Abstract: Aiming at the problem that traditional analysis methods are difficult for fault diagnosis of bearings under variable speeds， a fault diagnosis method was proposed for variable speed bearings based on GADF(Gramian angular difference field) and ResNet34 model introduced transfer learning. Firstly， a one-dimensional time-series vibration signal was encoded by using GADF and converted into a two-dimensional image to generate the corresponding fault maps， which were then input into a residual network (ResNet) using transfer learning to automatically extract and classify fault features. To verify the effectiveness of the method， the results of comprehensive comparison with other methods showed that the proposed method performs better on the Western Reserve University bearing dataset. Finally， the variable speed bearing dataset from the University of Ottawa in Canada was diagnosed to examine its classification performance in the variable speed case. The results showed that a high diagnostic accuracy can be achieved in the variable speed case.

Key words: fault diagnosis; variable speed bearing; Gramian angular field; residual network(ResNet); transfer learning