Transformer fault diagnosis method based on the three-stage lightweight residual neural network

The fault diagnosis method for dissolved gas in transformer oil based on deep learning has the problems of complex structure, over-parameterization, and high resolution in practical application. This paper presents a three-stage lightweight residual neural network method for transformer fault diagnosis. In the first stage, based on the 50-layer residual networks, the residual block is enhanced using the inverted bottleneck idea, and the Swish activation function and a simple, parameter-free attention module are incorporated to optimize the model structure and performance. In the second stage, an adaptive channel pruning method is proposed, selectively eliminating redundant filters and channels based on the fault data complexity during the training process, thereby realizing network lightweight. In the third stage, a quantization-aware method is introduced, which converts all 32-bit floating point parameters in the network to 8-bit integers, reduces the bit width of each parameter, and accomplishes a reduction in parameter size. The experimental results for the transformer oil dissolved gas fault dataset indicate that the three-stage lightweight model, sizing at 2.20 MB—only 1.51 % of the original—achieves a fault diagnosis accuracy of 97.64 %, 1.49 % higher than the original, achieving a well-balanced between accuracy and complexity.