LMFE: Learning-Based Multiscale Feature Engineering in Partial Discharge Detection.

Area of Science:

• Electrical Engineering
• Power Systems Analysis

Background:

• Partial discharge (PD) detection is crucial for power distribution stability.
• Existing feature engineering methods for PD detection face limitations in fault identification, temporal representation, and multiscale feature integration.

Purpose of the Study:

• To develop a learning-based multiscale feature engineering (LMFE) framework for improved PD detection in three-phase power systems.
• To address limitations of existing methods, including fault-related pulse identification, inner-phase temporal representation, and multiscale feature integration.

Main Methods:

• Preprocessing three-phase measurements to identify pulses and surrounding waveforms.
• Extracting global-scale features (phase-level, measurement-level) and local-scale features (waveform, inner-phase temporal information).
• Utilizing a recurrent neural network (RNN) for feature extraction and merging multiscale features for classification.

Main Results:

• The proposed LMFE framework successfully extracts multiscale features, integrating global and local information.
• The LMFE framework demonstrates superior performance compared to existing approaches in PD detection.
• The VSB ENET dataset analysis confirms LMFE as a state-of-the-art solution.

Conclusions:

• The developed LMFE framework offers a significant advancement in partial discharge detection accuracy.
• LMFE effectively addresses key challenges in PD detection, enhancing reliability in power distribution.
• This approach provides a robust and state-of-the-art solution for identifying faulty signals.