When a support vector machine (SVM) is applied to transformer fault diagnosis, the optimum values of its kernel function parameters g and c are difficult to select from manual experience, and the accuracy of fault diagnosis is low. The bald eagle search (BES) algorithm has the defects of easily falling into a local optimum and low convergence accuracy. Given these problems, a transformer fault diagnosis model based on an improved bald eagle search algorithm (Ct-GBES) is proposed to optimize the parameters g and c of the SVM. To improve the convergence speed and search ability of the algorithm, the three stages of the BES algorithm are improved and optimized using a tent chaotic map, adaptive t-distribution and dynamic selection, and the golden sine algorithm. The superiority of the Ct-GBES algorithm is verified by comparing it with the optimization tests of the original BES, the cuckoo algorithm (CS) and the firefly algorithm (FA). The Ct-GBES-SVM model is compared with the SVM, FA-SVM, CS-SVM models in fault diagnosis experiments, and is compared with the BES-SVM model in stability experiments. The results show that the proposed model has higher accuracy, better stability, shorter running time and offers better fault diagnosis.