Elsevier

Neurocomputing

Volume 306, 6 September 2018, Pages 119-129
Neurocomputing

Real-time incipient fault detection for electrical traction systems of CRH2

https://doi.org/10.1016/j.neucom.2018.04.058Get rights and content

Abstract

Electrical traction systems in a high-speed train are the core parts to provide traction force for the whole train. Due to performance degradation of electronic components and the prolonged operation under variously complicated operating environments, incipient faults will inevitably happen and will evolve into faults or failures if they are not successfully detected. Currently, the univariate control charts are used to monitor electrical traction systems of high-speed trains. However, this primitive solution is unable to deal with incipient faults with satisfactory performance. In this paper, a Kullback–Leibler divergence (KLD) and independent component analysis (ICA)-based method is proposed to perform incipient fault detection (FD) in electrical traction systems. Compared with the existing ICA-based methods, the proposed strategy is more sensitive to incipient faults; meanwhile it has low computational load because estimating the probability density functions (PDFs) of the derived independent components and the residuals is avoided. On the experimental platform of the traction system for China Railway High-speed 2-type (CRH2) trains, three typical incipient faults are successfully injected, and the proposed method is successful in detecting these incipient faults.

Introduction

The electrical traction system of a high-speed train provides traction force for the whole train via the highly efficient information transmission systems with highly integrated net communication, online monitoring, and control technologies. Successful detection of incipient faults is of fundamental importance to ensure optimum reliability and maximum safety. Incipient faults in traction systems are often caused by some inevitable reasons, such as insulation degradation of the windings [1], aging of mechanical and electronic components [2], prolonged operations [3], etc. When the high-speed train is running under multiple operations together with uncertain noises, the effects caused by incipient faults are not obvious and therefore their features are difficult to be extracted [4]. In recent years, detection and diagnosis of incipient faults in high-speed trains are active but difficult issues in academic research and practical applications [1], [2], [3], [5], [6].

Until now, there is no standard definition for the incipient faults. In the existing literature [7], [8], [9], [10], several common characteristics of incipient faults can be summarized as follows: (i) the slight anomalies caused by incipient faults are such small that they will not trigger any presetting alarms; (ii) as time goes on, incipient faults will evolve into failures if no preventive actions are taken; (iii) incipient faults are conceivable in the constant deviations in one short-time window because of their slow variation trends. Therefore, comparing with detection and diagnosis of regular faults, more attention should be paid on improving sensitivity to the incipient fault features.

Some efforts on incipient fault detection and diagnosis (FDD) in electrical traction systems are reported in [1], [2], [3], [5], [6], [11], [12]. The dominant method is model-based FDD where precise physical models of electrical systems are required [13,14]. For example, for detecting and isolating incipient voltage-sensor faults, sliding mode and adaptive estimation techniques are proposed in [6]. Similarly, based on the model of a high-speed train, a torque and speed-based observer was designed for 1.2 MW traction motor to detect incipient faults [11]. Following this work, a nonlinear observer was developed to detect faults in the d-q frame in [12]. However, in practice, it is extremely difficult to obtain accurate mathematical models of high-speed trains, which prevents the adoption of the aforementioned theoretical research results into real-world applications.

Another popular FDD method in electrical systems is signal processing techniques [13], including fast Fourier transform [5], wavelet transform [15], artificial intelligence [16,17], etc. Using signal processing techniques, the FDD can be achieved by directly analyzing the collected signals, such as voltage, current, speed, vibration, and external magnetic field. As a matter of fact, the classical signal-analysis-based FDD methods may not be suitable for detecting incipient faults [9].

Thanks to the great advancement of sensor, data storage, and data analysis techniques, the other strategies for electrical traction systems can be generalized into data-driven methods which have been a new research avenue [18]. They were originated from chemical engineering, and recent development on the data-driven FDD approaches for largely industrial systems is reviewed in [19], [20], [21], [22], [23]. For mechatronic systems, a kernel principal component analysis (PCA) method for detection of incipient faults is developed for multiple modes of traction systems caused by switching insulated-gate bipolar transistors (IGBTs) [2]. Aiming at the non-Gaussian current and voltage signals, the original signals are projected into the rotatory principal component and residual subspaces under the PCA framework for better incipient-fault estimation [7]. For the incipient air-brake faults in electric units, two smoothing techniques integrated with PCA are analyzed in details in [24]. Besides, deep PCA, by sufficiently capturing the information of both systematic variation and noise, is used to detect and diagnosis incipient sensor faults in traction systems [10]. For monitoring some specific faults, fault-relevant PCA is proposed in [25] which cooperates with fault information. The improved fault sensitivity can be hence achieved via mining the information hidden in normal data and fault data. Whilst, the artificial intelligent based methods, such as the multiple layer artificial neural network [26], using normal data and fault data are also popular to diagnose incipient faults. If the fault data can be not easily obtained, a limitation on the application of methods depending on fault data will be posed. Then, a sparse dissimilarity analysis algorithm without any priori on fault information is developed in [27] to monitor and isolate incipient faults.

