Round-link chain defect detection method and system based on deep learning
By constructing an enhanced dataset of downhole environmental features and an improved deep learning model, the real-time performance and accuracy issues of circular link chain detection in complex downhole environments were resolved. This enabled efficient and accurate detection of circular link chain defects, ensuring the safe operation of the scraper conveyor.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NINGXIA TIANDI BENNIU IND GRP
- Filing Date
- 2025-12-17
- Publication Date
- 2026-04-17
AI Technical Summary
Existing detection methods are not suitable for the complex downhole environment and cannot achieve real-time online detection of circular links, resulting in a high risk of missed detections and insufficient detection accuracy.
An enhanced dataset was constructed, and a deep learning target detection model was trained on downhole circular chain images. An improved YOLO model and Focal Loss loss function were adopted to improve the model's adaptability and detection accuracy in the downhole environment, thereby achieving real-time detection of defects in circular chains.
It achieves efficient, accurate, and real-time detection of defects in circular chain conveyors, provides precise equipment maintenance support, and ensures the safe and stable operation of scraper conveyors.
Smart Images
Figure CN121883374A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of scraper conveyor inspection technology, specifically to a method and system for detecting defects in circular chain conveyors based on deep learning. Background Technology
[0002] In coal mine production, scraper conveyors play a crucial role in material transportation, and the circular link chain, as the core transmission component of the scraper conveyor, directly affects the safety and efficiency of coal mine production. However, current methods for detecting defects in circular link chains mostly rely on manual inspection or traditional non-destructive testing (NDT) methods. Manual inspection is not only labor-intensive and inefficient, but also prone to missed defects due to human error in the complex and dimly lit underground environment. Traditional NDT methods, such as ultrasonic testing, penetrant testing, and magnetic particle testing, are generally only suitable for static testing and cannot achieve dynamic real-time detection while the scraper conveyor is running. Therefore, there is an urgent need for a circular link chain defect detection method that can adapt to the complex underground environment, perform real-time online detection, and offer high accuracy. Summary of the Invention
[0003] In view of this, the present invention provides a method and system for detecting defects in circular ring chains based on deep learning, in order to solve the technical problems that existing detection methods cannot adapt to complex downhole environments and cannot perform real-time online detection.
[0004] The technical solution adopted by this invention to solve its technical problem is:
[0005] A deep learning-based method for detecting defects in circular chain rings includes:
[0006] Images of circular chains under different downhole conditions were collected to construct an augmented dataset;
[0007] The augmented dataset is fed into the object detection model for training.
[0008] Based on the trained target detection model, the real-time acquired images of the circular chain of the scraper conveyor are used to detect defects in the circular chain.
[0009] Preferably, the image of the circular chain is obtained in the following manner:
[0010] An image acquisition device is installed above the tail of the scraper conveyor. Under various working conditions in the mine, the image acquisition device is controlled to move along the length of the scraper conveyor to acquire images of the circular chain under different working conditions and at different angles. The images of the circular chain include those with cracks and those without cracks.
[0011] Preferably, the augmented dataset is obtained in the following ways:
[0012] Preprocess the acquired images of the circular chain;
[0013] Data augmentation is performed on the preprocessed circular chain image to form an augmented dataset.
[0014] Preferably, the crack regions in the enhanced dataset are labeled, and the labeling information includes the coordinates of the crack bounding box and feature attributes.
[0015] Preferably, the step of inputting the augmented dataset into the improved object detection model for training includes:
[0016] The augmented dataset is randomly divided into training dataset, validation dataset, and test dataset;
[0017] Build an object detection model and input the training dataset into the object detection model for training;
[0018] By introducing the Focal Loss function, the parameters of the object detection model are adjusted using a validation dataset to optimize the training results of the object detection model.
[0019] Preferably, the construction of the target detection model includes: increasing the number of 3×3 convolution kernels in the first to third shallow convolutional layers of the basic backbone network of the target detection model.
