关键词: 多目标优化, 人工神经网络, 非支配排序遗传算法, 小样本数据, 初始种群

Abstract: Aiming at the high calculation density of complex product optimization design,an improved Artificial Neural Network-Genetic Algorithm (ANN-GA) Multi-objective Optimization (MOO) framework was proposed from the perspective of sample management.By considering the small sample size problem during the orthogonal experiment process,a limited disturbance based virtual training samples extraction method was adopted to expand the sample set which could enhance the precision of ANN model.Range analysis model with entropy weight was used for sensitivity analysis of decision variables,ANN models with different decision variables were created and compared to find a good model with high cost performance,and the selected ANN was used to calculate the finesses for subsequent GA searching.A new initial population construction method was proposed to achieve a Pareto front with better smoothness and uniformity.The feasibility and effectiveness of the proposed framework were validated by the optimization of high pressure turbine disc.

Key words: multi-objective optimization, artificial neural network, non-dominated sorting genetic algorithmⅡ, small sample data, initial population