Transformer fault diagnosis method and device based on large model

Abstract

The invention discloses a transformer fault diagnosis method and device based on a large model, relates to the technical field of electric power, and mainly aims to solve the problem of poor fault diagnosis accuracy of an existing transformer. Comprising the following steps: when an oil chromatography online monitoring state is determined to be normal operation, predicting online oil chromatography data based on a prediction model to obtain a prediction result of gas dissolved in oil; calling a transformer field knowledge base, transformer oil quality monitoring data, transformer operation condition data and off-line oil chromatography data; performing multi-modal coding on the transformer field knowledge base, the online oil chromatography data, the offline oil chromatography data, the prediction result of the gas dissolved in the oil, the transformer oil quality monitoring data and the transformer operation condition data, and processing the modal data obtained after coding based on the pre-trained large language model; obtaining a transformer monitoring result; and performing three-level evaluation on the transformer monitoring result based on confidence diagnosis to obtain a fault diagnosis result of the transformer.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric power, in particular to a transformer fault diagnosis method and device based on a large model. BACKGROUND

[0002] Oil-immersed power transformers are the most important power equipment in power grids, known as the industrial pulsating heart, and are super bridges connecting high-voltage power transmission and low-voltage power distribution. Accurate condition assessment and fault diagnosis are a solid guarantee for the safe and reliable operation of transformers, and therefore, the condition assessment of transformers is developing towards intelligence, relying on advanced sensors, big data analysis, and artificial intelligence and other frontier technological progress, to continuously improve the refinement, accuracy, and efficiency of fault diagnosis and condition assessment.

[0003] At present, existing oil-immersed power transformer fault diagnosis methods rely on various state monitoring parameters of transformers, including online, offline, live, running, and inspection data of various types, and therefore, learning models such as decision trees, fuzzy comprehensive diagnosis, neural networks, and image recognition can be constructed to make decision judgments. However, the aforementioned learning models are small parameter models, which are difficult to quickly update and adjust weights when facing new fault types or data features of transformers, and have problems such as gradient disappearance or gradient explosion in processing high-dimensional data. Moreover, some fault diagnosis rules are not integrated into the models, and there is a lack of reasonable optimization and updating, so small parameter models have been difficult to meet the increasingly complex field application requirements, and therefore, a transformer fault diagnosis method based on a large model is urgently needed to solve the aforementioned problems. SUMMARY

[0004] Therefore, the present application provides a transformer fault diagnosis method and device based on a large model, which mainly aims to solve the problem of poor accuracy of existing transformer fault diagnosis.

[0005] According to one aspect of the present application, a transformer fault diagnosis method based on a large model is provided, comprising: After determining that the online oil chromatography monitoring state is normal based on the online oil chromatography data of the transformer, the online oil chromatography data is predicted based on a trained prediction model to obtain a dissolved gas prediction result in oil, and the prediction model is obtained by incrementally learning and training a long short-term memory network based on the online oil chromatography data; The transformer domain knowledge base, transformer oil quality monitoring data, transformer operation data, and offline oil chromatography data are retrieved, the transformer domain knowledge base is constructed based on transformer domain specification knowledge, and the offline oil chromatography data is determined after capability evaluation based on experimental detection data; The transformer monitoring result is obtained by performing multi-modal coding on the transformer field knowledge base, the online oil chromatographic data, the offline oil chromatographic data, the dissolved gas prediction result in the oil, the transformer oil quality monitoring data and the transformer operation condition data based on a neural network model with introduced self-attention mechanism after model training is completed, and processing the modal data obtained after coding based on a large language model after pre-training. The transformer fault diagnosis result is obtained by performing three-level evaluation on the transformer monitoring result based on confidence.

[0006] According to another aspect of the present application, a transformer fault diagnosis device based on a large model is provided, comprising: The prediction module is configured to predict the online oil chromatographic data based on a trained prediction model to obtain a dissolved gas prediction result in the oil when determining that the online oil chromatographic monitoring state is normal based on the online oil chromatographic data of the transformer, wherein the prediction model is obtained by incrementally learning and training a long short-term memory network based on the online oil chromatographic data. The calling module is configured to call a transformer field knowledge base, transformer oil quality monitoring data, transformer operation condition data and offline oil chromatographic data, wherein the transformer field knowledge base is constructed based on transformer field specification knowledge, and the offline oil chromatographic data is determined after capability evaluation based on experimental detection data. The processing module is configured to perform multi-modal coding on the transformer field knowledge base, the online oil chromatographic data, the offline oil chromatographic data, the dissolved gas prediction result in the oil, the transformer oil quality monitoring data and the transformer operation condition data based on a neural network model with introduced self-attention mechanism after model training is completed, and process the modal data obtained after coding based on a large language model after pre-training to obtain a transformer monitoring result. The evaluation module is configured to perform three-level evaluation on the transformer monitoring result based on confidence to obtain a transformer fault diagnosis result.

[0007] According to still another aspect of the present application, a storage medium is provided, wherein at least one executable instruction is stored in the storage medium, and the executable instruction causes a processor to perform operations corresponding to the above-mentioned transformer fault diagnosis method based on a large model.

[0008] According to still another aspect of the present application, a terminal is provided, comprising a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface complete communication with each other through the communication bus. The memory is configured to store at least one executable instruction, and the executable instruction causes the processor to perform operations corresponding to the above-mentioned transformer fault diagnosis method based on a large model.

[0009] By means of the technical solutions described above, the technical solutions provided by the embodiments of the present application have at least the following advantages: The present application provides a transformer fault diagnosis method and device based on a large model. Compared with the prior art, the embodiments of the present application determine that the oil chromatogram online monitoring state is normal by means of the online oil chromatogram data of the transformer, predict the online oil chromatogram data based on the completed prediction model, obtain the dissolved gas prediction result in the oil, and the prediction model is obtained by incremental learning and training of a long short-term memory network based on the online oil chromatogram data; retrieve the transformer domain knowledge base, transformer oil quality monitoring data, transformer operation condition data, and offline oil chromatogram data, the transformer domain knowledge base is constructed based on the transformer domain specification knowledge, and the offline oil chromatogram data is determined after capability evaluation based on experimental detection data; the neural network model with the introduced self-attention mechanism based on the completed model training is used to perform multi-modal coding on the transformer domain knowledge base, the online oil chromatogram data, the offline oil chromatogram data, the dissolved gas prediction result in the oil, the transformer oil quality monitoring data, and the transformer operation condition data, and the modal data obtained after coding is processed based on the large language model that has completed pre-training to obtain a transformer monitoring result; the transformer monitoring result is evaluated in three levels based on the confidence to obtain a transformer fault diagnosis result, realize the intelligent analysis purpose based on the large language model, ensure the reliability of the transformer operation state analysis, avoid the diagnosis of transformer faults by single knowledge, improve the intelligent level of transformer operation and maintenance, and thus ensure the accuracy and reliability of the transformer state monitoring conclusion.

