Data-Centric Analysis of Complex Industrial Systems
Abstract
Modern industrial Cyber-Physical Systems (CPSs) combine distributed computing,
software and sensors, producing rich streams of operational data. This talk explores
how these data can support anomaly detection and identification, using semiconductor
photolithography machines as a case study. Repetitive execution phases provide a
basis for describing normal behaviour and identifying deviations, while selective
data collection helps balance useful observation against processing overhead.
We present behavioural fingerprinting approaches that combine electrical metrics
with machine learning. Phase-based regression models and power passports support
traditional classifiers, while convolutional neural networks offer an alternative
with less feature engineering. The talk compares these approaches in terms of
accuracy, preprocessing effort and explainability, and discusses how the available
data and domain knowledge shape the choice of solution. It concludes with challenges
in knowledge incorporation and generalising data-centric methods across systems.
Metadata
Links
Licence
Artefacts shared as PDF are licenced under
CC BY 4.0.