Automated Fault Diagnosis Using Maximal Overlap Discret Wavelet Packet Transform and Principal Components Analysis

Bearings and gears are components most susceptible to failure in electromechanical systems, especially rotating machines. Therefore, fault detection becomes a crucial step, as well as fault diagnosis. Over decades, substantial progress in this field has been observed and numerous methods are now proposed for feature extraction from monitoring data. Among these data, vibration signals are most used. However, in the presence of non-Gaussian noise, most conventional methods may be inefficient. In this paper, a hybrid methodology is proposed to address this potential issue. The proposed methodology uses a combination of the Maximal Overlap Discrete Wavelet Packet Transform (MODWPT) and Principal Component Analysis (PCA) techniques. First, the MODWPT technique decomposes the vibration signal with uniform frequency bandwidth, facilitating effective signal processing and introducing diversity for enhanced time-frequency signals. Then, to identify significant patterns and characteristics related to faults, PCA is used for 3D dimensional representation of system health state by capturing the variance in the extracted features. Subsequently, a self-organizing map (SOM) is used for system state classification for diagnostics. This technique is applied to open-access test bench data containing vibration signals with non-Gaussian noise.