Symmetric Radial Vectors for uncertainty-aware rotary machinery fault diagnosis

Abstract

Rotary machinery fault diagnosis faces persistent challenges in handling class imbalance, adapting to evolving fault patterns, and quantifying diagnostic uncertainty. Traditional deep learning approaches often struggle with these issues, particularly when dealing with a large number of fault categories or limited samples for rare fault types. This paper introduces a novel fault diagnosis framework based on Symmetric Radial Vectors (SRVs), specifically designed to address these technical hurdles. Our method predefines fixed, normalized vector embeddings for each fault type, serving as stable reference points in the feature space. By minimizing the spherical distance between input feature embeddings and their corresponding fault-type SRVs, we achieve robust classification even with imbalanced datasets. The predefined nature of SRVs allows for efficient handling of numerous fault categories without increasing model complexity, crucial for comprehensive fault coverage. Furthermore, the geometric properties of SRVs enable natural uncertainty quantification, as the distances to different fault-type vectors provide a direct measure of diagnostic confidence. We demonstrate the efficacy of our approach on benchmark datasets of rotary machinery faults, showing improved accuracy for rare fault classes and well-calibrated uncertainty estimates. Our method also exhibits strong adaptability to newly emerging fault types, a critical feature for evolving industrial systems.