Influenced by indistinguishable data samples and poor data balance, transformer state identification models using vibro-acoustic signals often have low accuracy. To address this problem, Focal loss is introduced to dynamic feedback weights according to the accuracy of the sample training process, thus constituting a Focal-XGBoost optimization model. Firstly, a set of filters that fit the transformer spectrum are used to fully extract the effective information of the vibro-acoustic signal, and then XGBoost-PCA is used to reduce the dimensionality of the samples. Then, the Softmax objective function in the original model is optimized using Focal loss to form the Focal-XGBoost model. After inputting the above samples, the hyperparameters of Focal are optimized based on the accuracy wave action, and the transformer state recognition results are output. The experimental results of 10 kV and 110 kV transformers show that Focal-XGBoost can reduce the number of samples compared with traditional SVM and KNN models. Focal-XGBoost reduces the misspecification of difficult samples in XGBoost test samples by 44.7%, which results in higher model recognition accuracy. In addition, non-uniform extraction compresses the sample space by 50% on the basis of average accuracy loss below 0.5%, which further reduces the model training cost.