Main transformer fault prediction method based on multivariable data fusion

Abstract

The present invention relates to a transformer fault prediction method, and specifically to a main transformer fault prediction method based on multivariable data fusion. First, multi-source data fusion technology, including Kalman filtering or Bayesian network, is used to integrate various data sources including temperature, load current, dissolved gas analysis data in oil and insulation material conditions collected from the main transformer to obtain a state description, and information gain evaluation is performed to optimize the data fusion strategy; then, advanced machine learning algorithms for processing nonlinear and multidimensional data, including random forests, support vector machines or deep learning networks, are selected to build a fault prediction model, and historical fault data and normal operation data are used to train and verify the model; finally, a real-time data processing system is introduced to continuously receive and process transformer monitoring data, and an early warning system is established. When the fault probability output by the prediction model exceeds the set early warning threshold, the early warning mechanism is triggered to trigger a timely early warning of the fault.

Description

Technical Field

[0001] The invention relates to a transformer fault prediction method, and in particular to a main transformer fault prediction method based on multivariable data fusion. Background Art

[0002] In the existing field of distribution transformer fault prediction, there have been many studies and technologies trying to solve the problem of transformer failure, such as a distribution transformer fault prediction method proposed in Chinese invention patent application number CN202410079928.7. This method uses the grey model and Markov chain principle combined with historical distribution transformer fault outage data for analysis and prediction. Although this method is innovative and effective in some aspects, it still has some shortcomings and disadvantages in the face of wider application and technical needs. First of all, this method mainly relies on historical fault data to build a prediction model. This dependence may lead to a decrease in prediction accuracy and reliability when the data quality and data volume are insufficient. The quality and integrity of historical data have a great impact on the output of the model. If there are problems such as record errors, missing or incomplete data in the historical data, it will directly affect the accuracy of the grey model.

[0003] In addition, historical data may not fully capture all the dynamic changes in transformer operation, especially in today's rapidly changing transformer technology and operating environment, relying solely on historical data may be difficult to adapt to new fault modes and conditions. Secondly, although the gray model mentioned in this method is suitable for system analysis with a small amount of data, it has significant limitations when dealing with complex multivariable systems. The gray model is based on a certain degree of simplification and assumptions of the data, which may ignore the important interactions and nonlinear relationships between the data in some cases, thus affecting the accuracy and generalization of fault prediction. For example, transformer failures are often the result of multiple factors and conditions, and a single model may be difficult to fully express these complex relationships. Furthermore, although this method uses a Markov chain to adjust the initial fault prediction results in order to optimize the accuracy of the prediction results, the application of the Markov chain model is also insufficient. The Markov chain assumes that the future state depends only on the current state (Markov property), which may not always hold true in practical applications. In the actual scenario of transformer fault prediction, the states of multiple historical moments may jointly affect the occurrence of future faults. This neglect of "memory" may reduce the comprehensiveness and accuracy of the model prediction. Summary of the invention

[0004] The purpose of the present invention is to provide a main transformer fault prediction method based on multivariable data fusion, so as to solve some of the drawbacks and shortcomings pointed out in the background technology.

[0005] The present invention solves the above-mentioned technical problems by adopting the following technical solutions: A main transformer fault prediction method based on multivariable data fusion, comprising: firstly, using multi-source data fusion technology, including Kalman filtering or Bayesian network, to integrate various data sources including temperature, load current, oil dissolved gas analysis data and insulation material conditions collected from the main transformer to obtain a state description, and to perform information gain evaluation to optimize the data fusion strategy;

[0006] The continuous time series data collected from the main transformer is processed using the Kalman filter to control the continuous estimation of the real-time state of the transformer, and then the classified or non-continuous data is processed by the Bayesian network to estimate the fault probability under different states using its probabilistic reasoning ability;

[0007] By analyzing the information gain of each data source, that is, the contribution of the collected data to the fault prediction model, key data indicators are identified, and the data fusion model is adjusted according to the information gain results, the model structure and input characteristics are optimized, and the output of the control model reflects the actual operating status of the transformer;

[0008] Then, we select advanced machine learning algorithms that process nonlinear and multidimensional data, including random forests, support vector machines, or deep learning networks, to build a fault prediction model, and use historical fault data and normal operation data to train and verify the model.

[0009] Finally, a real-time data processing system is introduced to continuously receive and process the monitoring data of the transformer, and an early warning system is established. When the failure probability output by the prediction model exceeds the set warning threshold, the early warning mechanism is triggered to trigger a timely warning of the failure.

[0010] Furthermore, the Kalman filter technology is used through the formula:

[0011] X t+1 =X t +α[(Y t -X t )+βΔt]+γ∈

[0012] Where X t is the state estimate at time t, Y t is the observation value at the same time, Δt is the time step, α and β are the coefficients for adjusting the process dynamics, γ is the noise adjustment coefficient, ∈ represents the noise in the process to control the continuous estimation of the real-time state of the transformer;

[0013] The Bayesian network model is used to process categorical or non-continuous data, including dissolved gas analysis DGA and insulation material status, and the formula is:

[0014]

[0015] Where, F represents the predicted fault type; E i It is evidence indicating the status of a certain aspect of the transformer, including DGA test results; P(F|E 1 ,E 2 ,…,E n ) The function formula provides a failure probability assessment by considering all predicted failure types;

[0016] Finally, a comprehensive evaluation formula is used to combine the outputs of the above two models to comprehensively evaluate the health status of the transformer. The formula is:

[0017]

[0018] Among them, α and β are weight parameters, which are used to adjust the influence of continuous state estimation and fault probability to control the comprehensive judgment of transformer state and fault warning.

[0019] Furthermore, the method for identifying key data indicators involves the following steps:

[0020] Firstly, the key data indicators in the main transformer operation data, including temperature, current, oil quality and insulation status, are determined through information gain analysis. Decision tree analysis and mutual information calculation are used to quantify the contribution of each data to the prediction of different types of transformer faults.

[0021] Then, the data index that has the greatest impact on the prediction of fault types is identified based on the obtained information gain results, and the data fusion model structure is adjusted and optimized based on the data index;

[0022] Machine learning algorithms including enhanced random forests or neural networks are used to handle nonlinear relationships of data indicators, and input features of the model are optimized through feature engineering techniques including principal component analysis or encoders.

[0023] Furthermore, the information gain analysis process includes the following steps:

[0024] First, collect and preprocess the key operating data of the main transformer, including temperature, current, oil quality analysis and insulation status, using the formula:

[0025] P prep (D) = ∫exp(-|D-μ| 2 / σ 2 )dD

[0026] To optimize the data preprocessing process, where D represents the original data set, μ and σ are the mean and standard deviation of the data respectively;

[0027] Then apply the mutual information calculation formula:

[0028]

[0029] To evaluate the mutual information between different data types and fault occurrence, so as to quantitatively determine the statistical importance of each data in predicting faults;

[0030] Then analyze the key data indicators through the decision tree, using the formula:

[0031]

[0032] To map input data to output decisions, where c(x) represents the classification of data point x in the tree, N is the number of leaf nodes in the tree, and v n is the predicted value of the nth leaf node, δ is the Kronecker function, which means that if x falls on leaf node n, the value is 1, otherwise it is 0.

