A dynamic calibration method and system for multi-physical field data fusion of an oil-immersed transformer

Abstract

The application relates to the technical field of power systems and is a dynamic calibration method and system for oil-immersed transformer multi-physical field data fusion, which comprises a data calibration unit, adopts a cross-physical field coupling calibration model to dynamically calibrate the multi-physical field data of the transformer, and evaluates the calibration effect based on a target function; wherein the multi-physical field data comprises temperature data, vibration data, UHF signal data and gas data; the cross-physical field coupling calibration model comprises a temperature-vibration compensation equation and a gas-UHF triggering mechanism; the application fuses data from multiple sensors, dynamically calibrates through the cross-physical field coupling calibration model, generates an adversarial sample through an edge GAN, triggers cross-physical field coupling calibration model updating in combination with confidence, and intelligently analyzes comprehensive characteristic parameters through a hierarchical fusion architecture based on Shapley values, so that the health state of the transformer can be accurately judged, and the accuracy of fault diagnosis is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, and particularly relates to a dynamic calibration method and system for multi-physical field data fusion of oil-immersed transformers. BACKGROUND

[0002] As important equipment in power systems, oil-immersed transformers are widely used in power transmission and distribution networks, especially in high-voltage and large-capacity power conversion scenarios. Their main function is to dissipate heat, protect and insulate through oil-immersed insulation, so monitoring and maintaining their operating state is crucial. However, due to their complex working environment, long-term operation and susceptibility to external factors, oil-immersed transformers may experience various faults such as overheating, partial discharge, mechanical damage, etc. If not detected and addressed in a timely manner, it may lead to equipment failure or even major safety incidents.

[0003] Traditional transformer fault monitoring and diagnosis methods mainly rely on single detection means such as temperature monitoring and gas monitoring. While these methods can monitor equipment operating conditions to some extent, they lack sufficient comprehensiveness and cannot accurately determine the health status of the equipment in a timely manner. They are also susceptible to external interference, leading to misdiagnosis or missed diagnosis. Therefore, how to comprehensively analyze multiple physical field data to achieve intelligent monitoring, health assessment and fault warning of oil-immersed transformers has become an important problem facing the power industry.

[0004] Currently, with the rapid development of smart grids, the Internet of Things, artificial intelligence and other technologies, transformer health monitoring methods are gradually developing towards multi-dimensional data fusion and intelligent diagnosis. In order to improve the operational safety and maintenance efficiency of transformers, there is an urgent need for a transformer fault monitoring technology based on dynamic calibration and intelligent data fusion of multi-physical field collaborative perception. SUMMARY

[0005] The present application provides a dynamic calibration method and system for multi-physical field data fusion of oil-immersed transformers, which overcomes the shortcomings of the prior art. It effectively solves the problem that existing transformers can only be diagnosed by single diagnostic methods such as temperature and gas monitoring, which cannot accurately determine the health status of the transformer in a timely manner.

[0006] To solve the above problems, one of the technical solutions of the present application is achieved by the following method: a dynamic calibration method for multi-physical field data fusion of oil-immersed transformers, comprising the following steps:

[0007] The multi-physical field data of the transformer is dynamically calibrated by using a cross-physical field coupling calibration model, and the calibration effect is evaluated based on a target function; wherein the multi-physical field data includes temperature data, vibration data, UHF signal data and gas data; the cross-physical field coupling calibration model includes establishing a temperature-vibration compensation equation and a gas-UHF triggering mechanism;

[0008] Through the multi-physical field data of the transformer, an adversarial sample is generated based on an edge GAN, and the cross-physical field coupling calibration model is updated in combination with confidence triggering;

[0009] The dynamic features of each physical field data after dynamic calibration are extracted, comprehensive feature parameters are generated, intelligent analysis is performed on the comprehensive feature parameters based on a Shapley value hierarchical fusion architecture, key spatiotemporal features are strengthened through a double-channel attention mechanism, a comprehensive diagnosis score is calculated and the fault type is judged.

[0010] The above-mentioned temperature-vibration compensation equation includes the following calculation equation: ,

[0011] In the formula, V is the original vibration characteristic value, the unit is mm / s; V' is the vibration value after temperature compensation, the unit is mm / s; T is the real-time monitored temperature value, the unit is ℃; T0 is the reference temperature; α is the material thermal expansion coefficient, the unit is ; β is the transient response factor, the unit is mm / s·℃; t represents the time variable of temperature change, the unit is s; wherein, α and β are calibrated through finite element thermal coupling simulation, , .

[0012] The above-mentioned gas-UHF triggering mechanism includes,

[0013] Setting a gas threshold value;

[0014] Judging whether the gas data exceeds the threshold value;

[0015] In response to yes, the UHF signal of the transformer is high-frequency sampled to capture the partial discharge pulse signal.

