Zero-shot rolling bearing fault diagnosis from vibration measurements using a dual-path channel-attention autoencoder and variable-weight attribute matrices derived from latent Dirichlet allocation (LDA)
Review of Scientific Instruments · 01.11.2026 · 97 (11)
DOI: 10.1063/5.0341030 ↗Kurz gefasst
Fachartikel in der Zeitschrift Review of Scientific Instruments, online erschienen am 20.11.2026. Verfasst von Lunpan Wei und 2 weiteren Autor:innen. Die Arbeit nennt 42 Literaturangaben. Gefördert von: National Natural Science Foundation of China.
Abstract (Originalsprache)
Vibration measurement systems for rotating machinery condition monitoring must reliably identify fault signatures even when rare or newly emerging bearing failure modes produce signal patterns absent from the instrument’s calibration database. Existing vibration-based diagnostic instruments employing deep learning classifiers require large, fully labeled vibration datasets for every fault category—an impractical constraint in industrial deployments where novel fault modes arise unpredictably. To address this, we propose a semantic-guided zero-shot learning framework capable of diagnosing unseen faults without labeled training specimens. The framework combines two elements in a configuration tailored to the requirements of the zero-shot setting. First, the proposed dual-path efficient channel attention stacked autoencoder places an efficient channel attention branch in parallel with a one-dimensional convolutional branch at the bottleneck of a stacked autoencoder, capturing local cross-channel and channel-specific discriminatory representations from multi-domain vibration features. Second, a data-driven, variable-weight fault attribute matrix is constructed. Latent Dirichlet allocation extracts semantic attributes and importance weights from a unified textual fault corpus, while term frequency–inverse document frequency vectorization yields a compact binary matrix. A nearest-neighbor Bayesian mapping then projects vibration features into this semantic space. Experiments on the Case Western Reserve University benchmark and on an independent test rig give unseen-class recognition accuracies of 88.7% and 86.3%, respectively, exceeding the strongest of the six baselines compared by 1.8 and 2.2 percentage points and the classical direct attribute prediction baseline by 17.4 and 17.1 points.
Förderung
- National Natural Science Foundation of China
Zitieren
APA
Wei, L., Tan, Y., & Peng, X. (2026). Zero-shot rolling bearing fault diagnosis from vibration measurements using a dual-path channel-attention autoencoder and variable-weight attribute matrices derived from latent Dirichlet allocation (LDA). Review of Scientific Instruments, 97(11). https://doi.org/10.1063/5.0341030
Vancouver
Wei L, Tan Y, Peng X. Zero-shot rolling bearing fault diagnosis from vibration measurements using a dual-path channel-attention autoencoder and variable-weight attribute matrices derived from latent Dirichlet allocation (LDA). Review of Scientific Instruments. 2026;97(11). doi:10.1063/5.0341030
BibTeX
@article{Wei2026,
author = {Wei, Lunpan and Tan, Yinchao and Peng, Xiuyan},
title = {Zero-shot rolling bearing fault diagnosis from vibration measurements using a dual-path channel-attention autoencoder and variable-weight attribute matrices derived from latent Dirichlet allocation (LDA)},
journal = {Review of Scientific Instruments},
year = {2026},
volume = {97},
number = {11},
doi = {10.1063/5.0341030}
}
EUVIMED