To enhance the low precision of transformer fault diagnosis, a model based on multi-strategy improved sparrow algorithm (MISSA) and bidirectional long short-term memory network (BiLSTM) is proposed. Based on dissolved gas analysis (DGA) technology in oil, the uncoded ratio method is used to extract 9-dimensional fault features of the transformer as the input of the model for network training. The Softmax function is used to obtain fault diagnosis types in the output layer. The sparrow search algorithm ( SSA) is improved by logistic chaos mapping, uniformly distributed dynamic adaptive weights and dynamic Laplacian operator. In the initial solution set, the multi-strategy improved Sparrow algorithm (MISSA) is used to optimize the target hyperparameters. In this way, the transformer fault diagnosis accuracy is optimized, and the kernel principal component analysis (KPCA) is used to reduce the dimension of fault feature indexes, and the convergence speed of the model is accelerated. Compared with PSO-BiLSTM, GWA-BiLSTM and SSABILSTM fault diagnosis models, the diagnostic accuracy of the proposed model is 94% , which is 11. 33% , 8. 67% and 6% higher than those of PSO-BiLSTM, GWA-BiLSTM and SSA-BiLSTM fault diagnosis models, respectively. It is verified that the proposed method can effectively improve the performance of transformer fault diagnosis.