Physics inspired LSTM anomaly detector for time-series data

Project Summary

This project demonstrates working principles of anomaly detection in a physics inspired LSTM network for a time series data.

Dataset

• Simulated simple harmonic oscillator data
• Clean simple harmonic oscillator over limited time window for training
• Longer signal with the same frequency with noise and known anomalies for testing

Methodology

• Choice of network: LSTM for better preservation of features in time series data
• Implmentation of physics loss function as SHM ODE
• Tune F1 score to achive optimal anomaly detection performance
• Comparison of standard LSTM and PI-LSTM networks

Results

• Visulization of loss curves for both PI-LSTM and standard LSTM networks
• Visualization of predicted vs actual anomalies
• Comparison of standard LSTM and PI-LSTM network performances

How to Run

1. Clone repo
2. Create environment: pip install -r requirements.txt
3. Run main.py for detailed analysis and model training.

Future Work

• Implement uncertainty quantification.
• Implement frequency estimation
• Inject more realistic signals with drift, glitches etc.

Technologies

Python, pytorch, pandas, scikit-learn, matplotlib

---

Feel free to reach out or check my portfolio: suchitakulkarni.github.io