Multi-view contextual performance profiling in rotating machinery

Abstract

Nowadays, most industrial assets are equipped with a multitude of different sensors continuously examining the asset's status and health. For a reliable estimation of an asset's performance it is crucial though to consider that most assets are exposed to different and typically varying contexts during their operations. These contexts are defined by both internal and external factors and complicate the task of asset condition monitoring and profiling. In this article, an unsupervised approach for asset performance profiling is proposed based on multi-view representation and matrix decomposition. It enables one to derive specific fingerprints characterising asset performance behaviour in a context-sensitive fashion. The data is processed in two separate data views: 1) the process view, in which variables related to the asset's internal working are processed and partitioned such that each measurement point is associated with a specific label representing the context; and 2) the vibration view, where vibration profiles are extracted via non-negative matrix decomposition. Subsequently, the two views are linked together allowing to derive characteristic fingerprints using a suitable contextual representation and performance-related indicators. The proposed methodology is validated on a real-world industrial data set, consisting of vibration and operational sensor measurements of feedwater pumps. The obtained results illustrate that the profiling methodology is able to deliver a meaningful risk assessment estimation associated to different operating contexts.