Estimation of Remaining Useful Life of Rolling Element Bearings Using Wavelet Packet Decomposition and Artificial Neural Network

Rolling element bearings (REBs) are usually considered among the most critical elements of rotating machines. Therefore, accurate prediction of remaining useful life (RUL) of REBs is a fundamental challenge to improve reliability of the machines. Vibration condition monitoring is the most popular method used for diagnosis of REBs and this is a motivating fact to use recorded vibration data in RUL prediction too. However, it is necessary to extract appropriate features from vibration signal that represent actual damage progress in the REB. In this paper, wavelet packet transform is used to extract signal features and artificial neural network is applied to estimate RUL of the REB. To obtain more accurate results, a method is proposed to find appropriate mother wavelet, optimal level and optimal node for signal decomposition. The desired features were extracted from the decomposed wavelet coefficients. To reduce random fluctuations, which is essential in real-life tests, a preprocessing algorithm was applied on the raw data. A multilayer perceptron neural network was selected and trained by preprocessed input data as well as non-processed input data, and results are compared. A series of accelerated life tests were conducted on a group of radially loaded bearings and vibration signals were acquired in whole life cycle of the tested REBs. Comparison of the experimental results with the output of the trained neural network shows enhanced prediction capability of the proposed method.