In order to improve the accuracy of transformer fault diagnosis, a transformer fault diagnosis method based on adaptive deep learning model is proposed. The method uses dissolved gas in oil as fault diagnosis feature, and builds diagnostic model based on deep learning theory. To solve the shortcomings of slow convergence speed and low convergence precision in the training process of traditional fixed learning rate based deep learning model, an adaptive deep learning model construction method is proposed. This method adaptively adjusts the learning rate according to the changing characteristics of the iterative process, and effectively improves the training accuracy and speed of the deep learning model. Parameters of adaptive deep learning model for transformer fault diagnosis such as hidden layer number, learning rate adjustment coefficient are proposed. The experimental results show that the proposed method has a strong ability of feature extraction and analysis and has better convergence speed and convergence precision, which can effectively improve the accuracy of transformer fault diagnosis.