[0016] The above-mentioned through the multi-physical field data of the transformer, an adversarial sample is generated based on an edge GAN, and the cross-physical field coupling calibration model is updated in combination with confidence triggering, including,

[0017] The edge GAN generates an adversarial sample according to the multi-physical field data and outputs it to the cross-physical field coupling calibration model;

[0018] The edge GAN receives new multi-physical field data and adversarial samples in real time, and calculates the confidence of the new multi-physical field data through a discriminator;

[0019] The confidence calculation formula is as follows: ,

[0020] In the formula, is the confidence of new data, the value range is [0, 1]; is the model prediction output value; is the real label value; is the maximum value for normalizing relative error;

[0021] determine whether the confidence of new multi-physical field data is less than 0.85;

[0022] in response to yes, trigger the cross-physical field coupling calibration model update, that is, realize the cross-physical field coupling calibration model update by compressing the cross-physical field coupling calibration model through the knowledge distillation method.

[0023] The above hierarchical fusion architecture based on Shapley value performs intelligent analysis on comprehensive feature parameters, and strengthens key space-time features through a double-channel attention mechanism, including,

[0024] The Shapley value is used to quantify the contribution of each feature, and the calculation method is as follows:

[0025] ,

[0026] In the formula, represents the average marginal contribution value of the i-th feature in all feature combinations; F represents the feature set, which is temperature, vibration, gas and UHF physical field features here; |F| is the number of features in set F; S is a subset of F; |S| is the number of features in set S; feature subset, represents a part of feature combinations taken from F; is the prediction value or score function of the model when using the feature subset S; represents the feature i prediction value of the model when the feature is not added to the subset S; the score item is a weighting coefficient, which represents the probability of occurrence of different feature combinations;

[0027] The double-channel attention mechanism includes:

[0028] The spatial attention module: calculate the sensor position weight ws=softmax(MLP(xyz)),

[0029] In the formula, xyz is a three-dimensional coordinate feature vector of the sensor; is a multi-layer perceptron, which extracts spatial features through nonlinear mapping; ws is the attention weight of each sensor in the spatial dimension, and the larger the value is, the greater the influence of the position at the current time is; : normalization function, used to map the output to the probability distribution interval [0, 1];

[0030] Time attention module: LSTM is used to extract time importance ,

[0031] In the formula, ht is the hidden state of LSTM at time step t, which represents the feature representation at the current time; ct is the cell state of LSTM at time step t, which represents the long-term memory feature; W is a trainable weight matrix for linear transformation of [ht; ct]; [ht; ct] is the splicing vector of hidden state and cell state; wt is the attention weight in the time dimension, and the larger the value, the greater the contribution of the time step to the final prediction of the model; is a sigmoid activation function, which maps the value to the interval [0, 1].

[0032] The above extracts the dynamic features of each physical field data after dynamic calibration, including:

[0033] The sliding window technique is used to calculate the trend change, volatility and change rate of temperature data;

[0034] Frequency domain analysis and time domain statistical quantity extraction are performed on the vibration data;

[0035] Wavelet transform is performed on the UHF signal data to extract partial discharge features;

[0036] The concentration change trend and key gas component ratio of the gas data are calculated.

[0037] The above calculates the comprehensive diagnosis score and judges the fault type, including:

[0038] The comprehensive diagnosis score is calculated in the following way:

[0039] Let each physical field feature be , , , , each single-domain fault indicator function be , , , , and the weight be , , , ; the bias term is b;

[0040] The comprehensive diagnosis score formula is:

[0041] ,

[0042] For fault determination, the following logic function is used:

[0043] ,

[0044] In the formula, P(fault|X) is the probability of failure under the input feature X; D is the comprehensive diagnostic score; is an exponential function, which is used to realize the S-shaped curve of the logistic regression;

[0045] When the temperature and gas weight part in the comprehensive diagnostic score D are significantly high, and , the overheat failure is determined;

[0046] When the vibration and UHF part of the comprehensive diagnostic score exceeds the set threshold, the mechanical failure is determined;

[0047] When the score of the UHF part is significant, and there is a joint effect of the temperature and gas anomaly index, the insulation failure is determined;

[0048] When the comprehensive diagnostic score D is significantly high, that is, , and each single-domain index is abnormal at the same time, the multi-factor coupling failure is determined.

[0049] The second technical scheme of the present application is realized by the following way: an oil-immersed transformer multi-physical field data fusion dynamic calibration system uses an oil-immersed transformer multi-physical field data fusion dynamic calibration method, which comprises:

[0050] The data calibration unit adopts a cross-physical field coupling calibration model to dynamically calibrate the multi-physical field data of the transformer, and evaluates the calibration effect based on a target function; wherein the multi-physical field data includes temperature data, vibration data, UHF signal data and gas data; the cross-physical field coupling calibration model includes establishing a temperature-vibration compensation equation and a gas-UHF triggering mechanism;

[0051] The edge GAN unit generates an adversarial sample based on the edge GAN through the multi-physical field data of the transformer, and combines the confidence to trigger the cross-physical field coupling calibration model update;

[0052] The analysis and diagnosis unit extracts the dynamic characteristics of each physical field data after dynamic calibration, generates comprehensive feature parameters, intelligently analyzes the comprehensive feature parameters based on the hierarchical fusion architecture of the Shapley value, strengthens the key spatiotemporal features through the double-channel attention mechanism, calculates the comprehensive diagnostic score and determines the fault type.

