Research on Teaching of Structural Safety Judgment for Railway Vehicles Driven by Controlled Errors of Generative Artificial Intelligence

Abstract

Aiming at the common problems in railway vehicle structure courses: students tend to overemphasize calculation results while ignoring applicable conditions, prioritize software operation over evidence verification, and easily develop judgment dependence after the introduction of generative artificial intelligence, this paper proposes a teaching framework for structural safety judgment driven by controlled errors. Taking welded joints of bogie frames, box structural responses under wheel polygon excitation, and local stress interpretation as research carriers, a case generation mechanism integrating retrieval-augmented generation, targeted error mutation, independent program verification and instructor review is designed. Teaching activities include independent preliminary judgment, evidence checking, revision and argumentation, and AI-free transfer evaluation. Teaching examples demonstrate that incorrectly substituting stress amplitude into stress-range-based fatigue curves can lead to an 87.5% underestimation of damage in this case. Under the specified single-mode near-resonance condition, the estimation of responses using a constant dynamic load factor is significantly low. The above results are used to characterize error consequences and do not represent classroom teaching effectiveness. Furthermore, an evaluation rubric covering error detection, correct adoption of information, evidence sufficiency and confidence calibration, together with three groups of control schemes, is put forward. This study develops reproducible professional cases and implementable evaluation routes, providing instructional design references for cultivating engineering judgment capabilities of railway vehicle engineers in the generative AI era. The learning benefits and applicable boundaries still need to be validated by real classroom data.