YOLOv8-MSDA network model and conveyor belt damage monitoring method

Abstract

The present invention discloses a YOLOv8-MSDA network model and a conveyor belt damage monitoring method, which relates to the field of image monitoring technology, and in particular to image monitoring of conveyor belts. The model is improved based on the YOLOv8 network model, which is conducive to improving the accuracy and precision of detection. The method steps include establishing belt damage image data; establishing belt damage segmentation image data; building YOLOv8-MSDA network model detection training; based on DeepLabv3+ network model segmentation training; belt damage visual monitoring system data collection; input into the YOLOv8-MSDA network model detection after processing, and output the result; obtain the risk level assessment result according to the test result evaluation; update information according to the risk level assessment result and make processing. The method detects and monitors the belt damage risk level in real time and warns on the basis of meeting the accuracy and detection precision, and makes corresponding processing.

Description

Technical Field

[0001] The invention relates to the technical field of image monitoring, in particular to image monitoring of a conveyor belt. Background Art

[0002] Belt conveyors are indispensable equipment in mines, ports and power companies. They are used to efficiently transport bulk materials (such as blocks and powders) and can achieve uninterrupted transportation of large volumes of materials at high speeds and long distances at low cost. However, in the long-term operation process, with the continuous increase in material load and operation cycle and other unexpected factors, the belts are severely worn and damaged. During this period, accidents such as longitudinal tearing of the belts often occur. Such accidents will have a huge impact on the normal operation of the company's production, the stability of equipment and the safety of personnel.

[0003] At present, there are related solutions and technologies for the safe operation and life monitoring of belts, such as a machine vision-based intelligent safety protection system for belt conveyors disclosed in Application No. 201710454337.3. In this case, when the belt conveyor is unloaded, machine vision technology is used to detect whether there are cracks in the belt image, and the image is pre-processed by filtering and enhancing; the image is binarized by using iterative adaptive threshold segmentation method; and the morphological algorithm is used to remove small area noise, so as to complete the detection and identification of crack targets. In this scheme, the belt conveyor is mainly detected in the unloaded state, and the traditional threshold-based segmentation processing algorithm is greatly affected by environmental changes. Therefore, the problem with this scheme is that it will be seriously disturbed by the environment during use, and is limited to the unloaded state of the belt conveyor and the detection of belt cracks in this state.

[0004] Another example is a belt conveyor belt monitoring system and method based on image recognition disclosed in application number 202210432580.6. In this scheme, image recognition is used to detect the uniformity and irregularity of belt loads. By constructing a belt life prediction model of image recognition rules and influencing factors, the prediction model is corrected and the belt life is predicted based on historical data and belt life prediction deviation data sets. At the same time, corresponding early warnings are made based on the set belt life safety indicators. This scheme mainly indirectly judges the impact of the load on the belt life by detecting the belt load. It is difficult to determine the measurement standard of the specific degree of impact. In addition, the coefficients of various influencing factors in constructing the belt life prediction model are difficult to determine. The parameter error has a greater impact on the results of the life prediction model, and cannot directly reflect the health status of the conveyor belt. Summary of the invention

[0005] The purpose of the present invention is to provide a YOLOv8-MSDA network model based on the improvement of the YOLOv8 network model, which is conducive to improving the accuracy and detection precision of the belt damage type detection. Another purpose of the present invention is to provide a conveyor belt damage monitoring method, through which the belt damage risk level can be monitored and warned in real time on the basis of meeting the detection accuracy, and the warning function of the belt damage risk level can be realized according to the wear state, tear length, damage area and other indicators.

