Unsupervised Pitting Defect Detection on Ball Screw Surfaces Based on Collaborative Reconstruction with Dual AutoEncoders

Early pitting defects on the surface of machine tool ball screws are characterized by small size, susceptibility to oil contamination, and limited sample availability, which pose significant challenges for accurate detection. To address this issue, this paper proposes an unsupervised defect detection method based on dual autoencoder collaborative reconstruction. The proposed method constructs a cascaded architecture consisting of a simple and a complex autoencoder. The simple autoencoder performs a masked reconstruction task to generate a pixel-level difficulty score map, which is used to quantify the reconstruction difficulty of different regions. This score map further guides the complex autoencoder to focus on high-difficulty regions, enabling high-fidelity reconstruction while suppressing over-generalization on minor anomalies.