[0010] The above description is only a summary of the technical solutions of the present application. In order to enable a clearer understanding of the technical means of the present application, the content of the specification can be implemented, and in order to enable the above and other purposes, features and advantages of the present application to be more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0011] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are for purposes of illustration only and are not considered a limitation thereof. Moreover, like reference numerals designate identical components throughout the several views. In the drawings: Figure 1 A flowchart of a transformer fault diagnosis method based on a large model provided by the embodiments of the present application is shown; Figure 2 A large language model application architecture schematic diagram provided by the embodiments of the present application is shown; Figure 3Another flow chart of a transformer fault diagnosis method based on a large model is shown; Figure 4 An abnormal state detection flow chart is shown; Figure 5 A basic structure of an LSTM memory unit is shown; Figure 6 An incremental learning training flow chart is shown; Figure 7 Another incremental learning training flow chart is shown; Figure 8 A three-level evaluation flow chart is shown; Figure 9 A transformer fault diagnosis device based on a large model is shown; Figure 10 A terminal structure is shown. DETAILED DESCRIPTION

[0012] Exemplary embodiments of the present disclosure will be described more fully hereinafter with reference to the accompanying drawings; however, they are not limited to the embodiments set forth herein but can be implemented in various forms. The present disclosure will be described herein with reference to exemplary embodiments.

[0013] It should be noted that the terms "first", "second", and the like in the description and claims of the present application and the above-described drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or apparatus including a series of steps or units need not be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to such processes, methods, products or apparatus.

[0014] The embodiments of the present application can acquire and process related data based on artificial intelligence technology. Among them, artificial intelligence (AI) is to use digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0015] The basic technology of artificial intelligence generally includes technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics, etc. The software technology of artificial intelligence mainly includes computer vision technology, robot technology, biometric technology, speech processing technology, natural language processing technology, and machine learning / deep learning, etc.

[0016] Based on this, in one embodiment, the embodiments of the present application provide a transformer fault diagnosis method based on a large model. Taking the application of the method to a computer device such as a server as an example, the server can be a standalone server, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and basic cloud computing services such as big data and artificial intelligence platforms, etc.

[0017] The embodiments of the present application provide a transformer fault diagnosis method based on a large model, as shown in Figure 1 The method comprises: 101. After determining that the online oil chromatogram monitoring state is normal by the online oil chromatogram data of the transformer, the online oil chromatogram data is predicted based on the trained prediction model to obtain the dissolved gas prediction result in the oil.

[0018] In the embodiments of the present application, the current execution end as the execution subject of transformer fault diagnosis can be a cloud server, a terminal device, or a fault diagnosis subsystem embedded in a power system, so as to obtain related parameters of the transformer and diagnose the fault result. The online oil chromatogram data is an oil chromatogram detection sequence obtained by detecting the oil chromatogram in real time, so as to determine the online monitoring state of the oil chromatogram, that is, the running state of the online equipment of the transformer is monitored through the online oil chromatogram data. At this time, if the online monitoring state of the oil chromatogram is normal, it means that the online monitoring equipment of the oil chromatogram is stable, that is, the online oil chromatogram data is normal data, and the prediction model is used for prediction to obtain the prediction result of the dissolved gas in the oil. The prediction model is obtained by incremental learning and training of a long short-term memory network based on the online oil chromatogram data. The long short-term memory network (LSTM) is combined with online oil chromatogram data for incremental learning (IL) during model training, effectively keeping the original data set (the historical oil chromatogram data can include online oil chromatogram data and offline oil chromatogram data) under the premise of incremental learning IL, continuously acquiring new knowledge features from the online oil chromatogram data set, realizing dynamic fitting of the oil chromatogram data, and fully utilizing the real-time nature of online monitoring and the accuracy of offline monitoring.

[0019] 102, retrieve transformer domain knowledge base, transformer oil quality monitoring data, transformer operation data and offline oil chromatogram data.

[0020] In the embodiments of the present application, after obtaining the prediction result of the dissolved gas in the oil, the current execution end calls the transformer domain knowledge base, the transformer oil quality monitoring data, the transformer operation condition data and the offline oil chromatography data. The transformer domain knowledge base is constructed based on the specification knowledge in the transformer field, and is used to represent the knowledge information and specification information in the transformer field. The transformer domain knowledge base can be stored in the form of text and can also be stored in the form of data. The offline oil chromatography data is determined after the capability evaluation based on the experimental detection data. For example, the data obtained through the offline oil chromatography test platform is verified and evaluated with the laboratory data. The transformer oil quality monitoring data is the insulation oil test data other than the oil chromatography that needs to be detected regularly during the operation of the transformer, and can also reflect the operation state of the transformer. The test data is also needed for judging the transformer fault. For example, the test data includes but is not limited to breakdown voltage, dielectric loss factor, moisture, flash point, interfacial tension, acid value, gas content, furfural, copper ion content in oil, particle size in oil and the like. The transformer operation condition data is the data generated during the operation of the transformer. For example, the transformer operation condition data includes but is not limited to winding temperature, oil temperature, load, voltage fluctuation, infrared temperature measurement, partial discharge, core clamp ground current, winding deformation detection, transformer winding joint sleeve capacitance, transformer short-circuit impedance data, transformer winding frequency response characteristic curve and the like. The embodiments of the present application are not limited specifically.

[0021] It should be noted that the transformer oil quality monitoring data, the transformer operation condition data and the offline oil chromatography data can be input or collected through the power system. The transformer domain knowledge base can be constructed in advance. The transformer domain knowledge base not only contains the static description of the online equipment, but also contains the knowledge of the dynamic operation state of the transformer, has good logical consistency and knowledge expression ability, and forms a knowledge connection framework from the account to the fault cause. The embodiments of the present application are not limited specifically.

[0022] In some embodiments, the transformer domain knowledge base can include the following data: transformer equipment account, non-power-off inspection items, power-off inspection items, inspection measures, fault cases, and transformer-related standards, etc., which are not specifically limited in the embodiments of the present application. Among them, the transformer equipment account can include equipment model, voltage grade, capacity, voltage regulation mode, cooling mode, oil brand, oil weight, operation date, manufacturer, connection group, etc.; the non-power-off inspection items include temperature, oil level, leaked oil, abnormal sound and vibration, core grounding, DC bias, etc.; the power-off inspection items include the insulation resistance of the core and the clamp; the inspection measures include the inspection and measures of abnormal sound of the transformer body, the inspection and measures of insulation dampness abnormality, the inspection and measures of overheat abnormality, the inspection and measures of discharge abnormality, the inspection and measures of winding deformation, the inspection of tap changer, etc.; the fault cases include circuit overheat fault cases, magnetic circuit overheat fault cases, overheat fault cases caused by peripheral accessories, arc discharge fault cases, spark discharge fault cases, etc.; the transformer-related standards include JB 501-2021 "Guidelines for Testing of Power Transformers", DL / T 984-2018 "Guidelines for Judging Insulation Aging of Oil-immersed Transformers", DL / T 596-2021 "Procedures for Preventive Testing of Electrical Equipment", DL / T 572-2021 "Operation Procedures for Power Transformers", DL / T 573-2021 "Guidelines for Overhaul of Power Transformers", DL / T 393-2021 "Procedures for Condition-based Maintenance Testing of Power Transmission and Transformation Equipment", GB / T 14542-2017 "Guidelines for Maintenance and Management of Operating Transformer Oil", DL / T 722-2014 "Guidelines for Analysis and Judgment of Dissolved Gases in Transformer Oil", GB / T 7595-2017 "Quality of Transformer Oil in Operation", etc., which are not specifically limited in the embodiments of the present application.

