Comparative Analysis of Anomaly Detection Models for Irregular and Incomplete Time Series in Cyber-Physical Systems

Anomaly detection in cyber-physical systems (CPS), such as smart electric grids and sensor driven industrial platforms, presents unique challenges due to irregular sampling, missing data, and severe class imbalance. In this thesis, we conduct a comparative evaluation of four distinct modeling approaches for anomaly detection on two real-world datasets: the SGCC smart meter dataset and the NAB machine sensor dataset. The models include two classical baselines — XGBoostand Isolation Forest, and two neural architectures designed for temporal modeling under irregularconditions: Time-aware LSTM (TLSTM) and Neural Controlled Differential Equations (NCDE).Our results show that TLSTM consistently outperforms classical models across both datasets, particularly in handling missingness and variable time intervals. Surprisingly, the theoretically grounded NCDE architecture underperforms in practice, exhibiting training instability and scal ability issues on larger data regimes. XGBoost performs competitively even in the presence of missing data, but lacks the capability to handle temporal data. These findings suggest that time aware discrete models like TLSTM offer a strong trade-off between complexity and performance in real-world CPS anomaly detection tasks. We also outline limitations of current approaches and propose future directions including precursor-of-anomaly and novelty detection, early warning scoring, and integration of control-theoretic priors.