Speaker
Description
The controls group at ISIS has been exploring anomaly detection and its associated challenges. This overview highlights the challenges faced, unsuccessful attempts, and lessons learned. Initially, the group implemented a machine learning anomaly detection system on the methane moderator for Target Station 1. The anomaly detection work began before the system upgrade, rendering previous challenges and solutions obsolete. Initially, the focus was on a single temperature channel, with the main challenge being data labelling of normal and anomalous. Solutions included dimensionality reduction, initial classification in lower dimensions, and convolutional neural networks. Post-upgrade, the machine’s behavior changed, introducing new challenges such as analyzing differences between moderator versions, feature engineering for time series, data relabeling, and developing new evaluation methods. The project’s scope expanded to detecting anomalies across multiple channels and understanding their interactions, aiding users in identifying root causes rather than just symptoms.
