🌀 Wind Turbine Fault Detection with Deep Learning

A machine learning–assisted framework for early fault detection in operational wind turbines using SCADA data.

📌 Project Overview

This project was developed in collaboration with the Renewable Energy Research Group under the supervision of Dr. Ahmaid. The primary objective was to build an intelligent system that detects early-stage faults in wind turbines to optimize inspection schedules and maintenance operations.

We leveraged SCADA (Supervisory Control and Data Acquisition) sensor data from wind turbines, focusing on minimal supervision and interpretability.

---

⚙️ Key Features

• ✅ Unsupervised Pretraining: 1D convolutional autoencoder trained solely on normal turbine data to learn healthy operating patterns.
• 🚀 Transfer Learning: Latent representation from the encoder is used to train a classifier.
• 📊 Classification Performance:
  • Test Accuracy: ~80%
  • F1 Score: ~85%
• 🧠 SHAP Feature Importance: Identifies critical sensor readings influencing model decisions.
• ⚖️ Class Imbalance Handling:
  • The dataset was ~90:10 (normal:faulty).
  • We used ADASYN oversampling to rebalance training data effectively.

---

🗂️ Repository Structure

WindTurbine_fault_detection/
│
├── ClassifierHead/               # Classifier built on top of encoder
├── EncoderBlock/                 # Autoencoder model and training
├── Notebooks/                    # Jupyter notebooks for exploration
├── artifacts/                    # Saved models, logs, and metrics
├── config/                       # YAML config files for schema and parameters
├── src/                          # Core source code
│   ├── components/               # Data transformation, training, evaluation
│   ├── pipeline/                 # Training and evaluation pipelines
│   ├── entity/                   # Config entity classes
│   └── utils/                    # Utility functions
│
├── params.yaml                   # Model and training hyperparameters
├── schema.yaml                   # Data schema (feature names, labels)
├── main.py                       # Entry point for pipeline execution
└── README.md                     # Project documentation

---

📈 Workflow Summary

1. Data Preparation

   • Load SCADA data from CSV
   • Drop all-zero (non-informative) sensor columns
   • Impute and scale using KNNImputer and StandardScaler
   • Address class imbalance with ADASYN

2. Pretraining (Autoencoder)

   • Train an unsupervised autoencoder on normal-only samples
   • Extract latent codes for each input sample

3. Transfer Learning (Classifier)

   • Freeze the encoder
   • Train a classification head using balanced training data
   • Save model, architecture, and training metrics

4. Evaluation

   • Use held-out validation data
   • Log precision, recall, F1-score, and AUC
   • Save confusion matrix and SHAP importance plots

---

📊 Results

Metric	Test Set Value
Accuracy	~80%
F1 Score	~85%
AUC Score	~97% (train) / ~98% (val)*
Precision	~94%
Recall	~91%

Note: The AUC anomaly during validation suggests a threshold mismatch or class imbalance during slicing. Use thresholds= and match data dimensions when evaluating.

---

📦 Installation

git clone https://github.com/Shahriyar-1988/WindTurbine_fault_detection.git
cd WindTurbine_fault_detection
pip install -r requirements.txt

Ensure the following key dependencies are installed:

• TensorFlow >= 2.12
• scikit-learn
• imbalanced-learn
• matplotlib
• pandas
• mlflow

---

🧠 SHAP Analysis

Feature importance was computed using SHAP values on the classifier’s predictions, offering insights into which SCADA sensor readings most influenced fault predictions.

---

📌 Future Improvements

• Incorporate real-time anomaly alerting
• Extend to multi-class fault detection
• Optimize SHAP for batch-wise inference
• Consider interpretable models like LIME or attention-based networks

---

🙏 Acknowledgments

This project was conducted as part of an academic collaboration with the Renewable Energy Research Group, supervised by Dr. Ahmaid. Special thanks to the SCADA data providers and the MLflow community.

---

📬 Contact

Shahriyar
GitHub: @Shahriyar-1988

---

📝 License

This repository is open-source and available under the MIT License.