Abstract:  In order to give full play to the advantages of deep learning recognition in two-dimensional image in rolling bearing fault detection, an intelligent fault diagnosis model based on Gram angle difference field (GADF) and improved convolutional neural network (CNN) was proposed. Firstly, the one-dimensional time series vibration signal was transformed into a two-dimensional image through the gram angle difference field, and the target image features were extracted and input into the improved CNN model. Secondly, the global pooling layer instead of the traditional full connection layer was used by the improved CNN model, which effectively solved the problem of parameter explosion of the traditional CNN model. Finally the ideal detection through iterative training with Adam small batch optimization method accuracy was achieved. The test and comparison results show that the diagnosis method is more rapid and accurate in feature extraction, fully demonstrates the nonlinear expression ability of CNN model, and the detection accuracy is better than the other intelligent diagnosis algorithms.

Key words: Gram angle difference; field fault diagnosis; convolution neural network; deep learning