Fault Detection in a Multi Sensors Context by 3D Object Descriptors Method
Résumé
The monitoring of an asset in an industrial context is a real challenge today, as data are more and more available, and computation power becomes cheaper with time. However, if we want to use data from different sensors to detect if there are anomalies of any kind, it is usually needed to individually consider a whole time series, or the values of several time series at a particular moment. In this article, we propose an adaptation of 3D objects descriptors to the detection of unknown faults in a multi-sensors context for features extraction. Then, classical outliers detection methods such as Local Outlier Factor1 and isolation forests 2 are used. This allows us to detect an unknown problem to come on an asset monitored by several sensors. To our knowledge, this problem has not been completely solved yet, and opens new opportunities in class disequilibrium contexts. Final performances confirm the interest of the proposed approach adapted to a real time industrial context, and allow to consider a new way for extracting features in the pretreatment of multi-time series.