Abstract: Glowing contact faults in electrical circuits caused by poor connections are highly prone to triggering electrical fires, but the locations of glowing contact faults are often unknown. To solve the detection trouble, this paper builds a glowing contact experimental platform, focusing the detection on the loop current of the experimental circuit, and extracts its time-domain, frequency-domain, and time-frequency domain features. To address the problems of small sample size and data imbalance in actual scenarios, a Wasserstein Gradient Penalty Generative Adversarial Network is established for data augmentation. Finally, a one-dimensional convolutional neural network is built to identify new samples obtained in the laboratory. Experimental results show: glowing contact faults and normal loop currents can be distinguished in the characteristic frequency band of 5~15 kHz. The built glowing contact fault identification model achieved a training accuracy of 99.93%, successfully realizing fault identification for 15 sets of loop states obtained from the laboratory.

Key words: electrical fire, glowing contact, neural network, multi-feature fusion, fault identification, arc