🔧 Rub Detection in Aeroderivative Gas Turbines

Rotor–stator rub is a common but critical malfunction in turbomachinery, particularly in aero-derivative gas turbines where design constraints limit sensor placement. Due to compact geometries and the use of ball bearings instead of hydrodynamic ones, proximity sensors are impractical, and vibration must be monitored indirectly via accelerometers on the casing.

These signals are affected by high levels of noise, making rub detection in early stages difficult with traditional frequency-domain techniques such as Fourier analysis.

🧠 Our Deep Learning-Based Solution

In our work, we present a novel Intelligent Fault Detection (IFD) system using deep neural networks (DNNs) trained with synthetic vibration data from a calibrated finite element (FE) model of a rotating machine.

This data-driven approach avoids the need to collect dangerous real-world fault data and opens the door to safe, scalable training of deep models. The key features of our methodology include:

Synthetic Dataset Creation: An FE model simulates rub events of varying intensity and frequency (e.g., light rub: 1×; heavy rub: ½×).
Transfer Learning: The DNN trained on simulated data generalizes well to real experimental signals, achieving up to 91% G-Mean and ~86% accuracy.
Data Preprocessing: Accelerometer signals are resampled, windowed, and transformed via FFT to emphasize relevant harmonic content.
Regularization & Optimization: Grid search was used to identify optimal architectures with L1/L2 regularization, improving generalization.

🖼️ Simulation and Experimental Setup

🧩 Finite Element Model Simulation

Here are two visualizations from the FE model used to generate synthetic training data for rub scenarios:

FE model example 1
Finite Element model simulating single-point rub in high-speed shaft.

FE model example 2
FE mesh used to simulate both light and heavy rub events.

🧪 Real Experimental Setup

Real experimental rig
Experimental test bench with mounted accelerometers and proximity sensors used to validate model predictions.

📊 Results and Transfer Learning Performance

We tested multiple DNN architectures with different regularization techniques (L1, L2, ElasticNet) to study how they affect generalization from synthetic to real-world signals.

📋 Result Summary

Result Table
Performance of different DNN models on real test signals after training on synthetic data.

📈 AUC/ROC Curves Comparison

AUC ROC Curves
Comparison of AUC-ROC performance across models with varying regularization. Shows impact of regularization on robustness and transfer learning accuracy.

Squicciarini, A., Zarzo, A., González-Guillén, C.E., Muñoz-Guijosa, J.M.
Application of Deep Neural Networks for Automatic Rub Detection in Aero-Derivative Gas Turbines,
Advanced Engineering Informatics, vol. 62, 2024, Art. no. 102607.
📎 DOI: 10.1016/j.aei.2024.102607

⬅️ Back to Projects