Method for establishing gas model on oil surface under sudden serious fault of oil immersed transformer

By installing gas chromatography-mass spectrometry combined equipment on the top of the oil tank of the oil-immersed transformer and combining multi-layer neural network models to monitor and analyze the oil surface gas composition in real time, the dynamic trend of gas concentration of oil surface of oil-immersed transformer in the prior art was solved. The problems of inaccurate and poor sensitivity to trace gas changes have achieved high-precision real-time monitoring and accurate diagnosis of sudden failures.

CN119044343APending Publication Date: 2024-11-29ELECTRIC POWER RESEARCH INSTITUTE OF STATE GRID NINGXIA ELECTRIC POWER COMPANY
View PDF 0 Cites 5 Cited by

Patent Information

Application Number
CN202411039961.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-31
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The prior art is difficult to accurately capture the dynamic trend of gas concentration in oil-infiltrated transformer oil surface, especially in the case of sudden and serious failure, which has poor sensitivity to trace gas changes, resulting in inaccurate fault diagnosis.

Method used

Using a combination of high-frequency sampling and multi-layer neural network model, a dynamic gas composition model is established, including high-frequency feature subnetwork and low-frequency features, by installing gas chromatography-mass spectrometry equipment on the top of the fuel tank. Subnetwork, main gas characteristic subnetwork, trace gas characteristic subnetwork, time characteristic subnetwork and spatial characteristic subnetwork, data preprocessing and time series analysis are carried out to predict gas change trends.

Benefits of technology

High-precision real-time monitoring of sudden serious failures of oil-immersed transformers is realized, multi-scale feature extraction capability for gas concentration changes is improved, sensitivity to trace gas changes is enhanced, and the accuracy and response speed of fault diagnosis is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119044343A_ABST
    Figure CN119044343A_ABST
Patent Text Reader

Abstract

The invention provides a method for establishing a gas model on an oil surface under the sudden serious fault of an oil immersed transformer, and belongs to the technical field of oil immersed transformer faults, and the method comprises the following steps: firstly, installing high-precision gas chromatography-mass spectrometry equipment at the top of an oil tank of the transformer, and setting a sampling interval for sampling once every 5-10 minutes; reference data for at least 24 hours is collected in a normal operating state. Once a sudden serious fault is detected, a high-frequency sampling mode is started immediately, and sampling is performed once every 15 seconds to 1 minute. Collected gas component data are subjected to preprocessing such as denoising, standardization and abnormal value detection and then modeled into a time sequence, and main gas components and change trends are analyzed. On the basis, a dynamic gas composition prediction model is established, the model is continuously optimized by utilizing a verification set, and finally the model is applied to real-time data flow and continuous updating optimization, so that the technical problems that the dynamic change trend of gas concentration is difficult to capture and the sensitivity to trace gas change is poor in the prior art are solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of oil-immersed transformer faults, and in particular, relates to a method for establishing a gas model on an oil surface of an oil-immersed transformer under sudden severe faults. Background Art

[0002] Oil-immersed transformers are key equipment in power systems, and their safe and stable operation is directly related to the reliability of the entire power grid. During transformer operation, faults such as insulation aging, local overheating, or arc discharge can cause abnormal changes in the dissolved gas in the oil. Therefore, real-time monitoring and analysis of dissolved gas in transformer oil has become an important means of fault diagnosis.

[0003] Traditional methods for analyzing dissolved gases in oil mainly include regular offline sampling and analysis and online monitoring. Offline sampling and analysis usually uses gas chromatography, and sampling and testing are performed every few months. Although this method has high accuracy, the sampling frequency is low and sudden failures cannot be discovered in time. Online monitoring technology has been widely used in recent years. Common methods include optical sensor method, electrochemical sensor method and gas permeation membrane method. These methods can achieve continuous monitoring, but they still have limitations in monitoring accuracy, gas type coverage and anti-interference ability.

[0004] Existing gas analysis models are mainly based on empirical thresholds or simple statistical methods, such as the three-ratio method and the Duhan ratio method. These methods have the following problems in practical applications: 1) It is difficult to accurately capture the dynamic trend of gas concentration, especially in the case of sudden serious faults; 2) Insufficient consideration of the complex interactions between multiple gases, resulting in limited diagnostic accuracy; 3) Lack of sensitivity to trace gas changes may miss early fault signs; 4) It is difficult to adapt to the differences between different types of transformers and operating environments, and the generalization ability is insufficient.

[0005] In summary, the existing technology has technical problems such as difficulty in capturing the dynamic change trend of gas concentration and poor sensitivity to changes in trace gases, resulting in inaccurate gas model on the oil surface under sudden serious faults of oil-immersed transformers. Summary of the invention

[0006] In view of this, the present invention provides a method for establishing a gas model on the oil surface under sudden severe faults of an oil-immersed transformer, which can solve the technical problems of the prior art that it is difficult to capture the dynamic change trend of gas concentration and has poor sensitivity to changes in trace gases.

[0007] The present invention is achieved in that:

[0008] A first aspect of the present invention provides a method for establishing a gas model on the oil surface of an oil-immersed transformer under a sudden severe fault, which comprises the following steps:

[0009] S10. Install a gas chromatography-mass spectrometry device on the top of the oil tank of the oil-immersed transformer to monitor and collect the gas composition on the oil surface in real time;

[0010] S20, according to the fault development progress of the transformer, set the sampling interval of the acquisition to be once every 5 to 10 minutes under normal operation;

[0011] S30. Under normal operating conditions of the transformer, collect gas composition data for at least 24 hours;

[0012] S40, when a sudden serious fault of the transformer is detected, immediately start the high-frequency sampling mode to collect the gas composition, and shorten the sampling interval to once every 15 seconds to 1 minute;

[0013] S50, preprocessing the collected data of the gas composition, including denoising, standardization and outlier detection;

[0014] S60, performing time series analysis on the changes of the pre-processed data of various gas components over time, identifying the main gas components and their change trends, and obtaining a gas time series matrix;

[0015] S70, based on the gas time series matrix and the moment of sudden severe fault, establish a training data set, the input of the training data set is the gas time series matrix, the output is the moment of occurrence of sudden severe fault, the severe fault category and the occurrence area, establish and train a dynamic gas composition model, and predict the change of various gas components over time;

[0016] S80, using part of the collected gas composition data as a verification set to verify and optimize the dynamic gas composition model;

[0017] S90, applying the optimized dynamic gas composition model to the real-time data stream of the oil-immersed transformer collected on site, continuously updating the monitored gas composition and prediction results, and setting a threshold for triggering an alarm mechanism according to the prediction results, so as to timely detect and predict abnormal gas changes.

[0018] Among them, the dynamic gas composition model adopts a long memory network structure, including a high-frequency characteristic sub-network, a low-frequency characteristic sub-network, a main gas characteristic sub-network, a trace gas characteristic sub-network, a time characteristic sub-network, a space characteristic sub-network and a fusion characteristic sub-network; the gas components include the main gas and trace gas.

[0019] Among them, the high-frequency feature sub-network is used to extract the characteristics of rapid changes in gas concentration in a short period of time. The structure of the high-frequency feature sub-network is composed of multiple layers of LSTM units, and the input is the data of gas composition collected in the high-frequency sampling mode in step S40.

[0020] Among them, the low-frequency feature sub-network is used to capture the long-term change trend of gas concentration. The structure of the low-frequency feature sub-network is composed of multiple layers of GRU units, and the input is the data of gas composition collected in the low-frequency sampling mode of the sampling interval sampling in step S20.

[0021] The main gas characteristic subnetwork is used to analyze the concentration change characteristics of the main gases, which include hydrogen, methane and ethylene. The structure of the main gas characteristic subnetwork is composed of a fully connected layer and an LSTM layer.

[0022] The trace gas characteristic subnetwork is used to analyze the concentration change characteristics of the trace gas, the trace gas includes: acetylene and carbon monoxide, and the structure of the trace gas characteristic subnetwork is composed of an attention mechanism and an LSTM layer.

[0023] Furthermore, the temporal feature subnetwork is used to extract the temporal pattern of gas concentration changes, and its structure is composed of a temporal convolutional network.

[0024] Furthermore, the spatial feature subnetwork is used to analyze the relationship between different gases, and its structure is composed of a graph neural network.

[0025] Furthermore, the fusion feature sub-network is used to integrate the output features of each sub-network, and its structure is composed of a multi-head self-attention mechanism and a fully connected layer.

[0026] Furthermore, the input and output relationship of each sub-network is that each sub-network receives the corresponding part of the gas timing matrix as input, and the output feature vector is passed to the fusion feature sub-network for integration, and finally the prediction results of the fault occurrence time, serious fault category and occurrence area are output.