[0006] To achieve the above purpose, the technical solution of the present invention is: an improved YOLOv8-MSDA network model, which is improved based on the YOLOv8 network model, and the improvement includes the C2f module in the head part of the YOLOv8 network model, by adding multi-scale hole attention in the C2f module, using the MSDA structure to sparsely select key points and values ​​in the sliding window, and by controlling the expansion rate, it is possible to select keys and values ​​at different scales to simulate local sparse patch interactions, and use linear layers to aggregate semantic information of different scales to reduce the redundancy of the standard attention mechanism. The specific improvement content is that the Bottleneck structure of the C2f module in the head part of the YOLOv8 network model is replaced by Dila The DilateBlock structure is replaced by the DilateBlock structure. The network data flow is designed in the DilateBlock structure. The input data is transmitted to the MSDA structure after passing through LayerNorm. In the MSDA, the head data is processed by the sliding window dilated attention (SWDA). At the same time, the output feature data is spliced ​​according to the feature dimension in the Concat layer, and then the aggregation of the feature data of each scale is completed through the Linear layer. The result of the Linear layer aggregation is transmitted to the DropPath layer for randomly deleting some connections in the structure to prevent overfitting of the model training. Finally, the DropPath output data and the input data of the LayerNorm layer are added to the feature map and output.

[0007] A method for monitoring damage of a conveyor belt, the monitoring method steps are as follows:

[0008] S1. Establish belt damage image data; establish belt damage segmentation image data; build a YOLOv8-MSDA network model based on the improved YOLOv8 network model, and perform belt damage type detection training on the belt damage image data; perform belt damage area segmentation training on the belt damage segmentation image data based on the DeepLabv3+ network model;

[0009] S2. During the conveyor belt damage detection work, the belt damage visual monitoring system performs real-time image data acquisition;

[0010] S3. The belt damage visual monitoring system decomposes the real-time image data obtained by the camera to obtain a belt image, pre-processes the belt image and inputs it into the YOLOv8-MSDA network model in step S1 to detect the belt damage type, and outputs the detection result; the custom settings of the belt damage type include damaged, undamaged, worn and torn types; when the detection result is undamaged or worn, enter step S4; when the detection result is torn, save the detection frame information of the damaged ROI area to calculate the tear length and then enter step S4; when the detection result is damaged, the YOLOv8-MSDA network model crops the damaged ROI area image and calls the DeepLabv3+ network model in step S1 to segment the damaged area of ​​the belt on the damaged ROI area image to obtain a damaged area image and save it, calculate the damaged area area through the damaged area image and then enter step S4;

[0011] S4. Conduct belt risk level assessment based on the test results and obtain the risk level assessment result;

[0012] S5. Update the warning belt damage risk level information in real time according to the risk level assessment results and take corresponding measures according to the risk level assessment results.

[0013] The loss function for training the belt damage detection model defined in the YOLOv8-MSDA network model includes a classification loss function based on the belt damage type classification and a regression loss function for the belt damage location area.

[0014] The belt damage visual monitoring system collects real-time image data through imaging by a visual device; the visual device is arranged below the inner surface of the conveyor belt and is equipped with a light source for supplementing the shooting environment and a reflective mirror for reflecting the conveyor belt imaging information and transmitting it to the visual device.

[0015] The lens sight line of the visual device is set horizontally, and the installation setting of the reflective mirror is set according to the angle between the conveyor belt and the horizontal angle to set the angle between the reflective mirror and the conveyor belt. The visual device adopts a line array camera, and the working distance from the line array camera to the reflective mirror is set by the pixel equivalent of the line array camera. The pixel equivalent calculation formula of the line array camera is: ,in, They represent the pixel equivalent of the visual system, the imaging field of view width, and the number of photosensitive chip pixels corresponding to the imaging field of view width in this solution. The calculation formula for the corresponding lens focal length of the linear array camera is: ,in, They represent the focal length and the camera working distance respectively; the imaging signal of the linear array camera is set to trigger the linear array camera to take pictures according to the frame signal.

[0016] In the belt damage type detection in step S3, the detection includes determining the damage ROI area in the belt image, obtaining the belt damage type and location of each damage ROI area, and outputting the damage type and location information of each damage position for storage; the tear length is calculated by approximating the longest side of the detection box in the detection box information of the damaged ROI area, and the calculation formula is: , where is the longest side of the detection box, is the working distance from the belt surface to the linear array camera, is the pixel size of the linear array camera, is the imaging focal length of the linear array camera; The actual physical length.

[0017] The damaged area is calculated by obtaining the mask area of ​​the damaged area based on the damaged area image and combining it with the actual pixel equivalent in the image to calculate the actual damaged area. The calculation formula is: , where Image mask area; is the physical size of the damaged area, is the working distance from the belt surface to the linear array camera, is the pixel size of the line array camera.

