Most of existing domain adaptation-based scenarios used to estimate inter-domain distribution discrepancy can ignore the impact aroused by conditional distribution discrepancy between classes in two domains on the measurement of inter-domain distribution discrepancy, which can make the accuracy of knowledge transfer drop dramatically. A fault diagnosis method based on one-dimensional convolutional auto-encoder and conditional distribution domain adaptation is proposed for rolling bearings across rotating machinery. An inter-domain dissimilarity criterion (T-LCORAL) based on threshold local subdomain correlation alignment is constructed, which is embedded into the output layer of the autoencoder to guide the model for feature extraction effectively. Then, pseudo labels are employed to divide the source and target domains into multiple subdomains for fine-grained alignment within the domain, Moreover, the pseudo labels are screened by reliability thresholds to improve the discriminability of the boundary distribution between classes in the domain. Finally, three bearing datasets were used for cross mechanical bearing fault transfer diagnosis experiments, and the results showed that the proposed method can achieved better performance.