Aiming at the problems of low accuracy, insufficient real-time performance and continuous operation of high load of existing arc fault detection devices, a fault arc detection algorithm based on Convolutional Neural Networks ( CNN ) and characteristic period change rate is proposed. The algorithm is divided into two detection modes : power-on and process. In the power-on mode, multiple features are imported into CNN for detection. In the process mode, the concept of periodic change rate of different state eigenvalues of current is proposed, and the change rate of eigenvalues is partitioned. CNN is combined for progressive detection, which reduces the complexity of the algorithm while ensuring the accuracy. The algorithm uses stm32H743 as the processor, with conditioning, data acquisition and other circuits to form a real-time arc fault diagnosis system. After experimental testing, the average accuracy of the device for AC arc fault detection is 97.43%, and the fastest detection time is as low as 0.045 s. It can provide theoretical support and reliable reference for the development of arc fault detection device.