Condition monitoring and fault detection of induction motor based on wavelet denoising with ensemble learning

This paper puts forward a novel fault diagnosis scheme to detect incipient stator fault in induction motor, in addition to estimating the failure severity. Early detection of the stator fault during the motor running can improve the operational efficiency, minimize the risk of further damage to the phase winding, and ensure machine availability. The sensitive fault features are often immersed in random noise and then hard to capture. An improved method to extract the distinctive features by wavelet threshold denoising is developed in this work. Stationary wavelet transform (SWT) is employed to analyze the raw current signals in the time domain. Next, SWT denoising by thresholding is applied to acquired coefficients to accomplish the noise elimination process. Apart from this, statistical norm L-1 is computed from the error signal that clearly demonstrates unique characteristics linked to the fault. The parameter values are arranged and then inputted into the classifier to determine the motor status and qualify the fault intensity. As additional functionality, aiming at the minority training samples cannot be efficiently diagnosed when these samples are imbalanced and limited, ensemble AdaBoost decision tree (EADT) is used to implement the classification task. The EADT can improve the accuracy and solve several problems of traditional machine learning algorithms. The adopted approach is further tested under various loading situations to validate its effectiveness and robustness. To verify the practical feasibility, multiple hardware experiments are carried out on the motor. The results obtained from both Simulink and experiment affirm the superiority of the proposed method as compared to other related works, with a higher level of correctness, reaching 98.48%.