[0023] In another embodiment of the present application, the current execution end can also obtain transformer oil quality monitoring data based on inter-laboratory comparison method, at this time, each capacity verification organizer will uniformly distribute the samples that need to be verified, i.e. transformer oil quality monitoring data, within the specified period, according to the requirements, simultaneously detect in each oil laboratory, and fill in the experimental detection data through the ".csv" template, upload to the current execution end, and each capacity verification organizer evaluates the transformer oil quality monitoring data. At this time, the general robust Z-score method can be used, The score as a capacity evaluation statistic quantity, the calculation formula is represented as: ; Among them, is the test result of the participating laboratory, is the robust mean value, is the robust standard deviation. The detection results of each laboratory are evaluated according to the following criteria: , indicating a satisfactory result; , indicating a questionable result (suspicious value) that needs to be found out and retested; , indicating an unsatisfactory result (outlier value) that may have a major problem and needs to be found out or corrected.

[0024] Wherein, the robust mean value is determined by the consensus value of each proficiency testing organizer, and the proficiency evaluation standard deviation is determined by the planned data of each proficiency testing organizer in the same round. The calculation process includes: arranging the collected detection data of a certain gas component in ascending order, represented as: , wherein the robust mean value and the robust standard deviation of the data are and , the initial values of and are calculated, and the calculation formula is represented as: ; ; According to the following steps, update and , and calculate: ; Then, for each , calculate: ; Recalculate the new value, and the formula is represented as: ; .

[0025] Finally, the robust estimate value and are obtained by iterative calculation, and according to the values after multiple updates, until the process converges. When the third significant digit of the robust standard deviation and the robust mean value no longer changes in two consecutive iterations, it is considered that the process is convergent. After confirming the convergence, the robust standard deviation and the robust mean value of a certain gas component are obtained. From this, the Z value of each component is calculated in turn, and the proficiency evaluation result of the detection data of the corresponding component is given according to the distribution range of the Z value. For example, as shown in the following Table 1, the main robust parameters are summarized, and as shown in Table 2, the off-line oil chromatography proficiency testing results of Alashan Power Supply Company in 2025 are shown.

[0026] Table 1

[0027] Table 2

[0028] 103. The transformer monitoring result is obtained by inputting the encoded modal data into the pre-trained large language model based on the transformer field knowledge base, the online oil chromatographic data, the offline oil chromatographic data, the dissolved gas prediction result in the oil, the transformer oil quality monitoring data, and the transformer operation condition data.

[0029] In the embodiments of the present application, in order to realize accurate fault detection of the transformer, the transformer field knowledge base, the online oil chromatographic data, the offline oil chromatographic data, the dissolved gas prediction result in the oil, the transformer oil quality monitoring data, and the transformer operation condition data are combined with the transformer operation state knowledge graph and the large language model for monitoring. At this time, since the transformer field knowledge base, the online oil chromatographic data, the offline oil chromatographic data, the dissolved gas prediction result in the oil, the transformer oil quality monitoring data, and the transformer operation condition data are multi-modal data, for example, including image modal, data modal, and text modal, before processing based on the large language model, multi-modal encoding can be performed, and the encoded modal data is input into the pre-trained large language model for processing to obtain the transformer monitoring result. The transformer state monitoring state model is constructed based on the large language model, for example, the large language model can be LLaMa-3, Qwen2.5, and when multi-modal data such as pictures are involved, MiniCPM-V 2.0, BLIP2, and other multi-modal large language models can be selected to realize extraction, fusion, and reasoning of multi-dimensional information. In the embodiments of the present application, the multi-modal large language model BLIP2 is preferred, without specific limitation.

[0030] In some embodiments, the large language model in the embodiments of the present application is constructed by a self-attention mechanism neural network and a large language model. When processing, the encoded modal data is first input into the self-attention mechanism neural network as input, and the modal data is encoded and decoded, that is, after inputting into the multi-layer self-attention mechanism neural network Transformer, the modal feature alignment is first performed, and the nonlinear feature operation is performed, that is, the multi-modal reconstruction decoder is input into the self-attention mechanism neural network Transformer to realize reconstruction of information of different modalities, and finally the output layer obtains the encoded modal data.

[0031] In some embodiments, for the processing process of the large language model, the current execution end can preform model pre-training and post-training, that is, the training of the large language model is performed in a hybrid stereo parallel training manner. At this time, the capability training of the basic large language model can include pure text data and multi-modal data, and the post-training purpose is to enhance the large model capability and preference alignment, including instruction fine-tuning, supervised learning, human feedback reinforcement learning, and human preference alignment. The supervised instruction fine-tuning enhancement is the ability to understand and follow human instructions in specific scenarios. Artificial feedback reinforcement learning is to explicitly introduce human feedback into the training process, align the model output to human values and preferences through human feedback, and improve the answer quality and reliability of the model.

[0032] 104、based on the confidence, the transformer monitoring result is evaluated in three levels to obtain a fault diagnosis result of the transformer.

[0033] In the embodiments of the present application, in order to improve the accurate diagnosis of the fault condition of the transformer, the current execution end calculates the confidence of the fault type corresponding to the transformer monitoring result as the basis for three-level evaluation. The first-level evaluation confidence is used to represent the basis for the initial evaluation of the fault type, the second-level evaluation confidence is used to divide the confidence of different fault types relative to the symptom state, and the third-level evaluation confidence is used to integrate all the second-level evaluation confidences to obtain the final fault diagnosis result.

[0034] In a specific implementation scenario, as shown in Figure 2 To achieve lightweight and edge computing deployment purposes, the current execution end can deploy a large language model LLM locally, and use knowledge distillation technology based on feature relationship preservation and model quantization technology based on gradient precision analysis to achieve lightweight processing of the large model. At this time, the state perception and operation and maintenance of a single transformer can be completed locally and efficiently by edge computing technology, and the state data processing of the transformer station layer can also be completed. The cloud edge collaborative technology is required, and each edge terminal uploads the data required by the cloud under the unified scheduling of the center server, and implements the perception, diagnosis and prediction of the overall state in the cloud. In the process of implementing the method based on multi-modal data in steps 101-104, as a processing system, it can be divided into system levels, that is, the data layer is used for multi-modal data management, such as data acquisition, generation, cleaning, validity judgment and other preprocessing processes, the model layer is used for the construction and learning training of the large language model, and the application layer is used for fine-tuning using business to adapt to different use scenarios, such as fault evaluation in the embodiments of the present application. It can also include intelligent question answering, intelligent prediction and other scenarios, which are not limited in the embodiments of the present application.

