Elsevier

Measurement

Volume 184, November 2021, 109885
Measurement

Rolling bearing fault diagnosis based on multi-channel convolution neural network and multi-scale clipping fusion data augmentation

https://doi.org/10.1016/j.measurement.2021.109885Get rights and content

Highlights

  • Multi-scale clipping fusion guarantees the use of CNN with relatively limited data samples.

  • Multi-channel CNN makes the best of sensor data and alleviate data heterogeneity.

  • MCNN with MSCF is suited for bearing fault diagnosis under variable working condition.

Abstract

Deep learning has evolved to a prevalent approach for machinery fault diagnosis in recent years. However, the high demanding for training data amount refrains its implementation. In this study, we proposed a novel rolling bearing fault diagnosis strategy based on multi-channel convolution neural network(MCNN) combining multi-scale clipping fusion(MSCF) data augmentation technique. The fault signals were augmented using MSCF before transformed to time–frequency images through short-time Fourier transform, then the multi-sensor derived image data were fused by MCNN for feature extraction and fault pattern classification. Experiments validate that the combination of MSCF and MCNN is good at making the best of the information contained in each single sensor recording, leading to a significantly improved fault pattern classification accuracy and cluster effect. The proposed approach is low complexity but effective and robust, it is well suited for bearing fault diagnosis in case limited sensor data and/or variable working condition is presented.

Introduction

As an key component of rotating machinery, such as wind power equipment, aerospace engine, high-precision CNC machine tools [1], [2], rolling bearings are likely to suffer various faults due to its long time operation under harsh conditions and consequently lead to unexpected breakdown of machines accounting for about 40% of all accidents [3]. Therefore early and accurate fault diagnosis of rolling bearings is critical to maintain normal running of rotating machinery. With the increasing number of monitoring positions and sampling frequency, the amount of sensor data recorded in the whole service cycle of the mechanical equipment is extremely huge, which makes fault diagnosis enter into “big data” era [4], [5], [6]. It is therefore Machine learning(ML) and data-driven machinery fault diagnosis methods have become the mainstream in the past decade [7].

The pipeline of typical ML based fault diagnosis techniques consists of three steps [4], [6]: firstly, signal pre-processing is performed to reduce noise and normalize the multi-source fault signals; secondly, feature extraction is executed from the preprocessed signals to reflect the health conditions of machines. Thirdly, pattern recognition is employed by feeding the extracted fault features into a classifier to predict faults types. Typical classification techniques used for fault diagnosis include artificial neural networks (ANN), support vector machines (SVM), decision tree, random forest, k-nearest neighbor(KNN) and extreme learning machine(ELM) etc [8], [9], [10], [11], [12], [13]. Implementation of these methods usually requires the handcrafted feature extraction from fault signal in time domain, frequency domain, time–frequency domain and nonlinear domain. Specifically, nonlinear features represented by various definition of entropy and time–frequency features represented by Hilbert-Huang transform(HHT), short-time Fourier transform(STFT), and continuous wavelet transform(CWT) have been widely exploited [8], [12], [14], [15], [16], [17]. Among them, Zheng et al. [8], [12] applied ensemble SVM and ELM to identify the rolling bearing faults by extracting multi-scale fuzzy entropy and weighted multi-scale variation entropy from vibration fault signal. Pan et al. [13] proposed an ELM based two-stage method to predict the remaining useful life of rolling-element bearings. Wang et al.[14] used ensemble local mean decomposition and fast kurtogram-based time–frequency analysis to detect rotating machine failures. Yu et al. [16] employed HHT and time–frequency entropy to detect the fault of gear. Sun et al. [15], [18] extracted the bearing fault features using modified orthogonal matching pursuit algorithm and exploited structured sparse time–frequency analysis to achieve a good gear faults detection. Methods above are easy to implemented and have proved effective but still show some drawbacks in: 1) manual feature extraction from noised signals requires great deal of prior knowledge; 2) The three sections of fault diagnosis process is somewhat isolated and likely to lose the coupling information, resulting in low accuracy and weak generalization.

