Aiming at the low accuracy of transformer fault diagnosis, a multi-strategy integrated model is proposed. Firstly, the high dimensional nonlinear indivisible transformer fault data is reduced by isometric mapping (Isomap). The hybrid kernel based extreme learning machine (HKELM) is used for training and learning. Considering that the HKELM model is easily affected by parameters, the northern Goshawk optimization (NGO) algorithm is used to optimize its parameters. However, due to the slow convergence rate of NGO, it is easy to fall into local optimal, and Chebyshev chaotic mapping, optimal learning, and adaptive T-distribution joint strategies are introduced to improve it. At the same time, in order to improve the overall accuracy of the model, the Adaboost-INGO-HKELM transformer fault identification model is constructed by combining the Adaboost integrated algorithm. Finally, the test accuracy of the proposed Adaboost-INGO-HKELM model is compared with that of INGO-HKELM model without dimensionality reduction, Isomap-INGO-KELM model, Adaboost-Isomap-GWO-SVM model. The Adaboost-INGO-HKELM model proposed in this paper can achieve an accuracy of 96%, which is higher than other models, which verifies that the model has a good effect on transformer fault identification.