A transformer operation fault detection method and system

Abstract

The application is suitable for the technical field of fault detection, and particularly relates to a transformer operation fault detection method and system.The method comprises the following steps: finding out transformers within a preset range, setting a plurality of monitoring indexes, editing evaluation rules, comparing the monitoring indexes and the evaluation rules, and clustering the transformers into stable groups and risk groups; creating a global model, collecting operation parameters of the transformers, obtaining historical data of operation parameters of all transformers in the stable groups, training the global model, and obtaining a parameter group composed of a plurality of sets of model parameters, wherein each transformer corresponds to a set of model parameters.The application can implement macroscopic monitoring of operation parameters of all transformers by using a global perception model through construction of a multi-level monitoring architecture, realize preliminary screening of abnormalities, trigger a local monitoring model to perform in-depth detailed analysis on relevant operation parameters when an abnormal signal is detected, and greatly improve the accuracy and reliability of fault identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault detection, in particular to a transformer operation fault detection method and system. BACKGROUND

[0002] The transformer operation fault refers to the phenomenon that the operation state is abnormal or even fails due to the comprehensive action of electrical, thermal, mechanical and chemical factors during the long-term operation of the transformer. Common operation faults include winding inter-turn short circuit and inter-phase short circuit caused by insulation aging, moisture or breakdown.

[0003] Under the existing data conditions, a machine learning model based on historical operation data is usually used to diagnose the transformer fault. However, for some transformers with good operation environment and long-term stable equipment state, the abnormal data samples are extremely scarce, which makes it difficult for the model to effectively learn the specific fault features and affects the diagnosis accuracy. On the contrary, in some areas with complex operation environment or serious aging, transformer faults and abnormal events occur frequently, and the abnormal data is relatively abundant, providing more effective samples for model training.

[0004] Therefore, how to use the rich abnormal data of other areas to improve the fault diagnosis capability of the area in the case of scarce abnormal samples is a technical problem to be solved by the present application. SUMMARY

[0005] The present application aims to provide a transformer operation fault detection method and system to solve the problem of how to use the rich abnormal data of other areas to improve the fault diagnosis capability of the area in the case of scarce abnormal samples.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0007] A transformer operation fault detection method, the method comprising:

[0008] Finding out the transformers within a preset range, setting a plurality of monitoring indicators, editing evaluation rules, comparing the monitoring indicators and the evaluation rules, and clustering the transformers into stable groups and risk groups;

[0009] Creating a global model, collecting the operation parameters of the transformers, obtaining the historical data of the operation parameters of all the transformers in the stable groups, and training the global model to obtain a parameter group composed of a plurality of model parameters, wherein each transformer corresponds to a set of model parameters, and each model parameter at least includes a weight coefficient and a threshold value;

[0010] Collect historical data of all transformers in the risk group, select abnormal data, construct a training set, train the pre-constructed conditional generation model, extract the running parameters of each transformer in the stable group under normal state, obtain the target parameter set, add random noise, input into the conditional generation model, output to obtain simulated data, and train the global model to correct the parameter set;

[0011] Based on the parameter set, calculate the mean coefficient and deploy it to the global model to obtain the global perception model, deploy the model parameters to the global model to obtain the local detection model, integrate the global perception model and the local detection model, generate a multi-level monitoring architecture, and write the running parameters in real time.

[0012] Further, the step of finding out the transformer within the preset range and setting a plurality of monitoring indicators and editing the evaluation rule comprises:

[0013] Based on the evaluation rule, set the fluctuation range of each monitoring indicator, compare the running parameters and the fluctuation range, and select the abnormal equipment;

[0014] Collect attribute data of each transformer, wherein the attribute data at least includes position and type, write the attribute data of the abnormal equipment into a preset template, generate a maintenance task, and send it to a preset terminal.

[0015] Further, the step of creating a global model, collecting running parameters of the transformer, obtaining historical data of running parameters of all transformers in the stable group, and training the global model comprises:

[0016] Calculate the mathematical features of the running parameters, wherein the mathematical features at least include mean, variance and median;

[0017] Based on the mathematical features corresponding to each transformer in the stable group, the stable group is refined into a plurality of subgroups, and a parameter sharing mechanism based on the subgroups is constructed.

