Elsevier

Neurocomputing

Volume 430, 21 March 2021, Pages 24-33
Neurocomputing

Fault diagnosis with synchrosqueezing transform and optimized deep convolutional neural network: An application in modular multilevel converters

https://doi.org/10.1016/j.neucom.2020.11.037Get rights and content

Abstract

High voltage direct current (HVDC) transmission mode with modular multilevel converters (MMC) topology is the future direction of transmission engineering, and security is their fundamental issue. Submodule fault of MMC in HVDC is the most common problem, nevertheless, traditional time–frequency based diagnosis technology can’t achieve high accuracy. To solve this pain spot, a new diagnosis strategy based on the synchrosqueezing transform (SST) and genetic algorithm optimized deep convolution neural network (GA-DCNN) is proposed in this paper. Firstly, the time–frequency representations (TFRs) of the raw signals which is synthesized by ac current and inner circulating current of the MMC are calculated with SST. Then, DCNN is introduced to learn the underlying features from the TFRs, and its key hyperparameters are optimized with genetic algorithm. Meanwhile, batch normalization, dropout and data augment technologies are explored to prevent DCNN model from overfitting and improve model performance. Compared to traditional SVM and BP-based algorithms, SST-GA-DCNN achieve high diagnosis accuracy. The experimental results show the feasibility and applicability of the proposed fault diagnosis framework.

Introduction

With the rapid development of power electronics in recent years, modular multilevel converters (MMC) have been widely used in various engineering fields, e.g., high-voltage direct current (HVDC) transmission, high-voltage electric driving system, renewable energy [1]. As a new topology of voltage source converters, MMC plays an significant role in the power system [2]. However, security of power electronic switches is one of the most important challenges in MMCs for the exposure to tough working environments. Specially, the insulated gate bipolar transistor (IGBT) of submodules (SMs) is most prone to suffer failures because of device aging, overloading and unexpected operating conditions [3]. Moreover, MMC contains a large number of SMs, which will increase the risk of potential failure. Therefore, effective fault diagnosis (FD) technique are helpful to enhance operational reliability and shorten cost savings of MMC [4].

In the past few years, fault diagnosis of the MMC become a new focus [5], [6]. Generally, there are mainly two methods of FD including model-based method [7], [8], [9], [10], [11] and data-driven method [12], [13], [14]. Model-based methods have been widely used in FD of MMC [15], [16], [17], however, these methods is limited by requiring prior knowledge of the physical model which is difficult to be represented accurately for the complex condition. Alternative to model-based methods, data-driven one becomes more and more popular in the field of fault diagnosis in recent years [18], [19], [20]. Different from utilizing physical law to estimate the process, data-driven method employ the observed data to reveal the underlying correlations and causalities of systems. It should be noted that so far, the FD of MMC has mainly focused on model-based methods, and there are little pioneering work in the literature on data-driven methods [21].

The data-driven method usually extracts the features of the sampled data firstly, and then uses machine learning methods such as decision tree, support vector machine (SVM), neural network and other classifiers for fault classification and recognition. The feature extraction process that is called feature engineering [22], which directly determines the complexity and effect of FD. As one of the key steps in FD, feature extraction contains time domain, frequency domain and time–frequency representations (TFRs) methods. For example, Li [23] introduced a time domain fault feature extraction method for MMC, and then diagnosed it with an improved support tensor machine method; Xiao [24] proposed a method for extracting fault features combining time and frequency domains, and then used deep learning to predict the remaining life of rolling bearings; Zhao [25] developed a algorithm based on time–frequency representations (TFRs) combined with CNN to classify the faults of planet bearings. Compared with the traditional filtering method [26], [27], [28], [29], continuous wavelet transform (CWT) and short time fourier transform (STFT), it has been verified that synchrosqueezing transform (SST) is more effective in characterizing signals and it’s not limited by Heisenberg uncertainty principle. As the variant of standard time–frequency reassignment method, SST can be used to calculate the directional reassignment vector in both time and frequency directions from the magnitude of STFT or the Wigner-Ville distribution, and then remap the energies of the TFR by using the calculated directional reassignment vector [30], [31]. Therefore, by using the time–frequency resolution of SST, the fault feature can be effectively characterized.

Deep Learning has been paid much attention and achieved much success in many fields, such as image processing, machine vision and speech recognition, etc. [32]. As one of the best efficient deep learning model with multiple hidden layers, deep convolutional neural network (DCNN) is mainly used in computer vision tasks, and it achieves efficient classification results even on very small data sets. To be specific, by switching the low-level features to the high-level features through the feature transfer of layer by layer, the feature learning and expression can be realized. Compared with BP neural network and other shallow networks, DCNN has stronger learning and expression ability for complex features, faster computing speed, and avoids training falling into local extremum. Not surprisingly, as a deep model with excellent recognition performance, DCNN has been gradually applied in fault diagnosis [33], [34].

Inspired by the above discussion, this paper presents a fault diagnosis framework of MMC with SST and the optimized deep convolution neural network. The main contributions are highlighted as follows: 1) The TFRs of different MMC faults are calculated by SST, then the clearer feature maps can be obtained with its high time–frequency resolution; 2) DCNN is employed to mine deep fault features, batch normalization and dropout technologies are introduced to prevent DCNN model from overfitting, and the GA and cross validation algorithm are utilized to optimize the hyperparameters of DCNN; and 3) By synthesizing the SST and optimized DCNN model, more accurate classification of MMC faults can be achieved.