[0018] The belt damage risk level assessment in step S4 is divided into the following levels: when the detection result is wear or no damage, the risk level is assessed to be safe; when the detection result is damage or tear, the assessment is performed based on the calculation results of the damage area or tear length. For example, in the test example of this embodiment, when the damage area is not greater than Or the tear length is not greater than , it is judged as low risk; when the damaged area is greater than and not greater than Or the tear length is greater than and not greater than , it is judged as medium risk; when the damaged area is greater than Or the tear length is greater than , determined to be high risk.

[0019] The preprocessing of the belt image in step S1 includes filtering and equalizing the belt image; and / or, the establishment of the belt damage image data includes collecting conveyor belt damage defect samples at the work site of the long-distance conveyor belt machine, collecting belt damage images in various states, and also including collecting various belt damage images of some belts on the Internet, and then using a labeling tool to mark the damage position and type of each collected belt damage image, wherein the damage position and type marking is to mark the damage position and damage type of belt images with different damage types according to the target detection method, and at the same time, the belt damage image with the damage type of broken belt is semantically segmented and marked with the belt damage area.

[0020] By adopting the above technical scheme, the beneficial effects of the present invention are as follows: the above monitoring method is constructed and trained by improving the fusion multi-scale void attention mechanism based on YOLOv8 to obtain the YOLOv8-MSDA network model; by collecting and annotating the belt skin damage area segmentation image data set, a belt damaged area segmentation model based on the DeepLabv3+ network is trained, which is used to further perform segmentation processing on the damaged area when the damage type detection result is damaged; the visual monitoring system is triggered by the frame signal to collect the image data of the bottom of the conveyor belt and input it into the belt damage type detection model, which can improve the accuracy and detection precision of the belt damage type detection.

[0021] The next step of processing is carried out according to the model detection results. When the detection result is no damage or the damage type is wear, it is directly input into the risk level assessment model; if the detection result is tearing, the tear length is approximately calculated according to the tear area detection frame, and the tear length calculation result is input into the risk level assessment model; if the detection result is damage, the belt damage ROI area image is cropped according to the damaged area detection frame, and the damaged area is segmented by inputting the damaged ROI area image into the belt damaged area segmentation model based on DeepLabv3+, and the damaged area area is calculated according to the area segmentation result, and the calculation result is input into the risk level assessment model; according to the assessment result of the risk level assessment model, real-time warning is issued, and the processing strategy is automatically generated according to the risk level, so as to realize the warning function of the level of belt damage risk and realize the intelligent monitoring and early warning of the operation status of long-distance conveyor belts, which is particularly suitable for monitoring and operation and maintenance of the operation status of long-distance belt conveyors in smart mine construction.

[0022] In summary, the above-mentioned purpose of the present invention is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The invention discloses a flow chart of a conveyor belt damage monitoring method.

[0024] Figure 2This is a schematic diagram of the structure of the C2f module defined in the YOLOv8 network model.

[0025] Figure 3 It is a schematic diagram of the structure of the C2f module defined in the improved YOLOv8-MSDA network model.

[0026] Figure 4 This is a schematic diagram of the Bottleneck structure in the C2f module of the YOLOv8 network model.

[0027] Figure 5 This is a schematic diagram of the DilateBlock structure in the C2f module of the improved YOLOv8-MSDA network model.

[0028] Figure 6 The figure is a schematic diagram of the visual imaging layout structure of the belt damage visual monitoring system in the present invention within a narrow space away from the conveyor belt. DETAILED DESCRIPTION

[0029] In order to further explain the technical solution of the present invention, the present invention is described in detail below through specific embodiments.

