A neuro-wavelet based approach for diagnosing bearing defects

https://doi.org/10.1016/j.aei.2020.101172Get rights and content

Highlights

  • Designing a precise fault diagnosis system with an improved accuracy of 99.48%.

  • Providing exhaustive experimentation of 27 integrated diagnostic schemes.

  • A comparative study between variational mode decomposition and wavelet based methods.

  • Examining projective methods and manifold modeling to remove redundant features.

Abstract

In recent years advanced signal processing techniques are used increasingly to excavate the nonstationary vibration signals and extract elemental-fault information. However, managing and analyzing a multicomponent signal mixed with background noise using only a single analysis tool is not a simple task and may lead to low diagnostic accuracy and a delayed diagnosis. This paper introduces a novel intelligent neuro-wavelet based system with high diagnostic accuracy based on nonrecursive variational mode decomposition (VMD) and wavelet-based neural network, which mainly consists of three steps (i.e. feature extraction (FE), dimension reduction (DR), and fault classification). Firstly, the vibration signals are segmented and processed by a novel nonrecursive VMD, which can decompose the nonstationary signals into a series of discrete modes adaptively, to extract informative features from vibration signals. Multi-Class generalized discriminant analysis is then used in the second step that aims to reduce the dimension of the feature set and improve the computational burden by selecting meaningful information and removing redundant features. In the next step, the obtained features vector is fed to a state-of-the-art hierarchical multi-resolution classifier, so-called wavelet neural network (WNN), which possesses the advantages of both wavelet transform and artificial neural networks for the decision-making. Additionally, to evaluate the information extraction capability of VMD, the subsequent DR method and the calculation accuracy of WNN, other state-of-the-art techniques are used in this work. In this regard, the superiority of the proposed approach is also confirmed through an experimental comparison with published works in the literature.

Introduction

In the golden age of technology, induction motors (IMs) have been the main tool for transforming electrical energy into mechanical energy [1]. Considering their robustness, simplicity, reliability and low price, induction motors are extensively utilized in many industries. However, sudden unpredictable failures and aging components in IMs result in considerable maintenance cost and economic losses [2], [3].

Generally, induction motors’ failures can be divided into two main groups of electrical and mechanical faults [4]. However, almost 41% of all unexpected machine failures are due to bearing defects [5]. Hence, this type of fault should be detected as early as possible to decrease friction losses and increase the system efficiency [2].

The literature on process monitoring and fault diagnosis (PM-FD) methods is huge and diverse. The PM-FD methods can be classified into signal processing methods [2] and redundancy-based methods [6], [7]. Analytical redundancy methods use some redundant knowledge to diagnose faults in the specific process. In this approach, the fault detection system uses the input and output signals to produce the residual signal carrying the fault information [7]. The signal processing-based methods extract fault symptoms from specific signals carrying fault information of concerned components. Although contemporary diagnostic schemes are commonly based on the mathematical model [8], [9], they rely on an exact or highly reliable model of the process which is not easy to obtain.

In this regard, data-driven diagnostic schemes have gained a substantial amount of research attention for fault detection in recent years. In 2018, Jiang et al. presented a complete review of the recent advances in data-driven monitoring, fault diagnosis and control approaches with their applications to the autonomous vehicles and the smart grid [10]. In [11], an efficient MATLAB Data-based key-performance-indicator oriented fault detection toolbox (DB-KIT) was developed to present a series of advanced schemes for process monitoring and fault diagnosis. Liu et al. in [12], proposed a hybrid fault classification algorithm constructed based on three steps: extraction of features using redundant second generation wavelet package transform (RSGWPT), selection of dominant features based on kernel principal component analysis (KPCA) and applying twin support vector machine (TWSVM) for classification phase. Yan and Jia [13] suggested a hybrid fault diagnosis system integrating VMD, Laplace score and particle swarm optimization-based support vector machine (PSO-SVM). Motivated by the above discussions, an automated precise diagnostic system is introduced based on three fundamental blocks to detect the existing or potential bearing faults by resorting to the vibrational signals.

