Method for detecting incomplete slurry laying in capacitor production process based on YOLO model

Abstract

The invention relates to a capacitor production process slurry spreading incomplete detection method based on a YOLO model. The method comprises the following steps: S1, acquiring real-time image data of a slurry spreading procedure in a capacitor production process on line by adopting a high-resolution industrial camera; s2, marking defect categories and defect target coordinates for the acquired slurry spreading process image data, and constructing a data set; s3, expanding a slurry laying image data set by adopting a data preprocessing method, and improving the robustness and the detection precision of a detection model; s4, on the basis of the expanded slurry laying process data set, model parameters are updated through a back propagation algorithm, and training of a YOLO model is completed; and S5, deploying the trained model in high-performance edge computing equipment, and realizing real-time online detection of the incomplete slurry laying defect of the capacitor production line by combining an industrial camera, a light source and a three-color alarm lamp. According to the method, a YOLO-based single-stage target detection framework is used, and efficient identification of the slurry paving defect is realized through high-precision image acquisition, adaptive data enhancement and optimization of a model training strategy.

Description

Technical Field

[0001] This invention relates to the technical field of industrial defect detection, and in particular to a method for detecting incomplete slurry spreading in capacitor production processes based on the YOLO model. Background Technology

[0002] Currently, capacitor slurry quality inspection mainly relies on manual visual inspection or traditional image processing methods. Manual visual inspection suffers from low efficiency, high cost, strong subjectivity, and susceptibility to fatigue, leading to high rates of missed detections and false detections. Traditional image processing methods, typically based on threshold segmentation and edge detection algorithms, are sensitive to factors such as lighting changes and noise interference, require manually designed feature extraction rules, have weak generalization ability, and are difficult to adapt to complex and changing production environments. Therefore, there is an urgent need to propose a solution to address the shortcomings in quality inspection of the slurry spreading process in current capacitor production. Summary of the Invention

[0003] To address the technical problems existing in the prior art, this invention provides a method for detecting incomplete slurry spreading in the capacitor production process based on the YOLO model. This method addresses the shortcomings of existing quality inspection of the slurry spreading process in capacitor production by using a YOLO-based single-stage target detection framework. Through high-precision image acquisition, adaptive data augmentation, and optimized model training strategies, it achieves efficient identification of slurry spreading defects.

[0004] This invention employs the following technical solution: a method for detecting incomplete slurry spreading in the capacitor production process based on the YOLO model, comprising the following steps: S1. A high-resolution industrial camera is used to collect real-time image data of the slurry spreading process during capacitor production online, and the imaging quality is ensured by combining the light source illumination system. S2. Label the defect categories and target coordinates of the collected grouting process image data, and construct a dataset containing complete grouting and partial grouting defects; S3. Data preprocessing methods are used to expand the slurry spreading image dataset to improve the robustness and detection accuracy of the detection model; S4. Based on the expanded slurry spreading process dataset, the model parameters are updated through the backpropagation algorithm to complete the training of the YOLO model; S5. Deploy the trained model on a high-performance edge computing device, and combine it with an industrial camera, light source and three-color alarm light to realize real-time online detection of defects in incomplete slurry spreading in capacitor production lines.

[0005] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. This invention improves model performance by constructing a multi-scale feature fusion network and optimizing the Mosaic data augmentation strategy, adaptive anchor frame calculation, and CIoU loss function, thereby achieving high-precision, real-time slurry quality detection.

[0006] 2. This invention effectively solves the problems of low efficiency and high missed detection rate of traditional manual inspection, and provides a reliable automated solution for quality control in the capacitor production process. Attached Figure Description

[0007] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is the detection result image for test set 1; Figure 3 This is the detection result image for test set 2; Figure 4 This is the detection result image for test set 3. Detailed Implementation

[0008] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0009] Example like Figure 1 As shown in this embodiment, a method for detecting incomplete slurry spreading in the capacitor production process based on the YOLO model includes the following steps: S1. A high-resolution industrial camera is used to collect real-time image data of the slurry spreading process during capacitor production online, and the imaging quality is ensured by combining the light source illumination system. S2. Label the defect categories and target coordinates of the collected grouting process image data to construct a dataset containing complete grouting and partial grouting defects. S3. Data preprocessing methods are used to expand the slurry spreading image dataset to improve the robustness and detection accuracy of the detection model; S4. Based on the expanded slurry spreading process dataset, the model parameters are updated through the backpropagation algorithm to complete the training of the YOLO model; S5. Deploy the trained model on a high-performance edge computing device, and combine it with an industrial camera, light source and three-color alarm light to realize real-time online detection of defects in incomplete slurry spreading in capacitor production lines.

