Hybrid Deep Learning for Gearbox Fault Diagnosis Using Raw Vibration Time‐Series Data
Structural Control and Health Monitoring · 2026-01 · 2026 (1)
DOI: 10.1155/stc/8787490 ↗Kurz gefasst
Fachartikel in der Zeitschrift Structural Control and Health Monitoring, online erschienen am 10.08.2026. Verfasst von Yıldırım Özüpak. Die Arbeit wurde bisher 7-mal zitiert (Crossref) und nennt 39 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Reliable gearbox fault diagnosis is an important component of condition monitoring and predictive maintenance in industrial systems. Vibration‐based approaches are widely used for this purpose; however, raw vibration signals are often noisy, nonstationary, and affected by changing operating conditions. In this study, a hybrid deep learning framework combining a convolutional neural network (CNN), bidirectional long short‐term memory (BiLSTM), and gated recurrent unit (GRU) is investigated for gearbox fault diagnosis directly from raw vibration time‐series data. Within the proposed framework, the CNN module is used to extract localized fault‐related patterns, the BiLSTM module models bidirectional temporal dependencies, and the GRU layer provides compact temporal summarization for classification. The model is evaluated on an industrial‐grade gearbox vibration dataset using stratified cross‐validation, out‐of‐fold (OOF) predictions, and an independent holdout set. Performance is assessed using ROC–AUC, F1 score, and accuracy, together with threshold optimization, bootstrap‐based confidence interval estimation, statistical comparison, and interpretability analysis. On the considered benchmark, the CNN–BiLSTM–GRU model achieves strong diagnostic performance, reaching a ROC–AUC of 0.998, an F1 score of 0.979, and an accuracy of 0.980, while generally outperforming standalone and reduced hybrid variants. Additional analyses using bootstrap confidence intervals, the DeLong test, LIME, and SHAP further support the stability and interpretability of the obtained results. The findings suggest that the proposed hybrid structure is effective for raw vibration‐based gearbox fault diagnosis on the evaluated binary benchmark. However, the results should be interpreted within the scope of the employed dataset, which represents a relatively constrained fault scenario. Accordingly, the main contribution of the study lies in the integration of hybrid temporal modeling, leakage‐aware evaluation, threshold analysis, and interpretability within a single diagnosis framework, rather than in claiming universal generalization across all industrial gearbox conditions.
Zitieren
APA
Özüpak, Y. (2026). Hybrid Deep Learning for Gearbox Fault Diagnosis Using Raw Vibration Time‐Series Data. Structural Control and Health Monitoring, 2026(1). https://doi.org/10.1155/stc/8787490
Vancouver
Özüpak Y. Hybrid Deep Learning for Gearbox Fault Diagnosis Using Raw Vibration Time‐Series Data. Structural Control and Health Monitoring. 2026;2026(1). doi:10.1155/stc/8787490
BibTeX
@article{zpak2026,
author = {Özüpak, Yıldırım},
title = {Hybrid Deep Learning for Gearbox Fault Diagnosis Using Raw Vibration Time‐Series Data},
journal = {Structural Control and Health Monitoring},
year = {2026},
volume = {2026},
number = {1},
doi = {10.1155/stc/8787490}
}
EUVIMED