A fault detection method for oil-immersed transformers based on gases on the oil surface

Abstract

The application provides an oil-immersed transformer fault detection method based on gas on the oil surface, and belongs to the technical field of oil-immersed transformer fault detection, and comprises the following steps: the method collects historical operation parameters and internal gas data of the oil-immersed transformer in real time, establishes a mathematical model of the transformer operation state and the gas concentration change, analyzes the correlation between the two, and constructs a gas-operation parameter change matrix. A training set is established based on the data of fault-free and known fault transformers, and a neural network model capable of identifying the fault type of the transformer to be detected is trained. Finally, the transformer to be detected is monitored and diagnosed in real time. The method fully utilizes the internal data of the transformer, establishes a complete fault diagnosis system, and solves the technical problems that the existing methods only consider static gas characteristics, ignore the dynamic relationship between the transformer state and the gas concentration, and are difficult to accurately capture the transformer fault.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of oil-immersed transformer fault detection, and in particular, relates to an oil-immersed transformer fault detection method based on gas on the oil surface. BACKGROUND

[0002] Oil-immersed transformers are key equipment in power systems, and their reliability directly affects the safe and stable operation of power grids. With the development of smart grids and the growing demand for electricity, there is a higher demand for transformer operation state monitoring and fault diagnosis. Traditional transformer fault diagnosis methods mainly rely on periodic offline detection and manual experience judgment, which has problems such as diagnosis lag and low efficiency, and is difficult to meet the needs of modern power grids for real-time monitoring and predictive maintenance.

[0003] In recent years, with the rapid development of sensing technology, communication technology and artificial intelligence technology, transformer fault diagnosis methods based on online monitoring data have been widely studied and applied. These methods mainly include dissolved gas analysis (DGA) based fault diagnosis, electrical parameter based fault diagnosis, and vibration and sound signal based fault diagnosis. Among them, the DGA method has attracted widespread attention because it can detect internal transformer faults early. Existing DGA fault diagnosis methods mainly include ratio methods (such as the Rogers ratio method and the Duval triangle method) and artificial intelligence methods (such as expert systems, neural networks, support vector machines, etc.).

[0004] However, there are still some problems in the existing technology: first, single DGA data or electrical parameters are difficult to fully reflect the operation state of the transformer, which can easily lead to one-sidedness and inaccuracy of the diagnosis results. Second, although the traditional ratio method is simple and intuitive, it is difficult to handle complex fault patterns and multiple faults coexisting. Third, although artificial intelligence methods have improved the accuracy of diagnosis to some extent, most methods only consider static features, ignoring the dynamic change process of transformer parameters, making it difficult to capture the trend of fault development. Finally, existing methods generally lack explainability in the model decision-making process, making it difficult to gain the trust of users and promote in practical applications.

[0005] For example, Chinese patent No. CN202011102982.7 discloses a transformer oil gas fault diagnosis method, which uses a threshold judgment method to compare the concentration of each gas dissolved in the oil with the normal limit value to determine whether the transformer has a fault. Alternatively, according to the gas production rate, the gas concentration and the gas production rate of each component are combined, and if the content of each component gas rapidly increases in a short period of time, even if it does not exceed the fixed attention value, it can be determined as a fault. This method only considers static gas features, ignores the dynamic relationship between transformer state and gas concentration, and is difficult to accurately capture transformer faults. SUMMARY

[0006] Therefore, the present application provides an oil-immersed transformer fault detection method based on gases on the oil surface, which can solve the technical problem that the existing method only considers static gas characteristics and ignores the dynamic relationship between transformer state and gas concentration, making it difficult to accurately capture transformer faults.

[0007] The present application is implemented as follows:

[0008] The first aspect of the present application provides an oil-immersed transformer fault detection method based on gases on the oil surface, comprising the following steps:

[0009] S10, real-time acquisition of historical operating parameters of the oil-immersed transformer;

[0010] S20, collecting gases on the oil surface in the oil-immersed transformer according to a preset sampling interval, and obtaining a gas parameter vector at each sampling time using gas chromatography, which is used to represent the composition ratio of the gas at the sampling time;

[0011] S30, establishing and fitting a transformer operating state equation set according to the operating parameters, including a voltage equation, a current equation, a power equation, and a temperature equation;

[0012] S40, establishing and fitting a gas concentration change equation set according to the gas parameter vector, including a hydrogen change equation, a methane change equation, an ethylene change equation, an acetylene change equation, and a trace gas change equation;

[0013] S50, using a multiple linear regression analysis method and combining principal component analysis to obtain the correlation between the transformer operating state equation set and the gas concentration change equation set, and establishing a gas-operating parameter change matrix based on grey correlation analysis and the correlation;

[0014] S60, establishing a training data set based on the gas-operating parameter change matrix of multiple oil-immersed transformers without faults or with known faults, wherein the input of the training data set is the gas-operating parameter change matrix, and the output is the fault category;

[0015] S70, training a multi-branch parallel neural network model using the training data set to obtain an oil-immersed transformer fault diagnosis model;

[0016] S80, real-time monitoring of the oil-immersed transformer to be detected using the steps of S10-S50 to obtain a gas-operating parameter change matrix of the oil-immersed transformer to be detected, inputting the fault diagnosis model, and outputting a fault vector of the oil-immersed transformer to be detected.

[0017] The operating parameters include primary side voltage, primary side current, secondary side voltage, secondary side current, frequency, and harmonic;

[0018] The fault categories include no fault or overheating fault, discharge fault, and insulation aging fault.

[0019] The fault diagnosis model adopts a multi-branch parallel structure and includes a macro feature subnetwork, a micro feature subnetwork, and a fusion feature subnetwork.

[0020] The macro feature subnetwork is configured to perform time series feature extraction and global feature analysis on the overall state feature change matrix and capture the overall change trend of the transformer state over time.

[0021] The micro feature subnetwork is configured to perform abnormality detection and local feature extraction on elements in the state feature change matrix that do not conform to the Bayesian distribution, perform fine-grained fault feature analysis, and further identify potential local faults.

[0022] The fusion feature subnetwork is configured to fuse macro features and micro features, comprehensively consider global and local information, and finally output a fault diagnosis result.

