A transformer fault online diagnosis method based on a graph neural network

Abstract

The present application relates to a kind of transformer fault online diagnosis method based on graph neural network.By mechanism analysis to the typical fault of transformer, the characteristic parameter capable of fully representing the fault is obtained, then all characteristic parameters are constructed into a graph network, and the fusion processing of all parameter data is completed.After the establishment of graph network, the dynamic weight of each parameter is continuously trained, then the center node is selected to carry out node aggregation and construct the adjacency matrix of each node, the center node is gradually aggregated from the outer layer, and finally the probability of the occurrence of a certain fault of transformer is output, the construction of graph network needs to continuously train weight.The present application can establish simple graph network using multi-source fusion data to diagnose transformer fault, solve the pain point problem that large sample size leads to difficult data collection and single data type leads to low diagnosis accuracy, in addition, it also combines the correlation between faults under actual operating conditions, and outputs the probability of multiple faults possibly occurring.

Description

Technical Field

[0001] This invention belongs to the field of transformer fault handling, and specifically relates to an online fault diagnosis method for transformers based on graph neural networks. Background Technology

[0002] Existing transformer fault diagnosis technologies typically involve collecting and processing abnormal data from transformer faults, determining the matching correspondence between input feature parameters and output quantities, designing a neural network, and continuously training on a large database. Finally, when any fault data occurs, the neural network outputs a specific type of transformer fault. Other improved neural network methods follow the same principle, only continuously improving in speed and accuracy. While these neural network methods can quickly determine the type of transformer fault based on abnormal data, they are limited by the availability of a large fault database. Data on equipment malfunctions under actual operating conditions is extremely limited, making it impossible to obtain a large number of effective fault data samples. Furthermore, the data is primarily structured data such as online monitoring oil chromatography data and core and clamp current, generally requiring on-site inspection by maintenance personnel for final confirmation. In addition, most transformer faults are highly correlated; even with different faults, data such as oil temperature and winding temperature can fluctuate significantly. Therefore, existing technologies have the following drawbacks:

[0003] ① Most existing technologies are limited by the need for a large fault database, while it is impossible to obtain a large number of effective fault data samples under actual field operating conditions.

[0004] ② Currently, the data used for online diagnosis of transformer faults are mainly online monitoring oil chromatography data, core and clamp current and other structured data. The data types are relatively simple and cannot fully reflect the authenticity of the fault. Generally, on-site maintenance personnel need to conduct manual inspection and verification again, and the accuracy still needs to be further improved.

[0005] ③ In addition, for transformer faults, most faults are highly correlated. Even when different faults occur, data such as oil temperature and winding temperature will fluctuate significantly and abnormally. Current diagnostic methods cannot output the probability of multiple faults at the same time.

[0006] Therefore, the online diagnostic technology for field transformers aims to achieve a new diagnostic method with richer sample data types, smaller data volume, and the ability to consider the correlation of transformer faults. Summary of the Invention

[0007] The purpose of this invention is to provide an online transformer fault diagnosis method based on graph neural networks. This method can use multi-source fusion data to build a simple graph network for transformer fault diagnosis, solving the problems of traditional methods where large sample sizes lead to difficult data collection and single data types lead to low diagnostic accuracy. In addition, it also takes into account the correlation between faults under actual on-site operating conditions, and can simultaneously output the probability of multiple faults.

[0008] To achieve the above objectives, the technical solution of this invention is: an online transformer fault diagnosis method based on graph neural networks. This method obtains feature parameters that can comprehensively characterize the typical transformer faults by performing mechanism analysis on typical transformer faults. Then, all feature parameters are constructed into a graph network. After the graph network is established, the dynamic weights of each feature parameter are continuously trained. Then, the central node is selected for node aggregation and the adjacency matrix of each node is constructed. The central node is aggregated step by step from the outer layer, and finally, the probability of a certain transformer fault occurring is output.

[0009] Compared to existing technologies, this invention has the following advantages: It overcomes the need for large anomaly (fault) databases in existing diagnostic methods; it improves fault feature parameters by incorporating structured, semi-structured, and unstructured data; and it solves the problem of strong correlation between typical transformer faults in the field, enabling the online diagnostic model to simultaneously output the probabilities of multiple possible faults when the same anomaly (fault) data arrives. This application proposes a novel diagnostic method with richer sample data types, lower data requirements, and the ability to consider the correlation of transformer faults. Attached Figure Description

[0010] Figure 1 A graph network is established for the operation of a transformer pressure relief valve in an example of the present invention.

