On-Line Monitoring and Fault Diagnosis of Box Transformer Substation Based on VPRS-RBFNN

: Box transformer substation (BTS) is an important power distribution environment. To ensure the safe and stable operation of the power distribution system, it is critical to monitor the BTS operation and diagnose its faults in a reliable manner. In the Internet of Things (IoT) environment, this paper aims to develop a real-time and accurate online strategy for BTS monitoring and fault diagnosis. The framework of our strategy was constructed based on the IoT technique, including a sensing layer, a network layer and an application layer. On this basis, a BTS fault diagnosis method was established with variable precision rough set (VPRS) as the pre-network and the radial basis function neural network (RBFNN) as the back-fed network. The VPRS and the RBFNN were selected, because the BTS faults have many characteristic parameters, with complex nonlinear relationship with fault modes. Finally, a prototype of our strategy was developed and applied to the fault diagnosis of an actual BTS. The results fully demonstrate the effectiveness and feasibility of our strategy.


INTRODUCTION
As the world's largest power generation, the most complex power system, China's power system contains power generation, transmission and transformation, distribution, power consumption and dispatching [1]. As a key link connecting power transmission and users, the safe and reliable operation of distribution network is of great significance to the stability of power system and users' lives and properties. Therefore, it is an urgent task to comprehensively improve the information and automation level of the distribution network, enhance the ability of analysis, early warning and disaster prevention of power distribution links, realize the real-time perception and fine control of the operation status of the power distribution network [2][3][4][5].
Domestic and foreign scholars have done a lot of research work on transformer fault diagnosis. Hu et al. [6] mined the casing data of oil-immersed transformers and used Apriori algorithm and Tanimoto coefficient to evaluate the relationship between state parameters; based on Pearson correlation coefficient, the fault diagnosis matrix was constructed, which judges the fault diagnosis mode of the equipment. Feng et al. [7] used the method of data fusion to establish the equipment diagnosis model of power transformer, which reduced the uncertainty of fault diagnosis; Lin et al. [8] found the discrete model of topology structure and state of power system of transformer from a large number of historical data, and proved the synchronous fault principle of transformer. In order to overcome the limitations of DGA method, Abu-Siada and Hmood [9] introduced fuzzy logic to identify transformer faults more accurately. Velásquez and Lara [10] designed an adaptive decision system based on principal component analysis and fuzzy logic to realize early fault diagnosis of transformer and obtain degradation rate and health index. Malik and Jarial [11] used fuzzy logic to evaluate four state conditions of transformer deterioration, and evaluated the quality of transformer. The edge paper degradation was estimated and appropriate maintenance scheme was proposed. Chen et al. [12] used rough set to reduce the fault decision table and establish the fault diagnosis rules and decision table of transformer; Li Hui et al. [13] transformed the characteristic value of the gas in the transformer into two-dimensional data and input it into the convolutional neural network for optimization. In order to calculate the health index of transformer, Mominul et al. [14] proposed a generalized regression neural network GRNN for the state evaluation of transformer.
In the troubleshooting of other power distribution equipment, domestic and foreign scholars have also achieved some research results. Wang et al. [15] proposed a grid fault tracking method based on big data platform to find faulty components and give the cause of faults; Mansour et al. [16] used the Petri net to establish a fault diagnosis model for the fault items of a large power station, so that the monitoring personnel could accurately diagnose the fault. Based on cloud-IoT technology, Meloni et al. [17] proposed SE architecture solution to support the effective evaluation of the distribution network. The results of some studies have shown that online monitoring of power distribution equipment is more reliable with the support of IoT technology [18][19][20].
To sum up, the research on fault diagnosis of power system is booming, and most of these researches focus on fault diagnosis of power transmission and transformation, and the fault diagnosis of oil immersed power transformer is the core. However, with the continuous development of the Internet of things technology, ubiquitous sensors enable us to master the operation status of distribution equipment more comprehensively. Therefore, it is urgent to use the perspective of big data for intelligent real-time diagnosis of distribution equipment.
The BTS is an important power distribution equipment in the distribution network, which transforms electric energy from high-voltage system above 10 kV to lowvoltage system. Most of the BTS are installed outdoors, the working environment is changeable, the conditions are complicated, and natural disasters and external forces are seriously damaged. In the long-term operation process, various failure problems will occur. When a fault occurs, relying on the manual troubleshooting method, the workload is large, the accuracy is low, and the fault diagnosis state of the BTS cannot be diagnosed and processed in real time, which may cause major safety accidents such as large-scale blackouts and explosions.
Therefore, we propose a real-time monitoring and fault diagnosis strategy based on variable precision rough setradial basis function neural network (VPRS-RBFNN) algorithm in the IoT environment. Firstly, a real-time online monitoring platform for BTS equipment based on IoT technology is constructed to realize real-time monitoring of the operation information of the BTS. Secondly, based on the historical data and real-time data of the BTS operation, a fault diagnosis model based on VPRS-RBFNN algorithm is constructed. Finally, the validity and feasibility of the proposed strategy are verified by the specific project of the BTS production enterprise.

