Abstract: The unsupervised domain adaptation method uses the knowledge learned from the source domain label data to classify the target domain unlabeled data,which has become the mainstream method to solve the feature alignment of the two domains in transfer learning.In view of the fact that the amount of labeled data is small and the quality is not high,the extracted features are incomplete,this paper proposes an unsupervised domain adaptation method based on the optimal selection of self-supervised tasks.In order to make the features have stronger semantic information,multiple self-supervised tasks are used on the unlabeled data in the two domains.In addition,a new intelligent combination optimization strategy is proposed to adaptively select effective features for self-supervised tasks.Finally,the two domains are approached along the task-related direction so that the classifier trained on the source domain label data can be better promoted to the target domain.The simulation experiment conducts a comparative analysis on the six public benchmark datasets from three aspects:classification accuracy,training data volume,and self-supervised task use effect.Experimental results show that the proposed method outperforms the existing advanced methods in three aspects.The classification accuracy is improved by 8% when using the same datasets,and the amount of data used is reduced by 12% under the same classification accuracy requirements.Compared with a single self-supervised task,the accuracy is improved by 11%.

Key words: Intelligent combination optimization strategy, Self-supervised task, Semantic information, Transfer learning, Unsupervised domain adaptation