Research on Trajectory Recognition and Control Technology of Real-Time Tracking Welding

Xiaohui Zhao, Yaowen Zhang, Hao Wang, Yu Liu, Bao Zhang, Shaoyang Hu

TL;DR

This paper introduces a robust real-time seam tracking system for robotic welding that leverages structured light vision. It tackles challenges like arc splash and fluctuating laser stripes with a novel adaptive feature extraction algorithm and an embedded Pauta criterion for outlier removal, achieving an average tracking error of just 0.1 mm.

What Problem Does This Paper Address?

Robotic welding, while increasingly intelligent, still faces significant challenges in accurately tracking the welding seam in real-time. Existing image processing algorithms often struggle with the dynamic and noisy environments of welding, particularly due to arc splash, fluctuating lighting, and imperfect laser stripe data. This can lead to inaccurate seam tracking, resulting in poor weld quality, system interruptions, and even safety hazards. The paper aims to improve the robustness and reliability of real-time seam tracking by addressing these issues.

Methods

The proposed system employs a multi-stage approach:

  • Region of Interest (ROI) Localization: A Convolutional Neural Network (CNN) model is used to accurately identify and localize the laser stripe region of interest (ROI) in real-time, even under varying levels of arc splash.
  • Adaptive Feature Extraction: An adaptive feature extraction algorithm is developed to extract the seam center from the laser stripe. This algorithm utilizes row and column scanning and is designed to handle continuous, discontinuous, and fluctuating laser stripes, overcoming limitations of traditional methods.
  • Prior Model for Noise Reduction: A "prior model" is employed to further refine the extracted stripe information. This model uses angle and width criteria from previous frames to filter out noise and interference, ensuring the stripe centerline is more consistent with the theoretical position.
  • Embedded Pauta Criterion for Outlier Filtering: To enhance robustness, an embedded Pauta criterion is used to segmentally process the seam center point data stream. This criterion cyclically identifies and removes outliers, preventing them from causing welding torch instability or system errors.

Datasets

The study utilized a dataset of 4134 laser images for training the CNN model. The ROI object tracking test was performed on 472 images from actual welding scenarios. The experiments were conducted using a custom-built structured light vision welding system on V-shaped seam butt-jointed specimens.

Key Insights and Ideas

  • Robustness in Harsh Environments: The combination of CNN for ROI detection, adaptive feature extraction for stripe processing, and the Pauta criterion for outlier removal creates a highly robust system capable of handling significant noise and interference common in welding.
  • Adaptive Seam Extraction: The adaptive feature extraction algorithm's ability to handle various stripe defects (holes, bulges, missing points, fluctuations) is crucial for real-world welding applications where ideal laser stripes are rare.
  • Proactive Outlier Management: The embedded Pauta criterion acts as a safety mechanism, preventing erroneous data points from impacting the welding process, thus ensuring stability and preventing accidents.
  • Pixel-Level Accuracy: The system achieves a high level of accuracy, with an average tracking error of 0.1 mm, demonstrating its potential for precise robotic welding.

Key Results

  • The CNN model achieved a predictive validation accuracy of approximately 97.0% for identifying the target area.
  • The adaptive feature extraction algorithm demonstrated strong anti-interference capabilities and adaptability to various laser stripe defects.
  • The embedded Pauta criterion effectively removed outliers, with 41 outliers removed from 870 Te_c points in one experiment.
  • The overall tracking error was maintained within ±0.2 mm, with an average error of 0.1 mm across different welding currents.
  • Experimental results showed that the tracked trajectory closely matched the seam centerline, even under challenging welding conditions.

Limitations and Future Work

  • Limitations: While the system demonstrates high robustness, the paper acknowledges that the working environment of the optical sensor is demanding, and achieving absolute correctness in every extracted center point remains a challenge. The study focused on V-shaped seams, and its performance on other joint types might require further validation.
  • Future Work: The authors suggest optimizing the matching relationship between the robot control cycle and the image processing cycle to further enhance the robustness of the welding application.

Summary updated

Abstract

Real-time tracking welding with the assistance of structured light vision enhances the intelligence of robotic welding, which significantly shortens teaching time and guarantees accuracy for user-customized product welding. However, the robustness of most image processing algorithms is deficient during welding practice, and the security regime for tracking welding is not considered in most trajectory recognition and control algorithms. For these two problems, an adaptive feature extraction algorithm was proposed, which can accurately extract the seam center from the continuous, discontinuous or fluctuating laser stripes identified and located by the CNN model, while the prior model can quickly remove a large amount of noise and interference except the stripes, greatly improving the extraction accuracy and processing speed of the algorithm. Additionally, the embedded Pauta criterion was used to segmentally process the center point data stream and to cyclically eliminate outliers and further ensure the accuracy of the welding reference point. Experimental results showed that under the guarantee of the above-mentioned seam center point extraction and correction algorithms, the tracking average error was 0.1 mm, and even if abnormal trajectory points existed, they did not cause welding torch shaking, system interruption or other accidents.
BibTeX
@inproceedings{zhao2022research,
  author = {Xiaohui Zhao and Yaowen Zhang and Hao Wang and Yu Liu and Bao Zhang and Shaoyang Hu},
  title = {{Research on Trajectory Recognition and Control Technology of Real-Time Tracking Welding}},
  booktitle = {Sensors},
  year = {2022},
  doi = {10.3390/s22218546},
}

Loading related content...

Figure