[0018] Further, the step of collecting historical data of all transformers in the risk group, selecting abnormal data, and constructing a training set comprises:

[0019] Determine the fault type corresponding to each abnormal data, based on the fault type, divide the training set into a plurality of subsets;

[0020] Insert a label generated by the fault type into the subset.

[0021] Further, the step of calculating the mean coefficient based on the parameter set and deploying it to the global model to obtain the global perception model, deploying the model parameters to the global model to obtain the local detection model comprises:

[0022] Collecting computing resources in each preset range, and selecting an edge node;

[0023] Mounting the global perception model to the edge node and mounting the local detection model to the preset cloud node.

[0024] Further, the step of generating a multi-level monitoring architecture and writing the running parameters in real time includes:

[0025] Editing a federated learning rule and deploying it to the multi-level monitoring architecture;

[0026] Establishing a mapping between the preset range and the multi-level monitoring architecture and performing federated aggregation.

[0027] Further, the system includes:

[0028] A clustering module for finding transformers in a preset range, setting a plurality of monitoring indicators, editing evaluation rules, comparing the monitoring indicators and the evaluation rules, and clustering the transformers into stable groups and risk groups;

[0029] A obtaining module for creating a global model, collecting running parameters of the transformers, obtaining historical data of running parameters of all transformers in the stable groups, and training the global model to obtain a parameter group composed of a plurality of model parameters, wherein each transformer corresponds to a set of model parameters, and each model parameter at least includes a weight coefficient and a threshold value;

[0030] A correction module for collecting historical data of all transformers in the risk groups, selecting abnormal data, constructing a training set, training a pre-constructed conditional generation model, extracting running parameters of each transformer in the stable groups under normal state to obtain a target parameter set, adding random noise, inputting into the conditional generation model, outputting to obtain simulation data, and training the global model to correct the parameter group;

[0031] A writing module for calculating a mean coefficient according to the parameter group and deploying it to the global model to obtain a global perception model, deploying model parameters to the global model to obtain a local detection model, integrating the global perception model and the local detection model, generating a multi-level monitoring architecture, and writing the running parameters in real time.

[0032] Further, the clustering module includes:

[0033] A setting unit for setting a fluctuation range of each monitoring indicator according to the evaluation rules, comparing the running parameters and the fluctuation range, and selecting abnormal equipment;

[0034] The collection unit is used for collecting attribute data of each transformer, wherein the attribute data at least includes position and model number, writing the attribute data of the abnormal device into a preset template, generating a maintenance task, and sending to a preset terminal.

[0035] Further, the obtaining module comprises:

[0036] The calculation unit is used for calculating mathematical features of the operation parameters, wherein the mathematical features at least include mean value, variance and median value;

[0037] The sharing unit is used for refining the stable group into a plurality of subgroups according to the mathematical features corresponding to each transformer in the stable group, and constructing a parameter sharing mechanism based on the subgroups.

[0038] Further, the correction module comprises:

[0039] The cutting unit is used for determining the fault type corresponding to each abnormal data, and cutting the training set into a plurality of subsets based on the fault type;

[0040] The insertion unit is used for inserting labels generated by the fault type into the subsets.

[0041] Compared with the prior art, the present application has the following beneficial effects:

[0042] By setting the monitoring indicators, the key parameters can be determined, and the operation state of each transformer can be accurately identified, thereby improving the detection efficiency. By constructing the global model and setting the model parameters, individualized parameters can be provided for each transformer to reflect its unique operation characteristics and environmental differences, thereby realizing individualized diagnosis and improving the accuracy of fault detection. By constructing the conditional generation model and training it using the target parameter set, the data of the area rich in abnormal data can be used to train the conditional generation model, so that the conditional generation model has the ability to simulate abnormal data, and the accuracy of the simulated abnormal data is improved. By using the correction parameter group, the simulated abnormal data can be used to train the global model, thereby improving the identification ability of the global model for abnormal data and alleviating the problem of abnormal data scarcity. By constructing the multi-level monitoring architecture, the global perception model can be used to macroscopically monitor the operation parameters of all transformers, and preliminary screening of abnormalities can be performed. When an abnormal signal is detected, the local detection model is triggered to perform in-depth detailed analysis on the related operation parameters, thereby greatly improving the accuracy and reliability of fault identification. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The flowchart of the transformer operation fault detection method provided by the embodiment of the present application is shown in the figure.

[0044] Figure 2The first sub-flow block diagram of the transformer operation fault detection method provided by the embodiment of the present application is shown in the figure.

[0045] Figure 3 The second sub-flow block diagram of the transformer operation fault detection method provided by the embodiment of the present application is shown in the figure.

[0046] Figure 4 The third sub-flow block diagram of the transformer operation fault detection method provided by the embodiment of the present application is shown in the figure.

[0047] Figure 5 The fourth sub-flow block diagram of the transformer operation fault detection method provided by the embodiment of the present application is shown in the figure.

[0048] Figure 6 The component block diagram of the transformer operation fault detection system provided by the embodiment of the present application is shown in the figure.

[0049] Figure 7 The component block diagram of the clustering module in the transformer operation fault detection system provided by the embodiment of the present application is shown in the figure.

[0050] Figure 8 The component block diagram of the obtaining module in the transformer operation fault detection system provided by the embodiment of the present application is shown in the figure.

[0051] Figure 9 The component block diagram of the correction module in the transformer operation fault detection system provided by the embodiment of the present application is shown in the figure.

[0052] Figure 10 The component block diagram of the writing module in the transformer operation fault detection system provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0054] In embodiment 1, Figure 1 The transformer operation fault detection method implementation process provided by the embodiment of the present application is shown in the figure, and the following is described in detail as follows:

[0055] S100: find out the transformer in the preset range, set several monitoring indexes, edit evaluation rules, compare the monitoring indexes and evaluation rules, and cluster the transformer into stable groups and risk groups.

[0056] Find the transformer in the preset range, where the preset range refers to the range that needs to be detected for faults, for example, a certain transformer substation or administrative region; set several key monitoring indicators for the transformer, such as temperature, load, current, voltage fluctuation, and partial discharge, etc., and formulate evaluation rules, which are usually based on industry standards, equipment historical performance, and safety thresholds, etc., and can also be formulated by equipment manufacturers and operation and maintenance teams, etc., and the evaluation rules are mainly used for abnormal judgment of monitoring indicators; by comparing the real-time values of the monitoring indicators with the evaluation rules, the health status of each transformer is identified, and the transformers are divided into stable groups and risk groups.

[0057] For example, assuming that a certain power operation and maintenance company has demarcated a certain area in the power grid of a city as a preset range, 100 transformers in the area are found, and monitoring indicators are set, specifically: transformer oil temperature, load current, and partial discharge value, and the real-time values of each monitoring indicator are collected, wherein the evaluation rule corresponding to the transformer oil temperature is: the temperature exceeds 85℃, and the duration is more than 15 minutes, which has a safety risk, the load current exceeds 90% of the rated capacity, which is considered as high load operation, and the partial discharge value exceeds a certain threshold, which indicates a potential fault; wherein 10 transformers have appeared safety risk or potential fault in the evaluation rule in the last 3 days, and the 10 transformers are divided into a risk group, and the remaining 90 are divided into a stable group.

[0058] S200: Create a global model, collect the operating parameters of the transformers, obtain the historical data of the operating parameters of all transformers in the stable group, and train the global model to obtain a parameter group composed of a plurality of model parameters, wherein each transformer corresponds to a set of model parameters, and each model parameter at least includes: weight coefficient and threshold value.

[0059] Local features of time series data or image data are extracted using multiple convolution kernels, and these features are fused through multiple convolution and pooling operations to capture information at different scales and levels, and a global model is established through feature integration and abstract representation, wherein the global model is a fault analysis model, which is mainly used to identify potential abnormalities in operating parameters; collect the operating parameters of all transformers, it should be noted that the operating parameters here are the real-time values of the monitoring indicators described above; continuously collect the historical operating data of all transformers in the stable group under different operating states and environmental conditions, and train the global model, during the training process, repeatedly adjust the weight of each convolution kernel in the model, fine-tune the threshold of each monitoring indicator, and the set of weight and threshold obtained is defined as model parameter, wherein each transformer corresponds to a set of model parameters, and the model parameters corresponding to all transformers are integrated to obtain a parameter group.

[0060] In actual production environment, due to the geographical location, system architecture and load condition of the transformer, there are certain subtle differences in the threshold of each monitoring index; therefore, after training the global model based on the historical operation data of each transformer, the obtained parameters will also be different, in this application, the global model is similar to the initial model, the historical data of each transformer is used to train the initial model, a set of parameters is obtained, and according to the geographical location, system architecture and load condition of the transformer, the threshold of each monitoring index is adjusted, the set of parameters and threshold obtained after training is defined as model parameters, each transformer corresponds to a set of model parameters, when used, the set of model parameters is written into the initialized global model, and a unique and personalized fault analysis model is obtained.

[0061] S300: Collect the historical data of all transformers in the risk group, select abnormal data, construct a training set, and train the pre-constructed conditional generation model, extract the running parameters of each transformer in the stable group under normal state, obtain the target parameter set, add random noise, input into the conditional generation model, output the simulated data, and train the global model to correct the parameter set.

[0062] Collect the historical data of all transformers in the risk group, and define the historical data under the state of safety risk or potential fault as abnormal data, integrate all abnormal data to generate a training set; create a conditional generation model, which can simulate corresponding abnormal data according to the running parameters of the transformer under normal state under certain conditions (geographical location, system architecture and load condition, etc.); the conditional generation model learns the distribution of normal data to simulate the running parameters, wherein the conditional generation model can be a conditional generative adversarial network in the prior art, or can be obtained by adjusting the conditional constraint transformation or conditional embedding of a general generation model; for example, in transformer monitoring, the condition is the running parameter of the transformer, the simulated data conforming to the current state characteristics is generated, and the conditional generation model is trained by a large amount of real data, which can more accurately simulate the running parameters under abnormal state.

[0063] Collect the running parameters of each transformer in the stable group under normal state, integrate to generate a target parameter set, add random noise to the target parameter set, simulate the fluctuations of the transformers in the stable group in the actual environment, input the target parameter set with noise into the conditional generation model, output the simulated data through the learned distribution characteristics of the running parameters under stable state, wherein the simulated data refers to the simulated running parameters of the transformers in the stable group, and the simulated data is used to train the global model to correct the corresponding parameter set.

[0064] In the present application, after determining the model parameters corresponding to each transformer, the operating parameters of each transformer in the stable group under normal state are used to generate a condition model, simulate abnormal data of the transformers in the stable group, and each transformer in the stable group corresponds to a set of abnormal data. The corresponding global model is trained using the abnormal data, and the model parameters are adjusted. In other words, the global model is trained using real stable data and simulated abnormal data, which can effectively make up for the problem of insufficient number of abnormal samples of transformers in the stable group, improve the generalization ability of the global model in identifying abnormal states, and finally through continuous iteration training, the global model can accurately capture the feature changes of the transformer under different operating states, and improve the accuracy of fault identification.

[0065] S400: Based on the parameter set, the mean coefficient is calculated and deployed into the global model to obtain the global perception model. The model parameters are deployed into the global model to obtain the local detection model. The global perception model and the local detection model are integrated to generate a multi-level monitoring architecture, and the running parameters are written in real time.

[0066] The average value of the weight coefficients of all parameter sets is calculated and defined as the mean coefficient. The mean coefficient can highlight the commonality of the overall operating state. The calculated mean coefficient is used as a new weight coefficient and deployed into the initialized global model to obtain the global perception model. The global perception model can roughly monitor and preliminarily warn the operating state of the transformer within the entire preset range and has strong generalization ability. The global model is initialized again, and the model parameters corresponding to each transformer in the stable group are deployed into the global model to obtain the local detection model. The local detection model has strong pertinence and flexibility, can analyze the subtle abnormal features of the transformer in combination with the geographical location, system architecture and load condition of the specific transformer, effectively identify early fault signals, and integrate the global perception model and the local detection model to generate a multi-level monitoring architecture.

