Difference Matrix and Self-Attention Mechanism for Series AC Arc Fault Detection

The long-term use of electrical equipment inevitably results in the occurrence of aging or breakage of the cable insulation. In the event of aging or breakage, the arc may be produced, which may result in the occurrence of electrical fires. Therefore, it is necessary to accurately detect the arc fault. To solve this problem, this article proposes an arc fault detection framework based on a difference matrix and a self-attention mechanism. The framework realizes arc fault detection and load type identification by converting current data into a difference matrix image and then inputting it into a network model. In comparison with the conventional arc current time-frequency feature extraction approach, the difference matrix is a method that converts time series data into images. The method is not constrained by the threshold setting and exhibits a high degree of self-adaptation. The difference matrix has been shown to effectively characterize the time domain features in arc current by downscaling the current data and calculating the difference signals at different scales. For the difference matrix image, this article constructs a convolutional neural network (CNN) model that includes a self-attention mechanism. This model is capable of capturing the correlation of different scales of differential signals in the difference matrix image, with the objective of weighting key features and enhancing the detection performance of the model. Finally, the effectiveness and superiority of the proposed method are validated through experimental analysis. The results demonstrate that the proposed method can achieve accurate detection of arc faults and load types, with a detection accuracy of 98.85%.