[0033] Furthermore, the information gain analysis adopts a nonlinear integral formula:

[0034]

[0035] The amount of information between various data types including temperature, current, oil quality, insulation status and the occurrence of transformer faults is calculated, where I(X;Y) represents information gain, which measures the mutual information between variables X and Y; X, Y represent random variables of the main transformer operation data, including X may be a temperature reading and Y is a fault type; p(x,y) represents the joint probability density function, which represents the probability that variables X and Y take a certain value at the same time; p(x) and p(y) represent marginal probability density functions, which represent the probability of X and Y taking any value respectively; Represents the log-likelihood ratio, which is used to quantify the degree of dependence between X and Y; dx,dy represents the integral symbol, indicating that all possible X and Y values ​​are integrated to calculate the overall information gain;

[0036] Then optimize the data fusion model structure based on the information gain results and apply the advanced adjustment formula:

[0037]

[0038] Among them, M new and M old Represents the adjusted and original model structures respectively; α(x) represents the dynamic learning rate function, which depends on the input data x and is used to adjust the rate of model parameter update; represents the loss function, which evaluates the performance of the model M under the input x; Represents the gradient of the loss function with respect to the model parameter M; dx represents the integral variable, and the adjustment process involves all input data x.

[0039] Furthermore, the enhanced random forest algorithm is expressed by the formula:

[0040]

[0041] Where T ij represents the transition function of the jth decision node in the i-th tree, Θ ij is the parameter of the node, γ i and β ij is the coefficient that adjusts the contribution of each tree and its nodes, and h is a nonlinear activation function used to process the data structure and identify the precursors of transformer failure;

[0042] Then use a deep convolutional neural network:

[0043]

[0044] Where W k and b k are the weight and bias of the kth convolutional layer, * represents the convolution operation, R k is the response function, σ is the high-level activation function used to extract and learn the deep features of the transformer operating status from time series data;

[0045] Finally, extended principal component analysis is applied:

[0046]

[0047] Where V k are the principal components extracted from the expanded covariance matrix of the data, Λ k It is the inverse of the eigenvalue matrix associated with the principal components and is used to dynamically adjust the principal components of the data for dimensionality reduction and noise reduction.

[0048] Furthermore, the method for constructing a fault prediction model:

[0049] S1. First, use the multidimensional data standardization formula:

[0050]

[0051] Where X represents the raw data collected from multiple sensors, μ is the expected value of the data, P is the probability measure of the data distribution, and the data scale is adjusted by calculating the standard deviation and mean of the entire data set;

[0052] S2, then by enhancing the random forest model, the function is expressed as:

[0053]

[0054] Where N is the number of trees, is the prediction function of the ith tree, Mi and σ 2 are the average response and variance of the i-th tree, Θ i is the parameter of the tree, used to handle nonlinear and high-dimensional data;

[0055] Combining deep convolutional neural networks, the function expression is:

[0056]

[0057] Where W k ,b k are weights and biases, * represents convolution operation, R is ReLU activation function, and tanh is hyperbolic tangent activation function, which is used to learn and extract key features from temporal and spatial data;

[0058] S3. Finally, high-dimensional cross-validation:

[0059]

[0060] Where k is the number of folds, model j is the model trained on the jth fold, is the test set, is the loss function used to control the stability of the model on unseen data.

[0061] Furthermore, the real-time data processing system is established including:

[0062] S1. Preprocess the data continuously collected from multiple sensors using mathematical formulas.

[0063]

[0064] Normalize high-dimensional data, where x represents any sensor data collected in real time, μ(x) is the moving average of the data, and D processed (x) Enhance the nonlinear expression of data features through logarithmic and exponential processing;

[0065] S2. Next, deploy the hybrid machine learning model through the integral transformation formula:

[0066]

[0067] Calculate the failure probability, where T is the time window length, w i is the feature weight, g i (x(t)) is the nonlinear characteristic function of the time series data x(t), P fault (x) Enhance the model’s response to dynamic changes in data through time integration;

[0068] S3. Once the model outputs the failure probability P fault Exceeding the preset threshold θ, through the formula:

[0069] Alert=∫Θ(P fault (x)-θ)dx

[0070] Triggering the early warning mechanism, where P fault (x) represents the failure probability calculated in S2; θ represents the preset threshold. If the failure probability exceeds this value, an early warning needs to be triggered; Θ represents the Heaviside step function, which is used to create a threshold effect. fault When (x) is greater than θ, the function value is 1 and an alarm is triggered, otherwise it is 0 and no alarm is triggered.

[0071] Beneficial effects of the present invention:

[0072] 1. Improve the accuracy of fault prediction: By fusing multiple data from different sensors (such as temperature, current, oil quality, insulation status, etc.), this method can capture a more comprehensive transformer operating status. The use of this comprehensive data can reveal potential faults and anomalies more effectively than a single data source, thereby significantly improving the accuracy of fault prediction.

[0073] 2. Real-time monitoring and dynamic response: This method realizes real-time monitoring of transformers. By continuously collecting and analyzing data, abnormal changes can be discovered in a timely manner. In addition, the model can dynamically adjust and respond to changes in data to ensure the timeliness and relevance of the prediction results.

[0074] 3. Reduce maintenance costs and extend equipment life: Early fault detection and accurate fault prediction reduce unexpected repairs and downtime, thereby reducing maintenance costs. In addition, excessive losses caused by faults are avoided, which helps to extend the service life of transformers and related equipment.

[0075] 4. Optimize maintenance and operational decisions: The data-driven insights provided support smarter operational decisions, such as optimizing maintenance plans and resource allocation based on prediction results. The insights provided by this approach can help operators optimize operation and maintenance strategies and achieve more efficient resource utilization and cost control. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 This is a flow chart of the main transformer fault prediction method based on multivariable data fusion of the present invention.

[0077] Figure 2 This is a flow chart of the Kalman filtering technology used in the present invention.

[0078] Figure 3 A flow chart of a method for identifying key data indicators according to the present invention.

[0079] Figure 4 A flow chart of the method for constructing a fault prediction model for the present invention. DETAILED DESCRIPTION

[0080] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0081] See attached Figure 1 S1, the present invention uses multi-source data fusion technology, including Kalman filtering or Bayesian networks, to process and integrate multiple data types collected from the main transformer. Specifically, it includes but is not limited to the temperature readings of the transformer, the load current, the dissolved gas analysis (DGA) in oil, and the condition of the insulating materials. These data sources are processed comprehensively to form a comprehensive description of the equipment status. Kalman filtering is used for continuous data updating and noise filtering, and is suitable for data such as temperature and current. The state estimate of the transformer is continuously adjusted through prediction and correction steps. Bayesian networks are used to process more classified or discrete data, such as DGA results and insulating material tests, and probabilistic reasoning is used to evaluate the nature of different states and their impact on the health of the transformer. After data fusion, information gain evaluation is performed to determine which data sources contribute the most to the prediction of transformer failures. Through information gain evaluation, the data fusion strategy can be optimized to ensure that the prediction model more accurately reflects the actual operating status of the transformer, thereby effectively guiding maintenance decisions and operations.

[0082] Kalman filtering and Bayesian networks process continuous time series data and categorical or non-continuous data, respectively, to achieve comprehensive monitoring of transformer status and fault prediction. Kalman filtering is used for real-time processing of continuous time series data such as temperature, load current, etc. collected from the main transformer. The state at the next time point is estimated using a prediction model, and the state estimate is updated by comparison with the actual observed data, thereby continuously estimating the real-time state of the transformer. This method can effectively smooth out the noise in the data and provide accurate information about the current working condition of the transformer. Categorical or non-continuous data, such as dissolved gas analysis (DGA) results in oil and the condition of insulation materials, which are usually related to specific types of potential transformer faults, are then processed by Bayesian networks. Bayesian networks evaluate the probability of different types of transformer faults given current and historical data through probabilistic reasoning. This includes calculating the conditional probability of fault occurrence and updating these probabilities using prior knowledge and real-time data.

