Autoencoders Beat PCA for Low-Dimension DGA-based Fault Diagnosis of Power Transformers

Thales Wulfert Cabral, Eduardo de Lima, José Cândido Silveira Santos Filho, Luís Geraldo P. Meloni

DOI: 10.14209/sbrt.2024.1571036788

Evento: XLII Simpósio Brasileiro de Telecomunicações e Processamento de Sinais (SBrT2024)

Keywords: Machine learning Autoencoders Dimensionality reduction DGA

Abstract

Energy utility companies are investing in advanced monitoring systems using efficient data processing methodologies to mitigate the impacts of power transformer malfunctions on supply stability. Reducing data processing volume is crucial for achieving efficiency. In this context, our contributions include (i) proposing a fault diagnosis system that maintains high performance even under severe dimensionality reduction, (ii) introducing two Autoencoder structures, (iii) conducting pioneering tests of the Adafactor optimizer in dissolved gas analysis using Autoencoders, and (iv) comparing our solution with Principal Component Analysis (PCA), one of the most well-established techniques in the literature. Results confirm that our proposed system outperforms PCA, particularly in scenarios requiring severe dimensionality reduction.

Download