Belt tearing detection method based on Mobile U-Net and K-means clustering

Abstract

The invention discloses a belt tearing detection method based on Mobile U-Net and K-means clustering, and belongs to the technical field of industrial detection and target recognition. According to the method, firstly, a Mobile U-Net deep learning model is used for processing a video frame picture collected by a camera, the edges of the two sides of a belt are detected in real time, a belt area is accurately extracted, a belt area mask is manufactured, and therefore false alarms caused by the environment and light problems of the area outside the belt are reduced; carrying out clustering analysis on the video frame picture by adopting a K-means clustering algorithm to obtain a new image with a re-colored connected domain, and extracting and analyzing an abnormal region with a tearing feature through a connected domain mark; and finally, calculating the abnormal area and the belt area mask to remove the interference of the area outside the belt so as to realize the belt tearing abnormity detection. According to the method, additional installation of a sensor and laser equipment is not needed, hardware overhead is low, and the problems that an existing belt tearing detection method is large in error, high in omission ratio, complex in installation and debugging, high in maintenance cost and slow in technology updating iteration can be effectively solved.

Description

[0001] This invention relates to the field of industrial inspection and target recognition technology, specifically a belt tear detection method based on Mobile U-Net and K-means clustering. Background Technology

[0002] In industrial production, belt conveyors are widely used in mining, ports, metallurgy, and chemical industries as crucial equipment for material handling. During long-term operation, belts are prone to longitudinal tearing due to material impact, wear, and aging. If a belt tear is not detected and addressed promptly, it can lead to material leakage, equipment damage, and even safety accidents, resulting in significant economic losses. Therefore, achieving rapid and accurate detection of belt tears is of great practical importance. Currently, various solutions have been proposed in the research and development of belt longitudinal tear monitoring devices both domestically and internationally. Existing belt tear detection equipment is mainly divided into two categories: one category uses mechanical devices for detection, including linear detectors and material leakage detectors; the other category uses computer vision technology for detection, mainly including computer vision detection based on deep learning and visual detection methods based on line lasers and CCD cameras. Linear detectors are installed below trough-shaped conveyor belts, with a metal or nylon wire running along the belt's contour. A spring-loaded limit switch is attached to one end of the wire. When material punctures the belt and gets caught on the wire, it breaks or increases the tension, causing the corresponding limit switch to activate and stopping the conveyor. However, this type of detector is highly sensitive to the location and extent of belt tears. If the tear doesn't catch the metal or nylon wire, it may miss the detection. Furthermore, the metal or nylon wire is susceptible to wear and corrosion over long-term use, requiring regular replacement and resulting in high maintenance costs. The material leakage detector consists of a tray, a fulcrum, a counterweight, and a switch. When the conveyor belt tears longitudinally, the material on the belt leaks into the tray through the tear. The weight of the material overcomes the weight of the counterweight, causing the entire device to rotate around the fulcrum, forcing the limit switch to activate and stopping the conveyor. This detector has a simple structure and is convenient to use, but it has significant drawbacks: It only functions when there is material on the conveyor belt and the tear is large enough to leak material. If there is no material on the conveyor belt or the tear is too small to leak, it will miss the detection. Furthermore, when dust accumulates on the idler rollers, it increases the weight of the tray, causing malfunctions and affecting detection accuracy.

[0003] Deep learning-based detection methods based on computer vision learn tear features in belt images through training a model to achieve detection. However, this method requires a large amount of labeled data, which is difficult to obtain in real-world industrial scenarios. Furthermore, after model training, the model has poor adaptability to belt tear detection in different scenarios. When the belt material, color, or operating environment changes, the detection accuracy drops significantly, and the technology updates slowly. The visual inspection method based on line laser and CCD camera obtains three-dimensional information of the object's surface by projecting a line laser onto it and capturing the projection of the laser line onto the surface using a CCD camera, thus enabling belt tear detection. While this method provides relatively detailed and accurate belt surface information, improving detection accuracy, it requires the additional installation of a line laser emitter and a high-precision CCD camera, resulting in high hardware costs, complex installation and debugging, and the laser equipment is easily damaged in harsh industrial environments (such as dust, vibration, and high temperatures), making maintenance difficult. Given the many shortcomings of existing belt tear detection methods, there is an urgent need for a new belt tear detection method to solve problems such as large errors, high false negative rate, complex installation and debugging, high maintenance cost, and slow technology update and iteration speed. Summary of the Invention

