A feature fusion network with time-frequency cross-attention for rolling bearing fault diagnosis

To address the limitations of existing rolling bearing fault diagnosis methods in time-frequency feature coordination, cross-domain information interaction, and fault feature extraction under noise interference, a Time-Frequency Cross-Attention Network (TFCANet) is proposed. The proposed method employs a dual-branch architecture to extract complementary features from raw vibration signals and their Fast Fourier Transform (FFT) spectra, capturing transient fault impulses, dynamic response patterns, and spectral characteristics. Squeeze-and-Excitation (SE) modules are incorporated into both branches to adaptively emphasize fault-sensitive channels and suppress redundant information. Furthermore, a multi-head time-frequency cross-attention mechanism is developed, using temporal features as queries and spectral features as keys and values to facilitate explicit cross-domain interaction and adaptive feature fusion. Experiments conducted on the Case Western Reserve University (CWRU) bearing dataset demonstrate that TFCANet achieves an accuracy of 0.9954, precision of 0.9951, recall of 0.9948, and F1-score of 0.9949, outperforming the compared methods under the original evaluation protocol. Additional cross-load evaluations, attention-direction comparisons, extended ablation studies, and repeated training experiments are performed to assess diagnostic performance and stability. Across 20 cross-load tests, TFCANet achieves an overall accuracy of 0.980 ± 0.036 and a macro-F1 score of 0.979 ± 0.039. In the extended ablation experiments, the complete model achieves an accuracy of 0.982 ± 0.020 and a macro-F1 score of 0.981 ± 0.022. These results demonstrate that the proposed time-frequency fusion framework effectively extracts discriminative bearing fault features and maintains favorable diagnostic performance across different load conditions, while revealing variations in model performance and stability under different operating conditions.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-18
DOI
https://doi.org/10.1038/s41598-026-80829-y
Primary Topic
Machine Fault Diagnosis Techniques
Type
article
Field-Weighted Citation Impact
0.00

Funders

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

A feature fusion network with time-frequency cross-attention for rolling bearing fault diagnosis

Yi Zhai, Anhua Bian, Wei Zhang, Zhengwei Tang
2 links
Scientific Reports
Machine Fault Diagnosis Techniques
article

A feature fusion network with time-frequency cross-attention for rolling bearing fault diagnosis

Yi Zhai, Anhua Bian, Wei Zhang, Zhengwei Tang
article en

Abstract

To address the limitations of existing rolling bearing fault diagnosis methods in time-frequency feature coordination, cross-domain information interaction, and fault feature extraction under noise interference, a Time-Frequency Cross-Attention Network (TFCANet) is proposed. The proposed method employs a dual-branch architecture to extract complementary features from raw vibration signals and their Fast Fourier Transform (FFT) spectra, capturing transient fault impulses, dynamic response patterns, and spectral characteristics. Squeeze-and-Excitation (SE) modules are incorporated into both branches to adaptively emphasize fault-sensitive channels and suppress redundant information. Furthermore, a multi-head time-frequency cross-attention mechanism is developed, using temporal features as queries and spectral features as keys and values to facilitate explicit cross-domain interaction and adaptive feature fusion. Experiments conducted on the Case Western Reserve University (CWRU) bearing dataset demonstrate that TFCANet achieves an accuracy of 0.9954, precision of 0.9951, recall of 0.9948, and F1-score of 0.9949, outperforming the compared methods under the original evaluation protocol. Additional cross-load evaluations, attention-direction comparisons, extended ablation studies, and repeated training experiments are performed to assess diagnostic performance and stability. Across 20 cross-load tests, TFCANet achieves an overall accuracy of 0.980 ± 0.036 and a macro-F1 score of 0.979 ± 0.039. In the extended ablation experiments, the complete model achieves an accuracy of 0.982 ± 0.020 and a macro-F1 score of 0.981 ± 0.022. These results demonstrate that the proposed time-frequency fusion framework effectively extracts discriminative bearing fault features and maintains favorable diagnostic performance across different load conditions, while revealing variations in model performance and stability under different operating conditions.

Scientific Reports
Qilu University of Technology (CN), Yancheng Teachers University (CN), Shandong Academy of Sciences (CN), Yancheng Institute of Technology (CN), National Supercomputing Center in Jinan (CN)
Natural Science Foundation of Yangzhou City, Government of Jiangsu Province
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