A YOLO-based large block detection system and method for discharging slag of a thermal power plant slag conveyor
By combining CycleGAN and the improved YOLOX model, the reliability and environmental adaptability issues of large object detection in thermal power plant slag removal machines were solved, achieving efficient and accurate detection in complex environments, reducing false detection rates, and improving system safety and efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA THREE GORGES CORPORATION
- Filing Date
- 2026-02-27
- Publication Date
- 2026-06-19
AI Technical Summary
Existing technologies for detecting large objects in thermal power plant slag removal machines suffer from poor reliability, high maintenance costs, and poor environmental adaptability. They are difficult to achieve stable, economical, and accurate detection in complex and harsh environments, resulting in a high false detection rate and affecting the safety and efficiency of the slag removal system.
A YOLO-based system for detecting large objects in slag discharge from thermal power plant slag removal machines is adopted. The system includes an image preprocessing module that uses the CycleGAN model to remove smoke interference, combined with an improved YOLOX model for large object detection, and improves image clarity through adversarial training of the generator and discriminator. The improved YOLOX model is used to enhance the perception of cross-channel and spatial context information to achieve accurate detection.
It significantly reduced the false detection rate, improved the accuracy of detection and the safety of system operation, and increased the operating efficiency of the slag discharge system.
Smart Images
Figure CN122243881A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of machine vision technology, and in particular relates to a YOLO-based detection system and method for detecting large slag particles discharged from thermal power plant slag removal machines. Background Technology
[0002] In thermal power plants, ash removal machines are crucial equipment for removing ash and slag from the bottom of boilers. Large foreign objects, such as solidified coke slag and detached refractory bricks, can clog the ash removal port or damage the equipment during the ash removal process, seriously affecting production safety. Therefore, real-time and accurate detection of large objects in the ash removal process is essential. Currently, detection technologies for this problem are mainly divided into two categories: physical contact detection and optical detection.
[0003] 1. Physical contact testing, including: Scraper skewness measurement device detection: This method determines the presence of large objects by detecting the skewness of the scraper. Its advantage is that it directly reflects the situation of encountering large objects during the operation of the slag remover, and the detection results are relatively intuitive and reliable. Resistance detection: Sensors are installed on the transmission system or scraper of the slag remover to monitor changes in operating resistance. Its advantage is that it allows for real-time monitoring of the slag remover's operating status and early detection of abnormal resistance caused by large objects.
[0004] 2. Optical inspection, including: Infrared thermal imager detection: Infrared thermal imagers are used to scan and monitor the slag discharge area. The advantages are non-contact detection of large objects in the slag, no interference with the normal operation of the slag remover, and rapid acquisition of temperature distribution information in the slag discharge area, helping to promptly detect anomalies. LiDAR scanning: LiDAR is used to scan the slag discharge port and downstream area. The advantages are high detection accuracy, accurately measuring the size and location of large objects, providing detailed data support for subsequent processing.
[0005] The existing technology has the following drawbacks: The scraper skewness measurement device requires the installation of an additional detection device on the scraper of the slag remover, which increases the complexity and cost of the equipment to some extent. In addition, the detection device may be affected by the working environment of the slag remover, such as water vapor and impurities, and requires regular maintenance. In resistance detection, resistance changes may be caused by a variety of factors, such as equipment wear and debris blockage. Relying solely on resistance detection may lead to misjudgment, and it is necessary to combine it with other detection methods for comprehensive judgment.
[0006] Infrared thermal imaging inspection: The price is relatively high, and the inspection results may be affected by factors such as changes in ambient temperature and steam. Accurate analysis and processing of the inspection data are required. LiDAR scanning: The equipment cost and maintenance cost are relatively high. It has high requirements for the installation environment and accuracy. In complex field environments, it may be affected by dust, debris and other interference.
[0007] Therefore, based on the aforementioned widespread technical problems, it is necessary to propose a YOLO-based detection system and method for detecting large slag particles discharged from thermal power plant slag removal machines to solve these problems. Summary of the Invention
[0008] The technical problem to be solved by this invention is to provide a YOLO-based detection system and method for large slag particles in thermal power plant slag removal machines. The aim is to solve the problem that existing technologies, due to poor reliability and high maintenance costs, or poor environmental adaptability and high implementation costs, are unable to achieve stable, economical and accurate detection of large slag particles in the complex and harsh working environment of thermal power plant slag removal machines. This will improve detection capabilities, enhance the operational safety of the slag removal system, significantly reduce the false detection rate, and improve the operational efficiency of the slag removal system.
[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A YOLO-based system for detecting large slag particles in ash removal equipment used in thermal power plants, comprising: The image preprocessing module is used to remove smoke from images collected in the slag discharge area of the slag remover. The large object detection module uses an improved YOLOX model to detect large objects in the preprocessed image. The output module is used to output the category and location information of large objects.
[0010] Preferably, the image preprocessing module uses a CycleGAN model for image desmoke processing. The CycleGAN model includes: Generator G is used to convert a smoky image into a desmoky image; Generator F is used to restore a sharp image to a smoky image; The discriminator Dx is used to distinguish between the generated image and the real, clear image; The discriminator Dy is used to distinguish between generated images and real images with smoke.
[0011] Preferably, the network structure of the generator G and generator F is an encoder-decoder structure, the input image size is 256×256 pixels, the encoder consists of a multi-level smoke feature extraction module, and the decoder includes transposed convolution and skip connections.
[0012] Preferably, the bulk object detection module adopts the YOLOX model and includes: The backbone feature extraction network employs residual convolutional networks, CSPNet, and Focus structures. The feature fusion network employs the Path Aggregation Feature Pyramid Network (PAFPN). The Spatial Context Awareness Module (SCAM) is used to enhance contextual information across channels and spaces. The prediction head includes a classifier and a regressor, which perform data classification and regression calculations, respectively.
[0013] Preferably, the Spatial Context Awareness Module (SCAM) includes three branches: The first branch integrates global information using Global Average Pooling (GAP) and Global Max Pooling (GMP). The second branch uses 1×1 convolution to generate a linear transformation result of the feature map; The third branch uses 1×1 convolutions to simplify queries and keyword multiples; The first and third branches are multiplied by the second branch, and the resulting two branches represent the context information across channels and space, respectively. Finally, the output of SCAM is obtained by broadcasting the Hadamard product.
[0014] Preferably, the prediction head of the YOLOX model separates the classifier and the regressor, runs them separately, and finally integrates the results in the prediction stage.
[0015] Preferably, the detection method of the YOLO-based slag removal machine large object detection system for thermal power plants includes the following steps: Collect image data of the slag discharge area of the slag remover; The CycleGAN model was used to preprocess the images for smoke removal. Large object detection is performed on the preprocessed image using an improved YOLOX model; Output the detection results and take corresponding measures.
[0016] Preferably, the training steps of the CycleGAN model include: Input images with and without smoke; they do not need to be paired. The generator and discriminator are trained alternately, and the generative adversarial loss and cycle consistency loss are calculated. The effectiveness of smoke removal was evaluated using a validation set.
[0017] Preferably, the training steps of the YOLOX model include: Use pre-processed, clear images and their annotation information as training data; Features are extracted through the backbone network and then fused and enhanced using the PAFPN and SCAM modules. The classifier and regressor run separately, calculate the loss and backpropagate to update the parameters; Use the validation set to evaluate detection performance.
[0018] Preferably, an electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the YOLO-based method for detecting large slag particles discharged from a thermal power plant slag removal machine.
[0019] The beneficial effects of this invention are as follows: This solution improves image recognition by preprocessing with a smoke removal and anti-interference algorithm, effectively increasing the recognition rate of subsequent detections. The improved YOLOX detection model enhances the perception of cross-channel and spatial contextual information, improving detection capabilities. Overall, the solution significantly improves the operational safety of the slag discharge system, significantly reduces the false detection rate, and increases the system's operational efficiency. Attached Figure Description
[0020] Figure 1 This is a schematic diagram of the flowchart of the present invention; Figure 2 This is a schematic diagram of the CycleGAN model structure in an embodiment of the present invention; Figure 3 This is a schematic diagram of the generator structure in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the YOLOX model, the large object detection model in this embodiment of the invention; Figure 5 This is a schematic diagram of the SCAM module of the YOLOX model in an embodiment of the present invention. Detailed Implementation
[0021] Example 1: like Figure 1 As shown, a YOLO-based system for detecting large slag particles in the ash discharge of a thermal power plant ash removal machine includes: The image preprocessing module is used to remove smoke from images collected in the slag discharge area of the slag remover. The large object detection module uses an improved YOLOX model to detect large objects in the preprocessed image. The output module is used to output the category and location information of large objects.