[0023] Further, the voltage equation is a composite voltage equation and is specifically expressed as:

[0024] U1=U 1n +k1·I1+k2·T+k3·f+k4·THD;

[0025] wherein U1 is the primary side voltage, U 1n is the primary side rated voltage, I1 is the primary side current, T is the temperature, f is the frequency, THD is the total harmonic distortion rate, and k1, k2, k3, and k4 are coefficients.

[0026] The current equation is a composite current equation and is specifically expressed as:

[0027]

[0028] wherein I2 is the secondary side current, U2 is the secondary side voltage, η is the transformer efficiency, and k5, k6, and k7 are coefficients.

[0029] The power equation is a composite power equation and is specifically expressed as:

[0030] P=U1·I1·cosφ+k8·T+k9·f+k 10 ·THD;

[0031] wherein P is the power, cosφ is the power factor, k8, k9, and k 10 are coefficients.

[0032] wherein the temperature equation is specifically expressed as:

[0033]

[0034] wherein T a is the ambient temperature, k 11 is the coefficient of thermal conductivity, k 12 is the coefficient of heat capacity, k 13 is the coefficient of heat convection, k 14 is the coefficient.

[0035] wherein the hydrogen variation equation is specifically expressed as:

[0036]

[0037] the methane variation equation is specifically expressed as:

[0038]

[0039] the ethylene variation equation is specifically expressed as:

[0040]

[0041] the acetylene variation equation is specifically expressed as:

[0042]

[0043] the trace gas variation equation is specifically expressed as:

[0044]

[0045] wherein H2, CH4, C2H4, C2H2, G m represent the concentrations of hydrogen, methane, ethylene, acetylene and trace gas respectively, a i , b i , c i , d i , e i (i = 1, 2, …, 6) are the corresponding coefficients.

[0046] For each gas G j (j = 1, 2, …, 5, representing H2, CH4, C2H4, C2H2, G m ), the following regression equation is established:

[0047] G j = β j0 + β j1 U1+ β j2 I1+ β j3 U2+ β j4 I2+ βj5 f+β j6 THD+β j7 T+

[0048] β j8 P+∈ j

[0049] wherein, β jk (k = 0, 1, …, 8) are regression coefficients, ∈ j is an error term.

[0050] Principal component analysis:

[0051] The independent variables of the above regression equation are subjected to principal component analysis to obtain principal components PC1, PC2, …, PC m , wherein m ≤ 8.

[0052] wherein, the gas-operation parameter change matrix is specifically represented as:

[0053]

[0054] wherein, wherein element m ij represents the i-th principal component PC i and the j-th gas concentration rate of change gray correlation degree: γ represents a gray correlation degree calculation function.

[0055] wherein, the fault vector contains three elements, respectively representing the probability of no fault or overheating fault, discharge fault, and insulation aging fault.

[0056] wherein, the fault vector is specifically represented as:

[0057] F = [p nf , p d , p i ];

[0058] wherein, p nf represents the probability of no fault or overheating fault, p d represents the probability of discharge fault, and p i represents the probability of insulation aging fault, and satisfies: p nf +p d +p i = 1.

[0059] In the training process of the oil-immersed transformer fault diagnosis model, the input and output of each sub-network are allocated as follows:

[0060] a) Macroscopic feature sub-network:

[0061] Input: complete gas-operation parameter change matrix M

[0062] Output: Macroscopic feature vector F macro

[0063] b) Microscopic feature subnetwork:

[0064] Input: Elements in gas-operation parameter change matrix M that do not conform to Bayesian distribution

[0065] Processing method:

[0066] 1. Perform Bayesian distribution test on each element m in matrix M ij

[0067] 2. Extract elements that do not conform to Bayesian distribution to form a new matrix M abnormal

[0068] 3. Perform local feature extraction on M abnormal

[0069] Output: Microscopic feature vector F micro

[0070] c) Fusion feature subnetwork:

[0071] Input: Macroscopic feature vector F macro and microscopic feature vector F micro

[0072] Output: Final fault vector F

[0073] This design allows the model to consider both the overall trend and local anomalies of the transformer state, thereby improving the accuracy and reliability of fault diagnosis.

[0074] The step S10 specifically includes:

[0075] Step 101, real-time acquisition of transformer operation parameter data through voltage sensors, current sensors, temperature sensors, frequency sensors and harmonic analyzers installed on oil-immersed transformers;

[0076] Step 102, preliminary processing of the collected data using a data acquisition system, including data cleaning, outlier detection and compensation;

[0077] Step 103, using the standard deviation method to detect outliers, marking data points exceeding three times the standard deviation range of the mean as outliers;

[0078] Step 104, interpolation processing of missing or abnormal data points, using linear interpolation or spline interpolation method;

[0079] Step 105, storing the processed data into a special database; ​​

[0080] Step 106, establish real-time data flow, ensure that the system can continuously obtain the latest operating parameters.

[0081] The step S20 specifically comprises:

[0082] Step 201, according to the operating characteristics and fault development law of the transformer, set the sampling frequency of once an hour, and appropriately increase the sampling frequency when the load changes sharply or abnormality is found;

[0083] Step 202, using a special gas sampling device to collect gas samples from the gas space at the top of the transformer oil tank, the gas sampling device includes a gas-tight sampling bottle, a gas extraction pump and a connecting pipeline;

[0084] Step 203, the collected gas samples are sent into a gas chromatograph for analysis, mainly focusing on the content of hydrogen, methane, ethylene, acetylene and other trace gases;

[0085] Step 204, according to the gas chromatography analysis results, calculate the gas parameter vector at each sampling time, the gas parameter vector contains five elements, respectively representing the volume percentage concentration of five kinds of gas.

[0086] The step S30 specifically comprises:

[0087] Step 301, based on the operating parameter data obtained in step S10, establish a transformer operating state equation set, the equation set includes voltage equation, current equation, power equation and temperature equation;

[0088] Step 302, using the least square method to fit the equation set, determine the corresponding coefficients;

[0089] Step 303, in the fitting process, use regularization technique to avoid overfitting problem;

[0090] Step 304, evaluate the fitting results, use the coefficient of determination and root mean square error as evaluation index;

[0091] Step 305, if the fitting effect is not ideal, recheck the data quality, increase the data amount or adjust the equation form;

[0092] Step 306, save the fitted equation set as the basis for subsequent analysis.

