Bayesian and Classical Feature Ranking for Interpretable BLDC Fault Diagnosis

Abstract

This paper compares Bayesian and classical feature ranking methods for interpretable fault diagnosis of brushless DC (BLDC) motors. Two Bayesian approaches, spike-and-slab and ARD logistic ranking, are evaluated against three classical baselines on a public BLDC benchmark in binary and multiclass settings using current-based, rotational-speed-based, and combined feature sets. The strongest overall results are obtained for the combined representation. In binary classification, ReliefF achieves the highest balanced accuracy of 0.923, while ARD logistic and spike-and-slab remain very close at 0.919 and 0.920 with much smaller subsets ($k=5$). In multiclass classification, ARD logistic performs best for the combined variant with balanced accuracy 0.914, followed closely by LASSO (0.913) and spike-and-slab (0.912). The results show that Bayesian ranking is particularly competitive for current-only and combined descriptors, while ReliefF remains especially effective for speed-based ranking. Because the benchmark consists of short segmented observations from a limited number of experimental conditions, the findings are interpreted primarily as benchmark-specific evidence rather than strong claims of fault generalization.

Figures (2)

• Figure 1: Compact comparison of the best predictive performance obtained by all ranking methods across tasks and feature variants.
• Figure 2: Summary of top-ranked features and agreement between Bayesian and classical methods.