This paper aims to propose a domain-incremental learning method for weld seam tracking robots, to address the catastrophic forgetting problem that arises when seam samples are sequentially introduced, especially under production demands characterized by small batches and multiple varieties.
The weld seam tracking robot uses a light-weight semantic segmentation network, ENetx74, to extract laser stripes and uses the efficient convolution operator, independent of deep learning, to locate feature points. The proposed method uses distillation with old samples for rehearsal to retrain ENetx74 on new data sets while maintaining its performance on old ones and a memory blur strategy to reduce storage costs.
Compared to the state-of-the-art methods, the Backward Transfer, which measures the model’s performance on old tasks after learning new ones, increased by 34.8%–71.7%, while storage costs dropped by 48.45%, demonstrating the method’s effectiveness in achieving stronger generalization with less storage cost. Furthermore, seam tracking experimental results show mean tracking errors below 0.3 mm and maximum tracking errors smaller than those of the baseline, effectively mitigating the forgetting problem.
This paper proposes a domain-incremental learning method based on rehearsal for weld seam tracking robot, combining a memory blur strategy to reduce storage cost. Guided by the generalization error bound based on Vapnik–Chervonenkis dimension, the reduction in storage costs facilitates increasing the number of old samples used for rehearsal, enhancing the robot’s ability to generalize.
