Answers are grounded in this paper and cite their sources.
Conventional data-driven fault diagnosis for rolling-element bearings often uses static dimensionality-reduction methods, such as principal component analysis (PCA). Such methods may preserve load-related static variation and produce inconsistent evidence when several classifiers are fused. This paper proposes a dynamic feature extraction and evidence fusion method based on the probabilistic reduced-dimensional vector autoregressive (PredVAR) model proposed in [Mo and Qin, 2025, Automatica]. The raw vibration signal is first reconstructed by delay embedding, and PredVAR is then used to extract predictable dynamic latent variables (DLVs). The extracted DLVs are converted into compact dynamic feature vectors for learning vector quantization (LVQ) and decision tree (DT) classifiers, whose outputs are combined by Dempster-Shafer (D-S) evidence theory. Experiments on a mixed-load CWRU bearing dataset show that the proposed PredVAR + D-S framework achieves an average accuracy of 99.91% with a standard deviation of 0.0014 under the main 20-run protocol. The D-S conflict coefficient is reduced from 0.0195 to 0.0056 compared with PCA-based features, indicating more consistent classifier evidence. The proposed method can be positioned as a compact and interpretable dynamic-fusion approach. In a limited-label setting with 50 labeled windows per class, PredVAR + D-S obtains the highest mean accuracy among the compared methods, suggesting an advantage when labeled fault data are scarce.
No takes yet. Share an insight, caveat, or question.
Zhu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: