4 4.4. Experimental evaluation 79 (see Jimenez-Roa, Heskes, Tinga, et al., 2023 for details). This work improves upon it by introducing parallelisation, enabling the use of multiple system cores for FT generation. - Output: An expanded FT population featuring new FTs. (3) - Input: The expanded FT population. - Process: Each FT in the population is processed to calculate the 17 metrics listed in Table 4.1. Caching is used to avoid recalculating metrics for previously evaluated FTs, thus enhancing e"ciency. Parallelisation is implemented to further improve this process. - Output: The enlarged FT population with corresponding metric values. (4) - Input: The enlarged FT population with corresponding metric values. - Process: The NSGA-II algorithm (Deb, Pratap, Agarwal, et al., 2002) and Crowding-Distance (Martí, Segredo, (nchez-Pi, et al., 2017) are utilised for multi-objective optimisation to construct Pareto fronts of non-dominated FTs based on the metrics. - Output: The topNFTs—where Nis a user-defined parameter—are selected for the next generation. (5) - Input: The top N FTs. - Process: Evaluate convergence criteria: (i) the maximum number of generations is reached, or (ii) the best FT candidate remains unchanged for a specified number of generations. If neither condition is fulfilled, the top N FTs are used as input for Step 2, and Steps 2 to 5 are repeated until one of the convergence criteria is met. - Output: The inferred FT, FD, identified as the best FT candidate in the first Pareto front. 4.4 Experimental evaluation Case studies. We evaluate our approach on six FTs stemming from various application areas. The Data-driven Fault Tree (ddFT) (Lazarova-Molnar, Niloofar, and Barta, 2020) was obtained from time series data. The Mono-propellant Propulsion System (MPPS) (NASA, 2002) is used for a small space flight vehicle. The COVID-19 FT (Bakeli, Hafidi, et al., 2020) is used in infection risk management. Table 4.2: Overview of case studies. Case |BEs| |F| |D| |CD| All FTs ddFT 8 19 256 6 83,600 MPPS 8 14 256 7 73,200 COVID-19 9 13 512 6 60,400 TS1 10 21 1,024 16 127,200 GPT12BE 12 25 4,096 13 139,200 GPT15BE 15 27 32,768 10 108,000 The Truss System (TS1) (JimenezRoa, Volk, and Stoelinga, 2022) models a symmetric truss bridge system. The two FTs GPT12BE and GPT15BE were generated with GPT-4 (OpenAI, Achiam, Adler, et al., 2024), representing larger FTs designed to test scalability. The prompts used for generation included examples of existing FTs, the number of nodes, and the number of
RkJQdWJsaXNoZXIy MjY0ODMw