A Comparative Study: Pre-whitening for Informative Kurtogram based Preprocessing in Bearing Fault Diagnosis using Convolutional Neural Networks

Recent technology has made it possible for wind turbines to grow both in number and size, yet still their future depends on a competitive energy production price. An effective condition monitoring not only increases the lifetime expectancy of wind turbines but also reduces their downtimes, making wind energy a more reliable and efficient sources of energy. Despite their vast usage, roller element bearings as element that carry key functionalities in wind turbines remain prone to faults, leading to failures and downtime. Many research activities have been dedicated during the last decade to unveil the true bearing health condition. However, bearing fault diagnosis remains an open challenge. The recent development of machine learning techniques enables us to tackle more complex systems such as the multi-dynamic machinery of wind turbines. Yet, an appropriate preprocessing technique is advantageous in separating effects of various dynamics to highlight fault signature. Kurtogram has been proven a powerful tool for in bearing fault detection (Nader Sawalhi and Robert B. Randall)as they provide an informative filter to distinguish the fault dynamics signature from the rest of the spectra by suggesting the bandwidth with highest kurtosis value. However, they remain susceptible to impulsiveness caused by other dynamics (e.g., imbalance in system) and the result might be deceptive detection of fault band width and center frequency. The issue would be crucial in an automated preprocessing technique for a machine learning based diagnostic system due to the absence of expert supervision who decides on exact parameters of filters. Consequently, a preliminary process is necessary in order to remove non relevant dynamics before looking at the kurtograms and a candidate for such a process is so-called Pre-Whitening. Due to the variety of techniques, a comparative study is necessary to assess advantages and disadvantages of each approach. Thus, this work has been dedicated to evaluate performance of well know prewhitening techniques using a CNN model, and highlight their obstacle and opportunities. Approach Herein an attempt has been made to outline few of available pre-whitening (Sawalhi 2007) techniques. Meanwhile, two case studies of signals with ambiguous kurtogram have been selected to evaluate the chosen pre-whitening techniques. The methodologies have been applied on the MAFAULDA dataset, which is composed of 1951 multivariate time-series acquired by a triaxial accelerometer, a microphone, and a tachometer on a SpectraQuest's Machinery Fault Simulator (MFS), simulating different states including normal operation as well as inner and outer race bearing faults. The effectiveness of the methodologies is evaluated and compared with a state-of-the-art CNN network with limited samples of vibration data. The accuracy of CNN network can represent a comparative understanding between the methodology's functionalities. Additionally, advantages and disadvantages of each technique based on the author experience have been reported.