Validating Prognostic Algorithms: A Case Study Using Comprehensive Bearing Fault Data

The ultimate goal of prognostics is to accurately predict remaining useful life (RUL) based on sensor data, system usage, and prior knowledge of fault-to-failure progression rates (i.e. a model). One of the key components necessary for developing a prognostic algorithm is a diagnostic severity metric. This paper presents an evaluation of a number of standard vibration-based diagnostic metrics, utilizing a large set of experimental fault-to-failure progression data for bearings. These experiments included over 40 complete bearing failure progressions with 10 to 30 ground truth data points per bearing. Additional data supporting the potential of using oil debris monitoring in conjunction with vibration monitoring is also presented. Once a prognostic algorithm has been developed, the next critical step is to validate how well the algorithm performs. Conceptually, this seems like a simple task. However, there are many criteria to be considered, including convergence rate, accuracy, and stability of the RUL prediction. The paper includes an evaluation of prognostic algorithms based on vibration-based diagnostics that feed into a model-based prediction of future spall propagation and thus remaining life. Methods for objectively measuring the quality of the predictions are proposed. The results presented herein help demonstrate the capabilities and limitations of predictive prognostics at the current state-of-the-art.