Cyber-physical systems (CPS) that support critical infrastructures increasingly rely on structured digital communication, exposing them to complex and potentially high-impact cyber threats. To strengthen their resilience, it is essential to optimize testing efforts by focusing on transactions that are most likely to reveal vulnerabilities. This paper presents a methodology for leveraging deep packet inspection and machine learning to prioritize test cases based on behavioral patterns observed in network traffic. The approach reconstructs request-response pairs from raw packet data and extracts structured features for supervised modeling tasks, such as predicting response time. Experimental results using an open dataset from electric vehicle charging systems demonstrate the feasibility of this method, highlighting key features that contribute to latency and interaction complexity. The findings support the use of predictive models for intelligent test suite optimization and open the path for more adaptive, protocol-agnostic security testing strategies in CPS environments.