Efficient Detection of Series Arc Fault at the Back End of Frequency Converter Using KTDM-Optimized Lightweight Model

Series arc fault (SAF) is one of the main causes of electrical fire. There is no effective detection method for SAF at the back end of frequency converter (BEFC). A lightweight SAF identification model was proposed, and an on-line SAF detection device was developed. First, the SAF experiments at the BEFC were carried out. Second, the SAF identification model was designed based on EfficientNet, multiple lightweight modules mobile inverted bottleneck convolution (MBConv) were applied to extract the SAF features, and the global average pooling layer was used to improve the model’s complexity. Third, a lightweight design method
$k$
-means++ and t-SNE Design Method (KTDM), which was proposed based on the idea of reducing feature mapping redundancy, was used to optimize the identification model. The specific steps include: 1) visualize the feature mappings output from the convolutional layers of the model by using the t-distributed stochastic neighbor embedding (t-SNE) algorithm; 2) cluster the visualization results by using
$k$
-means++ algorithm; and 3) analyze the clustering results by using the elbow method and silhouette coefficient, and determine the optimal channel number of the convolutional layers. Finally, the optimized model was deployed to an embedded device to detect SAF under various operation conditions in real time. Both identification accuracy and real-time performance meet the IEC62606 standard. KTDM can improve the complexity and real-time performance of the model without obviously lowering its identification accuracy and robustness.