[0053] The above edge GAN unit comprises a sample generation module, a calculation module, a judgment module and a response module;

[0054] The sample generation module generates an adversarial sample output to the cross-physical field coupling calibration model according to the multi-physical field data of the edge GAN;

[0055] The calculation module receives new multi-physical field data and adversarial samples in real time through the discriminator, and calculates the confidence of the new multi-physical field data.

[0056] a judgment module, judging whether the new multi-physical field data confidence is less than 0.85;

[0057] a response module, in response to yes, triggering cross-physical field coupling calibration model updating, that is, realizing cross-physical field coupling calibration model updating by compressing the cross-physical field coupling calibration model by adopting a knowledge distillation method.

[0058] The present application fuses multiple sensor data, dynamically calibrates by adopting a cross-physical field coupling calibration model, generates an adversarial sample by an edge GAN, triggers cross-physical field coupling calibration model updating in combination with confidence, and intelligently analyzes comprehensive feature parameters based on a Shapley value-based hierarchical fusion architecture, so as to accurately judge the health state of a transformer and improve the accuracy of fault diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0059] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

[0060] Figure 1 It is a method flowchart of the embodiment 1 of the present application.

[0061] Figure 2 It is a flowchart of the gas-UHF triggering mechanism in the embodiment 1 of the present application.

[0062] Figure 3 It is a flowchart of model updating based on an edge GAN in the embodiment 1 of the present application.

[0063] Figure 4 It is a system block diagram of a dynamic calibration system of oil-immersed transformer multi-physical field data fusion in the embodiment 2 of the present application.

[0064] Figure 5 It is a block diagram of an edge GAN unit in the embodiment 2 of the present application. DETAILED DESCRIPTION

[0065] The present application is not limited by the following embodiments, and the specific implementation can be determined according to the technical solution of the present application and the actual situation.

[0066] Embodiment 1: as shown in the following table, the embodiment of the present application discloses a dynamic calibration method for oil-immersed transformer multi-physical field data fusion, including the following steps: Figure 1

[0067] S101, dynamically calibrating the multi-physical field data of the transformer by adopting a cross-physical field coupling calibration model, and evaluating the calibration effect based on a target function; wherein the multi-physical field data includes temperature data, vibration data, UHF signal data and gas data; the cross-physical field coupling calibration model includes establishing a temperature-vibration compensation equation and a gas-UHF triggering mechanism;​

[0068] S102, generating an adversarial sample based on the edge GAN through the multi-physical field data of the transformer, and triggering the cross-physical field coupling calibration model update in combination with the confidence;

[0069] S103, extracting the dynamic features of the dynamic calibrated physical field data, generating comprehensive feature parameters, intelligently analyzing the comprehensive feature parameters based on the Shapley value hierarchical fusion architecture, strengthening the key spatiotemporal features through the double-channel attention mechanism, calculating the comprehensive diagnosis score and judging the fault type.

[0070] The above also includes real-time collection of multi-physical field data of the oil-immersed transformer through multiple sensors, including temperature data, vibration data, UHF signal data and gas data; Specifically, the temperature information of the transformer is obtained in real time through the temperature sensor installed on the transformer winding, oil temperature and other parts, the vibration sensor is installed on the transformer shell and other parts to collect the vibration signal during the operation of the transformer, the UHF signal sensor is used to capture the partial discharge signal that may be generated inside the transformer, and the gas sensor is used to detect the composition and concentration of the dissolved gas in the transformer oil, such as hydrogen, methane, ethylene, etc.

[0071] The dynamic calibration can also include:

[0072] Linear regression (for processing linear relationship data), support vector regression (suitable for non-linear data), random forest regression (an integrated model that can handle complex relationships between features), gradient boosting machine (for solving high-dimensional, non-linear problems) or neural network (suitable for highly complex data patterns) are used as candidate calibration algorithms; In practical applications, appropriate algorithms can be selected for preliminary calibration according to the characteristics of the data and the calibration requirements;

[0073] The calibration accuracy is evaluated by mean square error, weighted mean square error or adaptive error measurement; For example, the mean square error of the data before and after calibration is calculated to evaluate the calibration effect;

[0074] The mean square error (MSE) is used to measure the difference between the calibrated data and the true value, and the goal is to minimize the error, the smaller the MSE, the better the calibration effect, and the specific calibration method is as follows:

[0075] ,

[0076] In the formula, represents the actual data, represents the calibrated data, and N is the number of samples;

[0077] The weighted mean square error (WMSE) gives different weights to different sensor data to highlight the contribution of key sensors, and the specific method is as follows:

[0078] ,

[0079] wherein, is the weight of sensor i, represents the actual data, represents the calibrated data, and N is the number of samples;

[0080] The adaptive error metric (AEM) measures the error metric method under specific sensors and environmental changes, with adaptability and dynamic adjustment, in the following specific ways:

[0081] ,

[0082] wherein, is a dynamic adjustment function based on environmental variables, represents the actual data, represents the calibrated data, and N is the number of samples;

[0083] Combine reinforcement learning (optimize calibration strategy through interaction between agent and environment, maximize cumulative reward, optimize model parameters), genetic algorithm (simulate natural selection, select optimal solution through iteration, optimize model hyperparameters) or particle swarm optimization (simulate the process of bird foraging, find the optimal solution in the solution space, optimize the parameters of the calibration algorithm) to dynamically adjust the calibration model parameters; taking reinforcement learning as an example, through continuous trial and error and feedback, the parameters of the calibration model are optimized to improve the calibration accuracy.

