Steel surface defect real-time detection system and method based on dynamic calibration and deep learning

Abstract

The invention relates to a steel surface defect real-time detection system and method based on dynamic calibration and deep learning, and aims to solve the technical problems of low steel surface defect detection speed and low precision. According to the technical scheme, the device comprises a conveying device, a dynamic calibration module, a control module, an image acquisition module, an image processing module, a display and alarm module and a data storage and analysis module, the method comprises the steps of steel conveying, image acquisition starting, position calibration, image acquisition, image processing, visual presentation, data analysis and data storage. According to the method, the high-resolution image acquisition equipment is combined with the deep learning algorithm, so that real-time detection and classification of the steel surface defects are realized, the problems of low steel surface defect detection efficiency and insufficient precision can be effectively solved, and an efficient and reliable solution is provided for steel production quality management.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial inspection, and particularly relates to a real-time detection system and method for steel surface defects based on dynamic calibration and deep learning. Background Art

[0002] During the steel production process, surface defects such as cracks, scratches, pits, and oxidation spots have an important impact on product quality. In existing detection technologies, the UNet model relies on segmentation algorithms and has a slow detection speed, which is not suitable for high-speed production lines; the strain gauge detection device only targets corrosion defects and has insufficient versatility. Although other solutions mention real-time performance, they do not solve the coordination problem between algorithm efficiency and hardware integration. Summary of the Invention

[0003] The purpose of the present invention is to solve the above problems and provide a real-time detection system and method for steel surface defects based on dynamic calibration and deep learning.

[0004] To solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A real-time detection system for steel surface defects based on dynamic calibration and deep learning includes a conveying device, a dynamic calibration module, a control module, an image acquisition module, an image processing module, a display and alarm module, and a data storage and analysis module;

[0006] After the dynamic calibration module senses that the steel reaches the detection area, it sends a signal to the control module to start the detection process and calibrates the detection area to ensure that the image acquisition and conveying speeds are strictly matched;

[0007] The control module receives signals from the dynamic calibration module and the image acquisition module, coordinates the operation of the conveying device, the image processing module, and the data storage and analysis module to ensure the real-time performance and data integrity of the detection process;

[0008] The image acquisition module collects the surface image of the steel through a high-resolution line-scan camera, enhances the image quality with multi-angle light sources, eliminates the interference of oil stains and water droplets, and transmits the collected image to the image processing module through a high-speed image acquisition card;

[0009] The image processing module uses the deep learning YOLOv7 algorithm based on embedded GPU hardware, and its network architecture includes three parts: a backbone network, a feature pyramid, and a detection head;

[0010] The display and alarm module is used to display the detection results in real time and trigger an audible and visual alarm;

[0011] The data storage and analysis module is used to store the detection data and supports defect distribution statistics, historical data traceability, and quality report generation;

[0012] The dynamic calibration module is installed at the entrance of the detection area on the conveying device, the image acquisition module is installed above the detection area, the control module is electrically connected to the conveying device, the dynamic calibration module, the image acquisition module, the image processing module, the display and alarm module, and the data storage and analysis module respectively. The image acquisition module is electrically connected to the image processing module, the image processing module is electrically connected to the display and alarm module and the data storage and analysis module, and the data storage and analysis module is electrically connected to the dynamic calibration module, the image acquisition module, and the display and alarm module respectively.

[0013] Furthermore, the multi-angle light sources in the image acquisition module are respectively set at 30°, 60°, and 90°, and the angles are adjustable.

[0014] Furthermore, a vibration isolation device is provided on the dynamic calibration module. After vibration isolation, the image acquisition is clear at a conveying speed of 120 m / min, and the synchronous precision error is ≤0.1 ms.

[0015] Furthermore, the backbone network of the image processing module adopts the Extended Efficient Layer Aggregation Network (E-ELAN), and a multi-gradient flow structure is constructed through grouped convolution and channel rearrangement;

[0016] The feature pyramid structure adopts the BiFPN structure to improve the detection ability for micro defects;

[0017] Three detection heads are set, corresponding to different feature scales respectively. The detection heads adopt a decoupled structure to separate the classification and regression branches, improve the anchor box parameters according to the characteristics of steel defects, and use the EIoU loss function to replace the original CIoU loss function;

[0018] A real-time detection method for steel surface defects based on dynamic calibration and deep learning includes the following steps:

[0019] Step 1) Steel conveying: The steel to be detected is placed on the conveying device and is stably and continuously transmitted to the detection area, and the transmission speed is controlled by a servo motor;

