PROGNOSTIC ANALYSIS OF ROLLING BEARINGS USING ONE DIMENSIONAL CONVOLUTIONAL NEURAL NETWORK

Abstract

The accurate prediction of the remaining useful life (RUL) of bearings plays a vital role in ensuring the reliability and efficiency of industrial machinery. The main aim of this study is prognostic analysis of rolling bearing for RUL prediction. In this study, signal processing methods are used to denoise the signal and features are extracted using both time-domain and frequency-domain analysis. Prognostic sensitive features are selected using feature ranking metrics namely, monotonicity and trendability. The respective contribution of feature ranking metric is taken into account using equal weighting factor. Further, feature fusion is carried out using PCA (principal component Analysis) technique. The reduced dimensional features are then utilized to obtain unidirectional health indicator (HI). The bearing lifetime response using HI is then modeled using Exponential degradation model which is then utilized to make new predictions. The exponential degradation model suffers with two major limitations namely, subjective selection of the first fault predicting time (FFPT) and the reduced accuracy due to random errors in the stochastic process. To overcome these issues, the Wasserstein distance (WD) method is implemented to compute FFPT. The WD method measures the difference between the probability distributions of normal condition features and faulty condition features thereby enabling accurate estimation of FFPT for accurate RUL estimations. Subsequently, One-dimensional convolutional neural network (1D-CNN) is implemented for online condition monitoring of bearing data. The 1D-CNN is employed to extract informative features from the sequential sensor data collected from the bearings. By employing multiple convolutional layers with varying filter sizes, the 1D-CNN effectively captures spatial dependencies within the sensor readings, enabling it to learn complex patterns related to bearing degradation. Besides this, the long short-term memory (LSTM) network is utilized to model the long term temporal dependencies for sequential time-series for the extracted features. The LSTM network leverages its memory cells and gating mechanisms to retain relevant information and capture the sequential dependencies, further enhancing the accuracy of the RUL predictions. To optimize the performance of the LSTM model, Bayesian optimization is implemented to optimally select the both model and training parameters for better accuracy of RUL predictions. Furthermore, the performance of the deep learning models is quantified using performance assessment metrics such as root mean square error (RMSE), and mean absolute error (MAE). In addition to this, extensive comparison study of the different RUL prediction models is carried out to demonstrate pattern learning ability of various machine lending models.