The Potentiality of Integrating Model-Based Residuals and Machine-Learning Classifiers: An Induction Motor Fault Diagnosis Case
Widagdo Purbowaskito, Chen-Yang Lan, Kenny Fuh
Abstract
In the recent development of induction motors fault diagnosis, machine-learning algorithms have been implemented to replace the need for experts in fault diagnostic decisions. In industrial practice, faults exhibit symptoms but not in the early stage. This condition limits the availability of fault datasets for machine-learning classifier training. Therefore, the classifiers must be retrained and updated over time when the new fault datasets become available and after the classifiers have failed to diagnose faults, which can lead to catastrophic and dangerous situations. This study proposes an integrated redundant fault diagnosis framework using model-based diagnosis and machine-learning classifiers. The model-based diagnosis provides residual signals for the classifier training and early diagnosis that defines whether the fault is known or unknown to the existing classifiers. The experiments from an actual industrial centrifugal pump validate the approach and demonstrate the strength of the integration of model-based residuals and machine-learning classifiers.