[0067] In the multi-level monitoring architecture, the global perception model is responsible for comprehensive rough monitoring and trend analysis of the operating state of the transformer within the entire preset range, and quickly identifies potential risk areas. The local detection model is aimed at a single transformer and combines its specific operating environment and historical behavior data to perform deep abnormal detection and fault diagnosis. The multi-level monitoring architecture refers to a layer-by-layer progressive monitoring system formed by a macroscopic monitoring layer composed of the global perception model and a microscopic detection layer composed of the local detection model. The real-time operating parameters generated by the transformer are analyzed and processed using the multi-level monitoring architecture.

[0068] In embodiment 2, Figure 2The transformer operation fault detection method provided by the embodiment of the present application is shown, and the steps of finding out the transformers within the preset range, setting a plurality of monitoring indexes, and editing evaluation rules are described in detail as follows:

[0069] S101: Based on the evaluation rules, set the fluctuation range of each monitoring index, compare the operation parameters and the fluctuation range, and select the abnormal equipment.

[0070] According to the evaluation rules, set a corresponding fluctuation range for each monitoring index, and define the transformer whose operation parameter exceeds the fluctuation range as an abnormal equipment.

[0071] S102: Collect the attribute data of each transformer, wherein the attribute data at least includes: position and type, write the attribute data of the abnormal equipment into a preset template, generate a maintenance task, and send it to a preset terminal.

[0072] Collect the attribute data of each transformer, when the abnormal equipment is identified, write the attribute data of the abnormal equipment into a preset task template, the template defines the key fields required by the maintenance task in advance, such as equipment position, type, fault type and recommended treatment scheme, etc., by writing the content in the attribute data into the corresponding position of the preset template, generate a maintenance task, and send it to a preset terminal, wherein the preset terminal can be a mobile workstation, tablet computer or mobile terminal of the operation and maintenance personnel, etc.

[0073] In embodiment 3, Figure 3 The transformer operation fault detection method provided by the embodiment of the present application is shown, and the steps of creating a global model, collecting the operation parameters of the transformer, obtaining the historical data of the operation parameters of all transformers in the stable group, and training the global model are described in detail as follows:

[0074] S201: Calculate the mathematical features of the operation parameters, wherein the mathematical features at least include: mean, variance and median.

[0075] Calculate the mean, variance and median of each monitoring index in the operation parameters, in the specific calculation process, the mathematical features of the operation parameters under each time window can be calculated by constructing a time window and a multi-point sampling method.

[0076] S202: Based on the mathematical features corresponding to each transformer in the stable group, refine the stable group into a plurality of subgroups, and construct a parameter sharing mechanism based on the subgroups.

[0077] According to the mathematical characteristics, the transformers in the stable group are further divided into multiple subgroups; for example, the transformers with the root mean square of temperature, load and current in the preset range are combined into a subgroup, and a parameter sharing mechanism is constructed, that is, the model parameters in the subgroup are uniformly managed and applied, so that the transformers in the same group can share the calculation results and model parameters in model training, running state prediction and abnormality detection, so as to reduce repeated calculation, improve the generalization ability of the global model while improving the overall running analysis efficiency and maintenance decision level of the stable group.

[0078] In embodiment 4, Figure 4 The transformer operation fault detection method implementation process provided by the embodiment of the application is shown, and the steps of collecting the historical data of all transformers in the risk group, selecting abnormal data and constructing a training set are described in detail as follows:

[0079] S301: Determine the fault type corresponding to each abnormal data, and cut the training set into several subsets based on the fault type.

[0080] Determine the fault type corresponding to each abnormal data, for example, winding short circuit, insulation aging, oil temperature too high or partial discharge anomaly, etc. The data containing different fault types are cut from the original training set according to categories to form several relatively independent subsets, each of which contains only abnormal data of one fault type.

[0081] S302: Insert the label generated by the fault type into the subset.

[0082] Insert the label generated by the fault type into the subset, which directly shows the fault type corresponding to each subset. By cutting the training set, the convergence speed and training accuracy of the global model can be greatly improved, and the fault recognition ability of the global model can be enhanced.

[0083] In embodiment 5, Figure 5 The transformer operation fault detection method implementation process provided by the embodiment of the application is shown, and the steps of calculating the mean coefficient based on the parameter group and deploying it into the global model to obtain the global perception model, and deploying the model parameters into the global model to obtain the local detection model are described in detail as follows:

[0084] S401: Collect the computing resources in each preset range and select the edge node.

