Multisensor Vibration-Based Wind Turbine Gearbox Anomaly Detection Using Cost-Sensitive Actor--Critic Contextual-Bandit Learning
Description
Accurate gearbox anomaly detection is essential for improving wind turbine reliability and maintenance planning. This study develops MDA-RL-AC, a cost-sensitive actor–critic contextual-bandit framework for file-level gearbox condition classification from multisensor vibration data. Seven vibration channels from the National Renewable Energy Laboratory Gearbox Reliability Collaborative benchmark were segmented into non-overlapping 1-s windows. Each window was represented by 105 statistical and Fast Fourier Transform-derived features, while a five-step causal temporal context with record-wise left zero-padding was used to preserve short-term vibration history. Because the classification action does not influence subsequent vibration contexts, the implemented learning problem is treated as a contextual bandit rather than as a controlled Markov decision process. Bayesian hyperparameters were selected using a strictly nested grouped-validation design, with outer Leave-One-File-Pair-Out evaluation and an inner nine-fold grouped procedure for model selection. Across ten outer folds, MDA-RL-AC achieved fold-macro Accuracy of 96.25%, Precision of 97.01%, Recall of 95.50%, F1-score of 96.22%, ROC–AUC of 0.9848, PR–AUC of 0.9846, and a normalized classification cost of 0.0425. The results demonstrate strong file-level separability under the evaluated benchmark condition. However, because the same two physical gearboxes contribute records to all folds, the findings do not establish cross-turbine or cross-gearbox generalization. Ablation results further indicate that feature-representation performance is criterion dependent rather than uniformly improved by combining statistical and spectral descriptors.
Files
Steps to reproduce
The data used in this study consist of vibration signals from the GRC–NREL benchmark dataset for gearbox anomaly analysis in wind turbines. The dataset was organized into two operating conditions: healthy (`Healthy_CSV`) and damaged (`Damaged_CSV`). Each condition comprises 10 independent records, denoted as H1–H10 for the healthy condition and D1–D10 for the damaged condition. Seven vibration measurement channels were considered for each record: AN3, AN4, AN6, AN7, AN8, AN9, and AN10. Each CSV file contains 2,400,000 samples, corresponding to 60 s of vibration data acquired at a sampling frequency of 40 kHz. Therefore, the complete experimental dataset consists of 20 records × 7 channels, yielding a total of 140 CSV files. To ensure full reproducibility of the experiments, the dataset must preserve the following directory structure: Healthy_CSV/ ├── H1/ │ ├── AN3.csv │ ├── AN4.csv │ ├── AN6.csv │ ├── AN7.csv │ ├── AN8.csv │ ├── AN9.csv │ └── AN10.csv ├── H2/ │ └── ... └── H10/ └── ... Damaged_CSV/ ├── D1/ │ ├── AN3.csv │ ├── AN4.csv │ ├── AN6.csv │ ├── AN7.csv │ ├── AN8.csv │ ├── AN9.csv │ └── AN10.csv ├── D2/ │ └── ... └── D10/ └── ... The main experimental pipeline is implemented in `wind_turbine_gearbox_actor_critic.py`. Python 3.10 or later is recommended. The required dependencies can be installed using: pip install numpy pandas scipy scikit-learn torch scikit-optimize Before running the complete experiment, the software and dataset configuration can be validated using the built-in self-check procedure: python wind_turbine_gearbox_actor_critic.py --self-check-only The complete experimental pipeline can subsequently be executed as follows: python wind_turbine_gearbox_actor_critic.py The implementation automatically performs data loading, segmentation of the vibration signals into non-overlapping 1-s windows, extraction of statistical and frequency-domain features using the Fast Fourier Transform (FFT), feature scaling based exclusively on the training data, model training, and final performance evaluation. The classification stage is based on a cost-sensitive Actor–Critic contextual-bandit framework designed to account for the asymmetric consequences of anomaly-detection errors. The pipeline also automatically generates the corresponding experimental result files. For exact experimental reproducibility, the directory hierarchy, record identifiers, channel names, CSV filenames, and sampling configuration should remain unchanged. Maintaining this standardized organization ensures that the data-loading, preprocessing, feature-extraction, training, and evaluation procedures can be executed without manual modifications to the source code.
Institutions
- Universidad Complutense de MadridMadrid, Madrid
- CUNEF UniversidadMadrid, Madrid
- Pontificia Universidad Católica del Ecuador Sede AmbatoTungurahua Province, Ambato