A Robust Fault Diagnosis Method for Rolling Bearings Based on Deep Convolutional Neural Network

Fault diagnosis of rolling bearings has been an important and challenging research issue. The existing conventional algorithms for diagnosing rolling bearings rely on artificial feature extraction requiring a wealth of knowledge of signal handling, expertise and human efforts, lacking adaptability. With the capacity for learning features from original signals automatically, deep learning methods can solve the shortcomings of conventional diagnosis methods. This paper puts forward a deep learning method for rolling bearing fault diagnosis based on two-dimensional convolutional neural network, called 2dCNN-FD. The method has the capability of automatic feature extraction and robust noise tolerance. Experimental results demonstrate that comparing with other traditional diagnosis algorithms, the 2dCNN-FD model achieves high diagnosis accuracy and better robustness under different noise conditions.

I. Introduction

Being widely used in industrial equipment, rolling bearings play a crucial role in modern rotating machinery and their health conditions has a great impact on the machinery health and its remaining lifetime [1]-[3]. With an amount of effort putted into the topic, fault diagnosis of rolling bearings has been a one of research hotspots for decades [4]. The most popular way for bearing fault diagnosis is intelligent methods combined shallow machine learning (ML) with artificial feature extraction. An intelligent fault diagnosis algorithm usually contains three main steps: feature extraction, feature selection and feature classification. In general, the original vibration signal collected by the sensors contains a lot of useless noise, so the purpose of feature extraction is to obtain helpful information from original vibration data for fault classification. The widely-used methods of feature extraction include fast Fourier transform (FFT) [5], discrete wavelet transform (DWT) [6], empirical mode decomposition (EMD) [7] and so on. Feature selection is carried out after feature extraction to discard the insensitive and useless features and further reduce the number of effective features. Principle component analysis (PCA) [8], independent component analysis (ICA) [9] and manifold learning [10] are widely applied to select features. The last step, the selected features are inputted to classifiers such as multi-layer perceptron (MLP) [11], hidden Markov model (HMM) [12], support vector machine (SVM) [13] and so on for training to realize fault classification. The conventional methods mentioned above have been widely used to diagnosing faults of rolling bearings. Lin et al. [14] introduced a novel feature extraction approach based on Morlet wavelet and applied it for mechanical fault diagnosis. Cai et al. [15] combined EMD with genetic neural network adaptive boosting (GNN-AdaBoost) to realize fault feature extraction and classification. Kankar et al. [16] extracted features of the bearing original data by statistical methods and then made use of artificial neural network and SVM respectively for feature classification to complete diagnosis.