DMAC: Discovering Multi-Attribute Correlations of Deep Quality Features for Defect Detection in Mechanical Components

Abstract

This paper introduces DMAC (Discovering Multi-Attribute Correlations of Deep Quality Features), a novel framework that enhances defect detection in mechanical components by leveraging multimodal quality-guided deep/shallow feature fusion techniques. The DMAC model addresses challenges such as the variability of defect patterns under different operational conditions and the complexity of sensor data. It combines deep and shallow learning methods to discover high-quality features that reflect the unique characteristics of mechanical operations and defect occurrences. DMAC advances the state-of-the-art by constructing a large-scale hypergraph to model the complex, multi-attribute relationships among defect features, uncovering high-order interactions that traditional methods cannot detect. Additionally, a Gaussian Mixture Model (GMM)-based posterior probability is used to rank candidate patches for defect detection, prioritizing high-risk areas and minimizing false positives. Experimental results on datasets from multiple mechanical component assessments demonstrate that DMAC effectively uncovers complex feature interactions, offering adaptive, data-driven optimization for a responsive defect detection system. By integrating hypergraph modeling and advanced feature fusion, DMAC presents a robust and accurate solution for real-world defect detection in mechanical systems, setting the stage for future advancements in multi-attribute fusion techniques and real-time adaptive learning.

Keywords

• Defect detection
• deep learning
• multi-attribute correlations
• hypergraph modeling
• GMM