Smart machine fault diagnostics based on fault specified discrete wavelet transform

This study examines the impact of the mother wavelet, sensor selection, and machine learning (ML) models for smart fault diagnosis of rotating machines via discrete wavelet transform (DWT). The ability of Daubechies, Haar, Biorthogonal (Bior), Symlets (Sym), and Coiflets (Coif) wavelets is measured in terms of distinguishing imbalance, horizontal/vertical misalignment, and overhang/underhang bearing ball, cage, and outer race faults. For this purpose, single-step and two-step fault monitoring (SSFM and TSFM) approaches are proposed. In SSFM, the ML models detect the fault type by the healthy and faulty signals. In TSFM, the built models first determined whether the machine is faulty or not. If it is, then the models detect the fault type. As ML models, Random Forest (RF), AdaBoost with C4.5 (AB-C4.5), and two artificial neural network algorithms are trained by the features of DWT. Besides, the effect of the sensor type on the fault diagnosis is measured by considering the tachometer, microphone, and two accelerometers individually and combined. The results are interpreted regarding the evaluation metrics such as accuracy, precision, recall, confusion matrix, and model built time. It is concluded that Bior3.1 and Haar wavelets distinguish the fault type more accurately than other wavelets. Besides, the RF-Bior3.1 give the best results for SSFM and TSFM by accuracy values of 99.80% and 99.98%, respectively. It is also found that the sensor type is correlated with the selected mother wavelet.