The most critical steps for designing the data-driven intelligent diagnostic system are usually the use of an effective strategy for FE and a powerful classification algorithm for building a predictive model during the training phase that can consequently predict the future states of the system in a timely manner. The aim of FE phase is to reveal latent and meaningful information from the vibration signals and consequently, preparing a reliable evidence for evaluating the healthy and faulty bearings [14], [15], [16]. In order to preserve important aspects of data in FE phase, it is better to extract as much information as possible and filter out the most relevant features. This approach is naturally followed by redundancy and information sparseness. Therefore, there is a need to reduce and compress the massive data resources for the ease of decision making and eventually improving the performance of the system [17], [4], [18].

The conventional FE methods such as fast Fourier transform (FFT), wavelet packet transform (WPT) and lifting wavelet transform are mostly based on basis expansion. However, their performance is greatly restricted because explicit bases are constructed only by prior information of signals and these expansion-based methods are subject to orthogonality constraint. According to the above limitations, adaptive mode decomposition (AMD) techniques have been recruited in this area due to their ability in analyzing complicated signals [19]. For a complete review on AMD methods and their application in fault diagnosis, one can refer to [20].

As one of the most widely used AMD methods, empirical mode decomposition (EMD), proposed by Huang, is well-known for its great ability in decomposing non-stationary and noisy signals which can adaptively divide a complicated signal into several intrinsic mode functions (IMFs) and a residue [19]. Albeit EMD has great adaptability in decomposing non-stationary signals, it suffers from various shortages like sensitivity to sampling, lack of mathematical idea, mode mixing problem and low efficacy due to cubic spline interpolation [20]. [21] has introduced an improved version of EMD method called ensemble empirical mode decomposition (EEMD) for alleviating the mode mixing problem of EMD and it was applied for fault detection of rolling bearings, gears and rotors. In order to overcome the envelope errors caused by the cubic spline method used in EMD, local mean decomposition (LMD) and its improved versions were introduced [22], [23].

A recently proposed AMD method attracting great attention in the field of fault diagnosis is variational mode decomposition (VMD), which can adaptively decompose a signal into a set of finite band modal functions, where their center frequencies are frequently updated and all modes are concurrently extracted [24], [25], [26]. Compared with EMD, EEMD and LMD, this entirely non-recursive decomposition technique which is based on the spectrum segmentation seems to be more suitable showing superior performance in signal decomposition [27], [28], [24], [20].

Actually, many researchers use the above-mentioned adaptive-based decomposition approaches for diagnosing faults based on the spectra domain analysis which increases the complexity of methods due to extracting the fundamental, sub-harmonics and super-harmonics components and analyzing the signal responses. Thus, it is critical to introduce automatic reliable adaptive-based algorithms to reduce the interventions of humans for determining the system condition. Various papers have been published to motivate and support this hypothesis [12], [13], [29], [30].

As we discussed earlier, by using a dimensionality reduction (DR) block, we extract low-dimensional feature information. Given that all the DR techniques even the good ones show strong variations in their outcomes among different data sets, choosing the proper DR technique has always been a challenging task to reach the desired performance level. In this regard, we examine a large collection of DR techniques with different hyper-parameters to find the best one which leads to satisfactory accuracy and precision.

The classification block is the final step for the design of the proposed automatic diagnostic system. Among computational intelligence techniques, various researches have witnessed the efficiency of artificial neural networks (ANNs) as a powerful tool for diagnosing bearing faults [31], [5], [32], [33], [30], [34], [35]. However, the classical neural networks usually employ the global sigmoid functions and need to modify the network parameters over all the interacting nodes in order to achieve the desired output, which is a very time-consuming task. Moreover, the trained parameters may occasionally converge to local minima on the error surface [36], [37]. In addition, in many cases, it is not easy to attribute a physical interpretation to the model generated by training feed-forward error back-propagation neural networks. Besides, there is no guarantee for the convergence of the learning algorithms and the initial weights should be randomly selected at the beginning of the training phase which can slow down the training process. In spite of the efficiency and widespread applications of combined ANN algorithms, these hybrid systems are heuristic structures. In addition, many underlying factors serve a critical role in the effectiveness of the proposed networks, for instance, choosing an appropriate learning rate, topology and so forth.

