Abstract: In order to improve the reliability of series fault-arc detection, an experimental platform is built according to the standard, and a real-time data acquisition device is made to collect both faulty and non-faulty current data generated by 4 types of linear or non-linear loads, including incandescent lamps, fluorescent lamps, air compressors and blowers. A one-dimensional convolutional neural network (1D-CNN) is proposed to inspect current signals in the circuit and determine whether fault arcs occur. Testing results show that the average detection accuracy of the model for all kinds of loads reaches 100%, and the loss is below 0.0007. We also transplant the model into an embedded system, and the accuracy reaches 96.25%. It is proven that the designed convolutional neural network structure can successfully detect series fault arcs and reduce the risk of fire.

Key words:  fault arcs, convolutional neural networks, fire risk