Introducing wireless interface(WI) in traditional wired network on chip has pushing performance of on-chip system to new limits. While integrating more components like antenna and transceiver also increases system complexity, which causes system more susceptible to failures. In this paper we propose a run-time fault diagnosis mechanism based on artificial neural network. We build our dataset by collecting data from partially faulty NoC and trained the neural network offline. Then Ann assembled in NoC can recognize if there is any failure happened and locate the failure position with run-time date. To evaluate this mechanism, we compare the performance by utilizing several different neural networks and testing them in different fault situation(locations, fault numbers, fault categories, ANN category). Results show that CNN takes more advantages and achieves detection rate up to 83% in 2D mesh topology.