Research on transformer vibration monitoring and diagnosis based on Internet of things

1 Introduction

The continuous and rapid development of China’s economy has led to the rapid development of the power industry. Transformer is one of the most important equipment in the power grid and plays a very important role in the whole power system. The electric power system is a real-time energy delivering system that generates the power, transmits it, and supplies it when the power is required [1]. The transformer is used in the electrical power system to increase and decrease the voltage to a level required by the consumer equipment [2]. For practical applications, power is transmitted at higher voltages to reduce the conductor cost and improve the voltage regulation along with the line loss reduction [3]. Distribution transformers play an important role in a power distribution network, and they operate 24 h a day. Thus, any kind of fault in such transformers may lead to great inconvenience to a large number of consumers causing the financial loss of utilities. Therefore, it is required to avoid such transformer failures, and the distribution transformers are needed to be operated under the rated conditions [4]. The life of the transformer can be increased with proper maintenance and continuous monitoring of distribution transformers. Efficient prompt correction of any abnormal functioning can improve the lifespan of a transformer saving heavy investments and replacements of utilities [5]. The major reasons of failure in distribution transformers are overloading or unbalanced loading, its inadequate oil level or overheating of transformer oil, and moisture [6]. The unbalanced distribution transformer loading may lead to increased copper loss [7,8]. Thus, all these transformer parameters are needed to be taken care of for continuous monitoring of transformer performance and avoid unexpected failures [9,10]. Hence, there is a need for continuous monitoring and control for a distribution transformer.

Maintaining the normal operation of the transformer is the first prerequisite for the security of the power grid. Long-term operation may lead to aging and even some internal winding deformation and bulge [11,12,13,14]. Once a sudden accident occurs in the operation of large-scale power transformer, it may cause a large area of power outage, causing significant social effects and economic losses. Moreover, the maintenance cost of power transformer is high and the maintenance period is long. This article is based on the study of power transformer vibration diagnosis and fault diagnosis. A software system for transformer vibration monitoring and diagnosis based on Android platform is developed and implemented by applying the increasingly mature technology and architecture of Internet of things (IoT) [15]. The system can realize live, real-time monitoring and fault diagnosis of power transformer all day long and remotely through cloud server, so as to improve the safety and stability of transformer operation and also the economy and reliability of power grid [16].

The traditional monitoring systems for distribution transformers involve manual monitoring. An individual manually visits the transformer site at regular intervals of time, to record transformer parameters. However, this type of method is not suitable for monitoring the occasional transformer health involving sudden overloads, oil and winding temperature, etc. [17,18]. IoT-based solutions have made better monitoring and control of distributed transformers by accurate monitoring of voltage, current, and other parameters which can increase the lifespan of transformer maintaining the stability of the grid [19,20].

In this article, the naive Bayesian classification model and the support vector machine model are studied, and the transformer fault diagnosis algorithm based on the naive Bayesian model and the support vector machine model, respectively, is proposed. Through the comparison and analysis of the experimental results and the field measured data of the transformer, the more effective machine learning model and its diagnosis algorithm for live monitoring and fault diagnosis of power transformer are obtained.

The rest of this article is organized as follows: Section 2 provides the literature review of the existing techniques for transformer monitoring schemes proposed by various researchers. The research processes involving sample processing and designing of vibration monitoring and diagnosis software system based on Android platform are discussed in Section 3. Section 4 provides the results and discussion followed by Section 5 depicting the conclusion and future scope of this work.

