A deep transfer maximum classifier discrepancy method for rolling bearing fault diagnosis under few labeled data

https://doi.org/10.1016/j.knosys.2020.105814Get rights and content

Abstract

Rolling bearing fault diagnosis is closely related to the safety of mechanical system. In real-world diagnosis, it is difficult to obtain abundant labeled data due to varying operation conditions, complex working environment and inevitable indirect measurement, which will affect the ability of diagnosing. To tackle this problem, a deep transfer maximum classifier discrepancy method is proposed under few labeled data, which utilizes fully deep learning and transfer learning. Firstly, a batch-normalized long-short term memory (BNLSTM) model which can learn the mapping relationship between two kinds of datasets is designed to generate some auxiliary samples. Then, a transfer maximum classifier discrepancy (TMCD) method, which considers the characteristics of each data type by an adversarial strategy, is applied to align probability distributions of auxiliary samples generated by BNLSTM and unlabeled data from target domain. Sufficient experimental results indicate the effectiveness of the proposed method under few labeled data.

Introduction

Rolling bearing widely used in modern industry is one of the critical component of rotating machinery. Rolling bearing fault diagnosis can not only ensure the smoothness and effectiveness of mechanical equipment, but also make for detecting and eliminating unexpected fault in time. Rolling bearing fault diagnosis is conducive to prevent the occurrence of major accidents to a certain extent. Therefore, more and more scholars aim to research it in recent years [1], [2], [3].

In past decades, artificial neural network (ANN) and support vector machine (SVM) as two typical traditional intelligent diagnosis methods are utilized frequently to fault recognition [4], [5]. ANN and SVM have made some achievements on the basis of manual feature extraction. For instance, Kumar et al. applied ANN to diagnose three kinds of health status in the case of extracting the statistical features from the dominant wavelet coefficients [6]. Dhamanda et al. used ANN to detect combined gear-bearing fault in single stage spur gear box after extracting representative features [7]. Zhou et al. extracted features based on hierarchical entropy and utilized SVM to diagnose rolling bearing fault [8]. Chen et al. constructed a multi-kernel SVM with chaotic particle swarm optimization, which employed local tangent space alignments to select salient features [9]. To get satisfying classification accuracies, it is significant to extract representative features. On one hand, manual feature extraction relies on the engineering experience and takes lots of time, on the other hand, the generalization of extracted features is difficult to guarantee. These weaknesses limit the further application of traditional intelligent diagnosis methods in rolling bearing fault diagnosis.

To avoid manual feature extraction, deep learning has been introduced to address this problem in the last few years. Deep learning is able to excavate automatically data features, which greatly improves recognition accuracies and extremely promotes the development of intelligent diagnosis methods [10], [11]. With the advent of deep learning using a large amount of available labeled data, rolling bearing fault diagnosis has achieved a great deal of successes. Shao et al. designed an adaptive deep belief network (DBN) with dual-tree complex wavelet packet to diagnose rolling bearing fault [12]. Yang et al. used hierarchical symbolic analysis and convolutional neural network (CNN) for fault diagnosis of rotating machinery, which performed superior classification accuracies compared with shallow methods [13]. Yang et al. proposed an intelligent fault diagnosis scheme based on LSTM for rotating machinery fault diagnosis [14]. Yu et al. applied a deep stacked auto-encoder (SAE) to gearbox fault diagnosis [15]. Though, deep learning methods have made massive achievements under a general assumption: the training and the test data are drawn from the same feature space and the same distribution [16]. In some cases especially in real-world applications, it is hard or even impossible to recollect abundant labeled fault data due to complicated and changeable working environment. However, the knowledge from relevant data is often underutilized. Therefore, how to use few labeled data to diagnose rolling bearing fault is of great significance.

Transfer leaning is a prospective method to figure out this problem. There are two kinds of domain in transfer leaning: source domain and target domain. The essence of transfer learning in rolling bearing fault diagnosis is to diagnose target domain data with the help of the knowledge from source domain data. Source domain and target domain are different but related to each other, the distribution of source domain data have some differences from that of target domain data [17]. So far, transfer learning has been applied to emotion analysis [18], text recognition [19] as well as fault classification [20]. For example, Wen et al. designed a three-layer SAE to extract features of training data and test data, which came from different working conditions of the same bearing [21]. Guo et al. applied a deep convolution transfer learning method to bearing fault recognition [22]. Hasan et al. introduced acoustic spectral imaging technology to fault recognition, which transformed raw signals into spectrum imaging [23]. However, their proposed methods confront with some weaknesses as follows. Firstly, most of previous studies still needed sufficient labeled data and extracted features from frequency spectrum data rather than raw vibration data. Secondly, transfer knowledge is only transferred from one working condition to another. Thus, the close connection between different working conditions makes it possible to get great classification accuracies. Finally, the generalization abilities of their proposed methods have not been demonstrated due to the data collected for transfer learning in the same bearing in their methods.

