Elsevier

Computers in Industry

Volume 106, April 2019, Pages 142-153
Computers in Industry

Automated bearing fault diagnosis scheme using 2D representation of wavelet packet transform and deep convolutional neural network

https://doi.org/10.1016/j.compind.2019.01.008Get rights and content

Highlights

  • This paper proposes an adaptive deep convolutional neural network (ADCNN).

  • ADCNN utilizes a two-dimensional visualization of the raw acoustic emission (AE) signal.

  • ADCNN provides bearing health state information to automate the feature extraction and classification process.

  • The proposed approach outperforms existing state-of-the-art algorithms for multi-fault classification of bearings.

Abstract

Bearings are one of the most crucial components in many industrial machines. Effective bearing fault diagnosis is essential for normal and safe machine operation. Existing fault diagnosis methods are mostly limited to manual feature characterization and traditional machine learning schemes such as support vector machines (SVM) and k-nearest neighbors (k-NN) algorithms. Unfortunately, the interpretation and engineering of such features require substantial human expertise. This paper proposes an adaptive deep convolutional neural network (ADCNN) that utilizes a two-dimensional visualization of the raw acoustic emission (AE) signal to provide bearing health state information, which serves to automate and better generalize the feature extraction and classification process. This 2D visualization tool applies a discrete wavelet packet transform (WPT) and quantifies each sub-band of the signal by defining a new evaluation metric—the degree of defectiveness ratio (DDR)— to precisely represent each fault condition, henceforth called a DDRgram. The motivation for using this DDRgram-based preprocessing scheme is that valuable information regarding rotating components is distributed across discrete frequencies, and thus bearing health conditions can be revealed by those frequency spectra. Furthermore, the proposed ADCNN is trained using an adaptive learning method to achieve improved diagnostic performance. The efficiency of the proposed fault diagnosis methodology (DDRgram + ADCNN) is verified using AE data, collected from a benchmark bearings testbed. The experimental studies demonstrate that the proposed approach outperforms existing state-of-the-art algorithms for the multi-fault classification of bearings with good accuracy.

Introduction

Bearings are essential elements in modern machines as diverse as aircraft, automobiles, steel mills, and wind turbines [1,2]. Bearings are the components that fail the most frequently because they must support heavy loads at a stationary rotating speed, accounting for approximately 51% of all failures [3,4]. Thus, fault diagnosis of bearings is of paramount importance to keep machines in normal operation and lessen the breakdown time.

Fortunately, since 1990, the rapid development of data mining, data acquisition techniques, and machine learning techniques has provided the ability to collect and store an immense amount of process data to extract the useful information inherent in such a highly significant amount of recorded data (e.g., acoustic emission (AE), vibration, and current) for the purpose of reliable fault diagnosis in major important industries [2,3,5,6]. It is, therefore, necessary to develop fault diagnosis methods that can efficiently process a massive amount of data to self-learn fault features and obtain accurate diagnosis results. Vibration signals and motor current analysis techniques have been widely exploited in the field of bearing fault diagnosis. These methods primarily established high reliability for high-speed bearings fault diagnosis, with rates ranging from a few hundred to a few thousand revolutions-per-minute (RPM), due to the difficulty of capturing intrinsic information regarding low-speed bearing defects [[7], [8], [9]]. In contrast with a vibration-based analysis, this study utilizes an AE signals-based diagnosis method, since AE techniques are highly effective for early-stage fault detection when operating at low speeds.