[0084] The above temperature-vibration compensation equation includes the following calculation equation: ,

[0085] wherein, V is the original vibration characteristic value, unit: mm / s; V' is the vibration value after temperature compensation, unit: mm / s; T is the real-time monitored temperature value, unit: ℃; T0 is the reference temperature; α is the material thermal expansion coefficient, unit: ; β is the transient response factor, unit: mm / s·℃; t represents the time variable of temperature change, unit: s; wherein, α and β are calibrated through finite element thermal coupling simulation, , .

[0086] As shown in Figure 2 , the above gas-UHF triggering mechanism includes,

[0087] S201, set a gas threshold value;

[0088] S202, determine whether the gas data exceeds the threshold value;

[0089] S203, in response to yes, high-frequency sampling is performed on the UHF signal of the transformer to capture the partial discharge pulse signal.

[0090] S204, in response to no, high-frequency sampling is not performed on the UHF signal of the transformer.

[0091] The UHF high-frequency sampling is activated to capture the partial discharge pulse signal, specifically, the system automatically starts the high-frequency data acquisition module to capture the partial discharge pulse signal in the range of 100MHz-1GHz, and outputs the collected high-frequency signal for extraction of dynamic characteristics in the subsequent process.

[0092] As shown in Figure 3 the above, the multi-physical field data of the transformer is used to generate an adversarial sample based on edge GAN, and a confidence degree is used to trigger the update of the cross-physical field coupling calibration model, including,

[0093] S301, the edge GAN generates an adversarial sample according to the multi-physical field data and outputs it to the cross-physical field coupling calibration model;

[0094] S302, the edge GAN receives new multi-physical field data and an adversarial sample in real time, and calculates the confidence degree of the new multi-physical field data through a discriminator;

[0095] The definition formulas of the generator and the discriminator of the edge GAN are as follows:

[0096] ,

[0097] In the formula, G(·) is the generator model; D(·) is the discriminator model; z is random noise or latent variable, input to the generator; x fake is a virtual sample output by the generator; x is a real multi-physical field data sample; θ g is a set of model parameters of the generator; θ d is a set of model parameters of the discriminator.

[0098] The confidence degree calculation formula is as follows: ,

[0099] In the formula, is the confidence degree of the new data, with a value range of [0, 1]; is the model prediction output value; is the real label value; is the maximum value for normalizing the relative error;

[0100] S303, whether the confidence degree of the new multi-physical field data is less than 0.85 is judged.

[0101] S304, in response to yes, triggering cross-physical field coupling calibration model update, that is, realizing cross-physical field coupling calibration model update by compressing the cross-physical field coupling calibration model through the knowledge distillation method.

[0102] S305, in response to no, not triggering cross-physical field coupling calibration model update.

[0103] The Shapley value-based hierarchical fusion architecture intelligently analyzes the comprehensive feature parameters, and strengthens key space-time features through a double-channel attention mechanism, including,

[0104] The Shapley value is used to quantify the contribution of each feature, and the calculation method is as follows:

[0105] ,

[0106] In the formula, represents the average marginal contribution value of the i-th feature in all feature combinations; F represents the feature set, which is temperature, vibration, gas and UHF physical field features here; |F| is the number of features in set F; S is a subset of F; |S| is the number of features in set S; The feature subset represents a part of the feature combination taken from F; is the prediction value or score function of the model when using the feature subset S; represents the feature i prediction value of the model without joining the subset S; the score item is a weighting coefficient, representing the probability of occurrence of different feature combinations;

[0107] The double-channel attention mechanism includes:

[0108] The spatial attention module: calculate the sensor position weight ws=softmax(MLP(xyz)),

[0109] In the formula, xyz is a three-dimensional coordinate feature vector of the sensor; is a multilayer perceptron that extracts spatial features through nonlinear mapping; ws is the attention weight of each sensor in the spatial dimension, and the larger the value, the greater the influence of the position at the current time; : normalization function, used to map the output to the probability distribution interval [0, 1];

[0110] The time attention module: uses LSTM to extract time importance ,

[0111] In the formula, ht is the hidden state of LSTM at time step t, which represents the feature representation at the current moment; ct is the cell state of LSTM at time step t, which represents the long-term memory feature; W is a trainable weight matrix for linear transformation of [ht; ct]; [ht; ct] is a spliced vector of the hidden state and the cell state; wt is the attention weight in the time dimension, and the greater the value, the greater the contribution of the time step to the final prediction of the model; is a Sigmoid activation function, which maps the value to the interval [0, 1].

