Residual stress assessment in shot-peened titanium alloys using LSTM-attention neural networks

The residual stress induced by shot peening is a key factor affecting the efficacy of shot peening, with its magnitude and distribution directly influencing the fatigue performance of materials. Therefore, accurately predicting residual stress is crucial for optimising the shot peening process. Taking titanium alloy as the research object, based on the finite element-discrete element (FEM-DEM) coupled model, the data set of shot peening residual stress was generated by using orthogonal experimental design, and the long and short-term memory (LSTM) network prediction model incorporating the attention mechanism was proposed to address the issue of long-range dependent information loss with the increase of the sequence length. The model takes process parameters (shot diameter, shot peening speed, shot peening coverage) and residual stress layer depth as inputs and the shot peening residual stress values as outputs. To evaluate the model performance, orthogonal tests screen the optimal hyper-parameter combinations, and the results are compared with those of feed-forward neural networks, support vector regression models, and ordinary long and short-term memory network models. The results show that the proposed model performs optimally in shot peening residual stress prediction, with the coefficient of determination R 2 exceeding 0.98, and the other performance indexes are improved by nearly 60% compared with the comparison algorithms. The neural-network prediction results show that the proposed model is efficient and accurate, providing a new approach for intelligent prediction of shot peening surface integrity.