[0022] Preferably, the image preprocessing module uses a CycleGAN model for image desmoke processing. The CycleGAN model includes: Generator G is used to convert a smoky image into a desmoky image; Generator F is used to restore a sharp image to a smoky image; The discriminator Dx is used to distinguish between the generated image and the real, clear image; The discriminator Dy is used to distinguish between generated images and real images with smoke.
[0023] Preferably, the network structure of the generator G and generator F is an encoder-decoder structure, the input image size is 256×256 pixels, the encoder consists of a multi-level smoke feature extraction module, and the decoder includes transposed convolution and skip connections.
[0024] Preferably, the bulk object detection module adopts the YOLOX model and includes: The backbone feature extraction network employs residual convolutional networks, CSPNet, and Focus structures. The feature fusion network employs the Path Aggregation Feature Pyramid Network (PAFPN). The Spatial Context Awareness Module (SCAM) is used to enhance contextual information across channels and spaces. The prediction head includes a classifier and a regressor, which perform data classification and regression calculations, respectively.
[0025] Preferably, the Spatial Context Awareness Module (SCAM) includes three branches: The first branch integrates global information using Global Average Pooling (GAP) and Global Max Pooling (GMP). The second branch uses 1×1 convolution to generate a linear transformation result of the feature map; The third branch uses 1×1 convolutions to simplify queries and keyword multiples; The first and third branches are multiplied by the second branch, and the resulting two branches represent the context information across channels and space, respectively. Finally, the output of SCAM is obtained by broadcasting the Hadamard product.
[0026] Preferably, the prediction head of the YOLOX model separates the classifier and the regressor, runs them separately, and finally integrates the results in the prediction stage.
[0027] Preferably, the detection method of the YOLO-based slag removal machine large object detection system for thermal power plants includes the following steps: Collect image data of the slag discharge area of the slag remover; The CycleGAN model was used to preprocess the images for smoke removal. Large object detection is performed on the preprocessed image using an improved YOLOX model; Output the detection results and take corresponding measures.
[0028] Preferably, the training steps of the CycleGAN model include: Input images with and without smoke; they do not need to be paired. The generator and discriminator are trained alternately, and the generative adversarial loss and cycle consistency loss are calculated. The effectiveness of smoke removal was evaluated using a validation set.
[0029] Preferably, the training steps of the YOLOX model include: Use pre-processed, clear images and their annotation information as training data; Features are extracted through the backbone network and then fused and enhanced using the PAFPN and SCAM modules. The classifier and regressor run separately, calculate the loss and backpropagate to update the parameters; Use the validation set to evaluate detection performance.
[0030] Preferably, an electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the YOLO-based method for detecting large slag particles discharged from a thermal power plant slag removal machine.
[0031] Example 2: This embodiment provides an architecture for detecting large objects in the slag discharge of a thermal power plant slag remover based on YOLO. The architecture consists of two parts: an image preprocessing layer and a large object detection layer. The specific process is as follows: Figure 1 As shown.
[0032] 1. CycleGAN anti-interference preprocessing: During the slag removal process of a thermal power plant slag remover, smoke or steam generated by high-temperature combustion may permeate the slag removal area, causing significant smoke interference in the images captured by the camera. Therefore, image desmearing technology is needed to remove the influence of smoke on the images. Since it is difficult to collect paired images with and without smoke during the slag removal process, CycleGAN is used to achieve image anti-interference; the overall structure is as follows. Figure 2 As shown, During the training phase, the input images are images with and without smoke. Generator G is used to convert the input foggy image x into a smoke-free image, while generator F is used to restore the clear image y to a smoke-filled image. Discriminator Dx is used to distinguish the difference between the generated image and the real clear image y, while discriminator Dy is used to distinguish between the generated image and the real smoke-filled image x.
[0033] The generator network structure is as follows: Figure 3 As shown, the overall architecture follows an encoder-decoder structure based on the generator, with an input image of 256×256 pixels. The encoder consists of multi-level smoke feature extraction modules, and the decoder can output an image of the same size as its input, including transposed convolutions and cascaded skip connections for upsampling.
