A conveyor belt laser vision tear detection system and method

Abstract

The present application belongs to the technical field of industrial detection, and particularly relates to a conveying belt laser vision tear detection system and method, comprising a layered storage architecture to construct a dynamically updated cross-model conveying belt basic database; a working condition adaptive dynamic enhancement model is designed to perform data enhancement processing on the constructed cross-model conveying belt basic database; an improved YOLOv13 vision tear detection neural network model is constructed to realize cross-model conveying belt laser vision tear intelligent detection; and a multi-dimensional three-level evaluation index system is established based on the vision tear intelligent detection result. The present application effectively improves the range, intelligent degree and detection accuracy of the conveying belt laser vision tear detection, and realizes digital quantifiable evaluation of the conveying belt tear degree.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial detection, and particularly relates to a conveying belt laser vision tearing detection system and method. BACKGROUND

[0002] As the core equipment for continuous transportation of bulk materials, conveying belts are widely used in industrial fields such as mines, ports, power and metallurgy. In actual working conditions, the conveying belt is prone to tearing failure due to factors such as material impact, foreign object scratches and roller wear, which not only causes equipment damage, but also may cause production interruption and environmental pollution. With the development of conveying belts towards "wide width, high speed and heavy load", the traditional manual inspection efficiency is low (about 0.5 km / h), the recognition accuracy is less than 60% in harsh working conditions, and false and missed judgments occur frequently. Therefore, developing efficient and accurate automatic detection technology has become the key to ensuring production safety.

[0003] Current conveying belt tearing detection technologies mainly include contact type and non-contact type. The contact type technology obtains state information by contact between the detection element and the conveying belt, and is widely used in early stage, mainly including pressure sensor type and steel wire rope core electromagnetic detection. The pressure sensor type technology arranges a pressure sensor array under the conveying belt, and the pressure is stable under normal working conditions. When tearing occurs, the pressure drops suddenly due to material leakage, so as to judge the fault. At present, some domestic short-distance conveying belts still use this technology, and the sensor response time is about 50 ms. The steel wire rope core electromagnetic detection technology is used for steel wire rope core conveying belts with a proportion of more than 60%, which uses the principle of electromagnetic induction to capture the magnetic field distortion caused by the breaking of steel wire rope through the excitation coil and Hall sensor, and can detect steel wire rope defects with a diameter of ≥0.5 mm.

[0004] The non-contact type technology realizes non-contact detection through optical and acoustic means, and has become a research hotspot in recent years, mainly including four types: the infrared thermal imaging technology uses the temperature difference of the tearing part to capture the temperature field image through the infrared thermal imager, identifies the high temperature area to determine the tearing, and the detection distance can reach 8-10 meters; the ultrasonic wave technology is based on the penetration and reflection characteristics of sound waves, and captures the sound wave attenuation caused by tearing through the transmitter and receiver, which can detect conveying belt defects with a thickness of ≥5 mm, and the accuracy is about ±2 mm; the ordinary machine vision technology uses an industrial camera with a pixel of 200-500 million to obtain images, and combines with algorithm to identify tearing features. Some systems cover wide conveying belts through multi-camera splicing, and the recognition accuracy is about 85%; the early laser vision technology introduces a line laser projection module to form a continuous laser stripe, calculates three-dimensional coordinates through image acquisition by an industrial camera and the principle of triangulation, and judges tearing by stripe breaking or deviation. At present, it is still in the research and development stage and has not been industrialized on a large scale.

[0005] From the industry application, the prior art also has three common bottlenecks: first, the detection range is limited, the detection capability is low, and it is mostly suitable for specific type or working condition of the conveyor belt; second, the intelligent degree is low, and it can only judge whether it is torn, and cannot divide the defect level and predict the failure trend; third, the system integration is poor, and there is lack of linkage with the conveyor belt control system, and manual shutdown is required after detection, and the delayed response is easy to expand the failure. These defects restrict the detection reliability, and also provide a direction for the optimization and upgrading of the laser vision detection technology.

