Oscillation Detection on High-Resolution Time Series Data

Abstract

Field test vehicles at Volvo Group log between 500 and 1000 signals at resolutions up to 100 Hz, some of which originate from actuators. Oscillations in actuator signals can cause system instabilities and lead to components being worn out prematurely. This thesis investigates the suitability of unsupervised anomaly detection techniques for identifying such oscillations by framing them as a specific type of anomaly. Par ticularly, the focus is on evaluating the performance of a One-Class Support Vector Machine (OC-SVM) and a transformer-based model (TranAD). The available data is unlabeled and consists of high-frequency time series data collected from two main sources: field test vehicles and test cells. To complement this, a number of man ual data recordings were provided by domain experts at Volvo Group, containing examples of oscillations. These recordings, combined with synthetically generated oscillations, were used to create a labeled test set. OC-SVM and TranAD were trained on both field test data and test cell data, with the best OC-SVM model be ing trained on field test data and the most effective TranAD model being trained on test cell data. Although both models are able to detect oscillations, they also cap ture other types of anomalies and sometimes misclassify normal data as anomalous. Overall, TranAD demonstrates the most promising result in detecting oscillatory behaviour. Since both OC-SVM and TranAD were able to detect oscillations, but also other types of anomalies, a valuable extension to this work would therefore be some sort of clustering as a postprocessing step. Despite some limitations, the mod els successfully identified oscillatory patterns that had not previously been detected at Volvo Group.