OVERVIEW OF THE PROPOSED APPROACH
The box transformer usually consists of high-voltage equipment, transformer and low-voltage equipment [21], and its structure is shown in Fig. 1. In the actual operation process, the BTS's operation state is affected by many factors, excessive humidity, frost, dew, snow, fog, etc., which will cause short circuit and leakage and discharge of circuit; generally, the space in the cabinet is narrow, the equipment operates under high load for a long time, and the ambient temperature and electrical contact temperature are easy to be on the high side. If the monitoring and treatment are not timely, it is easy to cause fire or explosion accidents, which seriously affects the stability and safety of the BTS [22].
At present, with the development of power IoT technology, the fault diagnosis model is established based on the big data from BTS's on-line monitoring, the historical operation data of BTS is analyzed and mined, and the real-time risk assessment of operation state is carried out by using artificial intelligence methods such as machine learning, so as to discover the weak links in time and realize the transform from diagnosis based on traditional sensor to intelligent system. So, it has great significance to ensure the reliable operation of the BTS. Therefore, the first step is to establish the BTS's online monitoring fault diagnosis system framework as shown in Fig. 2. According to the typical three-tier architecture of IoT application, it is composed of perception layer, transport layer and application layer. The perception layer mainly solves the problem of information perception and collection, which is the core infrastructure; the transport layer mainly carries out reliable long-distance transmission of information; the application layer is used to support the platform and application services, which mainly support information collaboration, sharing and interworking.

Perception Layer
The perception layer is designed with temperature and humidity controller, harmonic meter, YZ80 multi-function instrument, KY-8180A microcomputer transformer protection intelligent monitoring unit, smart sensors as the device for collecting operational data of box-type substation. The collected data includes different types, such as temperature and humidity, phase voltage, phase current, active power, reactive power, harmonic factor, and insulation factor. These data are divided into different indicators such as environment, electrical, power quality, energy efficiency, and fault. In order to properly collect the above monitoring data, the sensor deployment information is shown in Tab. 1.

Transport Layer
The intelligent monitoring units and sensors designed by the perception layer adopt a combination of wired RS485 and wireless Zigbee and Wi-Fi to convert the original data into the data collector through the Modbus protocol for further analysis and processing. After the encapsulation of TCP/IP protocol, the data is transmitted to the server's database or data center via GPRS.

Application Layer
The application layer is usually rich in content. This framework only focuses on real-time online monitoring and fault diagnosis of the BTS. It consists of two parts: the real-time online monitoring and the fault diagnosis based on VPRS-RBFNN. The essence of BTS's real-time online monitoring is real-time collection and monitoring of operation data of BTS equipment system based on IoT technology. Through this, real-time collection and supervision of the operation parameters of the BTS equipment and its external environment information can be realized, which is the basis and data source for online diagnosis of BTS faults. The fault diagnosis of BTS based on VPRS-RBFNN is based on the real-time monitoring of box transformer operation information. The main process is to establish a fault diagnosis model between the key fault components and the fault characteristic parameters of the BTS. Based on the VPRS-RBFNN fault diagnosis algorithm, the model is trained by large amount of BTS's historical fault data, the failure modes are mined from the real-time data, and finally the fault diagnosis of the BTS can be realized.

