Semi-supervised edge energy consumption anomaly detection and classification

The researcher will compare energy consumption models based on model and data-driven approaches and noisy run-time state observations. As prediction is in general more difficult than anomaly detection, (1) a first objective will  be to determine normal and abnormal energy consumption profiles. Energy consumption anomalies could be hardware failures, as well as abnormal communication environment conditions (such as jammers of signal blockers). (2) A second objective will then be to label or classify the anomalies, first  unsupervised and later with expert feedback.  Expert feedback will be possible by training the model to work with interpretable features. (3) In a third objective, the normal energy consumption profiles will be used for energy consumption prediction. These models will work on the interpretable features, so ideally derived prediction models will be explainable to users and algorithm  designers.