[0083] Information gain analysis is used to evaluate the contribution of different data sources to the fault prediction model, thereby optimizing the model's structure and input features. First, various data collected from the transformer monitoring system, including temperature, current, dissolved gas analysis (DGA), and insulation material status, are analyzed to determine the specific contribution of each data type to the effectiveness of the prediction model. This analysis process involves calculating information gain, which quantifies the amount of information each data has in distinguishing fault conditions. Through analysis, key data indicators that have the greatest impact on the prediction model can be identified. Then, based on these information gain results, the data fusion model is adjusted and optimized, including reconfiguring the input layer in the model, selecting a more suitable data processing algorithm, or changing the feature selection method.

[0084] like Figure 1 As shown in S2, advanced machine learning algorithms suitable for processing nonlinear and multi-dimensional data are selected, including random forests, support vector machines (SVMs), and deep learning networks. Random forests effectively improve the accuracy and generalization ability of the model by building multiple decision trees and voting on the output results, and are suitable for processing data with a large number of input variables. Support vector machines separate different categories by finding the best hyperplane in high-dimensional space, and are suitable for processing classification problems with unclear boundaries. Deep learning networks, especially convolutional neural networks (CNNs) and recurrent neural networks (RNNs), are able to learn complex patterns and features from data, especially for time series data such as continuously monitored current and temperature data, and can extract time-dependent characteristics and potential connections in the data. The fault prediction models built using these advanced algorithms are trained and validated using historical fault data and normal operating data collected from transformers, which provide the actual conditions and scenarios required for model learning and parameter adjustment. Through this training, the model is able to identify patterns and precursors that lead to failures, while the validation process ensures that the model has good predictive performance and robustness on new or unseen data, thereby ensuring the accuracy and timeliness of failure predictions, reducing downtime and maintenance costs caused by failures.

[0085] like Figure 1As shown in S3, a real-time data processing system is finally introduced to continuously receive and process data from multiple monitoring sensors of the transformer, such as temperature, current, vibration, etc. These data are continuously fed into previously trained fault prediction models, such as random forests, support vector machines, or deep learning networks, to analyze and calculate the fault probability in real time. At the same time, an early warning system is set up in the system to monitor the fault probability output by the model. When this probability exceeds a preset threshold, the early warning system will automatically trigger the early warning mechanism, including sending real-time alerts to the operation and maintenance team or triggering automated response measures, thereby achieving timely early warning of the fault. This setting reduces or avoids downtime and maintenance costs caused by faults, and enhances the reliability and efficiency of the entire power system.

[0086] Embodiment 1:

[0087] like Figure 2 As shown, in this embodiment, Kalman filtering technology is used to monitor and predict the status of the transformer, and the model is dynamically adjusted according to the collected data to predict potential faults.

[0088] Assume that a power company needs to monitor the health of its high-voltage transformer to prevent failures and maintenance interruptions. The transformer is equipped with multiple sensors that collect temperature and current data in real time. These data are continuous time series data updated every minute.

[0089] Kalman filtering technique is used to estimate and update the current state (temperature and current) of the transformer.

[0090] The model formula is defined as follows:

[0091] X t+1 =X t +α[(Y t -X t )+βΔt]+γ∈

[0092] in:

[0093] X t is the state estimate at time t, such as X t =[100℃,200A] T .

[0094] Y t is the actual observation value collected from the sensor at the same time t, such as Y t =[102℃,198A] T .

[0095] Δt is the time step, which is set to 1 minute here.

[0096] α and β are coefficients for adjusting the process dynamics, and are set here as α = 0.5 and β = 0.05.

[0097] γ is the noise adjustment coefficient, which is set to γ=0.1 here.

[0098] ∈ represents the noise in the process, which is set to be normally distributed with a mean of 0 and a standard deviation of 1, for example ∈~N(0,1).

[0099] Initial state: X 0 =[100℃,200A] T

[0100] First minute observation received: Y 1 =[102°C,198A] T

[0101] Apply the Kalman filter formula to calculate the next state:

[0102] X 1 =[100℃,200A] T +0.5[([102℃,198A] T -[100℃,200A] T )+0.05×1]+0.1×∈

[0103] Set ∈ = [0.2, -0.1] T

[0104] X 1 =[100℃,200A] T +0.5[[2℃,-2A] T +[0.05,0.05] T ]+[0.02℃,-0.01A] T

[0105] X 1 =[100℃,200A] T +[1.025℃,-0.975A] T +[0.02℃,-0.01A] T

[0106] X 1 =[101.045℃,199.015A] T

[0107] The calculation demonstrates the use of Kalman filtering to dynamically update the transformer state estimate based on real-time observation data. Through continuous updates, the system is able to detect abnormal trends in advance, thereby triggering early warning mechanisms in time to prevent potential failures from developing into more serious problems.

[0108] The Bayesian network model is introduced to process categorical or non-continuous data such as dissolved gas analysis (DGA) and insulation material status, thereby providing an accurate probabilistic assessment of the fault type. This step is performed after the Kalman filter model performs real-time state estimation on continuous time series data such as temperature and current, aiming to use the probabilistic reasoning ability of the Bayesian network to comprehensively process and analyze data provided by different sensors and detection technologies.

[0109] Specifically, the Bayesian network model is applied to process and infer the nature of potential transformer faults, which is defined by the following formula:

[0110]

[0111] in:

[0112] F represents the predicted fault type.

[0113] E i It is evidence indicating some aspect of the transformer's condition, such as DGA test results.

[0114] P(F|E 1 ,E 2 ,…,E n ) is the posterior probability of fault type F given all the evidence.

[0115] Assume that a transformer has three potential fault types: insulation aging (F1), overheating (F2), and current unbalance (F3). Based on the DGA test, there are two types of evidence: high methane content (E1) and high acetylene content (E2). For a simple example, the following probability values ​​can be set:

[0116] P(E1|F1)=0.1, P(E1|F2)=0.8, P(E1|F3)=0.1

[0117] P(E2|F1)=0.7, P(E2|F2)=0.2, P(E2|F3)=0.1

[0118] Prior probability P(F1) = 0.2, P(F2) = 0.5, P(F3) = 0.3

[0119] When both methane and acetylene levels are high, the posterior probability of the fault type can be calculated:

[0120]

[0121] For each fault type, the calculation is as follows:

[0122] For F1:

[0123] For F2:

[0124] For F3:

[0125] The calculation results show that the most common fault type is overheating (F2), with a probability of about 73%. The practicality of this method lies in its ability to integrate multiple pieces of evidence to provide a comprehensive and accurate assessment of transformer fault types, making maintenance decisions more effective and timely.

[0126] Finally, a comprehensive evaluation formula is used to combine the outputs of the Kalman filter model and the Bayesian network model to comprehensively evaluate the health of the transformer. This comprehensive evaluation model not only takes into account the real-time state estimation of the transformer, but also incorporates the probabilistic analysis of the fault type. In this way, potential faults can be predicted and warned more accurately.

[0127] The formula is as follows:

[0128]

[0129] in:

[0130] R: represents the comprehensive rating of the transformer's health status.

[0131] X T : Represents the real-time state estimation value of the transformer obtained from the Kalman filter.

[0132] P(F): represents the probability of various fault types calculated by the Bayesian network.

[0133] α and β: are weight parameters used to balance the contribution of continuous state estimation and failure probability to the final health score.

[0134] ∫…dT: represents the integration over the entire observation period to obtain the total health score.

[0135] It is assumed that the necessary data has been obtained from the Kalman filter and the Bayesian network within a specific time window. Set the following parameters and values:

[0136] α = 0.7, β = 0.3, these two weights reflect the higher emphasis on real-time state estimation while also considering the importance of failure probability.