As a new branch of ML, the use of deep learning(DL) algorithm in machinery fault diagnosis research has increased exponentially over the past decade due to its high accuracy and efficient data representation, automatic feature extraction and selection [19]. Unlike traditional ML, DL algorithms express complex objective functions by stacking multi-level structures to discover the intrinsic relationship between variables and improve model generalization performance [4], [19], [20]. Typical DL algorithms used in machinery fault diagnosis include auto-encoder(AE) [21], [22], [23], [24], [25], convolution neural network (CNN) [26], [27], [28], [29], [30], [31], deep belief network(DBN) [24], [30], deep residual network(DRN) [32], [33], and recursive neural network(RNN)or its variant long short time memory network(LSTM) [34], [35]. Among the all DL algorithms, CNN and its improved versions has attracted more attention in machinery fault diagnosis. In practice, CNN is best suited to combine with other ML/DL methods, such as ELM [29], DBM [30], RNN [35] and DRN [32], [33], to improve the network performance. Particularly, time–frequency techniques such as STFT [28], [29], [33] and CWT[26] are usually employed to convert the 1D sensor signal to 2D image format as the input of CNN. Although CNN can produce satisfactory performance in machinery fault diagnosis, its network training is often a challenge in that a large training samples is needed to avoid over-fitting in testing samples [36]. Therefore it necessitate the employment of data augmentation techniques in case the amount of raw data is not large enough.

To the best of our knowledge, in the field of rotating machinery fault diagnosis, early work were mainly focused on constant operating conditions which is hard to reflect the overall running state of equipment. In real practice, however, it is common to face the issue that the equipments frequently change the working conditions with different loads and rotation speeds, leading to a limited or imbalanced fault data on each single working condition and an increased discrepancy of data distribution. Recently, transfer learning(TL) partnered with CNN provide an effective way to solve the problem of cross-domain fault diagnosis, where the training and testing data are collected under different working conditions [7], [37], [38], [39], [40], [41]. In addition, in the past couple of years, multi-modal sensor data fusion strategy based on 1D or 2D CNN is also widely used to comprehensively describe the state of equipment and facilitate the implementation of fault diagnosis under variable working conditions [31], [42], [43], [44], [45].

Based on the inspiration of the aforementioned studies, this paper introduce a novel approach combining data augmentation and multi-channel technique that can give full play to CNN in bearings fault diagnosis in case ‘big data’ is not available and variable working condition is presented. The implementation of the proposed approach can be summarized as follows. First, taking the three accelerometer sensors of bearing as signal sources, the raw fault data was augmented by so called multi-scale clipping fusion(MSCF), and the data collected under different motor loads was mixed to simulate the operation state of variable working conditions. Then STFT is used to convert the processed data into time–frequency images so as to obtain discriminative features of diverse sources of failures. Finally, the multi-sensor recording derived image data is fused by multi-channel CNN(MCNN) to comprehensively represent the state of equipment and consequently improve the overall classification performance.

The rest of this paper is arranged as follows: in section 2, the introduction of STFT, MCNN, MSCF data augmentation, fault diagnosis procedure are presented in sequence; section 3 is the description of selected datasets; experimental results and corresponding discussions are presented in section 4 to validate the effectiveness of the proposed method. Finally, some conclusions are given in section 5.

Section snippets

Short-time Fourier Transform(STFT)

STFT is a typical time–frequency analysis technique for processing non-stationary stochastic signals. The STFT of 1D time series signal is 2D spectrum representing the joint information both in time domain and frequency domain. Essentially, STFT can be regarded as Fourier transform with a fix length time window function, as the window slides, the Fourier transform at different time points can be obtained. The definition of STFT is given follow:STFTX(τ,f)=-+x(t)h(t-τ)e-jωtdt

Where x(t) denotes

Dataset 1(CWRU dataset)

Dataset 1 is downloaded from Bearing Data Center website at Case Western Reserve University(CWRU). CWRU dataset is the most classical public source and has become the standard reference to verify different fault diagnosis methods in the past decade. The model of faulty rolling bearings on CWRU experimental platform is 6205–2 RS JEM SKF. The apparatus for data collection is shown in Fig. 6. Relevant parameters are shown in Table 2.

