A Bearing Remaining Useful Life Prediction Method based on Inception-Resnet Module and Attention Mechanism

Abstract

In production activities, predicting the remaining useful life (RUL) of the bearing and grasping the health status is one of the prerequisites ensuring the safe and reliable operation of mechanical equipment. In order to improve the accuracy of bearing RUL prediction, a bearing RUL prediction method based on Inception-Resnet model and attention mechanism (AM) is proposed. The proposed method improves the convolutional neural network (CNN) in three aspects: network width, network depth, and enhanced features. First, large-stride convolution instead of pooling is used to perform feature compression and shallow feature learning to reduce the amount of network calculations. Then, multiple-size convolution kernels of the Inception structure are adopted for multi-scale deep feature extraction to obtain richer degradation information. As such, when increasing the network depth, the jump connection of the residual network (Resnet) can powerful relief the disappearance of gradient and network degradation. In addition, the attention mechanism is introduced to re-calibrate the deep features by giving greater weight to the more important degraded features. Finally, the re-calibrated deep features is input into fully connected network to map to get the RUL value. The experimental verification is performed through public bearing data sets, and results show that the prediction performance of proposed method is superior to other methods.