Imbalanced Multifault Diagnosis via Improved Localized Feature Selection

In real-world industrial processes, various fault samples are often much less than normal samples, which makes fault diagnosis an imbalanced multiclass classification problem. Feature selection (FS) has shown its potential in handling class imbalance and reducing redundant features to enhance the model generalization. However, conventional FS is usually combined with a balancing strategy to handle imbalanced data, which prevents the feature subset from estimating the true data distribution. At the same time, influenced by the varying distribution of samples from different local regions, the selection of a globally optimal feature subset for all samples often suffers from poor generalization. In this article, an improved localized FS (LFS) approach based on multiobjective binary particle swarm optimization (LFS-MOBPSO) is proposed to address fault diagnosis from a novel perspective that takes advantage of the local distribution of data without the need to use balancing strategies. Different from existing LFS algorithms, LFS-MOBPSO considers optimizing two conflicting objectives, simultaneously. A derivative-free algorithm, multiobjective binary particle swarm optimization (MOBPSO) is developed to solve the LFS problem without any formulation transformation or convex relaxation, where a novel solution selection strategy out of Pareto solutions as the final local feature subset is applied. Simulation results on benchmark datasets and real-world case studies demonstrate the superiority of LFS-MOBPSO to the state-of-the-art LFS and imbalanced ensemble classifiers under different settings of imbalanced ratio (IR).

Automated summary

In this study, an improved localized feature selection method based on multiobjective binary particle swarm optimization was proposed to address fault diagnosis by utilizing the local distribution of data without the need for balancing strategies.