[0011] Figure 2 This is a network number for an example diagram of the present invention. Detailed Implementation

[0012] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0013] This invention discloses an online fault diagnosis method for transformers based on graph neural networks. By performing mechanistic analysis on typical transformer faults, feature parameters that can comprehensively characterize the corresponding typical transformer faults are obtained. Then, all feature parameters are constructed into a graph network. After the graph network is established, the dynamic weights of each feature parameter are continuously trained. Then, the central node is selected for node aggregation and the adjacency matrix of each node is constructed. The central node is aggregated step by step from the outer layer, and finally the probability of a certain fault occurring in the transformer is output.

[0014] The following is a detailed implementation process of the present invention.

[0015] This invention obtains more comprehensive fault characteristic state quantities (structured, unstructured, and semi-structured data) by analyzing the mechanisms of typical transformer faults, and then performs fusion processing on multi-source data; subsequently, all state parameters are constructed into a graph network and dynamic weights are trained, and central nodes are selected for layer-by-layer aggregation to obtain the probability of different fault occurrences. The technical solution proposed in this application is as follows:

[0016] 1. Establishment of a fault characteristic parameter diagnostic model;

[0017] 2. Establishment of graph networks;

[0018] 3. Node aggregation in graph networks;

[0019] Details are as follows:

[0020] 1. Establishment of fault characteristic parameter diagnostic model:

[0021] Based on the mechanism of transformer failure and real-time monitoring data during equipment anomalies (faults), characteristic parameters of typical transformer faults are identified to establish diagnostic models for these faults. Here, only the light gas activation and pressure relief valve activation of the transformer body are used as examples. The corresponding structured, semi-structured, and unstructured data types and their characteristic parameters are shown in Tables 1 and 2, providing a richer and more comprehensive evaluation of the data types.

[0022] Table 1

[0023]

[0024] Table 2

[0025]

[0026]

[0027] As can be seen from the two common typical transformer faults mentioned above, there is a high degree of repetition of the same characteristic parameters, indicating a strong correlation. Previous conventional diagnostic methods could only output the occurrence of a specific fault by matching the data, but could not accurately output the probability of several faults occurring simultaneously. This invention establishes a more comprehensive diagnostic model using multi-type data characteristic parameters and employs a graph neural network method to address this urgent problem in the field.

[0028] 2. Establishment of graph networks:

[0029] The aforementioned fault characteristic parameters that characterize the light gas operation (body) of a transformer are arbitrarily constructed into a graph network. The weights are continuously adjusted to fit the current graph network by outputting the probabilities of various typical faults based on a given anomaly (fault) data. Taking the operation of a transformer pressure relief valve as an example, its graph network is as follows: Figure 1 As shown.

[0030] Because the graph network is built randomly, it is necessary to obtain parameter weights suitable for each type of network through training. The training process mainly relies on a set of abnormal (fault) data, threshold data, and normal data from the field, as shown in Table 3.

[0031] Table 3

[0032]

[0033] This explanation uses only two faults as examples; the diagnostic method remains identical when other faults are introduced. Let Y0 be the probability of the normal state, Y1 be the probability of the transformer pressure relief valve actuating, and Y2 be the probability of the transformer light gas actuating. X0 is the total score of each characteristic parameter under the normal state, X1 is the total score of each characteristic parameter under the transformer pressure relief valve actuating, and X2 is the total score of each characteristic parameter under the transformer light gas actuating.

[0034] X0=∑x 0i w 0i

[0035] X1=∑x 1i w 1i

[0036] X2=∑x 2i w 2i

[0037] Where x 0i This is the real-time data of each parameter under normal conditions, w 0i These are the dynamic weights corresponding to each parameter; x 1i These are abnormal (fault) data of various parameters under the action of the transformer pressure relief valve. 1i These are the dynamic weights corresponding to each parameter; x 2i These are abnormal (fault) data of various parameters under the light gas operation of the transformer, w 2i These are the dynamic weights corresponding to each parameter, where i is the i-th feature parameter under this fault. The dynamic adjustment of the weights in the training graph network is based on the abnormal (fault) data of the transformer pressure relief valve operation at the site. The output target requires ensuring that Y1>Y2>Y0 after continuous training. The dynamic adjustment of the weights during the training process is carried out according to backpropagation, as shown in the following formula. It is necessary to take the derivative of each parameter under this fault to obtain the dynamic weights of each parameter under the graph network, and then train iteratively to ensure that the output Y1 is maximized under this abnormal (fault) data.

[0038]

[0039] If backpropagation is used during the dynamic weight adjustment process, then...

[0040]

[0041] In the above formula, n is the approximation value. When the input is abnormal (fault) data under this fault, n is 1; otherwise, it is 0. Here, the example is abnormal (fault) data input under this fault, so n = 1. To solve for the dynamic weight value, it is necessary to perform Y1 versus w in the above formula. 1i Taking the derivative, we have

[0042] Therefore, the dynamic weight w′ 1i Values

[0043]

[0044] w 1i This refers to the weights of various parameters under the action of the transformer pressure relief valve. The initial weights here are manually assigned based on a comparison of threshold data and normal data. w′ 1i These are the dynamic weights of various characteristic parameters under the action of the transformer pressure relief valve. η is a constant, typically set to 0.01. During iterative training, the weights are continuously updated to ensure that the probability value of all faults is maximized by Y1. Even if the network order is different, the graph neural network of this invention can still output the same fault probability result through training with the above weights.

[0045] 3. Node aggregation in graph networks:

[0046] The graph network composed of the above feature parameters is numbered, and the numbering order does not affect the final output. Then, the central node is selected for layer-by-layer aggregation, and finally the probability of transformer fault occurrence is obtained. The graph network numbering is shown in the figure below.

[0047] First, construct the adjacency matrix of the 12 nodes under this fault, that is, the adjacent nodes are set to 1 (including themselves), and the non-adjacent nodes are set to 0. The adjacency matrix of each node is shown below.

[0048]

[0049] Next, we arbitrarily select node ② as the center node and search for adjacent nodes around ②, resulting in nodes ①③⑥⑦⑩. Then, we continue searching for lower-level adjacent nodes based on the found adjacent nodes ①③⑥⑦⑩. According to the graph network, the lower-level nodes corresponding to ① are (4), (5), (11), and (12); the lower-level nodes corresponding to ③ are (8) and (9); and there are no corresponding nodes for ⑥⑦⑩. Here, we first retain the data for the second-level nodes ①③⑥⑦⑩ and then aggregate them using the average aggregation function. For nodes ①③⑥⑦⑩, we have...

[0050]

[0051]

[0052] Finally, the result obtained by aggregating the data from the central node ② is the probability x of the failure occurring. ② Each of the other faults has its own feature parameter graph network. After training dynamic weights, it is possible to output the probability of multiple faults occurring simultaneously under a set of abnormal data for faults with strong correlation with transformers.

[0053]

[0054] This section will only take the action of the transformer pressure relief valve as an example for detailed explanation. When conducting online diagnosis on site, all typical transformer faults are considered at the same time. Each fault corresponds to a graph network and is output simultaneously to obtain the probability of the fault occurring. For transformer faults with strong correlation, this provides very important early warning instructions for on-site operation and maintenance personnel.