[0030] A conveyor belt damage monitoring method, such as Figure 1 The flow chart shows the following monitoring steps:

[0031] S1. Pre-establishment includes S1.1. Establishing belt damage image data; S1.2. Establishing belt damage segmentation image data; S1.3. Building a YOLOv8-MSDA network model based on the improved YOLOv8 network model, and performing belt damage type detection training on the belt damage image data, and saving the training results; S1.4. Based on the DeepLabv3+ network model, performing belt damage area segmentation training on the belt damage segmentation image data, and saving the training results;

[0032] The establishment of the belt damage image data in the above step S1.1 can be achieved by collecting conveyor belt damage defect samples at the work site of the long-distance conveyor belt machine, and collecting belt damage images in various states. When collecting samples, it is necessary to pay attention to that images of various defect states should be collected in place to ensure the balance of samples as much as possible. In order to expand the data set, multiple belt damage images of some belts can also be collected on the Internet at the same time, and then the collected belt damage images are annotated with the damage position and type through the annotation tool. Here, the damage position and type are annotated for annotating the belt images with different damage types according to the target detection method (ROI area detection) and then saved as a belt damage image data file for network model training in step S1.3. At the same time, the belt damage image with the damage type of broken is annotated with the belt damage area semantic segmentation. Here, the belt damage area semantic segmentation is annotated for cropping the belt damage area ROI area for semantic segmentation and annotation, and then saved as a belt damage segmentation image data file for network model training in step S1.4.

[0033] The improvement of the YOLOv8-MSDA network model in the above step S1.2 based on the YOLOv8 network model includes the C2f module in the head part of the YOLOv8 network model, by adding multi-scale void attention in the C2f module, using the MSDA structure to sparsely select key points and values ​​in the sliding window, and by controlling the expansion rate, it is possible to select keys and values ​​at different scales to simulate local sparse patch interactions, and use linear layers to aggregate semantic information of different scales to reduce the redundancy of the standard attention mechanism. The above improvements can improve the ability to aggregate semantic information at different scales, so as to better capture multi-scale features and improve the target detection accuracy and detection precision of the model.

[0034] Specifically, the Bottleneck structure of the C2f module in the head part of the YOLOv8 network model (such as Figure 2 ) is replaced with a DilateBlock structure (as shown Figure 3 As shown in Figure 2), the Bottleneck structure in the C2f module of the YOLOv8 network model is as follows Figure 4 As shown in the figure, the input data is processed by two CBS structures and then added to the original data and output. The DilateBlock structure in the C2f module of the improved YOLOv8-MSDA network model is as follows: Figure 5As shown in the figure, the network data flow is designed in the DilateBlock structure, and the input data is transmitted to the MSDA structure after passing through LayerNorm. In the MSDA, each head data is processed by sliding window dilated attention (SWDA), and the output feature data is spliced ​​according to the feature dimension in the Concat layer. Then, the aggregation of feature data of each scale is completed through the Linear layer, and the result of the Linear layer aggregation is transmitted to the DropPath layer for randomly deleting some connections in the structure to prevent overfitting of the model training. Finally, the DropPath output data and the input data of the LayerNorm layer are added to the feature map and output.

[0035] The loss function for training the belt damage detection model defined in the YOLOv8-MSDA network model improved by S1.3 above includes a classification loss function based on the belt damage type and a regression loss function for the belt damage location area.

