Transformer Vibration Recognition Method Based on Lightweight Deep Learning Algorithm

In order to enhance the prediction speed of the transformer vibration recognition diagnostic algorithm, as well as to improve recognition accuracy in low signal-noise ratio environments, a method for transformer operating condition recognition based on multi-point vibration detection and lightweight deep learning algorithms is proposed. A depthwise separable convolution algorithm adapted for multi-point vibration detection is introduced: preprocessing of multi-point vibration signals generates three-dimensional data structures; the primary algorithm employs a depthwise and pointwise convolution architecture combined with dilated convolution kernels for multi-scale feature extraction from data streams; the secondary algorithm utilizes a fully connected network for sample recognition. The research findings indicate that, compared to conventional algorithms, the deep separable convolution algorithm under the multi-point detection scheme demonstrates outstanding performance. When the signal-to-noise ratio is 10dB, the recognition accuracy remains stable at 90%. Additionally, this lightweight algorithm exhibits shorter recognition prediction time, which facilitates deployment on edge computing devices to enhance processing speed.