Fault Diagnosis of Bearings under Multiple Operating Conditions Based on Modular Neural Network

Abstract

Fault diagnosis for rolling bearings are important since they are critical components in rotating machinery used in intelligent production process. Although deep learning is efficient for fault diagnosis of rolling bearing, once the equipment operates in multiple conditions it is difficult to diagnosis all faults accurately due to its bad generalization ability. In order to overcome the above problems, this paper designs a modular neural network with dynamic routing technology as the module fusion mechanism (DR-MNN). We first divide the data collected in multiple conditions into different modules according to different working conditions, and perform preliminary feature extraction according to each working condition. Then, reorganize the extracted features of different working conditions. The innovation of our work is that we use vector neuron ideas to further divide the reorganized features into modules. Dynamic routing technology is used to adaptively assign weights to different fault modules to achieve module fusion. Finally, to multi- condition fault diagnosis on the original bearing data. Moreover, the efficiency of this method is verified by using the bearing dataset of Case Western Reserve University.