Comparative Deep Learning-Based Fault Diagnosis of a Two-Stage Gearbox Using Vibration Data

İşci M., Yıldırım B.

VI. International BRITISH Congress on Interdisciplinary Scientific Research & Practices, London, İngiltere, 18 - 21 Haziran 2026, ss.1, (Özet Bildiri)

  • Yayın Türü: Bildiri / Özet Bildiri
  • Basıldığı Şehir: London
  • Basıldığı Ülke: İngiltere
  • Sayfa Sayıları: ss.1
  • Kayseri Üniversitesi Adresli: Evet

Özet

This study addresses the modeling of vibration signals using deep learning algorithms, based on the “Gearbox Fault Diagnosis Data” set shared in the OEDI database. The relevant dataset, derived with the SpectraQuest simulator, includes vibration data for both intact and structurally damaged conditions in a two-stage gear mechanism. Collected using accelerometers from four different axes, this dataset was recorded under load variations ranging from 0% to 70%. The existing dataset provides a suitable foundation for developing algorithms for fault diagnosis, given its distinctive spectral components and modulation sidebands. Given the time-series nature of the vibration data, Bi-LSTM and Bi-GRU architectures were designed and their performance compared. In the modeling hierarchy, signals were first passed through a moving-average smoothing filter and then normalized using the Min-Max technique. 120-step input windows were defined to capture temporal correlation; early stopping and dynamic learning rate reduction protocols were implemented during the training phase to prevent overlearning. Both network structures were optimized with two-layer bidirectional recurrent units and dropout regularization. The predictive performance of the models was evaluated using eight-fold cross-validation with MSE, RMSE, MAE, and R2. Based on the findings, the MSE, RMSE, MAE, and R2 scores for the Bi-LSTM model were 0.4353, 0.6584, 0.4992, and 0.9104, respectively. In the Bi-GRU model, these values ​​remained at 0.6382, 0.7770, 0.5914, and 0.8787, respectively. Numerical data demonstrate that the Bi-LSTM model exhibits a more stable graph than the Bi-GRU, with both lower error margins and better data explanation capabilities. In particular, while high error deviations were observed in some test folds of the Bi-GRU architecture, the Bi-LSTM architecture produced much more stable results. Consequently, this comparative analysis of OEDI data confirms that the bidirectional LSTM architecture is a more reliable alternative for vibration-based condition monitoring and fault diagnosis.