[0093] The step S40 specifically comprises:

[0094] Step 401, based on the gas parameter vector data obtained in step S20, establish a gas concentration change equation set, which includes a hydrogen change equation, a methane change equation, an ethylene change equation, an acetylene change equation, and a trace gas change equation;

[0095] Step 402, use a generalized additive model to fit the equation set, and use a smooth spline function to capture the nonlinear relationship;

[0096] Step 403, automatically select the smoothing parameter through the generalized cross-validation method;

[0097] Step 404, evaluate the fitting results, and use the explained variance ratio and the mean absolute error as evaluation indicators;

[0098] Step 405, if the fitting effect of a certain gas is not ideal, recheck the data of the gas, increase the data volume or adjust the equation form;

[0099] Step 406, save the fitted equation set as the basis for subsequent analysis.

[0100] The step S50 specifically includes:

[0101] Step 501, use a multiple linear regression analysis method to establish a relationship model between the transformer operating state parameters and the gas concentration change rate;

[0102] Step 502, estimate the regression coefficients using the least squares method;

[0103] Step 503, perform principal component analysis on the independent variables, and select the principal components with a cumulative contribution rate of 95% as new features;

[0104] Step 504, calculate the grey correlation degree between the principal components and the gas concentration change rate;

[0105] Step 505, construct a gas operating parameter change matrix, where the elements represent the grey correlation degree between the principal components and the gas concentration change rate.

[0106] The step S60 specifically includes:

[0107] Step 601, collect historical operating data and fault records of multiple oil-immersed transformers, including transformers without faults and transformers with known different types of faults;

[0108] Step 602, label the fault state of each transformer, which is divided into three categories: no fault or overheating fault, discharge fault, and insulation aging fault;

[0109] Step 603, combine the gas operating parameter change matrix of all transformers and the corresponding fault labels into a training data set;

[0110] Step 604, using oversampling technology to expand the minority class samples, balancing the number of samples of different classes;

[0111] Step 605, divide the data set into training set, validation set and test set, the proportion is set to 6:2:2.

[0112] The step S70 specifically comprises:

[0113] Step 701, design a multi-branch parallel neural network structure, including a macro feature subnetwork, a micro feature subnetwork, and a fusion feature subnetwork;

[0114] Step 702, use a bidirectional long short-term memory network to process the entire gas operating parameter change matrix and capture time series features;

[0115] Step 703, use a local anomaly detection algorithm to identify abnormal elements in the matrix, and perform feature extraction through a two-dimensional convolutional neural network;

[0116] Step 704, splice the outputs of the macro feature subnetwork and the micro feature subnetwork, and perform feature fusion through a fully connected layer;

[0117] Step 705, use an Adam optimizer to train the network, with a learning rate of 0.001 and a batch size of 64;

[0118] Step 706, use cross-entropy as the loss function, and introduce L2 regularization to prevent overfitting;

[0119] Step 707, adopt an early stopping strategy, and stop training when the loss on the validation set does not improve for 10 consecutive epochs;

[0120] Step 708, use the test set to evaluate the model performance, and calculate the accuracy, precision, recall, and F1 score, etc.

[0121] The step S80 specifically comprises:

[0122] Step 801, real-time monitor the oil-immersed transformer to be detected, collect data every hour according to the method of steps S10 to S50, and update the gas operating parameter change matrix;

[0123] Step 802, normalize the obtained gas operating parameter change matrix, using the min-max scaling or standardization method;

[0124] Step 803, input the processed matrix into the trained oil-immersed transformer fault diagnosis model;

[0125] Step 804, make fault judgment according to the output fault vector, set the probability threshold to 0.6;

[0126] Step 805, compare the diagnosis result with historical data to analyze the fault development trend;

[0127] Step 806, if the diagnosis result is consistent for several times in succession and the fault probability shows an upward trend, the alarm level is raised and further inspection or maintenance is suggested;

[0128] Step 807, save the diagnosis result, original data and intermediate calculation result to the database for subsequent analysis and model optimization.

[0129] Optionally, the method further comprises the following steps:

[0130] Step 901, calibrate various sensors and gas sampling devices regularly to ensure the accuracy of data;

[0131] Step 902, set simple rule to detect obvious outliers, mark data points exceeding 3 times of standard deviation of normal range as suspicious data;

[0132] Step 903, use incremental learning method, fine-tune the model using newly collected data every month or every quarter;

[0133] Step 904, if the performance of the model is found to be significantly decreased, retrain the entire model.

[0134] Compared with the prior art, the oil-immersed transformer fault detection method based on oil surface gas provided by the application has the following beneficial effects:

[0135] 1. Real-time monitoring combined with dynamic analysis:

[0136] The application establishes transformer operation state equation set and gas concentration change equation set by acquiring transformer operation parameters and oil surface gas data in real time. This dynamic model not only considers static gas characteristics, but also takes into account real-time changes of transformer operation parameters, which can more accurately reflect the actual operation state of the transformer. This method significantly improves the real-time performance and accuracy of fault detection, so that the system can timely discover potential fault hidden dangers.

[0137] 2. Multi-source information fusion:

[0138] The application uses multiple linear regression analysis and principal component analysis method to establish a gas-operation parameter change matrix. This method realizes comprehensive analysis of transformer electrical parameters and gas characteristics in oil, overcoming the limitation of traditional methods which only focus on single characteristics. Through the fusion of multi-source information, the method can more comprehensively evaluate the health status of the transformer, greatly reducing the probability of misjudgment and missed judgment.

[0139] 3. Intelligent diagnostic model:

[0140] The present application uses a multi-branch parallel neural network model as the core algorithm for fault diagnosis. This structure includes a macro feature subnetwork, a micro feature subnetwork, and a fusion feature subnetwork, which can simultaneously capture the overall trend and local abnormal features of the transformer state. This design greatly improves the model's ability to recognize complex fault patterns, especially when dealing with early fault signs and multiple faults.

[0141] 4. Adaptive learning ability:

[0142] Through training using a large amount of historical data and real-time monitoring data, the diagnostic model of the present application has strong adaptive learning ability. As data continues to accumulate and the model continues to optimize, the system can continuously improve its diagnostic accuracy and reliability, adapting to different types of transformers and various operating environments.

