DFPSE: Diagnostic-Feature-Preserving Signal Enhancement for Robust Bearing Fault Diagnosis under Compound Noise
DOI:
https://doi.org/10.54097/5rk1xm44Keywords:
Bearing fault diagnosis, signal enhancement, knowledge distillation, compound noise, deep learningAbstract
Rolling bearing fault diagnosis using deep learning (DL) achieves high accuracy under clean laboratory conditions, yet performance collapses under the compound noise typical of industrial environments — Gaussian white noise, harmonic interference, transient impulses, and sensor drift. Existing signal denoising methods optimize for waveform reconstruction fidelity, but we demonstrate a critical and counterintuitive finding: higher signal-to-noise ratio (SNR) does not imply better diagnostic accuracy, and reconstruction-only enhancement can actively degrade classification performance below the unenhanced noisy baseline. To address this disconnect, we propose Diagnostic-Feature-Preserving Signal Enhancement (DFPSE), a framework in which a frozen, pretrained diagnostic model (teacher) guides a lightweight Residual Temporal Convolutional Network (Residual TCN) enhancer via three complementary losses: waveform reconstruction, diagnostic classification, and feature-level alignment. Trained under a single Gaussian drift noise condition (SNR in [-10, 8] dB), the enhancer is evaluated across six unseen noise conditions spanning five Gaussian drift levels and a compound interference scenario. Experiments on the Huazhong University of Science and Technology (HUST) and Paderborn University (PU) bearing datasets with three architecturally diverse teachers — ResNet1D (CNN), TCN (dilated convolution), and Vanilla Transformer (self-attention) — demonstrate that DFPSE consistently recovers diagnostic accuracy with gains of 22–29 percentage points on HUST (average across conditions) and 9–22 points on PU. Critically, on the HUST dataset at -2 dB input SNR, the reconstruction-only variant achieves 5.59 dB output SNR but only 65.4% accuracy (below the 86.7% noisy baseline), whereas DFPSE achieves 2.72 dB SNR with 98.4% accuracy — a 33-point gap that reveals a fundamental trade-off between waveform fidelity and diagnostic discriminability. This finding challenges the conventional assumption that better denoising necessarily yields better diagnosis and argues for task-aware evaluation in signal enhancement research.
Downloads
References
[1] Zhu, Z., Lei, Y., Qi, G., Chai, Y., & Li, N. (2023). A review of the application of deep learning in intelligent fault diagnosis of rotating machinery. Measurement, 206, 112346. https://doi.org/10.1016/j.measurement.2022.112346
[2] Tama, B.-A., Vania, M., Lee, S., & Lim, S. (2023). Recent advances in the application of deep learning for fault diagnosis of rotating machinery using vibration signals. Artificial Intelligence Review, 56(5), 4667–4709. https://doi.org/10.1007/s10462-022-10293-3
[3] Matania, O., Dattner, I., Bortman, J., Kenett, R.-S., & Klein, R. (2024). A systematic literature review of deep learning for vibration-based fault diagnosis of critical rotating machinery. Journal of Sound and Vibration, 590, 118562. https://doi.org/10.1016/j.jsv.2024.118562
[4] Li, Y., Cheng, G., Liu, Y., & Chen, X. (2021). Research on bearing fault diagnosis based on spectrum characteristics under strong noise interference. Measurement, 169, 108509. https://doi.org/10.1016/j.measurement.2020.108509
[5] Wu, Z., Jiang, H., Zhao, K., & Li, X. (2020). An adaptive deep transfer learning method for bearing fault diagnosis. Measurement, 151, 107227. https://doi.org/10.1016/j.measurement.2019.107227
[6] Li, X., Jiang, H., Wang, R., & Niu, M. (2021). Rolling bearing fault diagnosis using optimal ensemble deep transfer network. Knowledge-Based Systems, 213, 106695. https://doi.org/10.1016/j.knosys.2020.106695
[7] Wang, H., Liu, Z., Peng, D., Qin, Y., & Shi, J. (2022). Attention-guided joint learning CNN with noise robustness for bearing fault diagnosis and vibration signal denoising. ISA Transactions, 128, 470–484. https://doi.org/10.1016/j.isatra.2021.11.028
[8] Hong, S., & Kim, J.-M. (2023). 1D convolutional neural network-based adaptive algorithm structure with system fault diagnosis and signal denoising. Mechanical Systems and Signal Processing, 197, 110395. https://doi.org/10.1016/j.ymssp.2023.110395
[9] Shang, Y., Zhao, X., Yan, R., & Chen, X. (2023). Denoising Fault-Aware Wavelet Network: A signal processing informed neural network for fault diagnosis. Chinese Journal of Mechanical Engineering, 36(1). https://doi.org/10.1186/s10033-023-00838-0
[10] Kim, S., & Kim, J.-M. (2024). Physics-informed time-frequency fusion network with attention for noise-robust bearing fault diagnosis. IEEE Access, 12, 12517–12532. https://doi.org/10.1109/ACCESS.2024.3355268
[11] Han, T., Shen, C., Wang, D., Kong, D., & Li, Y. (2025). A novel dual-domain adversarial method for vibration signal denoising in bearing fault diagnosis. IEEE Transactions on Instrumentation and Measurement, 74, 1–12. https://doi.org/10.1109/TIM.2025.3551836
[12] van den Ende, M., Lior, I., Ampuero, J.-P., Sladen, A., Ferrari, A., & Richard, C. (2023). A self-supervised deep learning approach for blind denoising and waveform coherence enhancement in distributed acoustic sensing data. IEEE Transactions on Neural Networks and Learning Systems, 34, 3371–3384. https://doi.org/10.1109/TNNLS.2021.3132832
[13] Hinton, G., Vinyals, O., & Dean, J. (2015). Distilling the knowledge in a neural network [Preprint]. arXiv. https://doi.org/10.48550/arXiv.1503.02531
[14] Sepahvand, M., Abdali-Mohammadi, F., & Taherkordi, A. (2022). Teacher–student knowledge distillation based on decomposed deep feature representation for intelligent mobile edge computing. Expert Systems with Applications, 202, 117474. https://doi.org/10.1016/j.eswa.2022.117474
[15] Lu, R., Liu, S., Gong, Z., Xu, C., Ma, Z., Zhong, Y., & Li, B. (2024). Lightweight knowledge distillation-based transfer learning framework for rolling bearing fault diagnosis. Sensors, 24(6), 1758. https://doi.org/10.3390/s24061758
[16] Zhou, Y., Wang, J., & Wang, Z. (2022). Bearing faulty prediction method based on federated transfer learning and knowledge distillation. Machines, 10(5), 376. https://doi.org/10.3390/machines10050376
Downloads
Published
Issue
Section
License
Copyright (c) 2026 Academic Journal of Applied Sciences
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