[0027] Wherein, the step S10 specifically includes:

[0028] Step 101, selecting a gas chromatography-mass spectrometry device with high sensitivity, high resolution and fast response capability to ensure that the device can meet the detection requirements of common gas components in transformer oil;

[0029] Step 102, determining an installation position near the gas sampling port on the top of the fuel tank to ensure that the most representative gas sample is collected;

[0030] Step 103, installing and debugging the equipment, including setting up gas line connection, power supply and data transmission interface;

[0031] Step 104: calibrate and verify the performance of the installed gas chromatography-mass spectrometry equipment to ensure the accuracy and reliability of the measurement results;

[0032] Step 105, setting the automatic sampling program of the device to ensure continuous sampling at predetermined time intervals;

[0033] Step 106: Establish a data transmission and storage system to achieve real-time transmission and long-term storage of collected data;

[0034] Step 107: Develop equipment maintenance and regular calibration plans to ensure long-term monitoring stability and data quality.

[0035] Wherein, the step S20 specifically includes:

[0036] Step 201, analyzing the development speed characteristics of different types of transformer faults, including insulation aging, local overheating, arc discharge and other fault types;

[0037] Step 202: Estimate the typical time span from the initial stage to the serious stage of each type of failure based on historical data and experience;

[0038] Step 203, considering the kinetics of gas generation and diffusion, and evaluating the time constant of gas concentration change;

[0039] Step 204, weigh the relationship between sampling frequency and data storage and processing capabilities, and select a suitable sampling interval;

[0040] Step 205: For a medium-sized oil-immersed transformer, the sampling interval is set to 7 minutes;

[0041] Step 206: Setting a sampling program on the gas chromatography-mass spectrometry device to ensure automatic sampling at set time intervals;

[0042] Step 207: Establish a dynamic adjustment mechanism for the sampling interval to automatically adjust the sampling frequency according to the transformer operating status.

[0043] Wherein, the step S30 specifically includes:

[0044] Step 301, ensuring that the transformer is in a stable normal operating state, including that parameters such as load, temperature and voltage are within normal ranges;

[0045] Step 302, starting the continuous sampling mode of the gas chromatography-mass spectrometry device, and collecting data according to the sampling interval set in step S20;

[0046] Step 303, record the collected data, including the concentration values ​​of various gases and sampling timestamps;

[0047] Step 304: Perform a preliminary analysis on the collected data and calculate statistical indicators such as the mean, standard deviation and coefficient of variation of each gas concentration;

[0048] Step 305: preliminarily determine the fluctuation range of the gas composition under the normal state of the transformer through statistical indicators;

[0049] Step 306: Store the benchmark data in a database as a reference for subsequent anomaly detection and model training;

[0050] Step 307: Establish a regular update mechanism for the reference value data to ensure that the reference value can reflect the long-term change trend of the transformer.

[0051] Wherein, the step S40 specifically includes:

[0052] Step 401, setting fault trigger conditions based on multiple indicators such as sudden change in gas concentration, abnormal gas generation rate, gas ratio analysis and abnormal electrical parameters;

[0053] Step 402, develop a real-time monitoring program to continuously compare the deviation of current gas data with normal baseline values;

[0054] Step 403: When the deviation exceeds a preset threshold, the high-frequency sampling mode is automatically triggered;

[0055] Step 404: In the high frequency sampling mode, dynamically adjust the sampling interval according to the severity of the fault type;

[0056] Step 405: For a suspected arc discharge fault, the sampling interval is set to 15 seconds; for a suspected local overheating fault, the sampling interval is set to 45 seconds;

[0057] Step 406, ensuring that the data acquisition system can handle the large amount of data generated by high-frequency sampling, including data storage, real-time transmission and preliminary processing;

[0058] Step 407: Establish an automatic termination mechanism for the high-frequency sampling mode, and restore the normal sampling frequency when the fault state is stabilized or eliminated.

[0059] Wherein, the step S50 specifically includes:

[0060] Step 501, remove noise from the original data, and use the wavelet transform method to perform multi-scale decomposition on the gas concentration time series;

[0061] Step 502: Select a suitable threshold to perform soft threshold processing, and then reconstruct the signal to obtain denoised data;

[0062] Step 503: Perform data standardization processing, and perform z-score standardization on the concentration data of each gas;

[0063] Step 504: perform outlier detection, using an improved local anomaly factor algorithm to detect abnormal points in the multidimensional gas data;

[0064] Step 505: introduce the concept of time window and only consider the data of the latest N time points to adapt to the dynamic change characteristics of gas concentration;

[0065] Step 506: Process the detected abnormal points, and select to delete or replace them using an interpolation method;

[0066] Step 507: Establish a data quality assessment mechanism to comprehensively score the pre-processed data to ensure the data quality for subsequent analysis.

[0067] Wherein, the step S60 specifically includes:

[0068] Step 601: Perform trend analysis on the time series data of each gas, and use the seasonal trend decomposition method to decompose the time series into three components: trend, seasonality and residual;

[0069] Step 602: Perform principal component analysis to identify the gas component that contributes most to the transformer state change;

[0070] Step 603: extract time series features of main gas components and calculate statistical features within the sliding time window;

[0071] Step 604, calculating the rate of change and acceleration of gas concentration to capture short-term change characteristics;

[0072] Step 605: Perform gas ratio analysis to calculate the ratio of key gas pairs;

[0073] Step 606: Integrate the analysis results into a gas time series matrix, where each row of the matrix represents a time point, and the columns include the original concentration, trend value, principal component score, statistical characteristics, change rate, acceleration, and gas ratio of each gas;

[0074] Step 607: Establish a dynamic update mechanism for the gas timing matrix to ensure that the matrix can reflect the latest changes in gas composition in real time.

[0075] Wherein, the step S70 specifically includes:

[0076] Step 701: construct a training data set, divide the gas time series matrix into several time windows in chronological order, and set the length of each window to 2 hours with a step length of 10 minutes;

[0077] Step 702: Mark each time window to see whether it contains a sudden serious fault moment. If so, further mark the fault type and occurrence area.

[0078] Step 703, designing a network structure of a dynamic gas composition model, including a high-frequency characteristic sub-network, a low-frequency characteristic sub-network, a main gas characteristic sub-network, a trace gas characteristic sub-network, a time characteristic sub-network, a space characteristic sub-network, and a fusion characteristic sub-network;

[0079] Step 704: define a loss function and an optimization strategy, use mean square error loss for prediction of fault occurrence time, and use cross entropy loss for prediction of fault category and occurrence area;

[0080] Step 705: Perform model training, using 80% of the data as a training set and 20% as a validation set, using a small batch gradient descent method, and setting the batch size to 64;

[0081] Step 706: Use the early stopping strategy to stop training when the loss on the validation set does not improve for 10 consecutive epochs.

[0082] Step 707: Test and evaluate the trained model, using the remaining data as a test set to evaluate the performance of the model on unseen data.

[0083] Wherein, the step S80 specifically includes:

[0084] Step 801: prepare a verification data set, randomly select 20% of the collected data as the verification set, and ensure that the verification set contains different types of fault cases and normal operation data;

[0085] Step 802: Perform model validation, input validation set data into the trained model, and obtain prediction results;

[0086] Step 803: Calculate various performance indicators on the validation set, including the root mean square error and mean absolute percentage error of the prediction of the fault occurrence time, the accuracy, precision, recall and F1 score of the prediction of the fault category and occurrence area;

[0087] Step 804: Analyze the model performance, compare the validation set performance with the training set performance, and evaluate whether the model has overfitting or underfitting problems;

[0088] Step 805: Optimize the model according to the verification results, including adding regularization terms, adjusting the network structure, performing feature engineering, using ensemble learning technology, and applying transfer learning;

[0089] Step 806: Perform cross validation, using a 5-fold cross validation method, divide the data set into 5 parts, and use 4 of them as training sets and 1 as validation set in turn;

[0090] Step 807: Based on the cross-validation results, the model is fine-tuned, and the optimal hyperparameter combination is automatically searched using methods such as grid search or Bayesian optimization.

[0091] Wherein, the step S90 specifically includes:

[0092] Step 901: deploy the optimized model to a real-time data processing server to ensure that the model can efficiently process the input gas time series matrix;

[0093] Step 902: Develop a data stream processing module to receive data from the gas chromatography-mass spectrometry device in real time and convert it into the input format required by the model;

[0094] Step 903: Use sliding window technology to maintain a data window with a fixed length of 2 hours, and update the window content every time new data arrives;

[0095] Step 904: Whenever the data window is updated, the latest gas time series matrix is ​​input into the model to obtain the prediction result;

[0096] Step 905: Based on the model prediction results, set a multi-level alarm threshold, including three levels: slight abnormality, moderate abnormality and severe abnormality;

[0097] Step 906: Implement an adaptive threshold mechanism and dynamically update the normal range of each gas using an exponentially weighted moving average method;

[0098] Step 907: Develop a visualization interface to display the concentration, change trend, prediction results and alarm status of various gases in real time.

