Transformer defect identification method based on statistical characteristic inversion and mechanism constraint

Abstract

The application discloses a transformer defect identification method based on statistical characteristic inversion and mechanism constraint. The method first collects transformer data, and constructs a feature set through preprocessing. Then, the electromagnetic field physical equation constraint is embedded in a physical information neural network, a magnetic circuit forward model is trained by fusing a multi-class loss function, and subsequently, a weighted adaptive joint function is constructed by using skewness and kurtosis, which is used as a particle swarm optimization algorithm fitness function, and a function error characteristic change rate and a gradient dynamic adjustment optimization strategy are integrated to realize intelligent switching of search mode and accurate inversion of magnetic circuit defect parameters. Finally, the inversion parameters are compared with the benchmark parameters to complete defect diagnosis. The application has mechanism interpretability and inversion accuracy, realizes online and quantitative diagnosis of defects, adapts to the intelligent demand of power grids, and can be effectively applied to online monitoring and diagnosis of oil-immersed power transformer magnetic circuit defects.

Description

Technical Field

[0001] This invention relates to the field of transformer condition monitoring and defect identification technology, and in particular to a transformer defect identification method based on statistical feature inversion and mechanism constraints. Background Technology

[0002] Transformers are core equipment in the power grid transmission and distribution process. As a core component of transformers, the magnetic circuit system is highly susceptible to defects such as core magnetic permeability degradation, core loosening, increased air gap, and magnetic saturation distortion. These defects can lead to electromagnetic field distortion, increased losses, and localized overheating. In mild cases, they can affect transformer operating efficiency; in severe cases, they can cause insulation breakdown, equipment burnout, or even power grid outages.

[0003] Traditional methods for detecting magnetic circuit defects in transformers largely rely on qualitative judgment of electrical quantities and offline testing, which suffers from low sensitivity, weak anti-interference ability, and difficulty in quantitative identification, failing to meet the needs of intelligent and online operation and maintenance of power grids. Existing data-driven fault identification methods are mostly pure data fitting, lacking physical mechanism constraints, resulting in poor model generalization and insufficient accuracy. To address these industry pain points, this technical solution uses PINN as a high-precision mechanism-driven forward modeling basis, embedding the physical laws of transformer magnetic circuit-electromagnetic field coupling to construct a skewness and kurtosis joint statistical inversion model. Through core algorithm optimization, it achieves efficient, accurate, and quantitative identification of transformer magnetic circuit defects. Summary of the Invention

[0004] This invention aims to address the problems of poor generalization, low diagnostic accuracy, and weak interpretability in existing transformer magnetic circuit defect identification technologies, where data-driven models suffer from poor physical mechanism constraints and traditional detection methods struggle to quantitatively identify defects. Specifically, pure data fitting models are prone to black-box predictions, and diagnostic conclusions violate electromagnetic physics laws. To achieve the above objectives, this invention provides a transformer defect identification method based on statistical feature inversion and mechanistic constraints, comprising:

[0005] Step 1: Collect transformer operating data, process the raw data, and construct a feature set;

[0006] Step 2: Construct a forward model of the transformer's magnetic circuit by embedding electromagnetic field physical equation constraints into the physical information neural network;

[0007] Step 3: Construct the inversion objective function KSE using kurtosis and skewness, and integrate the error characteristic change rate and gradient in the function into the particle swarm optimization algorithm to dynamically adjust the optimization strategy. The magnetic circuit defect parameters are then inverted through the particle swarm optimization algorithm.

[0008] Step 4: Compare the magnetic circuit parameters obtained from the inversion with the reference parameters to complete the diagnosis of transformer magnetic circuit defects.

[0009] Optionally, in step one, real-time operating data is collected through a transformer online monitoring sensor array. This operating data includes electrical and fault characteristics, such as three-phase voltage, three-phase current, active power, reactive power, and partial discharge. The raw collected data is then sequentially processed through time alignment, noise reduction, feature extraction, and normalization to construct a feature set Y. obs .