[0143] 5. Fine-grained analysis of anomaly detection:

[0144] The micro feature subnetwork is specifically designed to analyze abnormal elements that do not conform to the Bayesian distribution, making the system capable of capturing subtle abnormal changes. This is of great significance for the identification of early faults and the early warning of potential problems, helping maintenance personnel to take preventive measures before the fault expands.

[0145] In summary, the present application solves the technical problem of existing methods that only consider static gas characteristics, ignoring the dynamic relationship between transformer state and gas concentration, making it difficult to accurately capture transformer faults. BRIEF DESCRIPTION OF DRAWINGS

[0146] Figure 1 A flowchart of the method provided by the present application. DETAILED DESCRIPTION

[0147] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0148] As Figure 1 shown is a flowchart of an oil-immersed transformer fault detection method based on oil surface gas provided by the present application, the method includes the following steps: which includes the following steps:

[0149] S10, real-time acquisition of historical operating parameters of the oil-immersed transformer;

[0150] S20, collecting the gas on the oil surface in the oil-immersed transformer at each sampling time according to a preset sampling interval, and obtaining a gas parameter vector at each sampling time by using gas chromatography, for representing the composition proportion of the gas at the sampling time;

[0151] S30, establishing and fitting a transformer operation state equation set according to the operation parameters, including a voltage equation, a current equation, a power equation and a temperature equation;

[0152] S40, establishing and fitting a gas concentration change equation set according to the gas parameter vector, including a hydrogen change equation, a methane change equation, an ethylene change equation, an acetylene change equation and a trace gas change equation;

[0153] S50, using a multiple linear regression analysis method and combining principal component analysis to obtain a correlation between the transformer operation state equation set and the gas concentration change equation set, and establishing a gas-operation parameter change matrix based on grey correlation degree analysis and the correlation;

[0154] S60, establishing a training data set based on the gas-operation parameter change matrix of multiple oil-immersed transformers without faults and / or with known faults, wherein the input of the training data set is the gas-operation parameter change matrix, and the output is a fault category;

[0155] S70, training a multi-branch parallel neural network model by using the training data set to obtain an oil-immersed transformer fault diagnosis model;

[0156] S80, using the steps of S10-S50 to perform real-time monitoring on a to-be-detected oil-immersed transformer, obtaining a gas-operation parameter change matrix of the to-be-detected oil-immersed transformer, inputting the gas-operation parameter change matrix into the fault diagnosis model, and outputting a fault vector of the to-be-detected oil-immersed transformer.

[0157] The specific implementation of the above steps is described in detail as follows:

[0158] The specific implementation of step S10 is as follows: first, real-time acquisition of the operating parameter data of the transformer through various sensors installed on the oil-immersed transformer. These sensors include but are not limited to voltage sensors, current sensors, temperature sensors, frequency sensors, and harmonic analyzers. Second, the collected data is preliminarily processed through the data acquisition system, including data cleaning, outlier detection, and compensation. Data cleaning mainly removes obviously erroneous data points, such as values exceeding the physically possible range; outlier detection uses the standard deviation method, i.e., data points exceeding the range of plus or minus 3 times the standard deviation of the mean are marked as outliers; data compensation is the interpolation of missing or abnormal data points, common methods include linear interpolation and spline interpolation. Then, the processed data is stored in a special database for subsequent analysis and use. Finally, a real-time data stream is established to ensure that the system can continuously obtain the latest operating parameters. The main purpose of this step is to provide reliable basic data for subsequent analysis and fault diagnosis, ensuring the real-time, integrity, and accuracy of the data.

[0159] The specific implementation of step S20 is as follows: first, according to the operating characteristics and fault development law of the transformer, set an appropriate sampling interval. For large oil-immersed transformers, a sampling frequency of once per hour is usually adopted, but in the case of rapid load changes or abnormality, the sampling frequency can be appropriately increased. Second, use a special gas sampling device to collect gas samples from the gas space at the top of the transformer oil tank. The sampling device usually includes a gas-tight sampling bottle, a gas extraction pump, and connecting pipelines. Attention should be paid to avoid air mixing during sampling to ensure the purity of the sample. Then, the collected gas samples are sent to a gas chromatograph for analysis. Gas chromatography is a highly efficient and accurate gas component analysis method that uses the difference in the distribution coefficient of different gases between the stationary phase and the mobile phase to realize the separation and quantification of gases. In this invention, the focus is on the content of hydrogen (H2), methane (CH4), ethylene (C2H4), acetylene (C2H2), and other trace gases. Finally, according to the gas chromatography analysis results, the gas parameter vector at each sampling time is calculated. The vector usually contains 5 elements, representing the volume percentage concentration of the above 5 gases. The main purpose of this step is to obtain accurate concentration data of the dissolved gases in the transformer oil, which is an important basis for determining whether a fault has occurred inside the transformer and the type of fault.

[0160] The specific implementation of step S30 is as follows: First, based on the operating parameter data obtained in step S10, establish a set of transformer operating state equations. This equation set includes voltage equations, current equations, power equations, and temperature equations. The voltage equation describes the relationship between the primary side voltage and the rated voltage, primary side current, temperature, frequency, and total harmonic distortion; the current equation describes the relationship between the secondary side current and the primary side current, transformation ratio, temperature, frequency, and total harmonic distortion; the power equation describes the relationship between power and primary side voltage, primary side current, power factor, temperature, frequency, and total harmonic distortion; and the temperature equation describes the rate of change of transformer temperature and power, ambient temperature, frequency, and total harmonic distortion. Second, use the least squares method to fit these equations. The least squares method is a commonly used parameter estimation method that determines the optimal parameter value by minimizing the sum of squares between predicted values and actual observations. In this invention, for each equation, the corresponding coefficients (such as k1, k2, etc.) need to be determined. During the fitting process, regularization techniques such as ridge regression or LASSO can be used to avoid overfitting problems. Then, the fitting results are evaluated, usually using the coefficient of determination (R 2 ) and the root mean square error (RMSE) as evaluation indicators. If the fitting effect is not ideal (for example, R 2 <0.8 or RMSE> preset threshold), the data quality needs to be rechecked, and the data volume may need to be increased or the equation form needs to be adjusted. Finally, the fitted equation set is saved as the basis for subsequent analysis. The main purpose of this step is to establish a mathematical model of the transformer operating state, providing a theoretical basis for subsequent fault diagnosis. Through these equations, the mutual relationship between various operating parameters and their influence on the transformer state can be better understood.