[0099] Optionally, the high-frequency feature subnetwork is composed of 3 layers of long short-term memory network units, each layer has 64 neurons, which are used to extract the characteristics of rapid changes in gas concentration in a short period of time; the low-frequency feature subnetwork is composed of 2 layers of gated recurrent units, each layer has 32 neurons, which are used to capture the long-term change trend of gas concentration; the main gas feature subnetwork contains a fully connected layer and a long short-term memory network layer, the fully connected layer has 64 neurons, and the long short-term memory network layer has 32 neurons, which are used to analyze the concentration change characteristics of the main gases.

[0100] Optionally, the trace gas characteristic subnetwork includes a multi-head attention layer and a long short-term memory network layer. The multi-head attention layer has 4 heads, each head has 16 dimensions, and the long short-term memory network layer has 32 neurons, which are used to analyze the concentration change characteristics of trace gases; the time feature subnetwork adopts a temporal convolutional network, including 3 dilated convolutional layers, the convolution kernel size is 3, and the dilation rates are 1, 2, and 4 respectively, which are used to extract the temporal pattern of gas concentration changes; the spatial feature subnetwork adopts a graph attention network structure, including 2 graph attention layers, each with 16 attention heads, which are used to analyze the relationship between different gases.

[0101] Optionally, the fusion feature subnetwork includes a multi-head self-attention mechanism and two fully connected layers. The multi-head self-attention mechanism has 8 heads, each head is 32 dimensions, and the two fully connected layers have 128 and 64 neurons respectively, which are used to integrate the output features of each subnetwork; the optimizer of the dynamic gas composition model adopts the Adam algorithm, and the initial value of the learning rate is set to 0.001, which decays by 10% every 50 epochs; the early stopping strategy is used during the model training process, and the training is stopped when the loss on the validation set has not improved for 10 consecutive epochs.

[0102] Optionally, the wavelet transform method in step S50 is specifically: using the db4 wavelet basis, selecting 4-5 layers as the decomposition layer, and adaptively selecting the threshold using the VisuShrink method; the formula for z-score standardization is: the standardized value is equal to the original value minus the mean divided by the standard deviation; the improved local anomaly factor algorithm introduces the concept of time window, and only considers the data of the most recent N time points, where N ranges from 100 to 200.

[0103] Optionally, the seasonal trend decomposition method in step S60 can handle nonlinear trends and multiple periodicities; the principal component analysis selects the principal component with a cumulative contribution rate of 85% as the main gas component; the size of the sliding time window is set to 30 minutes, and it slides every 5 minutes; the key gas ratios include the ratio of acetylene to ethylene, the ratio of methane to hydrogen, and the ratio of ethylene to ethane.

[0104] Optionally, the fault categories in step S70 include insulation aging, local overheating, and arc discharge; the occurrence area includes coils, cores, and bushings; the judgment condition of the early stopping strategy is that the loss on the validation set has not improved for 10 consecutive epochs; the model performance evaluation indicators include the mean absolute error of the prediction of the fault occurrence time, the accuracy of the prediction of the fault category and occurrence area, and the F1 score.

[0105] Optionally, the model optimization method in step S80 includes: adding an L2 regularization term, and gradually adjusting the regularization coefficient starting from 0.001; using dropout technology, and setting the dropout rate between 0.2 and 0.5; using feature importance analysis methods, such as permutation importance, retaining important features, and removing or combining secondary features; using model integration technology, such as random forest or gradient boosting tree, combined with a deep learning model to form a hybrid model.

[0106] Optionally, the multi-level alarm thresholds in step S90 are set as follows: a minor abnormality is when the predicted concentration of a certain gas exceeds 1.5 times the normal range in the next 30 minutes; a moderate abnormality is when the predicted concentration exceeds 2 times the normal range, or minor abnormalities occur in multiple gases at the same time; a serious abnormality is when the predicted concentration exceeds 3 times the normal range, or the model predicts that there is a high probability of a serious failure; the normal range is determined based on statistical analysis of historical data using the 3σ principle.

[0107] Optionally, the adaptive threshold mechanism in step S90 adopts an exponentially weighted moving average method, and the weight factor is set to 0.1; the visualization interface uses different colors and icons to represent different levels of abnormalities; the method also includes an alarm push function, which sends alarm information to relevant personnel through multiple channels such as SMS, email, and mobile application push; the alarm information contains key information such as abnormal gas type, concentration value, predicted trend, possible fault type and location.

[0108] Optionally, the method also includes establishing an alarm log and feedback mechanism to record detailed information of each alarm, including trigger time, alarm level, related data, etc.; designing a feedback interface to allow operators to confirm, classify and annotate the alarm; the method also includes a regular evaluation and update mechanism to conduct a comprehensive evaluation and update of the model every 3 months based on actual operating data and feedback information to ensure continuous optimization and adaptability of model performance.

[0109] Compared with the prior art, the method for establishing a gas model on the oil surface of an oil-immersed transformer under sudden severe faults provided by the present invention has the following beneficial effects:

[0110] 1. High-precision real-time monitoring: By installing a high-precision gas chromatography-mass spectrometry device on the top of the oil tank and combining it with an adaptive sampling strategy, high-frequency and high-precision monitoring of the gas composition on the oil surface is achieved. Compared with traditional methods, the present invention can shorten the sampling interval to 15 seconds to 1 minute, greatly improving the response speed to sudden failures.

[0111] 2. Multi-scale feature extraction: A complex neural network structure that includes multiple sub-networks such as high-frequency feature sub-networks and low-frequency feature sub-networks can simultaneously capture the short-term rapid changes and long-term slow changes in gas concentration. This multi-scale feature extraction method significantly improves the model's ability to identify different types of faults.

[0112] 3. Gas relationship analysis: The introduction of graph neural networks to process the relationships between different gases overcomes the limitation of traditional methods that only consider a single gas or a simple ratio. This innovation enables the model to more comprehensively understand the complex gas composition change pattern and improve the accuracy of fault diagnosis.

[0113] 4. Trace gas sensitivity: Through a dedicated trace gas feature sub-network and attention mechanism, the model's sensitivity to trace gas changes is enhanced. This enables the system to capture early signs of failure and achieve advanced warning of failures.

[0114] 5. Dynamic adaptive capability: The use of a dynamic gas composition model, combined with real-time data stream processing and model update mechanisms, enables the system to continuously learn and adapt to changes in transformer operating conditions. This greatly improves the generalization ability and long-term reliability of the model.

[0115] In summary, the solution of the present invention solves the technical problems existing in the prior art that it is difficult to capture the dynamic change trend of gas concentration and has poor sensitivity to changes in trace gases. BRIEF DESCRIPTION OF THE DRAWINGS

[0116] Figure 1 A flow chart of the method provided by the present invention. DETAILED DESCRIPTION

[0117] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0118] like Figure 1 As shown, it is a flow chart of a method for establishing a gas model on the oil surface of an oil-immersed transformer under a sudden severe fault provided by the present invention. The method comprises the following steps:

[0119] S10. Install a gas chromatography-mass spectrometry device on the top of the oil tank of the oil-immersed transformer to monitor and collect the gas composition on the oil surface in real time;

[0120] S20, according to the fault development progress of the transformer, set the sampling interval of the acquisition to be once every 5 to 10 minutes under normal operation;

[0121] S30. Under normal operating conditions of the transformer, collect gas composition data for at least 24 hours;

[0122] S40, when a sudden serious fault of the transformer is detected, immediately start the high-frequency sampling mode to collect the gas composition, and shorten the sampling interval to once every 15 seconds to 1 minute;

[0123] S50, preprocessing the collected data of the gas composition, including denoising, standardization and outlier detection;

[0124] S60, performing time series analysis on the changes of the pre-processed data of various gas components over time, identifying the main gas components and their change trends, and obtaining a gas time series matrix;

[0125] S70, based on the gas time series matrix and the moment of sudden severe fault, establish a training data set, the input of the training data set is the gas time series matrix, the output is the moment of occurrence of sudden severe fault, the severe fault category and the occurrence area, establish and train a dynamic gas composition model, and predict the change of various gas components over time;

[0126] S80, using part of the collected gas composition data as a verification set to verify and optimize the dynamic gas composition model;

[0127] S90, applying the optimized dynamic gas composition model to the real-time data stream of the oil-immersed transformer collected on site, continuously updating the monitored gas composition and prediction results, and setting a threshold for triggering an alarm mechanism according to the prediction results, so as to timely detect and predict abnormal gas changes.

[0128] The specific implementation methods of the above steps are described in detail below:

[0129] The specific implementation method of step S10 is: install a high-precision gas chromatography-mass spectrometry device on the top of the oil-immersed transformer tank to monitor the gas composition on the oil surface in real time. The specific implementation process of this step includes: first, select a suitable gas chromatography-mass spectrometry device, which should have high sensitivity, high resolution and rapid response capability to meet the needs of real-time monitoring. The selection of the equipment should consider the detection range of common gas components in transformer oil, such as hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), acetylene (C2H2), carbon monoxide (CO) and carbon dioxide (CO2). Secondly, determine the installation location, usually near the gas sampling port on the top of the oil tank to ensure that the most representative gas sample is collected. Then, install and debug the equipment, including gas line connection, power supply, data transmission interface, etc. Finally, calibrate and verify the performance of the equipment to ensure the accuracy and reliability of the measurement results. The main purpose of this step is to establish a reliable gas monitoring system to provide a basis for subsequent data acquisition and analysis.

