Marcos Medina
Leer en español
All work

04/04 Master’s thesis × Research

Partial Discharge Localization in 3D

Locating electrical faults inside high-voltage equipment in three dimensions, with convolutional neural networks and optimization methods.

Role
Master’s thesis (TFM)
Context
MSc in Telecommunication Engineering, UPV/EHU
Focus
EM simulation, deep learning, 3D localization
  • 3D Source localization inside a volume
  • CST Electromagnetic simulation
  • CNN Combined with optimization methods

Interactive demo — Illustrative simulation — not experimental results

Overview

Partial discharges are small electrical breakdowns inside the insulation of high-voltage equipment. They do not short-circuit the system, but they degrade it, and they are among the earliest warnings of a future failure. Knowing where they happen is key to maintenance.

Each discharge radiates an electromagnetic pulse. Captured by several sensors, the differences between the received signals encode where the source is. This thesis explores how to recover that position in three dimensions using convolutional neural networks and optimization methods.

The challenge

Real environments are hostile to clean measurements. Metallic enclosures reflect the pulse again and again (multipath), and every measurement carries noise. Both degrade localization.

Collecting labelled measurements for every possible source position is impractical, so the problem is approached through electromagnetic simulation first.

Architecture

CST Studio EM propagation Sensor signals Multichannel data Noise & multipath Robustness tests CNN Deep learning Optimization Position estimation (x, y, z) 3D source position CST Studio EM propagation Sensor signals Multichannel data Noise & multipath Robustness tests CNN Deep learning Optimization Position estimation (x, y, z) 3D source position
Hover or focus a component to trace its connections.

Key decisions

  1. 01

    Simulation as the data source

    Electromagnetic propagation is simulated with CST Studio, producing the signals that sensors would observe for discharges at known positions.

  2. 02

    Convolutional networks for localization

    Convolutional neural networks learn the relationship between the signals received by the sensors and the three-dimensional position of the source.

  3. 03

    Learning and optimization together

    The work combines neural networks with optimization methods to estimate the position of the discharge in three dimensions.

  4. 04

    Stress-testing robustness

    The approach is studied under noise and multipath, the two conditions that most degrade localization in practice.

Stack

  • CST Studio Suite
  • Convolutional neural networks
  • Optimization methods
  • Signal processing

Notes

The 3D visualisation on this site is an illustrative simulation built for the portfolio. It does not reproduce experimental results or metrics from the thesis.