[0004] The purpose of this invention is to provide a belt tear detection method based on Mobile U-Net and K-means clustering, so as to solve the technical problems of existing belt tear detection methods mentioned in the background art, such as large error, high false negative rate, complex installation and debugging, high maintenance cost, and slow technology update and iteration speed.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a belt tear detection method based on Mobile U-Net and K-means clustering, comprising the following steps: S1: Capture video frames during the belt operation using a 1080P high-definition camera, ensuring that there are no obstructions in front of the camera to avoid affecting the image acquisition quality. S2: Perform image preprocessing on the video frames acquired in step S1, specifically including noise removal, size normalization, and image enhancement operations; S3: Use the Mobile U-Net deep learning model to perform belt edge detection on the video frames preprocessed in step S2, accurately extract the belt region, and create a belt region mask. The Mobile U-Net deep learning model uses the lightweight MobileNetV3 network as its backbone. MobileNetV3 is specifically designed to address computational limitations in real-world applications, offering advantages such as fewer parameters, higher speed, and moderate depth. The bneck module from MobileNetV3 is ported to the encoding part of the U-Net network. Different depths of bneck modules are used to extract features based on the location within the U-Net encoding part; deeper locations require more bneck modules, resulting in more sophisticated abstract features. These extracted features are then sequentially connected to corresponding locations in the decoder for feature fusion. During the connection process, the number of channels and the size of the feature map are kept consistent with the corresponding locations in the decoder, thus accurately extracting the belt region and creating a belt region mask. This belt region mask only retains the belt region in the video frame, shielding areas outside the belt, effectively reducing false alarms caused by environmental and lighting issues in areas outside the belt. S4: Use the K-means clustering algorithm to perform clustering analysis on the video frames after preprocessing in step S2 to obtain new images with recolored connected components. Perform connected component label extraction analysis on the new images to extract abnormal regions with tearing characteristics. The K-means clustering algorithm uses Euclidean distance as an indicator to measure the similarity between data objects. The similarity is inversely proportional to the distance between data objects. Its specific execution process is as follows: S41: Randomly select k samples as the mean vector of the initial cluster class; S42: Divide each sample dataset into the clusters that are closest to it; S43: Update the cluster mean vector based on the cluster to which each sample belongs; S44: Repeat steps S42 and S43. When the set number of iterations is reached or the mean vector of the clusters no longer changes, output the clustering results. The formula for calculating the Euclidean distance between data objects and cluster centers in space is:

[0006] Where X is the data object; Ci is the i-th cluster center; m is the dimension of the data object; Xj and Cij are the j-th attribute values ​​of X and Ci, respectively. The formula for calculating the sum of squared errors (SSE) for the entire dataset is:

[0007] Where SSE represents the quality of the clustering result; k is the number of clusters; S5: Perform calculations on the abnormal region extracted in step S4 and the belt area mask created in step S3 to remove interference from the outer area of ​​the belt and realize belt tear anomaly detection.

[0008] Preferably, the belt area mask created in step S3 only retains the belt area in the video frame and blocks the area outside the belt.

[0009] Preferably, the value of k in step S4 is preset according to the actual operating scenario of the belt, and the value range is 3-5.

[0010] Preferably, the number of iterations set in step S44 is in the range of 50-100.