[0112] The time attention model is used to extract the importance weight wt in the time sequence, which is fused by the hidden state ht and the cell state ct output by the LSTM network to calculate the importance of each time step to the overall task.

[0113] The spatial attention model is used to calculate the spatial weight ws of different sensor positions, which maps the three-dimensional coordinate information (x, y, z) of the sensor to a weight distribution through a multi-layer perceptron (MLP), indicating the importance of each sensor position in the current analysis task.

[0114] Therefore, the dual-channel attention mechanism determines which sensor position is more critical through the spatial attention module, and identifies which time period is more important through the time attention module, thereby enhancing the contribution of key spatio-temporal features in the final diagnostic model, and strengthening the abnormal signals (such as temperature sudden rise or partial discharge pulse) of certain sensor nodes at a certain time period, improving the accuracy of fault diagnosis.

[0115] Wherein, the new multi-physical field data confidence is less than 0.85; the cross-physical field coupling calibration model is updated, that is, the cross-physical field coupling calibration model is compressed by using the knowledge distillation method to realize the cross-physical field coupling calibration model update, including:

[0116] The following formula is used to dynamically update the cross-physical field coupling calibration model,

[0117] ,

[0118] In the formula, the KL divergence term KL(p teacher || p student) measures the difference between the output distribution of the student model and the teacher model, which is used to ensure that the student model can imitate the feature judgment ability of the teacher model as much as possible;

[0119] CrossEntropy term: measure the difference between the prediction result of the student model and the true label value, which is used to ensure that the student model can correctly classify or predict the actual multi-physical field data state;

[0120] : weight parameter, used to balance the influence of teacher model guidance (KL term) and real label supervision (CrossEntropy term).

[0121] The student model and the teacher model are used, and specifically:

[0122] The teacher model (TeacherModel): a high-precision, complex benchmark model, used for training based on a large amount of historical data in the early stage, and capable of accurately depicting the multi-physical field characteristics and fault modes of oil-immersed transformers.

[0123] In the present application, the prediction result pteacher of the teacher model is the guiding standard for dynamic calibration and adversarial sample generation.

[0124] The student model (StudentModel): a lightweight model, used for fast operation on real-time edge devices. It learns the behavior of the teacher model through knowledge distillation (KnowledgeDistillation), maintaining high reasoning speed and certain accuracy.

[0125] The output probability distribution pstudentp of the student model is consistent with the teacher model through the Loss function.

[0126] The above, extracting the dynamic features of each physical field data after dynamic calibration, includes:

[0127] Using the sliding window technique to calculate the trend change, volatility and change rate of temperature data;

[0128] Frequency domain analysis and time domain statistical quantity extraction are performed on the vibration data;

[0129] Wavelet transform is performed on the UHF signal data to extract partial discharge features;

[0130] The concentration change trend and key gas component ratio of the gas data are calculated.

[0131] Among them, the temperature data extraction process can include: calculating the change trend (linear regression slope), first-order difference (change rate), and volatility (standard deviation);

[0132] The vibration data extraction process can include: combining time domain statistics (mean, variance) with frequency domain (FFT) and wavelet transform to extract local time-frequency features;

[0133] The gas data extraction process can include: extracting the concentration trend and volatility, combined with the change of the gas component ratio;

[0134] The UHF signal data extraction process can include: extracting the time-frequency features and abnormal mutation indicators of the partial discharge signal through wavelet transform.

[0135] The above calculating the comprehensive diagnosis score and determining the fault type comprises:

[0136] The comprehensive diagnosis score is calculated by the following way:

[0137] Let each physical field feature be , , , , each single-domain fault indicator function be , , , , the weight be , , , ; the bias term be b;

[0138] The comprehensive diagnosis score formula is:

[0139] ,

[0140] For fault determination, the following logic function is adopted:

[0141] ,

[0142] In the formula, P(fault|X) is the probability of fault occurrence under the input feature X; D is the comprehensive diagnosis score; is an exponential function, which is used to realize the S-shaped curve of the logistic regression;

[0143] When the temperature and gas weight part of the comprehensive diagnosis score D is significantly high, and , the overheat fault is determined;

[0144] When the comprehensive diagnosis scores of vibration and UHF part exceed the set threshold, the mechanical fault is determined;

[0145] When the score of UHF part is significant, and there is a joint effect of temperature and gas abnormal indicators, the insulation fault is determined;

[0146] When the comprehensive diagnosis score D is significantly high, that is, corresponding to , and each single-domain indicator is abnormal at the same time, the multi-factor coupling fault is determined.

[0147] Wherein, the weight is , , , for measuring the importance of each physical field feature; is the fault indicator function of the temperature feature, which represents the nonlinear mapping or score of , is the fault indicator function of the vibration feature, which represents the nonlinear mapping or score of Nonlinear mapping or scoring The fault index function is a gas characteristic, representing the fault index function for... Nonlinear mapping or scoring The fault index function is a characteristic of UHF, representing the fault index function for UHF. The nonlinear mapping or scoring; bias term b, used to adjust the baseline of the comprehensive diagnostic score.

