To improve the generalization capability of domain adaptive fault diagnosis methods on unlabeled target domains, a fault diagnosis model to uncover the latent causal relationships within vibration signals was proposed. Based on a sequential variational autoencoder (VAE), robust feature representations associated with fault types were learned using fault labels. Guided by the principles of domain labels, factors unrelated to the faults yet reflective of the domain’s data distribution were separated, thereby progressively decoupling the vibration signals causally. The evidence lower bound (ELBO) of the sequential VAE was restructured to effectively guide the decoupling process. By incorporating domain adaptation techniques, the shared fault features’ distance between the source and target domains was brought close in the feature space, enhancing the generalization of the features learned by the encoder. Experimental results on the CWRU and IMS datasets show that the feature decoupling capability of the causal network complements domain adaptation methods, improving the performance of the fault diagnosis model in domain adaptation tasks.