In view of the problems of poor generalization ability and insufficient diagnostic capability of traditional Convolutional Neural Network (CNN) model due to the data distribution discrepancy in strong noise environment and across working conditions, a fault diagnosis method for rolling bearings based on parallel convolution kernel and channel attention mechanism is proposed. Using this method, a parallel network structure with different convolution kernel scales was designed to effectively extract feature information from data while suppressing noise interference. Meanwhile, channel attention mechanism was added to enhance the feature extraction capability of the convolutional layer, and improve the anti-noise performance of the model and the adaptive ability in across working conditions. Diagnosis effects were trained and tested by using bearing data set of Case Western Reserve University. The proposed method was compared with peer approaches under different Signal-to-Noise Ratio (SNR) cases and across working conditions, it was shown that the proposed method achieves an average diagnosis accuracy rate of 97.3% in across working conditions and in the variable noise experiment on the bearing dataset from Case Western Reserve University the diagnostic accuacy rate is beyond 93.8%, which are obviously higher than the competing methods; the proposed method have better noise resistance and generalization ability under complex and variable working conditions．