2 Literature review

For transformer fault diagnosis, many research studies at home and abroad are based on oil chromatogram analysis. At present, the mature diagnosis algorithm based on machine learning is also based on oil chromatography analysis. Compared with this, the research on transformer fault diagnosis based on vibration analysis is slightly insufficient. It has gone through the process of experience classification, diagnosis method based on cylinder single linear classification and machine learning intelligent diagnosis. In the early stage of fault diagnosis, linear classification after extracting vibration features is the main method. Based on the research status at home and abroad, there are two main methods for transformer vibration signal feature extraction [21]. The first is based on time series. Based on the analysis of the time-frequency characteristics of the transformer, the time-domain characteristics of the transformer are extracted based on the analysis of the vibration characteristics of the transformer. Li et al. applied multiscale entropy to feature extraction of winding vibration signals. Therefore, it is very suitable for processing transformer vibration signals with fast characteristic information attenuation, high electromagnetic interference, and high data correlation [22]. Therefore, the variation of vibration signal sample entropy in different scales can be used to reflect the change of transformer winding operation state, and the sample entropy value can be used as the characteristic quantity reflecting the mechanical structure change of winding. The software system of live vibration monitoring and fault diagnosis for transformer is designed, developed, and implemented. It can effectively monitor the healthy operation status of power transformer in remote and real time, so as to improve the safety, stability, and reliability of transformer operation [23,24]. Bossi et al. [25] studied various faulty as well as non-faulty conditions of power transformers. Domain generation algorithms (DGAs) were proven to be the most power diagnostic tools for the determination of dielectric, thermal, and chemical aging problems in the transformers. The offline methods such as flow response and partial discharge analysis were utilized for analyzing the transformer condition by Höhlein et al. [26], but some of these approaches require expert analysis. This requirement makes such methods costly and not suitable for monitoring the distribution systems. The application of ultra-high frequency range for the monitoring purpose has been explored by Judd et al. [27], which was considered good for transformer diagnosis. Jaiswal et al. [28] developed and implemented a transformer inter-turn fault detection system for power transformers. This eliminates the requirement current transformers at the secondary side, and the load profile is needed to be reorganized as a prerequisite for the determination of turning error in this case. To detect the presence of inter-turn fault, the authors have investigated the approach demonstrating its applicability on no-load and very light load conditions [29]. It is investigated that under high load conditions, the analysis based on the no-load harmonics is not used as the current at this stage is constant regardless of the load [30]. An algorithm was proposed by Al-Haj [31] to calculate the errors and twisting of line. A simple and stable approach was proposed by Venikar et al. [32] based on symmetric components before troubleshooting. Fuzzy set theory was implemented by Judd et al. [27] to calculate the health index for oil-immersed transformers. This method uses furan analysis and DGA. A method to examine the transformer health index based on the load data and different transformer tests has been proposed in ref. [33]. Authors use the information of asset owners and utility engineering to detect the abnormality by using the SMS system [34].

To sum up, although the research on power transformer vibration monitoring and fault diagnosis has made some progress in recent years, it still needs to be further improved. The main innovations of this article are as follows: (1) the vibration mechanism of power transformer and its extraction algorithm based on vibration; (2) based on the vibration eigenvalue, the vibration mechanism of power transformer and its extraction algorithm are described.

3 Research process

The entire research process is divided into three different steps: a sample processing step, a naive Bayesian model-based algorithm, and the designing of a vibration monitoring and diagnosis software system based on Android platform.

3.1 Sample processing

Therefore, in the training model, transformer data are selected every other hour, which independently represents all the vibration characteristics of a transformer. Besides the invalid data, two-thirds of them are selected as training samples. The remaining one-third is used as the validation sample. The statistical table of sample library is given in Table 1.

Table 1

Statistical table of sample database

	Normal	Ageing	Fault
Number of transformers	17	8	8
Sample size	305	160	160
Training sample	204	106	106
Validation sample	101	54	54

In this article, the three-dimensional (3D) feature of the sample cannot be shown in the four-dimensional (4D) distribution as shown in Figure 1.

Figure 1

3D distribution of feature set samples.

Taking frequency complexity analysis (FCA), determinant (DET), eigenvector dimension reduction (EDR), and model predictive controller (MPC) as the three axes of the 3D graph, it can be seen from the figure that the normal, aging, and fault transformers are linear and inseparable from the perspective of 3D eigenvalues, that is, there is no plane to accurately divide the 3D sample. Therefore, the 4D feature vector is more linear and inseparable. In this article, the naive Bayesian model and support vector machine (SVM) model will be established based on the feature vector, so that the model can better complete the transformer fault diagnosis.

3.2 Naive Bayesian model-based algorithm

The method of establishing the probability model of power transformer based on the prior probability model of the power transformer is established based on the process of Bayesian probability calculation in Figure 2.

Figure 2

Training process of naive Bayesian model.

4 Results and discussion

The results are analyzed for the proposed model in this section of the article. Initially, the results are presented for the Bayesian model on the basis of transformer fault diagnosis and system function.

4.1 Bayesian model test based on transformer fault diagnosis

The Bayesian networks overcome the difficulties of conceptual rule-based systems by creating a relationship between the data much clearer. Bayesian model for the fault diagnosis takes into account the transformer operation and its abnormality behaviors. The transformer fault is diagnosed in time using the Bayesian model which utilizes an association rule for the statistical analysis of correlation between the operating conditions, abnormality signs of the transformer, and the type of fault occurred. When the transformer fails, data are provided by D = { X i , X j , … , X k } , where X i ∼ X k is the occurrence of abnormal sign for a fault set F of the transformer [35]. Therefore, the joint probability of abnormal sign of transformer determined by the Bayesian network is provided by the following equation:

(1) P { X i , X j , … , X k } =   ∏ i P ( X i \ P a i ) .