[0020] Preferably, the Focal Loss function is defined as follows: ;
[0021] in, The improved weighted Focal Loss value; As a category balance factor; For sample weights; The target probability predicted by the model; For focusing parameters.
[0022] This invention also provides a deep learning-based circular chain defect detection system, applied to the deep learning-based circular chain defect detection method described above, comprising:
[0023] An image acquisition device is installed on the tail of the scraper conveyor and is used to acquire images of the circular chain of the scraper conveyor.
[0024] An image processing device is used to process the acquired images of the circular chain to obtain defect detection results of the circular chain images.
[0025] Preferably, the image processing device includes a dataset construction module, a model training module, and a defect detection module;
[0026] The dataset construction module is used to construct an enhanced dataset based on the circular chain images of different downhole working conditions acquired by the image acquisition device;
[0027] The model training module is used to input the augmented dataset into the object detection model for training.
[0028] The defect detection module is used to detect defects in the circular chain image of the scraper conveyor acquired in real time based on the trained target detection model, and obtain the defect detection results of the circular chain.
[0029] Preferably, the deep learning-based circular chain defect detection system further includes a support frame, which is mounted on the tail of the scraper conveyor. A slide rail is mounted above the support frame, and the slide rail is located directly above the running path of the scraper conveyor chain. The image acquisition device is movably mounted on the slide rail.
[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0031] This invention first acquires images of circular link chains under different downhole operating conditions to construct an enhanced dataset containing downhole environmental features. Then, the enhanced dataset is input into a target detection model for training, thereby improving the model's adaptability and robustness in complex downhole environments, enabling accurate location and detection of defects in the circular link chains. Finally, the trained target detection model is deployed to a backend processing device to detect real-time acquired images of the circular link chains of the scraper conveyor, outputting the defect detection results. This achieves real-time online detection of the scraper conveyor's circular link chains. By constructing an enhanced dataset containing downhole environmental features and a target detection algorithm, this invention improves the accuracy of the target detection model in detecting defects in circular link chains, achieving efficient, accurate, and real-time detection of scraper conveyor circular link chains. This provides precise data support for equipment maintenance and ensures the safe and stable operation of the scraper conveyor. Attached Figure Description
[0032] Figure 1 This is a flowchart of the method for detecting defects in circular ring chains based on deep learning according to the present invention.
[0033] Figure 2 This is a schematic diagram of the structure of the deep learning-based circular chain defect detection system of the present invention.
[0034] In the figure: scraper conveyor tail 10, image acquisition device 100, image processing device 200, bracket 300, slide rail 400. Detailed Implementation
[0035] The technical solutions and effects of the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.
[0036] Please refer to Figure 1 A method for detecting defects in circular ring chains based on deep learning, comprising:
[0037] Images of circular chains under different downhole conditions were collected to construct an augmented dataset;
[0038] The augmented dataset is fed into the object detection model for training.
[0039] Based on the trained target detection model, the real-time acquired images of the circular chain of the scraper conveyor are used to detect defects in the circular chain.
[0040] This invention first acquires images of circular link chains under different downhole operating conditions to construct an enhanced dataset containing downhole environmental features. Then, the enhanced dataset is input into a target detection model for training, thereby improving the model's adaptability and robustness in complex downhole environments, enabling accurate location and detection of defects in the circular link chains. Finally, the trained target detection model is deployed to a backend processing device to detect real-time acquired images of the circular link chains of the scraper conveyor, outputting the defect detection results. This achieves real-time online detection of the scraper conveyor's circular link chains. By constructing an enhanced dataset containing downhole environmental features and a target detection algorithm, this invention improves the accuracy of the target detection model in detecting defects in circular link chains, achieving efficient, accurate, and real-time detection of scraper conveyor circular link chains. This provides precise data support for equipment maintenance and ensures the safe and stable operation of the scraper conveyor.
