Bearing Fault Diagnosis Based on Prominence Peak-Picking IMFs Selection and PSO-SSAE

Bearing fault diagnosis techniques have been shifted toward using the deep learning (DL) model due to its ability to process the raw vibration signal. However, most of the deep learning models use high-speed datasets for bearing fault diagnosis; thus, the performance of the DL model on low operational speed datasets still needs to be solved. Therefore, this research proposed the integration method of particle swarm optimisation (PSO), stacked sparse autoencoder (SSAE), and empirical mode decomposition (EMD) for diagnosing the bearing fault at three-speed conditions (60, 780, and 1800 rpm) from three different experiment dataset including INV, Mafaulda, and CWRU datasets. The PSO-SSAE models produced 100% accuracy at 1800 rpm without any signal processing method involved. Meanwhile, the PSO-SSAEs’ performance drops 64 and 70% on 780 and 60 rpm datasets, respectively. The EMD method is used to preprocess these signals, and the prominence peak selection method is proposed for IMFs selection. Hence, the performance of PSO-SSAE models with EMD increased to 97% for 780 and 60 rpm datasets.