[0036] S2. In the long-distance conveyor belt damage detection work, a belt damage visual monitoring system based on linear array camera imaging is used to collect real-time image data. In order to accurately obtain the health status of the conveyor belt during the load-bearing operation, the belt damage visual monitoring system must be set below the long-distance conveyor belt loading section. Due to the small internal space of the upper and lower belts of the long-distance conveyor belt machine, the working space and imaging distance are limited. The existing single area array camera or video camera cannot provide a large field of view to achieve coverage in the width direction of the conveyor belt. To solve this problem, the present invention uses reflective mirror reflection to transmit the conveyor belt imaging information to the linear array camera, and realizes the transformation of the camera imaging direction, so as to obtain a suitable camera working distance, reduce the imaging field of view angle, and reduce imaging distortion. The specific structure is as follows: Figure 6 As shown, according to the structural design of the long-distance conveyor belt machine combined with the visual imaging layout structure, the imaging field angle required for the working distance and belt width of the linear array camera is calculated. For example, in the test example of this embodiment, the conveyor belt width is 1000mm. According to the design of the long-distance belt working surface and the horizontal angle α=30°, the design of the angle β between the reflective mirror surface and the belt surface is 45°. In order to improve the detection accuracy of conveyor belt tearing, this embodiment adopts the Hikvision robot MV-CL043-A1GC linear array camera, and its pixel equivalent is calculated as follows: ,in, They represent the pixel equivalent of the visual system, the imaging field of view width, and the number of photosensitive chip pixels corresponding to the imaging field of view width in this solution. The pixel equivalent is calculated to be 0.244mm / pixels. According to the internal space structure design of the long-distance conveyor belt, the working distance from the camera to the reflective mirror is set to 300mm. The corresponding lens focal length is calculated as follows: ,in, Respectively represent the focal length and the camera working distance. After calculation, the focal length of the visual imaging system is about 8.6mm. According to the lens options provided by existing manufacturers, the LensC11-0824-12M-P industrial lens is selected, which has a focal length of 8.5mm and a final imaging field of view width of 1012mm. Due to the small internal space of the conveyor belt machine equipment and the dim environment, in this embodiment, a strip light source with a length of 1m is used to maintain a 30° angle with the horizontal direction to achieve fill light for the visual monitoring system at the bottom of the conveyor belt, ensuring the stability of the lighting environment for the visual system imaging. According to the design of the visual monitoring system in this embodiment, this embodiment can achieve full coverage of the conveyor belt with a width of 1000mm by the linear array camera at a working distance of 300mm. Through the above setting, the information of the bottom of the conveyor belt is transmitted to the linear array camera through the reflective mirror, and the conversion of the camera imaging direction in the visual system is realized to obtain the ideal camera working distance, which can ensure the imaging quality of the visual system and the image information is not distorted.

[0037] In this embodiment, the corresponding line frequency parameters are set according to the principle of the linear array camera vision system and the running speed of the conveyor belt to ensure that the image height direction is not stretched or compressed. According to the working characteristics of the conveyor belt, the conveyor belt generally maintains a uniform speed during normal cargo transportation. In this embodiment, the linear array camera imaging signal is set to trigger the linear array camera to take pictures according to the frame signal, thereby ensuring the consistency of the output image format for subsequent processing.

[0038] S3. The belt damage visual monitoring system decomposes the real-time image data obtained by the linear array camera to obtain a belt image, pre-processes the belt image and inputs it into the YOLOv8-MSDA network model in the called step S1 to detect the belt damage type and output the detection result.

[0039] The above-mentioned preprocessing of the belt image includes improving the image quality by filtering and equalizing the belt image. The filtering algorithm suppresses part of the sampling noise. In this embodiment, the filtering algorithm uses a Gaussian filter kernel to perform Gaussian filtering on the belt image to suppress part of the Gaussian noise in the image. The equalization algorithm performs histogram equalization on the belt image to improve the contrast of some pixels in the image, thereby achieving the purpose of belt image enhancement.

[0040] After the above-mentioned belt image is preprocessed, the input is called into the YOLOv8-MSDA network model to perform belt damage type detection. The detection includes determining the damaged ROI area in the belt image, obtaining the belt damage type and location of each damaged ROI area, and outputting the damage type and location information of each damaged position for storage. In this embodiment, the custom settings of damage types include broken, undamaged, worn and torn types. In order to more accurately divide the risk level of belt damage, this embodiment adopts the following different processing strategies according to the damage type to facilitate the subsequent risk level assessment step.

[0041] When the detection result is no damage or wear, go to step S4;

[0042] When the detection result is tearing, the detection frame information of the damaged ROI area is saved to calculate the tear length and then enter step S4. In this embodiment, the tear length is calculated by approximating the longest side of the detection frame in the detection frame information of the damaged ROI area. The calculation formula is: , where is the longest side of the detection box, is the working distance from the belt surface to the linear array camera, is the pixel size of the linear array camera, is the imaging focal length of the linear array camera; The actual physical length.

