A multi-objective evolutionary algorithm with mutual-information-guided improvement phase for feature selection in complex manufacturing processes

Complex manufacturing processes (CMP) involve numerous features that impact product quality. Therefore, selecting key process features (KPF) is crucial for effective quality prediction and control in CMPs. This paper proposes a KPF (feature) selection method for the high-dimensional CMP data. The KPF selection problem is formulated as a bi-objective combinatorial optimization task of maximizing the geometric mean measure and minimizing the number of selected features. To solve this challenging high-dimensional KPF selection problem, we propose a novel multi-objective evolutionary algorithm (MOEA) called NSGAII-MIIP. NSGAII-MIIP applies an improvement phase (called MIIP) to purify the non-dominated solutions obtained by genetic operators during the iteration process to improve the FS performance. The improvement phase is guided by a mutual-information-based feature importance measure considering both a feature’s relevance degree to class (product quality level) and its redundancy degree to selected features. This allows MIIP to efficiently update non-dominated solutions by selecting relevant features and eliminating redundant features. Moreover, MIIP is seamlessly integrated into the solution ranking process of NSGAII-MIIP so that solutions from the improvement phase can be ranked together with original solutions in the population efficiently. Experiments on eight datasets show that NSGAII-MIIP has better KPF selection performance than eight state-of-the-art multi-objective FS methods. Moreover, NSGAII-MIIP exhibits superior search performance compared to eight typical multi-objective optimization algorithms.