A knowledge-guided hybrid learning framework for semantic constraint integration in time series models

Current time series models often operate solely on sensor data, lacking the contextual understanding that domain knowledge provides. This limitation particularly exists in domains like maritime operations or medical monitoring, where sensor data are often noisy, incomplete, or ambiguous. To address this gap, this doctoral research proposes a hybrid learning framework that integrates semantic knowledge from ontologies, domain texts, and expert-defined rules into the modeling process as formal constraints. The framework comprises three main building blocks: (1) learning joint representations from heterogeneous sources such as time series, structured knowledge, and unstructured text; (2) extracting and formalizing semantic knowledge into symbolic or functional constraints; and (3) fusing these components into a hybrid framework, where formal constraints complement machine-learned patterns. Initial work has been conducted in the maritime domain and will be extended to medical datasets for cross-domain evaluation.