基于边缘2D-CNN轻量部署的ANPC逆变器多开路故障实时诊断

Multiple open-circuit (OC) faults have severely deteriorated the reliable operation of active neutral-point-clamped (ANPC) inverters. To achieve fast and accurate detection for multiple OC faults of three-level ANPC (3L-ANPC) inverter, this article deploys a lightweight convolutional neural network (CNN) into an edge computation board. Specifically, the proposed online diagnosis includes offline training and online deployment. During the offline training, a multisource information fusion is devised to promote the effective feature extraction, thereby enhancing the predictive accuracy of the designed CNN. Furthermore, a specific optimization framework is utilized to simplify the offline training process. For online deployment, the lightweight design of the trained CNN is realized by employing the TensorRT to reduce the computation cost and accelerate the diagnostic speed. Comparative studies are conducted to verify the advantages and the generalization ability of the designed CNN. Finally, the feasibility of the proposed edge-based diagnosis strategy is experimentally verified through the platform of the permanent magnet synchronous motor drives fed by 3L-ANPC inverter.