A New Fault Classification Scheme Using Vibration Signal Signatures and the Mahalanobis Distance

Abstract

Condition monitoring is a vital task in the maintenance of factory automation. Many feature extraction and selection algorithms have been studied to derive distinctive feature vectors. The common extraction methods for all fault classes can be ineffective in separating a new class out of existing classes or dealing with input signals under severely noisy environments. Thus, extraction and selection algorithms might need to be redesigned. Therefore, we propose a new approach to accurately identify fault classes from vibration signals even under severely noise conditions; our approach can also easily add a new group to the classification system. The proposed algorithm, a viable alternative to detect induction motor defects online, uses the differences in the fault-related harmonics of vibration signals to generate good feature vectors. This approach discriminates the harmonics for one specific fault to generate features and then classifies faults using a modified minimum distance classifier, which improves classification accuracy. In our experiments, the proposed technique shows a clear advantage over existing methods in classification performance in both noiseless and additive white Gaussian noise circumstances and demonstrates the capability to learn new signatures from unknown motor conditions.