[0020] Step 2) Image acquisition start and position calibration: When the steel enters the detection area, the dynamic calibration module detects and sends a signal to the control module. The control module controls the image acquisition module to start. At the same time, the dynamic calibration module calibrates the position of the detection area to avoid image acquisition deviation;

[0021] Step 3) Image acquisition: The line scan camera of the image acquisition module real-time acquires the steel surface image. The multi-angle light sources provide uniform shadowless illumination for the steel surface, enhance the image quality and improve the contrast of the defect area. The acquired image data is transmitted to the image processing module through a high-speed image acquisition card;

[0022] Step 4) Image processing: The image processing module uses the deep learning YOLOv7 algorithm to process and analyze the collected image data. The specific steps are as follows:

[0023] Step 4.1) Feature fusion: Use the bidirectional weighted feature pyramid BiFPN to enhance the detection ability of small defects by weighted fusion of feature layers at different scales;

[0024] Step 4.2) Use the K-means++ clustering algorithm to analyze the morphological distribution of steel defects and regenerate the anchor box sizes adapted to the dataset;

[0025] Step 4.3) Loss function optimization: Adopt EIoU Loss to replace the original CIoU Loss, and combine the normalized calculation of the center point distance and aspect ratio to optimize the bounding box regression accuracy;

[0026] Step 5) Visualization presentation: The output results of the image processing module are presented in real-time visualization on the display and alarm module;

[0027] Step 6) Data analysis: Evaluate the severity of the defects presented on the display and alarm module. The specific evaluation method is as follows:

[0028] Step 6.1) Obtain the depth of the defect crack through a laser rangefinder;

[0029] Step 6.2) Use the image segmentation algorithm to calculate the defect area:

[0030] S = 0.7×Depth_normalized + 0.3×Area_normalized;

[0031] In the formula, Depth_normalized = Depth / D_max;

[0032] Area_normalized = Area / A_max;

[0033] Where: D_max is the maximum allowable defect depth of the current steel, which is dynamically determined by the steel specifications;

[0034] A_max is the maximum allowable defect area of the current steel, which is dynamically determined by the steel specifications;

[0035] Step 6.3) Set the scoring rules according to the calculated defect area;

[0036] The specific scoring rules are: S < 0.3 is minor, 0.3 ≤ S < 0.7 is medium, and S ≥ 0.7 is severe;

[0037] Step 6.4) Display and trigger a three - level alarm in real - time according to the scoring result:

[0038] The alarm triggering rules are: green light - minor; yellow light - medium; red light - severe;

[0039] Step 7) Data storage: The detection result data is stored in the data storage and analysis module 7 in a structured form, including calibration information, defect type, location, severity, and alarm information.

[0040] Furthermore, the data stored in step 7) supports multiple subsequent operations:

[0041] Including defect distribution statistics: Analyze the frequency and location distribution of defects and generate a statistical report;

[0042] Quality traceability: Combine production batch information to provide a basis for tracing quality problems;

[0043] Trend analysis: Use historical data to monitor the quality trend of the production line and provide decision - making support for process optimization.

[0044] Compared with the prior art, the beneficial effects of the present invention are:

[0045] 1. By combining a high - resolution image acquisition device with a deep - learning algorithm, the present invention realizes real - time detection and classification of steel surface defects, effectively solving the problems of low detection efficiency and insufficient accuracy of steel surface defects, and providing an efficient and reliable solution for the quality management of steel production;

[0046] 2. The advantages of the present invention are high - speed detection, meeting the continuous detection requirements of the production line; high accuracy, capable of identifying tiny defects; real - time alarm and data storage, facilitating quick response and subsequent analysis; modular design, facilitating expansion and maintenance. Brief Description of the Drawings

[0047] Figure 1 It is a schematic diagram of the system structure of the present invention;

[0048] In the figure: 1. Conveyor device; 2. Dynamic calibration module; 3. Control module; 4. Image acquisition module; 5. Image processing module; 6. Display and alarm module; 7. Data storage and analysis module. Detailed Embodiment

[0049] The present invention will be further described below with reference to the drawings and embodiments.

[0050] This system is trained and tested on 50,000 steel surface defect samples (covering types such as cracks, scratches, and pits).

