Year: 2026
Type: dissertation
Abstract: This thesis investigated how machine learning-based anomaly detection can be applied to support predictive maintenance for a radar system using available monitoring data. An LSTM autoencoder was trained on selected healthy running operation and used to detect deviations in multivariate sensor windows based on reconstruction error. The anomaly score was combined with report information and a rule-based operating-state classification to separate potentially maintenancerelevant anomalies from deviations caused by known operating conditions. The results show that the method can detect both naturally occurring anomalous behaviour and simulated fault-like deviations, but also that a more complete predictive maintenance implementation would require richer sensor coverage, better operational metadata, and more direct information about system states.
Author: Thorgren, Klas
Institution: Lund University
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213856700