Rolling element bearing fault diagnosis using convolutional neural network and vibration image

Abstract

Detecting in prior bearing faults is an essential task of machine health monitoring because bearings are the vital components of rotary machines. The performance of traditional intelligent fault diagnosis methods depend on feature extraction of fault signals, which requires signal processing techniques, expert knowledge, and human labor. Recently, deep learning algorithms have been applied widely in machine health monitoring. With the capacity of automatically learning complex features of input data, deep learning architectures have great potential to overcome drawbacks of traditional intelligent fault diagnosis. This paper proposes a method for diagnosing bearing faults based on a deep structure of convolutional neural network. Using vibration signals directly as input data, the proposed method is an automatic fault diagnosis system which does not require any feature extraction techniques and achieves very high accuracy and robustness under noisy environments.

Introduction

Rolling Element Bearings (REBs) are the essential components of rotary machines. Health conditions of REBs have considerable impacts on machines. According to a literature review, 45–55% of broken machines are caused by bearing faults (Nandi, Toliyat, & Li, 2005). Thus, condition monitoring and fault diagnosis of bearings are significant tasks in the industry. The most popular way to diagnose faults of bearings is intelligent diagnosis which employs machine learning (ML) algorithms. A typical intelligent diagnosis method includes four steps as follows: data acquisition, feature extraction, feature selection, and feature classification (Huang, 1996).

Data acquisition step collects signal from machines by sensor systems. Until now, for bearing fault diagnosis, current methodologies often use acoustic emission signals (Chacon, Kappatos, Balachandran, & Gan, 2015), motor current signals (Singh, Kumar, & Kumar, 2014), and vibration signals (Zarei, Tajeddini, & Karimi, 2014). Among those types of bearing data, vibration signal based method is the most popular approach because vibration signals are easy to measure and can provide rich dynamic information reflecting bearing health status (Kharche & Kshirsagar, 2014).

Measured vibration signals from machines contain not only useful information reflecting machine conditions but also futile noisy signals. Consequently, it is important to extract only helpful features and avoid useless information. Originally, vibration signals are temporal signals in time domain, but they can be represented in frequency domain and time-frequency domain. Correspondingly, features of vibration signals can be extracted from time domain (Samanta & Al-Balushi, 2003), frequency domain (Malhi & Gao, 2004), and time-frequency domain (Lou and Loparo, 2004, Yen and Lin, 2000). In time domain, features can be extracted by Root Mean Square, Kernel Density Estimation, Crest factor, Crest-Crest Value and Kurtosis (Prieto, Cirrincione, Espinosa, Ortega, & Henao, 2013). In frequency domain, Fourier Transform is the most popular tool (Lin, Ye, Huang, & Su, 2016), while in time-frequency domain, besides Short-time Fourier Transform method (Li, Sanchez, Zurita, Lozada, & Cabrera, 2016), features can be extracted by Wavelet Packet Transform (Hemmati, Orfali, & Gadala, 2016), Dual-tree Complex Wavelet Transform (Van & Kang, 2016). In addition, there are some other methods for extracting features such as Intrinsic Mode Function (Pandya, Upadhyay, & Harsha, 2013), Hilbert Huang Transform (HHT) (Osman & Wang, 2016), and Empirical Mode Decomposition (Van & Kang, 2015).

After the feature extraction step, the dimensionality of the feature set should be reduced because there is no guarantee that all features are equally useful in reflecting machine health (Shen, Wang, Kong, & Peter, 2013) and a high dimensional feature set not only weakens the performance but also slows down the learning process of the adopted classifier in the system. To address this problem, there are a lot of discriminant feature analysis techniques have been proposed. In general, there are two approaches can be used to select discriminant features (Van & Kang, 2016). In the first way, a subset of the original feature set will be generated based on the transformation of the existing features. Some well-known methods of this approach are Principal Component Analysis (PCA) (Malhi & Gao, 2004), Independent Component Analysis (ICA) (Hyvärinen & Oja, 2000). In the second method, based on some evaluation criteria, the original features will be evaluated to select most discriminant features. In this approach, a lot of algorithms can be applied such as Sequential Forward Selection (SFS), Sequential Backward Selection (SBS) (Guyon & Elisseeff, 2003), or we can exploit Genetic Algorithm (GA) (Yang & Honavar, 1998) and Particle Swarm Optimization (PSO) (Xue, Zhang, & Browne, 2013).