[0011] Preferably, in step S2, noise removal is performed because industrial sites often experience interference such as floating dust and equipment vibration, which can easily cause salt-and-pepper noise or Gaussian noise in the acquired video frames, affecting edge detection accuracy. A Gaussian filtering algorithm is used to filter the video frames. By weighted averaging of image pixels, the noise caused by dust and vibration in the industrial environment is effectively eliminated while preserving the belt edge features, resulting in smoother grayscale transitions. Size normalization: The original video frames captured by the camera may have inconsistent sizes due to slight changes in shooting angle and distance. Directly inputting them into the model would increase computational complexity and affect detection stability. After noise removal, the video frames are uniformly adjusted to a preset size. The size scaling is achieved through bilinear interpolation to ensure that the spatial resolution of each frame is consistent, meeting the input specifications of the Mobile U-Net model. Image Enhancement: Frequent changes in lighting conditions in industrial settings (such as day-night cycles and localized shadows) can easily lead to low contrast between the conveyor belt and the background, and blurred features in the tear area. An adaptive histogram equalization algorithm (CLAHE) is used to enhance the brightness and contrast of the size-normalized video frames. By dividing the image into blocks and adjusting the histogram distribution of each block, the grayscale gradient differences at the belt edges and the texture features of the tear area are enhanced, enabling subsequent models to more clearly identify the target area.

[0012] Preferably, the Gaussian filter has a convolution kernel size of 3×3 and a standard deviation of 1.0; the preset size for size normalization is 512×512 pixels; the CLAHE clipping constraint parameter is 2.5, and the grid size is 8×8.

[0013] Compared with the prior art, the beneficial effects of the present invention are: (1) Low hardware overhead: This invention only requires a 1080P high-definition RGB camera for tear recognition, without the need for additional sensors, laser emission devices and other equipment, which reduces hardware procurement costs and installation and debugging difficulties. (2) High detection accuracy: The belt region is accurately extracted by the Mobile U-Net deep learning model and a belt region mask is made, which effectively reduces false alarms caused by environmental and lighting problems in areas outside the belt; combined with the K-means clustering algorithm, abnormal regions with tearing characteristics are extracted, and interference outside the belt is removed by operation with the belt region mask, which significantly reduces the false alarm rate and false alarm rate and improves the detection accuracy. (3) Strong adaptability: The Mobile U-Net deep learning model has the characteristics of few parameters and fast speed, which can adapt to the computing power limitations of industrial sites. Moreover, the K-means clustering algorithm does not require a large amount of labeled data and has good adaptability to belt tear detection of different materials, colors and operating environments. The technology is flexible in updating and iterating. (4) Low maintenance cost: Since no additional complex equipment is required, the number of equipment failure points is reduced, the maintenance difficulty and maintenance cost are reduced, and the periodic replacement of vulnerable parts in mechanical testing devices is avoided, thus extending the service life of the equipment. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram illustrating the principle of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Please see Figure 1-2 This invention provides a technical solution: a belt tear detection method based on Mobile U-Net and K-means clustering, the specific implementation steps of which are as follows: Step 1: Hardware Preparation and Image Acquisition A 1080P high-definition RGB camera was selected and installed in a suitable position on the belt conveyor, ensuring that the camera lens was directly facing the belt surface and that there were no obstructions in front of the camera to avoid affecting image acquisition. The camera was then turned on to capture video frames in real time during the belt's operation, with the frame rate set to 25fps to ensure real-time capture of the belt's running status. Step 2: Image Preprocessing Perform the following preprocessing operations on the captured 1080P video frames: Noise removal: The GaussianBlur function from the OpenCV library was used, with the convolution kernel size set to (3,3) and the standard deviation sigma=1.0. Each frame of the image was filtered to eliminate Gaussian noise caused by dust and vibration at the scene. After processing, the grayscale fluctuation of the image was reduced by 40%-60%. Size normalization: By using the resize function in combination with bilinear interpolation, the filtered image is uniformly scaled to 512×512 pixels to ensure that the aspect ratio and number of pixels of each frame are consistent, thus avoiding feature extraction deviations caused by size differences during subsequent model processing. Image Enhancement: The CLAHE function is called, with clipLimit set to 2.5 and tileGridSize set to (8,8), to enhance the normalized image. After processing, the grayscale gradient difference at the belt edge is increased by 30%-50%, and the texture details in the torn area are significantly improved, allowing the belt to be clearly distinguished from the background even in low-light conditions.

