Investigations on Multiclass Classification Model based Optimized Weights Spectrum for Rotating Machinery Condition Monitoring

Abstract

Machinery condition monitoring is beneficial to equipment maintenance and has been receiving much attention from academia and industry. Machine learning, especially deep learning, has become popular for machinery condition monitoring because that can fully use available data and computational power. Since significant accidents might be caused if wrong fault alarms are given for machine condition monitoring, interpretable machine learning models, integrate signal processing knowledge to enhance trustworthiness of models, are gradually becoming a research hotspot. A previous spectrum-based and interpretable optimized weights method has been proposed to indicate faulty and fundamental frequencies when the analyzed data only contains a healthy type and a fault type. Considering that multiclass fault types are naturally met in practice, this work aims to explore the interpretable optimized weights method for multiclass fault type scenarios. Therefore, a new multiclass optimized weights spectrum (OWS) is proposed and further studied theoretically and numerically. It is found that the multiclass OWS is capable of capturing the characteristic components associated with different conditions and clearly indicating specific fault characteristic frequencies (FCFs) corresponding to each fault condition. This work can provide new insights into spectrum-based fault classification models, and the new multiclass OWS also shows great potential for practical applications.