Leveraging the Transformer Architecture for Anomaly Detection on Multivariate Time Series Data. Master thesis Radboud University Nijmegen
Year: 2020
Type: dissertation
Abstract: Due to widespread digitalization, predictive maintenance and Digital Twinning are increasingly viable and popular options for industry aiming for smarter operation and maintenance of industrial equipment.Machine Learning will play a crucial role in this transition, augmenting existing expert knowledge.There is a growing need for robust and transparent methods for analyzing and acting upon the wealth of sensor data available in this field.In this thesis we study the feasibility of using a Transformer model for anomaly detection in an industrial setting.Leveraging the Transformer architecture for anomaly detection allows for studying the effects of neural attention distributions on predictions, a step towards more interpretable deep anomaly detection models.We compare the signal reconstruction and anomaly detection performance of an LSTM encoder-decoder model with that of a Transformer.We also apply our models to a real world use case, detecting anomalies in an industrial gas turbine.The lack of easy access to anomalous training data in industrial settings remains a challenge.In the future our approach can be extended to perform anomaly prediction, functioning as an early warning system or scheduling predictive maintenance as part of a Digital Twin.
Source: TNO Repository
Author: Grauw, M.J.J. de
Institution: Radboud University Nijmegen
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7132167902
