Converter Transformer Fault Diagnosis Based on Domain Adaptive Semi-Supervision, Lightweight and Model Compression

Aiming at the problem that deep learning cannot directly read one-dimensional dissolved gas data in oil, a wavelet transform data dimension conversion method based on the generalized Gaussian function is proposed to convert one-dimensional fault samples into two-dimensional The characteristic image is used to satisfy the data Diagnostic format of the converter transformer fault diagnosis method and at the same time, domain adaptive semi-supervised loss is introduced to optimize the model parameters. In view of the characteristics of existing deep learning models with many parameters and complex structures, a method based on lightweight neural networks and model compression is proposed to integrate and reduce deep learning models through convolutional layers and batch-normalization (BN) layers. The method of model compression by adjusting the bit-width of the parameters not only can effectively compress the deep learning model with almost no loss of model accuracy, but also improves the fault identification efficiency. In addition, the method can be successfully deployed on the edge computing platform, which realizes the effective utilization of resources and real-time response.