Remaining Useful Life Estimation of Turbofan Engines

This repository contains a machine learning project aimed at predicting the Remaining Useful Life (RUL) of turbofan engines using the PHM08 Prognostics Data Challenge Dataset provided by NASA. Accurate RUL Estimation can enable proactive maintenance, reduce operational costs, and prevent unexpected failures.

📊 Dataset Overview

The PHM08 dataset includes:

• 218 engine units: Each representing a unique turbofan engine.
• 21 sensor measurements: Including fan speed, pressure, temperature, and vibration parameters.
• Degradation trajectories: Each engine's data spans from normal operation to failure.
• Training and test sets: Provided for model development and evaluation.

Source: PHM08 Challenge Data Set

🚀 To get started with RUL predicton project:

1. Clone the repository:

   git clone https://github.com/palscruz23/rul-estimation.git
   cd rul-estimation
   

2. Install dependencies

   pip install -r requirements.txt
   

3. Run RUL using machine learning methods

   Change directory to machine-learning folder

   cd machine-learning
   

   Open streamlit app for the RUL Estimation visualization

   streamlit run src/Main.py
   

4. Run RUL using deep learning methods (RNN, LSTM, Seq2Seq, Informer models)

   Change directory to deep-learning folder

   cd deep-learning
   

   Start the FastAPI server using Uvicorn:

   uvicorn service.service:app --reload --host 127.0.0.1 --port 8000
   

   Run inference from service/predict.py script.

   python -m service.predict 
   

   (Optional) To check ML flow experiments, run then open http://127.0.0.1:5000/ in browser.

   cd src
   mlflow ui --port 5000
   

📚 Projects Overview

Remaining Useful Life Estimation using machine learning techniques

• Data Exploration
  • Load training and test data
  • Utilise unsupervised learning (K-means) to get operational parameter clusters
• Data Processing
  • Split data set into training and validation sets
  • Apply Standard Scaler to normalise different sensor measurements
• Remaining Useful Life Estimation
  • Construct Asset Health Indicator
  • Perform Linear Regression on each sensor to get trendability
  • Perform sensor fusion to develop health indicator
  • Develop Residual-similarity model using Degree-2 polynomial fit
  • Plot health indicator of validation engine from 5% to 100% operating life

📉 RUL Estimation using ML Demo

📃 References

[1] MATLAB Similarity-Based Remaining Useful Life Estimation

[2] A similarity-based prognostics approach for Remaining Useful Life estimation of engineered systems

Remaining Useful Life Estimation using deep learning techniques

• Data Processing
  • Load training data
  • Split data set into training, validation and test sets
  • Perform scaling using StandardScaler
  • Create PHM08RULDataset dataset class
• Remaining Useful Life Estimation
  • Initiate ML flow experiment
  • Create model classes for RNN, LSTM, Seq2Seq and Informer
  • Prepare training and validation loops
  • Perform grid search for hyperparameter tuning
  • Select best model
  • Perform bias vs variance analysis
  • Perform Estimation on test data.

📉 RUL Estimation using DL Demo

📃 References

[1] Recurrent neural networks for remaining useful life estimation

[2] Recurrent Neural Network (RNN)

[3] Understanding LSTMs

[4] Understanding the Seq2Seq Model

[5] Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting

📜 License

This repository is licensed under the MIT License. See the LICENSE file for more details.