[0051] Dynamic Calibration Module 2 Test: As shown in Table 1, based on the ISO 10816 vibration standard, within the vibration frequency range of 5 - 30 Hz, the synchronization error is controlled within 0.15 ms. When the vibration frequency ≤ 30 Hz, the error rate ≤ 0.5%, meeting the requirements of the high-speed production line (120 m / min);

[0052] Table 1 Dynamic Calibration Synchronization Accuracy Test

[0053]

[0054] As Figure 1 shown, a real-time steel surface defect detection system based on dynamic calibration and deep learning includes a conveying device 1, a dynamic calibration module 2, a control module 3, an image acquisition module 4, an image processing module 5, a display and alarm module 6, and a data storage and analysis module 7;

[0055] The dynamic calibration module 2 is provided with a vibration isolation device. After sensing that the steel reaches the detection area, it sends a signal to the control module 3 to start the detection process and calibrate the detection area to ensure that the image acquisition and the conveying speed are strictly matched. After vibration isolation, the dynamic calibration module 2 has no blurring in image acquisition at a conveying speed of 120 m / min, and the synchronization accuracy error ≤ 0.1 ms.

[0056] The control module 3 receives signals from the dynamic calibration module 2 and the image acquisition module 4, coordinates the operation of the conveying device 1, the image processing module 4, and the data storage and analysis module 7 to ensure the real-time nature and data integrity of the detection process;

[0057] The image acquisition module 4 acquires the steel surface image through a high-resolution line scan camera, and cooperates with multi-angle light sources respectively set at 30°, 60°, and 90° with adjustable angles to enhance the image quality, eliminate the interference of oil stains and water droplets, and the acquired image is transmitted to the image processing module 5 through a high-speed image acquisition card;

[0058] The image processing module 5 adopts the deep learning YOLOv7 algorithm based on embedded GPU hardware, and its network architecture includes three parts: a backbone network, a feature pyramid, and a detection head; the backbone network adopts the extended efficient layer aggregation network E-ELAN, and constructs a multi-gradient flow structure through grouped convolution and channel rearrangement; the feature pyramid structure adopts the BiFPN structure to enhance the detection ability for tiny defects; three detection heads are set, corresponding to different feature scales respectively. The detection head adopts a decoupled structure to separate the classification and regression branches, improves the anchor box parameters according to the characteristics of steel defects, and uses the EIoU loss function to replace the original CIoU loss function;

[0059] The display and alarm module 6 is used to display the detection results in real time and trigger an audible and visual alarm;

[0060] The data storage and analysis module 7 is used to store the detection data and support defect distribution statistics, historical data traceability, and quality report generation;

[0061] The dynamic calibration module 2 is installed at the entrance of the detection area on the conveying device 1, the image acquisition module 4 is installed above the detection area, the control module 3 is electrically connected to the conveying device 1, the dynamic calibration module 2, the image acquisition module 4, the image processing module 5, the display and alarm module 6, and the data storage and analysis module 7 respectively. The image acquisition module 4 is electrically connected to the image processing module 5, the image processing module 5 is electrically connected to the display and alarm module 6 and the data storage and analysis module 7, and the data storage and analysis module 7 is electrically connected to the dynamic calibration module 2, the image acquisition module 4, and the display and alarm module 6 respectively.

[0062] A real-time detection method for steel surface defects based on dynamic calibration and deep learning includes the following steps:

[0063] Step 1) Steel conveying: The steel to be detected is placed on the conveying device 1 and is stably and continuously transported to the detection area, and the transmission speed is controlled by a servo motor;

[0064] Step 2) Image acquisition start and position calibration: When the steel enters the detection area, the dynamic calibration module 2 detects it and sends a signal to the control module 3. The control module 3 controls the image acquisition module 4 to start. At the same time, the dynamic calibration module 2 calibrates the position of the detection area to avoid image acquisition deviation;

[0065] Step 3) Image acquisition: The line scan camera of the image acquisition module 4 real-time acquires the steel surface image. The multi-angle light source provides uniform shadowless illumination for the steel surface, enhances the image quality and improves the contrast of the defect area. The acquired image data is transmitted to the image processing module 5 through a high-speed image acquisition card;

[0066] Step 4) Image processing: The image processing module 5 uses the deep learning YOLOv7 algorithm to process and analyze the acquired image data. The specific steps are as follows:

[0067] Step 4.1) Feature fusion: Use the bidirectional weighted feature pyramid BiFPN to enhance the detection ability of small defects by weighted fusion of feature layers at different scales;

[0068] Step 4.2) Use the K-means++ clustering algorithm to analyze the morphological distribution of steel defects and regenerate the anchor box sizes adapted to the dataset. For example, the aspect ratio of the crack anchor box is 3:1, and the aspect ratio of the pit anchor box is 1:1;

[0069] Step 4.3) Loss function optimization: Replace the original CIoU Loss with EIoU Loss, and combine the normalized calculation of the center point distance and aspect ratio to optimize the bounding box regression accuracy;

