Transferring knowledge about gear systems to machines in order to improve diagnosis efficiency

摘要

Predictive maintenance through condition monitoring presents a huge interest for the important financial costs of component failure. The rise of Machine Learning technology is a new tool to improve efficiency of predictions and simulations. If Big Data approaches require mostly huge volumes of data to ensure efficiency, we think that building models on moderate amount of data enriched by human expertise and knowledge form a more suitable way to deal with most of the industrial situations (where available data is often scarce, especially labelled data). In this paper, we discuss the possibilities of diagnosis of geared system from a naive approach compared to a human-knowledge-enhanced one through dedicated indicators. Main issues and achievements through the different tools used for that purpose are discussed and results are based on a realistic measurement campaign found on the web thanks to the Prognostics and Health Management (PHM) society challenge of 2009. First, the procedure of features extraction is detailed and explained through its whole process oriented by the goal of the analysis: determine whether a given gear system is healthy or not. Then, the next level of default characterization is reached by using different condition indicators regarding the type of defect. Finally, classification of the dataset is studied.