A multi-stream multi-scale lightweight SwinMLP network with an adaptive channel-spatial soft threshold for online fault diagnosis of power transformers
Xiaoyan Liu, Yigang He
Abstract
Abstract Fault diagnosis of power equipment is extremely crucial to the stability of power grid systems. However, complex operating environments, high costs and limitations of single-modal signals are the biggest bottlenecks. To this end,a multi-tream, multi-scale lightweight Swin multilayer perceptron (MLP) network (MLSNet) with an adaptive channel-spatial soft threshold is proposed in this paper. First, a Res2net-based feature-enhanced method is used to learn the correlated features of vibration and voltage multi-modal signals. Second, a novel MLSNet is designed to combine the benefits of Swin transformers with an MLP with a lightweight convolutional neural network and employs a staged model to extract various scale features. Third, an adaptive deep fusion approach employing a channel-spatial soft threshold module is used to integrate and recalibrate staged information at different scales. The overall accuracy of the proposed method can reach 98.73% in various experiments, potentially making it an effective method for online fault diagnosis of power transformers.