[0006] A conveyor belt longitudinal tear visual detection system is disclosed in the prior art CN120235835A, which includes dynamically adjusting the window structure and the cleaning device by real-time monitoring of dust concentration, mud and water adhesion and illumination conditions, using a logical judgment algorithm, effectively reducing environmental interference, improving image acquisition quality, and ensuring the clarity of 3D image data; introducing a multiple detection mechanism and a secondary confirmation strategy, using a deep learning model to fuse historical data to construct a time series model, and optimizing the tear prediction result through an attention mechanism, finally generating a tear risk assessment index Tlzs, improving detection accuracy and reducing false alarm rate; supporting TCP / IP, Modbus Tcp and webSocket protocols, converting data formats uniformly and visualizing display, and realizing centralized management of multiple devices. However, the multiple detection mechanism and the secondary confirmation strategy used in this method result in high complexity of the system, and the response speed and detection efficiency are greatly reduced. At the same time, this invention only supports longitudinal tear detection, and the detection range and capability are insufficient. In the prior art CN112258398A, a conveyor belt longitudinal tear detection device and method based on TOF and binocular image fusion are disclosed, which includes a central processing unit, a data storage module, an image processing module, and an image acquisition module between the upper and lower belts of the conveyor belt; the power module is connected to the central processing unit, the image processing module is connected to the image acquisition module, and the central processing unit is connected to the data storage module and the image processing module; the image acquisition module transmits the collected images to the image processing module, and then transmits the processed data to the central processing unit; the central processing unit fuses the TOF parallax map and the binocular parallax map to calculate the depth map and analyze whether the conveyor belt is torn longitudinally according to the depth map, and the invention can more effectively detect the longitudinal tear of the conveyor belt. However, the intelligent degree of this invention is low, and the detection range and detection capability are also relatively insufficient. SUMMARY

[0007] To solve the above problems, the present application provides a conveyor belt laser vision tear detection method applied to a conveyor belt laser vision tear detection system to realize multi-type and strong intelligent conveyor belt laser vision tear detection. To achieve the above scheme, the present application provides the following technical solutions:

[0008] A conveyor belt laser vision tear detection method, comprising the following steps:

[0009] S1: Construct a dynamic updated cross-model conveyor belt database using a hierarchical storage architecture, including a laser stripe reference feature layer, a working condition interference feature layer, and a tear sample library layer;

[0010] S2: Design a working condition adaptive dynamic enhancement model to perform data enhancement processing on the cross-model conveyor belt database constructed in step S1;

[0011] S3: Construct an improved YOLOv13 visual tear detection neural network model using the cross-model conveyor belt database after data enhancement in step S2 as the data source, and complete model training;

[0012] S4: Based on the visual tear intelligent detection results in step S3, establish a multi-dimensional three-level evaluation index system to quantitatively reflect the impact of tears on the operation of the conveyor belt.

[0013] Further, the step S1 comprises:

[0014] S11: Design a laser stripe reference feature layer, and for each material conveyor belt, continuously collect m frame laser image data under non-interference working conditions, and after removing random noise by mean filtering, extract the laser stripe gray distribution , stripe continuity parameters , and three-dimensional profile reference values ; the calculation expression for removing random noise by mean filtering is:

[0015]

[0016] wherein, is the gray distribution of the filtered laser image, i.e., the reference gray value after removing random noise; is the original gray distribution of the laser stripe reference feature layer before filtering, which is directly collected by an industrial camera and contains sensor noise and environmental light interference random noise; x , y are pixel coordinates of the laser image, x axis corresponds to the width direction of the conveyor belt, y axis corresponds to the running direction of the conveyor belt; i , j are the offset amounts of the neighborhood pixel coordinates; m × n is the neighborhood size of the mean filtering;

[0017] S12: Design a working condition interference feature layer, covering three types of interference in industrial scenes, i.e., dust interference, light interference, and vibration interference, with 5 gradients for each type of interference, and collecting nFrame distortion samples, label interference type and intensity, construct interference-response mapping relationship, as subsequent anti-interference algorithm training data;

[0018] S13: Design tear sample library layer, according to "type-size-material" three-dimensional classification, cover longitudinal tear, transverse tear, irregular tear, each combination of l Frame laser three-dimensional image, while labeling tear position, geometric parameters and corresponding conveyor belt running parameters;

[0019] A preferred embodiment, step S11 described the non-interference working condition for dust concentration , light intensity 500-1000lux, vibration amplitude ≤0.5mm.

[0020] A preferred embodiment, step S11 described the three-dimensional profile reference value is calculated by triangulation method, unit: millimeter.

[0021] A preferred embodiment, step S12 described the 5 gradient respectively corresponding to: dust concentration, ; light intensity, 100 / 500 / 2000 / 5000 / 10000lux; vibration amplitude, 0.5 / 1 / 2 / 3 / 5mm.

