Automatic belt joint positioning method on crossheading belt

Abstract

The application belongs to the technical field of crossheading belt maintenance, and discloses a belt joint automatic positioning method on a crossheading belt, and the specific technical scheme is as follows: 1. a high-definition linear array camera is used to continuously scan the moving belt to obtain a belt surface scanning image; 2. the scanning image of the belt joint is scored to generate a list; 3. the position of each joint is positioned in real time; 4. a laser range finder is installed in a belt storage bin, and after being automatically started, the subsequent list comparison calculation is performed; 5. the comparison scoring list is compared to determine the position of the cut-off belt, two adjacent joints on the belt are cut off, the two belt joints are connected through a new joint, the belt length change is determined, and the positioning work is completed. The application obtains the image of the belt joint through the high-definition linear array camera, and trains and processes the image, so that the image of each belt joint can be directly displayed, the scoring condition of each belt joint can be obtained, the overall joint state can be more obviously evaluated, and the risk of belt breakage is reduced.

Description

Technical Field

[0001] This invention belongs to the field of roadway belt maintenance technology, specifically relating to an automatic positioning method for belt fasteners applied to roadway belts. Background Technology

[0002] As a key piece of equipment for underground coal mine transportation, long-distance PVG material belts are currently used in large coal mines in western China. Each belt segment is 50-200 meters long, and adjacent belt segments are spliced ​​together using metal buckles. The transportation distance is about 5 kilometers. A belt storage bin is set at the head of the machine, and the moving end of the belt storage bin is connected to a tensioning trolley. The belt is kept in a tensioned state by the traction force of the tensioning trolley.

[0003] After the conveyor belt in the roadway has been running for a period of time, a section of the belt needs to be cut off at the belt cutting point to shorten it. This is usually done manually, selecting sections with older, worn-out fasteners to prevent belt breakage during transport. After cutting off the old belt, the belt storage bin, tensioning trolley, and other equipment return to their initial state. This cutting process is repeated approximately dozens of times per working face, gradually shortening the conveyor belt until coal mining at the working face ends. The entire process takes about six months to a year.

[0004] Currently, during operation, the belt travels at a speed of approximately 5 meters per second, making it difficult to manually determine the wear level of each buckle. When shutting down for maintenance, manually assessing the wear of the belt buckles is time-consuming and labor-intensive. Furthermore, the belt's cyclical movement and subsequent cutting changes its length, making it difficult to pinpoint the real-time position of each buckle. Failure to replace severely worn buckles can easily lead to belt breakage, requiring production shutdowns and resulting in significant economic losses. Therefore, there is an urgent need for an automatic identification and positioning method for belt buckles on long-distance conveyor belts to accurately determine the wear level of each buckle. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides an automatic positioning method for belt buckles on a conveyor belt, which automatically identifies and positions the position of each belt buckle, providing a convenient and quick means of judgment for belt cutting.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is: an automatic positioning method for belt buckles on a roadway belt, the specific steps of which are as follows: 1. A high-definition line scan camera is used to continuously scan the moving belt to obtain images of the belt surface. 2. Score the scanned images of the belt buckles and generate a list; 3. The displacement L of the belt is continuously recorded by the displacement sensor. Based on the reading of the displacement sensor, the reading when the buckle C is found is set to L0, the belt length is set to D, the belt runs, and the current displacement sensor reading is recorded as L1. The distance between the buckle C and the displacement sensor is (L1-L0)%D. Similarly, the position of each buckle is located in real time. 4. Install a laser rangefinder in the tape storage bin to determine the distance between the moving end and the fixed end of the tape storage bin. If the distance between the moving end and the fixed end decreases significantly after each shutdown and before the next startup, it is determined that the tape cutting has been completed during this shutdown, and the subsequent list comparison calculation will be automatically started. 5. Compare the scoring list to determine the location where the belt was cut off.

[0007] Cut off two adjacent buckles on the belt and connect the two belt joints with new buckles to determine the belt length change and complete the positioning work; Determine the rotational displacement of the corresponding buckle, and match the lists before and after cutting one by one to identify the buckles that were cut off on the belt.

[0008] Let array A be the array before the belt is cut, and array B be the array after the belt is cut; Find the minimum value in array A and the two values ​​to its left and right, respectively, as feature group X. Calculate the similarity between feature group X and the five values ​​in a sub-interval of array B. If the sub-interval of array B extends beyond the end of array B, continue calculating from the beginning of array B. When the similarity is highest, determine the position of the corresponding sub-interval of feature group X and array B, which determines the rotation displacement of array B. The specific formula is as follows: d (X,Y) represents the Euclidean distance function between two intervals, where X is a subinterval of length 5 in array A and Y is a subinterval of length 5 in array B. Displacement = argmin id ( X , B [ i , i +∣ X ∣]) Where i is the starting index in array B, |X| is the length of feature group X, and B[i,i+|X|] is the sub-interval in array B starting from i. Then the displacement of AB matching is the value of i when the minimum Euclidean distance is reached, that is, the rotational displacement between B and A is obtained. Adjust the displacement of array B to correspond one-to-one with array A, continue to find the differences between array B and array A, and determine the cutting position of the belt; By comparing the list in array B with the list in array A, the new belt length is determined, and the new positioning is completed.