[0070] Step 5) Visualization: The output result of the image processing module 5 is visually presented in real time on the display and alarm module 6;

[0071] Step 6) Data analysis: Evaluate the severity of the defects presented by the display and alarm module 6. The specific evaluation method is as follows:

[0072] Step 6.1) Obtain the depth of the defect crack through a laser rangefinder;

[0073] Step 6.2) Calculate the defect area using an image segmentation algorithm:

[0074] S = 0.7×Depth_normalized + 0.3×Area_normalized;

[0075] Where Depth_normalized = Depth / D_max;

[0076] Area_normalized = Area / A_max;

[0077] Among them: D_max is the maximum allowable defect depth of the current steel, which is dynamically determined by the steel specifications;

[0078] A_max is the maximum allowable defect area of the current steel, which is dynamically determined by the steel specifications;

[0079] Step 6.3) Set the scoring rules according to the calculated defect area;

[0080] The specific scoring rules are: S < 0.3 is minor, 0.3 ≤ S < 0.7 is medium, and S ≥ 0.7 is severe;

[0081] Step 6.4) Display and trigger a three-level alarm in real time according to the scoring results:

[0082] The alarm trigger rules are: green light - minor; yellow light - medium; red light - severe;

[0083] This design ensures the timely discovery and response to abnormal situations.

[0084] Step 7) Data storage: The detection result data is stored in the data storage and analysis module 7 in a structured form, including calibration information, defect type, location, severity, and alarm information. The stored data supports multiple subsequent operations: including defect distribution statistics: analyzing the frequency and location distribution of defects and generating a statistical report; quality traceability: providing a basis for tracing quality problems in combination with production batch information; trend analysis: using historical data to monitor the quality trend of the production line and providing decision-making support for process optimization.

[0085] Compared with the traditional UNet model, the improved YOLOv7 algorithm has increased the crack detection accuracy from 78% to 92% (a 14% increase), and the detection speed reaches ≤10 ms / frame (the traditional method is 1 - 5 frames per second). Through the collaborative imaging of a high-resolution line scan camera (16k pixels) and multi-angle light sources (30°, 60°, 90°), the system can stably identify 0.1 mm micro-defects, with a three-fold improvement in accuracy compared to the traditional method (≥0.3 mm). The dynamic calibration module 2 (synchronization error ≤0.1 ms) supports a high-speed production line of 120 m / min and ensures clear imaging through a vibration isolation device (see Table 1). The anti-interference test shows that the multi-angle light source configuration can eliminate the influence of oil stains and water droplets on imaging, reducing the defect miss detection rate to less than 0.5%.

Claims

1. A real-time detection system for steel surface defects based on dynamic calibration and deep learning, characterized in that, It includes a conveying device (1), a dynamic calibration module (2), a control module (3), an image acquisition module (4), an image processing module (5), a display and alarm module (6), and a data storage and analysis module (7); After the dynamic calibration module (2) senses that the steel arrives at the detection area, it sends a signal to the control module (3) to start the detection process and calibrates the detection area to ensure that the image acquisition and conveying speed are strictly matched; The control module (3) receives signals from the dynamic calibration module (2) and the image acquisition module (4), coordinates the operation of the conveying device (1), the image processing module (4), and the data storage and analysis module (7), and ensures the real-time performance and data integrity of the detection process; The image acquisition module (4) acquires the surface image of the steel through a high-resolution line scan camera, and cooperates with multi-angle light sources to enhance the image quality, eliminate the interference of oil stains and water droplets, and the acquired image is transmitted to the image processing module (5) through a high-speed image acquisition card; The image processing module (5) adopts the deep learning YOLOv7 algorithm based on the embedded GPU hardware, and its network architecture includes three parts: a backbone network, a feature pyramid, and a detection head; The display and alarm module (6) is used to display the detection results in real time and trigger an audible and visual alarm; The data storage and analysis module (7) is used to store the detection data and supports defect distribution statistics, historical data traceability, and quality report generation; The dynamic calibration module (2) is installed at the entrance of the detection area on the conveying device (1), the image acquisition module (4) is installed above the detection area, the control module (3) is electrically connected to the conveying device (1), the dynamic calibration module (2), the image acquisition module (4), the image processing module (5), the display and alarm module (6), and the data storage and analysis module (7) respectively, the image acquisition module (4) is electrically connected to the image processing module (5), the image processing module (5) is electrically connected to the display and alarm module (6) and the data storage and analysis module (7), and the data storage and analysis module (7) is electrically connected to the dynamic calibration module (2), the image acquisition module (4), and the display and alarm module (6) respectively.

