Transformer fault diagnosis using continuous sparse autoencoder.

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Transformer fault diagnosis is critical for power system reliability.
• Traditional methods often struggle with complex, nonlinear data.
• Unsupervised feature learning offers a promising approach for improved diagnostic accuracy.

Purpose of the Study:

• To introduce a novel Continuous Sparse Autoencoder (CSAE) for unsupervised feature learning.
• To apply CSAE for enhanced transformer fault recognition using dissolved gas analysis data.
• To evaluate the effectiveness of CSAE in improving transformer fault diagnosis.

Main Methods:

• Developed a Continuous Sparse Autoencoder (CSAE) incorporating Gaussian stochastic units.
• Applied CSAE for unsupervised feature extraction from normalized IEC three ratios data.
• Integrated CSAE with a Back Propagation (BP) network for supervised fault classification.

Main Results:

• CSAE successfully extracted meaningful features from nonlinear dissolved gas analysis data.
• The proposed CSAE-BP network achieved a superior correct differentiation rate in transformer fault diagnosis.
• Comparative experiments validated the effectiveness of CSAE over existing methods.

Conclusions:

• The Continuous Sparse Autoencoder (CSAE) is a powerful tool for unsupervised feature learning in transformer fault diagnosis.
• CSAE significantly enhances the accuracy of transformer fault recognition.
• This approach offers a robust solution for reliable power transformer monitoring.