[0010] Specifically, in this embodiment, in step S1, the acquisition system is triggered synchronously with the proximity switch to accurately capture the key process point image sequence of each electrode sheet during the slurry laying process, providing comprehensive visual data support for subsequent defect detection. The acquisition scheme has been tested in practice and can still maintain a 100% effective image capture rate under normal production line conditions.

[0011] Specifically, in this embodiment, the specific process of step S2 is as follows: The collected image data of the slurry spreading process were annotated: Defect category labeling: Label the defect category as "incomplete grouting"; Defect target coordinate annotation: Use pixel-level precision rectangular box annotation to accurately mark the center coordinates (x, y) and width and height dimensions (w, h) of the defect area; A high-quality dataset containing 3,000 labeled samples was constructed, in which complete grouting samples and defective samples were evenly distributed in a 1:1 ratio. This dataset provides a reliable benchmark for subsequent model training.

[0012] Specifically, the data annotation tool used in step S2 is Labelimg, and the annotation content is the defect category, namely the coordinates of the upper left and lower right corners of the defect target and the incomplete grouting. The coordinate information file format after annotation is txt.

[0013] Specifically, in this embodiment, the specific process of step S3 is as follows: Preprocessing is performed using YOLO's built-in data augmentation methods. Mosaic four-image stitching is used to improve the ability to identify small defects. The prior box size is optimized through adaptive anchor box calculation, and limited HSV color perturbations are applied, specifically hue ±0.015, saturation ±0.7, and lightness ±0.4, as well as geometric transformations, specifically rotation ±10° and translation ±0.1. This data augmentation method, through Mosaic stitching and adaptive parameter optimization, significantly improves the model's accuracy and robustness in identifying small defects while ensuring real-time detection speed.

[0014] Specifically, in this embodiment, the core principle of the Mosaic data augmentation method used in step S3 is to randomly select four training images, perform scaling, flipping, and color adjustments on each, and then stitch them together on a large canvas to form a composite image containing multiple scenes. This hybrid approach not only increases the diversity of targets and the complexity of the background, but also allows the model to learn targets of different scales and locations simultaneously in a single training iteration, especially improving the detection capabilities for small targets and occluded scenes. Furthermore, Mosaic combines random cropping and stitching strategies to ensure richer training data, thereby significantly enhancing the model's generalization performance and making it one of the key technologies for efficient YOLO model training.

[0015] Specifically, in this embodiment, the specific process of step S4 is as follows: Based on the enhanced slurry spreading process dataset, the YOLO model is trained end-to-end using the backpropagation algorithm: Network optimization: A Focus module is introduced into the Backbone part of the network architecture to reduce computational complexity, while PANet feature pyramid is used to enhance multi-scale defect detection capabilities; Training strategy: SGD optimizer is used in conjunction with cosine annealing learning rate scheduling, and EMA (Exponential Moving Average) model weight averaging technique is introduced to improve training stability; Loss function: CIoU Loss (Complete-IoU) is used instead of traditional IoU, which comprehensively considers three indicators: overlapping area, center point distance and aspect ratio, effectively improving the accuracy of bounding box regression.

[0016] Specifically, in this embodiment, the YOLO model in step S4 is an efficient target detection model based on a single-stage detection framework. Its core architecture includes three key parts: Backbone, Neck, and Head. During training, Backbone first extracts multi-scale features through the CSPDarknet53 structure, which adopts the CSPNet design to enhance feature representation while reducing computational complexity. In the early stage of training, adaptive anchor box calculation is performed to automatically optimize the preset anchor box size to adapt to the specific dataset. The Neck part performs multi-level feature fusion by combining PANet and FPN, enabling the model to capture target information at different scales simultaneously. The Head part predicts the target's class probability, confidence, and bounding box offset based on the anchor box mechanism. During training, CIoU Loss is used to optimize the localization accuracy, and BCE Loss is used to process classification and confidence prediction. The calculation formulas are shown in formulas (1) and (4), respectively. The entire training process adopts mixed precision acceleration technology and introduces a cosine annealing learning rate scheduling strategy to optimize the convergence process. In the inference stage, the model processes the prediction results through non-maximum suppression and finally outputs efficient and accurate detection boxes. The specific formula process is as follows: (1) in, IoU It is a prediction box b and real frame b gt The intersection and union ratio; p It is the Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box; c It is the length of the diagonal of the smallest bounding rectangle; v It is a measure of aspect ratio consistency; α These are weighting coefficients; Specifically, v and α The calculation processes are as follows: formulas (2) and (3): (2) (3) in, w gt It is the width of the actual frame; hgt It is the actual height of the bounding box; w It is the width of the prediction box; h It is the height of the prediction box; (4) in, y Is it a real label 0 or 1? p It is the probability predicted by the model.