ML is often exploited to solve pattern recognition problems. There are two major types of ML, includes unsupervised learning and supervised learning. Unsupervised learning algorithms such as PCA and ICA are often used in feature extraction step of pattern recognition problems. Among supervised learning algorithms, Artificial Neural Network (ANN) is the most popular method with a huge number of models such as Recurrent Neural Network (Zaremba, Sutskever, & Vinyals, 2014), Flexible Neural Tree (Bao et al., 2014, Bao et al., 2017), Radial Basis Neural Network (Huang & Du, 2008), and Wavelet Neural Network (Pindoriya, Singh, & Singh, 2008). Besides ANN, Support Vector Machine (SVM), k-Nearest Neighbor (kNN), Hidden Markov Model (HMM), and ensemble learning (Bao et al., 2017, Brown, 2011) are also popular supervised algorithms. In the last step of a general intelligent diagnosis, those supervised learning algorithms are often employed. Amaral et al. (2013) made spectral images from vibration signals, then these images were enhanced by a 2-D averaging filter and binary image conversion. Finally, an ANN was exploited to classify those enhanced spectral images. Li, Meng, Ye, and Chen (2008) proposed the method using higher-order statistics features based on Discrete Wavelet Transform. After obtaining the higher-order statistics features, a kNN was used to identify types of bearing faults.

Traditional ML algorithms have been employed widely for a long time in machine fault diagnosis. However, in this approach, some drawbacks are existing. First, the classification accuracy mainly depends on the feature extraction step, which requires signal processing techniques and expert knowledge (Li et al., 2015). Thus, for every specific fault diagnosis task, feature extractor must be redesigned. The second drawback is that the traditional ML algorithms have shallow architectures with simple structures which limit the capacity of the classifiers to learn complex non-linear relationship in fault diagnosis issues (Jia, Lei, Lin, Zhou, & Lu, 2016).

Recently, deep learning (DL) emerged as the hottest trend in ML research. DL are algorithms which employ deep architectures that can learn multiple levels of data representations that correspond to different levels of abstraction (Deng et al., 2014). With the capacity of automatically learning multiple complex features from the input data, DL algorithms have great potential to overcome the disadvantages of traditional ML as mentioned above. There are a lot of DL models have been proposed such as Recurrent Neural Network (RNN), Deep Belief Network (DBN), Deep Boltzmann Machine (DBM), Stacked Auto Encoder (SAE), and Convolutional Neural Network (CNN). DL has great promise in many practical application (Dean et al., 2012), including computer vision (Krizhevsky et al., 2012, Szegedy et al., 2015), natural language processing (Mikolov, Karafiát, Burget, Cernockỳ, & Khudanpur, 2010), medical image analysis (Brosch et al., 2016, Greenspan et al., 2016), and machine health monitoring (Tamilselvan & Wang, 2013). In machine health monitoring, since 2015, a lot of researchers have tried to exploit DL models to diagnose bearing faults. Jia et al. used SAE to extracted features from signals in frequency domain (Jia et al., 2016). In an introductory paper, Chen et al. proposed three deep models, includes DBN, DBM and SAE for bearing fault diagnosis (Chen et al., 2017). They applied two approaches to learning fault features, the first one is to use deep model with raw signals in time domain directly. In the second approaches, low-level features were extracted from time domain, frequency domain, or time-frequency domain, then deep models were employed to learn higher-level features from those low-level features. L. Eren used a one-dimensional CNN to diagnose bearing faults, using raw signals directly in time domain (Eren, 2017).

In the family of DL algorithms, DBM and DBN are based on Restricted Boltzmann Machine (RBM), SAE is based on Autoencoder (AE), all of them are unsupervised learning algorithms, while CNN is a supervised learning method. Initially, with three key architectural ideas: local receptive fields, weight sharing, and sub-sampling in spatial domain, CNN is suitable for processing 2-D data (Phung & Bouzerdoum, 2009). In machine health monitoring, some researchers have tried to apply one-dimensional CNN models (Chen et al., 2017, Eren, 2017, Jia et al., 2016). However, it is much easier to extract information from data in a high dimension (Ding et al., 2017). Being motivated by this fact, in this paper, we propose a CNN model (VI-CNN), for diagnosing bearing faults using 2-D form of vibration signals. First, vibration signals in time domain are transformed into 2-D form, called vibration images. After that, a CNN will be used to identify faults of bearing through vibration image classification. To verify the effectiveness of the proposed method, we conduct experiments with bearing data from Case Western Reverse University (Loparo, 2005).

The remainder of this paper is structured as follows. Section 2 shortly introduces about CNN. The proposed fault diagnosis method is explained in Section 3. Section 4 describes experiments. Finally, conclusions are drawn in Section 5.