< Back to previous page

Project

Context-aware adaptive and continual learning for anomaly detection in dynamic industrial settings

An increasing number of industrial machines and processes is being monitored by a variety of sensors. This enables the identification of rare observations, which are often connected to an underlying problem with the equipment. Anomaly detection is the subfield of machine learning that is concerned with identifying these abnormal events from the data in an automated way. The application of anomaly detection algorithms in dynamic industrial settings is however challenging, due to the static nature of the learned model after deployment in combination with the stationarity assumption of the data which does not hold for heavily dynamic machine and process conditions. This distributional change is influenced by the context in which the data-generating process is operating. To address these challenges, the overall objective of this project is twofold. First of all, research will be performed on how how strategies for adaptive and continual learning can be employed to arrive at more accurate, robust and timely anomaly detection on time series data originating from dynamic industrial settings. Second, it will be investigated how these strategies can be adapted to explicitly account for the contexts in which the respective systems operate.
Date:1 Oct 2021 →  30 Sep 2023
Keywords:contextual anomaly detection, continual learning, time series analysis
Disciplines:Machine learning and decision making