[0161] The specific implementation of step S40 is as follows: First, based on the gas parameter vector data obtained in step S20, a system of gas concentration change equations is established. This system of equations includes a hydrogen change equation, a methane change equation, an ethylene change equation, an acetylene change equation, and a trace gas change equation. Each equation describes the relationship between the rate of change of the corresponding gas concentration and temperature, power, primary side voltage, primary side current, frequency, and total harmonic distortion rate. Second, these equations are fitted using a multivariate nonlinear regression method. Considering the possibility of nonlinear relationships in gas concentration changes, a generalized additive model (GAM) is used for fitting. GAM is a flexible nonparametric regression method that allows for nonlinear relationships between independent variables and dependent variables while maintaining the interpretability of the model. During the fitting process, a smoothing spline function is used to capture nonlinear relationships, and the smoothing parameter is automatically selected through the generalized cross-validation (GCV) method. Then, the fitting results are evaluated, and the explained variance ratio and mean absolute error (MAE) are used as evaluation indicators. If the fitting effect of a certain gas is not ideal (for example, the explained variance ratio is <0.7 or the MAE is > a preset threshold), the data of that gas needs to be rechecked, and the data quantity may need to be increased or the equation form may need to be adjusted. Finally, the fitted equation system is saved as the basis for subsequent analysis. The main purpose of this step is to establish a mathematical model of the concentration change of dissolved gases in transformer oil, providing an important basis for subsequent fault diagnosis. Through these equations, the influence of various operating parameters on gas generation can be better understood, and the internal state of the transformer can be more accurately judged.

[0162] The specific implementation of step S50 is as follows: First, a multivariate linear regression analysis method is used to establish a relationship model between transformer operating state parameters and gas concentration change rates. For each gas, a regression equation of the following form is established: where G j represents the concentration of the jth gas, β jk is the regression coefficient, and ∈ j is the error term. The least squares method is used to estimate the regression coefficients. Second, in order to reduce the influence of multicollinearity, principal component analysis (PCA) is performed on the independent variables. PCA can convert the original high-dimensional features into low-dimensional principal components while retaining most of the information. The principal components whose cumulative contribution rate reaches 95% are selected as the new features. Then, the grey correlation degree between the principal components and the gas concentration change rates is calculated. Grey correlation analysis is a multi-factor statistical analysis method that can quantify the correlation between factors in a system. For each principal component PC i and the concentration change rate of each gas G j , the grey correlation degree is calculated. The calculation formula of the grey correlation degree is: where n is the number of samples, p is the resolution coefficient, usually taken as 0.5. Finally, a gas-operation parameter change matrix M is constructed, where the element m ij represents the grey correlation degree between the ith principal component and the jth gas concentration rate of change. The main purpose of this step is to establish a quantitative relationship between the transformer operating state and the gas concentration change, providing more comprehensive feature information for subsequent fault diagnosis. Through multiple regression analysis, the influence of each operating parameter on gas generation can be understood; through principal component analysis, the correlation between features can be reduced, and more essential information can be extracted; through grey correlation analysis, the correlation degree between principal components and gas changes can be quantified, thereby constructing a feature matrix that comprehensively reflects the state of the transformer.

[0163] The specific implementation of step S60 is as follows: First, collect the historical operation data and fault records of multiple oil-immersed transformers. These transformers should include both fault-free operating transformers and transformers with known different types of faults. For each transformer, use the method of steps S10 to S50 to calculate its gas-operation parameter change matrix. Second, label the fault state of each transformer. The labeling is divided into three categories: no fault or overheating fault, discharge fault, and insulation aging fault. When labeling, reference can be made to the maintenance records of the transformer, the results of gas ratio analysis (such as the Duval triangle method), and expert diagnosis opinions. Then, combine the gas-operation parameter change matrix of all transformers and the corresponding fault label into a training data set. Each sample contains a matrix (input) and a fault category (output). In order to balance the number of samples of different categories, oversampling techniques (such as the SMOTE algorithm) can be used to expand the number of samples of minority classes. Finally, divide the data set into a training set, a validation set, and a test set, with a ratio of 6:2:2. The training set is used for model training, the validation set is used for adjusting hyperparameters and preventing overfitting, and the test set is used for evaluating the performance of the final model. The main purpose of this step is to construct a high-quality training data set to provide reliable learning materials for the subsequent neural network model. By collecting data from multiple transformers, the generalization ability of the model can be improved; by expert labeling and gas ratio analysis, the accuracy of the fault category can be improved; by data augmentation and sample balancing, the recognition ability of the model for minority class faults can be improved.

[0164] The specific implementation of step S70 is as follows: first, a multi-branch parallel neural network structure is designed. The network includes three main parts: a macro feature subnetwork, a micro feature subnetwork, and a fusion feature subnetwork. The macro feature subnetwork adopts a long short-term memory (LSTM) structure to capture time series features; the micro feature subnetwork adopts a convolutional neural network (CNN) structure to extract local features; and the fusion feature subnetwork adopts a fully connected layer structure to comprehensively analyze global and local information. Second, for the macro feature subnetwork, a bidirectional LSTM (Bi-LSTM) is used to process the entire gas-operation parameter change matrix. Bi-LSTM can consider both past and future information, better capturing long-term dependencies. The number of LSTM units is set to 128, and the number of layers is set to 2. Then, for the micro feature subnetwork, first, a local anomaly detection algorithm (such as Isolation Forest) is used to identify abnormal elements in the matrix. These abnormal elements and their neighborhoods are taken as input, and a 2D-CNN is used for feature extraction. The structure of the CNN includes 3 convolutional layers (with a kernel size of 3x3 and channel numbers of 32, 64, and 128, respectively) and 2 max-pooling layers. Next, for the fusion feature subnetwork, the outputs of the macro feature subnetwork and the micro feature subnetwork are spliced, and two fully connected layers (with node numbers of 256 and 128, respectively) are used for feature fusion. Finally, a softmax layer outputs the probability distribution of the three fault types. During network training, the Adam optimizer is used, the learning rate is set to 0.001, and the batch size is set to 64. Cross-entropy is used as the loss function, and L2 regularization (with a coefficient of 0.0001) is introduced to prevent overfitting. The early stopping strategy is adopted, and when the loss on the validation set does not improve for 10 consecutive epochs, training is stopped. Finally, the test set is used to evaluate the model performance, and indicators such as accuracy, precision, recall, and F1 score are calculated. If the performance does not meet the requirements (such as accuracy < 90%), the network structure or hyperparameters need to be adjusted and retrained. The main purpose of this step is to build and train a deep learning model that can accurately diagnose oil-immersed transformer faults. Through the multi-branch parallel network structure, the macroscopic trend and local abnormal features of the transformer state can be considered simultaneously; through the use of advanced neural network structures such as LSTM and CNN, time series and spatial features can be effectively extracted; and through the fusion feature subnetwork, information of different scales can be comprehensively analyzed, improving the accuracy and reliability of fault diagnosis.