[0137] Assume that at time point T, the estimated state value of the transformer is X T =120°C (indicates temperature), and the probability of the fault type is as follows:

[0138] Overheat P(F1)=0.7

[0139] Insulation aging P(F2)=0.2

[0140] Current unbalance P(F3)=0.1

[0141] According to the comprehensive evaluation formula, the transformer health score R can be calculated:

[0142] R=∫(0.7×120+0.3×(0.7×log0.7+0.2×log0.2+0.1×log0.1))dT

[0143] R=∫(84+0.3×(-0.3567-0.5004-0.3322))dT

[0144] R = ∫(84-0.3×1.1893)dT

[0145] R = ∫(84-0.35679)dT

[0146] R = ∫83.64321dT

[0147] Calculations show that after taking into account the real-time status of the transformer and the probability of potential failure, a comprehensive score is obtained, which can help the operation and maintenance team understand the overall health of the transformer and take preventive measures when necessary.

[0148] Embodiment 2:

[0149] Combination Figure 3 In this embodiment, a systematic information gain analysis process is used to identify key data indicators, and the contribution of these indicators to fault prediction is quantified through decision tree analysis and mutual information calculation. This process not only ensures the accuracy and practicality of data analysis, but also enhances the performance of the prediction model.

[0150] 1. Data collection and preprocessing:

[0151] Key operating data such as temperature, current, oil quality and insulation condition are collected from the transformer monitoring system. These data are the basis of the fault prediction model, which contains all the key performance indicators of transformer operation.

[0152] Use advanced data preprocessing formulas:

[0153] P prep (D) = ∫exp(-|D-μ| 2 / σ 2 )dD

[0154] This formula processes the data through a Gaussian function to make it smoother and remove outliers, where D is the original data set, μ and σ are the mean and standard deviation of the data, respectively, to ensure the quality and consistency of the data before further analysis.

[0155] 2. Information gain analysis and decision tree application:

[0156] Information gain analysis is used to determine which data indicators are most critical to the determination of fault types. By calculating the mutual information between each data indicator and the fault event, the contribution of each indicator can be quantified.

[0157] Decision tree analysis further helps clarify the relationship between data indicators and fault types in complex data environments. By building classification rules, the decision tree clearly shows the impact path and degree of different indicators on fault prediction.

[0158] The transformer data collected within a period of time is assumed to be as follows:

[0159] Temperature (D 1 ): [98℃, 102℃, 100℃]

[0160] Current (D 2 ): [150A, 160A, 155A]

[0161] Oil quality (D 3 ): good, good, deteriorated

[0162] Insulation status (D 4 ): good, deteriorated, deteriorated

[0163] For preprocessing, μ and σ are set to be the mean and standard deviation of temperature and current data respectively:

[0164] For temperature: μ temp =100℃,σ temp =2℃

[0165] For current: μ current =155A,σ current =5A

[0166] Preprocessing formula example:

[0167] P prep (D 1 )=∫exp(-|(98,102,100)-100| 2 / 4)dD 1

[0168] Calculate P prep (D 1 ) The corresponding results are processed similarly for current and other indicators.

[0169] Next, mutual information calculation is used to determine which data plays a key role in fault type identification, and these relationships are analyzed and visualized by building a decision tree model.

[0170] Mutual information calculation is applied to help determine the statistical correlation between different types of data (such as temperature, current, oil quality and insulation status) and the occurrence of transformer faults. This analysis process deepens the understanding of the relationship between data characteristics and faults, and optimizes the accuracy and response speed of the fault prediction model. The mutual information calculation formula is as follows, which is used to evaluate the statistical importance of each type of data in predicting faults:

[0171]

[0172] This formula reveals the interdependence between two variables by measuring the amount of shared information and is a powerful tool for quantifying the mutual information between variables. In practical applications, X can represent the type of operating data, such as temperature, and Y represents the occurrence of fault types.

[0173] The following actual data and steps are set up for implementation:

[0174] 1. Data Collection:

[0175] Temperature data: 98℃, 102℃, 100℃

[0176] Current data: 150A, 160A, 155A

[0177] Oil quality data: good, good, deteriorated

[0178] Insulation status: good, deteriorated, deteriorated

[0179] 2. Data preprocessing:

[0180] Using the formula P prep (D) Pre-processing of temperature and current.

[0181] Setting μ temp =100℃,σ temp =2℃;

[0182] Setting μ current =155A,σ current =5A.

[0183] 3. Mutual information calculation implementation:

[0184] Assumptions Preliminary analysis shows that the fault type is highly correlated with temperature and current, further calculation of the mutual information of these variables is required.

[0185] According to the data and related fault records, set X as the data set of temperature and current, and Y as the fault record (such as no fault, overheating, insulation degradation).

[0186] Using the above formula, for each pair (x, y) (ie, each combination of temperature or current value and corresponding fault type), the mutual information is calculated.

[0187] Consider the following simplified failure probability distribution and data probability distribution:

[0188] p(x) (the probability distribution of temperature / current data) is assumed to be uniformly distributed.

[0189] p(y) (probability distribution of fault type) is assumed to be set based on historical fault data, such as: no fault 60%, overheating 30%, insulation degradation 10%.

[0190] For the specific data point x = 102°C and fault type y = overheat:

[0191] p(x,y) can be estimated based on the actual observed frequency of overheating failures occurring at 102°C and is set to 0.03.

[0192] p(x)=0.33 (because there are only three temperature points, uniform distribution is assumed).

[0193] p(y)=0.3.

[0194] Insert the mutual information formula:

[0195]

[0196] The calculation results show that there is a certain degree of statistical correlation between specific temperatures and the fault type "overheating". This calculation can help understand which specific data types contribute most to fault prediction.

[0197] Then, through decision tree analysis, key data indicators in the main transformer operation data are identified and evaluated, and the contribution of various data to fault prediction is further quantified through mutual information calculation. After the comprehensive application of these technologies, the decision tree model is introduced to map input data to output decisions through formulas, enhancing the model's ability to process data and predict faults.

[0198] The role of decision tree analysis in this process is to classify the data points collected from the transformer through a series of decision rules, and ultimately determine the type of fault the data point points to. The following formula is used to achieve this mapping:

[0199]

[0200] in:

[0201] c(x) represents the classification of data point x in the tree.

[0202] N is the number of leaf nodes in the tree.

[0203] v n is the predicted value of the nth leaf node.

[0204] δ is the Kronecker function, which is used to determine whether x falls on a leaf node n. If so, it takes the value 1, otherwise it takes the value 0.

[0205] After the operating data of the assumed transformer is collected and preprocessed, its contribution to fault prediction needs to be evaluated.

[0206] 1. Data and model settings:

[0207] Data collected includes temperature, current, oil analysis and insulation condition.

[0208] It is assumed that the decision tree model has three leaf nodes determined by the previous analysis, corresponding to the three fault types of overheating, insulation aging and oil quality deterioration.

[0209] 2. Decision tree parameters and data examples:

[0210] The data point x at a specific time is set to [102℃, 150A, ′ Deterioration ′ ,'good ′ ].

[0211] Through decision tree analysis, x is classified into the leaf node of overheating fault, that is, c(x)=1.

[0212] The corresponding leaf node prediction value v 1 can be set as the failure probability, such as v 1 =0.8 (indicating an 80% probability of pointing to overheating failure).

[0213] 3. Calculate the comprehensive evaluation:

[0214] Apply Formula R tree (x) is calculated:

[0215]

[0216] Calculations show that a given data point x most likely belongs to the overheating fault category and has a high probability of failure. This approach accurately classifies and predicts faults for individual data points, and provides a quantitative method to evaluate the contribution of each data point to a specific fault type, making fault predictions more accurate and timely.

