Due to random uncertainty factors， such as the change in operating conditions of rolling bearings and noise interference， it is easy to lead to incomplete network feature extraction and the inability to capture local singular information of fault mutations. A parallel two-dimensional depthwise separable residual neural network （P2D-DSResNet） is proposed in this paper to address the aforementioned issues. By using Gramian angular difference field （GADF） and continuous wavelet transform （CWT）， one-dimensional vibration signals can be transformed into two-dimensional time frequency images while preserving the complete time-frequency domain information. The P2D-DSResNet uses depthwise separable convolution instead of ordinary convolution in residual modules， which enhances feature learning ability and improves the model’s feature extraction capability for better fault diagnosis in high noise and varying operating conditions. The bearing data obtained from a fault simulation testbed are used for experimentation and analysis. A comparative analysis with other convolutional neural network methods demonstrates that the optimized algorithm model has good generalization ability and high fault recognition accuracy.