[0165] The specific implementation of step S80 is as follows: first, the oil-immersed transformer to be detected is monitored in real time, the operating parameters and the oil-dissolved gas data of the transformer are acquired in real time according to the method of steps S10 to S50, and the gas-operating parameter change matrix is calculated. Specifically, data is collected every certain time interval (such as 1 hour) and the change matrix is updated. Second, the obtained gas-operating parameter change matrix is input into the trained oil-immersed transformer fault diagnosis model. Before input, the data needs to be normalized to ensure that the data range is consistent with the training data. The normalization method can use min-max scaling or standardization. Then, the model will output a fault vector F = [p nf ,p d ,p i ], where p nf represents the probability of no fault or overheating fault, p d represents the probability of discharge fault, and p i represents the probability of insulation aging fault.

[0166] Next, fault judgment is performed according to the output fault vector. A probability threshold (such as 0.6) can be set, and if the probability of a certain fault type exceeds this threshold, it is considered that the transformer may have this type of fault. If the probabilities of multiple fault types all exceed the threshold, the fault type with the highest probability is selected as the main fault type. Finally, the diagnosis result is compared with historical data to analyze the fault development trend. If the diagnosis results are consistent for several consecutive times and the fault probability shows an upward trend, the alarm level can be raised and further inspection or maintenance is recommended. At the same time, the diagnosis result, the original data and the intermediate calculation result are saved to the database for subsequent analysis and model optimization. The main purpose of this step is to realize real-time monitoring and diagnosis of oil-immersed transformer faults. Through continuous data collection and analysis, potential problems of the transformer can be found in time; through probability output, the health status of the transformer can be understood more carefully; through historical data comparison and trend analysis, the development direction of the fault can be predicted to provide basis for preventive maintenance.

[0167] During the implementation of the entire method, the following points need to be noted:

[0168] 1. Data quality control: In steps S10 and S20, various sensors and gas sampling devices need to be calibrated regularly to ensure the accuracy of the data. Some simple rules can be set to detect obvious outliers, for example, if the value of a certain parameter exceeds 3 times its normal range standard deviation, it will be marked as suspicious data and needs to be further confirmed.

[0169] 2. Model updating strategy: Over time, the characteristics of the transformer may change, so the model needs to be updated regularly. Incremental learning can be used, and the model can be fine-tuned using newly collected data every certain period of time (such as every month or every quarter). If the model performance is found to be significantly decreased, the entire model needs to be retrained.

[0170] 3. Parameter tuning: In steps S30 and S40, the parameters of the equation fitting (such as k1, k2, etc.) may need to be adjusted according to different transformer types and operating environments. Methods such as grid search or Bayesian optimization can be used to find the optimal parameters.

[0171] 4. Threshold setting: In step S80, the probability threshold for fault judgment can be adjusted according to the actual application scenario. If you want to detect faults more sensitively, you can lower the threshold; if you want to reduce false positives, you can raise the threshold. It is recommended to set the initial threshold to 0.6, and then fine-tune it according to the actual running effect.

[0172] 5. Feature importance analysis: After the model is trained, feature importance analysis methods (such as permutation importance or SHAP values) can be used to understand which features play a key role in fault diagnosis. This can help us better understand the fault mechanism and optimize the data collection strategy if necessary.

[0173] Specifically, the principles of the present application are:

[0174] 1. Multi-source data fusion principle: The running state of a transformer is a complex system, and a single data source cannot fully reflect its state. This method considers both electrical parameters (such as voltage, current, power) and dissolved gas data in oil, which reflect different aspects of the transformer: electrical parameters directly reflect the load and operating conditions of the transformer, while dissolved gases in oil reflect the thermal, electrical and chemical states inside the transformer. By establishing a set of transformer operating state equations and a set of gas concentration change equations, this method realizes the organic fusion of these two types of data, thereby obtaining a more comprehensive and accurate description of the transformer state.

[0175] 2. Dynamic feature extraction principle: Transformer faults are usually a gradual process, and it is difficult to detect potential faults in a timely manner relying solely on static features. This method correlates the rate of change of gas concentration with transformer operating parameters by establishing a set of gas concentration change equations, thereby capturing the dynamic change process of gas concentration. This dynamic feature not only contains the absolute level of gas concentration, but also reflects its trend, which is beneficial to early detection of abnormal states.

[0176] 3. Multi-scale analysis principle: Transformer faults can manifest as slow changes in overall state or sudden abnormalities in local parameters. This method employs a multi-branch parallel neural network structure, including a macro-feature subnetwork and a micro-feature subnetwork. The macro-feature subnetwork uses a long short-term memory network (LSTM) to process time series data, capturing the overall trend of transformer state changes; the micro-feature subnetwork uses a convolutional neural network (CNN) to process local abnormalities, identifying potential local faults. This multi-scale analysis method can comprehensively capture various abnormal states of the transformer.

[0177] 4. Deep learning model principle: Compared with traditional machine learning methods, deep learning models have stronger feature extraction and nonlinear mapping capabilities. The multi-branch parallel neural network structure used in this method can automatically learn complex feature representations, adapting to different types of fault patterns. At the same time, by fusing feature subnetworks, it realizes the organic combination of macro-features and micro-features, improving the generalization ability of the model.