To weaken the impact of these mentioned weaknesses and take extremely advantage of the knowledge of few labeled data from target domain. A deep transfer maximum classifier discrepancy method under few labeled data is proposed for rolling bearing fault diagnosis. In the proposed method, a batch-normalized long-short term memory (BNLSTM) model is designed to learn the mapping relationship of some labeled source domain data and few labeled target domain data. Then, generating some auxiliary samples using other labeled source domain data based on the pre-trained BNLSTM. Finally, transfer maximum classifier discrepancy (TMCD), which conducts an adversarial strategy that utilizes two different classifiers, is applied to align distribution between generated auxiliary samples and unlabeled target domain data. The superiority of the proposed method is demonstrated by sufficient comparative experiments. The main contributions of this paper are generalized as follows.

  • (1)

    To solve the fault recognition issue under few labeled data and fully use more information from abundant labeled source domain data, some auxiliary samples are generated by BNLSTM which builds the relationship about labeled source domain data and that of target domain.

  • (2)

    To construct a model that can work well on unlabeled target domain data, TMCD is introduced to align probability distributions of generated auxiliary samples and unlabeled target source data.

The former provides an approach when it comes to diagnosing target domain data under few labeled data. The latter can greatly improve classification accuracies by an adversarial manner, which considers the characteristics of each data type from different domains.

The arrangement of this paper is as follows. In Section 2, fundamental theories of LSTM and BN are introduced. Section 3 introduces the proposed method in details. In Section 4, sufficient experiments are conducted and results are analyzed. The general conclusions are summarized in Section 5.

Section snippets

The brief introduction of transfer learning

Traditional diagnosis methods promote the development of fault diagnosis, however, manual feature extraction limits its further applications. The advent of deep learning methods avoids manual feature extraction. The successes of deep learning methods and traditional methods could be attributed to the common prerequisite: the training and test data are obtained from same feature space and same distribution. However, it is difficult to meet mentioned prerequisite in practical applications.

The proposed method

In this paper, a deep transfer maximum classifier discrepancy method under few labeled data is proposed in which a pre-trained BNLSTM model generates some auxiliary samples and TMCD is applied to align distribution of generated auxiliary samples and unlabeled Xtar by using two different classifiers. Section 3 consists of three subsections. Section 3.1 describes the construction details of the BNLSTM model. In Section 3.2, TMCD is elaborately introduced. Section 3.3 depicts the general

Datasets introduction

To validate the feasibility and effectiveness of the proposed method, two different datasets obtained from two different but relevant domains are utilized to carry out sufficient bearing fault diagnosis experiments. Hereafter, dataset A as Xsrc is provided by Case Western Reserve University (CWRU) [40], dataset B as Xtar is collected from real-world railway locomotive. The two datasets are acquired from different bearings and various operation environments. Concrete information of two datasets

Conclusions

In this paper, a deep transfer maximum classifier discrepancy method is proposed to address bearing fault diagnosis problem under few labeled data. Firstly, the proposed method uses the knowledge from only few labeled target domain data to generate some auxiliary samples, and then adopts an adversarial strategy that introduces two different classifiers to classify fault types. Sufficient experiments results demonstrate that combining BNLSTM and TMCD is prospective for detecting bearing fault

CRediT authorship contribution statement

Zhenghong Wu: Methodology, Software, Investigation, Writing - original draft. Hongkai Jiang: Conceptualization, Formal analysis, Writing - review & editing. Tengfei Lu: Validation, Formal analysis, Software. Ke Zhao: Resources, Supervision.

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 research is supported by the major research plan of the National Natural Science Foundation of China (No. 91860124), the National Natural Science Foundation of China (No. 51875459) and the Aeronautical Science Foundation of China (No. 20170253003).

References (44)

  • ShaoH.D. et al.

    Rolling bearing fault diagnosis using adaptive deep belief network with dual-tree complex wavelet packet

    ISA Trans.

    (2017)
  • YangY.T. et al.

    A fault diagnosis scheme for rotating machinery using hierarchical symbolic analysis and convolutional neural network

    ISA Trans.

    (2019)
  • YangR. et al.

    Rotating machinery fault diagnosis using long-short-term memory recurrent neural network

    IFAC-PapersOnLine

    (2018)
  • QianW.W. et al.

    A novel transfer learning method for robust fault diagnosis of rotating machines under variable working conditions

    Measurement

    (2019)
  • HasanM.J. et al.

    Acoustic spectral imaging and transfer learning for reliable bearing fault diagnosis under variable speed conditions

    Measurement

    (2019)
  • WuZ.H. et al.

    An adaptive deep transfer learning method for bearing fault diagnosis

    Measurement

    (2020)
  • LiX.Q. et al.

    Rolling bearing health prognosis using a modified health index based hierarchical gated recurrent unit network

    Mech. Mach. Theory

    (2019)
  • ZhangB. et al.

    Bearing performance degradation assessment using long short-term memory recurrent network

    Comput. Ind.

    (2019)
  • WangJ.R. et al.

    Batch-normalized deep neural networks for achieving fast intelligent fault diagnosis of machines

    Neurocomputing

    (2019)
  • LeiJ.H. et al.

    Fault diagnosis of wind turbine based on long short-term memory networks

    Renew. Energy

    (2019)
  • YangR. et al.

    Rotating machinery fault diagnosis using long-short-term memory recurrent neural network

    IFAC PapersOnLine

    (2018)
  • ShaoH.D. et al.

    Enhanced deep gated recurrent unit and complex wavelet packet

    Knowl.-Based Syst.

    (2019)
  • Cited by (69)

    View all citing articles on Scopus
    View full text