[0130] The specific implementation method of step S20 is: according to the development speed of the transformer fault, the sampling interval is set to sample once every 5 to 10 minutes. The specific implementation process of this step includes: first, analyzing the development speed characteristics of different types of transformer faults, such as insulation aging, local overheating, arc discharge, etc. According to historical data and experience, the typical time span of various types of faults from the initial stage to the severe stage is estimated. Secondly, considering the kinetic process of gas generation and diffusion, the time constant of gas concentration change is evaluated. Then, the relationship between the sampling frequency and data storage and processing capabilities is weighed to select a suitable sampling interval. Generally, for medium-sized oil-immersed transformers, the sampling interval can be set to 7 minutes. This value can capture the development process of most faults without generating too much redundant data. Finally, a sampling program is set on the gas chromatography-mass spectrometry device to ensure automatic sampling at the set time interval. The role of this step is to optimize the utilization of system resources under the premise of ensuring data quality, and provide a sustainable data acquisition strategy for long-term monitoring.

[0131] The specific implementation of step S30 is: under the normal operating state of the transformer, collect at least 24 hours of gas composition data as a reference value. The specific implementation process of this step includes: first, ensure that the transformer is in a stable normal operating state, including parameters such as load, temperature, and voltage are all within the normal range. Secondly, start the continuous sampling mode of the gas chromatography-mass spectrometry device, and collect data according to the sampling interval set in step S20. The collected data should include the concentration values ​​of various gases and the sampling timestamp. Then, the collected data is preliminarily analyzed to calculate the statistical indicators such as the mean, standard deviation, and coefficient of variation of each gas concentration. Through these indicators, the fluctuation range of the gas composition under the "normal state" of the transformer can be preliminarily determined. Finally, these reference value data are stored in the database as a reference for subsequent abnormal detection and model training. The purpose of this step is to establish a gas composition baseline when the transformer is in normal operation, providing a comparison basis for subsequent fault diagnosis.

[0132] The specific implementation of step S40 is: when a sudden serious fault of the transformer is detected, the high-frequency sampling mode is immediately started, and the sampling interval is shortened to once every 15 seconds to 1 minute. The specific implementation process of this step includes: first, setting the fault trigger condition, which can be based on multiple indicators such as sudden change of gas concentration, abnormal gas generation rate, gas ratio analysis and abnormal electrical parameters. For example, when the concentration of hydrogen (H2) increases by more than 100ppm within 5 minutes, or the concentration of acetylene (C2H2) increases by more than 5ppm within 10 minutes, it can be regarded as a trigger condition. Secondly, develop a real-time monitoring program that continuously compares the deviation of current gas data from the normal baseline value. When the deviation exceeds the preset threshold, the high-frequency sampling mode is automatically triggered. Then, in the high-frequency sampling mode, the sampling interval is dynamically adjusted according to the severity of the fault type. For example, for suspected arc discharge faults, the sampling interval can be set to 15 seconds; for suspected local overheating faults, the sampling interval can be set to 45 seconds. Finally, ensure that the data acquisition system can handle the large amount of data generated by high-frequency sampling, including data storage, real-time transmission and preliminary processing. The purpose of this step is to quickly capture the dynamic changes in gas composition when a fault occurs, providing high-time-resolution data support for accurate diagnosis of the fault type and location.

[0133] It should be noted that one of the following methods is generally used to detect sudden serious faults in transformers:

[0134] Sudden change in gas concentration: By real-time monitoring of the concentration of dissolved gases in oil, when the concentration of one or more characteristic gases (such as hydrogen, acetylene, ethylene, etc.) rises sharply in a short period of time and exceeds the preset threshold, it is considered that a sudden serious fault has occurred.

[0135] Abnormal gas generation rate: Calculate the generation rate of characteristic gas. When the rate exceeds the normal range, it is judged as a sudden serious fault.

[0136] Gas ratio analysis: Use key gas ratios (such as C2H2 / C2H4, CH4 / H2, etc.) for analysis. When the ratio changes abnormally, it may indicate a serious fault.

[0137] Abnormal electrical parameters: Combined with abnormal changes in the transformer's electrical parameters (such as current, voltage, temperature, etc.), it helps determine whether a sudden serious fault has occurred.

[0138] The specific implementation of step S50 is: pre-processing the collected gas composition data, including denoising, standardization and outlier detection. The specific implementation process of this step includes: first, removing noise from the original data. The gas concentration time series is decomposed at multiple scales using the wavelet transform method, and a suitable threshold is selected for soft threshold processing, and then the signal is reconstructed to obtain the denoised data. Specifically, the db4 wavelet basis can be used, the number of decomposition layers is selected to be 4-5 layers, and the threshold can be adaptively selected using the VisuShrink method. Secondly, data standardization is performed. The concentration data of each gas is z-score standardized, that is, the mean is subtracted and divided by the standard deviation, so that the data of different gases are comparable. Then, outlier detection is performed. The improved local outlier factor (LOF) algorithm is used to detect outliers in multidimensional gas data. The LOF algorithm calculates the degree of anomaly of each data point based on the local density of the data point. The improvement is to introduce the concept of time window, and only consider the data of the most recent N time points to adapt to the dynamic change characteristics of gas concentration. LOF values ​​greater than 2 can be regarded as potential outliers. Finally, the detected outliers are processed and can be deleted or replaced using interpolation methods. The purpose of this step is to improve data quality, eliminate the impact of noise and outliers, and provide reliable input data for subsequent time series analysis and model building.

[0139] The specific implementation of step S60 is: perform time series analysis on the changes of various gas components over time, identify the main gas components and their change trends, and obtain a gas time series matrix. The specific implementation process of this step includes: first, perform trend analysis on the time series data of each gas. The seasonal trend decomposition (STL) method is used to decompose the time series into three components: trend, seasonality, and residual. The STL method can handle nonlinear trends and multiple periodicities and is suitable for complex gas concentration change patterns. Secondly, perform principal component analysis (PCA) to identify the gas component that contributes most to the state change of the transformer. PCA can reduce the data dimension and highlight the main change pattern. The principal component with a cumulative contribution rate of 85% is selected as the main gas component. Then, time series feature extraction is performed on the main gas components. Calculate statistical features within the sliding time window, such as mean, standard deviation, skewness, kurtosis, maximum value, minimum value, etc. The time window size can be set to 30 minutes, sliding every 5 minutes. In addition, the rate of change and acceleration of gas concentration are calculated to capture short-term change characteristics. Next, perform gas ratio analysis to calculate the ratios of key gas pairs, such as C2H2 / C2H4, CH4 / H2, C2H4 / C2H6, etc. These ratios are of great significance for fault type judgment. Finally, integrate the above analysis results into a gas time series matrix. Each row of the matrix represents a time point, and the columns include the original concentration, trend value, principal component score, statistical characteristics, change rate, acceleration, and gas ratio of each gas. The role of this step is to convert the original gas concentration data into structured data containing rich time series features, providing comprehensive input information for subsequent model establishment.

