BroadNet: A Novel Broad Learning System-Based Series AC Arc Fault Detection Approach With Zero-Phase Component Analysis Whitening Transformation
Junwei Duan, Linghao Zeng, Yajie Wu, Long Chen, C. L. Philip Chen
Abstract
Arc faults in domestic circuits are one of the main causes of home fires. Due to the complexity of arc faults, traditional circuit protectors cannot effectively detect series arc. In this paper, a novel broad learning system-based approach with time and frequency features (BroadNet) is proposed for the detection of series AC arc faults. In the proposed method, effective features are first extracted from the current signal in the time and frequency domain, respectively. Then, Zero Phase Component Analysis (ZCA) whitening transformation is applied to reduce the correlation and redundancy of the features. Finally, the detection of arc faults is conducted using the broad learning system (BLS). According to the experimental results obtained from 4000 training samples and 1000 testing samples, which were extracted from four types of appliances, the accuracy of arc fault detection using our proposed method can reach 98.6%. Furthermore, the training time is only 0.0803 seconds and testing time is a mere 0.0061 seconds. Compared to other methods, our BroadNet demonstrates faster computation speed and higher detection accuracy. This highlights the potential of our approach in the field of industrial anomaly detection.