2. The real-time steel surface defect detection system based on dynamic calibration and deep learning according to claim 1, characterized in that The multi-angle light sources in the image acquisition module (4) are respectively set at 30°, 60°, and 90°, and the angles are adjustable.

3. A real-time steel surface defect detection system based on dynamic calibration and deep learning according to claim 1, characterized in that The dynamic calibration module (2) is provided with a vibration isolation device. After vibration isolation, the image acquisition is not blurred at a conveying speed of 120 m / min, and the synchronization accuracy error ≤ 0.1 ms.

4. A real-time steel surface defect detection system based on dynamic calibration and deep learning according to claim 1, characterized in that, The backbone network of the image processing module (5) adopts an extended efficient layer aggregation network (E-ELAN), and constructs a multi-gradient flow structure through grouped convolution and channel rearrangement; The feature pyramid structure adopts a BiFPN structure to improve the detection ability for tiny defects; Three detection heads are provided, corresponding to different feature scales respectively. The detection head adopts a decoupled structure to separate the classification and regression branches, improves the anchor box parameters according to the characteristics of steel defects, and adopts the EIoU loss function to replace the original CIoU loss function.

5. The real-time detection method for steel surface defects based on dynamic calibration and deep learning according to any one of claims 1-4, characterized in that, It includes the following steps: Step 1) Steel transportation: The steel to be detected is placed on the conveying device (1) and stably and continuously transported to the detection area, and the transmission speed is controlled by a servo motor; Step 2) Image acquisition startup and position calibration: When the steel enters the detection area, the dynamic calibration module (2) detects it and sends a signal to the control module (3). The control module (3) controls the image acquisition module (4) to start. At the same time, the dynamic calibration module (2) performs position calibration on the detection area to avoid image acquisition deviation; Step 3) Image acquisition: The line scan camera of the image acquisition module (4) continuously acquires the surface image of the steel. The multi-angle light source provides uniform shadowless illumination for the steel surface, enhancing the image quality and increasing the contrast of the defect area. The acquired image data is transmitted to the image processing module (5) through a high-speed image acquisition card; Step 4) Image processing: The image processing module (5) uses the deep learning YOLOv7 algorithm to process and analyze the acquired image data. The specific steps are as follows: Step 4.1) Feature fusion: Use the Bidirectional Feature Pyramid Network (BiFPN) to enhance the detection ability of small defects by weighted fusion of feature layers at different scales; Step 4.2) Use the K-means++ clustering algorithm to analyze the morphological distribution of steel defects and regenerate the anchor box sizes adapted to the dataset; Step 4.3) Loss function optimization: Adopt EIoU Loss to replace the original CIoU Loss, and combine the normalized calculation of the center point distance and aspect ratio to optimize the boundary box regression accuracy; Step 5) Visualization presentation: The output result of the image processing module (5) is presented visually in real time on the display and alarm module (6); Step 6) Data analysis: Severity assessment is carried out on the defects presented on the display and alarm module (6). The specific assessment method is as follows: Step 6.1) Obtain the depth of the defect crack through a laser rangefinder; Step 6.2) Use an image segmentation algorithm to calculate the defect area: S = 0.7×Depth_normalized + 0.3×Area_normalized; Where, Depth_normalized = Depth / D_max; Area_normalized = Area / A_max; Among them: D_max is the maximum allowable defect depth of the current steel, which is dynamically determined by the steel specification; A_max is the maximum allowable defect area of the current steel, which is dynamically determined by the steel specification; Step 6.3) Set the scoring rules according to the calculated defect area; The specific scoring rules are: S < 0.3 is minor, 0.3 ≤ S < 0.7 is medium, and S ≥ 0.7 is severe; Step 6.4) Display and trigger a three-level alarm in real time according to the scoring result: The alarm trigger rules are: green light - minor; yellow light - medium; red light - severe; Step 7) Data storage: The detection result data is stored in the data storage and analysis module (7) in a structured form, including calibration information, defect type, location, severity, and alarm information.

6. The real-time steel surface defect detection method based on dynamic calibration and deep learning according to claim 5, characterized in that The data stored in step 7) supports a variety of subsequent operations: Including defect distribution statistics: Analyze the frequency and location distribution of defects and generate a statistical report; Quality traceability: Provide a basis for tracing quality problems in combination with production batch information; Trend analysis: Use historical data to monitor the quality trend of the production line and provide decision-making support for process optimization.

Images