For smart grids, intelligent and on-line fault diagnosis systems are required to keep power reliability and operational stability. This paper proposes a novel AI-based architecture for fault detection and classification, which overcomes issues faced by traditional methods like high computational complexity, poor robustness to operational changes and lack of interpretability. The proposed model is based on a light hybrid machine learning architecture to achieve the fault type and the zone classification performing at low latency. It is applicable to edge or cloud deployment and it works well with IoT-enabled monitoring solutions. Unlike prior GAN or GNN-based methods which suffer from the difficulty on generalizability, our method demonstrates its robustness on dynamic grid and on unseen faults pattern. It also provides explainable output for helping operators make decisions. Experiments on benchmark datasets demonstrate considerable gain in detection accuracy, efficiency, and robustness. This research aids in the development of smart grid intelligent fault management, providing a flexible and explainable method in the context of current complex energy systems.