Abstract: In order to improve the accuracy of transformer fault diagnosis with single neural network method, a transformer fault diagnosis method based on the fusion of improved neural network and ratio method is proposed. To solve the problem of adaptation between the deep one-dimensional convolution neural network (1D-CNN) and transformer dissolved gas data, an improved 1D-CNN is built as the basic classifier of fusion classification method. A fusion classification module (FCM) is suggested to identify in advance the samples that can potentially be misclassified by the network and switch to the traditional ratio method for individual data analysis. This aims to enhance the application performance of neural networks in transformer fault diagnosis. The simulation study is given to verify the operability and adaptability of the proposed method. The results show that, compared with conventional one-dimensional convolutional neural network and recurrent neural network, the improved 1D-CNN performs better as a basic classifier. FCM can improve the performance of basic classifiers under different data sets, and the improvement effect is more stable for basic classifiers with initial accuracy higher than 95%.