Cross-Scale Feature Blending Model for Surface Defect Identification in Machine Tool Elements Resilient to Contaminant Interference

Abstract

Ball screw drives (BSDs) play a crucial role in various industrial applications, particularly in CNC machines. However, they are vulnerable to external variables potentially leading to premature wear, degradation, and failure. Traditional fault diagnosis techniques reliant on dynamic modeling, vibration, and acoustics struggle to address the intricate dynamic parameters and operating conditions sensitivity like rotating speed, loads, external vibration influences, and ambient noise in real-world scenarios. In this context, vision-oriented techniques provide a promising alternative by directly processing visual data from camera-fitted sensors, providing a more comprehensive understanding of system conditions. Despite the merits of vision-oriented methods, challenges arise during image data measurement due to contaminating factors such as oil pollution or foreign material intrusion. To tackle issues related to surface damage detection and diagnosis resilient to contaminant interference in ball screw drives, a novel architectural framework called Cross-Scale Feature Blending (CSFB) is proposed. This framework, inspired by the fundamental U-Net architecture, is developed with three crucial modules: the Nonlinear Activation-Free (NAF) block, the Feature Blending Block(FBB), and the Focused Linear Attention (FLA) module. The CSFB architecture innovates a distinct mechanism to feature fusion, where each layer receives direct input from all previous layers to blend entire-scale features up to the top layer, addressing discrepancies between features visible at multiple scales. Furthermore, the CSFB framework incorporates assigned weights to guide the network during training, prioritizing features with a significant impact on the final decisions, thereby enhancing model performance and flexibility in handling complex image processing tasks. Two case studies for surface damage detection and diagnosis in ball screw drives are presented to validate the effectiveness of the CSFB model. These studies demonstrate the model’s robustness in various scenarios, validating its potential as a valuable tool for ensuring the reliability and longevity of ball screw drives in machine tool elements.

Keywords

• Ball screw drives
• defect identification
• fault diagnosis
• feature fusion
• image restoration