[0022] Further, the step S2 comprises:

[0023] S21: Light interference simulation enhancement, through the sine function to simulate light fluctuation, solve the traditional light enhancement only superimposed fixed intensity noise, can't simulate the periodic change of light in the running of the conveyor belt, the calculation expression is:

[0024]

[0025] Wherein, k is the light fluctuation coefficient, is the light period, calculated by the conveyor belt speed v And camera frame rate f , the formula is: ;

[0026] S22: According to the "locality" and "randomness" characteristics of dust shielding, the way of combining Gaussian noise superposition and local pixel shielding is adopted, and the calculation expression is:

[0027]

[0028] Wherein, is the Gaussian noise, is the mean, is the variance, is the interference weight, replace the pixel value of this area with the mean value of the neighborhood;

[0029] S23: Multimodal data fusion, fusion of laser three-dimensional data and conveyor belt running parameters, construct high-dimensional feature vector, formula is:

[0030]

[0031] Wherein, is the laser three-dimensional coordinate; is the stripe continuity; v t is the displacement of the conveyor belt in the detection time window; is the load correction value.

[0032] Further, the step S3 comprises:

[0033] S31: Adaptive laser parameter adjustment module design, design real-time adaptive laser parameter adjustment strategy for different types of conveyor belts, specifically including:

[0034] S311: Design laser line number and angle adjustment strategy, according to the belt width W Dynamic determination of the number of laser lines n , the calculation expression is:

[0035]

[0036] Wherein, is the upward rounding function; is the single laser coverage width;

[0037] The projection angle is calculated through geometric relationship , the calculation expression is:

[0038]

[0039] Wherein, H is the laser emitter installation height, the projection angle is calculated and the laser emitter angle is automatically adjusted through the stepping motor after the calculation is completed;

[0040] S312: Based on the real-time speed of the conveyor belt v , dynamically adjust the frame rate of the industrial camera f , so that the displacement of the conveyor belt in the adjacent frame image is less than or equal to 0.5 P Avoid motion blur, the calculation expression is:

[0041]

[0042] Wherein, P is the camera pixel accuracy, the real-time speed of the conveyor belt v is obtained through the encoder acquisition;

[0043] ​S32: Optimize and improve the YOLOv13 visual tearing detection neural network model from the aspects of feature extraction, fusion and detection head, specifically including:

[0044] S321: Embed CBAM attention mechanism modules after the 3 scale feature maps output by the Backbone of YOLOv13, strengthen the tearing features through channel attention and spatial attention, and suppress interference information, and the calculation expression is:

[0045]

[0046] Among them, is a channel attention map, is a spatial attention map; and the final output enhanced feature ;

[0047] S322: In the feature fusion layer of YOLOv13, the existing PANet only realizes one-way feature fusion, and has poor adaptability to multi-scale tearing targets; the present application uses BiFPN to replace PANet, and through bidirectional cross-scale connection and weighted feature fusion from top to bottom and from bottom to top, the feature fusion effect of small-scale tearing and large-scale tearing is enhanced, and the calculation expression of weighted feature fusion is:

[0048]

[0049] Among them, is the multi-scale feature layer fused by the feature fusion layer of YOLOv13; is a weight coefficient of a single feature map to be fused, which is obtained by model self-adaptive learning and is used to quantify the importance of different feature maps; is a small constant, which is used to avoid calculation abnormality caused by denominator tending to 0;

[0050] S323: Decoupled detection head design, the classification and regression shared head of the existing YOLOv13 is separated into independent branches, the classification branch adopts Softmax activation function and the regression branch adopts CIoU loss function, and the calculation formula of CIoU loss function is:

[0051]

[0052] Among them, is the center distance between the predicted box and the real box, c is the diagonal line length of the bounding box, is a balance coefficient, v is a length-width ratio consistency parameter;

[0053] S33: The improved YOLOv3 visual tear detection neural network model of step S32 is trained on the enhanced cross-model conveyor belt database of step S2, and the model weight is saved;

[0054] S34: The model weight saved in step S33 is loaded, and based on the real-time image data collected under the adaptive laser parameter adjustment strategy in step S31, a two-step method of anomaly screening and model reasoning is used for tear feature matching and identification, which specifically includes:

[0055] S341: Laser stripe anomaly detection, calculate real-time stripe continuity deviation from the reference value The calculation expression is:

[0056]

[0057] When , T L is the conveyor belt material adaptation threshold, marked as a suspected tear area, reducing the subsequent model reasoning range;

[0058] S342: Three-dimensional profile anomaly determination, calculate real-time three-dimensional coordinates deviation from the three-dimensional profile reference value The calculation expression is:

[0059]

[0060] When , T Z is a % of the thickness of the conveyor belt, determined as a profile anomaly, superimposed with the stripe anomaly area, the intersection area is retained, and the candidate tear area is further screened;

[0061] S343: Inference based on the improved YOLOv3 visual tear detection neural network model, the tear candidate area determined in steps S341 to S342 is cropped and scaled and input into the model for inference, and the tear type, bounding box and confidence are output.