[0140] The specific implementation of step S70 is: based on the gas time series matrix and the moment of sudden serious fault, a training data set is established, the input of the training data set is the gas time series matrix, and the output is the occurrence time, the serious fault category and the occurrence area, and a dynamic gas composition model is established and trained to predict the changes of various gas components over time. The specific implementation process of this step includes: first, constructing a training data set. The gas time series matrix is ​​divided into several time windows in chronological order, and the length of each window can be set to 2 hours, with a step length of 10 minutes. For each window, it is marked whether it contains the moment of sudden serious fault. If it does, the fault category and the occurrence area are further marked. The fault category can be divided into insulation aging, local overheating, arc discharge, etc., and the occurrence area can be divided into coils, cores, casings, etc. Secondly, the network structure of the dynamic gas composition model is designed. The model uses a long short-term memory network (LSTM) as the basic structure to capture the long-term dependence of gas concentration changes. Specifically, the model includes the following subnetworks: 1) High-frequency feature subnetwork: composed of 3 layers of LSTM units, each layer with 64 neurons, used to extract the characteristics of rapid changes in gas concentration in a short time. 2) Low-frequency feature subnetwork: It consists of 2 layers of gated recurrent units (GRU), with 32 neurons in each layer, which is used to capture the long-term trend of gas concentration changes. 3) Main gas feature subnetwork: It contains a fully connected layer (64 neurons) and an LSTM layer (32 neurons) to analyze the concentration change characteristics of the main gases. 4) Trace gas feature subnetwork: It contains a multi-head attention layer (4 heads, 16 dimensions each) and an LSTM layer (32 neurons) to analyze the concentration change characteristics of trace gases. 5) Temporal feature subnetwork: It uses a temporal convolutional network (TCN), which contains 3 dilated convolution layers with a convolution kernel size of 3 and dilation rates of 1, 2, and 4, respectively, to extract the temporal pattern of gas concentration changes. 6) Spatial feature subnetwork: It uses a graph neural network (GNN) with a graph attention network (GAT) structure, which contains 2 GAT layers with 16 attention heads in each layer to analyze the relationship between different gases. 7) Fusion feature subnetwork: Contains a multi-head self-attention mechanism (8 heads, each with 32 dimensions) and two fully connected layers (128 and 64 neurons, respectively) to integrate the output features of each subnetwork. Then, define the loss function and optimization strategy. For the prediction of the time of fault occurrence, the mean square error (MSE) loss is used; for the prediction of fault category and occurrence area, the cross entropy loss is used. The total loss is the weighted sum of the two parts. The optimizer uses the Adam algorithm, the initial value of the learning rate is set to 0.001, and the learning rate decay strategy is adopted, decaying by 10% every 50 epochs. Next, perform model training. Use 80% of the data as the training set and 20% as the validation set. Use the small batch gradient descent method, and the batch size is set to 64.An early stopping strategy is used during training, and training is stopped when the loss on the validation set has not improved for 10 consecutive epochs. Finally, the trained model is tested and evaluated. The remaining data is used as a test set to evaluate the performance of the model on unseen data. Indicators such as the mean absolute error (MAE) of the prediction of the moment of fault occurrence, the accuracy of the prediction of the fault category and occurrence area, and the F1 score are calculated. The purpose of this step is to establish a deep learning model that can accurately predict the dynamic changes of gas composition and provide strong support for fault diagnosis and early warning.

[0141] The following is a detailed explanation of each sub-network:

[0142] 1. High-frequency feature sub-network and low-frequency feature sub-network:

[0143] The outputs of these two sub-networks are used as the input of the temporal feature sub-network in a weighted fusion manner. The weights can be automatically adjusted through network training to balance the importance of short-term and long-term features.

[0144] The formula means:

[0145] I time =w h O h +w l O l

[0146] in:

[0147] I time : Input of the temporal feature subnetwork

[0148] O h : Output of the high-frequency feature sub-network

[0149] O l : Output of the low-frequency feature sub-network

[0150] w h , w l : The corresponding weight coefficient

[0151] 2. Main gas characteristic sub-network and trace gas characteristic sub-network:

[0152] The outputs of these two sub-networks are fused through the attention mechanism and used as the input of the spatial feature sub-network, which can dynamically adjust the importance of different gas features.

[0153] The formula means:

[0154] I space =Attention(O b , O s )

[0155] in:

[0156] I space : Input of the spatial feature subnetwork

[0157] O b : Output of the main gas characteristics subnetwork

[0158] O s : Output of the trace gas signature subnetwork

[0159] Attention(): attention mechanism function

[0160] 3. Temporal feature sub-network and spatial feature sub-network:

[0161] The outputs of these two sub-networks are merged through a concatenation operation and then passed through a fully connected layer for dimensionality reduction as the input of the fusion feature sub-network.

[0162] The formula means:

[0163] I fusion =FC([O time ;O space ])

[0164] in:

[0165] I fusion : Input of the fusion feature sub-network

[0166] O time : Output of the temporal feature subnetwork

[0167] O space : Output of the spatial feature subnetwork

[0168] [;]: indicates a connection operation

[0169] FC(): Fully connected layer function

[0170] 4. Fusion feature sub-network:

[0171] The fusion feature sub-network receives the output of all other sub-networks as input, performs feature fusion through the multi-head self-attention mechanism, and finally outputs the final prediction result through the fully connected layer.

[0172] The formula means:

[0173] O final =FC(MultiHeadAttention(I fusion , [O h , O l , O b , O s , O time , O space ]))

[0174] in:

[0175] O final : The final output of the model

[0176] MultiHeadAttention(): Multi-head self-attention mechanism function.

[0177] In oil-immersed transformers, the distinction between major and trace gases is usually based on their typical concentration levels in the transformer oil.

[0178] 1. Main gas (main gas):

[0179] Gases with concentrations exceeding 10 ppm (parts per million) are usually considered major gases. These gases include: hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), carbon monoxide (CO);

[0180] 2. Trace gases (trace gases):

[0181] Gases with concentrations generally below 10 ppm are considered trace gases. These gases include: acetylene (C2H2), propane (C3H8), propylene (C3H6), carbon dioxide (CO2) (although CO2 concentrations may be higher than 10 ppm, it is generally considered a trace gas);

[0182] The specific threshold can be expressed as follows:

[0183] T threshold =10ppm

[0184] For gas i, its concentration C i The classification can be expressed as:

[0185]

[0186] It should be noted that:

[0187] 1. This threshold value (10ppm) is a commonly used reference value, but it may be adjusted according to specific circumstances in actual applications.

[0188] 2. Some gases (such as CO2) may have concentrations much higher than 10ppm but are still classified as trace gases for other reasons (such as low importance for fault diagnosis).

[0189] 3. In some serious fault conditions, gases that are usually considered trace amounts (such as acetylene) may quickly increase to more than 10ppm, at which time its importance in fault diagnosis will be greatly increased.

[0190] 4. The relative proportion and generation rate of gas are usually more important than the absolute concentration. Therefore, in actual fault diagnosis, not only the absolute concentration of gas should be considered, but also the rate of change of gas concentration and the proportional relationship between gases should be analyzed.

[0191] Therefore, when building the model, this threshold should be considered as a tunable parameter that can be optimized based on the specific transformer and operating environment.

[0192] The following describes in detail the input and output of the training set, as well as the input and output of each sub-network during the training process of the dynamic gas composition model.

[0193] 1. Input and output of the training set:

[0194] The input includes: gas time series matrix G, time series vector T, sampling frequency information (used to distinguish high-frequency and low-frequency data), gas type information (used to distinguish major and trace gases);

[0195] The output includes: fault occurrence time t f , Serious fault category C f (can be a multi-category label), occurrence region R f (can be a multi-category label or a continuous value representing the coordinates);

[0196] 2. Output of each sub-network:

[0197] a. High-frequency feature sub-network:

[0198] Input: high-frequency sampled gas time series data;

[0199] Output: where d h is the dimension of the high-frequency feature vector;

[0200] b. Low-frequency feature sub-network:

[0201] Input: low-frequency sampled gas time series data;

[0202] Output: where d l is the dimension of the low-frequency feature vector;

[0203] c. Main gas characteristic subnetwork:

[0204] Input: Time series data of main gases;

[0205] Output: where d b is the dimension of the main gas eigenvector;

[0206] d. Trace gas characteristic subnetwork:

[0207] Input: Time series data of trace gases;

[0208] Output: where d s is the dimension of the trace gas feature vector;

[0209] e. Temporal feature sub-network:

[0210] Input: I time =w h O h +w l O l ;

[0211] Output: where d t is the dimension of the time feature vector;

[0212] f. Spatial feature sub-network:

[0213] Input: I space =Attention(O b ,O s );

[0214] Output: where d p is the dimension of the spatial eigenvector;

[0215] g. Fusion feature sub-network:

[0216] Input: I fusion =FC([O time ;O space ]);

[0217] Output: O final =[t f ,C f ,R f ], which is the final prediction result;

[0218] 3. Training process:

[0219] During the training process, you can use an end-to-end training method or a phased training strategy:

[0220] a. End-to-end training: Input the entire gas timing matrix and related information, and directly optimize the entire network to minimize the loss between the predicted output and the true label.

[0221] The loss function can be defined as:

[0222]

[0223] Where L timeIt can be the mean square error, L class It can be the cross entropy loss, L region It can be mean square error or cross entropy loss (depending on the region representation), λ1,λ2,λ3 are weight coefficients.

[0224] b. Training in stages:

[0225] 1. Pre-train each sub-network, using self-supervised learning or specific tasks

[0226] 2. Fix sub-network parameters and train fusion feature sub-network

[0227] 3. Fine-tune the entire network

[0228] 4. Optional, model evaluation: Evaluate model performance using an independent test set; calculate the mean absolute error (MAE) or root mean square error (RMSE) of the predictions at the time of fault occurrence; calculate the accuracy, precision, recall, and F1 score of the predictions for fault category and region; analyze the model's performance under different types of faults and different severities.

[0229] In this way, the gas data features of different scales and types can be fully utilized to improve the model's ability to identify complex fault modes. Each sub-network focuses on extracting a specific type of feature, while the fusion feature sub-network is responsible for integrating these features to obtain the final prediction result.

