Fault Diagnosis of Rolling Bearing Based on Improved Convolutional Neural Network

Abstract

Rolling bearing has an irreplaceable role in industrial production. Since most fault diagnosis methods for bearings require manual extraction of fault features, and convolutional neural networks are prone to problems such as gradient disappearance, this paper proposes an improved fault diagnosis algorithm for One-Dimensional Convolutional Residual Neural Networks. The extraction and compression of data fault features is first accomplished by convolutional pooling. An improved residual network is then added to avoid network degradation in the training model and uneven distribution within the data. It also uses Global Average Pooling to reduce the training model parameters and randomly deactivates neurons in the structure via Dropout techniques to prevent complex co-responses to the training data. The final four classification results are output and the model convergence rate is adjusted by dynamic learning rate throughout to prevent the emergence of local optima.