[0010] Optionally, in step two, the physical information neural network is used to construct the mapping relationship between the transformer magnetic circuit parameters and the monitoring data, so as to obtain the predicted value of the transformer data under specific magnetic circuit parameters. The input layer of the physical information network includes the core permeability degradation coefficient, air gap reluctance coefficient, magnetic saturation nonlinearity coefficient, and core loosening coefficient, and the output layer is the predicted value Y of the collected data. pre (θ), Y pre The dimension of (θ) is the same as the feature set Y formed by the collected data in step one. obs Maintain consistency;

[0011] Optionally, the loss function of the physical information neural network is composed of the sum of three parts: data-driven loss, physical constraint loss, and boundary condition loss. The training process of the physical information neural network is to select simulation data of transformer normal operation and different defect states, and after iterative training until the loss function converges, the final magnetic circuit forward model is obtained.

[0012] Optionally, in step two, the construction of the Physical Information Neural Network (PINN) includes utilizing the fitting ability of a neural network while incorporating physical information constraints during the fitting process. The loss function of PINN is a data-driven loss L. data Physical constraint loss L physics and boundary condition loss L boundary The loss function is L Loss =L data +L physics +L boundary L data It is the mean square error (MSE) loss between the measured features and the PINN simulated features, L boundary It is the sum of squared residuals of all embedded transformer boundary conditions. By quantifying the deviation between the model output value and the boundary condition requirements, the model output is constrained to conform to the actual structure and electromagnetic characteristics of the transformer core, windings, and ports. L physics This is the sum of squared residual loss from the multiphysics equations, where the electromagnetic field equations include:

[0013] ;

[0014] ;

[0015] In the formula: φ is the vector magnetic potential; Js is the source current density; μ is the magnetic permeability of the field medium. This is the eddy current region, where a conductive medium exists but there is no passive current. This is the non-eddy current region, which contains a given source current;

[0016] Optionally, in step two, the construction process of PINN is as follows: Simulation data of transformer normal operation and different magnetic circuit defect states are selected, and on-site measured data are added to enrich the training samples. The Adam adaptive optimization algorithm is used to dynamically adjust the learning rate. The dual objective loss function that integrates physical constraints and data matching is used as the optimization direction. The model is continuously iterated until the loss function is stably converged, and finally the PINN magnetic circuit forward model is obtained, which provides support for subsequent inversion calculation.

[0017] Optionally, in step three, the objective function is constructed, where skewness S characterizes the asymmetry of the residual sequence distribution, kurtosis K characterizes the steepness of the residual sequence distribution, and the kurtosis value of the standard normal distribution is 3, which is used as a benchmark for correction. ;

[0018] Where ei is the i-th value in the residual sequence. Let N be the mean of the residual sequence and N be the length of the residual sequence. Since the transformer operates at a 50Hz power frequency, the fundamental component will inherently interfere with kurtosis calculations and mask weak defect signals. Therefore, a power frequency harmonic correction term is introduced. To eliminate power frequency interference and improve the function's ability to identify early minor defects, the correction term ranges from 0 to 0.5, fine-tuned according to on-site operating conditions to adapt to transformers of various voltage levels. To balance the identification capabilities of skewness and kurtosis for different magnetic circuit defects, the two indicators are fused and their absolute values ​​are taken to eliminate the influence of positive and negative directions, unify the quantification of defect distortion, and perform baseline and power frequency corrections for kurtosis to align with actual transformer operating conditions. Finally, the weighted adaptive joint error function KSE is obtained.

[0019] ;

[0020] In the formula, These are skewness weighting coefficients. The kurtosis weighting coefficient satisfies Engineering standard values To balance and take into account various magnetic circuit defects Correction for the normal distribution baseline. As the power frequency harmonic correction term, KSE, as a function constructed from the skewness and kurtosis of the residual sequence, acts as the fitness function of the particle swarm algorithm during the inversion process. It evaluates the quality of defect parameters by quantifying the distribution difference between simulated features and measured features. The smaller the value, the closer the parameter is to the true state.