[0041] Furthermore, the images of the circular link chain are obtained in the following way: an image acquisition device is installed above the tail of the scraper conveyor. In a multi-condition underground environment, the image acquisition device is controlled to move along the length of the scraper conveyor to acquire images of the circular link chain under different underground conditions and at different angles. These images include both cracked and crack-free images of the circular link chain. Specifically, an industrial camera is first mounted on a bracket at the tail of the scraper conveyor. An adjustable slide rail is mounted above the bracket, positioned directly above the scraper conveyor chain's running path, with the slide rail's length parallel to the scraper conveyor's length. The image acquisition device is mounted on the slide rail and can move along it, with its lens facing the circular link chain to ensure a clear view of the entire chain in operation. Next, in a multi-condition underground environment, such as different light intensities and dust concentrations, the image acquisition device is used to acquire images of the circular link chain under different underground conditions and at different angles, obtaining a set of images of the circular link chain including both cracked and crack-free states. By improving the layout of the downhole image acquisition device and adopting an adjustable slide rail, a guarantee is provided for obtaining high-quality image data suitable for deep learning model training and detection.
[0042] In some implementations, the image acquisition device can be an industrial camera, which can operate continuously for extended periods and maintain stable operation in harsh industrial environments such as dust, high temperatures, and strong light, and is not easily damaged. The image acquisition device can also be an infrared camera, which can provide clear imaging in dark environments and may have better adaptability to low-light conditions underground. The corresponding camera parameters can be set according to the underground environment. In some implementations, to improve the universality of the subsequent target detection model, when using the image acquisition device to acquire images of the circular chain under different underground working conditions, in addition to acquiring images of the circular chain of the scraper conveyor in different underground environments, it is also necessary to acquire images of the circular chain on different scraper conveyors.
[0043] Furthermore, the augmented dataset is obtained through the following methods: preprocessing the acquired circular chain images; and performing data augmentation on the preprocessed circular chain images to form the augmented dataset. Specifically, a large number of circular chain images are acquired using the aforementioned image acquisition device under different downhole conditions. The acquired raw images are first preprocessed, including image cropping and adaptive denoising. Image cropping ensures that the size of each image is consistent, thereby ensuring the standardization of the data input to the target detection model. Denoising improves image quality, reduces interference information in the images, and makes target features clearer and easier to identify, thereby directly improving the performance of the target detection model. After preprocessing, data augmentation operations are performed on the circular chain images, such as adding Gaussian noise, adjusting image brightness and contrast, and performing random rotation and scaling, to simulate the complex and variable downhole environment and enhance the diversity and robustness of the dataset. Through data augmentation, the target detection model is exposed to more image features under complex downhole environments during training, improving the adaptability and robustness of the target detection model in actual downhole environments, enabling the target detection model to work stably under different lighting and dust conditions.
[0044] Furthermore, the cracked areas in the augmented dataset are labeled, with the labeling information including the coordinates of the crack bounding box and feature attributes. The acquired images of the circular chain include both cracked and crack-free images. To improve the accuracy of the target detection model in identifying cracks in the images, the cracked areas in the augmented dataset need to be accurately labeled, with the labeling information including the coordinates of the crack bounding box and feature attributes. Thus, when the trained model is deployed to the backend processing equipment, the circular chain of the scraper conveyor is detected in real time. The backend processing equipment receives the images of the circular chain acquired by the image acquisition device in real time, uses the trained model to detect the circular chain in the images, and directly outputs the coordinates and features of the cracked areas. This allows for quick and accurate location of the cracks in the circular chain, providing a precise basis for subsequent maintenance and replacement, helping to promptly troubleshoot faults and ensuring the safe operation of the scraper conveyor.
[0045] Furthermore, the augmented dataset is fed into the improved object detection model for training, including:
[0046] The augmented dataset is randomly divided into training dataset, validation dataset, and test dataset;
[0047] Build an object detection model and input the training dataset into the object detection model for training;
[0048] By introducing the Focal Loss function, the parameters of the object detection model are adjusted using a validation dataset to optimize the training results of the object detection model.