[0085] The computing resources in the preset range are collected and counted, including but not limited to servers, edge computing devices, industrial gateways and intelligent terminals with computing capability, and the nodes with low latency, high reliability and high availability are selected as edge nodes to process the operation parameters of the transformer near the data source.

[0086] S402: mount the global perception model into the edge node and mount the local detection model into the preset cloud node.

[0087] The global perception model is mounted into the edge node to meet the requirements of low latency and bandwidth saving and ensure the timeliness of fault monitoring. The more computationally intensive and numerous local detection models are deployed to the preset cloud node (in the present application, the number of local detection models is the same as the number of transformers or subgroups). On this basis, the global perception model can be subjected to lightweight processing and run on the edge device in the form of a lightweight service. The cloud end carries out the computing task of the personalized local detection model and uses stronger computing power to perform refined analysis and processing on the preliminarily detected fault data.

[0088] In embodiment 6, Figure 5 The transformer operation fault detection method provided in the embodiment of the present application is shown, and the steps of generating the multi-level monitoring architecture and writing the operation parameters in real time are described in detail as follows:

[0089] S403: edit the federated learning rule and deploy it into the multi-level monitoring architecture.

[0090] In the process of updating the model parameters of the global perception model by using the operation parameters, the operation parameters are not uploaded, but only the model parameters corresponding to each transformer are uploaded. The federated learning rule means that only the model parameters are uploaded, and the operation parameters are not uploaded.

[0091] S404: establish the mapping between the preset range and the multi-level monitoring architecture and perform federated aggregation.

[0092] In each preset range, a corresponding multi-level monitoring architecture is constructed, and the model parameters of the corresponding global perception model are updated. After the update, the model parameters of the global perception model in all preset ranges are calculated again to obtain new model parameters. The federated aggregation means that the model parameters obtained by local training in different preset ranges are combined according to certain rules to obtain a new global perception model. In other embodiments of the present application, the global perception model corresponds to a preset range (assuming that the preset range is a city), and the new global perception model corresponds to all preset ranges (a city or a larger area).

[0093] Figure 6 The composition structure block diagram of the transformer operation fault detection system provided in the embodiment of the present application is shown. The transformer operation fault detection system 1 comprises:

[0094] The clustering module 11 is configured to find transformers within a preset range, set a plurality of monitoring indexes, edit evaluation rules, compare the monitoring indexes and the evaluation rules, and cluster the transformers into stable groups and risk groups.

[0095] The obtaining module 12 is configured to create a global model, collect operation parameters of the transformers, obtain historical data of operation parameters of all transformers in the stable groups, and train the global model to obtain a parameter group composed of a plurality of sets of model parameters, wherein each transformer corresponds to a set of model parameters, and each set of model parameters at least includes a weight coefficient and a threshold value.

[0096] The correction module 13 is configured to collect historical data of all transformers in the risk groups, select abnormal data, construct a training set, train a pre-constructed conditional generation model, extract operation parameters of each transformer in the stable groups in a normal state to obtain a target parameter set, add random noise, input the target parameter set into the conditional generation model, output simulation data, and train the global model to correct the parameter group.

[0097] The writing module 14 is configured to calculate a mean coefficient according to the parameter group, deploy the mean coefficient into the global model to obtain a global perception model, deploy model parameters into the global model to obtain a local detection model, integrate the global perception model and the local detection model to generate a multi-level monitoring architecture, and write the operation parameters in real time.

[0098] Figure 7 A component structure block diagram of a transformer operation fault detection system provided by an embodiment of the present application is shown, and the clustering module 11 includes:

[0099] The setting unit 111 is configured to set a fluctuation range of each monitoring index according to the evaluation rules, compare the operation parameters and the fluctuation range, and select an abnormal device.

[0100] The acquisition unit 112 is configured to acquire attribute data of each transformer, wherein the attribute data at least includes a location and a type, write the attribute data of the abnormal device into a preset template, generate a maintenance task, and send the maintenance task to a preset terminal.

