solutions can be freely combined and superimposed.

[0127] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A belt tear detection method based on full-morphology perception and multi-scale reconstruction, characterized in that, Includes the following steps: Step S1: Acquire images of the tear on the surface of the conveyor belt and real-time operation video. Based on the spatial geometric characteristics of the defects, classify the tear images into five types of defects: obvious short tears, fine short tears, obvious long tears, fine long tears, and irregular tears. Step S2: The acquired original image is preprocessed and expanded in pixel space and latent space using a dual-space expansion enhancement strategy. The original image is then combined with the image to construct a full-morphology belt tear defect dataset. Boundary boxes are labeled for belt tears in the dataset, and the dataset is divided into a training set and a validation set. Step S3: Construct a full-morphological perception target detection model based on the improved YOLO26 architecture. In the feature pyramid fusion stage, use the multi-branch anisotropic feature reconstruction operator to perform morphological decoupling and reconstruction of deep features. Step S4: Full-morphology perception refers to the ability of the detection model to simultaneously extract features, locate boundaries, and identify categories for belt tear defects of different scales, aspect ratios, and textures, such as obvious short tears, fine short tears, obvious long tears, fine long tears, and irregular tears. This capability is mainly achieved through the multi-branch anisotropic feature reconstruction operator and full-morphology perception geometric loss. In this step, the full-morphology dataset used for training is input into the model, and the full-morphology perception geometric loss is used to calculate the comprehensive error between the predicted bounding box and the labeled bounding box, and backpropagation and parameter updates are performed. Step S5: Decode the real-time acquired video stream into continuous image frames or directly input the images into the trained detection model for forward inference, output the detection category, confidence level and location coordinates, and trigger an alarm signal in combination with multi-dimensional joint security warning logic.

2. The belt tear detection method based on full-morphology perception and multi-scale reconstruction according to claim 1, characterized in that: Step S3 introduces an anisotropic phantom reconstruction mechanism using a multi-branch anisotropic feature reconstruction operator. This mechanism utilizes mutually orthogonal long-range strip probe branches to capture long-range dependencies along the horizontal and vertical directions, generating phantom features with direction-aware capabilities. The process is as follows: use Convolution extracts intrinsic semantic features ;Will The parallel inputs are fed into two mutually orthogonal long-distance strip probe branches, one horizontal and one vertical, and then processed using... and Strip convolution generates horizontal phantom features With vertical phantom features The final output is obtained through channel splicing and residual connection. : in, express convolution, This indicates a splicing operation. To input a deep feature map, The span of the strip convolution is set according to the specific target, and is preferably set to 7. This indicates that a splicing operation is performed along the channel dimension.

3. The belt tear detection method based on full-morphology perception and multi-scale reconstruction according to claim 1, characterized in that: The full-morphological sensing geometric loss described in step S4 is calculated based on the area of ​​the labeled bounding box. With small target area threshold The determination result is divided into the following two adaptive calculations: (1) For regular and slender targets, calculate the anisotropic phantom penalty term. With dynamic focus factor regression loss : Let the center coordinates and width and height of the predicted bounding box be... The center coordinates and width and height of the bounding box are marked as follows: The square of the diagonal distance between the two smallest bounding rectangles is The square of the Euclidean distance between the center points is To enable the loss function to apply additional directional gradient penalties based on the specific direction of the belt crack's extension, the width and height of the predicted bounding box and the labeled bounding box are extracted, and the horizontal sensing weights are calculated accordingly. and vertical perception weights : Simultaneously, the adaptive adjustment coefficient of the aspect ratio of the labeled bounding box is calculated. And it performs an upper limit truncation operation, with the maximum value limited to 4.

0. Based on this, an anisotropic phantom penalty item is constructed. : Using outlier measure Finally, the bounding box regression loss for this branch is calculated as follows: In the formula, To predict the intersection-union ratio of the bounding box and the labeled bounding box, ϵ is a smoothed minimum value to prevent division by zero; The dynamic focusing adjustment parameter is set according to the specific target, and is preferably set to 1.9; (2) For extremely small targets, calculate the loss based on normalized Wasserstein distance. : For extremely small targets To address the problem of gradient vanishing due to its tendency to approach zero, the bounding box is modeled as a two-dimensional Gaussian distribution, and the square of the Wasserstein distance between the predicted and labeled bounding boxes is calculated. : Normalize it using an exponential function to obtain the NWD similarity measure, and then calculate the regression loss for small targets: The value is set according to specific objectives, with a preferred setting of 12.8; the model is based on... and The mask determination automatically summarizes the corresponding losses to complete the backpropagation.

4. A belt tear detection system based on full-morphology perception and multi-scale reconstruction, characterized in that, include: The image acquisition module acquires raw images of the torn belt surface and real-time video of the belt conveyor, and classifies the tears into five categories based on spatial geometric features: obvious short tears, fine short tears, obvious long tears, fine long tears, and irregular tears. The data augmentation module is used to preprocess and expand the pixel space and latent space of the acquired original images using a dual-space expansion augmentation strategy, and to jointly construct a full-morphology belt tear defect dataset with the original images; the full-morphology target detection module has a built-in full-morphology perception target detection model based on the YOLO26 architecture, which includes a backbone network for multi-scale feature extraction, a multi-branch anisotropic feature reconstruction operator embedded in the feature pyramid fusion stage for deep feature morphology decoupling and reconstruction, and a detection head for outputting predicted bounding boxes and class probabilities. The model training module is used to input the training dataset into the full-morphology target detection module, calculate the comprehensive error between the predicted bounding box and the labeled bounding box using full-morphology perception geometric loss, and perform backpropagation and parameter updates to obtain the trained detection model. The inference and early warning module is used to input the real-time acquired video stream or image into the trained detection model for forward inference, output the detection category and location coordinates of the belt tear, and trigger an alarm signal in combination with multi-dimensional joint safety early warning logic.

Images