In the process of load identification of ball mill under varying working conditions, a domain adaptation method based on domain antagonism and classification difference is proposed to solve the problem that domain adaptation method does not consider the target domain sample in the feature transfer between source domain and target domain. The method uses domain adversarial training to align the features between source domain and target domain. At the same time, two classifiers are introduced to detect samples far away from the target domain, and the inconsistency between the maximization and minimization of the classifiers is utilized to realize the adaptive matching of the features of the target domain and the source domain to achieve a better domain adaptation effect. In order to verify that the method of training the classifier error can consider the in-class boundary to improve the load recognition accuracy on the target domain, a migration experiment is designed to analyze the impact of its difference loss function on the model migration performance. The experiment shows: When the classifier loss value is greater than 0. 02, the accuracy of the prediction model will decrease by 0. 8% ~ 1. 2%, and the load accuracy is higher than that of the model without the classifier differential loss, which can reach 95. 78%. Compared with two classical transfer methods, the advantages of this method in the application of mill load identification under varying working conditions are verified.