[0043] When the detection result is damaged, the YOLOv8-MSDA network model cuts out the damaged ROI area image and calls the DeepLabv3+ network model in step S1 to perform segmentation processing on the cropped damaged ROI area image to obtain the damaged area image and save it. After calculating the area of ​​the damaged area through the damaged area image, step S4 is entered. Since the use environment of the long-distance conveyor belt is generally more complicated, the segmentation of the damaged belt area in this embodiment calls the belt damaged segmentation model based on deep learning to process the belt damaged ROI area image. Specifically, the damaged ROI area image cut out from the YOLOv8-MSDA network model for belt damage detection is passed to the DeepLabv3+ network model for segmentation processing of the belt damaged area. Since the average pooling strategy is used in the DeepLabv3+ network model to process input images of different sizes, it can accept input images of any size. Therefore, the damaged ROI area image cut out from the YOLOv8-MSDA network model can be directly transmitted to the belt damaged area segmentation model based on DeepLabv3+ for processing. In this embodiment, the damaged area is calculated by obtaining the mask area of ​​the damaged area based on the damaged area image and combining it with the actual pixel equivalent in the image to calculate the actual damaged area. The calculation formula is: , where Image mask area; is the physical size of the damaged area, is the working distance from the belt surface to the linear array camera, is the pixel size of the line array camera.

[0044] S4. Evaluate the belt risk level according to the test results and obtain the risk level evaluation result. The evaluation method of this embodiment is to evaluate the belt damage risk level according to the damage type test results, tear length and damaged area. In this embodiment, the belt damage risk level evaluation level is specifically divided into: when the test result is wear or no damage, the risk level is evaluated as safe; when the test result is damage or tear, the evaluation is performed based on the calculation results of the damaged area or tear length. For example, in the test example of this embodiment, when the damaged area is not greater than Or the tear length is not greater than , it is judged as low risk; when the damaged area is greater than and not greater than Or the tear length is greater than and not greater than , it is judged as medium risk; when the damaged area is greater than Or the tear length is greater than , it is judged as high risk, and can be expressed as follows:

[0045] .

[0046] S5. Update the warning belt damage risk level information in real time according to the risk level assessment results and make corresponding treatment according to the risk level assessment results. The corresponding treatment here can be a strategy formulation process pre-set according to the risk level assessment results. For example, when the risk is high, the conveyor belt machine should be stopped immediately to repair or replace the belt damage; when the risk is medium, the conveyor belt machine should be stopped and the belt damage should be repaired; when the risk is low, it does not affect the company's production, so it does not need to be stopped immediately, and the belt damage repair operation can be carried out after the belt conveyor belt machine is shut down normally.

[0047] The above embodiments and drawings do not limit the product form and style of the present invention. Any appropriate changes or modifications made by ordinary technicians in the relevant technical field should be deemed to be within the patent scope of the present invention.

Claims

1. A conveyor belt damage monitoring method, characterized in that: The monitoring method steps are as follows: S1. Establish belt damage image data; Establish belt damage segmentation image data; An improved YOLOv8-MSDA network model is included, wherein the YOLOv8-MSDA network model is trained by performing belt damage type detection on belt damage image data; S2. During the conveyor belt damage detection work, the belt damage visual monitoring system performs real-time image data acquisition; S3. The belt damage visual monitoring system decomposes the real-time image data obtained by the camera to obtain a belt image, pre-processes the belt image and inputs it into the YOLOv8-MSDA network model in step S1 to detect the belt damage type, and outputs the detection result; The belt damage type custom settings include broken, undamaged, worn and torn types; When the detection result is no damage or wear, go to step S4; When the detection result is tearing, the detection frame information of the damaged ROI area is saved to calculate the tearing length and then go to step S4; When the detection result is damaged, the YOLOv8-MSDA network model cuts out the damaged ROI area image and calls the DeepLabv3+ network model in step S1 to segment the damaged belt area in the damaged ROI area image to obtain the damaged area image and save it. After calculating the damaged area area through the damaged area image, it proceeds to step S4. S4. Conduct belt risk level assessment based on the test results and obtain the risk level assessment result; S5. Update the warning belt damage risk level information in real time according to the risk level assessment results and take corresponding measures according to the risk level assessment results; The improved YOLOv8-MSDA network model is improved based on the YOLOv8 network model. The improvement includes the C2f module in the head part of the YOLOv8 network model. By adding multi-scale hole attention to the C2f module, the MSDA structure is used to sparsely select key points and values ​​in the sliding window, and the expansion rate is controlled to select keys and values ​​at different scales to simulate local sparse patch interactions. The linear layer is used to aggregate semantic information of different scales to reduce the redundancy of the standard attention mechanism. The specific improvement content is that the Bottleneck structure of the C2f module in the head part of the YOLOv8 network model is modified with DilateBloc k structure replacement, the network data flow is designed in the DilateBlock structure, the input data is transmitted to the MSDA structure after passing through LayerNorm, and each head data in the MSDA is processed by sliding window expansion attention, and the output feature data is spliced ​​according to the feature dimension in the Concat layer, and then the aggregation of feature data of each scale is completed through the Linear layer, and the result of the Linear layer aggregation is transmitted to the DropPath layer for randomly deleting some connections in the structure to prevent overfitting of the model training, and finally the DropPath output data and the input data of the LayerNorm layer are added to the feature map and output; The loss function for training the belt damage detection model defined in the YOLOv8-MSDA network model includes a classification loss function based on the belt damage type classification and a regression loss function for the belt damage location area; The belt damage visual monitoring system collects real-time image data through imaging by a visual device; the visual device is arranged below the inner surface of the conveyor belt and is equipped with a light source for supplementing the shooting environment and a reflective mirror for reflecting the conveyor belt imaging information and transmitting it to the visual device.

