Résumé:
Large companies need to integrate a full fault detection system into their industry, especially electric
companies. Consequently, automatic fault detection has become of utmost importance to assess the
condition of electrical components and has been a long-standing challenge. Partial discharge is a
common indication of faults in power systems, such as generators and cables. These PDs can
eventually result in costly repairs and substantial power outages. The main goal of this work consists
of devising a monitoring system capable of detecting partial discharge patterns in signals acquired
from power lines. To accomplish this task, we have built and compared several detectors trained on
hand-crafted features extracted using basic statistics and signal processing-based methods. These
features are fed to a learning module; we have explored two kinds of learning paradigms: Sequential
Learning and Ensemble-based Learning. Specifically, we have invoked four deep sequential models:
Long-Short-Term-Memory (LSTM), Bidirectional LSTM (Bi-LSTM), Gated Recurrent Unit (GRU),
Recurrent Neural Network (RNN), and four ensemble learning approaches namely: AdaBoost,
Random Forest, LightGBM, and XGBoost. We have conducted our experiments on VSB power-linefaults-detection
challenge dataset. Our experimental findings indicate that signal processing-based
features do not improve the performance and that efficiently preprocessing signals achieves state-ofthe-art
performance compared to other features extraction techniques. In addition, we have found
that gated-based models (LSTMs and GRUs) have shown good results. Moreover, our analysis
reveals that the gradient boosting machine (LightGBM, XGBoost) exhibits high scores compared
with the other learning models. Most importantly, in order to build a successful fault detector, the
focus has to shift from model-centric to data-centric, i.e. understand the input training data and
preprocess it.
Keywords: Anomaly detection, Partial discharge, Ensemble learning, Long Short Term Memory, Signal Processing.
