Dual-stream 1D mobile-net with channel attention fusion for rolling element fault diagnosis
- 1School of Mechanical Engineering, Vellore Institute of Technology
Abstract
Accurate, lightweight, and real-time diagnostic devices are required because undetected rolling element bearing problems result in unscheduled industrial downtime and substantial financial loss. Inspired by coordinate attention concepts, this work offers DS-Mobile-Net-CA, a Dual-Stream 1D Mobile-Net with a 1D Channel Attention block. The model uses raw 1D vibration signals without any intermediate 2D image transformation to classify five bearing fault conditions: Healthy, Inner Race Fault (IRF), Outer Race Fault (ORF), Ball Fault, and Combined Fault. A Spectra Quest MFS-PKG5 test-rig attached with a Kistler 8763A50 IEPE accelerometer recording at sampling rate of 12,800 Hz was used to collect vibration data at three rotational speeds (1000, 1500, and 2000 rpm) and three load conditions (0, 6, and 12 g). In each stream, the architecture employs 1D depth-wise separable convolutions (Mobile-Net blocks), which are then concatenated and tuned via a channel attention technique. Through its ability to offer a mere 81,829 trainable parameters, 4.8 M FLOPs, and a time required for inference equal to 3.98ms and being the most computationally inexpensive among all other designs studied. The architecture is capable of delivering a classification accuracy of 99.80%, a weighted F1-score of 99.74%, as well as a Matthews Correlation Coefficient (MCC) value of 0.9974. In order to check the statistical validity of the study results, K-fold cross-validation (K = 5, average = 99.17%), ablation analysis, Confidence Interval 95% calculation, and Friedman ranking test (p = 0.0067) were conducted.
Related
- Cross-corpus consensus feature selection for language-invariant Parkinson’s disease detection: a dual-stream acoustic and spectrogram-attention framework
- Dual-stream attention and feature fusion for human activity recognition via wearable sensors
- Channel attention of the image for distinguishing ductal carcinoma in situ from benign mimics based on ultrasound images
- AVMD-ST: attention-variational mode decomposition based spectrogram transformer for differential diagnosis of asthma and COPD
- Interpretable attention-deficit hyperactivity disorder diagnosis from functional magnetic resonance imaging via hybrid convolutional neural network–transformer learning and multi-level graph-based comprehensibility
BibTeX
@article{kumar2026dualstream,
title = {{Dual-stream 1D mobile-net with channel attention fusion for rolling element fault diagnosis}},
author = {Raghav Kumar and Ramireddy Dheeraj Reddy and T. Narendiranath Babu and M. Pandiyan},
journal = {Scientific Reports},
year = {2026},
month = aug,
doi = {10.1038/s41598-026-64845-6},
url = {https://doi.org/10.1038/s41598-026-64845-6},
note = {CC BY},
}
RIS
TY - JOUR
TI - Dual-stream 1D mobile-net with channel attention fusion for rolling element fault diagnosis
AU - Raghav Kumar
AU - Ramireddy Dheeraj Reddy
AU - T. Narendiranath Babu
AU - M. Pandiyan
PY - 2026
DA - 2026/08/01
JO - Scientific Reports
DO - 10.1038/s41598-026-64845-6
UR - https://doi.org/10.1038/s41598-026-64845-6
AB - Accurate, lightweight, and real-time diagnostic devices are required because undetected rolling element bearing problems result in unscheduled industrial downtime and substantial financial loss. Inspired by coordinate attention concepts, this work offers DS-Mobile-Net-CA, a Dual-Stream 1D Mobile-Net with a 1D Channel Attention block. The model uses raw 1D vibration signals without any intermediate 2D image transformation to classify five bearing fault conditions: Healthy, Inner Race Fault (IRF), Outer Race Fault (ORF), Ball Fault, and Combined Fault. A Spectra Quest MFS-PKG5 test-rig attached with a Kistler 8763A50 IEPE accelerometer recording at sampling rate of 12,800 Hz was used to collect vibration data at three rotational speeds (1000, 1500, and 2000 rpm) and three load conditions (0, 6, and 12 g). In each stream, the architecture employs 1D depth-wise separable convolutions (Mobile-Net blocks), which are then concatenated and tuned via a channel attention technique. Through its ability to offer a mere 81,829 trainable parameters, 4.8 M FLOPs, and a time required for inference equal to 3.98ms and being the most computationally inexpensive among all other designs studied. The architecture is capable of delivering a classification accuracy of 99.80%, a weighted F1-score of 99.74%, as well as a Matthews Correlation Coefficient (MCC) value of 0.9974. In order to check the statistical validity of the study results, K-fold cross-validation (K = 5, average = 99.17%), ablation analysis, Confidence Interval 95% calculation, and Friedman ranking test (p = 0.0067) were conducted.
SN - 2045-2322
ER -