[0062] Further, the step S4 includes:

[0063] S41: Based on the cross-model conveyor belt laser visual tear intelligent detection result of step S3, a three-level tear grade evaluation index system is established, and the index definition and design includes:

[0064] S411: First-level index tear geometric parameters, including length L , width W , depth DThe boundary box output by step S343 and the laser three-dimensional coordinate data The calculation result is

[0065] S412: Secondary index tear propagation speed v s The calculation result is obtained by using the sliding window method, and the calculation expression is

[0066]

[0067] Among them, , The tear length and width change in the window are respectively The time interval is

[0068] S413: Tertiary index conveyor running parameters, including load F And running speed v The running speed v Is collected by the encoder; the greater the load and the higher the speed, the greater the possibility of tear propagation;

[0069] S42: Calculate the tear grade score by weighting and summing the tertiary tear grade evaluation indexes designed in step S41 S total The calculation expression is

[0070]

[0071] Among them, The weight coefficient is calibrated by AHP combined with historical fault data; S geo The geometric parameter score is , The maximum tear parameter allowed by the conveyor is The full score is equal to S ext The running parameter score is , The critical propagation speed is c The full score is S oper The running parameter score is , F max , v max The rated parameter of the conveyor is The full score is equal to

[0072] S43: Grade according to the score calculated in step S42

[0073] Stotal ≤ d Grade I slight tear, no impact on the safety of the conveyor belt structure, no need to stop;

[0074] d S total ≤ e Grade II moderate tear, there is slow expansion, but it will not break in the short term, ready to stop at any time for maintenance;

[0075] S total ≥ e Grade III severe tear, tear quickly expands or has damaged the core structure of the conveyor belt, immediately triggers a shutdown to avoid an accident.

[0076] In another aspect, a conveyor belt laser vision tear detection system is provided, which is applied to any of the conveyor belt laser vision tear detection methods, and the conveyor belt laser vision tear detection system comprises:

[0077] A laser data acquisition unit: the laser data acquisition unit is composed of a line laser emitter with multi-line output function, an industrial area array camera matched with a lens and a light supplement module, a dust sensor, an infrared temperature sensor, an encoder and a weighing sensor, and realizes laser data acquisition;

[0078] A data processing and inference unit: the core of the data processing and inference unit is an edge computing server, which is equipped with a CPU and a GPU card to meet the computing power requirement; an industrial-grade solid state disk is provided for storing algorithms and models;

[0079] A monitoring and management platform: the hardware of the monitoring and management platform is an industrial computer matched with a high-definition display, a network switch connected with each unit, a data server for historical data storage, and is equipped with a keyboard and a mouse input device, and can be externally connected with a printer, supports access to an enterprise local area network through an Ethernet, and realizes a remote access function.

[0080] Compared with the prior art, the conveyor belt laser vision tear detection system has the following beneficial effects:

[0081] 1. Improved YOLOv13 and improved Transformer are adopted, combined with multi-modal data enhancement, which can accurately identify multiple types of tears and give early warning, solve the problems of target missed detection, inaccurate detection and prediction lag of traditional methods, and adapt to complex working conditions.

[0082] 2. The data acquisition, processing, control and monitoring modules are integrated, the hardware selection is suitable for industrial scenes, supports real-time linkage and remote management, avoids manual intervention delay, and improves the convenience of tear detection and disposal. BRIEF DESCRIPTION OF DRAWINGS

[0083] Figure 1 ​A flow chart of a conveyor belt laser vision tear detection method of the present application.

[0084] Figure 2 A structure diagram of a conveyor belt laser vision tear detection system of the present application.

[0085] Figure 3 A structure diagram of an improved YOLOv3 neural network model in the present application. DETAILED DESCRIPTION

[0086] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0087] In the description of the present application, the terms "first", "second" are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0088] In the description of the present application, the term "for example" is used to indicate "as an example, illustration or description". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given in order to enable any person skilled in the art to implement and use the present application. In the following description, details are listed for the purpose of explanation. It should be understood that those skilled in the art can realize the present application without using these specific details. In other examples, well-known structures and processes will not be described in detail in order to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the shown embodiments, but is consistent with the broadest scope of principles and characteristics disclosed in the present application.

