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

MOS gas sensors offer significant potential for real-time dissolved gas analysis (DGA) in power transformer monitoring. However, their performance is often degraded in high-hydrogen (H₂) environments due to cross-interference, which impairs detection accuracy and limits practical deployment. To overcome these challenges, we propose a co-optimized sensing framework that integrates a MEMS-based hybrid sensor array with a CNN-LSTM-AM deep learning model. The hybrid array combines Pd-Au and MOS sensors to exploit their complementary gas-response behaviors, enabling reliable hydrocarbon detection even under H₂ saturation. On the algorithmic side, a 1D convolutional neural network (CNN) extracts subtle gas features from saturated MOS signals, while the LSTM-based attention mechanism (LSTM-AM) compensates for Pd-Au sensor drift by learning temporal dependencies. To further enhance robustness, a smooth-label training method is introduced to reduce prediction instability during abrupt concentration transitions. Experimental results demonstrate that our framework achieves a mean squared error (MSE) of 0.0020 on a custom datset (D1), outperforming the UCI-TGS benchmark by 87.3% (MSE: 0.0157). Moreover, the smooth-label strategy reduces prediction variance by 50% compared to conventional labeling. This integrated hardware-algorithm system not only improves Pd-Au sensor performance and reduces training loss by half but also provides an accurate and robust solution for real-time DGA, contributing to enhanced diagnostic reliability in smart grid applications.