[0230] The specific implementation method of step S80 is: use part of the collected data as a validation set to verify and optimize the dynamic gas composition model to improve the accuracy and generalization ability of the model. The specific implementation process of this step includes: first, prepare a validation data set. Randomly select 20% of the collected data as a validation set to ensure that the validation set contains different types of fault cases and normal operation data. Secondly, perform model validation. Input the validation set data into the trained model to obtain the prediction results. Calculate various performance indicators on the validation set, including the root mean square error (RMSE) and mean absolute percentage error (MAPE) of the prediction at the time of fault occurrence, and the accuracy, precision, recall and F1 score of the prediction of fault category and occurrence area. Then, analyze the model performance. Compare the validation set performance with the training set performance to evaluate whether the model has overfitting or underfitting problems. If the validation set performance is significantly lower than the training set performance, it means that there is an overfitting problem; if both performances are low, there may be an underfitting problem. Then, perform model optimization. According to the verification results, the following optimization strategies are adopted: 1) Regularization: If there is an overfitting problem, L2 regularization terms can be added to the model, and the regularization coefficient can be gradually adjusted starting from 0.001. 2) Adjust the network structure: If the model complexity is insufficient, the number of network layers or neurons can be appropriately increased; if the model is too complex, the number of layers can be reduced or dropout technology can be used. The dropout rate can be set between 0.2-0.5. 3) Feature Engineering: Analyze the contribution of different features to model performance. Feature importance analysis methods such as permutation importance can be used. According to the analysis results, important features are retained and minor features are removed or combined. 4) Ensemble learning: Model ensemble technology, such as random forest or gradient boosting tree, is used in combination with deep learning models to form a hybrid model. The prediction results of different models can be fused by soft voting or weighted averaging. 5) Transfer learning: If the verification data contains new types of faults, transfer learning technology can be used. Freeze the bottom layer of the pre-trained model and only fine-tune the top layer to adapt to the new fault type. 6) Data enhancement: If there are fewer samples of certain fault types, data enhancement techniques can be used, such as adding Gaussian noise, time series interpolation, and other methods to generate new samples. Then, cross-validation is performed. Using the k-fold cross-validation method (such as k=5), the data set is divided into 5 parts, and 4 of them are used as training sets and 1 as a validation set in turn. This can more comprehensively evaluate the performance of the model and reduce the risk of overfitting. Finally, the model is fine-tuned. According to the cross-validation results, the model is fine-tuned for the final time. Methods such as grid search or Bayesian optimization can be used to automatically search for the optimal hyperparameter combination. The purpose of this step is to improve the generalization ability and robustness of the model through a rigorous verification and optimization process, and to ensure that the model can reliably predict the dynamic changes of gas composition in practical applications.

[0231] The specific implementation method of step S90 is: apply the optimized model to the real-time data stream, continuously update the gas composition prediction results, and set the threshold to trigger the alarm mechanism to detect abnormal gas changes in time. The specific implementation process of this step includes: first, deploy the optimized model. Deploy the model to the real-time data processing server to ensure that the model can efficiently process the input gas time series matrix. Tools such as TensorFlow Serving or ONNX Runtime can be used for model deployment to improve the reasoning speed. Secondly, realize real-time data stream processing. Develop a data stream processing module that can receive data from the gas chromatography-mass spectrometry device in real time and convert it into the input format required by the model. Using sliding window technology, maintain a data window of a fixed length (such as 2 hours), and update the window content every time new data arrives. Then, perform real-time prediction. Whenever the data window is updated, the latest gas time series matrix is ​​input into the model to obtain the prediction results. The prediction results include the concentration change trend of various gas components in the future period (such as 30 minutes), as well as the possible fault types and locations. Next, set the alarm mechanism. Based on the model prediction results, set multi-level alarm thresholds: 1) Slight anomaly: triggered when the predicted concentration of a certain gas exceeds 1.5 times the normal range in the next 30 minutes. 2) Moderate anomaly: triggered when the predicted concentration exceeds 2 times the normal range, or when multiple gases have slight anomalies at the same time. 3) Severe anomaly: triggered when the predicted concentration exceeds 3 times the normal range, or when the model predicts a high probability (such as greater than 80%) of a serious failure. The normal range can be determined based on statistical analysis of historical data, such as using the 3σ principle. In addition, an adaptive threshold mechanism is implemented. The normal range of each gas is dynamically updated using the exponentially weighted moving average (EWMA) method, and the weight factor can be set to 0.1 to balance the impact of historical data and the latest trends. Then, a visual interface is developed. Design an intuitive dashboard to display the concentration, change trend, prediction results, and alarm status of various gases in real time. Use different colors and icons to represent different levels of anomalies, so that operators can quickly identify potential problems. Next, implement the alarm push function. When an alarm is triggered, the alarm information is sent to relevant personnel through multiple channels (such as SMS, email, and mobile application push). The alarm information should contain key information such as abnormal gas type, concentration value, predicted trend, possible fault type and location. Finally, establish an alarm log and feedback mechanism. Record detailed information of each alarm, including trigger time, alarm level, related data, etc. Design a feedback interface to allow operators to confirm, classify and annotate the alarm, which can be used for subsequent model optimization and system improvement. The purpose of this step is to apply the optimized dynamic gas composition model to the actual operating environment, realize real-time monitoring and early warning of transformer gas status, improve the timeliness and accuracy of fault detection, and provide strong guarantee for the safe operation of transformers.

[0232] Among them, the gas timing matrix is ​​specifically expressed as:

[0233]

[0234] Where G is the gas time series matrix; g ij is the concentration value of the i-th gas at the j-th time point (unit: μL / L or ppm); m is the number of monitored gas types; n is the length of the time series;

[0235] The steps to establish the gas timing matrix include:

[0236] 1. Data collection:

[0237] First, a high-precision gas chromatography-mass spectrometry device is installed on the top of the oil tank of the oil-immersed transformer. According to the operating status of the transformer, an appropriate sampling interval is set. Under normal operating conditions, sampling is performed every 5 to 10 minutes; when a sudden serious fault of the transformer is detected, the sampling interval is shortened to once every 15 seconds to 1 minute. The concentration data of multiple dissolved gases are continuously collected, including major characteristic gases such as hydrogen, methane, ethylene, acetylene, and trace gases such as carbon monoxide.

[0238] 2. Data preprocessing:

[0239] The collected raw data is preprocessed, including denoising, standardization and outlier detection. Denoising can remove high-frequency noise using methods such as moving average or wavelet transform. Standardization can unify the concentration data of different gases to the same scale. Outlier detection can identify and process data points that are significantly deviated from the normal range due to equipment failure or environmental interference.

[0240] 3. Time series construction:

[0241] The preprocessed data are arranged in the order of sampling time to construct the time series of each gas. Each time point corresponds to a concentration value, forming a two-dimensional data structure, in which rows represent different gas types and columns represent different time points.

[0242] 4. Trend function fitting:

[0243] A trend function is fitted for the concentration time series of each gas. This trend function is a composite function, which contains a constant term, a linear growth term, an exponential growth term, a Sigmoid function, and a sinusoidal fluctuation term. The parameters of the function are estimated by the nonlinear least squares method to make the fitting curve fit the actual data points to the greatest extent.

[0244] 5. Parameter Estimation and Optimization:

[0245] First, the initial parameter estimation is performed, and then the parameters are fine-tuned using optimization methods such as the Levenberg-Marquardt algorithm. This process requires repeated iterations until the preset convergence condition or the maximum number of iterations is reached.

[0246] 6. Model Validation:

[0247] Use part of the data as a validation set to calculate indicators such as the root mean square error (RMSE) between the model prediction value and the actual observation value to evaluate the model's fitting effect and prediction ability. If the validation result is not ideal, you may need to return to the parameter estimation step to adjust the model structure or optimize the method.

[0248] 7. Calculation of gas generation rate and ratio:

[0249] Based on the fitted trend function, the generation rate of each gas (the derivative of concentration with respect to time) and the concentration ratio between different gases are calculated. These derived indicators can provide additional fault diagnosis information.

[0250] 8. Matrix update mechanism:

[0251] Establish a sliding window mechanism to update the gas time series matrix as new data is continuously input. You can set a fixed-size time window to only retain data from the most recent period, or use methods such as exponential weighting to give recent data a higher weight.

[0252] The following is a detailed description of the calculation process involved in establishing the gas timing matrix:

[0253] (1) The time series is expressed as: T = [t1, t2, …, t n ];

[0254] Where T is the time series vector; t j is the jth sampling time point (unit: seconds, relative to the monitoring start time).

[0255] (2) The function of gas concentration changing with time is specifically expressed as:

[0256] C i (t) = f i (t)+∈ i (t);

[0257] In the formula, C i (t): the concentration of the i-th gas at time t (unit: μL / L or ppm); f i (t): concentration trend function of the i-th gas; ∈ i (t): Random error term, subject to normal distribution Variance of the measurement error of the i-th gas.