[0089] In the embodiments of the present application, a conveyor belt laser vision tear detection method, which can be specifically referred to the accompanying drawings Figure 1 , specifically includes:

[0090] S1: a layered storage architecture is used to construct a dynamic updated cross-type conveyor belt basic database, including a laser stripe reference feature layer, a working condition interference feature layer and a tear sample library layer;

[0091] S2: design a working condition adaptive dynamic enhancement model to perform data enhancement processing on the cross-model conveyor belt basic database constructed in step S1;

[0092] S3: construct an improved YOLOv13 visual tear detection neural network model based on the cross-model conveyor belt basic database after data enhancement in step S2, which can be seen in the attached Figure 3 , and complete model training;

[0093] S4: based on the visual tear intelligent detection result in step S3, establish a multi-dimensional three-level evaluation index system to quantitatively reflect the impact of the tear on the operation of the conveyor belt.

[0094] Referring to the attached Figure 2 , the application also provides a conveyor belt laser visual tear detection system, which comprises a laser data acquisition unit, a data processing and reasoning unit, and a monitoring and management platform.

[0095] The laser data acquisition unit is composed of a line laser emitter with multi-line output function, an industrial area array camera with lens and light supplement module, a dust sensor, an infrared temperature sensor, an encoder, and a weighing sensor, which realizes laser data acquisition.

[0096] The core of the data processing and reasoning unit is an edge computing server, which is equipped with a multi-core CPU and a GPU card to meet the computing power requirement; an industrial-grade solid-state hard disk is provided for storing algorithms and models.

[0097] The hardware of the monitoring and management platform is an industrial computer with a high-definition display, a network switch connected to each unit, a data server for historical data storage, and a keyboard and mouse input device, which can be connected to a printer, supports access to an enterprise local area network through Ethernet, and realizes remote access function.

[0098] In the specific embodiment, to solve the problems of limited detection range, low detection accuracy, low intelligence level, and poor system integration of the existing conveyor belt laser visual tear detection method and system, a conveyor belt laser visual tear detection system and method are proposed by a conveyor belt tear detection engineer. First, a dynamic updated cross-model conveyor belt basic database is constructed using a hierarchical storage architecture, including a laser stripe reference feature layer, a working condition interference feature layer, and a tear sample library layer. Second, a working condition adaptive dynamic enhancement model is designed to perform data enhancement processing on the constructed cross-model conveyor belt basic database. Third, an improved YOLOv13 visual tear detection neural network model is constructed based on the data enhanced cross-model conveyor belt basic database, the model training is completed, and cross-model conveyor belt laser visual tear intelligent detection is realized. Finally, based on the visual tear intelligent detection result, a multi-dimensional three-level evaluation index system is established to quantitatively reflect the impact of the tear on the operation of the conveyor belt.

[0099] Embodiment 1

[0100] In one embodiment, to solve the problems of low intelligence level, insufficient detection accuracy and easy to miss detection in the current conveyor belt laser vision tear detection system and method, the conveyor belt tear detection engineer proposes a dynamic updating cross-model conveyor belt basic database construction and data enhancement method, which specifically includes:

[0101] S11: Design a laser stripe reference feature layer. For each material conveyor belt, under the conditions of no interference, i.e. dust concentration , light intensity 500-1000 lux, vibration amplitude ≤0.5 mm, continuously collect 1000 frames of laser image data, remove random noise by mean filtering, and extract laser stripe gray distribution , stripe continuity parameters and three-dimensional profile reference value ; the calculation expression of mean filtering to remove random noise is:

[0102]

[0103] wherein, is the gray distribution of the filtered laser image, i.e. the reference gray value after removing random noise; is the original gray distribution of the laser stripe reference feature layer before filtering, which is directly collected by an industrial camera and contains sensor noise and environmental light interference random noise; x , y are pixel coordinates of the laser image, x axis corresponds to the width direction of the conveyor belt, y axis corresponds to the running direction of the conveyor belt; i , j are the offset amounts of the neighborhood pixel coordinates; m × n is the neighborhood size of mean filtering;

[0104] S12: Design a working condition interference feature layer, covering three types of interference in industrial scenes, i.e. dust interference, light interference and vibration interference, with 5 gradients for each type of interference, i.e. dust concentration, ; light intensity, 100 / 500 / 2000 / 5000 / 10000 lux; vibration amplitude, 0.5 / 1 / 2 / 3 / 5 mm; collect 500 frames of distortion samples for each gradient, label the interference type and intensity, and construct the interference-response mapping relationship as the basis for training data for subsequent anti-interference algorithms;

[0105] S13: Design a tear sample library layer, classified in three dimensions according to "type-size-material", covering longitudinal tears, transverse tears, and irregular tears. For each combination, acquire 200 frames of laser three-dimensional images, and label the tear location, geometric parameters, and corresponding conveyor belt operation parameters.

