At the end of 2016, China had built the world’s largest power grid and achieved a huge amount of construction (Liu et al., 2020). However, as the grid continues to grow, the number and variety of power devices in the grid continues to grow (Wang et al., 2021). Due to the large and complex system, the power grid is more likely and more severe than a simple system (Wang et al., 2019a). Faults occurrence also has the nature of randomness, secondary, concurrency, explosiveness, and obfuscation. In most cases, multiple faults will occur at the same time (Wang et al., 2019b). This type of faults is called concurrent faults (Qin et al., 2018) and is also known as compound faults or multiple faults. Concurrent faults in different scenarios are completely different, and the characteristics of the faults are extremely complex and difficult to diagnose (Ma et al., 2018). Therefore, research on how to diagnose concurrent transformer faults is crucial for the operation and maintenance of transmission and transmission equipment, and for the safe and reliable transmission of power systems.

Currently, there is a lot of research on fault diagnosis methods, which can be divided into two main types. One is the traditional method, namely the principal component analysis method (Wang and Xiao, 2004) and the physical feature diagnostic method (Lei et al., 2016). The other is based on artificial intelligence algorithms such as neural network algorithms (Ding et al., 2011; Al-Saud et al., 2019), petri networks (Pan et al., 2008), and fuzzy logic theorems (Lang et al., 2019). The former pays attention only to the fault characteristics, records a single record, and cannot judge the concurrent faults which make it inapplicable. The research on the concurrent faults mainly focuses on the latter. The authors in (Xu et al., 2010) proposed a method of concurrent fault diagnosis information fusion based on random set theory. First, a combination rule of single and concurrent faults is artificially constructed, and then fuzzy functions are used for pattern matching to diagnose concurrent faults. The authors in (Guan and Jiang, 2020) proposed a concurrent multi-fault diagnosis method for electromechanical systems based on the Elman network and ECOC-SVM. This method uses an Elman network instead of manual construction rules to classify faults, splits concurrent faults into single faults for analysis, and the efficiency is improved. But the disadvantage is that the relationship between every single fault is not considered, and the diagnosis effect is not ideal. In response to the above problems, the authors in (Hu et al., 2009) proposed a concurrent fault diagnosis method based on multiple regression LSSVM, which uses multiple regression least squares support vector machines to model concurrent faults as a multi-input and multi-output problem. In (Li and He, 2013), the authors proposed a kernel fuzzy clustering method used in the diagnosis of multiple faults in complex products. It is believed that the relationship between single failure modes in complex products is particularly important for diagnosing concurrent faults and kernelized clustering. However, the manual construction of rules in the above literature is relatively cumbersome, and the constructed combination rules are only applicable to specific scenarios, the model is relatively single, the generalization ability is poor, and it is difficult to adapt to complex scenarios.

Graph Convolutional Network (GCN) is an extension of traditional Convolutional Neural Network (CNN) in non-Euclidean space. It cannot only use multiple graph convolutional layers to automatically extract the features of input variables, as well as take into account the topology structure between the individual nodes (Zhang et al., 2019). GCN currently has excellent application effectiveness in the areas of link prediction, protein classification, drug synthesis, and cross-domain pedestrian detection (de Kleer and Williams, 1987), but its application for simultaneous disability diagnosis is still in its infancy (Cen, 2010). In the power system massive data scenario, the fault diagnosis of concurrent faults of electrical equipment has not been well resolved (Ryu et al., 2019).

In summary, existing methods primarily take two or more simultaneous faults as separate categories of pattern recognition. This is the same as judging multiple faults as a new fault type, ignoring the coupling and correlation between multiple faults. To further investigate the correlation between multiple faults, this paper selects the Graph Convolutional Neural Network (GCN), with strong topological feature expression ability as the basis, proposes a method for transformer concurrent fault diagnosis based on graph neural network and knowledge graph. The main contributions of this paper are summarized as follows:

1) This paper proposes a fault knowledge-related expression method based on a knowledge map. The fault type is used as the node and the correlation relationship between faults is used as the edge to describe the fault knowledge relationship.

