Battery SOH Prediction — From Simulink Simulation to Real-World Transfer Learning

End-to-end machine learning and deep learning pipeline for State of Health (SOH) estimation of Li-ion battery packs. Models are trained on simulated data (Battery A) and progressively adapted to real-world conditions (Battery B) through a 7-step transfer learning pipeline.

Dataset

Download the dataset from Google Drive and place in 0_Dataset/ folder:

Download Dataset (Google Drive)

File	Size	Description
recording_senaryo_1_Battery_A_N100.mat	~157 MB	Training data (100 cycles, UDDS)
recording_senaryo_1_Battery_B_SOH88.mat	~86 MB	Phase 1 test (GT=88%, UDDS)
recording_senaryo_1_Batarya_B_faz2.mat	~183 MB	Phase 2 test (mixed drive cycle)

Battery Configuration

• Pack: 5s1p (5 series, 1 parallel) LG 18650HG2 NMC cells
• Simulation: MATLAB/Simulink with Simscape Battery
• Drive Cycle: UDDS (training), mixed UDDS/WLTP/US06 (testing)
• SOH Range: 100% to 80% (End of Life)

---

Pipeline Architecture (7 Steps)

Step 1: ML Model Training (Battery A)
  |
Step 2: DL Model Training (Battery A)
  |
Step 3: ML vs DL Comparison
  |
Step 4: Battery B Phase 1 Inference (same drive cycle)
  |
Step 5: Battery B Phase 2 Zero-Shot Transfer (different drive cycle)
  |
Step 6: Anchor Point Fine-Tuning (improve Phase 2 results)
  |
Step 7: Domain Adaptation (self-training + meta-learner)

---

Step 1: Classical ML Training (1-ML/)

Train 6 ML models on Battery A data (100 cycles, UDDS drive cycle).

Models

Model	Type	Best MAE (%)	R2
Extra Trees	Bagging	0.064	0.9998
XGBoost	Boosting	0.083	0.9996
Random Forest	Bagging	0.106	0.9994
Gradient Boosting	Boosting	0.156	0.9987
CatBoost	Boosting	0.170	0.9983
LightGBM	Boosting	0.696	0.9727

Pipeline (9 sub-steps per model)

1. Raw signal visualization (EDA)
2. mV to V conversion
3. Cycle boundary detection (SOC peaks)
4. Cycle-level feature extraction (~85 domain features)
5. Trend features (rolling mean/std, delta, rate of change -> ~205 features)
6. Correlation analysis
7. Feature selection (All ~200 features OR Top 30 by correlation)
8. Model training (Baseline) + overfitting check + 5-Fold CV
9. Hyperparameter tuning (Grid Search)

Key Finding

• Bagging > Boosting on small data (100 cycles)
• Top 30 features often outperform all features (less noise)
• Tree-based models cannot extrapolate beyond training range

---

Step 2: Deep Learning Training (2-DL/)

Train 5 DL models on Battery A data using sequences of cycles.

Models

Model	Type	Best MAE (%)	R2
GRU	RNN (2 gates)	0.249	0.9970
LSTM	RNN (3 gates)	0.410	0.9889
Transformer	Attention	0.441	0.9881
MLP	ANN (flat)	0.787	0.9759
1D-CNN	CNN	0.923	0.9563

Key Differences from ML Pipeline

• No trend features -- sequence models (LSTM, GRU) learn temporal patterns from last 5 cycles
• Feature scaling required (StandardScaler) -- neural networks need normalized inputs
• Sequence input (samples, 5, features) instead of flat (samples, features)
• Training via epochs with early stopping, dropout, gradient clipping

Key Finding

• GRU > LSTM on small data (fewer parameters = less overfitting)
• Feature engineering still matters -- Top 30 improved GRU by 16%
• Hyperparameter tuning is critical -- MLP Baseline R2 = -5.58 (catastrophic)

---

Step 3: ML vs DL Comparison (3-MLvsDL/)

Direct comparison of best 5 ML and best 5 DL models.

	Best ML	Best DL
Model	ExtraTrees Top30	GRU Top30 Tuned
MAE	0.064%	0.249%
Winner	ML (on this dataset)	--

ML wins on small data (100 cycles). DL's advantages emerge with larger datasets, transfer learning, and extrapolation.

