Impact damage prediction method of glass/carbon fiber hybrid composites based on particle swarm optimization–backpropagation neural network

Abstract Damage to glass/carbon fiber reinforced hybrid composites from ballast impacts was a critical risk during high‐speed rail operation. In this study, strain sensors were employed to predict the maximum strain values induced by impact. Additionally, a finite element model (FEM) was established to obtain impact signal datasets. Train speed, maximum ballast diameter, and impact angle were selected as signal features. A particle swarm optimization–backpropagation neural network (PSO‐BPNN) damage prediction model was developed and optimized using the response surface methodology. The optimization improved prediction accuracy by 43% compared to the unoptimized model, reducing the mean squared error to below 0.0015%. Different sensor monitoring layout schemes were proposed and the results showed that the diagonal sensor layout achieved the highest monitoring accuracy, with a mean squared error of 0.0043%. The PSO‐BPNN model achieved a prediction success rate of 96%. Highlights Integrating FEM with machine learning enables real‐time composite damage prediction. Hyperparameter optimization enhanced the PSO‐BPNN model's accuracy. Model prediction accuracy exceeds 96%. The effect of different sensor layouts on prediction accuracy was tested.