Aiming at the bad fault classification accuracy of rolling bearing in motor due to different probability distributions and insufficient sample data, a fault classification method is proposed based on improved multi-channel convolutional network with bidirectional gated units for cross-working conditions. A multi-channel convolutional neural network with bottleneck and BiGRU module was designed to capture global fault information from raw vibration signals end-to-end, while reducing computational load through the bottleneck module and optimizing information flow through the BiGRU module. Local maximum mean discrepancy (LMMD) is adopted to complete subdomain adaptation, reducing the feature distribution differences between the source and target domains in the pre-trained models. Three scenarios were distinguished: different loads at the same speed, different loads at different speeds, and conditions with a wide range of variations. Twelve transfer tasks were designed on the SDUST, CWRU, and PU public datasets to experimentally validate the proposed method. The experimental results show that the average accuracy of the proposed method reaches 90.09%, 99.70%, and 91.75% respectively, significantly higher than comparison methods such as MMD, DANN, and CDAN. It maintains a top single-task accuracy of 99.99% under strong scenario shifts, showing high precision and generalization. The results of CWRU dataset show that, compared with other methods, such as DAMSCN-BiGRU, MSDAM, and improved DANN unsupervised domain adaptation model, the proposed method has also a higher accuracy quantity under cross-working conditions.