[0034] 2. Large object detection model: The overall structure of the model is as follows Figure 4As shown, the YOLOX algorithm, when faced with an input image, first performs feature extraction and feature enhancement, then performs feature analysis, and finally predicts the object corresponding to the feature points.
[0035] In the backbone extraction network step, the original input image is mainly converted into multi-layer feature maps for use in subsequent object detection tasks. The YOLOX network uses residual convolutional networks, CSPnet and Focus structures.
[0036] The multi-level weight fusion section in the middle typically uses a Feature Pyramid Network (FPN) to integrate features from different levels to improve detection accuracy. However, traditional FPN structures have problems during feature fusion, such as feature redundancy and conflicts. To address these issues, YOLOX also introduced a Path Aggregation Feature Pyramid Network (PAFPN) structure. PAFPN introduces an additional path aggregation module on top of FPN, enabling the establishment of more complex connections and more effective feature fusion. Furthermore, PAFPN employs an adaptive feature fusion mechanism, dynamically adjusting the fusion weights based on feature importance and relevance to further improve feature utilization. In addition, an SCAM module is added at the final connection point with the prediction head, with the structure shown below. Figure 5 As shown. Modeling the global relationship between the detected object and the background at this stage is more effective than in the main branch. Global contextual information can be used to represent the relationships between pixels across space, which suppresses unwanted background and enhances the distinction between the object and the background. SCAM consists of three branches. The first branch, GAP and GMP, integrates global information. The second branch uses 1×1 convolutions to generate a linear transformation of the feature map, and the third branch uses 1×1 convolutions to simplify the multiplication of queries and keywords. Subsequently, the first and third branches are matrix multiplied with the second branch, respectively. The resulting two branches represent the contextual information across channels and space, respectively. Finally, the output of SCAM is obtained by using the broadcastHadamard product on these two branches.
[0037] In the prediction section, YOLOX adopted a fundamental structure different from previous versions. In all previous algorithms, the functions of the classifier and regressor were implemented in a single 1×1 convolutional network. While this approach was concise, it was also complex and cumbersome. YOLOX believed that this approach hindered further optimization and upgrades, so it separated the prediction process into two parts: a classifier and a regressor. These two parts run concurrently. For a single input image, the prediction and regression layers perform both data classification and regression calculations simultaneously. Only in the final prediction step are these data analyzed together, effectively integrating the functions of the classifier and regressor at the last step. While this increases the complexity of the algorithm's structure, the simultaneous operation of classification and regression enhances the algorithm's data processing capabilities and reduces data processing time.
[0038] Example 3: This embodiment provides a specific implementation process for detecting large objects in the slag discharge of a thermal power plant slag removal machine based on YOLO. The process is as follows: 1. Image preprocessing layer: 1.1 Data Collection: Collect image data of the ash discharge area of the thermal power plant's ash removal machine, including both smoke-filled and smokeless images. Paired image acquisition is not required. Crop or scale the images to 256×256 pixels to fit the input requirements of CycleGAN.
[0039] 1.2. Model Training: CycleGAN Model Construction: Collected images with and without smoke are used as inputs to generators G and F, respectively, for training. The generator and discriminator are trained alternately until the model converges. The specific steps are as follows: Generator G is used to convert images with and without smoke, and the generative adversarial loss is calculated. Generator F is used to restore images with and without smoke from the images without smoke, and the cycle consistency loss is calculated. Discriminators Dx and Dy are used to distinguish between the generated images without and with smoke, and the discriminator parameters are updated. The generator parameters are updated based on the generative adversarial loss and the cycle consistency loss. The model's smoke removal performance is evaluated using a validation set to ensure that the generated images without smoke are of good quality and similar to real smoke-free images.
[0040] The trained generator G will be used for subsequent image desmoke preprocessing.
[0041] 2. Large object detection layer: 2.1 Data Preparation: Using preprocessed, clear images as input, the category and location information of large objects are labeled to form a training dataset.
[0042] 2.2 Model Training: Initialize the parameters of the YOLOX model. Input the training data into the model, extract features through the backbone network, and perform feature fusion and enhancement through the PAFPN and SCAM modules. In the prediction part, the classifier and regressor run simultaneously, performing data classification and regression calculations respectively. Calculate the loss value according to the loss function, and update the model parameters through backpropagation until the model converges. Use the validation set to evaluate the model's detection performance, including metrics such as detection precision and recall, to ensure that the model can accurately detect large objects in the ash discharge of the thermal power plant ash removal machine.