[0009] In step two, the first artificial intelligence model identifies all buckles on the belt surface, records the front-to-back relationship of the buckles, obtains the buckle scan image, records the reading of the displacement sensor, optimizes the scan image and inputs it into the first depth recognition model for training, monitors whether there are belt buckles on the trained scan image, if there are belt buckles, the corresponding scan image data is recorded and written into the database, and the scan image is saved.

[0010] The second artificial intelligence recognition model determines the degree of detachment on both sides of each buckle on the belt and scores the degree of detachment. The higher the score, the higher the risk of the buckle detaching.

[0011] For each scanned image with a buckle, it is necessary to train the second depth recognition model. First, the edge line of the belt is identified, and then the inner corner vertex lines at both ends of the metal buckle are identified. The pixel difference between the edge line of one side of the belt and the inner corner vertex line on the X-axis is taken as the first pixel difference, and the pixel difference between the edge line of the other side of the belt and the inner corner vertex line on the X-axis is taken as the second pixel difference. The first pixel difference and the second pixel difference are added together to get the score of the belt buckle.

[0012] In step five, if the belt cutting position falls into feature group X, and the minimum similarity also exceeds the preset value, the second smallest value in input A is reselected, and the second smallest value and the two values ​​to its left and right are set as feature group X.

[0013] Compared with the prior art, the specific beneficial effects of this invention are reflected in: I. This invention acquires images of belt buckles using a high-definition line scan camera, and trains and processes the images. The processed images can intuitively display the image of each belt buckle and obtain the score of each belt buckle, providing a more obvious assessment of the overall buckle status and reducing the risk of belt breakage.

[0014] Second, this invention cuts the belt fasteners according to the recommended fastener score, and automatically performs subsequent positioning and tracking without maintenance, which greatly improves the maintenance efficiency of the roadway belt. Attached Figure Description

[0015] Figure 1 This is a schematic diagram showing the operating status of the conveyor belt in the roadway.

[0016] Figure 2 This is an image for identifying belt buckles.

[0017] Figure 3 This is a diagram showing the position of the belt before and after cutting.

[0018] Figure 4 This is a flowchart of the processing of the second-depth recognition model.

[0019] Figure 5 This is a diagram illustrating the scoring positions for the belt buckle. Detailed Implementation

[0020] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0021] like Figure 1 As shown, the automatic positioning method for belt buckles on the conveyor belt is as follows: 1. Install a high-definition line scan industrial camera at the lower position of the belt where there is no material. The high-definition line scan industrial camera is installed downward to capture the non-working surface of the belt to reduce interference from material residue. The moving belt is continuously scanned by the high-definition line scan camera to obtain the image of the belt surface scan. 2. Score the scanned images of the belt buckles and generate a list; 3. The displacement L of the belt is continuously recorded by the displacement sensor. Based on the reading of the displacement sensor, the reading when the buckle C is found is set to L0, the belt length is set to D, the belt runs, and the current displacement sensor reading is recorded as L1. The distance between the buckle C and the displacement sensor is (L1-L0)%D. Similarly, the position of each buckle is located in real time. 4. Install a laser rangefinder in the tape storage bin to determine the distance between the moving end and the fixed end of the tape storage bin. If the distance between the moving end and the fixed end decreases significantly after each shutdown and before the next startup, it is determined that the tape cutting has been completed during this shutdown, and the subsequent list comparison calculation will be automatically started. 5. Compare the scoring list to determine the location where the belt was cut off; Cut off two adjacent buckles on the belt and connect the two belt joints with new buckles to determine the belt length change and complete the positioning work; Determine the rotational displacement of the corresponding buckle, and match the lists before and after cutting one by one to identify the buckles that were cut off on the belt. Let array A be the array before the belt is cut, and array B be the array after the belt is cut; Find the minimum value in array A and the two values ​​to its left and right, respectively, as feature group X. Calculate the similarity between feature group X and the five values ​​in a sub-interval of array B. If the sub-interval of array B extends beyond the end of array B, continue calculating from the beginning of array B. When the similarity is highest, determine the position of the corresponding sub-interval of feature group X and array B, which determines the rotation displacement of array B. The specific formula is as follows: d(X,Y) represents the Euclidean distance function between two intervals, where X is a subinterval of length 5 in array A and Y is a subinterval of length 5 in array B. Displacement = argmin id ( X , B [ i , i +∣ X ∣]) Where i is the starting index in array B, |X| is the length of feature group X, and B[i,i+|X|] is the sub-interval in array B starting from i. Then the displacement of AB matching is the value of i when the minimum Euclidean distance is reached, that is, the rotational displacement between B and A is obtained. Adjust the displacement of array B to correspond one-to-one with array A, continue to find the differences between array B and array A, and determine the cutting position of the belt; By comparing the list in array B with the list in array A, the new belt length is determined, and the new positioning is completed.

