In order to solve the problem of low diagnostic accuracy when conventional features are used as input features of transformer fault diagnosis model, a transformer fault diagnosis method based on Gaussian kernel Fisher discriminant analysis (GKFDA) and Newton-Raphson-based optimization (NRBO) algorithm optimized bidirectional long short-term memory network (BiLSTM) was proposed. Firstly, a 23-dimensional feature that can reflect transformer faults was constructed, and the optimal feature set was determined by GKFDA. Secondly, a transformer fault diagnosis model based on NRBO optimized BiLSTM was established. Then, the proposed method was verified by an example. The results show that the diagnostic accuracy of the proposed method is 98.4%, which is better than conventional feature method and other optimization diagnosis models. Finally, the proposed method was tested using IECTC10 data. The results show that the proposed method has good generalization in transformer fault diagnosis.