[0106] S14: Enhanced simulation of illumination interference. This uses a sine function to simulate illumination fluctuations, addressing the limitations of traditional illumination enhancement methods that only superimpose fixed-intensity noise and cannot simulate the periodic changes in illumination during conveyor belt operation. The calculation expression is as follows:

[0107]

[0108] in, k The light fluctuation coefficient, The illumination period is determined by the conveyor belt speed. v With camera frame rate f The calculation formula is as follows: ;

[0109] S15: To address the "locality" and "randomness" characteristics of dust occlusion, a combination of Gaussian noise superposition and local pixel occlusion is used. The calculation expression is as follows:

[0110]

[0111] in, It is Gaussian noise. The mean, For variance, As a disturbance weight, the pixel values ​​in this region are replaced with the neighborhood mean.

[0112] S16: Multimodal data fusion, fusing laser 3D data with conveyor belt operating parameters to construct a high-dimensional feature vector, the formula is:

[0113]

[0114] in, The laser's three-dimensional coordinates; For stripe continuity; v × t To detect conveyor belt displacement within the detection time window; This is the load correction value.

[0115] Example 2

[0116] In one embodiment, to address the issues of low intelligence and inaccurate detection accuracy in current conveyor belt laser vision tear detection systems and methods, conveyor belt tear detection engineers propose a conveyor belt laser vision tear detection method based on an improved YOLOv13 neural network model. Details can be found in the appendix. Figure 3, the specific steps include:

[0117] S31: The adaptive laser parameter adjustment module is designed, which designs a real-time adaptive laser parameter adjustment strategy for different types of conveying belts, specifically including:

[0118] S311: Designing a laser line number and angle adjustment strategy, according to the conveying belt bandwidth W Dynamically determine the number of laser lines n , the calculation expression is:

[0119]

[0120] wherein, is the ceiling function; is the single laser coverage width;

[0121] The projection angle is calculated through geometric relationship , the calculation expression is:

[0122]

[0123] wherein, H is the laser emitter installation height, and the laser emitter angle is automatically adjusted through the stepper motor after the projection angle calculation is completed;

[0124] S312: Based on the real-time speed of the conveying belt v , dynamically adjust the frame rate of the industrial camera f , so that the displacement of the conveying belt in adjacent frame images is less than or equal to 0.5 P Avoid motion blur, the calculation expression is:

[0125]

[0126] wherein, P is the camera pixel accuracy, and the real-time speed of the conveying belt v is obtained through the encoder acquisition;

[0127] S32: The YOLOv13 vision tearing detection neural network model is optimized and improved from the aspects of feature extraction, fusion, and detection head, specifically including:

[0128] S321: Embedding a CBAM attention mechanism module after the 3 scale feature maps output by the Backbone of YOLOv13, the feature map scales are 10x10, 20x20, and 40x40 respectively, the tearing features are strengthened through channel attention and spatial attention, and the interference information is suppressed, the calculation expression is:

[0129]

[0130] wherein, For channel attention map, For spatial attention map; final output enhanced feature ;

[0131] S322: In the feature fusion layer of YOLOv13, the existing PANet only realizes one-way feature fusion, and the adaptability to multi-scale tearing targets is poor; the BiFPN is used to replace the PANet, the feature fusion effect of small-scale tearing and large-scale tearing is enhanced through bidirectional cross-scale connection and weighted feature fusion from top to bottom and from bottom to top, and the calculation expression of weighted feature fusion is:

[0132]

[0133] Wherein, is the multi-scale feature layer fused by the feature fusion layer of YOLOv13; is the weight coefficient of a single feature map to be fused, which is obtained by model adaptive learning and is used to quantify the importance of different feature maps; is a small constant, which is used to avoid calculation abnormality caused by denominator approaching to 0;

[0134] S323: decoupled detection head design, the classification and regression shared head of the existing YOLOv13 are separated into independent branches, the classification branch adopts Softmax activation function, and the regression branch adopts CIoU loss function, and the calculation formula of CIoU loss function is:

[0135]

