A decomposition-based multi-objective particle swarm optimization algorithm with a local search strategy for key quality characteristic identification in production processes

Abstract

Identifying the key quality characteristics (KQCs) (including part and process parameters) in production processes is essential for quality control. In this paper, we propose a data-driven KQC identification method based on production process data. We model KQC identification as a multi-objective feature selection problem of maximizing the geometric mean (GM) and minimizing the number of selected QCs (features). GM can evaluate the importance of a QC subset by measuring its predictive ability for product quality. To solve this optimization model, we propose a multi-objective optimization algorithm called MOPSO-LS that combines particle swarm optimization (PSO) with a local search strategy. MOPSO-LS adopts a decomposition approach, i.e., Tchebycheff approach (TA), to update personal best positions (pbests) during the iterations. Thus, diversified and high quality solutions can be maintained by the pbests of particles. Moreover, the local search strategy aims to update the non-dominated set found by MOPSO-LS during the iterations with two basic local search steps, i.e., a) adding and b) removing a feature, which can improve the convergence performance of MOPSO-LS. We have verified the proposed method on four production datasets. The experimental results indicate that MOPSO-LS can select a few KQCs with a good capability for predicting product quality, which shows the effectiveness of MOPSO-LS for KQC identification. Further comparisons show that MOPSO-LS obtains better search performance than four typical multi-objective optimization algorithms.