🔋 Data-Driven Battery Diagnostics — NASA PCOE Dataset

End-to-end ML pipeline for SoC, SoH, and RUL prediction of lithium-ion batteries.

---

🎯 What This Project Does

Battery Management Systems (BMS) need answers to three critical questions:

#	Question	Technical Term	Our Model
1	"How much charge is left right now?"	State of Charge (SoC)	LSTM Neural Network
2	"How degraded is this battery?"	State of Health (SoH)	XGBoost (Gradient Boosting)
3	"How many cycles before it dies?"	Remaining Useful Life (RUL)	LSTM Neural Network

This pipeline takes raw NASA battery cycling data → parses .mat files → engineers physics-informed features → trains 3 models → outputs evaluation metrics + publication-ready plots.

---

📈 Results

Performance Metrics

Model	RMSE	MAE	R² Score	Interpretation
SoC (LSTM)	7.24%	2.82%	0.9499	Excellent — explains 95% of variance
SoH (XGBoost)	3.09%	2.76%	0.8260	Good — captures degradation trend
RUL (LSTM)	12.49 cycles	9.21 cycles	0.8699	Good — average error ~9 cycles

Capacity Fade Analysis

All 4 batteries show clear degradation over cycling. B0018 (4A discharge) degrades faster than B0005–B0007 (2A), consistent with Arrhenius kinetics.

Model A — SoC Estimation (LSTM)

The LSTM captures the full discharge curve trajectory. Predictions closely track the actual SoC with R² = 0.95.

Model B — SoH Estimation (XGBoost)

XGBoost tracks the overall degradation trend from 93% → 65% SoH. Feature importance reveals temperature variance (57%) and discharge time (29%) are the dominant aging indicators.

Model C — RUL Prediction (LSTM)

The LSTM learns the countdown pattern (120 → 0 cycles remaining). Error distribution is centered near zero, with most predictions within ±10 cycles.

Training Convergence

Both LSTM models converge smoothly with early stopping preventing overfitting.

---

📊 Dataset

NASA Prognostics Center of Excellence (PCoE) Battery Dataset — 4 × 18650 Li-ion cells cycled to failure.

Battery	Discharge Rate	~Cycles	EOL Reached?
B0005	2A	168	✅
B0006	2A	168	✅
B0007	2A	168	✅
B0018	4A	132	✅

• Charge: CC-CV at 1.5A → 4.2V cutoff
• Discharge: Constant current until 2.7V
• EOL threshold: 70% of rated capacity (1.4 Ah out of 2.0 Ah)
• Source: NASA PCoE Repository

---

🧠 Technical Approach

Pipeline Flow

NASA .mat files
    ↓  scipy.io.loadmat
Parse MATLAB structs → Extract discharge cycles only
    ↓  Current integration: Q(t) = ∫|I(τ)|dτ
Clean DataFrame (remove NaN/Inf)
    ↓
    ├── SoC Path: MinMax normalize → sliding window (50 steps) → LSTM
    │
    └── Cycle Path: Aggregate 8 features per cycle
            ├── SoH: XGBoost on tabular features
            └── RUL: Lookback window (10 cycles) → LSTM

Feature Engineering (8 Physics-Informed Features)

Feature	Formula	Physical Basis
max_capacity	NASA reported (Ah)	Direct health indicator
discharge_time	t_end − t_start	Shorter = degraded
avg_voltage	mean(V)	Drops with SEI growth
voltage_drop	V_max − V_min	Internal resistance proxy
avg_current	mean(|I|)	Operating condition
avg_temperature	mean(T)	Arrhenius aging factor
temp_variance	var(T)	Thermal instability
internal_resistance	ΔV / I_avg	Primary aging indicator

Model Architectures

SoC (LSTM)	SoH (XGBoost)	RUL (LSTM)
Input: 3 features × 50 steps LSTM(64) → Drop(0.2) LSTM(32) → Drop(0.2) Dense(1)	Input: 8 cycle features 200 trees, depth=5 lr=0.05, L1+L2 reg 5-fold CV	Input: 8 features × 10 cycles LSTM(128) → Drop(0.3) LSTM(64) → Drop(0.3) Dense(32, ReLU) → Dense(1)
Why LSTM? SoC depends on temporal voltage/current history — same voltage can mean different SoC depending on trajectory.	Why XGBoost? Only ~630 cycles of tabular data — trees outperform neural nets on small structured datasets.	Why LSTM? RUL requires detecting acceleration in degradation — the model needs to see trends over past cycles.

Key Design Decisions

• Chronological split (80/20) — prevents future data leakage (critical for time-series)
• MinMax scaling — bounded [0,1] range works better with LSTM's sigmoid/tanh activations
• Early stopping — prevents overfitting by monitoring validation loss
• max_capacity excluded from SoH features — avoids circular prediction (SoH is derived from capacity)

---

📁 Project Structure

battery-diagnostics-NASA/
├── data/
│   └── download_data.py         # Auto-downloads from NASA S3
├── src/
│   ├── data_loader.py           # .mat parsing + current integration
│   ├── preprocessing.py         # Clean, normalize, sliding windows
│   ├── feature_engineering.py   # 8 cycle features + SoH/RUL labels
│   ├── evaluation.py            # RMSE, MAE, R² metrics
│   ├── visualization.py         # 6 publication-ready plots
│   └── models/
│       ├── soc_model.py         # LSTM for SoC
│       ├── soh_model.py         # XGBoost for SoH
│       └── rul_model.py         # LSTM for RUL
├── plots/                       # Generated diagnostic plots
├── results/                     # Saved predictions & metrics (CSV/JSON)
├── main.py                      # ← Run this! End-to-end pipeline
└── requirements.txt

---

🚀 Quick Start

# Clone
git clone https://github.com/ARYANRAJ1121/battery-diagnostics-NASA.git
cd battery-diagnostics-NASA

# Setup
python -m venv venv && venv\Scripts\activate   # Windows
pip install -r requirements.txt

# Run entire pipeline (~5 min)
python main.py

Output: Metrics printed to console + 6 plots in plots/ + 7 result files in results/

---

🔬 Research Context

This project bridges data-driven ML and electrochemical domain knowledge:

Degradation Mechanism	Feature That Captures It
SEI layer growth	internal_resistance ↑, avg_voltage ↓
Lithium inventory loss	max_capacity ↓, discharge_time ↓
Thermal aging (Arrhenius)	avg_temperature ↑, temp_variance ↑
Electrode cracking	voltage_drop ↑

Relevant to: Battery Management Systems (BMS), Prognostics & Health Management (PHM), Electric Vehicle reliability, Grid-scale energy storage

Future Work

• Transformer-based models (PatchTST) for longer-range dependencies
• Physics-Informed Neural Networks (PINNs) combining data + electrochemical equations
• Transfer learning across battery chemistries
• Uncertainty quantification (confidence intervals on predictions)

---

📚 References

1. B. Saha, K. Goebel (2007). "Battery Data Set", NASA Ames Prognostics Data Repository
2. K. Severson et al. (2019). "Data-driven prediction of battery cycle life before capacity degradation", Nature Energy
3. M. Berecibar et al. (2016). "Critical review of state of health estimation methods", Renewable & Sustainable Energy Reviews
4. Y. Zhang et al. (2018). "LSTM RNN for remaining useful life prediction", Engineering Applications of AI
5. T. Chen, C. Guestrin (2016). "XGBoost: A Scalable Tree Boosting System", KDD

---

📄 License

MIT License — see LICENSE

---

Built for research internship application — IIT Jammu
Demonstrating proficiency in battery diagnostics, deep learning, and reproducible ML research