[0017] Specifically, in this embodiment, the specific process of step S5 is as follows: The trained YOLO model was deployed to the Horizon RDK x5 edge computing platform to build a complete intelligent detection system. Hardware integration: An industrial camera is used in conjunction with a strip LED light source to build a high-precision imaging unit, and I / O linkage with the alarm light is achieved through the GPIO interface; Software deployment: Develop a Java-based front-end web interface to visualize the test results; Performance verification: In continuous production line testing, the system achieved a detection accuracy of 98.5% (missed detection rate <0.5%), an average response time of 18ms, and can trigger an alarm signal in real time when a defect is detected.

[0018] Specifically, in this embodiment, the intelligent detection system in step S5 consists of an industrial camera, a bar light source, an edge computing board, a proximity switch, an alarm light, and a web front-end interface; An industrial camera, light source, and proximity switch constitute an image data acquisition system for the slurry spreading process, used to acquire images of the slurry spreading process in capacitor production in real time. The edge computing board serves as the central processor of the intelligent detection system. It is used to store images of the slurry spreading process collected by the acquisition system, execute forward inference operations of the YOLO model, and display the model inference results. The alarm light communicates with the calculation board, displaying green when there is no incomplete grouting and red when there is incomplete grouting. The web front-end interface is used to display the model detection results and statistical data of the detection results.

[0019] Specifically, the inspection result diagram of the present invention is as follows: Figure 2 , Figure 3 , Figure 4 As shown.

[0020] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for detecting incomplete slurry spreading in capacitor production process based on the YOLO model, characterized in that, Includes the following steps: S1. A high-resolution industrial camera is used to collect real-time image data of the slurry spreading process during capacitor production online, and the imaging quality is ensured by combining the light source illumination system. S2. Label the defect categories and target coordinates of the collected grouting process image data, and construct a dataset containing complete grouting and partial grouting defects; S3. Data preprocessing methods are used to expand the slurry spreading image dataset to improve the robustness and detection accuracy of the detection model; S4. Based on the expanded slurry spreading process dataset, the model parameters are updated through the backpropagation algorithm to complete the training of the YOLO model; S5. Deploy the trained model on a high-performance edge computing device, and combine it with an industrial camera, light source and three-color alarm light to realize real-time online detection of defects in incomplete slurry spreading in capacitor production lines.

2. The method for detecting incomplete slurry spreading in capacitor production process based on the YOLO model according to claim 1, characterized in that, In step S1, the acquisition system is triggered synchronously with the proximity switch to capture the image sequence of key process points in the slurry laying process of each electrode sheet, providing comprehensive visual data support for subsequent defect detection.

3. The method for detecting incomplete slurry spreading in capacitor production process based on the YOLO model according to claim 1, characterized in that, The specific process of step S2 is as follows: The collected image data of the slurry spreading process were annotated: Defect category labeling: Label the defect category as "incomplete grouting"; Defect target coordinate annotation: Use pixel-level precision rectangular box annotation to accurately mark the center coordinates (x, y) and width and height dimensions (w, h) of the defect area; A high-quality dataset containing several sets of labeled samples is constructed, in which complete grouting samples and defective samples are evenly distributed in a 1:1 ratio. This dataset provides the baseline ground values ​​for subsequent model training.

4. The method for detecting incomplete slurry spreading in capacitor production process based on the YOLO model according to claim 3, characterized in that, The data annotation tool used in step S2 is Labelimg. The annotation content is the defect category, namely the coordinates of the upper left and lower right corners of the defect target for incomplete grouting. The coordinate information file after annotation is in txt format.

5. The method for detecting incomplete slurry spreading in capacitor production process based on the YOLO model according to claim 1, characterized in that, The specific process of step S3 is as follows: Preprocessing is performed based on YOLO's built-in data augmentation methods. Mosaic four-image stitching is used to improve the ability to identify small defects. The prior box size is optimized through adaptive anchor box calculation, and limited HSV color perturbations are applied: hue ±0.015, saturation ±0.7, brightness ±0.4, and geometric transformations: rotation ±10°, translation ±0.

