Abstract: Industrial cybersecurity analysis often relies on textual incident descriptions that vary widely in structure, terminology, and level of technical detail. In practical manufacturing environments, the manual interpretation of such problem narratives by domain experts becomes increasingly difficult to sustain as system complexity and data volume grow in Cyber-Physical Systems (CPS). To address this issue, this study examines the use of a deep-learning-based framework, referred to as Deep-Learning-Based Use-Case Recommendation (DLUR), for mapping unstructured industrial problem descriptions to relevant analytical use-cases. The framework combines Doc2Vec-based document representations with Bidirectional Long Short-Term Memory (BiLSTM) models to capture contextual semantic information beyond keyword-level similarity. The resulting representations are further associated with a CPS-oriented hierarchical knowledge structure, enabling recommendations to be interpreted in relation to physical assets and security contexts. Experimental evaluation was conducted using a dataset of 3,548 industrial technical documents. The results indicate that DLUR provides more stable recommendation performance than conventional Support Vector Machine and Random Forest baselines under the examined conditions. These observations suggest that the proposed approach can support analysts during the early stages of cybersecurity assessment, particularly when prior domain experience is limited.

Keywords: Recommendation System, Deep Learning, Long Short-Term Memory, Cyber-Physical System, Use-Case, Natural Language Processing, Cybersecurity, Digital Twins

Pages: 317-325