Bearing Degradation Process Prediction based on Feedforward Neural Network

Abstract

As one of the most significant component of the rotary machinery, bearings play a vital role in smooth and reliable operation of the machinery. Estimating the remaining useful life (RUL) of bearings is essential for reducing the cost of maintenance and improving reliability. In this paper, a prognostics methodology based on artificial neural network (ANN) is proposed to improve the accuracy of RUL estimation for bearing. This is achieved by using features obtained from frequency, time and time-frequency domains. Popular techniques of Wavelet Packet Decomposition (WPD) and Fast Fourier Transform (FFT) are applied for feature extraction in time-frequency and frequency domains respectively. For effective prognostics, monotonicity and correlation-based feature selection criteria is used to discard redundant and unnecessary features. These features are then processed to be used as input into the ANN model. The model uses Feedforward Neural Network (FFNN) with the popular learning algorithm, Levenberg-Marquardt, for predicting the RUL. The results depict that this model is very effective for predicting the RUL of bearings. FFNN results are also compared with Gaussian Process Regression (GPR) algorithm results, showing the better performance of FFNN as compared to GPR.