Data-driven adaptive and stable feature selection method for large-scale industrial systems

Abstract

Data-driven modeling is a crucial technology for the real-time monitoring of large-scale industrial systems. However, it often suffers from the redundancy of input variables, resulting in low prediction and modeling accuracy. To address this issue, a novel feature selection method, namely adaptive and stable feature selection based on a reference vector-guided evolutionary multi-objective optimization algorithm (ASFS-RVEA), is proposed in this paper. The proposed ASFS-RVEA comprehensively considers four important objectives: the number of features, prediction accuracy, the dissimilarity of selected features, and the mitigation of feature redundancy.Considering the interaction and conflict among these four objectives, a multi-objective optimization problem with an unknown Pareto front is formulated to find an optimal balance among them, thereby obtaining promising and convincing results. Furthermore, Jensen–shannon divergence (JSD) is introduced to the RreliefF algorithm to account for the data distribution information between various input features and key output variables, guiding population crossover and mutation. This greatly enhances the robustness of the algorithm when handling data with different distributions. Next, a reference vector adapting strategy is proposed to update the generation based on dynamically changing distributions, which helps accelerate convergence in the optimization process. Finally, experiments conducted on datasets collected from the Dow process and the polyester polymerization process demonstrate the effectiveness of the proposed ASFS-RVEA.