---

Step 4: Battery B Phase 1 Inference (4-Inference_Battery_B_Phase_1/)

Apply trained models to Battery B (unseen battery, same UDDS drive cycle). No retraining -- models predict using saved weights.

Ground Truth: SOH = 88%

Rank	Model	SOH Pred	Error
1	DL_GRU_Tuned	87.93%	+0.07%
2	ML_ExtraTrees_Top30_TN	88.29%	-0.29%
3	ML_RF_Top30_TN	88.30%	-0.30%

Key Finding

• DL won on unseen battery! GRU generalized better than ML
• ML models were consistent (~88.3-88.6%)
• DL Top30 models collapsed (GRU_Top30_BL = 9.1%) -- scaler mismatch

---

Step 5: Battery B Phase 2 Zero-Shot Transfer (5-Inference_Battery_B_Phase_2/)

Apply same models to Battery B with different drive cycle (mixed UDDS/WLTP/US06 + random charge patterns). No adaptation.

Expected SOH: ~80% (no exact ground truth)

Rank	Model	SOH Pred	Error
1	DL_GRU_Top30_BL	80.4%	-0.4%
2	ML_XGBoost_Top30_BL	80.9%	-0.9%
3	ML_XGBoost_Top30_TN	81.2%	-1.2%
4-6	ML ExtraTrees, RF	86-87%	-6 to -7%
7-9	DL GRU_Top30_TN, LSTM	12-64%	Catastrophic

Key Finding

• Some models surprisingly robust -- GRU_Top30_BL and XGBoost zero-shot worked well
• Domain shift affects most models -- ExtraTrees dropped from 0.064% to 6% error
• Baseline often better than Tuned in cross-domain scenarios (less overfit to training domain)

---

Step 6: Anchor Point Fine-Tuning (6-Fine_Tuning_Battery_B_Phase_2/)

Improve Phase 2 results using "anchor points" -- full charge/discharge cycles that provide reliable reference data.

Method

1. Find anchors: Cycles where SOC_min <= 10% AND SOC_max >= 99.5%
2. Generate pseudo SOH labels: Fit quadratic polynomial on Battery A's proxy_IR vs SOH, apply to anchors
3. Incremental fine-tuning: As anchor points are encountered chronologically:
   • ML: Retrain with Battery A + accumulated anchors (5x weight)
   • DL: Fine-tune with low LR (5e-5), 50 epochs
4. Skip good models: GRU_Top30_BL, XGBoost_Top30_BL, XGBoost_Top30_TN already performing well -- fine-tuning could degrade them

Key Finding

• Fine-tuning improved most models but not all
• Skipping already-good models was critical -- fine-tuning GRU_Top30_BL degraded it from 80.4% to 91%
• Pseudo labels may be inaccurate, causing some models to shift in wrong direction

---

Step 7: Domain Adaptation (7-Domain_Adaptation_Phase_2/)

Most advanced adaptation using self-training loop + meta-learner calibration.

Method

1. Ensemble pseudo labels: All model predictions weighted (ML 2x, DL 1x) with confidence scores
2. Self-training (3 rounds):
   • ML retrain with Battery A + all pseudo labels (confidence-weighted)
   • DL fine-tune with confidence-weighted MSE loss
   • Update pseudo labels, check convergence
3. Meta-learner: Ridge Regression + Isotonic Regression for monotonic SOH calibration
4. Skip models in SKIP_FINETUNE set

