MGCM-net: a deep learning framework for robust reading of SF6 gas pressure gauges

Abstract The SF6 gas pressure gauge serves as a critical monitoring device in power systems, where its accurate reading is essential for grid security. To address the unique non-planar pointer structure and the requirement for high reading accuracy in low-pressure regions of SF6 gauges, this paper proposes an innovative deep learning framework named MGCM-Net. The framework incorporates a critical pressure augmentation (CPA) strategy to specifically enhance low-value samples by simulating visual degradation under real operating conditions. It utilizes a metadata-guided modulation (MGM) module to deeply integrate gauge range parameters into the visual feature extraction process, improving pointer position perception. Additionally, an adaptive penalty loss (APL) function is designed to optimize the training process dynamically. Experimental results demonstrate that the proposed method achieves an overall reference error of 1.160% and a low-pressure region error of 1.39%. Notably, the CPA strategy reduces the low-pressure region error by 8.92% points. Meanwhile, the framework maintains a computational efficiency of 81.55 FPS, thereby offering reliable technical support for smart grid development.