[0022] In step two, the first artificial intelligence model identifies all buckles on the belt surface, records the front-to-back relationship of the buckles, obtains the buckle scan image, records the reading of the displacement sensor, optimizes the scan image and inputs it into the first depth recognition model for training, monitors whether there are belt buckles on the trained scan image, if there are belt buckles, records the corresponding scan image data and writes it into the database, and saves the scan image. The second artificial intelligence recognition model determines the degree of detachment on both sides of each buckle on the belt and scores the degree of detachment. The higher the score, the higher the risk of the buckle detaching.

[0023] The AI ​​recognition model design and the judgment of the inner corner vertex of the buckle breakage include two recognition models. The first depth recognition model is used to find each buckle on the belt, and the second depth recognition model is used to finely determine the buckle condition. The two models are processed separately. Since the belt speed is relatively fast, real-time monitoring is required. First, the faster first depth recognition model is used to process all images of the belt and find the buckles. Then, the slower second depth recognition model is used to identify the detailed situation.

[0024] The first deep recognition model: It first uses the open-source YOLO V5 real-time object detection algorithm to find buckles in the scan results of the entire belt. Using specialized annotation tools such as Labelbox or VOTT, it draws bounding boxes for each belt buckle in the image and assigns a Buckle category label, which serves as the model's training data. The training process begins with the model weights obtained from the open source, and then iterates multiple times using the labeled dataset. In each iteration, the model attempts to predict the location and category of the target in the image. The difference between the prediction and the true label is quantified by a loss function. The model's goal is to minimize this loss. Through backpropagation and an optimizer, the model weights are continuously updated to improve prediction accuracy. For each belt image, YOLO V5 scans the entire image at once, predicting multiple possible Buckle bounding boxes. When the confidence level reaches 70% or higher, a buckle is considered to exist in the image.

[0025] The second depth recognition model identifies the detailed shape of the buckle and determines the position of the buckle edge and the inner corner vertex of the crack.

[0026] like Figure 4 As shown, to obtain the buckle shape, a segmentation model based on MobileNetV2 and DeepLabV3+ is used. MobileNetv2 is a lightweight network that requires little memory, has a fast computation speed, and can support the model to run on mobile or embedded devices. DeepLabV3+ is a semantic segmentation framework that combines multi-scale information and can segment image details well. Combining the two models to comprehensively consider segmentation accuracy and speed results in a lightweight and high-precision segmentation model.

[0027] The model input consists of partial images of the left and right sides of the joint, and the output is a segmentation result image of the same size as the input, in which the joint part is displayed in white and the background part is displayed in black.

[0028] The model employs an Encoder-Decoder structure. The Encoder extracts features from the input image and removes redundant information; the Decoder restores these features to pixel-level segmentation results, thus obtaining accurate segmentation. Both the Encoder and Decoder are deep learning models based on Convolutional Neural Networks (CNNs). This structure was chosen because CNNs can automatically extract high-level features of images without requiring manual feature design.

[0029] The Encoder consists of a backbone network and an AtrousSpatial Pyramid Pooling (ASPP) module with dilated convolutions.

[0030] The backbone network is Mobilenetv2, whose main structure is DS Conv (Depthly Separable Convolution). This structure can reduce the number of parameters and computational cost, resulting in a faster and less memory-intensive lightweight model that can run on mobile or embedded devices.

[0031] The main structure of ASPP is ArConv (dilated convolution), which can obtain contextual information, that is, the relationship between each pixel and its surrounding pixels, without increasing the number of parameters and computation, thus improving the recognition accuracy.

[0032] The Decoder combines the high-level and low-level feature maps output by the Encoder. The high-level feature maps lack structural information, while the low-level feature maps lack detailed information. The two complement each other and can improve the accuracy of segmentation boundary lines. All of the above models and structures are open source.

[0033] like Figure 5 As shown, after obtaining the specific shape of the buckle, the leftmost and rightmost pixel positions are the buckle edges, and the leftmost and rightmost concave points of the belt are the buckle positions.

[0034] The X-axis pixel difference between the left edge and the left concave point is the score for the crack on the left side, and the X-axis pixel difference between the right edge and the right concave point is the score for the crack on the right side. The two are added together to get the score for the whole image.

