Transfer learning (TL) technology have been successfully applied to address the domain adaptation (DA) problem in machinery fault diagnosis. However, partial DA problem is more suitable for industrial applications, where the target data only covers a subset of the source classes, which makes it difficult to know where to transfer the target data. To overcome this problem, a novel game theory enhanced domain adaptation network (GT-DAN) is proposed in this paper. Based on different metrics, including Maximum mean discrepancy (MMD), Jensen-Shannon (JS) divergence and Wasserstein distance, three attention matrices are constructed to describe the distribution discrepancies between the source domain and the target domain, and the optimal coordination between these attention matrices is achieved by the combination weighting based on game theory to generate the optimal probability weights, which can guide to filter out the irrelevant source examples in domain adaptation. Two experiments show that the proposed GT-DAN is superior to the existing methods in the partial DA diagnosis performance.