Rolling bearings in rotating machinery often work in the environment with variable loads and strong noise. Traditional fault diagnosis methods of rolling bearings are difficult to adaptively extract the favorable features under complex conditions, so a fault diagnosis method of rolling bearings with variable conditionsis proposedbased on improved AlexNet. Firstly, one-dimensional time-domain signals are translated into two-dimensional feature maps using transverse insert samples method to satisfy the requirements of the improved AlexNet input.Compared with the existing longitudinal insert samples method or two-dimensional spectrums method, the time series and correlation of vibration signals are preserved during feature extraction automatically. Secondly, the functional layer of AlexNet convolutional layer is improved and adjusted, andthe profitable characteristics for the state identification of rolling bearingscanbe automatically extracted via the convolution and sampling operations of improved AlexNet from the two-dimensional feature maps. Finally, the softmax cross entropy is considered as a loss function and Adam is used to realize the fault diagnosis of rolling bearings according to a small batch iterative optimization method. Compared the diagnosis effects with other methods for 12 kinds of states of different positions and damage degrees of rolling bearings under variable loads and strong noise, the results show that the proposed method has a higher accuracy of fault diagnosis of rolling bearing and its robustness is stronger.