[0217] Embodiment 3:

[0218] In this embodiment, information gain analysis is used to identify the data indicators that have the greatest impact on fault prediction. This analysis relies on a nonlinear integral formula to measure the mutual information between different operating data types (such as temperature, current, oil quality and insulation status) and fault occurrence.

[0219] formula:

[0220]

[0221] in:

[0222] I(X;Y) stands for information gain, which measures the mutual information between variables X and Y.

[0223] X, Y represent random variables of the main transformer operating data, for example, X is the temperature reading and Y is the fault type.

[0224] p(x,y) represents the joint probability density function, which indicates the probability that variables X and Y take a certain value at the same time.

[0225] p(x) and p(y) represent marginal probability density functions, which represent the probability of X and Y taking any value, respectively.

[0226] stands for log-likelihood ratio, which is used to quantify the degree of dependence between X and Y.

[0227] dx,dy represents the integral symbol, which means integrating all X and Y values ​​to calculate the overall information gain.

[0228] The temperature and current data of the set transformer and its fault history have been collected. The specific data set includes:

[0229] Temperature data X: 98℃, 102℃, 100℃

[0230] Current data X: 150A, 160A, 155A

[0231] Fault type Y: No fault, overheating, insulation aging

[0232] For the calculation of information gain, we need to estimate p(x,y), p(x) and p(y). Setting:

[0233] p(x) and p(y) are estimated based on the data distribution.

[0234] p(y) is set according to the frequency of occurrence of each type of fault in the fault record, for example, the probability of overheating fault is 30%, insulation aging is 20%, and no fault is 50%.

[0235] It is assumed that through specific data statistics and processing, a set of specific values ​​of p(x,y), p(x), and p(y) are obtained. Then they are substituted into the information gain formula for calculation. This requires actual data sets and statistical methods for specific calculations.

[0236] The information gain value calculated by this method will directly reflect the mutual information between data such as temperature and current and the fault type. The results can be used to adjust and optimize the data fusion model structure to ensure that the model more accurately reflects the actual operating status and potential fault risks of the transformer.

[0237] Following the information gain analysis, the obtained results are further used to optimize the data fusion model structure. By applying advanced adjustment formulas, the model parameters can be dynamically adjusted to more accurately reflect the actual operating status of the transformer and predict potential faults.

[0238] The high-level adjustment formula applied is as follows:

[0239]

[0240] in:

[0241] M new and M old Represent the adjusted and original model structures, respectively.

[0242] α(x) represents a dynamic learning rate function that depends on the input data x and is used to adjust the rate at which the model parameters are updated.

[0243] Represents the loss function, which evaluates the performance of the model M under the input x.

[0244] Represents the gradient of the loss function with respect to the model parameters M.

[0245] dx represents the integration variable, and the adjustment process involves all input data x.

[0246] The data monitoring system of the transformer collects data on temperature, current, oil quality and insulation status, and has identified temperature and current as key data indicators through information gain analysis. Based on the information gain results of these indicators, it is decided to optimize the weights of these indicators in the model.

[0247] 1. Set parameter values:

[0248] Set the original model M old The weights for temperature and current are set as w temp = 0.5 and w current =0.5.

[0249] Based on the information gain results, the dynamic learning rate α(x) is set as a function of temperature and current. For example, if the information gain shows that temperature is more critical to fault prediction, the learning rate of temperature is higher and α is set. temp =0.7,α current =0.3.

[0250] 2. Loss function and gradient calculation:

[0251] Setting the loss function is the sum of squares of the prediction errors, that is Where y is the actual fault status, is the failure state predicted by the model.

[0252] Calculate the gradient of the loss function Set to

[0253] 3. Model update:

[0254] Update the model's weight parameters: w temp,new =w temp,old +∫0.7×Gradient temp dx

[0255] Similarly update the current weight: w current,new =w current,old +∫0.3×Gradient current dx

[0256] This process ensures that the model structure constantly adapts to the latest operating data and environmental changes through actual monitoring data and dynamic adjustment of formulas, thereby improving the accuracy of fault prediction and the response speed of the model.

[0257] Embodiment 4:

[0258] In this embodiment, combining machine learning algorithms, especially enhanced random forests and feature engineering techniques, such as principal component analysis (PCA) or encoders, to optimize the input features of the model is a key step to improve the accuracy of fault detection. This method can effectively process and identify complex nonlinear relationships caused by various operating data (such as temperature, current, oil quality, insulation status, etc.), thereby more accurately predicting potential faults.

[0259] Algorithm formula:

[0260]

[0261] in:

[0262] T ij (X, Θ ij ) represents the transition function of the jth decision node in the i-th tree, which depends on the parameter Θ ij .

[0263] γ i and β ij It is the coefficient that adjusts the contribution of each tree and its nodes.

[0264] h is a nonlinear activation function used to process the structure of the data and identify the failure precursors of the transformer.

[0265] The transformer monitoring system collects detailed temperature, current, oil quality and insulation status data. These data are preliminarily processed through feature engineering techniques, such as PCA, to reduce the dimension and highlight the key features.

[0266] 1. Data preprocessing:

[0267] Use PCA to process data, for example to extract principal components from temperature and current data.

[0268] The principal components extracted by PCA were set to cover 90% of the data variance to ensure that key information was retained.

[0269] 2. Parameter setting and model training:

[0270] The number of trees in the random forest was chosen to be n = 100 to ensure the generalization ability and accuracy of the model.

[0271] Set the maximum depth of each tree Θ ij (For example, the maximum depth of each tree is 15) to control the complexity of the model.

[0272] Set the parameter β for each decision node ij , for example, the β of all nodes ij =1 / n, ensuring that each node has an equal influence on the final decision.

[0273] Select a suitable nonlinear activation function h, such as ReLU or Sigmoid, to enhance the model's ability to handle nonlinear data.

[0274] 3. Model application:

[0275] Use the trained random forest model to predict failures on newly collected data.

[0276] The enhanced random forest formula is applied to calculate the failure probability for each data point.

[0277] If the temperature suddenly rises to an abnormal value, the model will quickly identify such changes through enhanced node influence and regulation coefficients, and predict an overheating failure.

[0278] Through the application of this method, not only the performance of the prediction model is improved, but also the model structure and parameters are dynamically adjusted so that the model can more accurately reflect the actual operating status of the transformer.

[0279] The application of deep learning technology, especially deep convolutional neural network (CNN), has further enhanced the ability to recognize complex patterns in transformer operating status. Through deep convolutional neural network, deep features of transformer operating status can be extracted from time series data, which is crucial for early fault prediction and accuracy improvement.

[0280] Formula analysis:

[0281]

[0282] in:

[0283] W k and b k are the weight and bias of the kth convolutional layer, respectively.

[0284] ★ represents a convolution operation, which is used to apply a filter W on the input data X k .

[0285] R k It is a response function, usually an activation function such as ReLU or Leaky ReLU, which is used to introduce nonlinearity so that the network can handle complex data structures.

[0286] σ is a high-level activation function, such as softmax, used in the output layer to convert the result of the convolution into a probability distribution.

[0287] 1. Model structure design:

[0288] Design a network with multiple convolutional layers, each using filters of different sizes to capture short-term to long-term dependencies.

[0289] Each convolutional layer is followed by a pooling layer to reduce the feature dimension and enhance the generalization ability of the model.

[0290] 2. Parameter setting:

[0291] Set the weight W of each convolutional layer k and bias b k ,These parameters are optimized by the back-propagation algorithm during the training process.

[0292] Choose an appropriate response function R k , ReLU is usually chosen because it can effectively solve the gradient vanishing problem and accelerate convergence.

[0293] 3. Model training and application:

[0294] Use historical data (including normal operation data and data when failures occur) to train the model.

[0295] The trained CNN model is used to analyze the transformer data collected in real time to predict the nature of the fault.

