Abstract: For cross domain diagnosis of the label spaces of source domain and target domain are partially overlapped， that is to say， both the target domain and the source domain contain the classes that the other does not have， a cross domain adaptive fusion diagnosis method based on weighted adversarial learning is proposed. As entropy can be used to reflect the characteristics of the shared known classes and unknown classes， two convolutional neural networks with the same structure are introduced to carry out entropy-based weighted adversarial training， which is aim to enhance the ability to identify the shared known classes by extracting the domain-invariant features， as well as the binary cross schemes of the source domain and target domain sample outputs are used to isolate the unknown classes. In addition， the fully connected layer hidden features of these two convolutional neural networks are taken as the input of two label transfer models， and the probability outputs of these three diagnostic models are fused by voting rule. The failure test bench data of mechanical transmission components under variable working conditions and the damage data of self-priming centrifugal pump are used for analysis and verification， the experimental results show that the proposed cross domain adaptive fusion diagnosis method can distinguish the shared known classes and unknown classes in the target domain more accurately.