chiehsheshuai/rms-blind-spot-analysis

Beyond RMS: Power Quality Blind Spot Analysis

Feature engineering and machine learning framework for analyzing distorted electrical waveforms beyond conventional RMS measurements.

---

Overview

Electrical protection and monitoring systems commonly rely on RMS measurements to evaluate operating conditions. However, waveforms with similar RMS values may exhibit significantly different electrical stress due to harmonic distortion, impulsive behavior and waveform deformation.

This project investigates those blind spots using two real-world open datasets.

The workflow combines

• waveform analysis
• harmonic analysis
• crest factor analysis
• THD estimation
• machine learning
• feature importance analysis

to demonstrate why RMS alone is often insufficient for modern power-quality assessment.

---

Features

• High-resolution waveform analysis
• RMS / Peak / Crest Factor comparison
• THD computation
• FFT analysis
• Harmonic feature extraction
• High Impedance Fault (HIF) analysis
• Random Forest classification
• XGBoost classification
• Feature importance visualization
• Interactive Gradio demo

---

Repository Structure

Beyond-RMS/
│
├── app.py
├── README.md
├── requirements.txt
│
├── src/
│   ├── dataset_loader.py
│   ├── feature_engineering.py
│   ├── visualization.py
│   ├── models.py
│   └── inference.py
│
├── notebook/
│
├── datasets/
│
├── models/
│
├── figures/
│
└── assets/

---

Datasets

Dataset 1 — BidyutPi

Real instantaneous voltage and current measurements (10 kHz)

Devices

• 330 W Resistive Load
• 0.5 HP Induction Motor
• Ceiling Fan
• SMPS Mobile Charger

Used for

• RMS
• Peak
• Crest Factor
• THD
• FFT
• Moving RMS

---

Dataset 2 — HIF Power Quality Dataset

Real measurements obtained from controlled High-Impedance Fault experiments on medium-voltage covered conductors.

Contains

• 142 experimental recordings
• Harmonics
• Interharmonics
• RMS
• Power Factor
• Voltage
• Current
• Active / Reactive Power

Used for

• Harmonic feature extraction
• Feature importance
• Machine learning
• Blind-spot analysis

---

Machine Learning

Models

• Random Forest
• XGBoost

Performance comparison

Model	Accuracy
RMS-only Model	95.34 %
Harmonic Feature Model	98.01 %

The harmonic feature model improved the overall classification accuracy while providing considerably better macro F1 performance.

---

Research Question

Can two electrical waveforms having similar RMS values represent different electrical stress?

This repository demonstrates that harmonic information and waveform descriptors provide complementary information beyond RMS measurements.

---

References

1. BidyutPi Dataset
2. High Impedance Fault Power Quality Dataset
3. IEC 61000-4-7
4. IEC 61000-4-30
5. IEEE papers on High-Impedance Fault detection
6. IEEE papers on Power Quality

---

License

MIT License