TL;DR
This paper presents a novel algorithm for automatically recognizing the orientation of sealing bolts on large ball valves, crucial for automated adjustment systems. By combining wavelet transforms for noise reduction with mathematical morphology for edge detection, the method achieves robust and accurate angle recognition, even with blurred images.
What Problem Does This Paper Address?
Large ball valves, often found in high-altitude industrial settings like turbine inlets, require manual gap adjustment. This process is labor-intensive, time-consuming, and inconvenient due to the confined working space. To automate this, an auto-adjusting system is proposed, which relies on precisely determining the relative angle between a manipulator and the sealing bolt's nut. The challenge lies in accurately identifying this angle from digital images, especially given potential noise, blur, and the hexagonal shape of the bolt nut, which can lead to mismatches with the manipulator's geometry.
Methods
The core of the proposed solution is a two-stage approach:
- Image Preprocessing and Edge Detection:
- Color to Grayscale Conversion: Standard weighted average method.
- Image Segmentation: A combination of bimodal and multi-threshold methods is used to segment the image, addressing the issue of non-obvious peaks in the nut image histogram.
- Hole Filling: To maintain the integrity of nut features.
- Edge Detection: This is the key innovation. It leverages wavelet transforms for effective image denoising, followed by mathematical morphology for edge detection. Specifically, it uses a multi-shape/scale mathematical morphology operator combined with adaptive thresholding in the wavelet domain to reduce noise while preserving edge details. This approach aims to overcome the limitations of traditional edge detectors (like Sobel or Canny) which are sensitive to noise, and standard morphological operators which can have low noise-resisting ability.
- Angle Recognition:
- Coordinate Conversion: The image coordinate system is shifted to have its origin at the center of the image.
- Shape Feature Point Extraction: Key points (A, B, C, D, E, F) on the hexagonal nut are identified by calculating the distance from the center to the edge. The points with the greatest and equal distances are considered feature points.
- Angle Calculation: The relative angle between the manipulator and the bolt is determined using a rotation conversion matrix. The algorithm considers two cases: one feature point in the first quadrant (implying a 60-degree rotation) or two feature points (implying a 0-degree rotation). The angle is calculated using trigonometric functions based on the coordinates of these feature points.
Datasets
The paper does not explicitly mention specific datasets used for training or evaluation. However, it states that "The bolts with several different angles were identified and tested using the image processing methods above. The image used was 128×128 pixel in size." This suggests custom image acquisition from the Tai-shan pumped-storage power station.
Key Insights and Ideas
- Synergistic Edge Detection: The combination of wavelet transforms for denoising and mathematical morphology for edge detection proves highly effective. Wavelets handle the noise, allowing morphology to work on a cleaner image, thus improving edge integrity and reducing false positives.
- Robustness to Blur: The proposed method demonstrates good performance even on blurred images, which is a significant advantage in real-world industrial environments where image quality can be compromised.
- Feature Point Robustness: Identifying feature points based on maximum distance from the center provides a stable way to locate critical points on the hexagonal nut, even with variations in bolt position.
- Systematic Error Analysis: The paper goes beyond just calculating the angle by introducing and analyzing systematic errors, using correlation coefficients and a defined systematic error metric to quantify accuracy.
Key Results
- High Accuracy: The correlation coefficient (R) for both clockwise and counterclockwise rotations is consistently high, close to 1 (e.g., 0.990, 0.993, 0.991, 0.992).
- Low Systematic Error: The systematic error (γ) is reported to be less than 0.02, indicating high precision.
- Efficient Processing: The computing time for a 128x128 image was 29 seconds, which is considered feasible for an automated system.
- Qualitative Improvement: Figure 6 shows that the algorithm accurately detects edge details, resulting in clear, complete, and coherent outlines of the nut. Figure 8 illustrates that large errors are concentrated in the bolt's craft hole, with very small errors at the edges, confirming the algorithm's efficiency for edge recognition.
Limitations and Future Work
- Limitations:
- The paper acknowledges that systematic errors can arise from factors like sub-pixel reconstruction, illumination changes, thermal noise, and environmental conditions.
- While the method is robust to noise and blur, extremely complex or occluded images might still pose challenges.
- The paper mentions that the "large error mainly concentrates in the middle part of the image due to the bolt craft hole," suggesting that features within the hole itself are not the focus of this edge detection.
- Future Work:
- The paper implies that the system can accurately recognize the position of the bolt and that the manipulator can be exactly positioned, suggesting the system is ready for practical implementation.
- Further validation on a wider range of industrial conditions and potentially more complex valve designs could be a natural next step.
- Optimizing the processing speed further could be beneficial for real-time applications.
Summary updated