However, as pointed in [18], [21], PCA-based FDD methods only show its advantages when the measurements subject to Gaussian distributions. In other words, the above mentioned data-driven methods for detection of incipient faults in electrical systems are not exactly effective because of the non-Gaussian characteristics of measurements in traction systems of high-speed trains.

In recent ten years, a very influential method is independent component analysis (ICA) which is more advantageous than the PCA scheme because it can not only deal with non-Gaussian signals perfectly but also involve high-order statistics of signals [28]. Based on its advantages, classic ICA-based FDD [29] and its variant [30] are employed to detect and diagnose faults in non-electrical areas. However, these methods are unavailable to some extent if the fault’s amplitude is small, and applications on electrical systems are deficient at present.

In this paper, a new FD method for incipient faults, KLD-ICA-based FD method, is proposed to deal with the incipient faults in non-Gaussian electrical traction systems of high-speed trains. The ICA is adopted for extracting non-Gaussian latent variables while Kullback–Leibler divergence (KLD) is used to construct effective alarms for the detection of incipient faults. It does not require mathematical model of electrical traction systems and can overcome potential difficulties associated with multiple operation conditions and uncertainties caused by unknown noises and disturbances. Compared with the existing data-driven FD methods used in electrical traction systems, the proposed method has the following advantages:

  • 1.

    By integrating KLD into ICA framework, it is not only more sensitive to incipient faults than the existing methods, but also suitable for non-Gaussian traction system as expected.

  • 2.

    It works without the need for direct estimation of PDF of each latent independent components (ICs), which can achieve a remarkable computational cost procedure.

  • 3.

    A randomized algorithm is used to determine the thresholds, which can ensure the best fault detectability.

The rest of this paper is organized as follows. In Section 2, the electrical traction system of China Railway High-speed 2-type (CRH2) and ICA are briefly described, then the incipient FD problem is formulated. In Section 3, the proposed method is given in details, including KLD-ICA-based FD, computational cost reduction and determination of optimal thresholds. Furthermore, the proposed method is verified on an experimental platform of the traction system of CRH2. Comparative analysis with other data-driven methods is presented in Section 4. Finally, the conclusions are drawn in Section 5.

Section snippets

Preliminaries

In this section, the non-Gaussian electrical drives in CRH2 and the fundamentals of ICA-based FD method are first introduced, followed by the objective and design issues for incipient FD method.

The proposed FD methodology

In this section, the proposed method is discussed in details, including the KLD-based test statistics and the determination of appropriate thresholds.

Applications to electrical traction systems in CRH2

In this section, the proposed FD method for incipient faults is applied to a CRH2 train. It is developed by CRRC Zhuzhou Institute Co., Ltd., on which the incipient sensor fault injections can be easily achieved by simple modifications.

Conclusions

With the introduction of KLD in this paper, a new online incipient FD method under ICA frame has been studied and applied to the electrical traction systems of high-speed trains. It uses the higher-order statistical information of system signals, and is of high computational efficiency by avoiding calculating PDFs. Tests with three incipient faults have been carried out where the proposed KLD-ICA-based method is shown to be effective. Further work will be focused on incipient fault diagnosis

Hongtian Chen received the B.S. and M.S. degrees in School of Electrical and Automation Engineering from Nanjing Normal University, China, in 2012 and 2015, respectively, and he is currently pursuing the Ph.D. degree from the College of Automation Engineering in Nanjing University of Aeronautics and Astronautics, China.

His research interest covers incipient fault detection and diagnosis under multivariate statistical frame, machine learning based incipient fault detection and diagnosis, and

References (38)

  • D. Barater et al.

    Incipient fault diagnosis in ultrareliable electrical machines

    IEEE Trans. Ind. Appl.

    (2017)
  • H. Chen et al.

    Multi-mode KPCA based incipient fault detection for PWM inverter of CRH5

    Adv. Mech. Eng.

    (2017)
  • I. Bandyopadhyay et al.

    A combined image processing and nearest neighbor algorithm tool for classification of incipient faults in induction motor drives

    Comput. Electr. Eng.

    (2016)
  • A. Romanenko et al.

    Incipient bearing damage monitoring of 940-h variable speed drive system operation

    IEEE Trans. Energy Convers.

    (2017)
  • K. Zhang et al.

    Incipient voltage sensor fault isolation for rectifier in railway electrical traction systems

    IEEE Trans. Ind. Electron.

    (2017)
  • H. Chen et al.

    Data-driven incipient sensor fault estimation with application in inverter of high-speed railway

    Math. Probl. Eng.

    (2017)
  • L. Ren et al.

    Single-sensor incipient fault detection

    IEEE Sens. J.

    (2011)
  • H. Chen et al.

    Deep PCA based real-time incipient fault detection and diagnosis methodology for electrical drive in high-speed trains

    IEEE Trans. Vehicular Technol.

    (2018)
  • J. Guzinski et al.