[0049] This invention selects the YOLO object detection model as its foundation and improves it to address the characteristics of circular chain crack detection. Since chain cracks are typically small and their features are not obvious, the network structure of the object detection model is optimized to increase the extraction capability of shallow features, thereby improving the detection accuracy of small target cracks. Specifically, the object detection model is constructed by increasing the number of 3×3 convolutional kernels in the first three shallow convolutional layers of the basic backbone network. Specifically, the number of kernels in the second layer is increased from 32 to 64, and the number of kernels in the third layer is increased from 64 to 128. By stacking 3×3 convolutional kernels, the model can gradually expand its receptive field without increasing the number of parameters in a single layer, thus capturing a wider range of contextual information, which is particularly important for distinguishing small targets in object detection models. Simultaneously, depthwise separable convolutions are introduced to replace traditional ordinary convolutions, reducing computational complexity while maintaining feature extraction accuracy. By improving the target detection model to address the small target characteristics of circular chain cracks, the perceptual ability of the shallow feature extraction module can be enhanced, and the network's ability and efficiency in capturing subtle low-level features such as the edges and textures of circular chain cracks can be improved. This solves the problem of missed detection of small cracks caused by insufficient shallow feature expression in the traditional YOLO network.
[0050] Furthermore, the Focal Loss function is defined as follows: ;
[0051] in, The improved weighted Focal Loss value; As a category balance factor; For sample weights; The target probability predicted by the model; For focusing parameters.
[0052] This invention introduces the Focal Loss function to increase the weight of the crack region, making the target detection model pay more attention to crack features during training. The category balancing factor is a global category-level weight parameter used to balance the overall imbalance between the number of "cracked samples (positive samples)" and "non-cracked samples (negative samples)" in the dataset. Since the number of non-cracked samples is typically 5-10 times that of cracked samples in circular chain detection scenarios, without category balancing, the model will suffer from "underfitting crack features" due to the bias in training data distribution towards non-cracked samples. For example, in this invention... For positive samples (cracks), set as +=0.8, set to 0.8 for negative samples (non-cracks). -=0.2, by increasing the overall weight of the positive sample category, ensures that the model's learning resources for positive and negative samples are balanced during training, and avoids the model ignoring crack samples due to a large difference in the number of samples. To focus on parameters and control the loss contribution of easily classified samples. When As the value increases, the loss weight for easily classified samples decreases significantly, while the loss weight for difficult-to-classify samples increases relatively, causing the model to focus more on the difficult-to-classify samples. In this invention, easily classified samples are non-cracked samples, and difficult-to-classify samples are cracked samples; the focusing parameter is set accordingly. =2, reducing the loss weight of easily classified non-cracked samples.
[0053] Furthermore, Sample weights are local sample-level enhancement parameters used to prioritize the training of key samples within the same class. The "global category balance" function distinguishes between them. The setting only applies enhancement to positive samples (cracks), specifically: the weighting coefficient of crack samples. =2, the weighting coefficient for non-cracked samples =1. The core purpose of this design is to... Having already addressed the issue of "balancing the number of positive and negative samples," we further enhance the contribution of crack samples to loss calculation—even those that have already been identified as positive or negative. The balanced crack sample will generate twice the loss value of the non-crack sample for each prediction error, thus forcing the model to focus the training on crack targets, especially small, weak, and easily ignored difficult-to-classify crack samples. This solves the shortcomings of traditional Focal Loss, which only balances the categories and does not strengthen key samples, improves the model's detection sensitivity and localization accuracy of crack targets, and ensures that the trained model has the ability to identify small cracks in complex downhole environments.
[0054] In some implementations, after the target detection model is trained, it is deployed to the back-end processing equipment. When real-time detection of the scraper conveyor's circular link chain is required, an image acquisition device is installed on the tail of the scraper conveyor to be inspected, and the scraper conveyor and the image acquisition device are started. The image acquisition device moves along the slide rail according to the settings and acquires multi-angle images of the running circular link chain in real time according to the set camera parameters. Then, the multi-angle images of the circular link chain are transmitted to the back-end processing equipment, which inputs the multi-angle images of the circular link chain into the target detection model. The target detection model outputs the defect detection results of the scraper conveyor's circular link chain, thereby achieving efficient, accurate, and real-time detection of the scraper conveyor's circular link chain.