FAULT DIAGNOSIS BASED ON VPRS AND RBFNN
In the IoT environment, the BTS's fault diagnosis changes from the original manual diagnosis mode to the online fault diagnosis, the fault diagnosis decision is made based on the data of the device characteristic parameters. As an important power distribution equipment, the BTS has many characteristic parameters, which have a nonlinear and complex relationship with each other and with the fault mechanism. Therefore, it is impossible to directly establish fault models from data.

The BTS's Fault Diagnosis Process Based on VPRS and RBFNN
Neural networks are often used to solve nonlinear problems [22]. Among them, RBFNN has the advantages of simple structure, strong nonlinear fitting ability, fast approaching and high robustness. Compared with traditional neural networks, it can better overcome the shortcomings of local minimum and slow convergence speed, and is widely used in fault diagnosis, pattern recognition and other fields [23][24][25]. The input characteristic parameters of the BTS are various and the number is large, which undoubtedly greatly increases the complexity of the topology structure of the RBFNN, which increases the training time of the fault diagnosis model and increases the difficulty of network convergence.
The Variable Precision Rough Set (VPRS) can be used as a strategy to solve the nonlinear correspondence problem. It can mine potential knowledge and laws from massive data without prior knowledge [26]. However, if it is used as the theoretical basis for the fault diagnosis model of the BTS, a perfect decision table is needed to obtain a fault rule with high credibility. In this case, the number of rules generated will be far greater than the rules given by experts. This is undoubtedly a huge problem [27].
Considering the complexity of RBFNN modelling and the knowledge of variable precision rough set with the advantages of knowledge reduction, the variable precision rough set can be combined with the RBFNN algorithm. The variable precision rough set knowledge reduction ability can be used to reduce the feature, in the other words, the number of parameters is reduced , the original sample of the RBFNN is subtracted, the redundant information is removed, the scale of the RBF network is simplified, the training time is reduced, so that the fault diagnosis model is more real-time, rapid and accurate.
The core of the VPRS-RBFNN algorithm is to use the variable precision rough set as the front-end processing system of the RBFNN, to quantitatively simplify large numbers of original samples and find the main characteristic parameters that lead to box change fault. Then, the information reduced from the variable precision rough set is input into the neural network for training, and the fault diagnosis model is formed; then the test data is brought into the fault model to test the feasibility of the model. By integrating the variable precision rough set and the radial basis neural network, we give the overall process of box fault diagnosis based on VPRS-RBFNN. As shown in Fig. 3, the specific steps can be summarized as follows: Step 1: Extract fault history data and establish an original decision table; Step 2: Data preprocessing, processing the data in the original decision table; Step 3: Discretization and simplification of the conditional data, forming a decision table after reduction; Step 4: The data in the decision table after simplification is brought into the RBFNN for training; Step 5: Test the trained model with test data; Step 6: Output the result.

