Optimizing GIS partial discharge pattern recognition in the ubiquitous power internet of things context: A MixNet deep learning model

https://doi.org/10.1016/j.ijepes.2020.106484Get rights and content

Highlights

  • MixNet model is employed to optimize partial discharge (PD) pattern recognition.

  • Generative adversarial network is used to data enhancement.

  • The effectiveness of the proposed method has been verified by the PD dataset.

  • The results demonstrate its superiority over the other traditional methods.

Abstract

Gas-insulated switchgears (GISs) are an essential component of the power system, but in the event of a failure they may pose a serious threat to the safe operation of the entire power grid. The ubiquitous power Internet of Things (UPIoT), which is characterized by its online monitoring of failure samples for database building and further processing, is of great use in identifying potential insulation defects. We propose a MixNet deep learning model (MDLM) in the UPIoT context with the aim of optimizing partial discharge (PD) pattern recognition, after taking into account multiple indicators such as accuracy and effectiveness. Furthermore, a generative adversarial network was adopted for data enhancement to improve the model’s generalization ability and to solve such problems as noise jamming and the less clear effect of traditional spatial transformation methods on unified PD specification data. We found that an MDLM can effectively improve fault diagnosis accuracy while largely reducing calculation and storage costs. After validation, the recognition accuracy of an MDLM was 99.1%, significantly higher than that of other methods. The advantages of the proposed method were also demonstrated by the model feature extraction and the last hidden fully-connected layer using a visualization method.

Introduction

Gas-insulated switchgears (GISs) are widely used in high-voltage substations due to their small volume, high reliability, and maintenance-free features [1], [2], [3]. However, some defects are inevitable in a GIS after long-time operation, which may result in a failure. If a failure occurs, the on-site personnel may be in danger and very large economic losses will occur. For these reasons, strengthening GIS monitoring for real-time diagnosis is of great importance to ensure the safe and reliable operation of the entire power system [4], [5], [6]. GIS insulation defects occur in the form of partial discharge (PD), which hasten equipment aging. Because of the complex fault mechanism, it is very difficult to identify GIS insulation defects. Nevertheless, the advancement of the ubiquitous power Internet of Things (UPIoT) construction makes it possible to collect and save all GIS partial discharge signals by using intelligent terminals [7], [8]. Therefore, comprehensiveness and integrity of the fault information can be achieved. Moreover, digital intelligent diagnoses of faults can be achieved based on fault sample mining.

To build a reliable PD defect identification system, intensive research has been done over the past two decades, mainly in regard to two aspects: (1) Analyzing PD signals and constructing key characteristic parameters representing PD, mainly using Fourier transform, wavelet decomposition (WD), empirical mode decomposition (EMD), S-parameter transformation, envelope analysis, and energy analysis [9], [10], [11]; and (2) classifying key characteristic parameters by advanced pattern recognition classification methods, mainly involving support vector machines (SVM), decision trees, random forests (RF), back propagation neural networks (BPNN), and deep learning methods such as convolutional neural networks (CNNs), autoencoders (AE), and long short-term memory (LSTM)-based recurrent neural networks [12], [13], [14], [15]. Among them, CNN series have been widely adopted for PD pattern recognition because of their automatic feature extraction, which significantly reduces the need for expert experience in feature engineering.

Li et al. adopted a multiresolution CNN together with the LSTM method, which effectively and automatically captured the characteristics of ultrahigh-frequency (UHF) signals, thereby achieving multidata-fused GIS PD defect identification and location [16]. On the basis of laboratory experiments and field data, Song et al. constructed complex data sources and verified that the GIS PD pattern recognition accuracy of deep CNN in large data samples is significantly higher than that of traditional machine learning methods [17]. Duan et al. efficiently identified PDs using a sparse AE layer and a softmax layer. They designed a comprehensive blind test to verify the validity and robustness of their model [18]. Using a 1-D convolution instead of a 2-D convolution, Wan et al. successfully recognized patterns with applications in data mining and data use [19]. Han et al. proposed a novel adversarial learning framework to make feature representation robust, increase the generalization ability of the trained model, and avoid overfitting with a small labeled sample size for intelligent diagnosis of mechanical faults [20].

However, all the above methods have certain drawbacks. Some have low recognition accuracy resulting from multiple factors such as model depth, width, and input resolution ratio, and some are hard to train because the vanishing gradient. Also, they fail to consider embedding the models in intelligent terminals in the UPIoT context and the effect of data enhancement under deep learning circumstances. Furthermore, as a network specifically designed for mobile terminals, MixNet has never been used in the field of GIS PD defect recognition [21].

To this end, considering its high accuracy and effectiveness, we propose a MixNet deep learning model (MDLM) to identify GIS PD defects by using MixNet. Because the traditional spatial transformation data enhancement method cannot be applied to standard and normalized GIS PD data, we also adopted a generative adversarial network (GAN) for data strengthening. In addition to effectively improving the accuracy of fault diagnosis, it reduces the calculation and storage costs, which enables the proposed model to be deployed to a UPIoT intelligent terminal to achieve real-time, accurate, and rapid diagnoses of power equipment.