2. A conveyor belt damage monitoring method as claimed in claim 1, characterized in that: The lens sight line of the visual device is set horizontally, and the installation setting of the reflective mirror is set according to the angle between the conveyor belt and the horizontal angle to set the angle between the reflective mirror and the conveyor belt. The visual device adopts a line array camera, and the working distance from the line array camera to the reflective mirror is set by the pixel equivalent of the line array camera. The pixel equivalent calculation formula of the line array camera is: ,in, They represent the pixel equivalent of the visual system, the imaging field of view width, and the number of photosensitive chip pixels corresponding to the imaging field of view width in this solution. The calculation formula for the corresponding lens focal length of the linear array camera is: ,in, They represent the focal length and the camera working distance respectively; the imaging signal of the linear array camera is set to trigger the linear array camera to take pictures according to the frame signal.

3. A conveyor belt damage monitoring method as claimed in claim 1, characterized in that: In the belt damage type detection in step S3, the detection includes determining the damage ROI area in the belt image, obtaining the belt damage type and location of each damage ROI area, and outputting the damage type and location information of each damage position for storage; the tear length is calculated by approximating the longest side of the detection box in the detection box information of the damaged ROI area, and the calculation formula is: , where is the longest side of the detection box, is the working distance from the belt surface to the linear array camera, is the pixel size of the linear array camera, is the imaging focal length of the linear array camera; is the actual physical length; And / or, the damaged area is calculated by obtaining the mask area of ​​the damaged area based on the damaged area image and combining it with the actual pixel equivalent in the image to calculate the actual damaged area. The calculation formula is: , where Image mask area; is the physical size of the damaged area, is the working distance from the belt surface to the linear array camera, is the pixel size of the line array camera.

4. A conveyor belt damage monitoring method as claimed in claim 1, characterized in that: The belt damage risk level assessment level in step S4 is divided into: when the detection result is wear or no damage, the risk level is assessed to be safe; when the detection result is damage or tear, the risk level is assessed based on the calculation result of the damage area or tear length.

5. A conveyor belt damage monitoring method as claimed in claim 1, characterized in that: The preprocessing of the belt image in step S1 includes filtering and equalizing the belt image; and / or, the establishment of the belt damage image data includes collecting conveyor belt damage defect samples at the work site of the long-distance conveyor belt machine, collecting belt damage images in various states, and also including collecting various belt damage images of some belts on the Internet, and then using a labeling tool to mark the damage position and type of each collected belt damage image, wherein the damage position and type marking is to mark the damage position and damage type of belt images with different damage types according to the target detection method, and at the same time, the belt damage image with the damage type of broken belt is semantically segmented and marked with the belt damage area.

Images