[0055] Please refer to Figure 2 The present invention also provides a deep learning-based circular chain defect detection system, applied to the deep learning-based circular chain defect detection method described above, comprising:
[0056] Image acquisition device 100 is installed on the tail end 10 of the scraper conveyor and is used to acquire images of the circular chain of the scraper conveyor.
[0057] The image processing device 200 is used to process the acquired circular chain image to obtain the defect detection result of the circular chain image.
[0058] Specifically, in the early stages of circular chain defect detection, a target detection model needs to be built and trained. Therefore, an image acquisition device 100 is installed on the tail section 10 of the scraper conveyor to acquire images of circular chains under different downhole working conditions, serving as the dataset for the target detection model. Here, the scraper conveyor can be the one currently being detected or other scraper conveyors. Installing the image acquisition device 100 on different scraper conveyors allows for the acquisition of more diverse circular chain images, thereby improving the universality of the target detection model. After the target detection model is trained, the image acquisition device 100 is installed on the tail section of the scraper conveyor to be detected and acquires images of the circular chain of that scraper conveyor in real time, thus enabling real-time detection of the circular chain of the scraper conveyor to be detected. The image processing device 200 is the aforementioned back-end processing equipment. The image processing device 200 is electrically connected to the image acquisition device 100, enabling the image acquisition device 100 to transmit the acquired circular chain images to the image processing device 200 for processing. The deep learning-based circular link chain defect detection system of this invention enables real-time detection of circular link chains. This allows for timely and effective detection of cracks in the circular link chain when they first appear and are relatively minor, helping to eliminate faults promptly and ensuring the safe operation of the scraper conveyor.
[0059] Furthermore, the image processing device includes a dataset construction module, a model training module, and a defect detection module;
[0060] The dataset construction module is used to build an enhanced dataset based on images of circular chains under different downhole conditions acquired by the image acquisition device;
[0061] The model training module is used to input augmented datasets into the object detection model for training;
[0062] The defect detection module is used to detect defects in the circular chain images of the scraper conveyor that are acquired in real time based on the trained target detection model, and obtain the defect detection results of the circular chain.
[0063] In the early stage of circular chain defect detection, when the image processing device receives the circular chain image sent by the image acquisition device, the dataset construction module and the model training module need to be run to obtain the target detection model for circular chain defect detection. After the target detection model is trained, when the image processing device receives the circular chain image sent by the image acquisition device again, it needs to run the defect detection module to detect the circular chain image of the scraper conveyor that is acquired in real time and obtain the defect detection result of the circular chain.
[0064] In some implementations, to distinguish between the target detection model building stage and the real-time detection stage, the image processing device further includes a mode confirmation module. Specifically, when using the deep learning-based circular chain defect detection system of the present invention, the operator first selects the current mode through the image processing device, i.e., the target detection model building stage or the real-time detection stage. When the operator selects the target detection model building stage, the mode confirmation module controls the operation of the dataset construction module and the model training module, and controls the defect detection module to stop operating. In this case, if the image processing device receives a circular chain image sent by the image acquisition device, it will send the circular chain image to the dataset construction module for processing. When the operator selects the real-time detection stage, the mode confirmation module controls the operation of the defect detection module, and the dataset construction module and the model training module to stop operating. In this case, if the image processing device receives a circular chain image sent by the image acquisition device, it will send the circular chain image to the defect detection module for processing, and the defect detection module can directly output the defect detection result of the circular chain image.