[0035] In another embodiment of the present application, in order to further limit and illustrate, as shown inFigure 3 As shown, step 101, based on the prediction model after completing the training, the online oil chromatographic data is predicted to obtain the prediction result of the dissolved gas in the oil, before the method further comprises: 201, obtaining the online oil chromatographic data; 202, respectively determining the relative gas production rate, the sliding window mean, the volatility coefficient of variation, the outlier state and the concentration distribution rule of the online oil chromatographic data; 203, sequentially performing abnormal state detection according to the relative gas production rate, the sliding window mean, the volatility coefficient of variation, the outlier state and the concentration distribution rule; 204, when there is no abnormal state, determining that the online monitoring state of the oil chromatogram is normal.

[0036] In order to realize the abnormal judgment of the online equipment monitoring data, the current execution end obtains the online oil chromatographic data through the online oil chromatographic monitoring platform in advance, which can be used as the input oil chromatographic detection sequence. Further, based on the relative data in the online oil chromatographic data, the gas production rate, the sliding window mean, the volatility coefficient of variation, the outlier state and the concentration distribution rule are calculated, which are used as the basis for abnormal state detection. For example, when at least one of the gas production rate, the sliding window mean, the volatility coefficient of variation, the outlier state and the concentration distribution rule meets the preset state detection condition, it is determined that the online equipment running state is abnormal, and the abnormal point is determined. If it does not meet the preset state detection condition, it can be determined that the running state is normal, and then the online oil chromatographic data is used for the next step of processing.

[0037] In another embodiment of the present application, in order to further limit and illustrate, step 203, respectively determining the relative gas production rate, the sliding window mean, the volatility coefficient of variation, the outlier state and the concentration distribution rule of the online oil chromatographic data, and sequentially performing abnormal state detection according to the relative gas production rate, the sliding window mean, the volatility coefficient of variation, the outlier state and the concentration distribution rule comprises: Based on the adjacent node detection concentration in the online oil chromatographic data, the relative gas production rate is calculated, and when the relative gas production rate is non-zero or any concentration value is not a preset negative value, it is determined that the data is valid; Based on the time series and mean square deviation in the online oil chromatographic data, the volatility coefficient of variation is calculated, and when the volatility coefficient of variation is less than a preset volatility threshold, it is determined that the data fluctuation is normal; Based on the carbon dioxide concentration, carbon monoxide concentration and characteristic gas concentration in the online oil chromatographic data, the concentration is sorted, and when the concentration distribution rule obtained by sorting does not match the preset concentration distribution rule, it is determined that the regularity is normal; a sliding window mean value is calculated based on the time sequence in the online oil chromatogram data, and it is determined that the window statistics are normal when the sliding window mean value does not match a preset mean value; outlier detection is performed based on the online oil chromatogram data, an outlier state is determined, and it is determined that the outliers are normal when the outlier state does not match a preset outlier condition.

[0038] In order to accurately monitor the running state and ensure the effectiveness of the online oil chromatogram data, the current execution end respectively determines the relative gas production rate, the sliding window mean value, the volatility coefficient of variation, the outlier state and the concentration distribution rule of the online oil chromatogram data.

[0039] In some embodiments, when determining the relative gas production rate, the current execution end calculates the relative gas production rate based on the adjacent node detection concentration in the online oil chromatogram data. Specifically, a relative gas production rate calculator can be set to monitor the relative gas production rate at adjacent time points. At this time, two cases can be divided. One is that there is a complete failure anomaly in the historical data of the oil chromatogram sensor, that is, the device loop is abnormal or the internal digital circuit does not calculate the result, at this time, any concentration value is not a preset negative value, such as-99999 in the oil chromatogram data, and it is determined that the data is effective. The other is contrary to the dynamic characteristics of the oil chromatogram time series data, which is manifested as no change in long-term gas concentration. Therefore, the relative gas production rate can be calculated based on the relative gas production rate calculation formula , which indicates that when the relative gas production rate is non-zero, the data is determined to be effective, for example, if is always 0, it means that the data is completely abnormal. Wherein, the online oil chromatogram data detection concentration at a certain time node is , the adjacent time node detection concentration is , and the relative gas production rate calculation formula is: .

[0040] In some embodiments, when determining the volatility coefficient of variation, the current execution end calculates the volatility coefficient of variation based on the time sequence and the mean square deviation in the online oil chromatogram data, so as to be compared with the preset concentration distribution rule. Specifically, the current execution end can set a volatility coefficient of variation calculator at this time. When the online oil chromatogram data has volatility anomaly, it means that the oil chromatogram sensor is unstable. Therefore, the volatility coefficient of variation is calculated, which represents the difference between the mean square deviation and the mean value of the random variable, and is a dimensionless statistical quantity. It can effectively quantify the volatility of the oil chromatogram characteristic gas concentration. The larger the coefficient of variation, the greater the deviation of the time series data, which means that the sensor working condition is more unstable, that is, the data reliability is lower, in other words, the data quality is lower. The volatility coefficient of variation calculation formula is: ; wherein, is a coefficient of variation, is a mean square error, is a time series data mean, , is a time series. In addition, the preset fluctuation threshold can be a maximum value representing data fluctuation anomaly, so as to determine that the data fluctuation is normal when the coefficient of variation of the fluctuation rate is less than the preset fluctuation threshold.

[0041] In some embodiments, when determining the concentration distribution rule, the current execution end can set a concentration distribution rule determination criterion, and then perform concentration sorting based on the carbon dioxide concentration, carbon monoxide concentration and characteristic gas concentration in the online oil chromatogram data, and determine that the regularity is normal when the concentration distribution rule obtained by sorting does not match the preset concentration distribution rule. Preferably, under the normal operation of the transformer, the characteristic gas concentration has a clear rule: c(CO2)>c(CO) c(other characteristic gas), wherein the C2H2 content is the lowest, at this time, if the online oil chromatogram data violates the common sense rule, the regularity is abnormal. In addition, the preset concentration distribution rule representing concentration distribution anomaly can be set based on actual common sense criterion, which is not limited in the embodiments of the present application.

[0042] In some embodiments, when determining the sliding window mean, the current execution end can set a sliding window mean calculator, so as to calculate the sliding window mean based on the time series in the online oil chromatogram data, and determine that the window statistics are normal when the sliding window mean does not match the preset mean (which can be configured based on time series requirements). Specifically, the current execution end can set the sliding window mean calculator, at this time, the window mean anomaly means that the statistical values before and after a certain time point are significantly different, and the abnormality is found by setting a sliding window with appropriate length, and the time series data in the window The sliding window mean calculation formula is represented as: .