Fault Diagnosis Algorithm Based on VPRS-RBFNN
The realization of the algorithm based on VPRS-RBFNN mainly includes the following aspects: Step 1: Samples selection The selection of samples should follow the principle of representativeness and compactness, and select common faults and faultless representative data from the characteristic parameter database of the BTS. The selected data is used as the original decision table of the variable precision rough set. In this decision table, the sample object is represented by Y = {y 1 , y 2 , …, y i , …, y n }, n is the number of samples. The characteristic parameters are expressed as X = {x 1 , x 2 , …, x j , …, x m } and m is the number of characteristic parameters. The fault type is expressed as D = {d 1 , d 2 , …, d k , …, d h } and h is the number of fault types.
Step 2: Knowledge reduction. Variable precision rough set can be expressed as a quadruple S = <U, A, V, f>, where U = {y 1 , y 2 , …, y n } is the universe, which is a finite set consisting of sample object y i (i = 1, 2, …, n);   where E(x j ) is the mean value of the characteristic parameter x j in the original sample set, and D(x j ) is the standard deviation corresponding to the characteristic variable x j .
Step 2.2: Discretizing the data of condition and decision attribute of the original decision Step 2.5: Using Eq. (4) to calculate the dependence λ(C, D, β) of condition attribute on decision attribute: If the dependence λ(C i , D i , ) of a single attribute i is equal to λ(C, D, ), the attribute is redundant and deleted; otherwise, using Eq. (5) to calculate the importance SIG(C,{i}), when SIG(C,{i}) ≠ 0, taking this attribute as one of the attribute CORE, namely CORE = {C i ,}; otherwise, this attribute is a non-attribute CORE, let i = i + 1; Step 2.6: Judging whether the loop is terminating; if I > m, terminating the cycle and executing step 2.7; otherwise, return to step 2.3: Step 2.7: Obtaining attribute core set CORE and forming attribute reduction table.
Step 3: Training the sample set Part of the data after reduction in the decision table is selected as the training sample, and a RBFNN topology with Y i ' input vectors, X j ' hidden units and D k output units is designed. According to the algorithm flow of RBFNN, the sample set is trained. The specific steps are as follows: Step 3.1: Initializing parameters. Given the initial learning rate , momentum factor α, ∈(0, 1),α∈(0, 1); the termination precision ε: Step 3. Step 3.3: Initializing the connection weight from the hidden layer to the output layer. Use Eq. (6) to initialize the weight, and get W k = [w k1 , w k2 , …, w kn ] T ,(k = 1, 2, …, q); Step3.5: Initializing width vector, from Eq.
Step3.6: Calculating the distance from the sample to the cluster center; Step3.7: According to Eq. (8), Eq. (9), Eq. (10), the weight, center point and center width parameters are calculated iteratively; Step 3.8: Calculating root mean square error of output RMS, if RMS ≤ ε, turn to step 2.7, otherwise, finishing the training.
Step 4: Testing the sample set The remaining data in the decision table is selected as the test sample, and the RBFNN training results are compared with the test sample to obtain the fault diagnosis accuracy of the RBF fault diagnosis model.
According to the content of the algorithm, the specific steps of the VPRS-RBFNN fault diagnosis algorithm are summarized in Fig. 4.

EXAMPLE VERIFICATION AND RESULTS ANALYSIS
We select six key fault components from BTS's equipment list, which are respectively dry-type transformer, high-voltage circuit breaker, capacitor arrester, low-voltage outlet breaker, low-voltage incoming circuit breaker, high-voltage lightning arrester; then, the 24 characteristic parameters shown in Tab. 2 of the six key components are selected from BTS's historical data as the fault data source.
Based on VPRS and RBFNN, firstly, we establish a fault diagnosis model between six key fault components and 24 characteristic parameters, and then the fault diagnosis model is trained by fault characteristic parameter data, finally, the fault type of the BTS can be inferred by inputting parameter's real-time data.
The establishment process of fault diagnosis model is as follows: (1) Construct the original decision    (2) Standardization of original decision table Due to the different data dimensions, all data are normalized through the Z-core method according to Eq.
(1), so that the 24 characteristic parameters which can standardly measure the importance of the observed object are under the same dimension. The results obtained are shown in Tab. 4.
(3) Discretization Using the equal-frequency discrete method, the discrete frequency is 4. The discrete result obtained by the equal-frequency discrete method can set a breakpoint for each attribute by calculating the number of attribute values included, so that each interval includes the same number of objects, and re-obtaining a simpler discrete result representation.
The equal-frequency discrete representation is shown in Tab. 5.
It can be seen from the figure that the best model can be obtained by iterating 48 times, and the trained fault diagnosis model is closest to the actual situation. The test data is brought into the trained model to obtain the fault diagnosis results as shown in Tab. 8. The results in Tab. 8 show that, compared with 10 groups of actual fault types known in advance, there are 9 groups with the same results and 1 group with different results. That is, the diagnostic accuracy rate of the BTS's fault diagnosis model based on VPRS-RBFNN is 90%, indicating that this model is feasible.  Figure 5 Iterative results of VPRS-RBF

CONCLUSIONS
Using the IoT technology to fully perceive the operation state of BTS, and to quickly and intelligently diagnose the fault state can effectively improve the reliability of the BTS's operation. We construct a BTS's real-time online monitoring and fault diagnosis framework, and design the perception layer, transport layer and application layer in detail, which can meet the real-time monitoring requirements of the BTS's operation state; Based on the VPRS-RBFNN, fault diagnosis strategy is designed, the fault type and fault reason of BTS can be diagnosed in time. The accuracy of BTS's fault diagnosis is 90%.
This strategy not only improves the management and maintenance level of the BTS, but also provides support for the BTS's maintenance plan. In order to realize the prewarning of the BTS's fault, our future work will focus on the BTS's fault early warning based on big data.