[0178] 5. Explainability enhancement principle: Deep learning models are often criticized as "black boxes" and lack explainability. This method improves the model's explainability by introducing grey correlation analysis and feature importance analysis. Grey correlation analysis can quantify the degree of correlation between different features, while feature importance analysis can identify features that play a key role in fault diagnosis. These analysis results not only help understand the model's decision-making process, but also provide guidance for further optimizing data acquisition strategies.

[0179] 6. Incremental learning and transfer learning principle: The operating environment and inherent characteristics of transformers may change over time, making it difficult for fixed models to maintain high performance over time. This method uses an incremental learning strategy, regularly updating the model with new data, allowing the model to continuously adapt to new operating conditions. At the same time, through transfer learning technology, the knowledge learned from a large number of transformers is transferred to new transformers, solving the problem of insufficient data for new equipment.

[0180] In summary, the technical solution of the present invention, through multi-source data fusion, dynamic feature extraction, multi-scale analysis and other innovative designs, combined with the powerful capabilities of deep learning, realizes comprehensive, accurate and real-time evaluation of the state of oil-immersed transformers.

[0181] In order to better understand and implement the present invention, the following provides an embodiment of a specific application scenario of the present invention: This embodiment describes the process and results of applying the present invention method to a large oil-immersed transformer in a substation of a provincial power company. The transformer is a 220kV / 110kV three-phase dual-winding load regulation transformer with a rated capacity of 180MVA and a service time of 10 years.

[0182] 1. System deployment

[0183] The following sensors are installed on the transformer:

[0184] Voltage transformer: measures primary and secondary voltage

[0185] Current transformer: measures primary and secondary current

[0186] Temperature sensor: measures top layer oil temperature and winding temperature

[0187] Frequency measurement device

[0188] Harmonic analyzer

[0189] Online monitoring device for dissolved gases in oil: monitors hydrogen, methane, ethylene, acetylene, and carbon monoxide

[0190] At the same time, an edge computing server is installed next to the transformer for data preprocessing and preliminary analysis. The central server is located in the substation control room, responsible for model training and final fault diagnosis.

[0191] 2. Data collection and preprocessing

[0192] The system collects electrical parameter data at a frequency of every 5 minutes and dissolved gas data in oil at a frequency of every hour. Table 1 is an example of part of the data collected in 24 hours (rounded):

[0193] Table 1 Data collection table

[0194]

[0195] The system preprocesses the collected data, including outlier detection and data compensation. For example, the standard deviation method is used to detect that the secondary side current data at 16:00 is abnormal (more than 3 times the standard deviation of the mean), and the system automatically uses linear interpolation method to correct the data point.

[0196] 3. Establish and fit equation sets

[0197] Based on the preprocessed data, the system establishes and fits the transformer operating state equation set and the gas concentration change equation set. The following are some fitting results:

[0198] Transformer operating state equation set:

[0199] Primary side voltage equation: U1 = 219.5 + 0.002·I1 + 0.05·T + 0.1·f - 0.2·THD

[0200] Secondary side current equation: I2 = 2·I1 + 0.5·T + 0.1·f - 1·THD

[0201] Top layer oil temperature equation:

[0202] Gas concentration change equations (using hydrogen as an example):

[0203]

[0204] 4. Feature extraction and correlation analysis

[0205] The system performs principal component analysis on transformer operating state parameters and gas concentration change rates, selecting five principal components with a cumulative contribution rate of 95%. Then, the gray correlation degree between these principal components and each gas concentration change rate is calculated, and a gas-operating parameter change matrix is constructed. Some results are shown in Table 2:

[0206] Table 2 Principal component analysis data table

[0207] Principal component Hydrogen Methane Ethylene Acetylene Carbon monoxide PC1 0.85 0.78 0.92 0.71 0.63 PC2 0.72 0.81 0.75 0.89 0.58 PC3 0.63 0.69 0.81 0.76 0.72 PC4 0.57 0.62 0.68 0.82 0.79 PC5 0.51 0.55 0.59 0.67 0.85

[0208] 5. Model training

[0209] The system uses historical data to train a multi-branch parallel neural network model. The training data includes 3 years of operating data from 100 transformers, totaling approximately 2.6 million samples. Of these, 80% are used for training, 10% for validation, and 10% for testing. The model structure is as follows:

[0210] - Macro-feature subnetwork: 2 layers of Bi-LSTM, 128 units per layer

[0211] - Micro-feature subnetwork: 3 layers of 2D-CNN, convolution kernel size 3x3, channel numbers 32, 64, 128 respectively

[0212] - Fusion feature subnetwork: 2 layers of fully connected layers, node numbers 256 and 128 respectively

[0213] During training, the learning rate is set to 0.001, the batch size is 64, and the Adam optimizer is used. Training lasts for 100 epochs, and finally achieves an accuracy of 95.8% on the test set.

[0214] 6. Real-time monitoring and fault diagnosis

[0215] In actual operation, the system evaluates the transformer state every hour. As shown in Table 3 are some diagnostic results for a certain day:

[0216] Table 3 Diagnosis results table

[0217]

[0218]

[0219] The system issued a warning at 20:00, indicating that the transformer may have an overheat fault.

[0220] 7. Fault analysis and handling

[0221] Upon receiving the warning, the operation and maintenance personnel immediately conducted a detailed inspection of the transformer. Through analyzing the output of the model and the feature importance, it was found that the main factors leading to the rise in the probability of overheat fault were the abnormal rise in the top oil temperature and the rapid increase in the ethylene concentration.

[0222] Further inspection found that one of the transformer's cooling fans had failed, causing a decrease in cooling efficiency. At the same time, the load was high that day, exacerbating the transformer's heating.

