Multi objective optimization of stamping process parameters based on improved ASAE-GP-BPNN hybrid model and NSGA-II algorithm

Stamping-forming parameter optimization is a core technical hurdle in enhancing product quality. To tackle the low accuracy of existing surrogate models and the uneven Pareto fronts produced by current multi-objective algorithms, this paper introduces a novel framework that couples an Adaptive Sparse Autoencoder–Gaussian Process–Back-Propagation Neural Network (ASAE-GP-BPNN) hybrid surrogate model with an improved Non-dominated Sorting Genetic Algorithm-II (INSGA-II). ASAE-GP-BPNN achieves accurate modeling of complex nonlinear relationships through competitive sparse mechanism for dynamic feature extraction, Gaussian process uncertainty quantification, and BPNN deep fusion, showing 15–40% improvement over single models on test functions. INSGA-II introduces K-Nearest Neighbor(KNN) local density evaluation and adaptive parameter strategy, increasing the hypervolume of Pareto front by about 15%. Using TRIP780 double C-shaped parts as a case study, the surrogate model was trained with 120 groups of Latin hypercube sampling data. After optimization, the optimal process parameter combination was obtained achieving maximum thinning rate of 11.3%, thickening area of 4.6%, springback amount of 0.952 mm, and 20% reduction in overall defects. The research shows that the proposed method has both prediction accuracy and engineering applicability in high-dimensional nonlinear stamping process multi-objective optimization.