In this paper, a Joint Adversarial Domain Adaptive Network(JADAN) based cross working conditions fault diagnosis method is proposed to address the challenges of acquiring labeled fault data in real-world scenarios and the inconsistencies between training and test data distributions that render diagnostic models ineffective. Initially, domain adversarial training is employed to extract deep domain-invariant features from both source and target domains, enhancing the generalization capability of the diagnostic model in the target domain. A Softmax and structural prediction based pseudo-label strategy is proposed, enabling the model to generate pseudo-labels for unlabeled target domain data. Simultaneously, a class alignment module is incorporated to minimize the distance between class prototypes of the source and target domains, achieving joint alignment of domain and class. This effectively reduces the probability of samples near the decision boundary being misclassified.Additionally, a weight assignment mechanism for source domain samples is integrated into the domain discriminator, enabling adaptive weight assignment for each source sample. This effectively mitigates the negative transfer issue during model training and enhances model robustness. Experimental results demonstrate that the proposed method effectively addresses the problem of cross working conditions fault diagnosis.