By considering the localization property as well as transformation ability of wavelet and the high capability of NNs in learning, various types of the wavelet-based neural network have been introduced in different applications [38], [39], [40], [41], [42]. Therefore, in this work, we utilize an efficient method called wavelet neural network (WNN), in which neurons are activated by wavelets as activation functions, in order to make use of their benefits. The main features of this three-layer network are universal approximation property, high compression capability, fast convergence and local analysis. Additionally, for orthogonal wavelets, the nodes can be added to or removed from the trained network without retraining prior to testing [43], [44], [45], [40].

Compared to the existing results, the main contributions of this paper are summarized as follows: A novel and computable fault diagnosis system (27 integrated diagnostic schemes) is proposed with the contribution of VMD and WNN to automatically opt the foremost relevant features and to diagnose bearing faults with improved accuracy of 99.48% (see Table 6, Table 7). This paper also provides a flexible framework for hybrid methodologies which interacts with the classifier structure mainly because: (1) we examine projective methods and manifold modeling to remove redundant features. This increases the generalization ability of the proposed diagnostic system. (2) We also show that how VMD can achieve better performance compared to the wavelet-based methods through its unique features. (3) Taking the advantages of WNN as a classifier in this structure provides two main benefits: guaranteeing the convergence of the learning algorithms and enhancing the transparency of the models compared to the standard neural networks.

To validate the performance of the proposed diagnostic framework, two approaches are considered: (1) comparing the accuracy of the proposed hybrid system with existing methodologies under the same benchmark database. (2) considering a second case study with three class of bearing failures. This analysis is performed with respect to the Case Western Reserve University (CWRU) bearing datasets which has been a standard and widely used database in recent years [46], [47]. The results of both tests confirm the high performance of the proposed framework for diagnosing bearing faults.

The remainder of this paper is organized as follows: The description of the bearing dataset and the main objective of this work are represented in Section 2 and Section 3 respectively. Section 4 presents a complete description of the proposed intelligent system including FE, DR and classification approaches. Experimental results and extensive comparison results on the system’s performance against other existing methodologies are reported in Section 5. Finally, concluding remarks and future research can be found in Section 6.

Section snippets

Experimental setup and dataset

In this work, the experiment was carried out on shaft-end and fan-end bearings stand consisting of a 380 V, 1.2 kW, 50 Hz, three-phase four-pole induction motor and the nominal rotation speed of the motor is 1400 rpm [5].

The experiment was conducted with a smart experimental setup comprising (1) a three-phase induction motor; (2) Safety cover; (3) Test bearings with 6205-2z symbol and the characteristics listed in Table 1; (4) B&K 4395 accelerometer; (5) Shaft; (6) Loader; (7) Base; (8) Rubber

The proposed diagnostic system

This work proposes a diagnostic system for fault detection in ball bearings. Various states are considered into account that can be categorized into healthy (H), inner race defect (IRD) and outer race defect (ORD). The latent features of these states can be extracted through the pre-processing of the raw vibrational signals, which is a vital step in the process of diagnosing bearing faults in IMs.

The bearing vibration signals are noisy and nonlinear and usually have a non-stationary behavior,

Main components of the combined diagnostic system

In this paper, a combined system is proposed, which comprises three units: FE, DR and decision making, whose details are given in the next subsections. FE plays a key role in data processing that reveals hidden characteristics of the vibration signal. The proposed signal processing procedure is illustrated in Fig. 2. Initially, the measured signals are segmented and directed to the feature extraction unit. The next subsection includes three state-of-the-art signal processing methods introduced

Experimental results and discussion

Fig. 2 shows the methodological aspects used in this study, which mainly includes FE, DR and classification phase. We examined our approach with the dataset described in section 2 according to the combined diagnostic system reported in Section 4, which aims to find the best intelligent model based on classification performance. In the proposed setting, the acceleration sensor mounted on the bearing housing is used to collect the bearing vibration signals. The experimental signal of the bearing

Conclusion

The precise diagnosis of faults for rotating machinery at an early stage provides a key opportunity to minimize operational downtime and alleviating the maintenance costs [48]. In current work, a simple, reliable and accurate neuro-wavelet-based diagnostic system was designed for identifying bearing faults, in which a time–frequency domain demodulation module is defined to characterize elemental-fault information.

