An Improved Multiscale Convolutional Neural Network with Large Kernel for Bearing Fault Diagnosis

Abstract

We propose an improved end-to-end Multiscale Convolutional Neural Network with Large Kernel (LKMCNN) for bearing fault diagnosis in this paper. The LKMCNN is an end-to-end network, which can automatically extract features from the original vibration signal and accurately diagnose bearing fault without any manual feature selection operations. The LKMCNN can extract features at a wide-scale by using a large convolution kernel, which can effectively prevent information loss and improve the robustness of the model. Benefit from the adaptively features extraction of short-term, medium-term, and long-term periods by three parallel convolution operation with different kernel size, the adaptability and robustness of the model are improved. Compared with three excellent baseline models, the LKMCNN achieves state-of-the-art performance in bearing fault diagnosis by experiments using Paderborn bearing fault dataset.