Bearing Prognostics Using the PRONOSTIA Data: A Comparative Study

Vibration-based condition monitoring is a well-developed and established field in maintenance engineering. Failure prognosis is a crucial aspect of condition monitoring, particularly Remaining Useful Life (RUL) prediction. Over the last decade, the PRONOSTIA bearing dataset has become the standard reference for testing these prognostic algorithms. However, the lack of standardized comparisons makes it difficult to objectively assess the relative performance of different methods. This paper systematically compares three established data-driven artificial intelligence approaches: classical machine learning, deep learning, and transfer learning approaches. We analyze their application to the PRONOSTIA dataset, providing detailed discussions of their relative strengths, limitations, and achieved performance. While this study primarily serves as a benchmark for testing new prognostic algorithms in vibration monitoring, we hope the insights will also broadly apply to other condition-monitoring techniques.