Abstract: To address the problems of insufficient suppression of feature distribution offset and significant risk of negative migration of existing fault diagnosis models in rolling bearing cross-condition scenarios, a fault diagnosis method based on deep multi-source domain adaptation was proposed. Firstly, the dynamic weight allocation module was designed to quantify the difference between the source and target domain distributions through Wasserstein distance, and the softmax function was incorporated to adaptively fuse the multi-source knowledge to suppress noise interference and negative migration. Secondly, a multi-scale feature extraction network was constructed, and the parallel time-domain inflationary convolutional branch and the frequency-domain short-time Fourier transform branch were adopted to capture local transient features of the vibration signal and the global frequency-domain modes, and time-frequency feature interaction reinforcement was achieved through cross-scale attention mechanism. Finally, multi-discriminator adversarial training and maximum classifier difference criterion was introduced to jointly optimize domain-invariant feature alignment and classification discriminability. Experimental validation was carried out through the multi-source domain adaptation task, and the results show that the proposed method has a higher diagnostic accuracy and generalization ability than other traditional multi-source domain adaptation methods. The average diagnostic accuracy improved by up to 3.43%, while task-specific performance fluctuations were reduced by up to 40%. This provides a new way of thinking for cross-condition fault diagnosis of rolling bearings in complex industrial scenarios.