Nuclear Power Plant Fault Diagnosis Using Deep Learning

Exploring ML methods for nuclear power plant fault diagnosis and root cause analysis using open simulation data from the NPPAD dataset (Nuclear Power Plant Accident Data).

Overview

This project applies deep learning to time-series sensor data from pressurized water reactor (PWR) simulations to:

1. Anomaly Detection — Monitor reactor sensor streams and flag deviations from normal operating conditions using autoencoders, LSTMs, and transformers
2. Accident Classification & Root Cause Analysis — Classify flagged anomalies into accident types with interpretability via SHAP/attention visualization
3. Physics-Informed Digital Twin — Surrogate model approximating reactor dynamics with physics-informed loss constraints

Dataset

The NPPAD dataset contains time-series data from PCTRAN, a widely-used NPP simulator, covering:

• 96 operational parameters (temperature, pressure, neutron density, coolant flow, etc.)
• 18 operating conditions (1 normal + 17 accident scenarios)
• Multiple severity levels per accident type

Citation:

Qi, B., Xiao, X., Liang, J. et al. An open time-series simulated dataset covering various accidents for nuclear power plants. Sci Data 9, 766 (2022).

Project Structure

├── configs/            # YAML experiment configurations
├── data/
│   ├── raw/            # Raw NPPAD data (gitignored)
│   ├── processed/      # Preprocessed tensors (gitignored)
│   └── scripts/        # Download and preprocessing scripts
├── src/
│   ├── data/           # PyTorch datasets and transforms
│   ├── models/
│   │   ├── anomaly/    # Autoencoder, LSTM, Transformer detectors
│   │   ├── diagnosis/  # Accident classifier + SHAP (Module 2)
│   │   └── digital_twin/  # Physics-informed surrogate (Module 3)
│   ├── training/       # Training loop, callbacks
│   ├── evaluation/     # Metrics, uncertainty estimation
│   └── utils/          # Config system, logging
├── notebooks/          # EDA, training, evaluation notebooks
├── scripts/            # CLI entry points
└── tests/

Setup

pip install -r requirements.txt
pip install -e .

Usage

1. Download Data

python data/scripts/download.py

2. Preprocess

python data/scripts/preprocess.py

3. Train Anomaly Detectors

python scripts/train.py --config configs/autoencoder.yaml
python scripts/train.py --config configs/lstm.yaml
python scripts/train.py --config configs/transformer.yaml

4. Train Accident Classifiers

python scripts/train_classifier.py --config configs/cnn_classifier.yaml
python scripts/train_classifier.py --config configs/lstm_classifier.yaml
python scripts/train_classifier.py --config configs/transformer_classifier.yaml

5. Train Digital Twin

python scripts/train_digital_twin.py --config configs/digital_twin.yaml
python scripts/train_digital_twin.py --config configs/digital_twin_no_physics.yaml  # ablation baseline

Results

1: Anomaly Detection

Semi-supervised approach — models trained on normal operating data only (302 timesteps), evaluated on 7,316 test windows across 18 accident types. Anomalies are detected via reconstruction/prediction error exceeding a threshold optimized on the validation set.

Model	Precision	Recall	F1	AUROC	AUPRC
Autoencoder	0.999	1.000	0.9995	0.985	0.9999
LSTM	0.999	1.000	0.9995	0.987	0.9999
Transformer	0.999	1.000	0.9995	0.985	0.9999

All three architectures achieve near-perfect binary anomaly detection. The strong separation between normal and accident regimes in the NPPAD data makes coarse anomaly detection straightforward — the more challenging task is Module 2's multi-class accident classification and root cause attribution.

Training details:

• Window size: 30 timesteps, stride: 10
• Trained on 28 normal windows (semi-supervised)
• Early stopping with patience=10 on validation loss
• Hardware: Apple M-series GPU (MPS backend)

2: Accident Classification & Root Cause Analysis

Supervised multi-class classification across 18 accident types (LOCA, SGTR, steam line break, rod withdrawal, etc.). Trained on all 32,284 windows, evaluated on 7,316 test windows.

Model	Accuracy	F1 (macro)	F1 (weighted)	Precision (macro)	Recall (macro)
CNN	0.524	0.421	0.443	0.448	0.462
LSTM (bidirectional)	0.712	0.572	0.676	0.661	0.578
Transformer (CLS token)	0.707	0.538	0.662	0.538	0.582

The LSTM classifier achieves the best overall performance. Per-class analysis reveals:

• Near-perfect (F1 > 0.95): LLB, LR, MD, SLBOC, Normal
• Strong (F1 > 0.7): SLBIC, SGBTR
• Challenging (F1 < 0.5): LOCA/LOCAC confusion, SGATR, RW, low-sample classes (LACP, ATWS, SP)

The lower macro-F1 reflects class imbalance — some accident types have as few as 28-56 training windows vs 3,000+ for others. Root cause analysis via gradient attribution identifies the top contributing sensors per accident type (see notebooks/04_evaluate_classification.ipynb).

Training details:

• Supervised on all data (accident_type labels)
• 1D-CNN, bidirectional LSTM, and Transformer with [CLS] token
• Early stopping with patience=15

3: Physics-Informed Digital Twin

GRU encoder-decoder surrogate model that predicts future reactor states from a context window, constrained by reactor physics. Compares a physics-informed model (point kinetics + energy conservation constraints) against a data-only baseline. Trained on 32,284 windows, evaluated on 7,316 test windows.

Architecture: GRU encoder (context window) → GRU decoder (autoregressive) with residual connection. 384K parameters.

Metric	Physics-Informed	Data-Only Baseline
Physics Residual (point kinetics)	3,739	11,888
Conservation Violation (MW)	1,783	2,950
Physics Residual Reduction	68.6%	—
Conservation Violation Reduction	39.6%	—

The physics-informed model produces predictions significantly more consistent with point reactor kinetics equations and energy conservation laws, while maintaining comparable data-fitting accuracy. The key insight is that physics constraints act as a regularizer — they don't improve raw MSE much but ensure predictions respect physical laws (positive pressure/power, energy balance, neutron kinetics consistency).

Physics constraints:

1. Point kinetics residual — dn/dt consistency with simplified neutron kinetics ODE
2. Energy conservation — Q ≈ W · Cp · (T_hot - T_cold) energy balance
3. Physical bounds — soft penalties for negative pressure, power, temperature, flow, boron
4. Warmup schedule — physics terms ramp from 0 to full weight over 500 training steps

Training details:

• GRU encoder-decoder with residual connection (predictions = last_state + delta)
• Context: 20 timesteps → Predict: 10 timesteps ahead
• Learning rate: 5e-4, 150 epochs, patience: 15
• Hardware: Apple M-series GPU (MPS backend)

Tech Stack

Python, PyTorch, pandas, scikit-learn, matplotlib

License

MIT