Failure Prediction of Rolling Element Bearings Operated Under Non-Stationary Conditions

Abstract

Bearings constitute a majority of the components found in rotating machines. Though inexpensive, their failure can result in unnecessary downtime, losses in production, and propagation of failure to other critical components leading to expensive maintenance actions. Most of these rotating machinery are operated under adverse and varying conditions which result in difficulty in defining health indices from condition monitoring data. Predicting the failure of such machines is crucial to determine when the maintenance is required thereby leading to a reduction in maintenance costs and an improvement in the safety and reliability of the machines. Therefore, techniques for condition monitoring of rotating machinery operated under non-stationary conditions are necessary. This work employed a model-based condition monitoring approach to predict the failure of rotating machinery under non-stationary conditions. One of the advantages of model-based approach is the ability to incorporate physical understanding of the system monitoring. Firstly, the vibration model for rolling element bearing with fault was constructed in MATLAB/Simulink environment. An automatic parameter identification based on Particle Swarm Optimisation (PSO) algorithm was employed to identify the dynamic parameters of a rolling element bearing due to its ease of implementation and rapid convergence property. The optimized bearing parameters were then used in diagnosing bearing faults. To evaluate the feasibility of this approach, two publicly available data sets were employed. The results showed an improved average accuracy of 99.67% and 99.2% for bearing faults of Case Western Reserve University and University of Paderborn datasets, respectively. Additionally, the bearing model with estimated parameters was used to generate degradation data by varying the fault depth. Feature extraction was carried out where Root Mean Square (RMS) was determined as the appropriate health indicator. Lastly, Paris degradation model was employed to determine bearing damage and its evolution with time while factoring in speed, applied load, and bearing geometry. The remaining useful life of the bearing was found to be 1598 cycles. The prediction results and evaluation indexes demonstrated the effectiveness and superiority of the proposed method.