[0043] 3. Applications in large object detection: In actual monitoring of slag removal from thermal power plant slag removers, pre-processed images are input into a trained YOLOX model to detect large objects in the images in real time. Based on the detection results output by the model, the location and type of large objects are determined, and corresponding measures are taken, such as alarming or controlling the operation of the slag remover.
Claims
1. A YOLO-based detection system for large slag particles in ash removal machines used in thermal power plants, characterized in that, include: The image preprocessing module is used to remove smoke from images collected in the slag discharge area of the slag remover. The large object detection module uses an improved YOLOX model to detect large objects in the preprocessed image. The output module is used to output the category and location information of large objects.
2. The YOLO-based detection system for large slag particles in thermal power plant slag removal machines according to claim 1, characterized in that, The image preprocessing module uses the CycleGAN model for image desmearing. The CycleGAN model includes: Generator G is used to convert a smoky image into a desmoky image; Generator F is used to restore a sharp image to a smoky image; The discriminator Dx is used to distinguish between the generated image and the real, clear image; The discriminator Dy is used to distinguish between generated images and real images with smoke.
3. The YOLO-based detection system for large slag particles in thermal power plant slag removal machines according to claim 2, characterized in that, The network structure of generator G and generator F is an encoder-decoder structure. The input image size is 256×256 pixels. The encoder consists of a multi-level smoke feature extraction module, and the decoder includes transposed convolution and skip connections.
4. The YOLO-based detection system for large slag particles in thermal power plant slag removal machines according to claim 1, characterized in that, The large object detection module uses the YOLOX model and includes: The backbone feature extraction network adopts residual convolutional network, CSPNet and Focus structure; The feature fusion network employs the Path Aggregation Feature Pyramid Network (PAFPN). The Spatial Context Awareness Module (SCAM) is used to enhance contextual information across channels and spaces. The prediction head includes a classifier and a regressor, which perform data classification and regression calculations, respectively.
5. The YOLO-based detection system for large slag particles in thermal power plant slag removal machines according to claim 4, characterized in that, The Spatial Context Awareness Module (SCAM) comprises three branches: The first branch integrates global information using Global Average Pooling (GAP) and Global Max Pooling (GMP). The second branch uses 1×1 convolution to generate a linear transformation result of the feature map; The third branch uses 1×1 convolutions to simplify queries and keyword multiples; The first and third branches are multiplied by the second branch, and the resulting two branches represent the context information across channels and space, respectively. Finally, the output of SCAM is obtained by broadcasting the Hadamard product.
6. The YOLO-based detection system for large slag particles in thermal power plant slag removal machines according to claim 4, characterized in that, The YOLOX model's prediction head separates the classifier and regressor, runs them separately, and finally integrates the results during the prediction phase.
7. The detection method of a YOLO-based slag removal machine large object detection system for thermal power plants according to any one of claims 1-6, characterized in that, Includes the following steps: Collect image data of the slag discharge area of the slag remover; The CycleGAN model was used to preprocess the images for smoke removal. Large object detection is performed on the preprocessed image using an improved YOLOX model; Output the detection results and take corresponding measures.
8. The method for detecting large pieces of slag discharged from a thermal power plant slag removal machine based on YOLO, as described in claim 7, is characterized in that... The training steps of the CycleGAN model include: Input images with and without smoke; they do not need to be paired. The generator and discriminator are trained alternately, and the generative adversarial loss and cycle consistency loss are calculated. The effectiveness of smoke removal was evaluated using a validation set.
9. The method for detecting large pieces of slag discharged from a thermal power plant slag removal machine based on YOLO according to claim 7, characterized in that, The training steps for the YOLOX model include: Use pre-processed, clear images and their annotation information as training data; Features are extracted through the backbone network and then fused and enhanced using the PAFPN and SCAM modules. The classifier and regressor run separately, calculate the loss and backpropagate to update the parameters; Use the validation set to evaluate detection performance.
10. An electronic device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the YOLO-based method for detecting large slag particles discharged from a thermal power plant slag removal machine as described in any one of claims 7 to 9.