Transfer learning is a method of transferring knowledge from the source domain to solve the problem of the target domain, which can effectively solve the problem of inconsistent data distribution. Aiming at the problem that traditional methods lack a reasonable analysis of the transferability of multi-domain and the effective treatment of transfer effects in multi-domain transfer, the multi domain adaptation-manifold Regularization transfer learning algorithm is proposed to improve the effect of single source domain transfer while realizing effective transfer of multi-domain. Firstly, the transferability of multi-domain is analysed to select the source domains that can be transferred. Then, the marginal distribution and the conditional distribution is adapted and the balance factor is introduced to obtain the balanced distribution adaptation, and the manifold regularization is used to constrain the data structure to maximize the use of information in the single source domain. Finally, different domain classifiers are adaptively weighted by weighting factors, and the information of multi-domain is fully utilized to solve the target domain problem. This algorithm is applied to the optimization of the abrasive media in the mass finishing. By establishing the similarity matching method of the abrasive media, the multi-domain adaptation-manifold regularization of the abrasive media optimization model is constructed. A large number of comparative experiments show that the proposed method performs better, with an accuracy rate of up to 73.44%, which can provide more effective decision-making guidance for the optimization of the abrasive media in the mass finishing.