Although the distributed machine learning platform is very successful in the application of structured data, it is still very challenging in the absence of labeled data and a variety of data forms. Therefore, heterogeneous domain adaptation (HDA) emerged at the historic moment, with the goal of transferring knowledge between domains of different features and different distributions. Existing HDA methods only focus on the better alignment of the target domain with the source domain, while ignoring the rich semantic structure information of the target domain data itself, thereby affecting the performance of knowledge transfer. In order to solve the above problems, we propose a simple and effective self-training method oriented to the target domain. Specifically, we use the clustering algorithm to find the prototype of the target domain clusters, and select confident unlabeled data through a novel method of finding pseudo-labels that we propose. Under two self-training mechanisms, that is, single-stage Self-training (TST-SS) and multi-stage self-training (TST-MS), without introducing any additional network parameters, gradually train the transferable model. We conducted extensive experiments on cross-domain and cross-feature tasks to prove that our method is far superior to the existing heterogeneous domain adaptation methods.