Mapping CPS domains to suitable learning methods

Determine which cyber‑physical system (CPS) application domains align best with specific learning algorithms and ascertain whether unsupervised or semi‑supervised learning methods are appropriate for operation in unpredictable CPS environments, in order to guide principled method selection across CPS use cases.

Background

Within the systematic literature review, the authors observe substantial fragmentation in how learning-based methods are applied across CPS domains. While deep learning dominates many topics, the suitability of other paradigms—especially unsupervised and semi‑supervised learning—for dynamic, uncertain CPS is not established. This gap hampers principled method selection when designing CPS that must adapt to unpredictable environments.

The open question seeks to classify CPS applications by characteristics relevant to method performance and robustness, and to evaluate whether unsupervised or semi‑supervised approaches can reliably support monitoring and control in settings where labeled data are scarce or conditions change frequently.