Then, a DR module is proposed to filter out posterior features, remove redundant

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References (66)

  • A.A. Abdoos et al.

    Combined vmd-svm based feature selection method for classification of power quality events

    Appl. Soft Comput.

    (2016)
  • J. Zarei et al.

    Vibration analysis for bearing fault detection and classification using an intelligent filter

    Mechatronics

    (2014)
  • G.H. Bazan et al.

    Stator fault analysis of three-phase induction motors using information measures and artificial neural networks

    Electric Power Syst. Res.

    (2017)
  • R. Liu et al.

    Artificial intelligence for fault diagnosis of rotating machinery: A review

    Mech. Syst. Signal Process.

    (2018)
  • Y. Ding et al.

    Intelligent fault diagnosis for rotating machinery using deep q-network based health state classification: A deep reinforcement learning approach

    Adv. Eng. Inform.

    (2019)
  • A.K. Alexandridis et al.

    Wavelet neural networks: A practical guide

    Neural Networks

    (2013)
  • M.M. Ardestani et al.

    Human lower extremity joint moment prediction: A wavelet neural network approach

    Exp. Syst. Appl.

    (2014)
  • Z. Farajzadeh et al.

    Energy intensity and its components in Iran: Determinants and trends

    Energy Econ.

    (2018)
  • Z. Ebrahimi et al.

    Prediction therapy outcomes of hcv patients treated with interferon/ribavirin

    Biomed. Signal Process. Control

    (2018)
  • B.S. Saljooghi et al.

    Comparison of wavenet and ann for predicting the porosity obtained from well log data

    J. Petrol. Sci. Eng.

    (2014)
  • V. Zakeri et al.

    Developing real-time wave-net models for non-linear time-varying experimental processes

    Comput. Chem. Eng.

    (2009)
  • A. Safavi et al.

    Application of wavelet-based neural networks to the modelling and optimisation of an experimental distillation column

    Eng. Appl. Artif. Intell.

    (1997)
  • Y. Lei et al.

    Eemd method and wnn for fault diagnosis of locomotive roller bearings

    Exp. Syst. Appl.

    (2011)
  • X.-J. Wan et al.

    Fault diagnosis of rolling bearing based on optimized soft competitive learning fuzzy art and similarity evaluation technique

    Adv. Eng. Inform.

    (2018)
  • C. Lu et al.

    Intelligent fault diagnosis of rolling bearing using hierarchical convolutional network based health state classification

    Adv. Eng. Inform.

    (2017)
  • H. Qiu et al.

    Robust performance degradation assessment methods for enhanced rolling element bearing prognostics

    Adv. Eng. Inform.

    (2003)
  • R. Yan et al.

    Wavelets for fault diagnosis of rotary machines: A review with applications

    Signal Process.

    (2014)
  • Y. Tian et al.

    A subspace learning-based feature fusion and open-set fault diagnosis approach for machinery components

    Adv. Eng. Inform.

    (2018)
  • X. Zhang et al.

    Time–frequency analysis for bearing fault diagnosis using multiple q-factor gabor wavelets

    ISA Trans.

    (2018)
  • V. Migal et al.

    Reducing the vibration of bearing units of electric vehicle asynchronous traction motors

    J. Vib. Control

    (2020)
  • R. Razavi-Far et al.

    An integrated class-imbalanced learning scheme for diagnosing bearing defects in induction motors

    IEEE Trans. Industr. Inf.

    (2017)
  • S. Anwar et al.

    An analytical redundancy-based fault detection and isolation algorithm for a road-wheel control subsystem in a steer-by-wire system

    IEEE Trans. Veh. Technol.

    (2007)
  • H. Hao, Key performance monitoring and diagnosis in industrial automation processes, Ph.D. thesis...
  • Cited by (0)

    View full text