A Deep Learning Approach for Series DC Arc Fault Diagnosing and Real-Time Circuit Behavior Predicting

Electromagnetic interference (EMI) produced by series arc fault has long been a concern in dc distribution systems. As EMI fault may lead to serious safety risks, it is important to detect and predict arc disturbance for EMI management purposes. This article proposes a deep-learning-based approach for series dc arc diagnosis and circuit behavior prediction. Using time–frequency slices generated from power supply-side signals as a reference input, the proposed method is capable of extracting supply-side circuit features with a time–frequency feature extractor. The two-step feature extractor adopts convolutional neural network to extract static features on each time–frequency slice, and combines a long short-term memory network to capture dynamic time-varying signatures of time–frequency slice sequence. Fully connected network layers are lined up behind to implement mapping relation between extracted features and expected output. To further evaluate the performance of the proposed method, an experimental platform is established for series dc arc simulation and circuit data collection. Experimental data under normal switching operation and various arcing states are captured and employed to train the deep learning model. In this article, trained models show an overall accuracy of 98.43% in arc fault diagnosis, and give a time-domain prediction result resembling the actual signal.