An intelligent fault diagnosis method based on gramian angular field and multi-scale convolutional neural network (GAF-MCNN) was proposed to solve the non-stationary, nonlinear and easily disturbed background noise of motor bearing micro-fault signals. Firstly, the piecewise aggregation approximation algorithm is used to compress and reduce the dimension of the original vibration signals to reduce the data storage space and improve the computational efficiency. Then, the one-dimensional sequence signals are converted into two-dimensional matrix heat maps using the gramian angular field algorithm. The two-dimensional matrix strengthens the time relationship between the original vibration signals and encodes the time dimension into the matrix structure. Finally, a multi-scale convolutional neural network is designed to diagnose the fault efficiently and quickly. An example of motor bearing fault diagnosis shows that GAF-MCNN method not only overcomes the problem of low computational efficiency of traditional convolutional neural network diagnosis methods, but also has better diagnostic accuracy than single-scale convolutional neural network method, and has strong engineering practicability.