The extensive deployment of sensors in Cyber-Physical Systems (CPSs) enable continuous monitoring of physical processes, generating Correlated Time Series (CTS) data with both temporal and spatial correlations. Temporal correlations link consecutive measurements within a time series, such as heart rate fluctuations over time in healthcare monitoring. Spatial correlations connect multiple sensors, like wind speed and direction measurements across turbines in a wind farm.

However, existing deep learning-based methods struggle to balance accuracy and computational efficiency in CTS analytics, limiting their use in resource-constrained environments. This thesis addresses this challenge through five key contributions:

1. LightCTS – an efficient CTS forecasting framework leveraging lightweight temporal and spatial modules, achieving state-of-the-art accuracy with reduced computational overhead.
2. LightCTS⋆ – an extension incorporating model distillation to enhance efficiency without sacrificing accuracy.
3. E2USD – an efficient CTS state detection framework using a compact embedding encoder for contrastive learning and adaptive thresholds for online streaming detection, validated on MCUs.
4. ReCTSi – an efficient CTS imputation framework employing decoupled pattern learning for accurate imputation with minimal computational cost.
5. AdaCTSi – an adaptive CTS imputation framework dynamically tailoring its operations to on-demand scenarios, achieving state-of-the-art performance in both imputation accuracy and resource adaptability.

Together, these contributions advance CTS analytics by introducing efficient, effective, and adaptive deep learning methods, addressing critical efficiency challenges in CTS processing.

All interested parties are welcome. After the defence the department will be hosting a small reception in cluster 3.