---

Project Structure

calb_predict/
|-- 0_Dataset/                              # Raw .mat files
|   |-- recording_senaryo_1_Battery_A_N100.mat    # Training data
|   |-- recording_senaryo_1_Battery_B_SOH88.mat   # Phase 1 test (GT=88%)
|   |-- recording_senaryo_1_Batarya_B_faz2.mat    # Phase 2 test (GT~80%)
|
|-- 1-ML/                                  # Classical ML models
|   |-- Random_Forest/                     # RF (All + Top30)
|   |-- ExtraTrees/                        # ET (All + Top30)
|   |-- XGBoost/                           # XGB (All + Top30)
|   |-- GradientBoosting/                  # GB (All + Top30)
|   |-- LightGBM/                          # LGBM (All + Top30)
|   |-- Catboost/                          # CB (All + Top30)
|   |-- ML_Comparison.py                   # Compare all ML models
|   |-- ML_Comparison.csv / .png
|
|-- 2-DL/                                  # Deep Learning models
|   |-- LSTM/                              # BiLSTM (All + Top30)
|   |-- GRU/                               # BiGRU (All + Top30)
|   |-- CNN/                               # 1D-CNN (All + Top30)
|   |-- MLP/                               # MLP (All + Top30)
|   |-- Transformer/                       # Encoder-only (All + Top30)
|   |-- DL_Comparison.py                   # Compare all DL models
|   |-- DL_Comparison.csv / .png
|
|-- 3-MLvsDL/                              # ML vs DL head-to-head
|   |-- ML_vs_DL_Comparison.py
|
|-- 4-Inference_Battery_B_Phase_1/         # Same drive cycle inference
|   |-- Battery_B_Inference.py
|   |-- Output_Inference/
|
|-- 5-Inference_Battery_B_Phase_2/         # Zero-shot transfer
|   |-- Battery_B_Phase2_Inference.py
|   |-- Output_Inference_Phase2/
|
|-- 6-Fine_Tuning_Battery_B_Phase_2/       # Anchor point fine-tuning
|   |-- Battery_B_Phase2_FineTuning.py
|   |-- Output_FineTuning/
|
|-- 7-Domain_Adaptation_Phase_2/           # Self-training + meta-learner
|   |-- Battery_B_Phase2_DomainAdaptation.py
|   |-- Output_DomainAdaptation/
|
|-- kitap/                                 # Reference materials & guides
|   |-- ML_DL_Pratik_Rehber.pdf            # 28-page ML/DL learning guide
|
|-- README.md                              # This file
|-- requirements.txt

---

Key Metrics

Metric	Description
MAE	Mean Absolute Error (%) -- average prediction error
RMSE	Root Mean Squared Error (%) -- penalizes large errors
R2	R-Squared -- fraction of variance explained (1.0 = perfect)
CV MAE	5-Fold Cross Validation MAE -- more reliable than single split

---

Results Summary

Training Performance (Battery A)

Rank	Model	Type	MAE (%)	R2
1	ExtraTrees Top30 BL	ML	0.064	0.9998
2	XGBoost Top30 BL	ML	0.083	0.9996
3	GRU Top30 Tuned	DL	0.249	0.9970
4	GRU Tuned	DL	0.295	0.9954
5	LSTM Top30 Tuned	DL	0.410	0.9889

Battery B Phase 1 Inference (same drive cycle, GT=88%)

Rank	Model	SOH Pred	Error
1	DL_GRU_Tuned	87.93%	+0.07%
2	ML_ExtraTrees_Top30_TN	88.29%	-0.29%

Battery B Phase 2 Zero-Shot (different drive cycle, expected ~80%)

Rank	Model	SOH Pred	Error
1	DL_GRU_Top30_BL	80.4%	-0.4%
2	ML_XGBoost_Top30_BL	80.9%	-0.9%

---

Key Learnings

1. ML beats DL on small data (100 cycles) -- tree models need less data
2. DL generalizes better to unseen batteries -- GRU outperformed all ML on Battery B Phase 1
3. Feature engineering matters even for DL -- Top 30 selection improved results with limited data
4. Baseline can beat Tuned in cross-domain -- simpler models are more robust to domain shift
5. Fine-tuning doesn't always help -- already-good models can be degraded by incorrect pseudo labels
6. Tree models cannot extrapolate -- predictions bounded by training range
7. Bagging > Boosting on small data -- Extra Trees' randomness reduces overfitting

---

Requirements

pip install numpy pandas scikit-learn xgboost lightgbm catboost h5py matplotlib joblib torch fpdf2

---

How to Run

Train all models

# ML (12 pipelines)
cd 1-ML && for d in */; do cd "$d" && python *.py; cd ..; done

# DL (10 pipelines)
cd 2-DL && for d in */; do cd "$d" && python *.py; cd ..; done

# Compare
python 1-ML/ML_Comparison.py
python 2-DL/DL_Comparison.py
python 3-MLvsDL/ML_vs_DL_Comparison.py

Run inference pipeline

python 4-Inference_Battery_B_Phase_1/Battery_B_Inference.py
python 5-Inference_Battery_B_Phase_2/Battery_B_Phase2_Inference.py
python 6-Fine_Tuning_Battery_B_Phase_2/Battery_B_Phase2_FineTuning.py
python 7-Domain_Adaptation_Phase_2/Battery_B_Phase2_DomainAdaptation.py