Research on stress monitoring method for high-speed train carbody based on data-model hybrid-driven

Aiming at the challenges of short sensor service life and the difficulty in long-term full-field stress prediction, this article proposes a data-model hybrid-driven stress monitoring method leveraging easily accessible acceleration signals to achieve real-time, high-precision and full-field carbody stress prediction within engineering constraints. Line tests first revealed the coupled characteristics of quasi-static and vibratory fatigue loads on the carbody, confirming a strong correlation between acceleration and loads in time and frequency domains. By integrating dynamic theory and Pearson correlation analysis, the sensor layout was optimized from dense deployment to precision sensing, requiring accelerometers only at the bogie frame ends and bolster, which significantly reduced cost and data dimensionality. A temporal convolutional network model was then constructed for load identification, achieving an accuracy of R ²  > 0.7 for most loads and demonstrating strong robustness due to its dilated convolution and residual structure. Finally, the identified loads were input into a modal-testes calibrated finite element model to predict full-field stress using a hybrid strategy: quasi-static superposition handled low-frequency responses, while a fast Fourier transform-based mode superposition method calculated high-frequency responses. Compared with measured data, the prediction error was less than 4% with measured loads, and the maximum error was reduced to 3.08% with identified loads when a correction coefficient was applied. The research validates the engineering advantages of this hybrid paradigm, which allows independent verification of the load identification and stress prediction modules, reducing system debugging complexity while enhancing interpretability and compatibility with existing engineering systems. The method constructs a diagnosable and updatable mechanical interface prototype, laying the foundation for a closed-loop “monitoring-diagnosis-model updating” health management system.