For a certain type of fault C r in the set of fault C, the probability of its occurrence under the abnormal condition is provided by the following equation:

(2) P ( C r | X i , X j , … , X k ) = P ( X i , X j , … , X k | C r ) × P ( C r ) P ( X i , X j , … , X k ) .

The sampling criteria are described in Table 2, which determine the training and the validation samples on the basis of a prior probability of normal, ageing, and faulty transformer. The training sample number and prior probability corresponding to different transformer states are shown in Table 2.

Table 2

Training samples and prior probability table

	Normal	Ageing	Fault
Validation sample	204	106	106
Prior probability	0.49	0.25	0.25

Through the model training, the probability density curve based on the feature set can be obtained, which is the conditional probability part in Figures 3–6.

Figure 3

FCA probability density curve.

Figure 4

DET probability density curve.

Figure 5

EDR probability density curve.

Figure 6

MPC probability density curve graph.

Figures 3–6 represent the probability density curves of the following four eigenvalues: FCA, DET, EDR, and MPC. It can be seen from Figure 3 that the smaller the FCA value is, the higher the probability of transformer in the normal state. With the increase in FCA, the frequency complexity of transformer becomes higher, the harmonic component increases, and the possibility of transformer fault is higher. Similarly, for DET, the smaller the DET, the higher the probability of transformer faults. The linear separability of MPC single eigenvalue is lower.

According to the prior probability in Table 2 and the probability density (conditional probability) in Figures 3–6, the naive Bayesian algorithm model based on power transformer fault diagnosis can be obtained from the calculation method of naive Bayesian posterior probability proposed previously. The above is the process of building Gaussian naive Bayesian model, and the training process of polynomial naive Bayesian model is the same.

4.2 System function test results

The main test content is the feedback of each module under different events. The function test is carried out according to the modules, including the following items: user management module, information management module, real-time monitoring module, analysis and diagnosis module, and historical trend module.

4.2.1 Function test results of information management module

On the premise of successful login and good network signal, the monitoring terminal information management submodule can communicate with the server to obtain and display the information of the monitoring terminal. The transformer information management submodule can communicate with the server on the premise of successful login, download, and display the transformer information correctly. The transformer information is successfully modified. The complete process is depicted in Figure 7.

Figure 7

Function test process of information management module.

4.2.2 Function test of real-time monitoring module

In the good state of the network, the functions of the three submodules are normal. Users can normally register in the system, log in with the registered user name and password, and can modify the password under the correct operation steps.

4.2.3 Analysis of the functional test results of diagnostic module

The test results of the analysis of diagnostic function suggests that the methods were successful for transformer fault diagnosis, provides the detailed information query, and is also successful in switching the feature values. The results obtained are indicated in Table 3.

Table 3

Analysis of diagnostic function test results

Test content	Result
Transformer fault diagnosis	Success
Detailed information query	Success
Switching feature values in details	Success

4.2.4 Function test of historical trend module

The function test outcomes show that the system is able to switch between the historical data of the transformer and the Android network. The research processes involving sample processing and designing of vibration monitoring and diagnosis software system based on Android platform are used in this work to set the historical trend. This historical trend module meets the design requirements of the system, thereby providing a safe, stable, and reliable system.

5 Conclusion

At present, most of the online monitoring and fault diagnosis models of transformers based on vibration analysis method are simplified and empirical linear models. However, a small number of diagnosis models based on machine learning method are often affected by external factors (such as load changes and different measuring point positions), which need to be improved. The main achievements of this article are as follows:

1. The vibration mechanism of power transformer and the eigenvalue extraction algorithm based on vibration are described.

2. The application of SVM model in transformer fault diagnosis is improved. The vibration characteristics information at different measuring points are examined for the analysis and diagnosis of transformer, while considering the advantages of SVM model for classification of high-dimensional spatial features. In this article, four kinds of feature vectors are extracted from each measuring point to establish a 4D feature vector, which is used as the input of SVM diagnosis model for modeling and diagnosis, and a diagnosis algorithm model based on different kernel functions is established. Finally, through comparison, the diagnosis algorithm based on multimeasurement point information and Gaussian kernel SVM model is selected as the fault diagnosis algorithm of transformer.

3. On the basis of the above research results, the future perspective of the proposed system is the improvement of safety, stability, reliability, and the applicability of the system based on Android technology.