[0055] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A transformer fault online diagnosis method based on a graph neural network, characterized in that, The feature parameters capable of comprehensively representing the corresponding transformer typical fault are obtained by mechanism analysis of the transformer typical fault, and then all the feature parameters are constructed into a graph network. The dynamic weights of the feature parameters are continuously trained after the graph network is established. Then the central node is selected for node aggregation and the adjacency matrix of each node is constructed. The central node is gradually aggregated from the outer layer, and finally the probability of the occurrence of a certain fault of the transformer is output. The feature parameters capable of comprehensively representing the corresponding transformer typical fault are obtained by mechanism analysis of the transformer typical fault in the following way: according to the mechanism of the transformer typical fault and the real-time monitoring data of the transformer fault on site, the feature parameters of different transformer typical faults are sorted out, including: structured data, unstructured data, semi-structured data, and then the structured data, unstructured data and semi-structured data are fused to establish the feature parameter diagnosis model of the transformer typical fault; the feature parameters capable of representing the corresponding transformer typical fault are constructed into a graph network, and the probability of various transformer typical faults is output through a certain fault data to continuously adjust the weight suitable for the current graph network; the specific process is as follows: Suppose that two kinds of transformer typical faults are transformer pressure relief valve action and transformer light gas action; let Y0 be the probability of the normal state, Y1 be the probability of the transformer pressure relief valve action, and Y2 be the probability of the transformer light gas action; let X0 be the total score of each feature parameter under the normal state, X1 be the total score of each feature parameter under the transformer pressure relief valve action, and X2 be the total score of each feature parameter under the transformer light gas action; then we have: wherein x0i is real-time data of each characteristic parameter under normal state, w0i is the dynamic weight corresponding to each characteristic parameter; x1i is fault data of each characteristic parameter under action of transformer pressure relief valve, w1i is the dynamic weight corresponding to each characteristic parameter; x2i is fault data of each characteristic parameter under action of transformer light gas, w2i is the dynamic weight corresponding to each characteristic parameter, and i is the i-th characteristic parameter under typical fault of the corresponding transformer. 0i 2i wherein x0i is real-time data of each characteristic parameter under normal state, w0i is the dynamic weight corresponding to each characteristic parameter; x1i is fault data of each characteristic parameter under action of transformer pressure relief valve, w1i is the dynamic weight corresponding to each characteristic parameter; x2i is fault data of each characteristic parameter under action of transformer light gas, w2i is the dynamic weight corresponding to each characteristic parameter, and i is the i-th characteristic parameter under typical fault of the corresponding transformer.​ The dynamic adjustment of the weight of the graph network is based on the result output of the fault data of the transformer pressure relief valve action on site, and the output target requires that Y1>Y2>Y0 after continuous training; the dynamic adjustment of the weight during the training process is based on back propagation, as shown in the following formula, which needs to derive each feature parameter under the transformer pressure relief valve action to obtain the dynamic weight of each feature parameter under the graph network, and then cyclically trains to ensure that Y1 is the maximum when the fault data of the transformer pressure relief valve action is output: The back propagation method is used in the process of dynamic weight adjustment, so we have In the above formula, n is the approximation value, when the input is the fault data of the transformer pressure relief valve action, n takes 1, otherwise it takes 0; in this case, the fault data of the transformer pressure relief valve action is input, so n=1; to solve the dynamic weight value, Y1 needs to be derived with respect to w1i, so we have Thus the dynamic weight w' 1i is taken as w1i is the weight of each feature parameter under the transformer pressure relief valve action, which is initialized according to the comparison between the threshold data and the normal data; w′1i is the dynamic weight of each feature parameter under the transformer pressure relief valve action, η is a constant, and the weight is continuously updated during the cyclic training process to ensure that the probability value of all faults output is Y1 maximum.

2. The transformer fault online diagnosis method based on a graph neural network according to claim 1, characterized in that, The center node is selected for node aggregation and the adjacency matrix of each node is constructed, the center node is gradually aggregated from the outer layer, and finally the probability of occurrence of a certain fault of the transformer is output. The method is as follows: the graph network composed of all characteristic parameters is numbered, then the center node is selected for layer-by-layer aggregation, and finally the probability of occurrence of a certain fault of the transformer is obtained; the specific method is as follows: First, the graph network composed of all characteristic parameters under the corresponding fault is numbered, and it is assumed that the number is 12, i.e. the node is 12; Then, the adjacency matrix of 12 nodes under the corresponding fault is constructed, i.e. 1 for adjacent nodes and 0 for non-adjacent nodes, and the adjacency matrix of each node is as follows: Next, the number ② is randomly selected as the center node, and the adjacent nodes are found with ② as the center, so there are ①, ③, ⑥, ⑦ and ⑩, then the lower adjacent nodes are found according to the found adjacent nodes ①, ③, ⑥, ⑦ and ⑩; according to the graph network, it can be obtained that the lower nodes corresponding to ① are (4), (5), (11) and (12); the lower nodes corresponding to ③ are (8) and (9); the lower nodes corresponding to ⑥, ⑦ and ⑩ are null; here, the data of the second layer nodes ①, ③, ⑥, ⑦ and ⑩ is reserved first, and the average aggregation function is used for aggregation; there are Finally, the result obtained by aggregating the center node ② is the probability x② of occurrence of the corresponding fault, and the formula is as follows: each fault corresponds to a characteristic parameter graph network, and after training the dynamic weight, a group of abnormal data can output the probability of occurrence of multiple faults under the transformer correlation of the transformer is strong;

Images