Confidence-Guided Selective Refinement Network for Imbalanced Gearbox Fault Diagnosis

Abstract In intelligent gearbox fault diagnosis, class imbalance remains a major challenge: majority-class samples dominate the training process, whereas minority fault samples are scarce and difficult to characterize. Existing methods usually ignore differences in diagnostic difficulty among samples, resulting in insufficient recognition of minority and ambiguous fault categories. To address this problem, this paper proposes a confidence-guided selective refinement network, termed CGH-RMNet, for imbalanced gearbox fault diagnosis. The proposed framework decomposes the original diagnosis task into two stages: majority-class coarse screening and difficult-sample fine-grained refinement. In the first stage, a lightweight ResNet branch performs coarse discrimination and confidence estimation, allowing high-confidence majority-class samples to be directly accepted. In the second stage, samples predicted as complex categories or assigned low confidence are routed to a Mamba-based expert branch for refined full-class identification. Furthermore, an imbalance-aware optimization strategy is designed by combining class-weighted learning and label smoothing to enhance the representation of minority and hard samples. Experiments are conducted on the XJTU gearbox dataset and the MCC-THU Gearbox dataset under constant-speed, variable-speed, and variable-load conditions with different imbalance ratios. The results show that CGH-RMNet consistently outperforms several representative comparison methods in terms of diagnostic accuracy, confusion matrix distribution, and t-SNE-based feature separability. Ablation experiments further verify the effectiveness of the Mamba expert branch, the low-confidence routing mechanism, and the imbalance-aware learning strategy. The proposed method provides an efficient and robust solution for gearbox fault diagnosis under complex operating conditions and severe class imbalance.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-16
DOI
https://doi.org/10.21203/rs.3.rs-11579574/v1
Primary Topic
Machine Fault Diagnosis Techniques
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Confidence-Guided Selective Refinement Network for Imbalanced Gearbox Fault Diagnosis

唐贵基, Kai Sun, Xiao Wang, Wei Zhou et al.
1 links
Research Square
Machine Fault Diagnosis Techniques
preprint

Confidence-Guided Selective Refinement Network for Imbalanced Gearbox Fault Diagnosis

唐贵基, Kai Sun, Xiao Wang, Wei Zhou, Zhenli Xu, Lihong Cheng
preprint en

Abstract

Abstract In intelligent gearbox fault diagnosis, class imbalance remains a major challenge: majority-class samples dominate the training process, whereas minority fault samples are scarce and difficult to characterize. Existing methods usually ignore differences in diagnostic difficulty among samples, resulting in insufficient recognition of minority and ambiguous fault categories. To address this problem, this paper proposes a confidence-guided selective refinement network, termed CGH-RMNet, for imbalanced gearbox fault diagnosis. The proposed framework decomposes the original diagnosis task into two stages: majority-class coarse screening and difficult-sample fine-grained refinement. In the first stage, a lightweight ResNet branch performs coarse discrimination and confidence estimation, allowing high-confidence majority-class samples to be directly accepted. In the second stage, samples predicted as complex categories or assigned low confidence are routed to a Mamba-based expert branch for refined full-class identification. Furthermore, an imbalance-aware optimization strategy is designed by combining class-weighted learning and label smoothing to enhance the representation of minority and hard samples. Experiments are conducted on the XJTU gearbox dataset and the MCC-THU Gearbox dataset under constant-speed, variable-speed, and variable-load conditions with different imbalance ratios. The results show that CGH-RMNet consistently outperforms several representative comparison methods in terms of diagnostic accuracy, confusion matrix distribution, and t-SNE-based feature separability. Ablation experiments further verify the effectiveness of the Mamba expert branch, the low-confidence routing mechanism, and the imbalance-aware learning strategy. The proposed method provides an efficient and robust solution for gearbox fault diagnosis under complex operating conditions and severe class imbalance.

Research Square
North China Electric Power University (CN)
National Natural Science Foundation of China
Machine Fault Diagnosis Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon