ABB-EngineeredX-2.0

Transformer Digital Twin

Submission for ABB EngineeredX 2.0 · Problem Statement 1: Digital Twin of Transformer / Electrical Equipment

A working proof-of-concept digital twin that combines physics-based modeling with machine learning to monitor power transformer health, detect anomalies, classify dissolved-gas faults, and forecast remaining useful life.

---

🎯 The core idea — the hybrid loop

                ┌────────────────────────────────┐
                │     Digital twin core           │
                │                                 │
   Physical ──> │   Physics model ───┐            │
   sensors      │                    ├── ML layer │ ──> Operator
                │   (IEEE C57.91,    │            │     dashboard
                │   Arrhenius)       └── residual │
                │                                 │
                └────────────────────────────────┘

The physics model predicts what the transformer should be doing under current loading and ambient conditions. The ML layer interprets the residual — the gap between expected and observed — to flag faults and forecast life.

• Pure physics misses unmodeled degradation modes.
• Pure ML demands prohibitive amounts of training data.
• The combination is how industrial-grade digital twins actually work.

---

📁 Repository structure

transformer-digital-twin/
├── README.md                       ← you are here
├── requirements.txt
├── LICENSE
├── docs/
│   └── ABB_EngineeredX_PS1_DigitalTwin.pdf   ← the submission document
├── data/
│   ├── synthetic_transformer_telemetry.csv   ← 180 days, 21 channels, 4 320 rows
│   └── dga_training_set.csv                  ← 2 000 labeled DGA samples
├── src/
│   ├── data_generator.py           ← synthetic telemetry & DGA samples
│   ├── physics.py                  ← IEEE C57.91, Arrhenius, equivalent circuit
│   ├── ml_models.py                ← AnomalyDetector, DGAFaultClassifier, RULForecaster
│   └── visualization.py            ← reusable plotting helpers
├── notebooks/
│   ├── 01_data_generation.ipynb    ← walk through the synthetic data
│   ├── 02_physics_model.ipynb      ← thermal + aging models with validation
│   ├── 03_anomaly_detection.ipynb  ← Isolation Forest + hybrid-loop ablation
│   ├── 04_dga_fault_classifier.ipynb  ← XGBoost on Duval Triangle features
│   └── 05_rul_forecasting.ipynb    ← Remaining useful life predictions
└── dashboard/
    └── app.py                      ← Streamlit operator dashboard

---

🚀 Quickstart

# 1. Clone the repo
git clone https://github.com/<YOUR_USERNAME>/transformer-digital-twin.git
cd transformer-digital-twin

# 2. Install dependencies (Python 3.10+ recommended)
pip install -r requirements.txt

# 3. Walk through the notebooks in order
jupyter notebook notebooks/

# 4. Launch the operator dashboard
streamlit run dashboard/app.py

The dashboard opens in your browser at http://localhost:8501.

---

🔬 What's implemented

Physics layer (src/physics.py)

Component	Standard	Inputs	Outputs
ThermalModel	IEEE C57.91	load, ambient temp, cooling class	predicted hotspot, residual
AgingModel	Arrhenius / Montsinger	hotspot history	life consumed, F_AA
EquivalentCircuit	Standard T-model	V/I, impedance	losses, what-if simulations

Verified against textbook values: F_AA at 117°C equals 2.02 (Montsinger's rule says ageing should double every 7°C above the 110°C reference — ✓).

ML layer (src/ml_models.py)

Model	Algorithm	Input	Output
AnomalyDetector	Isolation Forest (200 trees) on sensors + physics residuals	11-dim vector	anomaly score ∈ [0,1]
DGAFaultClassifier	XGBoost (6-class) on gas + Duval Triangle features	13 features	fault type + confidence
RULForecaster	MLP regression on rolling window statistics	18-feature window	days to maintenance

Production note: The submission PDF specifies an LSTM with attention for RUL forecasting. PyTorch didn't fit in the build environment used to assemble this repo, so the demo uses scikit-learn's MLPRegressor on engineered window features. The pipeline and interface are identical; swap MLPRegressor for an LSTM in production:

# Demo (this repo):
self.model = MLPRegressor(hidden_layer_sizes=(64, 32), ...)

# Production:
self.model = LSTMWithAttention(input_dim=18, hidden_dim=128, n_heads=4)

Operator dashboard (dashboard/app.py)

Streamlit app with four views:

1. Operator dashboard — composite health gauge, KPI tiles, RUL trajectory, alerts
2. Fault diagnostics — interactive DGA fault classification with confidence
3. What-if simulator — overload scenario projection (hotspot + life consumed)
4. About — architecture and standards summary

---

📊 Validation results

Run notebook 03 to reproduce. Summary metrics on synthetic data:

Metric	Value
Anomaly detector score during faults	0.63 (vs 0.20 normal)
Anomaly detector precision @ 0.5 threshold	~0.85
DGA classifier test accuracy	1.00 (synthetic) / typical 0.85–0.92 on real data
Thermal model fit RMSE (after recalibration)	< 1.5°C

Caveat: Test accuracy of 1.00 reflects well-separated synthetic data, not realistic real-world performance. On real DGA fault libraries (e.g. the IEC TC10 database), gradient-boosted classifiers typically achieve 0.85–0.92 accuracy.

---

🏗️ How this extends to production

This is a proof-of-concept. The path to a real industrial deployment looks like:

Stage	Demo (this repo)	Production
Data ingestion	CSV file read	MQTT/OPC-UA streams via edge gateway, InfluxDB sink
Physics parameters	textbook defaults	calibrated against unit's heat-run test report
Training data	synthetic, 180 days	real fleet history, multi-year, with maintenance labels
RUL model	MLPRegressor	LSTM with attention or Temporal Fusion Transformer
Deployment	local Streamlit	Docker + Kubernetes on Azure Digital Twins (ABB Ability™ stack)
Anomaly detection	batch-trained	online, edge-deployed, with drift detection

The architecture also generalizes beyond transformers — swap the physics module to model a generator, exciter, or large motor. The data pipeline, ML layer, and operator dashboard are asset-agnostic.

---

📚 Standards referenced

• IEEE C57.91 — Guide for Loading Mineral-Oil-Immersed Transformers
• IEC 60599 — Mineral oil-impregnated electrical equipment in service — Guidance on the interpretation of DGA
• Duval Triangle Method — Graphical fault classification using normalized %CH₄ / %C₂H₄ / %C₂H₂
• Arrhenius / Montsinger's Rule — Paper insulation aging acceleration

---

⚠️ Limitations & honest disclosures

• Data is synthetic. The generator mimics realistic relationships but is not validated against real transformer telemetry.
• Physics models are simplified. Real deployment would use finite-element thermal models and full IEEE C57.91 Annex G load-cycle integration.
• RUL ground truth is synthetic. Real RUL labels require time-to-failure data from a fleet of similar assets — not available for a hackathon submission.
• The DGA classifier is trained on synthetic class profiles. Real training needs the IEC TC10 fault library or equivalent.

These limitations are inherent to a proof-of-concept assembled in 24 hours. The architecture and implementation pipeline are correct; calibration against real data is the production extension.

---

👤 Submission details

Candidate: Gagandeep Singh Choudhary

Institution: JECRC College

Batch: B.Tech CSE · 2027

Problem statement: Digital twin of transformer or any other electrical equipment

For the full design document, see docs/ABB_EngineeredX_PS1_DigitalTwin.pdf.

---

Submitted to ABB EngineeredX 2.0 · May 2026