An Industrial Case Study Using Vibration Data and Machine Learning to Predict Asset Health

Abstract

Over the years, there has be considerable progress in using condition monitoring of industrial assets to detect and predict failures. However, there are not many papers using real field data to validate such approaches. Our goal is to provide a proof-of-concept, which shows that the condition of industrial assets can be predicted using machine learning applied to field data from an industrial plant. In this paper, an extensive case study based on vibration monitoring is presented. Data collected from 30 industrial pumps in a chemical plant over a 2.5-year period is used to validate the concept. To do so, metrics derived from vibration data are predicted up to 7 days ahead using the well-established and quick-to-use Random Forest algorithm. The model's performance is benchmarked against a standard persistence technique. We detail the pre-processing steps taken to prepare the data for machine learning. In doing so, insights gained from the challenges that arise when applying machine learning to real-world industrial data are also mentioned. For some failures, we also physically verified their root-causes, which showed that such failures could have been prevented with reliable predictions. Thus, our findings are particularly useful for those interested in the applicability of machine learning in an industrial context.