[0136] Wherein, is the center distance between the predicted frame and the real frame, c is the diagonal line length of the bounding box, is a balance coefficient, v is a length-width ratio consistency parameter;

[0137] S33: the improved YOLOv13 visual tearing detection neural network model of step S32 is trained on the enhanced cross-model conveyor belt database of step S2, and the model weight is saved;

[0138] S34: load the model weight saved in step S33, based on the real-time image data collected under the adaptive laser parameter adjustment strategy in step S31, adopt the two-step method of abnormal screening and model reasoning for tearing feature matching and identification, which specifically includes:

[0139] S341: laser stripe anomaly detection, calculate the continuity of real-time stripe and the deviation of the reference value , the calculation expression is:

[0140]

[0141] When , mark as suspected tear area, reduce subsequent model inference range, T L The threshold value is adapted to the material of the conveyor belt, and the value of the steel wire core conveyor belt is T L = 10 pixels, canvas T L = 8 pixels, nylon T L = 6 pixels;

[0142] S342: Three-dimensional profile anomaly determination, calculate real-time three-dimensional coordinates and the deviation of the three-dimensional profile reference value , the calculation expression is:

[0143]

[0144] When , T Z The thickness of the conveyor belt is a %, and the profile anomaly is determined, which is superimposed with the stripe abnormal area, and the intersection area is reserved. Further filter the candidate tear area;

[0145] S343: Inference based on improved YOLOv13 visual tear detection neural network model, crop and scale the tear candidate area determined in steps S341 to S342 to 640x640 pixels, and input the model inference. Output tear type, bounding box and confidence.

[0146] Embodiment 3

[0147] In one embodiment, in order to overcome the defect that the current conveyor belt laser visual tear detection system and method do not have tear grade evaluation, the conveyor belt tear detection engineer establishes a three-level tear grade evaluation index system, which specifically includes:

[0148] S41: Based on the cross-model conveyor belt laser visual tear intelligent detection result in step S3, a three-level tear grade evaluation index system is established, and the index definition and design includes:

[0149] S411: First-level index tear geometric parameter, weight configuration is 0.5, including length L , width W , depth D , calculated by the bounding box and laser three-dimensional coordinate data output in step S343;

[0150] S412: Second-level index tear expansion speedv s , the weight configuration is 0.3, and the sliding window method is used to calculate, and the calculation expression is:

[0151]

[0152] Among them, , respectively, the tear length, width change amount in the window, is the time interval, which is set to 0.2;

[0153] S413: The three-level index conveyor belt running parameter, the weight configuration is 0.2, including load F and running speed v , the running speed v is collected by the encoder; the greater the load and the higher the speed, the greater the possibility of tear propagation;

[0154] S42: Calculate the tear grade score by weighting and summing the three-level tear grade evaluation indexes designed in step S41 S total , the calculation expression is:

[0155]

[0156] Among them, is the weight coefficient, which is calibrated by the analytic hierarchy process combined with historical failure data; S geo is the geometric parameter score, with a full score of 80, and the calculation expression is , is the maximum tear parameter allowed by the conveyor belt, which is equal to the full score value; S ext is the running parameter score, with a full score of 20, and the calculation expression is , is the critical propagation speed, c is the full score value; S oper is the running parameter score, with a full score of 20, and the formula is , F max , v max is the rated parameter of the conveyor belt, which is equal to the full score value;

[0157] S43: Grade according to the score calculated in step S42;

[0158] S total ≤ dFor Ⅰ level slight tear, no impact on the safety of the conveyor belt structure, no need to stop;

[0159] d S total ≤ e For Ⅱ level moderate tear, there is slow expansion, but it will not break in the short term, ready to stop at any time for maintenance;

[0160] S total ≥ e For Ⅲ level severe tear, the tear expands quickly or has damaged the core structure of the conveyor belt, immediately triggers a stop to avoid an accident.

[0161] Although preferred embodiments of the application have been described, those skilled in the art will appreciate that other alterations and modifications to these embodiments can be made without departing from the spirit and scope of the application. Therefore, it is intended that the appended claims encompass all such alterations and modifications as fall within the scope of the application.