[0101] Figure 8 A component structure block diagram of a transformer operation fault detection system provided by an embodiment of the present application is shown, and the obtaining module 12 includes:

[0102] The calculation unit 121 is configured to calculate mathematical features of the operation parameters, wherein the mathematical features at least include a mean value, a variance, and a median value.

[0103] The sharing unit 122 is configured to refine the stable group into a plurality of subgroups according to the mathematical features corresponding to each transformer in the stable group, and construct a parameter sharing mechanism based on the subgroups.

[0104] Figure 9 The component structure block diagram of the transformer operation fault detection system provided by the embodiment of the present application is shown, and the correction module 13 comprises:

[0105] The cutting unit 131 is configured to determine the fault type corresponding to each abnormal data, and cut the training set into a plurality of subsets based on the fault type.

[0106] The insertion unit 132 is configured to insert the label generated by the fault type into the subset.

[0107] Figure 10 The component structure block diagram of the transformer operation fault detection system provided by the embodiment of the present application is shown, and the write module 14 comprises:

[0108] The selection unit 141 is configured to collect the computing resources in each preset range, and select the edge node.

[0109] The mounting unit 142 is configured to mount the global perception model into the edge node, and mount the local detection model into the preset cloud node.

[0110] The editing unit 143 is configured to edit the federal learning rule, and deploy the federal learning rule into the multi-level monitoring architecture.

[0111] The aggregation unit 144 is configured to establish the mapping between the preset range and the multi-level monitoring architecture, and perform federal aggregation.

[0112] The clustering module 11 is mainly configured to complete the step S100, the obtaining module 12 is mainly configured to complete the step S200, the correction module 13 is mainly configured to complete the step S300, and the write module 14 is mainly configured to complete the step S400.

[0113] The setting unit 111 is mainly configured to complete the step S101, and the collection unit 112 is mainly configured to complete the step S102.

[0114] The calculation unit 121 is mainly configured to complete the step S201, and the sharing unit 122 is mainly configured to complete the step S202.

[0115] The cutting unit 131 is mainly configured to complete the step S301, and the insertion unit 132 is mainly configured to complete the step S302.

[0116] The selecting unit 141 is mainly configured to complete the step S401, the mounting unit 142 is mainly configured to complete the step S402, the editing unit 143 is mainly configured to complete the step S403, and the aggregating unit 144 is mainly configured to complete the step S404.

[0117] Any combination of the technical features of the above-described embodiments can be made, and for the sake of brevity, all possible combinations of the technical features in the above-described embodiments are not described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the description.

[0118] The above-described embodiments only express several implementation manners of the present application, the description is more specific and detailed, but it should not be understood as the limitation of the patent scope of the present application. It should be pointed out that, for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

[0119] The above-described is only the preferred embodiment of the present application, and does not limit the present application, any modification, equivalent replacement and improvement within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A transformer operation fault detection method, characterized in that: The method comprises: Find transformers within a preset range, set several monitoring indicators, edit evaluation rules, compare monitoring indicators and evaluation rules, and cluster transformers into stable and risk groups; Creating a global model, collecting transformer operating parameters, obtaining historical data of the operating parameters of all transformers in the stable group, and training the global model to obtain a parameter group consisting of several sets of model parameters, wherein each transformer corresponds to a set of model parameters, and each model parameter includes at least: a weight coefficient and a threshold; Collect historical data of all transformers in the risk group, select abnormal data, build a training set, and train a pre-built conditional generation model. Extract the operating parameters of each transformer in the stable group under normal conditions to obtain a target parameter set, add random noise, input it into the conditional generation model, output simulated data, train the global model, and calibrate the parameter set. Based on the parameter group, the mean coefficient is calculated and deployed to the global model to obtain a global perception model. The model parameters are deployed to the global model to obtain a local detection model. The global perception model and the local detection model are integrated to generate a multi-level monitoring architecture, and the operating parameters are written in real time.

2. The transformer operation fault detection method according to claim 1, characterized in that: The steps of finding transformers within a preset range, setting a number of monitoring indicators, and editing evaluation rules include: Based on the evaluation rules, a fluctuation range is set for each monitoring indicator, and the operating parameters are compared with the fluctuation range to select abnormal equipment; The attribute data of each transformer is collected, wherein the attribute data includes at least: location and model, the attribute data of the abnormal device is written into a preset template, a maintenance task is generated, and sent to a preset terminal.

