Scenario-Based Software Integration Testing of Railway Control Systems with SCSL Templates and Generative AI

Testing safety-critical railway control systems requires the systematic elaboration of large numbers of test scenarios covering both normal and exceptional behaviours. While scenario-based testing reduces the modelling effort compared to comprehensive model-based testing, the manual construction of complete scenario libraries remains time-consuming and error-prone. This article investigates how GPT-based large language models can support the automated production of formally specified software integration test scenarios from manually prepared templates. The presentation focus of this article is on experimental evaluation of the approach for software integration testing in the railway domain. To this end test scenarios are generated for software implementing a train detection system (TDS) conforming to the EULYNX standard. Templates and concrete scenarios are expressed in the formal the SCenario Specification Language (SCSL), which enables subsequent automated test data generation and test execution. The GPT is not used for open-ended formal specification synthesis. Instead, it is constrained to instantiate reviewed SCSL templates according to natural-language instantiation rules, generation targets derived from the domain standard, and a task description specifying the scenarios to be produced. The experiments show that example-based single-shot or multi-shot prompting is not required for this form of template-based scenario generation. Moreover, the availability of detailed EULYNX scenario descriptions makes it possible to use more abstract template parameters than in previous work: parameters may denote classes of values whose concrete instantiation depends on their occurrence context in the template. This increases template reuse and reduces the manual effort required to construct families of related test scenarios. Together with previous results on SCSL-based system testing and GPT-supported scenario generation for railway interlocking systems, the results presented here provide further evidence that SCSL is applicable across different domains and test levels, and that GPT-based template instantiation can produce rule-conformant scenarios with high reliability, provided that the instantiation rules have been carefully reviewed for completeness and unambiguity.