[0043] In some embodiments, when determining the outlier state, the current execution end can configure an outlier detection algorithm to perform outlier detection based on the online oil chromatogram data, determine the outlier state, and determine that the outlier is normal when the outlier state does not match the preset outlier condition. Specifically, the outlier detection algorithm is preferably an isolation forest algorithm, which has the characteristics of fast speed and high accuracy, so that the isolation forest algorithm can be used to detect outliers in the time series data, and then the outliers are compared with the preset outlier condition, for example, the preset outlier condition is that the outlier exceeds the normal range of the data sequence, that is, the outlier is determined to be normal.

[0044] Correspondingly, in the embodiments of the present application, the step further comprises: The online oil chromatogram data is determined to be valid based on that the data is valid, the data fluctuation is normal, the regularity is normal, the window statistics is normal, the outlier is normal, and the window outlier secondary check is normal.

[0045] In order to realize accurate monitoring of the state of the online device, as shown in the prior art, Figure 4 As shown in the prior art, when the data is valid, the data fluctuation is normal, the regularity is normal, the window statistics is normal, the outlier is normal, and the window outlier secondary check is normal, it is indicated that the online oil chromatogram data is valid data, and therefore, it is determined that the oil chromatogram online monitoring state is running normally. In the embodiment in which it is determined that the oil chromatogram online monitoring state is running normally, the above-mentioned data valid, data fluctuation normal, regularity normal, window statistics normal, outlier normal, and window outlier secondary check normal can be set to be simultaneously satisfied, so as to determine that the oil chromatogram online monitoring state is running normally, or a preset number of satisfied conditions can be set, for example, as long as the data is valid, the data fluctuation is normal, and the regularity is normal, the window statistics is normal, it can be determined that the oil chromatogram online monitoring state is running normally, and the embodiment of the present application is not limited.

[0046] It should be noted that, as shown in the prior art, Figure 4 When the window statistics value is abnormal and the outlier is abnormal, a secondary check method can also be used to judge the validity of the data, for example, a manual check can be performed, or the preset mean value of the window can be adjusted or the preset outlier condition can be adjusted, and the embodiment of the present application is not limited.

[0047] In another embodiment of the present application, in order to further limit and illustrate, before the step of obtaining the dissolved gas prediction result in oil based on the prediction model trained to predict the online oil chromatogram data, the method further comprises: building a long short-term memory network; obtaining online oil chromatogram samples and offline oil chromatogram samples in different time periods, and using the online oil chromatogram samples and the offline oil chromatogram samples to train the long short-term memory network in an incremental learning manner to obtain a prediction model trained.

[0048] To improve the accuracy of dissolved gas prediction in oil, the current execution unit utilizes online oil chromatography data as incremental learning to explore the data relationship between online and offline oil chromatography monitoring results. The current execution unit first constructs a Long Short-Term Memory (LSTM) network. This LSM network consists of an input layer, an output layer, and multiple hidden layers composed of memory cells. Each memory cell contains a cell state and three key gating mechanisms: an input gate, a forget gate, and an output gate. The cell state is represented as an information conveyor belt, capable of transmitting information sequentially without interference. The gating mechanisms control the flow and updating of information, determining which information is remembered, forgotten, or output in the network.

[0049] In some embodiments, such as Figure 5 The basic structure of the LSTM memory unit shown is illustrated, where the forget gate determines the cell state at the previous time step. Which information should be retained or forgotten? The calculation formula is expressed as follows: ; in, The output of the forget gate, with a value range of (0, 1), is used to control the proportion of information retained or forgotten. This represents the sigmoid activation function; The weight matrix representing the forget gate; This indicates the hidden state of the previous time step; This represents the input at the current time step; This represents the bias matrix of the forget gate. The input gate determines the information of the current input. Which elements need to be introduced into the cell state? The calculation formulas are expressed as follows: ; ; ; in, The output of the input gate, with a value range of (0, 1), is used to control the proportion of new information written into the cell state. Represents candidate memories, used to update cell state; represents the hyperbolic tangent activation function, with values ​​(-1, 1); This represents the cell state updated at the current time step. The output gate determines which information from the cell state affects the output at the current time step and is passed to the next time step. The calculation formulas are expressed as follows: ; ; wherein, represents an output gate weight, controlling the proportion of cell state output; represents the output of the output gate, i.e. the current hidden state.

[0050] It should be noted that when the current execution end uses the online oil chromatogram sample and the offline oil chromatogram sample to train the long short-term memory network in an incremental learning manner to obtain a trained prediction model, the training data set is continuously predicted and the existing weights of the LSTM neural network are updated based on the time memory characteristics and nonlinear fitting capability of the LSTM. At the same time, the training and prediction processes are alternately performed for online prediction of real-time data, as shown in Figure 6 At this time, the incremental learning IL can continuously learn and optimize the prediction model in the continuous data stream, continuously supplement and update the existing knowledge, correct the existing knowledge after new data input, automatically extract knowledge from the new data set, and update the training parameters of the model in a timely manner. In the incremental learning mode, the expansion and memory utilization of the system are improved.

[0051] In one model training scenario in the embodiments of the present application, as shown in Figure 7 The input layer of the LSTM neural network is a two-layer fully connected neural network (Neutral Network, NN), which uses a neural layer with a linear activation function for prediction. The part of the data set containing offline oil chromatogram samples is used to train the learning model offline, as a pre-training process. Based on the incremental learning prediction stage, the oil chromatogram data is predicted in real time online, and the training and prediction processes are alternately performed based on the online oil chromatogram samples. In the embodiments of the present application, an online algorithm is used, i.e. the training and prediction processes are performed in real time to ensure the accuracy of the prediction. The training steps include: ①Using the historical offline oil chromatogram data of the 0~t-1 day as the model sample for training to obtain a pre-training model; ②Updating the training model by incremental learning IL on the t day, and collecting the online oil chromatogram data collected on the t day to construct a fixed-length training sample. The training sample includes the latest NIL days of data from the t-NIL+1 day to the t day. On this basis, the training sample is used to train and update the parameters of the pre-training model; ③Predicting the data of the t+1 day on the t day, inputting the online oil chromatogram data monitored during the period into the model of ②, and outputting the data of the t+1 day. The steps of ② and ③ are repeatedly executed to continuously alternate the training and prediction processes.

[0052] In another embodiment of the present application, in order to further limit and illustrate, the step of performing multi-modal encoding on the transformer domain knowledge base, the online oil chromatogram data, the offline oil chromatogram data, the oil dissolved gas prediction result, the transformer oil quality monitoring data, and the transformer operation condition data based on the self-attention mechanism neural network model after the model training is completed comprises: performing symbol and / or sensitive word filtering, deduplication, and image cropping processing on the transformer domain knowledge base, the online oil chromatogram data, the offline oil chromatogram data, the oil dissolved gas prediction result, the transformer oil quality monitoring data, and the transformer operation condition data; performing modal unified encoding on the processed transformer domain knowledge base, the online oil chromatogram data, the offline oil chromatogram data, the oil dissolved gas prediction result, the transformer oil quality monitoring data, and the transformer operation condition data to obtain modal data.