[0017] Step 3: Mobile U-Net Model Training and Belt Region Extraction Dataset Preparation: Collect belt images under different materials (e.g., rubber, canvas), colors (e.g., black, green), and operating environments (e.g., normal lighting, low lighting, dusty environments), including normal belt images and belt images with varying degrees of tearing, to construct a dataset. Preprocess the images in the dataset, including image cropping, size normalization (adjusting image sizes to a uniform 512×512 pixels), and data augmentation (e.g., rotation, flipping, brightness adjustment) to improve the model's generalization ability. Model Training: A Mobile U-Net network model was built, using MobileNetV3-Small as the backbone. The bneck modules from MobileNetV3-Small were ported to the encoding part of the U-Net network. The encoding part consists of 5 stages: stage 1 with 2 bneck modules, stage 2 with 3 bneck modules, stage 3 with 4 bneck modules, stage 4 with 3 bneck modules, and stage 5 with 2 bneck modules. The decoder uses transposed convolutional layers to progressively restore image resolution. Features extracted from each stage of the encoding part are fused with features from the corresponding stage of the decoder through skip connections. Before fusion, the number of channels in the feature map is adjusted using a 1×1 convolutional layer to ensure that the feature map size and number of channels match. The cross-entropy loss function is used as the model's loss function, and the Adam optimizer is used for model training. The initial learning rate is set to 0.001, and the learning rate is gradually decreased with each training epoch. The preprocessed dataset is divided into training, validation, and test sets in a ratio of 7:2:1. The training rounds are set to 100 rounds. After each round of training, the model performance is validated on the validation set. When the loss function value of the model on the validation set no longer decreases for 10 consecutive rounds, training is stopped and the optimal model is saved. Belt region extraction: Input the video frames acquired in step one into the trained Mobile U-Net model. The model outputs a prediction mask for the belt region. The prediction mask is then binarized (the threshold is set to 0.5; regions with a prediction probability greater than 0.5 are identified as belt regions, and regions with a prediction probability less than or equal to 0.5 are identified as background regions) to obtain the final belt region mask. Step 4: K-means clustering and outlier region extraction Image preprocessing: The video frames acquired in step one are preprocessed to convert the RGB images to the Lab color space. The a and b channels are selected as clustering features (the Lab color space can better reflect the visual differences in color, and the a and b channels are more sensitive to color changes, which is helpful in distinguishing normal belt areas from torn areas). The feature data are normalized to the [0,1] interval to reduce the influence of data dimensions on the clustering results. K-means clustering: The k-value is set to 4 (based on the actual operation scenario of the conveyor belt in this embodiment, the image is divided into four clusters: normal conveyor belt area, background area, slightly torn area, and severely torn area). Four samples are randomly selected as initial cluster centers. The Euclidean distance between each sample and the four initial cluster centers is calculated, and each sample is assigned to the nearest cluster. The cluster centers are updated based on the mean of the features of all samples within each cluster. The above sample allocation and cluster center update process is repeated, with 80 iterations. Iteration stops when the number of iterations reaches 80 or the change in cluster centers is less than 0.001, yielding the clustering results. The image is recolored based on the clustering results, with different colors corresponding to different clusters, resulting in a new image with recolored connected components. Anomaly region extraction: A connected component labeling algorithm is used to label the connected components of the recolored image, calculating the morphological features of each connected component, such as area, perimeter, and circularity. Based on the morphological characteristics of the belt tear area (tear areas typically have small area, large perimeter, and low circularity), thresholds are set to filter out suspected tear connected components, i.e., anomaly regions with tear characteristics. For example, a connected component area threshold of 50-1000 pixels, a perimeter threshold of greater than 50 pixels, and a circularity threshold of less than 0.5 are set to filter out connected components that meet the criteria as anomaly regions. Step 5: Belt Tear Detection The abnormal region extracted in step three is ANDed with the belt area mask obtained in step two. The abnormal region within the belt area is retained, while interference from areas outside the belt is removed. The result is analyzed. If an abnormal region is found, it is determined that the belt has torn, an alarm signal is issued, and the belt conveyor is stopped. If no abnormal region is found, the belt is considered to be running normally, and the inspection continues.