[0296] The following actual data are set to be input:

[0297] The temperature and current data over a period of time have had their key features extracted through principal component analysis.

[0298] According to the set network parameters, for example, three convolutional layers are selected, and the weight of each layer is W 1 ,W 2 ,W 3 Initialize them separately, and adjust these weights through training data to minimize the loss function (e.g., cross entropy loss). During training, data is passed through these convolutional layers, each layer applies a ReLU activation function, and the last layer uses a softmax function to output the probability of the fault type.

[0299] Extended principal component analysis (PCA) is used to extract the most important features from complex sensor data while reducing noise and optimizing the model's input. This approach is particularly suitable for processing multiple data such as temperature, current, oil quality and insulation status, aiming to improve the performance and accuracy of fault detection models.

[0300] Extended PCA formula:

[0301]

[0302] in:

[0303] V k It is the principal component vector extracted from the covariance matrix of the data, responsible for capturing the main direction of change of the data.

[0304] is the principal component V k The inverse of the associated eigenvalue matrix is ​​used to adjust the contribution of each principal component in order to more accurately reflect the important features of the data.

[0305] X is the original data, μ is the mean of the data, and X-μ here represents the data after centralization.

[0306] e represents an exponential function, which is used to enhance the feature expression of the processed data to make it more suitable for subsequent machine learning processing.

[0307] 1. Data preprocessing and parameter setting:

[0308] Collect transformer operating data including temperature, current, oil quality and insulation condition.

[0309] The data are standardized, the mean μ and covariance matrix of the data are calculated, and then the eigenvalue decomposition is performed to obtain V k and Λ k .

[0310] 2. Principal component extraction and application:

[0311] The first k principal components are selected based on their cumulative contribution and are usually chosen to cover more than 90% of the total variance.

[0312] Use the extended PCA formula to reduce the dimensionality of the data. For example, if V k is a 3×3 matrix containing the first three principal components, and Λ k Corresponding to the eigenvalue matrices of these three principal components, for each data point X, the result obtained by PCA(X) calculation will be a new data vector that compresses and enhances the main features.

[0313] 3. Model training and fault prediction:

[0314] Use the processed data to train a machine learning model, such as a random forest or a deep neural network.

[0315] Apply the trained model to real-time or newly collected data to predict the type of failure.

[0316] Coefficient value range

[0317] V k :Determined according to the characteristics of the data and eigenvalue decomposition, the principal components corresponding to the first few (such as the first three) largest eigenvalues ​​are usually selected.

[0318] Λ k :The selection of eigenvalues ​​is based on the number of principal components retained, and generally eigenvalues ​​that cover most (>90%) of the data variance are selected.

[0319] μ: The mean of the data, calculated directly from the dataset.

[0320] Through this extended PCA method, complex sensor data can be effectively processed and simplified, improving the efficiency and accuracy of subsequent fault prediction models.

[0321] Embodiment 5:

[0322] Combination Figure 4 In this embodiment, building an effective fault prediction model involves accurately normalizing data from multiple sensors. This is a key step to ensure that the data has appropriate scale and statistical properties before entering the model. The multidimensional data normalization formula is at the core of the process, helping to reduce noise and bias in the data, thereby improving the prediction accuracy of the model.

[0323] Multidimensional data standardization formula:

[0324]

[0325] in:

[0326] X represents the raw data collected from multiple sensors.

[0327] μ is the expected value of the data, usually obtained by calculating the mean of the data set.

[0328] P is the probability measure of the data distribution and is used to calculate the statistical properties of the data.

[0329] In the denominator is the standard deviation of the data and is used to normalize the data.

[0330] 1. Data Collection:

[0331] Collect data including temperature, current, oil quality, and insulation status from different sensors on the transformer.

[0332] 2. Calculate the expected value μ and standard deviation of the data:

[0333] Calculate the average value μ for each data type. For example, the temperature data includes a series of measured values ​​such as 98°C, 102°C, and 100°C, and its average value is μ=100°C.

[0334] Calculate the standard deviation of the data. Use the formula Assuming the distribution of the temperature data is uniform, the standard deviation calculation shows the deviation of the data points around the mean.

[0335] 3. Apply the normalization formula:

[0336] Apply the normalization formula to each data point. For example, if the temperature data point is 102°C, its normalized value is

[0337] This normalization ensures that all data types are appropriately scaled, making their input to the model more consistent and comparable.

[0338] If the standard deviation is set to 2°C, the normalization of the temperature data point 102°C would be:

[0339]

[0340] The effectiveness of this method lies in its ability to make data from different sensors have the same scale and distribution, which is crucial for subsequent model training and fault prediction. Through such processing, the model can not only better understand and interpret the data, but also improve its generalization ability to unseen data.

[0341] Enhanced random forest models and deep convolutional neural networks are used to process complex data relationships to improve the accuracy and efficiency of fault prediction. These technologies combine advanced algorithms and feature engineering techniques to optimize model performance, enabling it to more effectively identify and predict potential faults.

[0342] Enhanced Random Forest Model Description

[0343] Model formula:

[0344]

[0345] in:

[0346] N is the number of trees.

[0347] is the prediction function of the i-th tree.

[0348] M i and σ 2 are the mean response and variance of the ith tree, respectively.

[0349] Θ i is a parameter of the tree, which is used to handle nonlinear and high-dimensional data.

[0350] 1. Data preparation and parameter setting:

[0351] Set the number of trees N=100.

[0352] The parameter Θ of each tree i , including the depth of the tree, splitting criteria, etc., need to be adjusted according to the actual data.

[0353] Set the M of each tree i and σ 2 Based on the statistical properties of the training data.

[0354] 2. Model training and application:

[0355] Train each tree Use an enhanced random forest formulation to handle nonlinear data features.

[0356] Apply the model to the newly collected data and calculate the failure prediction value for each data point.

[0357] Deep Convolutional Neural Network Description

[0358] Model formula:

[0359]

[0360] in:

[0361] W k ,b k are the weights and biases of the kth convolutional layer.

[0362] * indicates a convolution operation.

[0363] R is the ReLU activation function.

[0364] Tanh is a hyperbolic tangent activation function, which is used for nonlinear transformation of the output layer.

[0365] 1. Network architecture and parameter settings:

[0366] The number of convolutional layers K is usually set based on the complexity of the data and the required level of feature extraction.

[0367] Set weights W for each convolutional layer k and bias b k , these parameters are optimized by back-propagation during the training process.

[0368] 2. Model training and feature extraction:

[0369] The model is trained using a deep convolutional network to extract key features from the time series and spatial data of the transformer.

[0370] Apply trained models to new or real-time data for fault analysis and prediction.

[0371] The combination of these models can extract and learn deep features from multi-source data of transformers to optimize the accuracy of fault prediction. Through the implementation of this method, the reliability of transformer fault prediction can be significantly improved, thereby effectively guiding maintenance decisions and reducing unexpected downtime and maintenance costs.

[0372] Finally, high-dimensional cross-validation is implemented to ensure that the fault prediction model not only performs well on the training data, but also has high stability and reliability on unseen data. High-dimensional cross-validation evaluates the generalization ability of the model by repeatedly training and testing the model on different data subsets.

[0373] Description of High-Dimensional Cross-Validation

[0374] Formula analysis:

[0375]

[0376] in:

[0377] k is the number of folds, common choices such as 5 or 10, used to divide the dataset.

[0378] model j is the model trained on the jth fold.

[0379] It is a dataset used for testing, which is not seen during training.

[0380] is the loss function used to evaluate the model on the test set on performance.

[0381] 1. Data division and model training:

[0382] Suppose there is a set of data, containing a total of 1000 samples. Choose to perform 10-fold cross validation, that is, k = 10, and each fold contains 100 samples.