Traditional data-driven fault diagnosis methods are based on two main processes, namely manual fault feature extraction, by signal processing techniques, and identification of faults using the extracted features [3,10,11]. These manually-crafted features require careful engineering and substantial domain expertise that transform the raw data (e.g., time sample values of a one-dimensional (1D) bearing fault signal) into an appropriate feature vector or internal representation, from which a learning system can classify the patterns in the input. For example, Jack and Nandi implemented support vector machine and neural network models on extracted time-domain features for rolling element bearing fault detection and diagnosis [11]. Kang et al. [10] explained the utilization of parameter combinations through a hybrid feature extraction model, which aimed to extract sufficient information about bearing defects to define each bearing fault condition uniquely, and the selected features subset was further utilized by a k-NN classifier to identify different fault types. Therefore, the customized data-driven diagnosis methods discussed in [4,10,12] face major challenges in the age of big data and more generalized datasets. For example, features are selected for a specific diagnostic problem and might not be appropriate for different fault diagnosis problems. Another challenge is that the neural networks used in most approaches are constructed with only a single hidden layer. Such a shallow structure confines their capability to adaptively learn and explore coveted nonlinear information. Therefore, we incorporate a broadly-applicable deep learning technique to resolve these problems.

Recently, the manifold literature has reported rotating machinery fault diagnosis using deep learning approaches [[12], [13], [14], [15]]. Janssens et al. [12] proposed a fault diagnosis method using convolutional neural networks (CNN) and raw vibration signal pre-processing. In that approach, one-second windows of raw vibration signals were first extracted. For each window of obtained samples, the discrete Fourier transform (DFT) was applied. The amplitudes of the DFT decompositions were then used as training data samples for the CNN model. However, this kind of uncorrelated reshaping may not represent distinct bearing health states, and this process is very similar to using a 1D raw signal with a domain change. The issues with the existing method are that they do not consider the appropriate bearing health conditions and the physical characteristics of the faults, which might be the reason for sub-optimal classification performance.

In this paper, we develop an appropriate 2D visualization tool to represent bearing health states. Any crack or defect on a bearing raceway (outer, inner, or roller) creates a defect frequency, such as ball pass frequency outer race (BPFO), ball pass frequency inner race (BPFI), ball pass frequency (BPF), and harmonics in the envelope spectrum, depending on the crack placement in the raceway [1,16,17]. Several researchers have explored these effects using 2D visualizations obtained from various sub-band signals using time-frequency analyses, i.e., short-time Fourier transform [18] or multi-level bandpass filters [[19], [20], [21]]. An important issue is that it is difficult to correlate each sub-band signal in the time-frequency analysis with a unique fault pattern, which can be further utilized as an effective input to the classifier (e.g., in a CNN) in such a manner that all pertinent information in the original signal is preserved or emphasized. Wang et al. [18] developed a widely used wavelet transform (WT)-based kurtogram utilizing a kurtosis value (KV) for finding the useful sub-band signals through a 2D representation of WT nodes because a kurtogram can quantify the bearing defect frequencies (BPFO, BPFI, and BPF) and their harmonics. Nevertheless, these quantifying values are not strictly proportional to the degree of defectiveness of a bearing. To solve this problem, we apply a precise degree of defectiveness ratio (DDR) calculation using a Gaussian mixture model (GMM) around the defect frequency and the ratio between defective components to the residue components in the envelope spectrum of the wavelet packet transform (WPT) nodes. The key idea of our DDR calculation is that it first creates Gaussian windows around the BPFO, BPFI, BPF, and their harmonics and then computes DDRs around those defect frequencies. This evaluation metric is highly useful for accurately measuring the defectiveness of the bearing defect in comparison with the KV of a kurtogram [18]. Thus, the 2D visualization of these DDR values of WPT nodes is called a DDRgram, which highly efficiently represents each health state.

A DDRgram yields a highly efficient way to visualize each bearing health condition and enables easy application of a classifier for fault diagnosis, whereas most of the existing studies apply no classifier to diagnose fault types [[17], [18], [19]]. Once we obtain a DDRgram representing the bearing health state, an adaptive deep convolutional neural network architecture (ADCNN), a variant of the LeNet5 [22] architecture, is proposed to automate the feature extraction and optimal feature selection processes by recognizing patterns from the image pixels. Furthermore, we apply an adaptive learning rate for training the ADCNN. This adaptive learning rate is essential for optimized performance and avoiding convergence to a local minimum. The proposed DDRgram + ADCNN diagnosis methodology is validated in a low-speed bearing fault diagnosis application with four faults, including different crack sizes and operating speeds.

