A Resilient MEMS Sensor Array-AI System for DGA-Based Transformer Fault Monitoring in High-H₂ Environments

Area of Science:

• Electrical Engineering
• Materials Science
• Artificial Intelligence

Background:

• Metal-oxide-semiconductor (MOS) gas sensors are crucial for real-time dissolved gas analysis (DGA) in power transformers.
• High hydrogen (H2) environments degrade MOS sensor performance due to cross-interference, limiting accuracy in transformer monitoring.

Purpose of the Study:

• To develop a co-optimized sensing framework integrating a hybrid sensor array and deep learning model for robust DGA.
• To overcome the limitations of MOS sensors in high-hydrogen environments for improved power transformer diagnostics.

Main Methods:

• A hybrid sensor array combining Palladium-Gold (Pd-Au) and MOS sensors was developed.
• A deep learning model incorporating a 1D Convolutional Neural Network (CNN) and a Long Short-Term Memory network with an attention mechanism (LSTM-AM) was employed.
• A smooth-label training strategy was introduced to mitigate prediction instability.

Main Results:

• The proposed framework achieved a Mean Squared Error (MSE) of 0.0020 on a custom dataset (D1), significantly outperforming the UCI-TGS benchmark (MSE: 0.0157) by 87.3%.
• The smooth-label strategy reduced prediction variance by 50% compared to conventional methods.
• The integrated system demonstrated improved Pd-Au sensor performance and halved training loss.

Conclusions:

• The developed hardware-algorithm system provides an accurate and robust solution for real-time DGA in power transformers.
• This approach enhances diagnostic reliability for smart grid applications by overcoming hydrogen interference challenges.
• The study highlights the potential of hybrid sensing and deep learning for critical infrastructure monitoring.