Determination of neural-network topology for partial discharge pulse pattern recognition

Area of Science:

• Electrical Engineering
• Artificial Intelligence

Background:

• Partial discharge (PD) is a critical indicator of insulation degradation in electrical equipment.
• Accurate recognition of PD patterns is essential for predictive maintenance and preventing catastrophic failures.
• Existing methods for PD pattern recognition face challenges in handling complex, time-dependent data.

Purpose of the Study:

• To develop and evaluate a novel time-series approach for partial discharge pulse pattern recognition.
• To investigate the efficacy of different neural-network topologies for this application.
• To identify the most successful neural-network structure for differentiating between PD patterns.

Main Methods:

• Employed a time-series approach to design neural-network (NN) topologies.
• Implemented and compared various NN structures for PD pattern recognition.
• Focused on a cascaded output NN structure, utilizing indexed features from the first stage as input for the second stage.

Main Results:

• The cascaded output NN structure achieved the highest success rate in differentiating between two distinct PD patterns.
• The integration of indexed first-stage output features significantly improved the recognition performance.
• This approach demonstrated superior capability in handling time-dependent PD data.

Conclusions:

• Cascaded neural networks offer a highly effective solution for time-dependent partial discharge pattern recognition.
• The proposed method, leveraging sequential feature integration, enhances the accuracy and reliability of PD analysis.
• This advancement contributes to more robust condition monitoring and insulation diagnostics in electrical systems.