The remainder of this paper is as follows: Section 2 proposes how to conduct data enhancement using a GAN, Section 3 describes the MDLM, Section 4 states the data acquisition experimental details, Section 5 discusses the results, and conclusions are drawn in Section 6.

Section snippets

Data enhancement using a generative adversarial network

Data enhancement technology has long been an important way to deal with data scarcities through synthesizing or transforming, which can generate new data from limited data [22]. Traditional data enhancing methods include flip, rotation, scale, crop, and translation. For the PD dataset, however, traditional methods cannot significantly improve the generalization ability of models because the data are collected and stored using a unified standard method without involving processes such as picture

Convolutional neural network

As one of deep learning models designed for image recognition, CNN mainly consists of the input layer, convolutional layer, pooling layer, and fully connected layer [26]. With the further development of CNN models, the convolutional layer includes more than merely convolution operations, but batch normalization and non-linear activation. And currently, it is typical of a CNN model to possess all these three kinds of operations in the convolutional layer.

The convolution layer consists of

Data acquisition

Defects tend to occur in the process of GIS design, production, transportation, installation, operation, and maintenance. Four typical defects have the highest incidence, namely free-metal particle defects (M), metal tip defects (N), suspended electrode defects (O), and insulator air gap defects (P). Hence, we selected those four types of defects for GIS PD pattern recognition. The structure of the experimental cavity is shown in Fig. 7 (a). One of the most important devices in the experiment

Model performance analysis

For the Experimental and XFdtd simulation, 10,000 groups of data under the four typical defects were used for GIS PD pattern recognition, with 70% as the training set, 10% as the verification set, and 20% as the testing set. We trained our model with Keras (TensorFlow backend) on a machine equipped with a GeForce RTX 2060 GPU graphics card, an Intel i7-8700 CPU, and 16 GB of RAM. In the implementation of convolutional neural network and other model code, we use Pycharm and Anaconda combined

Conclusions

This paper proposes a novel method for GIS PD defect identification in a UPIoT context to achieve accuracy and effectiveness. Through simulation, many samples of four typical GIS PD defects were constructed that we also used for pattern recognition by an MDLM. After comparison with MobileNetV1, MobileNetV2, Xception, ResNet, and LeNet, the proposed MDLM outperformed the others with a 99.1% recognition rate. Furthermore, the relatively small parameters and storage space of the proposed model

Credit author statement

Y.W. and J.Y. conceived and designed the experiments; Y.W. wrote the paper; Z.Y. modified the code of the paper; Z.Y. revised the contents and reviewed the manuscript; T.L. provided the simulation data. Y. W., Z. Y. and J.Y. revised the manuscript.

Declaration of Competing Interest

The authors declared that there is no conflict of interest.

References (42)

  • Y. Wang et al.

    A Mobilenets Convolutional Neural Network for GIS Partial Discharge Pattern Recognition in the Ubiquitous Power Internet of Things Context: Optimization, Comparison, and Application

    IEEE Access

    (2019)
  • W. Hu et al.

    Selection of Cluster Heads for Wireless Sensor Network in Ubiquitous Power Internet of Things

    International Journal of Computers, Communications & Control (IJCCC)

    (2019)
  • X. Zhang et al.

    Fourier Transform Infrared Spectroscopy Quantitative Analysis of SF6 Partial Discharge Decomposition Components

    Spectrochim Acta Part A Mol Biomol Spectrosc

    (2014)
  • X. Peng et al.

    A Convolutional Neural Network-based Deep Learning Methodology for Recognition of Partial Discharge Patterns From High-voltage Cables

    IEEE Trans Power Delivery

    (2019)
  • J. Tang et al.

    Assessment of PD Severity in Gas-Insulated Switchgear with an Ssae

    IET Sci Meas Technol

    (2017)
  • Q. Khan et al.

    Partial Discharge Detection and Diagnosis in Gas Insulated Switchgear: State of the Art

    IEEE Electr Insul Mag

    (2019)
  • M.T. Nguyen et al.

    Recurrent Neural Network for Partial Discharge Diagnosis in Gas-Insulated Switchgear

    Energies

    (2018)
  • L. Gaoyang et al.

    Partial Discharge Recognition with a Multi-Resolution Convolutional Neural Network

    Sensors

    (2018)
  • H. Song et al.

    GIS Partial Discharge Pattern Recognition Via Deep Convolutional Neural Network under Complex Data Source

    IEEE Trans Dielectr Electr Insul

    (2018)
  • L. Duan et al.

    Identification of Partial Discharge Defects Based on Deep Learning Method

    IEEE Trans Power Delivery

    (2019)
  • X. Wan et al.

    Network

    (2018)
  • Cited by (0)

    View full text