[0148] The characteristic conditions of overheating failure include:

[0149] Temperature trend exceeds a set threshold (e.g., slope > 0.2);

[0150] Excessive temperature fluctuations (standard deviation exceeds the normal range);

[0151] This is accompanied by an increase in the concentration of flammable gases (such as hydrogen and acetylene) in the gas.

[0152] The characteristic conditions of mechanical failure include:

[0153] The vibration signal spectrum shows abnormal peaks (abnormal frequency components).

[0154] The volatility and rate of change of the vibration data deviate significantly from the historical average.

[0155] When a partial discharge phenomenon is observed in a UHF signal, it is easier to determine whether it is a mechanical fault or poor contact.

[0156] Among them, the characteristic conditions of insulation faults include:

[0157] The UHF signal was extracted using wavelet transform to reveal obvious partial discharge characteristics.

[0158] This may also be accompanied by a slight increase in temperature and changes in the concentration of abnormal gases in the gas (such as acetylene).

[0159] Vibration data may show slight anomalies, but are predominantly UHF.

[0160] The characteristic conditions of multi-factor coupled failures include:

[0161] Multiple physical fields exhibited abnormal indicators, such as continuous temperature rise, abnormal vibration, abnormal gas, and simultaneous partial discharge.

[0162] The fusion diagnostic engine gives a high overall score.

[0163] Specifically: when the transformer winding temperature continues to rise, and the concentration of carbon monoxide, carbon dioxide and other gases in the oil increases, it is judged as an overheating fault; when the vibration spectrum is abnormal and accompanied by a sudden change in the UHF signal, it is determined as a mechanical fault; if an abnormal peak value of a specific frequency appears in the vibration signal, and the UHF signal suddenly increases, it is judged as a mechanical fault; when the UHF signal is significantly abnormal and the gas composition changes, it is determined as a partial discharge fault; if the strength of the UHF signal increases significantly, and the concentration of hydrogen, acetylene and other gases in the oil increases, it is judged as a partial discharge fault; when multiple physical field data are abnormal at the same time, it is determined as a multi-factor coupling fault.

[0164] In summary, the present application fuses multiple sensor data, dynamically calibrates by using a cross-physical field coupling calibration model, generates an adversarial sample by using an edge GAN, triggers the update of the cross-physical field coupling calibration model in combination with the confidence, and intelligently analyzes the comprehensive feature parameters based on a Shapley value-based hierarchical fusion architecture, so as to accurately judge the health state of the transformer and improve the accuracy of fault diagnosis.

[0165] Embodiment 2: as shown in Figure 4 The present application discloses a dynamic calibration system for oil-immersed transformer multi-physical field data fusion, comprising:

[0166] The data calibration unit dynamically calibrates the multi-physical field data of the transformer by using a cross-physical field coupling calibration model, and evaluates the calibration effect based on a target function; wherein the multi-physical field data includes temperature data, vibration data, UHF signal data and gas data; the cross-physical field coupling calibration model includes a temperature-vibration compensation equation and a gas-UHF triggering mechanism;

[0167] The edge GAN unit generates an adversarial sample based on the edge GAN by using the multi-physical field data of the transformer, and triggers the update of the cross-physical field coupling calibration model in combination with the confidence;

[0168] The analysis and diagnosis unit extracts the dynamic features of each physical field data after dynamic calibration, generates comprehensive feature parameters, intelligently analyzes the comprehensive feature parameters based on a Shapley value-based hierarchical fusion architecture, strengthens the key spatio-temporal features by using a double-channel attention mechanism, calculates a comprehensive diagnosis score and judges the fault type, and generates a health evaluation result.

[0169] As shown in Figure 5 The above edge GAN unit includes a sample generation module, a calculation module, a judgment module and a response module;

[0170] The sample generation module generates an adversarial sample output to the cross-physical field coupling calibration model according to the multi-physical field data by using the edge GAN;

[0171] The computing module receives new multi-physical field data and adversarial samples in real time, and calculates the confidence of the new multi-physical field data through the discriminator;

[0172] The judging module judges whether the confidence of the new multi-physical field data is less than 0.85;

[0173] The response module responds to yes, triggers the cross-physical field coupling calibration model update, that is, compresses the cross-physical field coupling calibration model by using the knowledge distillation method, and realizes the cross-physical field coupling calibration model update.

[0174] The above-mentioned oil-immersed transformer multi-physical field data fusion dynamic calibration system further comprises an operation and maintenance management platform, which can comprise: a data visualization interface, which displays the multi-physical field data and the health status in real time, and visually displays the operation data and the health status of the transformer through charts, curves and the like; an early warning pushing module, which sends fault early warning information to operation and maintenance personnel, and timely sends early warning information to the operation and maintenance personnel through short messages, emails and the like when the transformer is abnormal; and an intelligent decision support module, which provides maintenance suggestions and optimization operation and maintenance strategies, and provides corresponding maintenance suggestions and operation and maintenance strategies according to the health status and the fault type of the transformer, thereby improving the operation and maintenance efficiency.

