Abstract:

Objective As the textile industry transitions toward smart technology, there is an urgent need for the automation of bobbin-changing operations in winding process. Aiming to address the challenges of yarn bobbin pose estimation and gripping, deep learning algorithms is adopted to predict the pose, so as to provide key technological support for the intelligent development of the textile industry, improve production efficiency, and promote sustainable development.

Method By building a system that includes a robotic arm, camera, and other components, real and synthetic datasets were created, and online data augmentation was performed. The Swin Transformer was adopted to process the yarn bobbin color information, while KPConv was employed to extract the geometric features. After local pixel fusion, the pose was predicted, and the robotic arm was controlled to grasp the yarn bobbin based on the predicted pose.

Results Using a trained network model, the input bobbin images were processed to generate six-degrees-of-freedom pose estimation results for the bobbins. Specifically, the model could predict the poses of five different types of bobbins. Experimental results demonstrated a high level of consistency between the predicted poses and the actual poses of the bobbins. During testing, 98.7% of the predictions resulted in an average distance error of less than 10% between the estimated model points and the actual model points, indicating that the network model exhibits high accuracy and stability in the bobbin pose estimation task, effectively addressing challenges such as lighting variations and color diversity that could interfere with pose estimation. Bobbin grasping experiments were conducted on five different types of bobbins, with each set of experiments comprising 100 grasping attempts. The grasping success rate ranged from 96% to 98% across different bobbins, with an average success rate of 96.8%. Overall, the grasping success rates for the various bobbins were consistently high, with minimal differences between them. Under the experimental grasping conditions, all types of bobbins could be reliably grasped. The average response time of the pose estimation for each image frame was found in the range of 0.11 s to 0.14 s, while the average grasping response time was in the range of 2.07 s to 2.21 s. The significantly higher grasping response time is primarily attributed to the use of Python for robotic arm control, which has relatively lower execution efficiency. However, the model's efficient performance in pose estimation suggests that deploying the system on higher-performance hardware platforms would leave substantial room for overall optimization.

Conclusion The study indicates that the higher grasping response time is primarily due to the robotic arm controlled by Python, which has relatively lower execution efficiency. However, the model's high efficiency in pose estimation suggests that if deployed on a higher-performance hardware platform, the overall system performance would have significant room for optimization. Overall, this method demonstrates high accuracy and stability in the tasks of bobbin pose estimation and grasping, providing valuable insights and references for the intelligent design and practical applications in the textile industry.

Key words: deep learning, pose estimation, bobbin grasping, color feature extraction, geometric feature extraction, intelligent replacement of bobbin