This paper divides the model layer of transformer fault knowledge graph into component layer and fault layer. Among them, the component layer defines all possible faulty components of the transformer and the connection relationship between the components. The component layer reflects the first-order connection relationship between the components of the transformer. The fault layer defines all the types of possible faults in the transformer, as well as the similarity and subordination between the types of faults. The fault layer is based on the transformer state evaluation guidelines, reflecting the logical relationship between the faults, and is essentially a knowledge graph with the transformer state evaluation guidelines as the data source.

Data layer triple extraction refers to obtaining structured knowledge such as entities and relationships between entities and attributes from unstructured data through a series of knowledge extraction methods under the guidance of the knowledge organization structure of the model layer. The fault knowledge graph in this paper has two entities: fault component and fault type. There are three relationships: fault component-fault component, fault component-fault type, and fault type-fault type. The knowledge graph is stored and represented in the form of triples. Compared with the traditional structured relational database, the relationship between fields and records requires complicated calculation and extraction. The triple representation of knowledge can explicitly express the relationship between the relationships of entities.

Since most of the model layers of the power transformer fault knowledge map can be determined by the transformer topology diagram and transformer operation inspection guidelines, the fault attribute nodes still need to be summarized and supplemented from the knowledge layer. Therefore, this paper uses a combination of top-down and bottom-up methods to construct a power transformer fault knowledge map. The construction process of the knowledge graph is shown in Figure 1

The specific steps are as follows:

First, determine the various components of the power transformer and their electrical and mechanical connections through the analysis of the transformer topology. Generate the first-level pattern diagram of the knowledge map. Determine the type of failure of each component, the operation mode, and the maintenance mode after the failure. A top-down approach is used to design the initial model layer of the knowledge map.

Then, under the guidance of the model layer, a bottom-up approach is used to perform entity, relationship, and relationship analysis of the operation and maintenance records based on the graph CNN method. The three elements of attributes are extracted to form a high-quality knowledge expression.

Finally, the extracted faulty components and fault time are added to the model layer as attribute nodes to complete the update of the model layer. So far, the construction of the power transformer fault knowledge map is completed, and the knowledge map is stored in the form of triples.

The model layer of the power transformer fault knowledge map is mainly composed of three core elements, including fault components, fault time, and fault location, as well as their interrelationships. Firstly, compile the first model layer of the fault knowledge map according to the power transformer structure diagram, and extract the relationship between the components of the transformer. Secondly, according to the transformer maintenance guide, extract the relationship between the components of the transformer and the fault. Finally, Form the second layer of the knowledge graph model layer.

The precision and recall as the evaluation indexes are introduced in the transformer concurrent fault diagnosis method based on graph neural network and knowledge graph. Among them, the accuracy rate refers to the proportion of the correct target in the total number of targets detected by the model, which is usually called the precision rate. The recall rate refers to the ratio of the number of correct targets detected by the model to the total number of correct targets, also called the recall rate.

As shown in Figure 5, suppose that in the correct sample library, the number of concurrent failures detected by the model is A, and the number of undetected concurrent failures is B. In the wrong sample library, the number of concurrent failures detected by the model is B. The number of failures is C, and the number of concurrent failures detected is D. The calculation formula for the accuracy rate p and the recall rate R is as follow: P = C C + D (1) R = C A + C (2)

Based on the PyTorch deep learning computing environment, a comparative experiment of lightweight graph convolution and standard graph convolution, and a comparative experiment of lightweight graph convolution with different layers are set to verify the approach proposed in this paper. The relevant parameters of the model are shown in Table 1. The model test verification of this paper is carried out on the server of the laboratory, the configuration of server hardware and software environment with 8-core CPU, 32 GB memory and an NVIDIA Tesla P4 graphics card with 8 GB video memory.

Responding to the problem of modeling and inference of safety hazards due to concurrency failures, this paper introduces the knowledge graph and uses its ability to model relational networks to extract a transformer fault relational network from operation and maintenance data. At the same time, based on graphs, the product neural network is marked by extracting a large number of sub-graphs of the transformer fault knowledge map, and a graph convolutional network is trained to concurrent fault mining. By comparing with the standard graph convolution network, the accuracy and recall rate of graph neural network mining on the knowledge graph are better than traditional algorithms such as convolutional neural networks, which can meet the effectiveness and robustness of concurrent fault mining. Currently, the research in this paper focuses primarily on the analysis and investigation of known companion faults. Subsequent research will consider the situation of unknown concurrent faults and study the discovery and update mechanism of unknown concurrent faults.