[0065] Furthermore, please refer again. Figure 2The deep learning-based circular link chain defect detection system also includes a support 300, which is mounted on the tail end 10 of the scraper conveyor. A slide rail 400 is mounted above the support 300, directly above the running path of the scraper conveyor chain. An image acquisition device 100 is movably mounted on the slide rail 400. Specifically, the length direction of the slide rail is parallel to the length direction of the scraper conveyor, allowing the image acquisition device 100 to move along the length direction of the slide rail 400 to ensure a clear image of the entire circular link chain. The moving speed of the image acquisition device 100 is greater than the running speed of the circular link chain, ensuring that the image acquisition device 100 can clearly capture the entire circular link chain even during operation.
[0066] The above-disclosed embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the invention. Those skilled in the art will understand that implementing all or part of the above-described embodiments and making equivalent changes in accordance with the claims of the present invention are still within the scope of the invention.
Claims
1. A deep learning-based chain defect detection method, characterized in that, include: Images of circular chains under different downhole conditions were collected to construct an augmented dataset; The augmented dataset is fed into the object detection model for training; Based on the trained target detection model, the real-time acquired images of the circular chain of the scraper conveyor are used to detect defects in the circular chain.
2. The method for detecting defects in circular chain rings based on deep learning according to claim 1, characterized in that, The image of the circular chain is obtained in the following way: An image acquisition device is installed above the tail of the scraper conveyor. Under various working conditions in the underground environment, the image acquisition device is controlled to move along the length of the scraper conveyor to acquire images of the circular chain under different underground working conditions and at different angles. The images of the circular chain include those with cracks and those without cracks.
3. The method for detecting defects in circular chain rings based on deep learning according to claim 2, characterized in that, Augmented datasets are obtained in the following ways: Preprocess the acquired images of the circular chain; Data augmentation is performed on the preprocessed circular chain image to form an augmented dataset.
4. The method for detecting defects in circular chain rings based on deep learning according to claim 3, characterized in that, The crack regions in the augmented dataset are labeled, and the labeling information includes the coordinates of the crack bounding box and feature attributes.
5. The method for detecting defects in circular ring chains based on deep learning according to claim 1, characterized in that, The step of inputting the augmented dataset into the improved object detection model for training includes: The augmented dataset is randomly divided into training dataset, validation dataset, and test dataset; Build an object detection model and input the training dataset into the object detection model for training; By introducing the Focal Loss function, the parameters of the object detection model are adjusted using a validation dataset to optimize the training results of the object detection model.
6. The method for detecting defects in circular ring chains based on deep learning according to claim 5, characterized in that, The construction of the target detection model includes: increasing the number of 3×3 convolution kernels in the first to third shallow convolutional layers of the basic backbone network of the target detection model.
7. The method for detecting defects in circular ring chains based on deep learning according to claim 5, characterized in that, The Focal Loss function is defined as follows: ； in, The improved weighted Focal Loss value; As a category balance factor; For sample weights; The target probability predicted by the model; For focusing parameters.
8. A deep learning-based circular chain defect detection system, applied to the deep learning-based circular chain defect detection method as described in any one of claims 1-7, characterized in that, include: An image acquisition device is installed on the tail of the scraper conveyor and is used to acquire images of the circular chain of the scraper conveyor. An image processing device is used to process the acquired images of the circular chain to obtain defect detection results of the circular chain images.
9. The deep learning-based circular chain defect detection system according to claim 8, characterized in that, The image processing device includes a dataset construction module, a model training module, and a defect detection module; The dataset construction module is used to construct an enhanced dataset based on the circular chain images of different downhole working conditions acquired by the image acquisition device; The model training module is used to input the augmented dataset into the object detection model for training. The defect detection module is used to detect defects in the circular chain image of the scraper conveyor acquired in real time based on the trained target detection model, and obtain the defect detection results of the circular chain.
10. The deep learning-based circular chain defect detection system according to claim 9, characterized in that, The deep learning-based circular chain defect detection system also includes a support frame, which is installed on the tail of the scraper conveyor. A slide rail is mounted above the support frame, and the slide rail is located directly above the running path of the scraper conveyor chain. The image acquisition device is movably mounted on the slide rail.