[0175] Through the collection, dynamic calibration, feature extraction and fusion analysis of the multi-physical field data, intelligent diagnosis and health status evaluation of the oil-immersed transformer fault are realized. At the same time, the operation and maintenance personnel are provided with intuitive data display and decision support through the operation and maintenance management platform, thereby improving the operation reliability and the operation and maintenance efficiency of the transformer.

[0176] Embodiment 3: A storage medium, wherein the storage medium stores a computer program readable by a computer, and the computer program is arranged to execute an oil-immersed transformer multi-physical field data fusion dynamic calibration method when running.

[0177] The above-mentioned storage medium can include but is not limited to: a U disk, a read-only memory, a mobile hard disk, a magnetic disk or an optical disk and various storage media that can store computer programs.

[0178] Embodiment 4: The embodiment of the present application discloses an electronic device comprising a processor and a memory, wherein the memory stores a computer program, and the computer program is loaded and executed by the processor to realize an oil-immersed transformer multi-physical field data fusion dynamic calibration method.

[0179] The above-mentioned electronic device further comprises a transmission device and an input and output device, wherein the transmission device and the input and output device are connected with the processor.

[0180] The above-described processor can be a Central Processing Unit (CPU), a general purpose processor, a Digital Signal Processor (DSP), an ASIC, a FPGA or other programmable logic device, transistor logic device, hardware component or any combination thereof. It can implement or execute various example logical blocks, modules and circuits described in connection with the disclosure. It can also be a combination of computing functions, such as a combination of one or more microprocessors, a combination of DSP and microprocessor, etc. The memory can include, but is not limited to, a U disk, a read-only memory, a mobile hard disk, a magnetic disk or an optical disk, and various computer program storage media that can store computer programs.

[0181] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present application can be implemented in various computer languages, such as object-oriented programming language Java and interpreted scripting language JavaScript.

[0182] The present application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems) and computer program products of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the computer or other programmable data processing device produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more flows and / or blocks.

[0183] These computer program instructions can also be stored in a computer-readable memory that can guide the computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more flows and / or blocks.

[0184] The application also provides a computer program product, which comprises a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to execute part or all of the steps of any method described in the above method embodiments. The computer program product can be a software installation package, and the computer comprises an electronic device.

[0185] Embodiment 5: The embodiment of the application discloses a terminal, comprising a processor, a memory, a communication interface, and one or more programs stored in the memory and configured to be executed by the processor, the program comprising instructions for performing steps in the dynamic calibration method of multi-physical field data fusion of oil-immersed transformers.

[0186] The above technical features constitute embodiments of the application, which have strong adaptability and optimal implementation effects. Non-essential technical features can be added or reduced according to actual needs to meet the needs of different situations.

Claims

1. A dynamic calibration method for oil-immersed transformer multi-physics field data fusion, characterized in that, The method comprises the following steps: The multi-physical field data of the transformer is dynamically calibrated by using a cross-physical field coupling calibration model, and the calibration effect is evaluated based on a target function; wherein the multi-physical field data comprises temperature data, vibration data, UHF signal data and gas data; the cross-physical field coupling calibration model comprises establishing a temperature-vibration compensation equation and a gas-UHF triggering mechanism; Based on the edge GAN, an adversarial sample is generated from the multi-physical field data of the transformer, and the cross-physical field coupling calibration model is updated in combination with the confidence; Dynamic features of each physical field data after dynamic calibration are extracted, comprehensive feature parameters are generated, intelligent analysis is performed on the comprehensive feature parameters based on a Shapley value hierarchical fusion architecture, key space-time features are strengthened through a double-channel attention mechanism, a comprehensive diagnosis score is calculated, and a fault type is determined; The temperature-vibration compensation equation is established, comprising: ， In the formula, V is the original vibration characteristic value, unit: mm / s; V' is the vibration value after temperature compensation, unit: mm / s; T is the real-time monitored temperature value, unit: ℃; T0 is the reference temperature; is the material thermal expansion coefficient, unit: . is the transient response factor, unit: mm / s·℃; t represents the time variable of temperature change, unit: s; wherein, and calibrated by finite element thermal coupling simulation, ; The gas-UHF triggering mechanism comprises, A gas threshold is set; It is judged whether the gas data exceeds the threshold; In response to yes, the UHF signal of the transformer is high-frequency sampled to capture the partial discharge pulse signal; The edge GAN generates an adversarial sample from the multi-physical field data and outputs it to the cross-physical field coupling calibration model; The edge GAN receives new multi-physical field data and adversarial samples in real time, and calculates the confidence of the new multi-physical field data through a discriminator; The confidence calculation formula is as follows: It is judged whether the confidence of the new multi-physical field data is less than 0.85; ， In the formula, is the confidence of new data, and the value range is [0, 1]; y pred is the model prediction output value; y true is the real label value; is the maximum value for normalizing the relative error; In response to yes, the cross-physical field coupling calibration model is updated, that is, the cross-physical field coupling calibration model is compressed by using a knowledge distillation method to realize the update of the cross-physical field coupling calibration model. The Shapley value hierarchical fusion architecture performs intelligent analysis on the comprehensive feature parameters, and strengthens key space-time features through a double-channel attention mechanism, comprising:

2. The dynamic calibration method of oil-immersed transformer multi-physical field data fusion according to claim 1, characterized in that, The Shapley value is used to quantify the contribution of each feature, and the calculation method is as follows: The double-channel attention mechanism comprises: ， wherein, represents the average marginal contribution value of the i-th feature in all feature combinations; F represents the full set of features, which are temperature, vibration, gas, and UHF physical field features; |F| is the number of features in the set F; S is a subset of F; |S| is the number of features in the set S; a subset of features, represents a partial feature combination taken from F; is the prediction value or score function of the model when using the feature subset S; represents the feature i is the prediction value of the model when the feature is not added to the subset S; score item is a weighting coefficient, representing the probability of occurrence of different feature combinations; The spatial attention module: calculate the sensor position weight ws=softmax(MLP(xyz)), wherein, xyz is a three-dimensional coordinate feature vector of the sensor; MLP(·) is a multi-layer perceptron, which extracts spatial features through nonlinear mapping; ws is the attention weight of each sensor in the spatial dimension, and the larger the value is, the greater the influence of the position at the current time is; softmax(·): normalization function, used to map the output to the probability distribution interval [0, 1]; The dynamic features of each physical field data after dynamic calibration comprise: Temporal attention module: LSTM is used to extract temporal importance , In the formula, ht is the hidden state of the LSTM at time step t, which represents the feature representation at the current moment; ct is the cell state of the LSTM at time step t, which represents the long-term memory feature; W is a trainable weight matrix for linear transformation of [ht; ct]; [ht; ct] is a spliced vector of the hidden state and the cell state; wt is the attention weight in the time dimension, and the greater the value, the greater the contribution of the time step to the final prediction of the model; is a Sigmoid activation function, which maps the value to the interval [0, 1].

3. The dynamic calibration method of oil-immersed transformer multi-physical field data fusion according to claim 1, characterized in that, The trend change, volatility and change rate of the temperature data are calculated using the sliding window technology; The vibration data is analyzed in frequency domain and the time domain statistics are extracted; The UHF signal data is wavelet transformed to extract the partial discharge feature; The concentration change trend and key gas component ratio of the gas data are calculated. The comprehensive diagnosis score is calculated by the following method:

4. The dynamic calibration method of oil-immersed transformer multi-physical field data fusion according to claim 1, characterized in that, The comprehensive diagnosis score formula is: ​ Let each physical field feature be , each single-domain fault indicator function be , the weight be ; the bias term be b; ​ ， For fault determination, the following logic function is adopted: ， wherein is the probability of failure under input feature X; D is the integrated diagnostic score; is the exponential function used to implement the S-shaped curve of the logistic regression; When the temperature and gas weight part in the comprehensive diagnosis score D is significantly high, and a superheat fault is determined. When the comprehensive diagnostic score of vibration and UHF part exceeds the set threshold, it is determined as mechanical failure; When the score of UHF part is significant, and there is a joint effect of temperature and gas anomaly index, it is determined as insulation failure; When the comprehensive diagnosis score D is significantly high, i.e. corresponding to , and each single-domain index is abnormal at the same time, it is determined as a multi-factor coupling fault.

5. A dynamic calibration system for multi-physical field data fusion of oil-immersed transformers, characterized in that, The dynamic calibration method of multi-physical field data fusion of an oil-immersed transformer according to any one of claims 1 to 4 comprises: A data calibration unit adopts a cross-physical field coupling calibration model to dynamically calibrate the multi-physical field data of the transformer, and evaluates the calibration effect based on a target function; wherein the multi-physical field data includes temperature data, vibration data, UHF signal data and gas data; the cross-physical field coupling calibration model includes establishing a temperature-vibration compensation equation and a gas-UHF triggering mechanism; An edge GAN unit generates an adversarial sample based on edge GAN through the multi-physical field data of the transformer, and combines the confidence to trigger the cross-physical field coupling calibration model update; An analysis and diagnosis unit extracts the dynamic characteristics of each physical field data after dynamic calibration, generates comprehensive feature parameters, and performs intelligent analysis on the comprehensive feature parameters based on a Shapley value hierarchical fusion architecture, strengthens key spatio-temporal features through a double-channel attention mechanism, calculates a comprehensive diagnostic score and determines a fault type.

6. The dynamic calibration system of a multi-physical field data fusion of an oil-immersed transformer according to claim 5, characterized in that, The edge GAN unit comprises a sample generation module, a calculation module, a judgment module and a response module; The sample generation module generates an adversarial sample output to the cross-physical field coupling calibration model according to the multi-physical field data of the edge GAN; The calculation module receives new multi-physical field data and adversarial samples in real time, and calculates the confidence of the new multi-physical field data through the discriminator; The judgment module judges whether the confidence of the new multi-physical field data is less than 0.85; The response module responds, and then triggers the cross-physical field coupling calibration model update, that is, the cross-physical field coupling calibration model is compressed by using a knowledge distillation method to realize the cross-physical field coupling calibration model update.

Images