TR2019-057

Fault Detection and Classification of Time Series Using Localized Matrix Profiles


Abstract:

We introduce a new primitive, called the Localized Matrix Profile (LMP), for time series data mining. We devise fast algorithms for LMP computation, and propose a fault detector and a fault classifier based on the LMP. A case study using synthetic sensor data generated from a physical model of an electrical motor is provided to demonstrate the effectiveness and efficiency of our approach.