The remaining sections of this paper are structured as follows. Section 2 presents a description of the data acquisition model. Section 3 describes the proposed methodology, including the DDRgram-based bearing health representation and the ADCNN, and Section 4 summarizes and analyzes the results. Finally, Section 5 contains concluding remarks.

Section snippets

Experimental testbed and bearing acoustic emission (AE) data

AE signals were recorded using a self-designed machinery fault simulator that simulates different fault conditions using a cylindrical roller element bearing (FAG NJ206-E-TVP2), as shown in Fig. 1. Bearings were seeded with cracks on different parts of the bearing. AE signals were collected for the bearings at the non-drive end of the simulator using a wide-band acoustic sensor (type WS α, from the Physical Acoustics Corporations [1]) and a PCI-2 based data acquisition system, which samples the

The proposed method for bearing fault diagnosis

The block diagram of the proposed ADCNN-based bearing fault diagnosis framework is presented in Fig. 3. The proposed method consists of three essential steps. After data acquisition, the first step is to obtain a 2D visualization (DDRgram) using a wavelet packet transform (WPT) with envelope analysis to pre-process the AE data. The main advantage of using such pre-processing is that valuable information regarding the rotating components is distributed within discrete frequencies. The second

Results and discussion

Investigations into the effects of a DDRgram-based health state visualization of a bearing fault and its application to fault diagnosis using the ADCNN are presented in this section.

To validate the performance, this paper utilizes four datasets with four fault types of OC, IC, RC, and a healthy bearing under various operating conditions (see Table 1). The effectiveness of the recorded AE signals can be seen in Fig. 9 for each bearing condition from dataset 1. As this study improves the

Conclusions

This study developed a highly efficient bearing fault diagnosis scheme using a 2-D visualization tool representing the bearing health state and an adaptive deep convolutional neural network (ADCNN). First, we applied the WPT to quantify each sub-band by defining a new evaluation metric, the DDR. The sub-bands are visualized as percentage DDR values, or DDRgrams, that represent unique fault information regarding the bearing health state. Once we obtained the DDRgram, we fed it into an ADCNN.

Conflicts of interest

The authors declare no conflict of interest.

Acknowledgments

This work was supported by the Korea Institute of Energy Technology Evaluation and Planning (KETEP) and the Ministry of Trade, Industry & Energy (MOTIE) of the Republic of Korea (Nos. 20181510102160). It was also funded in part by The Leading Human Resource Training Program of Regional Neo-Industry through the National Research Foundation of Korea (NRF) funded by the Ministry of Science, ICT and Future Planning (NRF-2016H1D5A1910564).

References (29)

  • M.M. Manjurul Islam et al.

    Reliable multiple combined fault diagnosis of bearings using heterogeneous feature models and multiclass support vector Machines

    Reliab. Eng. Syst. Saf.

    (2018)
  • M. Kang et al.

    A hybrid feature selection scheme for reducing diagnostic performance deterioration caused by outliers in data-driven diagnostics

    Ieee Trans. Ind. Electron.

    (2016)
  • M.M.M. Islam et al.

    A hybrid feature selection scheme based on local compactness and global separability for improving roller bearing diagnostic performance

    Artificial Life and Computational Intelligence

    (2017)
  • X. Dai et al.

    From model, signal to knowledge: a data-driven perspective of fault detection and diagnosis

    IEEE Trans. Industr. Inform.

    (2013)
  • Cited by (0)

    1

    Engineering building No. 1, Room No. 304-1, 93 Daehak-ro, Nam-gu, Ulsan, 44610, South Korea.

    View full text