An On-Line Detection Method and Device of Series Arc Fault Based on Lightweight CNN

Abstract

To quickly and accurately detect the series arc fault (SAF) in three-phase motor with frequency converter load (TMFCL) circuit, a SAF identification model based on convolutional neural network was proposed. The point-by-point isometric mapping was presented to construct input matrix. The lightweight design of the model was realized, respectively, by using bottleneck building block and depthwise separable convolution. A roofline model was used to analyze the complexity and theoretical runtime of the convolution operators. According to the runtime of the operators, the optimal lightweight SAF identification model was determined and labeled as SAFNet. A SAF on-line detection device was designed by deploying SAFNet to an embedded device. And its performance was evaluated by on-line tests. When the sampling frequency is 2.5 kHz, the accuracy is higher than 99.44%, and the runtime is less than 26.48 ms. It can be used to develop arc fault circuit interrupter for the TMFCL circuit.