Abstract: To address the diagnosis problem in the scenario of varying data distribution and inconsistent label space due to the change of wind turbine working conditions, a partial domain adaptation method（FWDAN） based on fusion weights domain adversarial is proposed for cross-working condition fault diagnosis of rotating machinery. The core idea of FWDAN is to apply the training weights at both the sample and category to weaken the role of outlier category samples in the adversarial training process and enhance the learning of shared category samples, thus facilitating the transfer of shared diagnostic knowledge between domains and improving the diagnostic performance. For sample-level weight generation, the label information is coupled into the sample data to fully explore the feature representation. Further, different statistical methods are applied to generate weights for assisting model training according to the differences between source and target domain data to achieve the purpose of promoting positive model transfer and reducing the risk of negative transfer. The experimental results of two diagnostic cases built on rolling bearing and gearbox datasets show that the proposed method has higher diagnostic accuracy and stronger generalization ability than other methods.