[0053] In order to ensure that the large language model can perform fault analysis on the multi-modal transformer domain knowledge base, online oil chromatogram data, offline oil chromatogram data, oil dissolved gas prediction result, transformer oil quality monitoring data, and transformer operation condition data, thereby improving the accuracy of fault diagnosis and the effectiveness of large model processing, the current execution end first performs symbol and / or sensitive word filtering, deduplication, and image cropping processing on the transformer domain knowledge base, online oil chromatogram data, offline oil chromatogram data, oil dissolved gas prediction result, transformer oil quality monitoring data, and transformer operation condition data. As a data cleaning step, the current execution end can include special symbol filtering, data deduplication, image cropping, and sensitive word removal for input data to the large model to ensure the effectiveness of data input. Further, for the cleaned transformer domain knowledge base, online oil chromatogram data, offline oil chromatogram data, oil dissolved gas prediction result, transformer oil quality monitoring data, and transformer operation condition data, when performing modal unified encoding, image encoding, electrical signal encoding, high-dimensional feature mapping, video key frame perception, etc. can be used to achieve multi-modal unified encoding.

[0054] In another embodiment of the present application, in order to further limit and illustrate, the step of performing three-level evaluation on the transformer monitoring result based on the confidence level to obtain the fault diagnosis result of the transformer comprises: determining the fault type of the transformer monitoring result, and calculating the first-level evaluation confidence of the transformer monitoring result based on the fault type; when the first-level evaluation confidence is less than a preset confidence threshold or the fault type is a preset fault type, calculating the second-level evaluation confidence based on the occurrence probability and symptom state of the fault type; The plurality of secondary evaluation confidences are weighted to obtain a tertiary diagnosis confidence, and a fault diagnosis result of the transformer is determined based on the tertiary diagnosis confidence.

[0055] To achieve accurate diagnosis of transformer faults, the current execution end performs three-level evaluation on transformer monitoring results based on confidence. Specifically, the fault type of the transformer monitoring result is first determined. At this time, the fault classification in the transformer domain knowledge base can be used for retrieval query, which is not limited in the embodiments of the present application. Then, the first-level evaluation confidence of the transformer monitoring result is calculated based on the fault type. The calculation formula of the first-level evaluation confidence is as follows: ; wherein, is a diagnosis query, represents the probability diagnosis result of the large model for the fault type . In the embodiments of the present application, when the first-level evaluation confidence is less than a preset confidence threshold (denoted as , is a preset confidence threshold), or the fault type is a preset fault type (denoted as , is a diagnosis subset), it indicates that the first-level evaluation cannot accurately diagnose the fault, and therefore, the second-level evaluation is entered, that is, the second-level evaluation confidence is calculated based on the occurrence probability of the fault type and the symptom state. The calculation formula is as follows: ; wherein, is the occurrence probability value of the fault type ; is a symptom state value set; is the conditional probability value of the symptom state value set to the fault type . Further, the plurality of secondary evaluation confidences are weighted to obtain a tertiary diagnosis confidence, and a fault diagnosis result of the transformer is determined based on the tertiary diagnosis confidence. The calculation formula of the tertiary diagnosis confidence is as follows: ; wherein n is the number of secondary evaluation confidences.

[0056] It should be noted that when the fault diagnosis result of the transformer is finally determined based on the tertiary diagnosis confidence, the tertiary diagnosis confidence can be guided to an intelligent robot that uses a large model for diagnosis, such as Figure 8As shown, in some embodiments, the fault type and the three-level diagnosis confidence are vectorized through the pre-trained model, the L2 distance is used to measure the correlation score between the query vector and the document information, and the final fault diagnosis result is determined from the transformer knowledge base. Among them, when determining the final fault diagnosis result by using the knowledge base, the correlation score index vector library can be used to retrieve the top k most similar text blocks in the retrieval result, and the retrieval result and the prompt text template are combined to input the case analysis intelligent agent, such as a large language model, to obtain a more concise and professional fault diagnosis result, fault cause analysis and operation and maintenance scheme.

[0057] The embodiment of the present application provides a transformer fault diagnosis method based on a large model. Compared with the prior art, the embodiment of the present application determines that the oil chromatogram online monitoring state is normal by using the online oil chromatogram data of the transformer, predicts the online oil chromatogram data based on the trained prediction model, obtains the dissolved gas prediction result in the oil, and the prediction model is obtained by incrementally learning and training a long short-term memory network based on the online oil chromatogram data; the transformer field knowledge base, transformer oil quality monitoring data, transformer operation data and offline oil chromatogram data are called, the transformer field knowledge base is constructed based on the transformer field specification knowledge, and the offline oil chromatogram data is determined after capability evaluation based on experimental detection data; the transformer field knowledge base, the online oil chromatogram data, the offline oil chromatogram data, the dissolved gas prediction result in the oil, the transformer oil quality monitoring data and the transformer operation data are encoded by a neural network model with a self-attention mechanism based on completed model training, and the modal data obtained after encoding is processed based on a large language model with completed pre-training, to obtain a transformer monitoring result; the transformer monitoring result is evaluated in three levels based on confidence, to obtain a transformer fault diagnosis result, realize the intelligent analysis purpose based on the large language model, ensure the reliability of the transformer operation state analysis, avoid the diagnosis of transformer faults by single knowledge, improve the intelligent level of transformer operation and maintenance, and thus ensure the accuracy and reliability of the transformer state monitoring conclusion.

[0058] Further, as to the above Figure 1 The embodiment of the present application provides a transformer fault diagnosis device based on a large model, as shown in the method. Figure 9 The device comprises: A prediction module 31 is configured to determine that the oil chromatogram online monitoring state is normal by using the online oil chromatogram data of the transformer, predict the online oil chromatogram data based on the trained prediction model, obtain the dissolved gas prediction result in the oil, and the prediction model is obtained by incrementally learning and training a long short-term memory network based on the online oil chromatogram data. The calling module 32 is configured to call a transformer field knowledge base, transformer oil quality monitoring data, transformer operation data, and offline oil chromatogram data. The transformer field knowledge base is constructed based on transformer field specification knowledge. The offline oil chromatogram data is determined based on experimental detection data after capability evaluation. The processing module 33 is configured to perform multi-modal encoding on the transformer field knowledge base, the online oil chromatogram data, the offline oil chromatogram data, the oil dissolved gas prediction result, the transformer oil quality monitoring data, and the transformer operation data based on a self-attention mechanism introduced neural network model that has completed model training, and perform processing on the modal data obtained after encoding based on a large language model that has completed pre-training, to obtain a transformer monitoring result. The evaluation module 34 is configured to perform three-level evaluation on the transformer monitoring result based on confidence to obtain a transformer fault diagnosis result.