[0383] The data is randomly divided into 10 non-overlapping subsets, and one subset is set aside each time as the test set. The remaining nine subsets are used to train the model j .

[0384] 2. Cross-validation process:

[0385] For each fold, the model is trained using the remaining 9 subsets, and then the model performance is evaluated on the reserved test set, calculating the loss function

[0386] The loss function can be mean squared error (MSE), cross entropy loss, or any other metric that is appropriate for the specific problem.

[0387] 3. Overall performance evaluation:

[0388] Calculate the average loss function value over all 10 folds to get the overall performance evaluation of the model.

[0389] For example, if the loss for each fold is 0.1, 0.2, 0.15, ..., calculate the average of these loss values.

[0390] Assume that in a certain fold, the model loss on the test set is 0.15, then calculate the average loss of all folds:

[0391]

[0392] The lower the final CV(X) value, the better the model performs on unseen data and the stronger its generalization ability.

[0393] Through this high-dimensional cross-validation method, the fault prediction model of the main transformer can be systematically evaluated and optimized.

[0394] Embodiment 6:

[0395] In this embodiment, the construction of a real-time data processing system is key, especially the effective pre-processing of data continuously collected from multiple sensors, which involves applying mathematical formulas to normalize the data to optimize the data quality and feature expression of subsequent model processing.

[0396] High-dimensional data normalization formula:

[0397]

[0398] in:

[0399] x represents any sensor data collected in real time.

[0400] μ(x) is a moving average of the data, which is often used to smooth the data and reduce the impact of temporary fluctuations.

[0401] D processed (x) is the processed data. The formula enhances the nonlinear expression of the data through logarithmic and exponential processing, so that the model can better capture complex data characteristics.

[0402] The scenario is processing data from a temperature sensor on a transformer. The sensor records the temperature every minute and this data needs to be processed in real time to detect potential anomalies or signs of failure.

[0403] 1. Data collection and preliminary processing:

[0404] Collect real-time temperature data, such as temperature sequence: 98℃, 100℃, 102℃, 100℃, 98℃.

[0405] Calculate the moving average μ(x) of the data and set the window size to 3, so the moving average of each point is 100℃.

[0406] 2. Apply the regularization formula:

[0407] Applying the formula to each data point, first calculating the deviation e of each point from its moving average x-μ(x) .

[0408] The integral of these deviation values ​​is calculated, which can be simplified to a summation (operating on a discrete data set).

[0409] Apply logarithmic function processing to enhance the nonlinear expression of data.

[0410] 3. Calculation example:

[0411] For a data point of 102°C, whose moving average is 100°C, calculate e 102-100 =e 2 ≈7.389.

[0412] Set the sum of the deviation value sequence (approximate integral) to 20 (this is a set value, which needs to be calculated in practice), then:

[0413]

[0414] In practical applications, the calculation of μ(x) depends on the selected window size, with common values ​​ranging from 3 to 10, depending on the update frequency of the sensor data and the required level of smoothing. processed (x) will be used in subsequent fault detection models, such as random forests or neural networks, to improve prediction accuracy and response speed.

[0415] Next, we deploy a hybrid machine learning model, specifically to calculate the probability of failure through the integral transformation formula. This approach enhances the model’s response to dynamic changes in data through time integration, making predictions more accurate and timely.

[0416] High-dimensional integral transformation formula description

[0417] Formula definition:

[0418]

[0419] in:

[0420] T is the time window length, which is used to define the time range of the data considered.

[0421] w i is the feature weight, which determines the i The influence of (x(t)) in the sum.

[0422] g i (x(t)) is a nonlinear feature function for time series data x(t), which is used to extract key information from the data.

[0423] P fault (x) is the calculated failure probability, and the integral result is converted into a probability value through a logic function.

[0424] 1. Feature function design and weight allocation:

[0425] Design nonlinear characteristic function g i ,For example, Fourier transform, wavelet transform and other methods can be used to ,extract the frequency domain or time-frequency features of the data.

[0426] Assign a weight w to each feature function i ,The selection of weights can be based on the impact of the ,features on the fault type and the historical data analysis.

[0427] 2. Data collection and time window setting:

[0428] Collect real-time data x(t) from sensors, such as temperature, current, etc.

[0429] Set the time window T, for example, select data within the past 24 hours for analysis.

[0430] 3. Model calculation and failure probability estimation:

[0431] For each time point t, calculate g i (x(t)) and apply weight w i .

[0432] Perform a time integration and sum the weighted characteristic function results at all time points.

[0433] Substitute the integration result into the logistic function to calculate the failure probability.

[0434] It is assumed that there is data from a temperature sensor with a sampling frequency of once per minute.

[0435] Set the time window T=1440 minutes (ie one day).

[0436] Select feature functions such as the moving average and standard deviation of temperature data, and set the weight w i , such as for the mean value w 1 =0.6, for standard deviation w 2 =0.4.

[0437] With this approach, the health status of the transformer can be effectively monitored and faults predicted in real time, using data collected from multiple sensors to improve the accuracy and timeliness of fault prediction.

[0438] Deploy a real-time data processing system, which uses an early warning mechanism combined with high-dimensional cross-validation to ensure the accuracy of the model and real-time response. After using mathematical formulas to normalize the real-time data in high dimensions, the fault warning is further triggered by a preset threshold.

[0439] Formula analysis:

[0440] Alert=∫Θ(P fault (x)-θ)dx

[0441] in:

[0442] P fault (x) represents the failure probability calculated from the model, which takes into account the data variation within the time window.

[0443] θ is the preset threshold. If the failure probability exceeds this value, an early warning needs to be triggered.

[0444] Θ represents the Heaviside step function, which is used to create a threshold effect, where when P fault When (x) is greater than θ, the function value is 1 and an alarm is triggered, otherwise it is 0 and no alarm is triggered.

[0445] 1. Settings and parameter adjustment:

[0446] Based on historical data analysis, a suitable threshold θ is determined, for example, 0.8, which means that an alarm is triggered when the failure probability is higher than 80%.

[0447] P fault The calculation of (x) is based on the aforementioned high-dimensional cross-validation model, ensuring that each calculation is based on the latest and most relevant data.

[0448] 2. Early warning trigger mechanism:

[0449] Run real-time data through the model and calculate P at each moment fault (x).

[0450] For each P fault The Heaviside function is applied to the (x) value to check whether it exceeds the threshold θ.

[0451] These results are integrated, which usually means counting the duration or number of consecutive threshold violations in practical applications.

[0452] 3. Data substitution and response:

[0453] Set the P value to be continuously monitored when abnormal temperature rise occurs. fault The value exceeds the threshold multiple times in a row.

[0454] Based on the integration results, if it is determined that the probability of failure continues to exceed the standard, the system automatically triggers an alarm and notifies the maintenance team to conduct inspection or intervention.