1. This data augmentation method improves the model's accuracy and robustness in identifying small defects while increasing real-time detection speed through Mosaic stitching and adaptive parameter optimization.

6. The method for detecting incomplete slurry spreading in capacitor production process based on the YOLO model according to claim 5, characterized in that, The Mosaic data augmentation method used in step S3 involves randomly selecting several training images, scaling, flipping, and color adjusting them respectively, and then stitching them together on a large canvas to form a composite image containing multiple scenes.

7. The method for detecting incomplete slurry spreading in capacitor production process based on the YOLO model according to claim 1, characterized in that, The specific process of step S4 is as follows: Based on the enhanced slurry spreading process dataset, the YOLO model is trained end-to-end using the backpropagation algorithm: Network optimization: A Focus module is introduced into the Backbone part of the network architecture to reduce computational complexity, while PANet feature pyramid is used to enhance multi-scale defect detection capabilities; Training strategy: SGD optimizer is used in conjunction with cosine annealing learning rate scheduling, and EMA model weight averaging technique is introduced to improve training stability; Loss function: CIoU Loss is used, which comprehensively considers three indicators: overlapping area, center point distance, and aspect ratio, to improve the accuracy of bounding box regression.

8. The method for detecting incomplete slurry spreading in capacitor production process based on the YOLO model according to claim 7, characterized in that, The core architecture of the YOLO model in step S4 consists of three key parts: Backbone, Neck, and Head. During training, Backbone first extracts multi-scale features through the CSPDarknet53 structure, which is designed using CSPNet to reduce computational complexity while enhancing feature representation. Adaptive anchor box calculation is performed in the early stage of training to automatically optimize the preset anchor box size to adapt to the specific dataset. The Neck part performs multi-level feature fusion by combining PANet and FPN, enabling the model to capture target information at different scales simultaneously. The Head part predicts the target's class probability, confidence, and bounding box offset based on the anchor box mechanism. CIoU Loss is used to optimize localization accuracy during training, and BCE Loss is used to process classification and confidence prediction. The calculation formulas are shown in formulas (1) and (4), respectively. The entire training process adopts mixed precision acceleration technology and introduces a cosine annealing learning rate scheduling strategy to optimize the convergence process. In the inference stage, the model processes the prediction results through non-maximum suppression and finally outputs efficient and accurate detection boxes. The specific calculation process is as follows: （1） in, IoU It is a prediction box b and real frame b gt The intersection and union ratio; p It is the Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box; c It is the length of the diagonal of the smallest bounding rectangle; v It is a measure of aspect ratio consistency; α These are weighting coefficients; Specifically, v and α The calculation processes are as follows: formulas (2) and (3): （2） （3） in, w gt It is the width of the actual frame; h gt It is the actual height of the bounding box; w It is the width of the prediction box; h It is the height of the prediction box; （4） in, y Is it a real label 0 or 1? p It is the probability predicted by the model.

9. The method for detecting incomplete slurry spreading in capacitor production process based on the YOLO model according to claim 1, characterized in that, The specific process of step S5 is as follows: The trained YOLO model was deployed to the Horizon RDK x5 edge computing platform to build a complete intelligent detection system. Hardware integration: An industrial camera is used in conjunction with a strip LED light source to build a high-precision imaging unit, and I / O linkage with the alarm light is achieved through the GPIO interface; Software deployment: Develop a Java-based front-end web interface to visualize the test results; Performance verification: During continuous operation testing on the production line, the system achieved a detection accuracy rate of 98.5%, an average response time of 18ms, and triggered an alarm signal in real time when a defect was detected.

10. The method for detecting incomplete slurry spreading in capacitor production process based on the YOLO model according to claim 9, characterized in that, The intelligent detection system in step S5 consists of an industrial camera, a bar light source, an edge computing board, a proximity switch, an alarm light, and a web front-end interface; An industrial camera, light source, and proximity switch constitute an image data acquisition system for the slurry spreading process, used to acquire images of the slurry spreading process in capacitor production in real time. The edge computing board serves as the central processor of the intelligent detection system. It is used to store images of the slurry spreading process collected by the acquisition system, execute forward inference operations of the YOLO model, and display the model inference results. The alarm light communicates with the calculation board, displaying green when there is no incomplete grouting and red when there is incomplete grouting. The web front-end interface is used to display the model detection results and statistical data of the detection results.

Images