[0059] Further, the device further comprises: The acquisition module is configured to acquire the online oil chromatogram data. The first determination module is configured to respectively determine a relative gas production rate, a sliding window mean, a volatility coefficient of variation, an outlier state, and a concentration distribution rule of the online oil chromatogram data, and perform abnormal state detection in sequence according to the relative gas production rate, the sliding window mean, the volatility coefficient of variation, the outlier state, and the concentration distribution rule. The second determination module is configured to determine that the oil chromatogram online monitoring state is normal when there is no abnormal state.

[0060] Further, The first determination module is specifically configured to calculate the relative gas production rate based on adjacent node detection concentrations in the online oil chromatogram data, and determine that the data is valid when the relative gas production rate is non-zero or any concentration value is not a preset negative extreme value. Calculate the volatility coefficient of variation based on time series and mean square deviation in the online oil chromatogram data, and determine that the data volatility is normal when the volatility coefficient of variation is less than a preset volatility threshold. Perform concentration sorting based on carbon dioxide concentration, carbon monoxide concentration, and characteristic gas concentration in the online oil chromatogram data, and determine that the regularity is normal when the concentration distribution rule obtained by sorting does not match a preset concentration distribution rule. Calculate the sliding window mean based on time series in the online oil chromatogram data, and determine that the window statistics are normal when the sliding window mean does not match a preset mean. Perform outlier detection based on the online oil chromatogram data to determine the outlier state, and determine that the outlier is normal when the outlier state does not match a preset outlier condition. The second determination module is specifically configured to determine, based on the online oil chromatogram data corresponding to the conditions of valid data, normal data fluctuation, normal regularity, normal window statistics, normal outlier, and normal window outlier secondary verification, that the oil chromatogram online monitoring state is normal operation.

[0061] Further, the device further comprises: The construction module is configured to construct a long short-term memory network, the long short-term memory network comprising an input layer, an output layer, and a plurality of hidden layers composed of memory cells; obtain online oil chromatogram samples and offline oil chromatogram samples of different time periods, and perform model training on the long short-term memory network in an incremental learning manner using the online oil chromatogram samples and the offline oil chromatogram samples to obtain a trained prediction model.

[0062] Further, The processing module is specifically configured to perform symbol and / or sensitive word filtering, deduplication, and image cropping processing on the transformer domain knowledge base, the online oil chromatogram data, the offline oil chromatogram data, the dissolved gas in oil prediction result, the transformer oil quality monitoring data, and the transformer operation condition data; and perform modal uniform coding on the processed transformer domain knowledge base, online oil chromatogram data, offline oil chromatogram data, dissolved gas in oil prediction result, transformer oil quality monitoring data, and transformer operation condition data to obtain modal data.

[0063] Further, the processing module is specifically configured to encode and decode the modal data through the self-attention mechanism neural network to obtain encoded modal data; and process the encoded modal data based on the large language model to obtain a transformer monitoring result.

[0064] Further, The evaluation module is specifically configured to determine a fault type of the transformer monitoring result, and calculate a first-level evaluation confidence of the transformer monitoring result based on the fault type; when the first-level evaluation confidence is less than a preset confidence threshold or the fault type is a preset fault type, calculate a second-level evaluation confidence based on an occurrence probability and a symptom state of the fault type; weight calculate a plurality of second-level evaluation confidences to obtain a third-level diagnosis confidence, and determine a fault diagnosis result of the transformer based on the third-level diagnosis confidence.

[0065] This application provides a transformer fault diagnosis device based on a large model. Compared with the prior art, this application provides a device that, after determining that the online oil chromatography monitoring status is normal based on the online oil chromatography data of the transformer, predicts the dissolved gas in the oil based on a trained prediction model. The prediction model is obtained by incrementally learning and training a long short-term memory network based on the online oil chromatography data. The device retrieves transformer domain knowledge base, transformer oil quality monitoring data, transformer operating data, and offline oil chromatography data. The transformer domain knowledge base is constructed based on transformer domain standard knowledge, and the offline oil chromatography data is determined after capability evaluation based on experimental detection data. The device is based on a trained model. The neural network model, employing a self-attention mechanism for training, performs multimodal encoding on the transformer domain knowledge base, online oil chromatography data, offline oil chromatography data, dissolved gas prediction results in the oil, transformer oil quality monitoring data, and transformer operating status data. Based on a pre-trained large language model, the encoded modal data is processed to obtain transformer monitoring results. These results are then evaluated in three levels based on confidence levels to obtain transformer fault diagnosis results. This achieves the goal of intelligent analysis based on a large language model, ensuring the reliability of transformer operating status analysis, avoiding diagnosing transformer faults using single knowledge sources, improving the level of intelligent substation operation and maintenance, and thus ensuring the accuracy and credibility of transformer condition monitoring conclusions.

[0066] According to one embodiment of this application, a storage medium is provided, the storage medium storing at least one executable instruction that can execute the transformer fault diagnosis method based on a large model in any of the above method embodiments.

[0067] Figure 10 The diagram shows a structural schematic of a terminal according to one embodiment of the present application. The specific embodiments of the present application do not limit the specific implementation of the terminal.

[0068] like Figure 10 As shown, the terminal may include: a processor 402, a communications interface 404, a memory 406, and a communications bus 408.

[0069] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408.

[0070] Communication interface 404 is used to communicate with other network elements such as clients or other servers.

[0071] The processor 402 is configured to execute the program 410, and specifically, can execute the related steps in the above-mentioned embodiments of the transformer fault diagnosis method based on a large model.

[0072] Specifically, the program 410 can include program codes including computer operation instructions.

[0073] The processor 402 can be a central processing unit (CPU) or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application. The one or more processors included in the terminal can be the same type of processors, such as one or more CPUs; or can be different types of processors, such as one or more CPUs and one or more ASICs.

[0074] The memory 406 is configured to store the program 410. The memory 406 can include a high-speed RAM memory, and can also include a non-volatile memory such as at least one disk memory.

[0075] The program 410 can be specifically used to cause the processor 402 to perform the following operations: After determining that the online oil chromatogram monitoring state of the transformer is normal based on the online oil chromatogram data of the transformer, the online oil chromatogram data is predicted based on the trained prediction model to obtain a dissolved gas prediction result in the oil, and the prediction model is obtained by incrementally learning and training a long short-term memory network based on the online oil chromatogram data; The transformer domain knowledge base, transformer oil quality monitoring data, transformer operation data, and offline oil chromatogram data are called, the transformer domain knowledge base is constructed based on transformer domain specification knowledge, and the offline oil chromatogram data is determined after capability evaluation based on experimental detection data; The transformer domain knowledge base, the online oil chromatogram data, the offline oil chromatogram data, the dissolved gas prediction result in the oil, the transformer oil quality monitoring data, and the transformer operation data are encoded based on the neural network model with an introduced self-attention mechanism after model training, and the modal data obtained after encoding is processed based on a large language model pre-trained to obtain a transformer monitoring result; The transformer monitoring result is evaluated in three levels based on confidence to obtain a transformer fault diagnosis result.