[0455] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. The main transformer fault prediction method based on multivariable data fusion is characterized by The method comprises the following steps: firstly, using multi-source data fusion technology, including Kalman filtering or Bayesian network, various data sources including temperature, load current, dissolved gas analysis data in oil and insulation material condition collected from the main transformer are integrated to obtain a state description, and information gain evaluation is performed to optimize the data fusion strategy; The continuous time series data collected from the main transformer is processed using the Kalman filter to control the continuous estimation of the real-time state of the transformer, and then the classified or non-continuous data is processed by the Bayesian network to estimate the fault probability under different states using its probabilistic reasoning ability; By analyzing the information gain of each data source, that is, the contribution of the collected data to the fault prediction model, key data indicators are identified, and the data fusion model is adjusted according to the information gain results, the model structure and input characteristics are optimized, and the output of the control model reflects the actual operating status of the transformer; Then, we select advanced machine learning algorithms that process nonlinear and multidimensional data, including random forests, support vector machines, or deep learning networks, to build a fault prediction model, and use historical fault data and normal operation data to train and verify the model. Finally, a real-time data processing system is introduced to continuously receive and process transformer monitoring data, and an early warning system is established. When the fault probability output by the prediction model exceeds the set early warning threshold, the early warning mechanism is triggered to trigger a timely early warning of the fault; The method of identifying key data indicators involves the following steps: Firstly, the key data indicators in the main transformer operation data, including temperature, current, oil quality and insulation status, are determined through information gain analysis. Decision tree analysis and mutual information calculation are used to quantify the contribution of each data to the prediction of different types of transformer faults. Then, the data index that has the greatest impact on the prediction of fault types is identified based on the obtained information gain results, and the data fusion model structure is adjusted and optimized based on the data index; Use machine learning algorithms including enhanced random forests or neural networks to handle nonlinear relationships in data indicators, and optimize the model’s input features through feature engineering techniques including principal component analysis or encoders; The enhanced random forest algorithm is expressed by the formula: Where T ij represents the transition function of the jth decision node in the i-th tree, Θ ij is the parameter of the node, γ i and β ij is the coefficient that adjusts the contribution of each tree and its nodes, and h is a nonlinear activation function used to process the data structure and identify the precursors of transformer failure; Then use a deep convolutional neural network: Where W k and b k are the weight and bias of the kth convolutional layer, ★ represents the convolution operation, R k is the response function, σ is the high-level activation function used to extract and learn the deep features of the transformer operating status from time series data; Finally, extended principal component analysis is applied: Where V k are the principal components extracted from the expanded covariance matrix of the data, Λ k It is the inverse of the eigenvalue matrix associated with the principal component, which is used to dynamically adjust the main components of the data for dimensionality reduction and noise reduction; X is the original data and μ is the mean of the data.

2. The main transformer fault prediction method based on multivariable data fusion according to claim 1 is characterized in that Using Kalman filtering technology, through the formula: X t+1 =X t +α[(Y t -X t )+βΔt]+γ∈ Where X t is the state estimate at time t, Y t is the observation value at the same time, Δt is the time step, α and β are the coefficients for adjusting the process dynamics, γ is the noise adjustment coefficient, ∈ represents the noise in the process to control the continuous estimation of the real-time state of the transformer; The Bayesian network model is used to process categorical or non-continuous data, including dissolved gas analysis DGA and insulation material status in oil, and the formula is: Where, F represents the predicted fault type; E i It is evidence indicating some aspect of the transformer status, including DGA test results; P(F|E1,E2,…,E n ) The function formula provides a failure probability assessment by considering all predicted failure types; Finally, a comprehensive evaluation formula is used to combine the outputs of the above two models to comprehensively evaluate the health status of the transformer. The formula is: Among them, α′ and β′ are weight parameters, which are used to adjust the influence of continuous state estimation and fault probability to control the comprehensive judgment of transformer state and fault warning; X T : represents the real-time state estimation value of the transformer obtained from the Kalman filter; P(F): represents the probability of various fault types calculated by the Bayesian network.

3. The main transformer fault prediction method based on multivariable data fusion according to claim 1 is characterized in that The information gain analysis process includes the following steps: First, collect and preprocess the key operating data of the main transformer, including temperature, current, oil quality analysis and insulation status, using the formula: P prep (D)=∫exp(-|D-μ| 2 / s 2 )dD To optimize the data preprocessing process, where D represents the original data set, μ and σ are the mean and standard deviation of the data respectively; Then apply the mutual information calculation formula: To evaluate the mutual information between different data types and fault occurrence, so as to quantitatively determine the statistical importance of each data in predicting faults; X represents the type of operation data, while Y represents the occurrence of fault type; p(x) is the probability distribution of temperature / current data; p(y) is the probability distribution of fault type; p(x,y) is the estimated value; x is the temperature or current value; y is the fault type; Then analyze the key data indicators through the decision tree, using the formula: To map input data to output decisions, where c(x) represents the classification of data point x in the tree, N is the number of leaf nodes in the tree, and v n is the predicted value of the nth leaf node, δ is the Kronecker function, which means that if x falls on leaf node n, the value is 1, otherwise it is 0.

4. The main transformer fault prediction method based on multivariable data fusion according to claim 1 is characterized in that The information gain analysis adopts the nonlinear integral formula: The amount of information between various data types including temperature, current, oil quality, insulation status and the occurrence of transformer faults is calculated, where I(X;Y) represents information gain, which measures the mutual information between variables X and Y; X, Y represent random variables of the main transformer operation data, and Y is the fault type; p(x,y) represents the joint probability density function, which represents the probability that variables X and Y take a certain value at the same time; p(x) and p(y) represent marginal probability density functions, which represent the probability of X and Y taking any value respectively; Represents the log-likelihood ratio, which is used to quantify the degree of dependence between X and Y; dx,dy represents the integral symbol, indicating that all possible X and Y values ​​are integrated to calculate the overall information gain; Then optimize the data fusion model structure based on the information gain results and apply the advanced adjustment formula: Among them, M new and M old Represents the adjusted and original model structures respectively; α(x) represents the dynamic learning rate function, which depends on the input data x and is used to adjust the rate of model parameter update; represents the loss function, which evaluates the performance of the model M under the input x; Represents the gradient of the loss function with respect to the model parameter M; dx represents the integral variable, and the adjustment process involves all input data x.

5. The main transformer fault prediction method based on multivariable data fusion according to claim 1 is characterized in that The method for constructing a fault prediction model: S1. First, use the multidimensional data standardization formula: Where X represents the raw data collected from multiple sensors, μ is the expected value of the data, P is the probability measure of the data distribution, and the data scale is adjusted by calculating the standard deviation and mean of the entire data set; S2, then by enhancing the random forest model, the function is expressed as: Where N is the number of trees, Tree i is the prediction function of the ith tree, M i and σ 2 are the average response and variance of the i-th tree, Θ i is the parameter of the tree, used to handle nonlinear and high-dimensional data; Combining deep convolutional neural networks, the function expression is: Where W k ,b k are weights and biases, * represents convolution operation, R is ReLU activation function, and tanh is hyperbolic tangent activation function, which is used to learn and extract key features from temporal and spatial data; S3. Finally, high-dimensional cross-validation: Where k is the number of folds, model j is the model trained on the jth fold, is the test set, is the loss function used to control the stability of the model on unseen data.

6. The main transformer fault prediction method based on multivariable data fusion according to claim 1 is characterized in that The real-time data processing system is established including: S1. Preprocess the data continuously collected from multiple sensors using mathematical formulas. Normalize high-dimensional data, where x represents any sensor data collected in real time, μ(x) is the moving average of the data, and D processed (x) Enhance the nonlinear expression of data features through logarithmic and exponential processing; S2. Next, deploy the hybrid machine learning model through the integral transformation formula: Calculate the failure probability, where T is the time window length, w i is the feature weight, g i (x(t)) is the nonlinear characteristic function of the time series data x(t), P fault (x) Enhance the model’s response to dynamic changes in data through time integration; S3. Once the model outputs the failure probability P fault Exceeding the preset threshold θ, through the formula: Alert=∫Θ(P fault (x)-θ)dx Triggering the early warning mechanism, where P fault (x) represents the failure probability calculated in S2; θ represents the preset threshold. If the failure probability exceeds this value, an early warning needs to be triggered; Θ represents the Heaviside step function, which is used to create a threshold effect. fault When (x) is greater than θ, the function value is 1 and an alarm is triggered, otherwise it is 0 and no alarm is triggered.

Images