Abstract
Machine learning prediction of structural responses is highly efficient. The use of surrogate models to assess the reliability of small probability events is significant for engineering safety evaluations. This paper proposes a structural reliability analysis method that combines machine learning surrogate models with physical information and Subset Simulation. By using CNN to extract features from load and structural frequencies and then employing LSTM to predict responses based on the feature vector, the surrogate model is integrated with the SS method. This approach aims to address the issues of computational efficiency and accuracy in structural reliability analysis, especially for small failure probabilities. Case studies on planar truss systems and steel frame structures under random dynamic loads demonstrate the effectiveness of the method. The results show that the CNN-LSTM model achieves better prediction accuracy than LSTM. In terms of reliability results, the failure probability predicted by CNN-LSTM is closer to the finite element-Subset Simulation results, while the failure probability error of LSTM-Subset Simulation is 12.50 percentage points higher than that of CNN-LSTM-Subset Simulation. This method is significant for evaluating structural reliability.

This work is licensed under a Creative Commons Attribution 4.0 International License.
Copyright (c) 2026 xinqi zhang, jun hu