Experimental assessment of artificial intelligence performance when choosing the type of repair work on bridges

Aim: To assess the quality of answers generated by modern neural network models when determining the type of repair work on bridges and to verify their applicability for engineering decision support. Methods: Based on the Russian regulatory framework, an original decision matrix was developed linking categories of technical condition, results of load-carrying capacity verification, and recommended types of work (routine repair, major repair, reconstruction). A unified engineering query on the criteria for choosing the type of work was posed to six popular AI models; their answers were qualitatively analyzed using three engineering markers: the role of load-carrying capacity verification, consideration of foundations and substructures, and the application of a risk-based approach. The quality of AI-generated answers is treated as a combination of three components: completeness (coverage of all necessary input data), relevance (compliance with current standards and the bridge management context), and engineering reliability (correctness of the recommended type of work from a risk perspective). Additionally, a cost–quality aspect was examined, taking into account the approximate cost of generating detailed answers. Results: Neural network models are effective in structuring input data, selecting appropriate terminology, and retrieving regulatory references; however, they do not provide a reliable independent choice of the type of work in cases of incomplete data on foundations and the absence of verification calculations. Conclusion: The results indicate that neural networks should be used as an auxiliary tool in engineering practice, with mandatory expert involvement and with comprehensive calculation data available; the comparison by the criteria "load-carrying capacity – foundations – risk – cost/quality" shows that low-cost models are suitable for quick text drafts, whereas tasks with strict regulatory requirements justify the use of more expensive models despite the increase in the approximate cost of a single detailed answer. The conclusions are compared with current approaches to digital twins and AI-supported management and decision-making.

Hali tarjima qilinmagan