Battery SOH Foundational Model

A state-of-the-art foundational model for battery State of Health (SOH) prediction using advanced time series decomposition and transformer-based methods.

🎯 Key Achievements

Our foundational model achieves exceptional performance on NASA battery degradation datasets:

Battery	Model	RMSE	MAE	R²	MAPE
B0005	TabPFN	0.0006	0.0004	0.9983	0.06%
B0006	TabPFN	0.0013	0.0010	0.9971	0.13%
B0007	TabPFN	0.0017	0.0013	0.9751	0.19%

Performance Comparison

Traditional ML models (XGBoost, RandomForest, GradientBoosting) achieved:

• RMSE: 0.013-0.033 (10-50x worse)
• R²: -4.5 to 0.79 (often negative)
• MAPE: 1.5-4.2%

TabPFN consistently outperforms by 10-50x across all metrics.

Features

Advanced Decomposition Methods

• CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise)

  • Separates capacity signal into intrinsic mode functions (IMFs)
  • Captures regeneration phenomena and noise

• Improved D3R (Dynamic Decomposition with Diffusion Reconstruction)

  • Custom loss functions for smooth, monotonic trend extraction
  • Spatial-temporal transformer architecture
  • Specialized heads for trend, seasonal, and noise components
  • Regularization: smoothness loss, monotonic loss, seasonal regularity

Temporal Feature Engineering

• ARIMA-based features

  • Lag features (1-10 lags)
  • Multi-horizon forecasts (1-5 steps ahead)
  • Autocorrelation and partial autocorrelation
  • Residuals and fitted values

• Rolling statistics

  • Multiple window sizes (3, 5, 10 cycles)
  • Mean, std, min, max, trend, skewness, kurtosis

• Degradation indicators

  • Capacity fade rate and acceleration
  • Cumulative degradation
  • Regeneration detection and counting
  • Internal resistance proxy

State-of-the-Art Models

• TabPFN: Pre-trained transformer for tabular data (primary model)
• Ensemble methods: XGBoost, GradientBoosting, RandomForest
• Automated model selection and hyperparameter optimization

Total Features: 100-150 per battery:

Raw measurements (7) voltage, current, temperature, timeRolling statistics (24) mean, std, min, max (windows: 3,5,10)Degradation8fade rate, cumulative fade, EWMAHealth indicators (5) voltage range, resistance proxy, temp increase Cycle features (4) normalized cycle, cycle², √cycle Statistical (6) skewness, kurtosis, trend Regeneration (3) increase flag, count, variance CEEMDAN (15-20IMF) energies, trend, seasonal, noiseD3R3trend, seasonal, noise ARIMA (40+) lags, forecasts, ACF, PACF, residuals

🏗️ Architecture

┌─────────────────────────────────────────────────────────────┐ │ Raw Battery Data (.mat) │ └────────────────────────┬────────────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────────┐ │ Data Loading & Preprocessing │ │ • Extract discharge cycles │ │ • Add timestamps │ │ • Truncate at 70% capacity (EOL) │ │ • Calculate SOH │ └────────────────────────┬────────────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────────┐ │ Feature Engineering (40+ features) │ │ • Rolling statistics │ │ • Degradation features │ │ • Health indicators │ │ • Statistical features │ └────────────────────────┬────────────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────────┐ │ CEEMDAN Decomposition │ │ • Trend extraction │ │ • Cyclical patterns │ │ • Noise separation │ │ • IMF energy features │ └────────────────────────┬────────────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────────┐ │ Improved D3R Decomposition (Transformer) │ │ • Smooth trend (degradation) │ │ • Seasonal patterns (regeneration) │ │ • Noise (residual) │ │ • Embedding features │ └────────────────────────┬────────────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────────┐ │ ARIMA Feature Engineering │ │ • Lag features │ │ • Forecasts (1-5 steps) │ │ • Autocorrelation │ │ • Differencing │ └────────────────────────┬────────────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────────┐ │ TabPFN + Ensemble Prediction │ │ • TabPFN (pre-trained transformer) │ │ • XGBoost, GradientBoosting, RandomForest │ │ • Ensemble averaging │ │ • Feature importance analysis │ └────────────────────────┬────────────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────────┐ │ SOH Predictions │ │ • RMSE: 0.0006-0.0017 │ │ • R²: 0.975-0.998 │ │ • MAPE: 0.06-0.19% │ └─────────────────────────────────────────────────────────────┘

License This project is licensed under the MIT License - see LICENSE file. 🙏 Acknowledgments

NASA PCoE for the battery dataset: https://www.nasa.gov/intelligent-systems-division/discovery-and-systems-health/pcoe/pcoe-data-set-repository/ TabPFN authors for the pre-trained model D3R authors (ForestsKing) for decomposition inspiration EMD-signal contributors for CEEMDAN implementation

📧 Contact

Author: [Your Name] Email: [[email protected]] GitHub: @yourusername LinkedIn: Your Profile

📖 Citation If you use this work, please cite: @software{battery_soh_foundational_2025, title={Battery SOH Foundational Model: Advanced Time Series Decomposition with TabPFN}, author={Your Name}, year={2025}, url={https://github.com/yourusername/battery_SOH_FoundationalModel} }