[0076] It should be apparent to those skilled in the art that the modules or steps of the application described above can be implemented with a general purpose computing device, which can be centralized on a single computing device or distributed on a network of multiple computing devices, and optionally implemented with program codes executable by a computing device, which can be stored in a storage device and executed by a computing device, and in some cases, the steps shown or described can be executed in an order different from that shown, or made into individual integrated circuit modules, or made into a single integrated circuit module. Thus, the present application is not limited to any particular hardware and software combination.

[0077] The preferred embodiments of the present application described above are intended to be illustrative only and the present application is not limited to the examples described above. Numerous modifications and variations are possible in light of the above teachings and within the scope of the present application, which is defined by the appended claims. Any and all modifications, variations or improvements not described above are considered to be within the scope of the present application.

Claims

1. A transformer fault diagnosis method based on a large model, characterized in that, include: Once the online oil chromatography data based on the transformer indicates that the online oil chromatography monitoring status is normal, the online oil chromatography data is used to predict the dissolved gas in the oil based on the trained prediction model. The prediction model is obtained by incrementally learning and training a long short-term memory network based on the online oil chromatography data. The system retrieves transformer knowledge base, transformer oil quality monitoring data, transformer operation data, and offline oil chromatography data. The transformer knowledge base is constructed based on standardized knowledge in the transformer field, and the offline oil chromatography data is determined after capability evaluation based on experimental test data. Based on a neural network model with a self-attention mechanism that has been trained, the transformer domain knowledge base, the online oil chromatography data, the offline oil chromatography data, the oil dissolved gas prediction results, the transformer oil quality monitoring data, and the transformer operation data are multimodally encoded. The encoded modal data are then processed based on a pre-trained large language model to obtain the transformer monitoring results. The transformer monitoring results are evaluated in three levels based on confidence level to obtain the transformer fault diagnosis results.

2. The method according to claim 1, characterized in that, Before obtaining the predicted dissolved gas in oil result by predicting the online oil chromatography data based on the trained prediction model, the method further includes: Obtain the online oil chromatography data; The relative gas production rate, sliding window mean, volatility coefficient of variation, outlier status, and concentration distribution pattern of the online oil chromatography data are determined respectively, and anomaly detection is performed sequentially according to the relative gas production rate, the sliding window mean, the volatility coefficient of variation, the outlier status, and the concentration distribution pattern. When there are no abnormal conditions, the online monitoring status of the oil chromatography is determined to be normal.

3. The method according to claim 2, characterized in that, The process of determining the relative gas production rate, sliding window mean, volatility coefficient of variation, outlier status, and concentration distribution pattern of the online oil chromatography data, and then performing anomaly detection sequentially based on the relative gas production rate, the sliding window mean, the volatility coefficient of variation, the outlier status, and the concentration distribution pattern, includes: The relative gas production rate is calculated based on the concentration detected at adjacent nodes in the online oil chromatography data. If the relative gas production rate is non-zero or any concentration value is not a preset negative extreme value, the data is determined to be valid. The volatility variation coefficient is calculated based on the time series and mean square error in the online oil chromatography data, and the data fluctuation is determined to be normal when the volatility variation coefficient is less than a preset volatility threshold. The concentrations of carbon dioxide, carbon monoxide, and characteristic gases in the online oil chromatography data are sorted, and if the sorted concentration distribution does not match the preset concentration distribution, the regularity is determined to be normal. The mean of the sliding window is calculated based on the time series data in the online oil chromatography data, and the window statistics are determined to be normal when the mean of the sliding window does not match the preset mean. Outlier detection is performed based on the online oil chromatography data to determine the outlier status, and if the outlier status does not match the preset outlier conditions, the outlier is determined to be normal.

4. The method according to claim 3, characterized in that, The method further includes: Based on the online oil chromatography data corresponding to valid data, normal data fluctuations, normal regularity, normal window statistics, normal outliers, and normal secondary verification of window outliers, the online oil chromatography monitoring status is determined to be normal.

5. The method according to claim 1, characterized in that, Before obtaining the predicted dissolved gas in oil result by predicting the online oil chromatography data based on the trained prediction model, the method further includes: Construct a Long Short-Term Memory (LSTM) network, which includes an input layer, an output layer, and multiple hidden layers composed of memory units; Online and offline oil chromatographic samples from different time periods are acquired, and the Long Short-Term Memory Network is trained using the online and offline oil chromatographic samples in an incremental learning manner to obtain a prediction model after training.

6. The method according to claim 1, characterized in that, The neural network model, based on a pre-trained model and incorporating a self-attention mechanism, performs multimodal encoding on the transformer domain knowledge base, the online oil chromatography data, the offline oil chromatography data, the dissolved gas prediction results in the oil, the transformer oil quality monitoring data, and the transformer operation data, including: The transformer knowledge base, the online oil chromatography data, the offline oil chromatography data, the dissolved gas prediction results in the oil, the transformer oil quality monitoring data, and the transformer operation data are processed by symbol and / or sensitive word filtering, deduplication, and image cropping. Modal unified encoding is performed on the processed transformer domain knowledge base, the online oil chromatography data, the offline oil chromatography data, the oil dissolved gas prediction results, the transformer oil quality monitoring data, and the transformer operation data to obtain modal data.

7. The method according to any one of claims 1-5, characterized in that, The three-level evaluation of the transformer monitoring results based on confidence level yields the following transformer fault diagnosis results: Determine the fault type of the transformer monitoring results, and calculate the first-level evaluation confidence level of the transformer monitoring results based on the fault type; When the confidence level of the first-level evaluation is less than the preset confidence threshold or the fault type is a preset fault type, the confidence level of the second-level evaluation is calculated based on the occurrence probability and symptom status of the fault type. The confidence scores of the multiple secondary evaluations are weighted to obtain the confidence score of the tertiary diagnosis, and the fault diagnosis result of the transformer is determined based on the confidence score of the tertiary diagnosis.

8. A transformer fault diagnosis device based on a large model, characterized in that, include: The prediction module is used to predict the dissolved gas in the oil based on the online oil chromatography data after the online oil chromatography monitoring status is determined to be normal based on the online oil chromatography data based on the transformer. The prediction model is obtained by incremental learning training of the long short-term memory network based on the online oil chromatography data. The retrieval module is used to retrieve transformer domain knowledge base, transformer oil quality monitoring data, transformer operation data, and offline oil chromatography data. The transformer domain knowledge base is constructed based on transformer domain standard knowledge, and the offline oil chromatography data is determined after capability evaluation based on experimental test data. The processing module is used to perform multimodal encoding on the transformer domain knowledge base, the online oil chromatography data, the offline oil chromatography data, the oil dissolved gas prediction results, the transformer oil quality monitoring data, and the transformer operation data based on a neural network model with a self-attention mechanism that has been trained. The module then processes the encoded modal data based on a large language model that has been pre-trained to obtain the transformer monitoring results. The evaluation module is used to perform a three-level evaluation of the transformer monitoring results based on confidence level to obtain the transformer fault diagnosis results.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

Images