Physics-Informed Machine Learning

Battery Degradation & Performance Prediction

This project implements physics-informed machine learning models to predict lithium-ion battery degradation and performance metrics, including Remaining Useful Life (RUL), State of Charge (SoC), and State of Health (SoH).

The project uses comprehensive battery cycling data from 22 high-density pouch cells, encompassing over 21,000 charge-discharge cycles under varied operational conditions.

Data Features:

• Time: timestamp of the measurement
• Voltage: battery voltage (V) - critical for SoC estimation
• Current: battery current (A) - key for power and thermal modeling
• Temperature: operating temperature (°C) - affects degradation patterns
• Cycle Number: charge/discharge cycle count - essential for RUL prediction
• Capacity: battery capacity (Ah) - primary indicator of SoH

File Structure

├── model.py                          # main entry point: command-line interface
├── processing.py                     # data loading and model training functions
├── visualization.py                  # graph generation and plotting functions
├── data.zip                          # battery dataset
├── requirements.txt                  # python dependencies
└── README.md

Script Description

model.py - main entry point with command-line interface for training pipeline configuration

processing.py - data loading and model training:

• loadData: loads battery data from CSV files and prepares features
• trainIndividualFiles: trains models per battery with train/test split
• trainCombinedFiles: trains on multiple batteries, tests on held-out file

visualization.py - graph generation and results plotting:

• plotPredictions: creates scatter plots of predicted vs true capacity
• plotTimeSeries: generates time series plots of battery degradation

Features

• Training Modes:
  • individual: single-file training with train/test split per battery
  • combined: multi-battery training for cross-cell generalization
• Feature Engineering: configurable inclusion of cycle number for different use cases
• Flexible Data Splits: customizable train/test ratios (20/80, 50/50, or any custom split)
• Optimized For Battery Health: supports RUL, SoC, and SoH prediction workflows
• Production-Ready: command-line interface for easy integration into ML pipelines

Installation

The dataset could not be uploaded to this repository because the compressed file exceeds size limits. The files can be downloaded from https://doi.org/10.1184/R1/14226830

# unzip the dataset
unzip data.zip

# install dependencies
pip install -r requirements.txt

Usage

Running The Script

# individual file training (80/20 split) with cycle number
python model.py --mode individual --include-cycle-num --test-size 0.2

# individual file training (50/50 split) with cycle number
python model.py --mode individual --include-cycle-num --test-size 0.5

# individual file training (80/20 split) without cycle number
python model.py --mode individual --test-size 0.2

# individual file training (50/50 split) without cycle number
python model.py --mode individual --test-size 0.5

# multi-file training with cycle number
python model.py --mode combined --include-cycle-num --test-file VAH07_r2.csv

# multi-file training without cycle number
python model.py --mode combined --test-file VAH01_r2.csv

# process specific files only
python model.py --mode individual --files VAH07_r2.csv VAH16_r2.csv --test-size 0.2

# custom configuration with more estimators
python model.py --mode individual --num-estimators 200 --seed 123

Command-Line Arguments

• --mode: training mode (individual or combined)
  • individual: train/test split on each file separately
  • combined: train on multiple files, test on one held-out file
• --include-cycle-num: include cycle number as input feature (flag, default: false)
• --test-size: test set proportion for individual mode (default: 0.2)
• --test-file: held-out test file for combined mode (default: VAH07_r2.csv)
• --num-estimators: number of trees in random forest (default: 100)
• --seed: random seed for reproducibility (default: 42)
• --files: specific files to process (default: all 22 battery files)

Model Parameters

• Algorithm: random forest regressor (suitable for real-time inference)
• Number of Estimators: 100 (configurable via --num-estimators)
• Random Seed: 42 (configurable via --seed)
• Feature Engineering: temporal normalization with optional cycle number inclusion
• Training Mode: individual or combined file training
• Data Splits: flexible train/test ratios for validation

Output

1. Performance Metrics: Mean Squared Error (MSE) & R² Score: Model Validation
2. Battery Health Indicators:
   • Remaining Useful Life (RUL) Predictions
   • State of Charge (SoC) Estimation Accuracy
   • State of Health (SoH) Degradation Tracking
3. Visualization Plots:
   • Predicted Vs True Capacity Scatter Plot (Confidence Intervals)
   • Degradation Curves Over Cycling History
   • Real-Time Inference Performance Metrics

Results Interpretation

• Feature Engineering Impact: toggle --include-cycle-num to compare models with and without cycle history
• Training Strategy Comparison: switch between individual and combined modes to evaluate different approaches
• Data Split Sensitivity: adjust --test-size to assess model robustness across different validation strategies
• Cross-Cell Generalization: use combined mode to evaluate performance across different battery chemistries
• Real-Time Performance: assess inference speed and accuracy for onboard integration requirements
• Degradation Modeling: analyze RUL prediction accuracy and SoH estimation reliability
• Temporal Analysis: understanding how model performance evolves with battery aging
• Hyperparameter Tuning: experiment with --num-estimators and --seed for optimization

Dependencies

• numpy: high-performance numerical computations for signal processing
• pandas: advanced data manipulation and time-series analysis
• matplotlib: professional visualization for battery performance analysis
• scikit-learn: machine learning algorithms optimized for regression tasks
• tqdm: progress monitoring for large-scale battery data processing

Future Enhancements

• Physics-Informed Neural Networks (PINNs): incorporating electrochemical constraints into deep learning models
• PyTorch Integration: advanced neural architectures for complex degradation pattern recognition
• Real-Time Optimization: edge computing deployment for ultra-low latency inference
• Multi-Modal Fusion: integration with thermal imaging and acoustic monitoring data
• Uncertainty Quantification: bayesian approaches for reliable confidence intervals in safety-critical applications
• Transfer Learning: domain adaptation across different battery chemistries and form factors
• Federated Learning: orivacy-preserving model updates across distributed battery fleets