关键词: 涡轮盘, 混合样本, 人工神经网络, 多目标粒子群优化算法, 增量学习

Abstract: To dealing with the computing-intensive problem of multi-objective optimization for turbine disc,an Artificial Neural Network(ANN)surrogate was proposed as fitness function of Multi-Objective Particle Swarm Optimization (MOPSO) algorithm based on the management of data sets of different precision.By the combination of discrepancy control of uniform design and error analysis of ANN,the management of Finite Element Analysis (FEA) samples with different precision was processed,and the noise-based virtual sample was used to enlarge the training set of ANN.Based on the error analysis,noise intensity-number of hidden nodes-virtual samples size was optimized collaboratively to minimize the generalization error and approximating error.The performance of MOPSO was improved with neighbourhood information of background grids.A forgetting mechanism based incremental learning method was used to enhance the guidance ability of optimization objects.The proposed strategy was validated with a multi-objective optimization of turbine disc.Experiment showed that the design cost could be reduced significantly with the proposed method under the premise of ensuring the multi-target optimization quality of turbine disk.

Key words: turbine disc, mixed data set, artificial neural network, multi-objective particle swarm optimization, incremental learning