[0258] (3) Gas concentration change trend function:

[0259]

[0260] In the formula, a i is the initial concentration constant (in μL / L or ppm); b i is the linear change rate (in μL / L / s or ppm / s); c i is the exponential growth coefficient (in μL / L or ppm); d i is the exponential growth rate (unit: 1 / s); e i is the amplitude of the Sigmoid function (in μL / L or ppm); k i is the steepness parameter of the Sigmoid function (unit: 1 / s); t 0i is the midpoint time of the Sigmoid function (in seconds); g i is the amplitude of the sine function (in μL / L or ppm); h i is the angular frequency of the sine function (in rad / s); ∈ i is the initial phase of the sine function (in rad).

[0261] (4) Gas generation rate:

[0262]

[0263]

[0264] Where R i (t) is the generation rate of the ith gas at time t (in μL / L / s or ppm / s).

[0265] (5) Gas concentration ratio:

[0266]

[0267] In the formula, S ij (t) is the concentration ratio of the ith gas to the jth gas at time t (dimensionless).

[0268] (6) The estimation steps of relevant parameters are:

[0269] a. Initial parameter estimates:

[0270] a i =C i (t1) is the concentration value at the beginning of the time series;

[0271] bi To estimate the overall trend using simple linear regression:

[0272]

[0273] in,

[0274] c i ,d i To estimate the exponential growth part using linear regression after log transformation:

[0275] ln(C i (t)-a i -b i t)≈ln(c i )+d i t;

[0276] e i ,k i ,t 0i To observe the mutation points in the data, the Sigmoid function parameters are preliminarily estimated;

[0277] g i ,h i ,φ i To estimate the periodic components using Fourier transform;

[0278] b. Nonlinear least squares fitting: Use the Levenberg-Marquardt algorithm for nonlinear least squares fitting.

[0279] The objective function is:

[0280] Among them, θ i =[a i ,b i ,c i ,d i ,e i ,k i ,t 0i ,g i ,h i ,φ i ] is the set of parameters to be optimized.

[0281] c. Model validation is

[0282] The root mean square error (RMSE) was used to evaluate the model fitting effect:

[0283]

[0284] d. Update the parameters to

[0285] The sliding window method is used to continuously update the model parameters. For a new data point C i (t n+1 ), update parameters:

[0286] Among them, k = max(1,n+2-w), w is the size of the sliding window.

[0287] 1. Function f i (t) can be a polynomial function, an exponential function, or other function suitable for describing the change in gas concentration.

[0288] 2. Generation rate R i (t) is calculated using numerical differentiation methods, such as the central difference method.

[0289] 3. Gas concentration ratio S ij (t) Used to analyze the relative change relationship between different gases.

[0290] Considering the complexity of the gas concentration variation in the oil-immersed transformer, a composite function containing multiple components can be used to represent f i (t).

[0291] f i The specific equation of (t) can be expressed as:

[0292]

[0293] Where: t is the time variable; a i is the initial concentration constant; b i is the linear rate of change; c i ,d i is the coefficient of the exponential growth term; e i ,k i ,t 0i is the parameter of the Sigmoid function, used to simulate the mutation process; g i ,h i ,φ i is the amplitude, frequency and phase of the sine function, which is used to simulate periodic fluctuations.

[0294] This equation contains the following components:

[0295] 1. Constant term: a i

[0296] 2. Linear growth term: b i t

[0297] 3. Exponential growth items:

[0298] 4. Sigmoid function:

[0299] 5. Sinusoidal wave term: g i sin(h i t+φ i )

[0300] Fitting method:

[0301] 1. Data preprocessing: denoise and standardize the original gas concentration data; convert the time series into relative time (starting from 0).

[0302] 2. Initial parameter estimation:

[0303] a i : Take the concentration value at the beginning of the time series;

[0304] b i :Use simple linear regression to estimate the overall trend;

[0305] c i ,d i : Estimate the exponential growth component using linear regression after log transformation;

[0306] e i ,k i ,t 0i : Observe the mutation points in the data and preliminarily estimate the parameters of the Sigmoid function;

[0307] g i ,h i ,φ i : Estimate periodic components using Fourier transform.

[0308] 3. Nonlinear least squares fitting: Use the Levenberg-Marquardt algorithm for nonlinear least squares fitting. The objective function is:

[0309]

[0310] Among them, y ij is the i-th gas at time t j The actual observed value, θ i is the set of parameters to be optimized.

[0311] Specifically, the principles of the present invention include:

[0312] 1. Principle of multi-source information fusion: The present invention comprehensively utilizes multi-source information such as gas concentration, electrical parameters, historical data, etc., and performs feature extraction and fusion through a deep learning model. This method meets the requirements of systematic and comprehensive fault diagnosis and can more accurately reflect the actual status of the transformer.

[0313] 2. Time series analysis principle: The use of recurrent neural network structures such as long short-term memory network (LSTM) and gated recurrent unit (GRU) can effectively capture the temporal characteristics of gas concentration changes. These network structures have memory capabilities and can learn long-term dependencies, which are very suitable for processing time series data.

[0314] 3. Multi-scale analysis principle: By designing high-frequency and low-frequency characteristic sub-networks, multi-scale analysis of gas concentration changes is achieved. This method conforms to the multi-time scale characteristics of transformer fault development and can capture both fast-changing and slowly evolving fault modes.

[0315] 4. Graph structure data processing principle: Graph neural network is introduced to process the relationship between different gases, and the gas composition is regarded as an interrelated system. This method conforms to the complex physical and chemical reaction mechanism inside the transformer and can better simulate the mutual influence between gases.

[0316] 5. Principle of attention mechanism: The attention mechanism is used in the trace gas feature subnetwork to automatically learn the importance weights of different gases for fault diagnosis. This method is consistent with the fact that different gases have different indicative effects on faults in actual situations.

[0317] 6. Transfer learning principle: By introducing transfer learning technology in the model optimization stage, the model can adapt to new types of faults and different types of transformers. This method improves the generalization ability and practicality of the model.

[0318] 7. Adaptive sampling principle: dynamically adjust the sampling frequency according to the transformer status, use a lower frequency in normal conditions, and quickly increase the sampling frequency in fault conditions. This strategy not only ensures the real-time nature of the data, but also optimizes the use of system resources.

[0319] In order to better understand and implement the present invention, an embodiment of a specific application scenario of the present invention is provided below: This embodiment takes a 300MVA oil-immersed main transformer of a 220kV substation as the object and describes in detail the specific implementation process of the present invention.

[0320] 1. Equipment Installation and Initial Setup

[0321] An Agilent 7890B gas chromatograph-5977B mass spectrometer is installed on the top of the transformer oil tank. The device has a detection limit of 0.1ppm and can simultaneously detect seven characteristic gases, including hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), acetylene (C2H2), carbon monoxide (CO) and carbon dioxide (CO2). The device is installed about 30cm from the oil surface and connected to the gas sampling valve on the top of the oil tank through a stainless steel pipeline.

[0322] The initial sampling interval is set to 7 minutes, and the data is transmitted to the on-site server in real time via industrial Ethernet. The server is configured with an Intel Xeon E5-2680 v4 CPU, 128GB RAM, and uses the Ubuntu 20.04LTS operating system.

[0323] 2. Benchmark data collection

[0324] Under normal operation of the transformer, gas data was collected continuously for 72 hours, and a total of 617 sets of samples were obtained. Each set of samples contains the concentration values ​​of 7 gases and the corresponding timestamps. After statistical analysis of these data, the baseline value range of each gas is shown in Table 1:

[0325] Table 1 Reference value range of each gas

[0326] gas Average value (ppm) Standard Deviation (ppm) Coefficient of variation (%) H2 15.3 2.1 13.7 CH4 8.7 1.4 16.1 C2H6 3.2 0.5 15.6 C2H4 12.6 1.8 14.3 C2H2 0.5 0.1 20.0 CO 245.8 28.3 11.5 CO2 1856.4 156.7 8.4

[0327] 3. Fault trigger sampling

[0328] The following fault trigger conditions are set:

[0329] H2 concentration increased by more than 10 ppm within 5 minutes

[0330] C2H2 concentration increased by more than 1 ppm within 10 minutes

[0331] C2H4 / C2H6 ratio exceeds 3

[0332] When any of the conditions is met, the system automatically switches to high-frequency sampling mode and the sampling interval is shortened to 30 seconds.

[0333] 4. Data Preprocessing

[0334] Taking 2 hours of data after a high-frequency sampling is triggered as an example, the data preprocessing process is demonstrated.

[0335] First, the original data is decomposed and reconstructed in 5 layers using db4 wavelet to remove high-frequency noise. Then z-score normalization is performed. The improved local anomaly factor algorithm is used to detect outliers on the standardized data, and the time window is set to 150 data points.