    Application of speed and load torque observers in high-speed train drive for diagnostic purposes

    IEEE Trans. Ind. Electron.

    (2009)
  • Cited by (27)

    • A supervised multisegment probability density analysis method for incipient fault detection of quality indicator

      2022, Journal of Process Control
      Citation Excerpt :

      In order to detect the incipient fault in complex dynamic system, the subspace identification and Kullback–Leibler divergence methods were used between the reference dataset and the online data [36]. Chen et al. proposed a Kullback–Leibler divergence and independent component analysis based method to perform the incipient fault detection [37]. In order to improve the fault sensitivity of the traditional canonical variate analysis indices, the canonical variate dissimilarity analysis (CVDA) method was proposed [38].

    • Slow feature analysis-aided detection and diagnosis of incipient faults for running gear systems of high-speed trains

      2022, ISA Transactions
      Citation Excerpt :

      In order to deal with incipient faults of high-speed trains, some studies have been focused on data-driven methods. For example, the improvement of independent component analysis (ICA) method not only overcomes the disadvantages of traditional algorithms, but also enables incipient FDD in electrical traction systems [25]. For incipient FDD in electrical drives, deep principal component analysis (DPCA) is used to extract slighter fault features.

    • Fault information mining with causal network for railway transportation system

      2022, Reliability Engineering and System Safety
      Citation Excerpt :

      Kernel PCA (KPCA) is adopted in [23] for the feature dimension reduction of the railway point machines dataset. Kullback-Leibler divergence and independent component analysis (ICA)-based method is proposed in [24] for fault detection of electrical traction systems. Previous feature extraction and section methods take into account only the correlations among monitoring variables rather than causalities.

    • Data-driven and deep learning-based detection and diagnosis of incipient faults with application to electrical traction systems

      2020, Neurocomputing
      Citation Excerpt :

      But, the integration of data-driven and deep learning-based methods for incipient FDD is not touched in the existing studies [19]. As synthetically described in [12,20–22], incipient faults are slightly unexpected deviations from system normal behaviors. If no remedial actions are taken, incipient faults will evolve into common faults and the system performance will be then risked seriously.

    View all citing articles on Scopus

    Hongtian Chen received the B.S. and M.S. degrees in School of Electrical and Automation Engineering from Nanjing Normal University, China, in 2012 and 2015, respectively, and he is currently pursuing the Ph.D. degree from the College of Automation Engineering in Nanjing University of Aeronautics and Astronautics, China.

    His research interest covers incipient fault detection and diagnosis under multivariate statistical frame, machine learning based incipient fault detection and diagnosis, and their applications to electrical traction system of high-speed trains.

    Bin Jiang obtained Ph.D. degree in Automatic Control from Northeastern University, Shenyang, China, in 1995. He had ever been postdoctoral fellow, research fellow and visiting professor in Singapore, France, USA and Canada, respectively.

    Now he is a Chair Professor of Cheung Kong Scholar Program in Ministry of Education and Dean of College of Automation Engineering in Nanjing University of Aeronautics and Astronautics, China. He currently serves as Associate Editor or Editorial Board Member for a number of journals such as IEEE Trans. On Control Systems Technology; Int. J. of Control, Automation and Systems; Nonlinear Analysis: Hybrid Systems; Acta Automatica Sinica; Control and Decision, Systems Engineering and Electronics Technologies. He is Chair of Control Systems Chapter in IEEE Nanjing Section, a member of IFAC Technical Committee on Fault Detection, Supervision, and Safety of Technical Processes. His research interests include fault diagnosis and fault tolerant control and their applications in aircrafts, satellites and high-speed trains.

    He has been the principle investigator on several projects of National Natural Science Foundation of China. He is the author of 8 books and over 200 referred international journal papers and conference papers. He won First Class Prize of Natural Science Award of Ministry of Education of China in 2015.

    Ningyun Lu received her Ph.D. degree from Northeastern University, Shenyang, China, in 2004. From 2002 to 2005, she worked as a research associate and postdoctoral fellow in Hong Kong University of Science and Technology. Currently, she is a full professor in College of Automation Engineering at Nanjing University of Aeronautics and Astronautics, Nanjing, China. Her research interest includes data-driven fault prognosis, data-driven diagnosis, health management techniques, and their applications to various industrial processes.

    Wen Chen received the Ph.D. degree from Simon Fraser University, Burnaby, BC, Canada, in 2004. From 2005 to 2007, he was a Postdoctoral Researcher at the University of Louisiana at Lafayette, Louisiana, USA. He then worked in industrial companies as a Control Systems Engineer. In 2009, he joined the Division of Engineering Technology, Wayne State University, Detroit, MI, USA. His teaching and research interests are control systems, alternative energy storage, and fault diagnosis of industrial systems.

    This work was supported by National Natural Science Foundation of China (61490703, 61374141, 61573180), Funding of Jiangsu Innovation Program for Graduate Education (KYLX16_0378).

    View full text