SecureRelayFL

Privacy-Preserving Adaptive Relay Settings via Federated Learning Over Industrial Communication Networks

Paper submitted to MAIN 2026 (Mediterranean Artificial Intelligence and Networking Conference), Palermo, Italy, July 1–3, 2026. IEEE ComSoc co-sponsored, IEEE Xplore indexed.

Overview

Industrial facilities operating heterogeneous electrical protection systems face a data silo problem: optimizing relay settings requires fault data that operators refuse to share. This is especially common for industrial parks, where there are a myriad of large facilities operating at maximum load. This work applies federated learning to train collaborative protection models without centralizing proprietary data, and systematically evaluates how realistic network impairments (latency, packet loss, bandwidth) and differential privacy affect protection accuracy.

The key insight is a dual communication layer problem: FL aggregation traffic shares the same industrial Ethernet carrying time-critical IEC 61850 GOOSE protection messages (sub-4 ms requirement). We analyze how these two communication requirements interact.

Quick Start

Prerequisites

• Python 3.10–3.12
• (Optional) CUDA 12.x for GPU training
• ~500 MB disk for generated data

Setup

# clone the repo
git clone https://github.com/sramharack/SecureRelayFL.git
cd SecureRelayFL

# create environment and install (pick one)
make env              # loose pins — latest compatible versions
make env-exact        # exact pins — bit-for-bit reproducibility
make env-dev          # loose pins + dev tools (pytest, ruff, jupyter)

# activate
source .venv/bin/activate

Or manually without Make

python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip setuptools wheel

# option A: exact reproducibility
pip install -r requirements.lock

# option B: latest compatible
pip install -r requirements.txt

# install project as editable package
pip install -e .

Generate Data

make data                          # default: 1000 samples/facility, seed=42
# or
python data/generator/generate.py --n-samples 1000 --seed 42

This creates data/generated/ with 5,000 synthetic fault waveforms (6 channels × 2,560 timesteps each) across 5 industrial facility profiles.

Run Tests

make test
# or
pytest tests/ -v

Project Structure

SecureRelayFL/
├── data/
│   └── generator/
│       └── generate.py          # Synthetic fault waveform generator
├── models/                      # 1D-CNN + MLP baseline (TBD)
├── fl/                          # Flower FL server + clients (TBD)
├── experiments/                 # Experiment scripts per axis (TBD)
├── analysis/                    # Publication figure generation (TBD)
├── configs/                     # Experiment YAML configs (TBD)
├── results/                     # Saved metrics & checkpoints (gitignored)
├── tests/                       # Pytest smoke tests
├── pyproject.toml               # Project metadata & deps
├── requirements.txt             # Loose-pin dependencies
├── requirements.lock            # Exact-pin dependencies
├── Makefile                     # One-command workflows
└── README.md

Data Generation

The waveform generator produces physics-based 3-phase voltage and current signals using analytical electromagnetic transient equations. No external simulation tool is required.

Facility Profiles (5 FL Clients)

Facility	Voltage	Fault MVA	X/R	Grounding	SNR	Key Protection
Data Center	13.8 kV	250	8	Solidly grounded	40 dB	ZSI (50/51)
Steel Plant	34.5 kV	500	15	Low-R grounded	25 dB	Distance (21) + 87B
Petrochemical	13.8 kV	150	6	High-R grounded	35 dB	ZSI (50/51)
Pharmaceutical	4.16 kV	100	5	Resistance grounded	38 dB	ZSI (50/51)
Cement Plant	34.5 kV	400	12	Low-R grounded	28 dB	ZSI (50/51)

Waveform Specifications

• Frequency: 60 Hz | Sample rate: 15,360 Hz | Channels: 6 (Va, Vb, Vc, Ia, Ib, Ic)
• Window: 10 cycles (3 pre-fault + 7 post-fault) = 2,560 samples
• Fault types: No-fault, SLG, LL, high-impedance arcing
• Labels: Fault type (4), fault zone (4), protection action (5)

Physics References

• Asymmetrical fault current: IEEE 551-2006, IEC 60909
• Grounding-dependent voltage sag/swell: IEEE C62.92
• High-impedance arcing faults: IEEE PSRC WG D15

Experiments

Axes (priority order)

1. Network impairment analysis — latency, packet loss, bandwidth
2. Privacy–accuracy trade-off — differential privacy (ε sweep)
3. Communication efficiency — gradient compression, GOOSE co-existence
4. FL strategy comparison — FedAvg vs. FedProx vs. SCAFFOLD

Baselines

• Centralized (pooled data) — upper bound
• Local-only (no collaboration) — lower bound
• Federated (ideal network) — isolates impairment effects

License

MIT

Citation

@inproceedings{securerelay2026,
  title     = {Privacy-Preserving Adaptive Relay Settings via Federated Learning Over Industrial Communication Networks},
  author    = {Shankar Ramharack},
  booktitle = {Proc. Mediterranean Artificial Intelligence and Networking
               Conference (MAIN)},
  year      = {2026},
  address   = {Palermo, Italy},
}