Intelligent fault diagnosis of rolling bearing using deep wavelet auto-encoder with extreme learning machine

Highlights

• Wavelet auto-encoder is designed with wavelet function to capture the signal characteristics.

• A deep wavelet auto-encoder is constructed with multiple wavelet auto-encoders to enhance the unsupervised feature learning ability.

• The proposed method effectively diagnoses the different fault types, different fault severities and different fault orientations of rolling bearing.

Abstract

Unsupervised feature learning from the raw vibration data is a great challenge for rolling bearing intelligent fault diagnosis. In this paper, a novel method called deep wavelet auto-encoder (DWAE) with extreme learning machine (ELM) is proposed for intelligent fault diagnosis of rolling bearing. Firstly, wavelet function is employed as the nonlinear activation function to design wavelet auto-encoder (WAE), which can effectively capture the signal characteristics. Secondly, a DWAE is constructed with multiple WAEs to enhance the unsupervised feature learning ability. Finally, ELM is adopted as the classifier to accurately identify different bearing faults. The proposed method is applied to analyze the experimental bearing vibration signals, and the results confirm that the proposed method is superior to the traditional methods and standard deep learning methods.

Introduction

As science and technology develop rapidly, modern rotating machinery becomes more high-speed, large-scale and integrated [1], which plays a more and more important role in different industries. Rolling bearing is not only a damageable but also the most significant component in the rotating machinery [2]. Thus, automatic and timely identification of bearing operating conditions has become increasingly important for reducing the unexpected downtime and economic losses.

Intelligent diagnosis is the new development of machinery fault detection technology, which can effectively analyze massive collected data and automatically provide reliable diagnosis results [3], [4]. Among different intelligent diagnosis methods, artificial neural network (ANN) and support vector machine (SVM) have been the most widely applied in the past decades. The major challenge of bearing intelligent fault diagnosis using ANN or SVM is to extract the sensitive features from the measured vibration signals [5], [6]. Li et al. [7] calculated 1634 features to reflect bearing conditions, and selected 12 sensitive features as the inputs of ANN for fault identification. Lei et al. [8] adopted wavelet packet transform (WPT) and empirical mode decomposition (EMD) for feature extraction, and then selected sensitive features for fault diagnosis based on ANN. Ali et al. [9] combined linear feature selection and ANN for online diagnosis of bearing. A feature vector based on 19 parameters was designed by Zhang et al. [10], and then SVM was carried out for bearing fault diagnosis. Lu et al. [11] extracted 71 features using EMD, and then the selected sensitive features were utilized as the inputs of SVM for bearing fault diagnosis. Zheng et al. [12] proposed multiscale fuzzy entropy as features and carried out SVM as the bearing fault classifier. Liu et al. [13] calculated 56 features based on redundant second generation wavelet package transform (RSGWPT), and then the reduced features are fed into SVM for bearing fault detection. Through the literature review, it can be found that the traditional intelligent diagnosis methods have two obvious limitations: (1) the diagnosis performance of traditional methods depends largely on feature extraction and selection [3], [14]. In order to accurately recognize different fault types, different fault severities and even different fault orientations, various advanced signal processing techniques must be well mastered. Worse still, it is a challenge to select the most sensitive features from the original feature set for different diagnosis tasks. Actually, feature selection usually cannot get rid of the dependence on engineering experience. (2) Shallow learning models are applied in the almost all the traditional diagnosis methods, which are difficulty to effectively solve very complex pattern recognition problems [15], [16]. Consequently, it is an urgent need to develop deep architectures for unsupervised feature learning from the raw vibration data.

Deep learning is a huge breakthrough in artificial intelligence, which can automatically learn the essential features from the raw data [17], [18]. In order words, deep learning can get rid of the reliance on various advanced signal processing techniques and manual feature extraction. Deep auto-encoder (DAE), convolutional neural network (CNN) and deep belief network (DBN) are three popular models in deep learning field [19], which have been gradually applied to mechanical fault diagnosis in the last three years [20], [21], [22]. Compared with DBN and CNN, DAE belongs to a purely unsupervised feature learning model, which can be trained more effectively and easily. However, there exists the challenge of applying the standard DAE directly for bearing intelligent fault diagnosis. One important reason is that the activation function of the standard DAE is usually selected as Sigmoid function, which is difficulty to establish the accurate mapping relationship between the various patterns and the raw input signals [23]. What's worse, in practical engineering, the vibration signals collected from bearing are always very complex and non-stationary with heavy noise [24]. Wavelet functions contain the scale factor and shift factor, which have been successfully employed as the new activation functions of shallow neural networks to construct the so-called wavelet neural networks (WNNs). The shift factor enables the wavelet to perform ergodicity analysis along the time axis of the signals. The scale factor aims to shrink and stretch wavelets to approximate the signals with different frequencies for each ergodicity analysis [25]. Therefore, wavelet functions have good time-frequency localization property and focal features through the change of different scale factors and shift factors. Lots of investigations have proved that WNNs usually show obvious advantages over the traditional neural networks in various applications [26], [27], [28]. However, there are few researches on wavelet functions applied in various deep learning models. Thus, it is meaningful to develop novel DAE models which combine the advantages of wavelet functions and deep learning.

In this paper, a novel method called deep wavelet auto-encoder (DWAE) with extreme learning machine (ELM) is proposed for intelligent fault diagnosis of rolling bearing. The proposed method can be divided into three major steps: Firstly, wavelet function is employed as the activation function to design wavelet auto-encoder (WAE), which can effectively capture the signal characteristics. Secondly, a DWAE is constructed with multiple WAEs to enhance the unsupervised feature learning ability. Finally, ELM is adopted as the classifier to accurately identify different bearing faults. The proposed method is applied to analyze the experimental vibration signals. The results show that the proposed method can get rid of the dependence on manual feature extraction, and it is more effective than the traditional methods and standard deep learning methods.

The remainder of this paper is organized as follows. The theory of standard auto-encoder is briefly reviewed in Section 2. In Section 3, the proposed method is introduced. In Section 4, the experimental results are analyzed and discussed. Finally, conclusions and further work are given in Section. 5.