[0021] Optionally, in step three, the particle swarm optimization includes using the KSE error characteristic change rate λ and gradient g as feedback parameters to adjust the PSO algorithm parameters to achieve intelligent switching between global and local search. λ reflects the optimization convergence speed, determining whether the algorithm is currently in the global or local search stage, dynamically adjusting the inertia weight to balance search breadth and accuracy, and g determines whether the algorithm has fallen into a local optimum trap. When the gradient approaches zero, particle mutation is triggered to escape local extrema, avoiding premature convergence and obtaining incorrect parameters. ;

[0022] A smaller KSE value indicates a higher degree of fit between simulated and measured features, higher particle fitness, and parameters that are closer to the actual defect state. Conversely, a larger KSE value indicates poorer particle quality and parameters that deviate from the actual fault. Through the KSE value, the PSO algorithm can clearly distinguish between high-quality and low-quality particles, establish the optimization direction, avoid blind searching, and guide the particle swarm to iterate towards the actual defect parameters. The KSE error change rate is used to determine the optimal particle size distribution. With gradient magnitude g, the inertia weight and learning factor are dynamically adjusted, and global search and local search modes are intelligently switched to balance the breadth and accuracy of optimization. In the global search stage ( > 0.1): The KSE error decreases rapidly, indicating the algorithm is in a large-scale optimization phase and has not yet locked onto the optimal interval. At this point, increasing the inertia weight is appropriate. To a value of 0.8~0.9, the particle search range is broadened, global search capability is improved, the entire parameter space is traversed quickly, and the globally optimal solution is avoided being missed. This allows for a more refined local search phase (0.01 < ...). ≤0.1): The KSE error decreases slowly, the algorithm has locked the optimal parameter range, and enters the fine-tuning optimization stage. At this time, reduce the inertia weight. Increase the cognitive learning factor c1 to 2.0-2.5 to 0.4-0.6, reduce the particle flight step size, focus on the optimal region to fine-tune parameters, improve the accuracy of defect inversion, and jump out of the stage ( ≤0.01 and g < 0.001): The KSE error tends to be stable and the gradient approaches zero. The algorithm is trapped in a local optimum. At this time, the particle local mutation operation is triggered to apply a small range of random perturbation to the global optimum particle, break the extreme value constraint, let the algorithm jump out of the local optimum, and restart the effective optimization.

[0023] Optionally, in step three, the particle swarm optimization includes determining the optimization variables: each particle represents a set of magnetic circuit defect parameters to be inverted, denoted as... It includes four core defect parameters: core magnetic permeability degradation coefficient, core air gap thickness, magnetic saturation coefficient, and core loosening displacement. The particle swarm size is set: the number of particles Np is set to 20~30. Too few particles are prone to getting trapped in local optima, while too many particles will increase the computation time. This range takes into account both optimization stability and computational efficiency.

[0024] Particle swarm velocity update formula ,

[0025] In the formula: Let be the velocity of the i-th particle in the (k+1)-th generation; For adaptive inertia weights; Use random numbers between 0 and 1 to increase the randomness of the optimization process; Let i be the position of the i-th particle in the k-th generation.

[0026] Particle position update formula ;

[0027] Optionally, in step four, the qualitative diagnosis of defects includes obtaining benchmark parameters based on the equipment's factory parameters, measured values ​​of healthy transformers of the same model, and industry standard values. The core benchmark parameter is the core magnetic permeability degradation coefficient benchmark value. Core air gap thickness reference value Magnetic saturation coefficient reference value Reference value of core loosening displacement Based on the inversion parameters and benchmark parameters, a preliminary diagnosis is made. Excessive core magnetic permeability degradation coefficient indicates aging of the silicon steel sheets, insulation damage, and decreased magnetic properties. Excessive core air gap thickness indicates core loosening, loose clamping bolts, and increased lamination gaps. Excessive magnetic saturation coefficient indicates localized magnetic saturation, over-excitation, and localized short circuits in the core. Excessive core loosening displacement indicates core structural displacement and failure of the clamping device. The deviation of the inversion parameters from the benchmark values ​​is calculated to quantify the severity of the defects and classify them into levels, forming a complete diagnostic conclusion.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] By integrating statistical feature inversion with physical mechanism constraints, the model output is forced to conform to the electromagnetic physical laws of transformers, effectively eliminating erroneous judgments that violate the operating logic of the magnetic circuit. The dual constraints of electromagnetic field equations and boundary conditions ensure a high degree of consistency between qualitative defect diagnosis and quantitative parameters. The dynamic optimization strategy and weighted adaptive joint function form multiple accurate inversion closed loops, significantly improving the stability and accuracy of defect identification. Attached Figure Description