As is presented in Fig. 6, the CWRU experiment apparatus includes

Results and discussions

All experimental results in this study were produced by running programs on a ThinkPad laptop with Windows10 64 bit Operation System, Intel(R) Core(TM) [email protected] GHz CPU and 24 GB RAM. Data augmentation and STFT time–frequency images were achieved through MATLAB 2018a. The CNN models were implemented by Python 3.7 compiler in Keras framework using TensorFlow as a backend.

Conclusions

The random variation of working load or rotation speed of mechanical equipment could lead to the poor consistency of collected data, which deteriorates the process of feature extraction and fault diagnosis; on the other hand, limited sample data restricts the application of DL in machinery fault diagnosis and data from single-marking sensor can not fully reflect the overall operation of the equipment.

We introduced a novel approach combining multi-channel CNN with MSCF data augmentation for the

CRediT authorship contribution statement

Ruxue Bai: Writing - original draft, Validation, Visualization, Writing - review & editing. Quansheng Xu: Formal analysis, Visualization, Writing - review & editing. Zong Meng: Conceptualization, Methodology, Investigation, Supervision, Validation. Lixiao Cao: Writing - review & editing, Validation. Kangshuo Xing: Data curation, Software. Fengjie Fan: Data curation, Writing - review & editing.

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.

Acknowledgments

This work was supported by the National Natural Science Foundation of China (No. 52075470, 61873227), the Natural Science Foundation of Hebei Province (No. E2019203448) and the Central government guides local science and technology development Foundation (No. 206Z4301G).

References (46)

  • R. Sun et al.

    Sparse representation based on parametric impulsive dictionary design for bearing fault diagnosis

    Mech. Syst. Sig. Process.

    (2019)
  • J.K. Sinha et al.

    A future possibility of vibration based condition monitoring of rotating machines

    Mech. Syst. Sig. Process.

    (2013)
  • X. Zhao et al.

    Deep Laplacian Auto-encoder and its application into imbalanced fault diagnosis of rotating machinery

    Measurement

    (2020)
  • Z. Meng et al.

    An enhancement denoising autoencoder for rolling bearing fault diagnosis

    Measurement

    (2018)
  • X. Kong et al.

    A multi-ensemble method based on deep auto-encoders for fault diagnosis of rolling bearings

    Measurement

    (2020)
  • H. Shao et al.

    A novel method for intelligent fault diagnosis of rolling bearings using ensemble deep auto-encoders

    Mech. Syst. Sig. Process.

    (2018)
  • P. Liang et al.

    Compound Fault Diagnosis of Gearboxes via Multi-label Convolutional Neural Network and Wavelet Transform

    Comput. Ind.

    (2019)
  • L. Jing et al.

    A convolutional neural network based feature learning and fault diagnosis method for the condition monitoring of gearbox

    Measurement

    (2017)
  • Y. Zhang et al.

    An enhanced convolutional neural network for bearing fault diagnosis based on time–frequency image

    Measurement

    (2020)
  • Z. Chen et al.

    Mechanical fault diagnosis using Convolutional Neural Networks and Extreme Learning Machine

    Mech. Syst. Sig. Process.

    (2019)
  • H. Shao et al.

    Rolling bearing fault feature learning using improved convolutional deep belief network with compressed sensing

    Mech. Syst. Sig. Process.

    (2018)
  • D. She et al.

    Wear indicator construction of rolling bearings based on multi-channel deep convolutional neural network with exponentially decaying learning rate

    Measurement

    (2019)
  • W. Zhang et al.

    Deep residual learning-based fault diagnosis method for rotating machinery

    ISA Trans

    (2019)
  • Cited by (83)

    View all citing articles on Scopus
    View full text