Semantische Suche

However, most approaches assume that the data's dimensionality, i.e., the number of attributes, stays constant over time. This assumption is unjustified in many real-world use cases, such as sensor networks or computer cluster monitoring. Feature-evolving data streams do not impose this restriction and thereby pose additional challenges.

In this thesis, we extend the well-known Local Outlier Factor (LOF) algorithm for outlier detection from the static case to the feature-evolving setting. Our algorithm combines subspace projection techniques with an appropriate index structure using only bounded computational resources. By discarding old observations our approach also deals with concept drift. We evaluate our approach against the respective state-of-the-art methods in the static case, the streaming case, and the feature-evolving case.

Although state-of-the-art approaches deliver good results, they are hard to train due to the high number of variables, lacking the ability to generalize, and need to make many simplifying assumptions. In contrast to theory-based models, purely data-driven approaches require large datasets and are often unable to produce physically consistent results. Theory-Guided Data Science (TGDS) aims at using scientific knowledge to improve the effectiveness of Data Science models in scientific discovery. This concept has been very successful in several domains including climate science and material research.

Our work is the first one to apply TGDS to battery systems by working together closely with domain experts. We compare the performance of different TGDS approaches against each other as well as against the two baselines using only theory-based EC-Models and black-box Machine Learning models.