[0035] For each scanned image containing a buckle, a second depth recognition model needs to be trained. First, the edge line of the belt is identified, then the inner corner vertices of both ends of the metal buckle are identified. The pixel difference on the X-axis between the edge line and the inner corner vertex line on one side of the belt is taken as the first pixel difference, and the pixel difference on the X-axis between the edge line and the inner corner vertex line on the other side of the belt is taken as the second pixel difference. The sum of the first and second pixel differences is used as the score for the belt buckle. As shown in Figure 2, the belt buckle has a 35-pixel crack on the left and a 42-pixel crack on the right, so the current buckle score is 77.

[0036] In step five, if the belt cutting position falls into feature group X, and the minimum similarity also exceeds the preset value, the second smallest value in input A is reselected, and the second smallest value and the two values ​​to its left and right are set as feature group X.

[0037] As shown in Figure 3, the belt segment between the two sides with buckle scores of 44 and 81 is cut off. After the cut is completed, the belt buckle is reconnected. The belt buckle with a score of 16 is the newly installed buckle.

[0038] This invention cuts the belt fasteners according to the recommended fastener score, and automatically performs subsequent positioning and tracking without maintenance, which greatly improves the maintenance efficiency of the conveyor belt.

[0039] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the scope of the present invention.

Claims

1. An automatic positioning method for belt buckles on a roadway belt, characterized in that, The specific steps are as follows:

1. A high-definition line scan camera is used to continuously scan the moving belt to obtain images of the belt surface.

2. The first artificial intelligence model identifies all buckles on the belt surface, records the front-to-back relationship of the buckles, obtains the buckle scan image, and records the readings of the displacement sensor. The second artificial intelligence recognition model determines the degree of separation on both sides of each buckle on the belt, scores the degree of separation, and generates a list.

3. Continuously record the belt displacement L using a displacement sensor. Based on the sensor readings, set the reading when buckle C is detected as L0, and the belt length as D. As the belt runs, record the current displacement sensor reading as L1. The distance between buckle C and the displacement sensor is... Similarly, the position of each buckle can be located in real time; 4. Install a laser rangefinder in the tape storage compartment to determine the distance between the moving end and the fixed end of the tape storage compartment. If the distance between the moving end and the fixed end decreases after each shutdown and before the next startup of the tape storage compartment, it is determined that the tape cutting has been completed during this shutdown, and the subsequent list comparison calculation will be automatically started.

5. Compare the scoring list to determine the location where the belt was cut off; Cut off two adjacent buckles on the belt and connect the two belt joints with new buckles to determine the change in belt length and complete the positioning work; Determine the rotational displacement of the corresponding buckle, and match the lists before and after cutting one by one to identify the buckles that were cut off on the belt. Let array A be the array before the belt is cut, and array B be the array after the belt is cut; Find the minimum value in array A and the two values ​​to its left and right, respectively, as feature group X. Calculate the similarity between feature group X and the five values ​​in a sub-interval of array B. If the sub-interval of array B extends beyond the end of array B, continue calculating from the beginning of array B. When the similarity is highest, determine the position of the corresponding sub-interval of feature group X and array B, which determines the rotation displacement of array B. The specific formula is as follows: The Euclidean distance function represents two intervals, where X is a subinterval of length 5 in array A and Y is a subinterval of length 5 in array B. ； Where i is the starting index in array B. Let X be the length of the feature set. If the subinterval in array B starting from i is the minimum Euclidean distance, then the displacement of AB matching is the value of i, which gives the rotational displacement between B and A. Adjust the displacement of array B to correspond one-to-one with array A, continue to find the differences between array B and array A, and determine the cutting position of the belt; By comparing the list in array B with the list in array A, the new belt length is determined, and the new positioning is completed.

2. The automatic positioning method for belt buckles on a grooved belt according to claim 1, characterized in that, In step two, the scanned image is optimized and input into the first depth recognition model for training. The system monitors whether there is a belt buckle on the trained scanned image. If there is a belt buckle, the corresponding scanned image data is recorded and written into the database, and the scanned image is saved. For each scanned image with a buckle, it is necessary to train the second depth recognition model. First, the edge line of the belt is identified, and then the inner corner vertex lines at both ends of the metal buckle are identified. The pixel difference between the edge line of one side of the belt and the inner corner vertex line on the X-axis is taken as the first pixel difference, and the pixel difference between the edge line of the other side of the belt and the inner corner vertex line on the X-axis is taken as the second pixel difference. The first pixel difference and the second pixel difference are added together to get the score of the belt buckle.

3. The automatic positioning method for belt buckles on a grooved belt according to claim 1, characterized in that, In step five, if the belt cutting position falls into feature group X, and the minimum similarity also exceeds the preset value, the second smallest value in array A is reselected, and the second smallest value and the two values ​​to its left and right are set as feature group X.

Images