[0018] The method described in this embodiment achieves an accuracy rate of over 95% for belt tear detection under different operating environments, with a false negative rate of less than 3% and a false positive rate of less than 2%, which can meet the real-time and accuracy requirements for belt tear detection in industrial settings.

[0019] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A belt tear detection method based on Mobile U-Net and K-means clustering, characterized in that, Includes the following steps: S1: Capture video frames during the belt operation using a 1080P high-definition camera, ensuring that there are no obstructions in front of the camera; S2: Perform image preprocessing on the video frames acquired in step S1, specifically including noise removal, size normalization, and image enhancement operations; S3: Use the Mobile U-Net deep learning model to perform belt edge detection on the video frames preprocessed in step S2, accurately extract the belt region, and create a belt region mask. The Mobile U-Net deep learning model uses the lightweight MobileNetV3 network as the backbone network. The bneck module in MobileNetV3 is ported to the encoding part of the Unet network. The bneck module of different depths is used to extract features according to the position of the Unet encoding part. The deeper the position, the more bneck modules are required and the higher the level of abstract features extracted. The extracted features are then connected to the corresponding positions of the decoder for feature fusion. During the connection process, the number of channels and the size of the feature map are kept consistent with the corresponding positions of the decoder. S4: Use the K-means clustering algorithm to perform clustering analysis on the video frames after preprocessing in step S2 to obtain new images with recolored connected components. Perform connected component label extraction analysis on the new images to extract abnormal regions with tearing characteristics. The K-means clustering algorithm uses Euclidean distance as an indicator to measure the similarity between data objects. The similarity is inversely proportional to the distance between data objects. Its specific execution process is as follows: S41: Randomly select k samples as the mean vector of the initial cluster class; S42: Divide each sample dataset into the clusters that are closest to it; S43: Update the cluster mean vector based on the cluster to which each sample belongs; S44: Repeat steps S42 and S43 until the set number of iterations is reached or the mean vector of the clusters no longer changes, then output the clustering results; The formula for calculating the Euclidean distance between data objects and cluster centers in space is:

2. Among them, X is a data object; Ci is the i-th cluster center; m is the dimension of the data object; Xj and Cij are the j-th attribute values ​​of X and Ci, respectively. The formula for calculating the sum of squared errors (SSE) for the entire dataset is:

3. Among them, The size of the SSE indicates the quality of the clustering result; k is the number of clusters; S5: Perform calculations on the abnormal region extracted in step S4 and the belt area mask created in step S3 to remove interference from the outer area of ​​the belt and realize belt tear anomaly detection.

4. The belt tear detection method based on Mobile U-Net and K-means clustering according to claim 1, characterized in that, The belt area mask created in step S3 only retains the belt area in the video frame and blocks the area outside the belt.

5. The belt tear detection method based on Mobile U-Net and K-means clustering according to claim 1, characterized in that, In step S4, the value of k is preset according to the actual operating scenario of the belt, and the value range is 3-5.

6. The belt tear detection method based on Mobile U-Net and K-means clustering according to claim 1, characterized in that, The number of iterations set in step S44 is 50-100.

7. The belt tear detection method based on Mobile U-Net and K-means clustering according to claim 1, characterized in that, Noise removal in step S2: Gaussian filtering algorithm is used to filter the video frame to eliminate image noise caused by dust and vibration in the industrial environment; Size normalization: The video frames after noise removal are uniformly adjusted to the preset size to ensure consistency in subsequent model processing; Image enhancement: The adaptive histogram equalization algorithm (CLAHE) is used to enhance the brightness and contrast of the size-normalized video frames, thereby improving the feature recognition of belt edges and tear areas.

8. The belt tear detection method based on Mobile U-Net and K-means clustering according to claim 5, characterized in that, The Gaussian filter has a convolution kernel size of 3×3 and a standard deviation of 1.0; the preset size for size normalization is 512×512 pixels; the CLAHE clipping constraint parameter is 2.5 and the grid size is 8×8.

Images