Intelligent CNN-based Fault Diagnosis of Rotating Machinery with Small Fault Samples

Up to now, convolution neural network (CNN) has been widely applied in fault diagnosis of rotating machinery. The CNN-based diagnostic methods often with the help of the fault data to implement the optimization. However, in real industrial applications, it is difficult and costly to obtain the fault data. In this paper, a ResNet-based diagnostic method combined a designed data transformation combination is proposed to achieve the fault diagnosis under small fault samples. Specifically, several data transformation of image are selected to deal with the small samples, where the parameter of the transformation is determined by the mutual information between the inputted sample and the corresponding transformed sample. Meanwhile, these obtained transformations are used as the input layer of the ResNet. Then, a fault diagnosis model is established by the constructed ResNet, and which is trained only by using small samples. Noting that the introduced data transformation is randomly used for the training samples to increase the complexity of the inputted samples in the training process for alleviating the overfitting risk. The bearing fault dataset is used to evaluated the effectiveness of the proposed method. From the experimental result, it is found the proposed has the capability to implement the effective fault diagnosis under small samples, and achieve a higher diagnostic performance than other existing methods.