Applied System Innovation

This paper proposes the Raid Team Strategy (RTS), a population-based metaheuristic inspired by the tactical organization of raid teams in search-and-attack operations. At each iteration, the best-ranked agents act as leaders and exploit their neighborhood through self-adaptive local search. In contrast, the remaining agents act as followers that are pulled toward the nearest leader along its direction of travel. The number of leaders is a fixed fraction of the population, while the number of teams that search separate regions changes during the run. Leaders that converge onto the same region are merged, and the weaker one is re-deployed at a random location; isolated but competitive followers are temporarily promoted to scouts, and stagnant followers are released to unexplored regions. RTS was evaluated under an equal budget of objective-function evaluations against ten methods, including the state-of-the-art L-SHADE, jSO and IPOP-CMA-ES and the deterministic DIRECT and DIRECT-L, on the CEC2017 (D = 10, 30, 50) and CEC2022 (D = 10, 20) suites with 51 independent runs, on the eight standard GKLS classes (800 randomly generated multi-extremal functions), and on six engineering design problems. RTS was significantly better than PSO, GWO, BA, and HHO on the CEC suites and statistically tied with SBOA, but it was clearly outperformed by L-SHADE, jSO, and IPOP-CMA-ES on the CEC suites and on the engineering problems. On the GKLS classes, RTS solved 638 of the 800 functions, compared with 260, 261, and 108 for L-SHADE, jSO, and IPOP-CMA-ES, and was exceeded only by the two DIRECT methods and, slightly, by HHO. An ablation study showed that the follower refinement and the focused probes are the components with a significant effect, and a sensitivity analysis showed that the results are robust to the parameter settings. The results position RTS as a method for landscapes with several separated basins rather than a fine local exploiter.

Monitoring the current technical condition of a marine internal combustion engine and diagnosing the causes of deviations of its operating parameters from reference values are important tasks for ensuring the efficient and safe operation of marine power plants and vessels in general. This study develops a digital twin framework for marine engine diagnostics and demonstrates its integration with the DEPAS 5.0W diagnostic system. The framework combines a calibrated 0D–1D engine-process model, sensor-based identification of the current operating mode, reference-state generation, and inverse identification of physical factors responsible for deviations from expected engine behavior. The mathematical model is adapted for digital twin operation through operating-mode-dependent parameterization, including a modified fuel-injection approximation. A four-factor constrained factorial numerical experiment with second-order response-surface models is used to investigate the inverse diagnostic problem. The approach is demonstrated first using virtual marine engines and subsequently using experimentally acquired data from a Weichai WP4 diesel generator with a controlled intake restriction introduced by a throttle valve. The results demonstrate that the proposed framework supports the identification of the most probable physical cause of the observed deviation when multiple potential fault factors are considered. At the same time, the results reveal the practical non-uniqueness of the inverse problem, as different factor combinations may produce similar observable engine responses. The Mahalanobis distance is shown to provide improved fault discrimination in the investigated case compared with individual pressure-based objective functions. The proposed architecture provides a basis for further development of automated digital twin-assisted diagnostics, including systematic uncertainty and identifiability analysis and validation under controlled faults.

Artificial intelligence (AI) is changing the scope of decision support from model-assisted analysis toward systems in which algorithms, data, domain knowledge, and human judgment interact throughout the decision process [...]