[0336] The comparison of C2H2 concentration data before and after treatment is shown in Table 2:

[0337] Table 2 C2H2 concentration data

[0338] Time (min) Original value (ppm) After denoising(ppm) Normalized value Is it abnormal? 0 0.52 0.51 -0.86 no 0.5 0.55 0.54 -0.64 no 1 0.61 0.60 -0.21 no ... ... ... ... ... 58.5 2.85 2.83 3.97 yes 59 2.91 2.89 4.12 yes 59.5 2.88 2.87 4.07 yes ... ... ... ... ... 119.5 4.62 4.60 7.86 no

[0339] 5. Time Series Analysis

[0340] The preprocessed data is decomposed into seasonal trends to extract trend components. Then principal component analysis is performed to select principal components with a cumulative contribution rate of 85%. In this example, the contribution rates of the first three principal components are 58.3%, 19.7% and 9.2% respectively, with a cumulative contribution rate of 87.2%.

[0341] Calculate the statistical characteristics within a 30-minute sliding window, including mean, standard deviation, slope, kurtosis, maximum and minimum values. Also calculate the rate of change and acceleration of each gas.

[0342] The key gas ratio analysis results are shown in Table 3:

[0343] Table 3 Key gas ratios

[0344] Time (min) C2H2 / C2H4 CH4 / H2 C2H4 / C2H6 0 0.041 0.569 3.938 30 0.078 0.542 4.156 60 0.153 0.518 4.375 90 0.228 0.497 4.594 120 0.303 0.478 4.813

[0345] 6. Model training and validation

[0346] The training data set is constructed by dividing the 2-hour data into 12 samples with a step length of 10 minutes. Each sample contains feature data of 240 time points, with labels such as whether it contains the fault moment (0 or 1), the fault category (arc discharge in this case) and the occurrence area (coil in this case).

[0347] The model structure is as follows:

[0348] High-frequency feature subnetwork: 3 layers of LSTM, 64 neurons per layer

[0349] Low-frequency feature subnetwork: 2 layers of GRU, 32 neurons per layer

[0350] Main gas feature subnetwork: 1 layer of full connection (64 neurons) + 1 layer of LSTM (32 neurons)

[0351] Trace gas feature subnetwork: 1 layer of multi-head attention (4 heads, 16 dimensions each) + 1 layer of LSTM (32 neurons)

[0352] Temporal feature subnetwork: 3-layer TCN, convolution kernel size is 3, and expansion rates are 1, 2, and 4 respectively

[0353] Spatial feature subnetwork: 2 layers of GAT, 16 attention heads per layer

[0354] Fusion feature sub-network: 1 layer of multi-head self-attention (8 heads, 32 dimensions each) + 2 layers of full connection (128 and 64 neurons)

[0355] The Adam optimizer is used, with an initial learning rate of 0.001 and a decay of 10% every 50 epochs. The batch size is set to 64, the maximum number of epochs is 500, and the early stopping strategy is adopted.

[0356] 80% of the data is used for training and 20% for validation. The loss curve during training is shown in Table 4:

[0357] Table 4 Training data table

[0358]

[0359]

[0360] The performance indicators of the model on the validation set are as follows:

[0361] Mean absolute error of failure prediction time: 1.8 minutes

[0362] Fault category prediction accuracy: 93.7%

[0363] Fault occurrence area prediction accuracy: 89.5%

[0364] 7. Model Optimization

[0365] To improve model performance, the following optimization measures were taken:

[0366] 1) Add L2 regularization with a coefficient of 0.001

[0367] 2) Add a Dropout layer after the fully connected layer with a ratio of 0.3

[0368] 3) Use feature importance analysis to remove features with a contribution of less than 1%

[0369] 4) Integrate the random forest model and perform weighted averaging with the deep learning model results, with a weight ratio of 3:7

[0370] After optimization, the model performance is improved as follows:

[0371] Mean absolute error of prediction of fault occurrence time: 1.2 minutes

[0372] Fault category prediction accuracy: 96.4%

[0373] Fault occurrence area prediction accuracy: 92.8%

[0374] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for establishing a gas model on the oil surface of an oil-immersed transformer under sudden severe fault, characterized in that: The following steps are involved: S10. Install a gas chromatography-mass spectrometry device on the top of the oil tank of the oil-immersed transformer to monitor and collect the gas composition on the oil surface in real time; S20, according to the fault development progress of the transformer, set the sampling interval of the acquisition to be once every 5 to 10 minutes under normal operation; S30. Under normal operating conditions of the transformer, collect gas composition data for at least 24 hours; S40, when a sudden serious fault of the transformer is detected, immediately start the high-frequency sampling mode to collect the gas composition, and shorten the sampling interval to once every 15 seconds to 1 minute; S50, preprocessing the collected data of the gas composition, including denoising, standardization and outlier detection; S60, performing time series analysis on the changes of the pre-processed data of various gas components over time, identifying the main gas components and their change trends, and obtaining a gas time series matrix; S70, based on the gas time series matrix and the moment of sudden severe fault, establish a training data set, the input of the training data set is the gas time series matrix, the output is the moment of occurrence of sudden severe fault, the severe fault category and the occurrence area, establish and train a dynamic gas composition model, and predict the change of various gas components over time; S80, using part of the collected gas composition data as a verification set to verify and optimize the dynamic gas composition model; S90, applying the optimized dynamic gas composition model to the real-time data stream of the oil-immersed transformer collected on site, continuously updating the monitored gas composition and prediction results, and setting a threshold for triggering an alarm mechanism according to the prediction results, so as to timely detect and predict abnormal gas changes.

2. The method for establishing a gas model on the oil surface of an oil-immersed transformer under sudden severe fault according to claim 1 is characterized in that: The dynamic gas composition model adopts a long memory network structure, including a high-frequency characteristic sub-network, a low-frequency characteristic sub-network, a main gas characteristic sub-network, a trace gas characteristic sub-network, a time characteristic sub-network, a space characteristic sub-network and a fusion characteristic sub-network; the gas components include the main gas and trace gas.

3. The method for establishing a gas model on the oil surface of an oil-immersed transformer under sudden severe fault according to claim 2 is characterized in that: The high-frequency feature sub-network is used to extract the characteristics of rapid changes in gas concentration in a short period of time. The structure of the high-frequency feature sub-network is composed of multiple layers of LSTM units, and the input is the data of gas composition collected in the high-frequency sampling mode in step S40.

4. The method for establishing a gas model on the oil surface of an oil-immersed transformer under sudden severe fault according to claim 3 is characterized in that: The low-frequency feature sub-network is used to capture the long-term change trend of gas concentration. The structure of the low-frequency feature sub-network is composed of multiple layers of GRU units, and the input is the data of gas composition collected in the low-frequency sampling mode of the sampling interval sampling in step S20.

5. The method for establishing a gas model on the oil surface of an oil-immersed transformer under sudden severe fault according to claim 4 is characterized in that: The main gas characteristic subnetwork is used to analyze the concentration change characteristics of the main gases, which include hydrogen, methane and ethylene. The structure of the main gas characteristic subnetwork is composed of a fully connected layer and an LSTM layer.

6. The method for establishing a gas model on the oil surface of an oil-immersed transformer under sudden severe fault according to claim 5 is characterized in that: The trace gas characteristic sub-network is used to analyze the concentration change characteristics of the trace gas, and the trace gas includes: acetylene and carbon monoxide. The structure of the trace gas characteristic sub-network is composed of an attention mechanism and an LSTM layer.

7. The method for establishing a gas model on the oil surface of an oil-immersed transformer under sudden severe fault according to claim 6 is characterized in that: The time feature sub-network is used to extract the time pattern of gas concentration changes, and the structure of the time feature sub-network is composed of a time convolutional network.

8. The method for establishing a gas model on the oil surface of an oil-immersed transformer under sudden severe fault according to claim 7 is characterized in that: The spatial feature subnetwork is used to analyze the relationship between different gases, and the structure of the spatial feature subnetwork is composed of a graph neural network.

9. The method for establishing a gas model on the oil surface of an oil-immersed transformer under sudden severe fault according to claim 8, characterized in that: The fusion feature sub-network is used to integrate the output features of each sub-network, and the structure of the fusion feature sub-network is composed of a multi-head self-attention mechanism and a fully connected layer.

10. The method for establishing a gas model on the oil surface of an oil-immersed transformer under sudden severe fault according to claim 9, characterized in that: The input and output relationship of each subnetwork is that each subnetwork receives the corresponding part of the gas timing matrix as input, and the output feature vector is passed to the fusion feature subnetwork for integration, and finally the prediction results of the fault occurrence time, serious fault category and occurrence area are output.

Citation Information

Cited By

  • High-sensitivity multi-parameter gas decomposition product long-term monitoring and analyzing system

    CN119534572A

  • Method and system for evaluating filtering effect of semiconductor preparation waste gas

    CN120012003A

  • A method and system for evaluating the filtering effect of waste gas in semiconductor preparation

    CN120012003B

  • Generator set relay protection debugging method and related device

    CN120109729A

  • Online monitoring system and method for dissolved gas in transformer oil

    CN121068818A