To ensure reliable oil-immersed transformer fault diagnosis, a transformer fault diagnosis method based on a multi-strategy improved dung beetle optimizer (MIDBO) optimized bi-directional long short-term memory network (BiLSTM) is proposed. There are shortcomings of the dung beetle algorithm, such as poor global search ability and it is easy for it to fall into a local optimum. First, through the Bernoulli chaotic map, the introduction of adaptive factors and the Levy flight strategy, the dynamic weight coefficient is improved and its performance is evaluated. Then, MIDBO is used to optimize many hyperparameters of BiLSTM to form the MIDBO-BiLSTM fault diagnosis model. Kernel principal component analysis (KPCA) is used to extract the eigenvalues, and the correlation between the eigenvalues and the fault types is analyzed in depth to improve the convergence speed of the model. The final experimental results show that the proposed MIDBO-BiLSTM transformer fault diagnosis method has high accuracy and strong generalizability. Its accuracy rate is as high as 94.67%. This is suitable for transformer fault diagnosis.