[0162] It is apparent that those skilled in the art can make various changes and modifications to the application without departing from the spirit and scope of the application. Thus, if these modifications and changes fall within the scope of the claims and their equivalents, the application is intended to include them.​

Claims

1. A laser vision tear detection method for conveyor belts, characterized in that, Includes the following steps: S1: A hierarchical storage architecture is used to build a dynamically updated cross-model conveyor belt basic database, including a laser stripe reference feature layer, an operating condition interference feature layer, and a tear sample library layer. S2: Design working condition adaptive dynamic enhancement model, which performs data enhancement processing on the cross-type conveyor belt basic database constructed in step S1; S3: Using the cross-model conveyor belt basic database after data augmentation in step S2 as the data source, construct an improved YOLOv13 visual tear detection neural network model and complete model training; S4: Based on the visual tear intelligent detection results in step S3, establish a multi-dimensional three-level evaluation index system to quantitatively reflect the impact of tearing on conveyor belt operation; Step S3 includes: S31: Design a real-time adaptive laser parameter adjustment strategy for different conveyor belt models; S32: Optimize and improve the YOLOv13 visual tear detection neural network model from three aspects: feature extraction, fusion, and detection head; S33: Train the improved YOLOv13 visual tear detection neural network model from step S32 on the enhanced cross-model conveyor belt basic database from step S2, and save the model weights. S34: Load the model weights saved in step S33, and based on the real-time image data collected under the adaptive laser parameter adjustment strategy in step S31, use the two-step method of anomaly screening and model inference to perform tear feature matching and identification. Step S32 includes: S321: Embed the CBAM attention mechanism module after the three scale feature maps output by the YOLOv13 Backbone, respectively, to enhance tearing features and suppress interference information through channel attention and spatial attention; S322: In the feature fusion layer of YOLOv13, BiFPN is used to replace PANet, and weighted features are fused through bidirectional cross-scale connections from top to bottom and bottom to top. S323: Design a decoupled detection head, separating the existing YOLOv13 classification and regression shared head into independent branches. The classification branch uses the Softmax activation function, and the regression branch uses the CIoU loss function.

2. The method for laser visual tear detection of conveyor belts according to claim 1, characterized in that, Step S1 includes: S11: Design a laser stripe reference feature layer to continuously collect data under interference-free operating conditions for each type of conveyor belt material. m After removing random noise from frame laser image data, the grayscale distribution of laser stripes is extracted. , stripe continuity parameter and three-dimensional contour reference value ; S12: Design condition interference feature layer, marking interference type and intensity, and constructing interference-response mapping relationship; S13: Design a tear sample library layer, categorized into three dimensions: "type-size-material," covering longitudinal tears, transverse tears, and irregular tears. Samples will be collected for each combination. l A frame of laser 3D image, simultaneously labeled with tear location, geometric parameters, and corresponding conveyor belt operating parameters.

3. The laser vision tear detection method for conveyor belts according to claim 1, characterized in that, Step S2 includes: S21: Enhanced simulation of illumination interference, using a sine function to simulate illumination fluctuations; S22: To address the "locality" and "randomness" characteristics of dust occlusion, a combination of Gaussian noise superposition and local pixel occlusion is adopted; S23: Multimodal data fusion, which integrates laser 3D data with conveyor belt operating parameters to construct a high-dimensional feature vector.

4. The laser vision tear detection method for conveyor belts according to claim 1, characterized in that, Step S4 includes: S41: Based on the intelligent laser vision tear detection results of cross-type conveyor belts in step S3, establish a three-level tear level assessment index system; S42: Calculate the tear level score by weighted summation of the three-level tear level assessment indexes designed in step S41. S total ; S43: Classify the grades based on the scores calculated in step S42.

5. The laser vision tear detection method for conveyor belts according to claim 4, characterized in that, Step S41 includes: S411: Level 1 tear geometry parameters, including length L ,width W ,depth D ; S412: Secondary indicator tearing expansion speed v s The result was obtained using the sliding window method. S413: Level 3 Indicator Conveyor Belt Operating Parameters, including load F With running speed v running speed v Data collected by the encoder.

6. A laser vision tear detection system for conveyor belts, characterized in that, The conveyor belt laser vision tear detection method according to any one of claims 1-5, wherein the conveyor belt laser vision tear detection system comprises: The laser data acquisition unit consists of a line laser emitter with multi-line output function, an industrial area array camera with lens and supplementary light module, a dust sensor, an infrared temperature sensor, an encoder and a weighing sensor. The data processing and inference unit is centered on an edge computing server, equipped with a CPU and GPU to meet computing power requirements; it is also equipped with an industrial-grade solid-state drive for storing algorithms and models. The monitoring and management platform consists of an industrial computer with a high-definition monitor, a network switch connecting each unit, a data server for storing historical data, a keyboard and mouse input device, an external printer, and supports remote access via Ethernet to the enterprise LAN.

Images