Dynamic Noise Reduction with Deep Residual Shrinkage Networks for Online Fault Classification.

Area of Science:

• Electrical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• High-voltage power plant assets generate fault signals captured via electromagnetic interference (EMI).
• These EMI signals often contain varying noise levels, complicating accurate analysis and classification.
• Existing methods struggle with effective de-noising and feature extraction in noisy EMI data.

Purpose of the Study:

• To develop and validate a deep residual shrinkage network (DRSN) for de-noising and classifying EMI fault signals.
• To address the challenge of varying noise levels in EMI signals for improved fault detection.
• To enhance the feature engineering of raw time-series signals using time-frequency decomposition.

Main Methods:

• Utilizing a deep residual shrinkage network (DRSN) with learned thresholds for de-noising.
• Implementing a time-frequency signal decomposition method for feature engineering.
• Training and validating multiple DRSN architectures on labeled EMI fault signals with controlled noise addition.

Main Results:

• DRSN architectures incorporating the residual-shrinkage-building-unit-2 (RSBU-2) demonstrated superior performance over RSBU-1 in low signal-to-noise ratio environments.
• The proposed de-noising and classification methods proved effective even with controlled noise injection at various levels.
• Accurate classification was achieved across different noise conditions, validating the model's robustness.

Conclusions:

• The DRSN with shrinkage methods and learned thresholds offers a robust solution for de-noising EMI fault signals.
• The developed approach is sufficient for real-world EMI fault classification and condition monitoring in high-voltage power plants.
• Implementing thresholding methods within de-noising techniques significantly improves performance in noisy signal environments.