ABSTRACT Understanding optimal feature engineering for motor diagnostics remains an open research area, as many signal analysis and processing strategies have yet to be systematically explored and explicitly linked to the behaviour of machine learning models. This paper explores optimal parameters and techniques of current signal acquisition and preprocessing for detecting different faults in induction motors using machine learning and motor current signature analysis. Conventional fault detection methods usually advocate for a higher frequency resolution in the motor current spectrum, which requires longer current signal measurements that are difficult and expensive to conduct and process in case of a real‐world scenario. Thus, this work aims to identify the limitations to frequency resolution for successful fault diagnosis when applying machine learning algorithms, here, a multilayer perceptron model. The goal is to provide recommendations on signal preprocessing for feature extraction to enhance the performance of the machine learning model. For example, a signal transformation method via signal multirate resampling is evaluated for the improvement of the algorithm accuracy results. The proposed methods boost not only the prediction accuracy but also may refine the system performance when considering industrial applications, as this work outlines in detail the signal acquisition and preprocessing logic. In this paper, the investigated faults include broken rotor bar, bearing fault, misalignment, eccentricity, imbalance, looseness and mixed faults. This article also proposes a new outlook on machine learning feature introspection with scatter plots, which are surprisingly rarely used in the domain of motor condition monitoring but offer a swift access to data insights at the stage of feature engineering.
Koveshnikov et al. (Thu,) studied this question.