Abstract: Aiming at the problem that fault diagnosis models cannot access target domain data during the training phase in practical industrial scenarios,leading to performance degradation under unknown working conditions,a multi-source domain generalization fault diagnosis method integrating data-level and feature-level approaches was proposed.Firstly,at the data level,an improved data augmentation strategy was proposed,which constructed augmented domains by mixing and generalizing multi-source domain data and their labels,achieving inter-domain information exchange and enhancing sample diversity.Secondly,at the feature level,a supervised contrastive learning mechanism was introduced into the multi-source domain adversarial network.While extracting domain-invariant features,this method employed contrastive loss to explicitly constrain the feature space,minimizing intra-class sample distances and maximizing inter-class sample distances,thereby extracting feature representations with stronger discriminative power and generalization capability.Experimental results show that this method can effectively address the feature shift issue in cross-condition scenarios.Compared with other mainstream domain generalization methods,it significantly improves diagnostic accuracy under unknown target working conditions.