3. The transformer operation fault detection method according to claim 1, characterized in that: The steps of creating a global model, collecting operating parameters of transformers, obtaining historical data of operating parameters of all transformers in a stable group, and training the global model include: Calculating mathematical characteristics of the operating parameters, wherein the mathematical characteristics include at least: mean, variance, and median; Based on the mathematical characteristics corresponding to each transformer in the stable group, the stable group is refined into several subgroups, and a subgroup-based parameter sharing mechanism is constructed.

4. The transformer operation fault detection method according to claim 1, characterized in that: The steps of collecting historical data of all transformers in the risk group, selecting abnormal data, and constructing a training set include: Determine the fault type corresponding to each abnormal data, and divide the training set into several subsets based on the fault type; Insert labels generated by fault types into the subset.

5. The transformer operation fault detection method according to claim 1, characterized in that: The steps of calculating the mean coefficient based on the parameter group and deploying it to the global model to obtain the global perception model, and deploying the model parameters to the global model to obtain the local detection model include: Collect computing resources within each preset range and select edge nodes; The global perception model is mounted on the edge node, and the local detection model is mounted on the preset cloud node.

6. The transformer operation fault detection method according to claim 5, characterized in that: The steps of generating a multi-level monitoring architecture and writing the operating parameters in real time include: Edit federated learning rules and deploy them into the multi-level monitoring architecture; Establish a mapping between the preset scope and the multi-level monitoring architecture and perform federated aggregation; including: Within each preset range, a corresponding multi-level monitoring architecture is constructed, and the model parameters of the corresponding global perception model are updated. After the update, the model parameters of the global perception model within all preset ranges are calculated again to obtain new model parameters. Federated aggregation means: the model parameters distributed in different preset ranges and obtained through local training are merged according to certain rules to obtain a new global perception model.

7. A transformer operation fault detection system, characterized in that: The system comprises: The clustering module is used to find transformers within a preset range, set several monitoring indicators, edit evaluation rules, compare monitoring indicators and evaluation rules, and cluster transformers into stable groups and risk groups; A module is obtained, which is used to create a global model, collect operating parameters of transformers, obtain historical data of operating parameters of all transformers in the stable group, and train the global model to obtain a parameter group consisting of several sets of model parameters, wherein each transformer corresponds to a set of model parameters, and each model parameter includes at least: a weight coefficient and a threshold; A correction module is used to collect historical data of all transformers in the risk group, select abnormal data, build a training set, and train a pre-built conditional generation model. The operating parameters of each transformer in the stable group under normal conditions are extracted to obtain a target parameter set, and random noise is added. The target parameter set is input into the conditional generation model, and the simulated data is output. The global model is trained to correct the parameter set. A writing module is used to calculate the mean coefficient based on the parameter group and deploy it to the global model to obtain a global perception model, deploy the model parameters to the global model to obtain a local detection model, integrate the global perception model and the local detection model to generate a multi-level monitoring architecture, and write the operating parameters in real time.

8. The transformer operation fault detection system according to claim 7, characterized in that: The clustering module includes: A setting unit, configured to set a fluctuation range for each monitoring indicator according to the evaluation rules, compare the operating parameters with the fluctuation range, and select abnormal equipment; The collection unit is used to collect attribute data of each transformer, wherein the attribute data includes at least: location and model, write the attribute data of abnormal equipment into a preset template, generate a maintenance task, and send it to a preset terminal.

9. The transformer operation fault detection system according to claim 7, characterized in that: The obtaining module includes: A calculation unit, configured to calculate mathematical characteristics of the operating parameters, wherein the mathematical characteristics include at least a mean, a variance, and a median; The sharing unit is used to refine the stable group into several subgroups according to the mathematical characteristics corresponding to each transformer in the stable group, and to build a parameter sharing mechanism based on the subgroups.

10. The transformer operation fault detection system according to claim 7, characterized in that: The correction module includes: A segmentation unit, configured to determine the fault type corresponding to each abnormal data, and based on the fault type, segment the training set into a plurality of subsets; The insertion unit is used to insert the label generated by the fault type into the subset.

Images