Section snippets

Operation principle of MMC

The main goal of this paper is to achieve efficient fault diagnosis of MMC. Before introducing the main algorithms, we first introduce its operating principles. The typical MMC topological structure is shown in Fig. 1(a), it is composed of three phases and six bridge arms, and the phase unit is constituted by upper and lower bridge arms. Each bridge arm is consisted of a electric reactor L and N SMs, where Uj and ij represent the phase voltage and current at the ac side of MMC, Udc and Idc

Theoretical framework

According to the operation data of MMC, we will utilize SST-GA-DCNN algorithm to realize high-efficiency fault diagnosis. In order to introduce the algorithm more clearly, the corresponding basic theory is described as follows.

Optimizing the hyperparameters of DCNN with GA

As is known to all, the architecture of CNN mainly contains weight parameters and hyperparameters. In the process of network training, the weight parameters are automatically updated by error back propagation algorithm. The network hyperparameters mainly include learning rate, sample batch size, dropout and number of neurons in each layer. Actually, small changes in hyperparameters will lead to a great impact on network training. Therefore, the intelligent optimization algorithm is urgent need

Experimental results and analysis

In this section, a half-bridge MMC topology is employed to construct a 31-level MMC system in the Matlab/Simulink platform with the carrier phase-shifted modulation strategy, and fault diagnosis method of MMC is verified be using the operation data with the parameters shown in Table 1.

Conclusion

In this paper, a integrated FD framework for the open-circuit fault of MMC is proposed based on SST-GA-DCNN algorithm. In this algorithm, the original 2-D time-domain fault signal of MMC is processed by wavelet hard threshold denoise and then transformed by SST, so that a more focused 3-D time–frequency map of time–frequency ridge can be obtained. After processing with SST, the time–frequency ridge line is more focused, which weakens the frequency aliasing phenomenon of current harmonic signals

CRediT authorship contribution statement

Longzhang Ke: Conceptualization, Methodology, Validation, Formal analysis, Software, Writing - original draft. Yong Zhang: Methodology, Formal analysis, Validation. Bo Yang: Formal analysis, Investigation. Zhen Luo: Investigation, Writing - review & editing. Zhenxing Liu: Supervision, Methodology, Writing - review & editing, Funding acquisition, Project administration.

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.

Acknowledgement

This work was supported by the National Natural Science Foundation of China [Grant 61873197].

Longzhang Ke received the M.Sc. degree in control theory and control engineering from Wuhan University of Science and Technology, Wuhan, China, in 2012, where he is presently working towards the Ph.D. degree. His research interests include machine learning, multilevel converters, and fault diagnosis.

References (35)

  • G. Thakur et al.

    The Synchrosqueezing algorithm for time-varying spectral analysis: robustness properties and new paleoclimate applications

    Signal Process.

    (2013)
  • W. Liu et al.

    A survey of deep neural network architectures and their applications

    Neurocomputing

    (2017)
  • F. Deng et al.

    Voltage-balancing method for modular multilevel converters under phase-shifted carrier-based pulsewidth modulation

    IEEE Trans. Ind. Electron.

    (2015)
  • P. Hu et al.

    Energy-balancing control strategy for modular multilevel converters under submodule fault conditions

    IEEE Trans. Power Electron.

    (2014)
  • Q. Song et al.

    A Steady-state analysis method for a modular multilevel converter

    IEEE Trans. Power Electron.

    (2013)
  • R. Picas et al.

    Reliable modular multilevel converter fault detection with redundant voltage sensor

    IEEE Trans. Power Electron.

    (2017)
  • Q. Yang et al.

    Analysis, detection, and location of open-switch submodule failures in a modular multilevel converter

    IEEE Trans. Power Delivery

    (2016)
  • Cited by (0)

    Longzhang Ke received the M.Sc. degree in control theory and control engineering from Wuhan University of Science and Technology, Wuhan, China, in 2012, where he is presently working towards the Ph.D. degree. His research interests include machine learning, multilevel converters, and fault diagnosis.

    Yong Zhang received the Ph.D. degree in control theory and control engineering from Huazhong University of Science and Technology, Wuhan, China, in 2010. From 2014 to 2015, he was a Visiting Scholar with the Department of Information Systems and Computing, Brunel University London, Uxbridge, U.K. He is currently an Professor with the School of Information Science and Engineering, Wuhan University of Science and Technology, Wuhan, China. He has authored over 20 papers in refereed international journals. His current research interests include remaining useful life prediction of key equipment, fault diagnosis and fault tolerant control of complex systems. Dr. Zhang is a very active Reviewer for many international journals.

    Yong Zhang received the Ph.D. degree in control theory and control engineering from Huazhong University of Science and Technology, Wuhan, China, in 2010. From 2014 to 2015, he was a Visiting Scholar with the Department of Information Systems and Computing, Brunel University London, Uxbridge, U.K. He is currently an Professor with the School of Information Science and Engineering, Wuhan University of Science and Technology, Wuhan, China. He has authored over 20 papers in refereed international journals. His current research interests include remaining useful life prediction of key equipment, fault diagnosis and fault tolerant control of networked ystems. Dr. Zhang is a very active Reviewer for many international journals.

    Bo Yang recived the Ph.D degree in nuclear science and technology from Tsinghua University, Beijing, China, in 2017. He is currently a senior engineer in the China Institute Marine Technology & Economy. His current research interests Intelligent manufacturing in industrial systems.

    Zhenxing Liu received the M.Sc. and Ph.D. degrees in electric engineering in 1990 and 2004, respectively, from Huazhong University of Science and Technology, Hubei, China. Currently, he is an Professor with the School of Information Science and Engineering from Wuhan University of Science and Technology, Wuhan, China. His research interests include monitoring and diagnosis of electrical machines and automatic devices.

    View full text