[0223] Based on this discovery, the operation and maintenance personnel took the following measures:

[0224] 1. Immediately repaired the malfunctioning cooling fan

[0225] 2. Temporarily reduced the load on the transformer

[0226] 3. Increased the monitoring frequency of the transformer, conducting a status assessment every 30 minutes

[0227] After taking these measures, the status of the transformer gradually returned to normal. The diagnosis results at 4:00 the next morning were as follows:

[0228] No fault probability: 0.94

[0229] Overheat fault probability: 0.04

[0230] Discharge fault probability: 0.01

[0231] Insulation aging fault probability: 0.01

[0232] Diagnosis result: Normal

[0233] 8. Long-term effect evaluation

[0234] After applying the method of the present invention for one year, the substation evaluated the performance of the system. The results showed:

[0235] 1. Fault detection accuracy: increased from 85% to 97%

[0236] 2. Fault warning time: an average of 36 hours in advance, 24 hours earlier than before

[0237] 3. False positive rate: decreased from 10% to 3%

[0238] 4. False negative rate: decreased from 8% to 1%

[0239] 5. Transformer availability: from 99.2% to 99.8%

[0240] 6. Maintenance costs: reduced by 15%

[0241] In addition, the adaptive ability of the system is also verified. When a new 110 kV transformer is introduced, through transfer learning, the system achieves a diagnostic accuracy of more than 90% using only one month of data.

[0242] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An oil-immersed transformer fault detection method based on gas on oil surface, characterized by, The method comprises the following steps: S10, real-time acquisition of historical operating parameters of the oil-immersed transformer; S20, according to a preset sampling interval, collecting the gas on the oil surface in the oil-immersed transformer at each sampling time, and obtaining a gas parameter vector at each sampling time by using gas chromatography, for representing the composition proportion of the gas at the sampling time; S30, according to the operating parameters, establishing and fitting a transformer operating state equation set, including a voltage equation, a current equation, a power equation and a temperature equation; S40, according to the gas parameter vector, establishing and fitting a gas concentration change equation set, including a hydrogen change equation, a methane change equation, an ethylene change equation, an acetylene change equation and a trace gas change equation; S50, using a multiple linear regression analysis method and combining principal component analysis, obtaining a correlation between the transformer operating state equation set and the gas concentration change equation set, and establishing a gas-operating parameter change matrix based on grey correlation degree analysis and the correlation; S60, based on the gas-operating parameter change matrix of multiple oil-immersed transformers without faults and / or with known faults, establishing a training data set, the input of the training data set being the gas-operating parameter change matrix, and the output being a fault category; S70, training a multi-branch parallel neural network model by using the training data set, to obtain an oil-immersed transformer fault diagnosis model; S80, using the steps of S10-S50 to perform real-time monitoring on a to-be-detected oil-immersed transformer, obtaining a gas-operating parameter change matrix of the to-be-detected oil-immersed transformer, inputting the fault diagnosis model, and outputting a fault vector of the to-be-detected oil-immersed transformer.

2. The oil-immersed transformer fault detection method based on gas above oil according to claim 1, characterized in that, The operating parameters include primary side voltage, primary side current, secondary side voltage, secondary side current, frequency and harmonics; The fault categories include no fault or overheating fault, discharge fault and insulation aging fault; The fault diagnosis model adopts a multi-branch parallel structure, including a macro feature subnetwork, a micro feature subnetwork and a fusion feature subnetwork; The macro feature subnetwork is used for time sequence feature extraction and global feature analysis of the whole state feature change matrix, to capture the overall change trend of the transformer state over time; The micro feature subnetwork is used for abnormal detection and local feature extraction of elements in the state feature change matrix that do not conform to the Bayesian distribution, to perform fine-grained fault feature analysis and further identify potential local faults; The fusion feature subnetwork is used for fusing macro features and micro features, including global and local information, and finally outputting a fault diagnosis result.

3. The oil-immersed transformer fault detection method based on gas above oil according to claim 2, characterized in that, The voltage equation is a complex voltage equation, and is specifically represented as: U1 = U 1n + k1 - I1 + k2 - T + k3 - f + k4 - THD; wherein U1 is a primary-side voltage, U 1n is a primary-side rated voltage, I1 is a primary-side current, T is a temperature, f is a frequency, THD is a total harmonic distortion, and k1, k2, k3, and k4 are coefficients.

4. The oil-immersed transformer fault detection method based on gas above oil according to claim 3, characterized in that, The current equation is a complex current equation, and is specifically represented as: wherein I2 is the secondary side current, U2 is the secondary side voltage, η is the transformer efficiency, and k5, k6 and k7 are coefficients.

5. The oil immersed transformer fault detection method based on gases above oil surface as claimed in claim 4, wherein, The power equation is a complex power equation, and is specifically represented as: P = U1* I1* cos φ + k8* T + k9* f + k 10 THD; where P is the power, cosφ is the power factor, k8, k9, k 10 are coefficients.

6. The oil immersed transformer fault detection method based on the gas above the oil surface according to claim 5, characterized in that, The temperature equation is specifically represented as: where T a is the ambient temperature, k 11 is the thermal conductivity of the material, k 12 is the thermal conductivity of the air, k 13 is the thermal conductivity of the water, and k 14 is a coefficient.

7. The oil-immersed transformer fault detection method based on gas above oil according to claim 6, characterized in that, The hydrogen change equation is specifically represented as: The methane change equation is specifically represented as: The ethylene change equation is specifically represented as: The acetylene change equation is specifically represented as: The micro-gas change equation is specifically expressed as: where H2, CH4, C2H4, C2H2, G m represent the concentrations of hydrogen, methane, ethylene, acetylene and trace gases, respectively, a i ,b i ,c i ,d i ,e i are the respective coefficients.

8. The oil-immersed transformer fault detection method based on gas above oil according to claim 7, characterized in that, The gas-operation parameter change matrix is specifically expressed as: wherein, wherein element m ij denotes the i-th principal component PC i The grey correlation degree of the concentration change rate of the j-th gas: γ denotes a grey correlation degree calculation function.

9. The oil immersed transformer fault detection method based on gases above oil surface as claimed in claim 8, wherein, The fault vector contains three elements, respectively representing the probabilities of no fault or overheat fault, discharge fault and insulation aging fault.

10. The oil-immersed transformer fault detection method based on gas above oil according to claim 9, characterized in that, The fault vector is specifically expressed as: F = [p nf ,p d ,p i ] ; where p nf represents the probability of a fault-free or overheat fault, p d represents the probability of a discharge fault, p i represents the probability of an insulation aging fault, and satisfies: p nf + p d + p i = 1.

Images