Anomaly Detection for Diagnosing Failures in a Centrifugal Compressor Train

Abstract

Predicting machine failures is of the utmost importance in industrial systems as it can turn expensive crashes and repair costs into affordable maintenance costs. To this end, this paper presents preliminary work for detecting failures in a centrifugal compressor train based on sensorial data. We show the detection capabilities of a two-step process consisting of: (1) a preprocessing step to reduce the dimensionality of the input data using Principal Component Analysis, and (2) an anomaly detection step using the Mahalanobis distance to detect anomalous observations on the sensors' data. The experiments using real-world data demonstrate the feasibility of our approach and the ability of the method to detect the failures eight days in advance.