Scale-Aware Domain Adaptation for Surface Defects Detection on Machine Tool Components in Contaminant Measurements

Machine tool elements, including ball screw drives (BSDs) in CNC machines, play a critical role but are vulnerable to wear due to external factors. Existing approaches in the field of fault diagnosis and health condition monitoring involve the examination of factors such as speed, loads, and external influences in real-world operating conditions. Specifically, these deep learning methods face challenges associated with the requirement for extensive historical datasets for training, a constraint that may not be available in numerous practical scenarios. Furthermore, their performance heavily depends on robust data correlations. On the other hand, vision-oriented techniques, utilizing visual data from sensors, provide a more comprehensive understanding and an advantage in accurately detecting surface damage. Nevertheless, they encounter challenges such as contamination by oil or other contaminants. To address these challenges, the scale-aware domain adaptation (SADA) framework has been proposed. This methodology incorporates three designed modules: the focused linear attention modules (FLAM), the scale-aware feature stacking (SAFS) block, and the nonlinear activation-free (NAF) block. Initially, SADA is trained on an open-access source domain dataset. Utilizing the SAFS block allows for the prioritization of influential features during the learning process through weighted assessments. Meanwhile, FLAM is designed to refine the efficiency of attention mechanisms for data sequences of considerable length. The NAF block aids in effortlessly fusing features at various scales. Upon this foundational training, the weights from the Encoder and Decoder are generalized for application to the target domain, specifically datasets concerning BSDs. SADA’s effectiveness is evidenced through case studies in accurately detecting surface damage on BSDs, contributing to the improved durability of machine tool components.