Multi-scale Joint Convolution Neural Network integrated CapsNet model for Machinery Fault Diagnosis

The utilization of feature extraction models based on convolutional neural network (CNN) has gained significant traction in fault diagnosis applications. However, conventional 1D convolution fails to fully capture the spatial information within vibration signals; likewise, the input of 2D convolution often loses crucial details due to irreversible conversion from the vibration signal into a time-frequency graph. In this paper, drawing inspiration from the characteristics of vibration signals, a Multi-scale Joint CNN (MJCNN) module that combines 1D and 2D convolution is proposed to extract more representative features. Subsequently, it is integrated into a capsule network for constructing the fault diagnosis model. Ultimately, the effectiveness of MJCNN module in feature extraction and the superiority of the fault diagnosis method are verified through a set of twelve diagnostic tasks.