[0030] Figure 1 This is a flowchart of the transformer defect identification method based on statistical feature inversion and mechanism constraints provided by the present invention. Detailed Implementation

[0031] Example 1

[0032] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments:

[0033] This patent discloses a transformer defect identification method based on statistical feature inversion and mechanistic constraints, applicable to online monitoring and diagnosis of transformer magnetic circuit defects. The method first collects data such as three-phase voltage, current, and partial discharge from the transformer through online monitoring. A standardized feature set is constructed through time alignment, denoising, feature extraction, and normalization preprocessing. Then, based on a physical information neural network, electromagnetic field physical equation constraints are embedded, and a loss function is constructed by fusing data-driven loss, physical constraint loss, and boundary condition loss. A magnetic circuit forward model is obtained through iterative training. A weighted adaptive joint error function (KSE) is constructed using skewness and kurtosis corrected by a normal distribution benchmark and power frequency harmonics, and this function is used as the fitness function for a particle swarm optimization algorithm. Simultaneously, a dynamic adjustment optimization strategy based on the KSE error feature change rate and gradient is incorporated to achieve intelligent switching between global search, local fine search, and local mutation, effectively avoiding local optima and accurately inverting core magnetic circuit defect parameters such as the core permeability degradation coefficient and air gap thickness. Finally, the inversion parameters are compared with the factory-made, same-model health equipment and industry standard benchmark parameters. Based on the type of parameter exceeding the standard, qualitative and quantitative diagnosis of magnetic circuit defects such as aging of iron core silicon steel sheets, iron core loosening, and local magnetic saturation is completed.

[0034] The feature set construction described in the above scheme involves collecting real-time operating data of oil-immersed transformers through an online transformer monitoring sensor array. This data covers two main categories: electrical characteristics and fault characteristics. The original collected data undergoes four main steps: time alignment, noise reduction, feature extraction, and normalization, to construct the feature set Y. obs ;

[0035] The above scheme describes the construction of a forward model based on a physical information neural network. The physical information neural network is used to construct the mapping relationship between transformer magnetic circuit parameters and monitoring data, obtaining predicted values ​​of the transformer's collected data under specific magnetic circuit parameters. The input layer of the physical information network includes the core permeability degradation coefficient, air gap reluctance coefficient, magnetic saturation nonlinearity coefficient, and core loosening coefficient. The output layer is the predicted value Y of the collected data. pre (θ), Y pre The dimension of (θ) is the same as the feature set Y formed by the collected data in step one. obs Maintain consistency;

[0036] In the above steps, the construction of the physical information neural network includes PINN, which utilizes the fitting ability of neural networks while incorporating physical information constraints during the fitting process. This avoids a pure data black box model and enhances the physical meaning and interpretability of the data. The loss function of PINN is a data-driven loss L. data Physical constraint loss L physics and boundary condition loss L boundary The loss function is L Loss =L data +L physics +L boundary L data Mean squared error (MSE) loss between measured and PINN simulated features is used to quantify the data fitting accuracy. L boundary It is the sum of squared residuals of all embedded transformer boundary conditions. By quantifying the deviation between the model output value and the boundary condition requirements, it constrains the model output to conform to the actual structure and electromagnetic characteristics of the transformer core, windings, and ports. Bn1, Bn2: Normal components of magnetic flux density on both sides of the boundary; Ht1, Ht2: Tangential components of magnetic field strength on both sides of the boundary. This ensures the forward model conforms to the actual transformer structure boundary, improving the reliability of defect identification. Lossphy is the sum of squared residuals from the multiphysics equations, quantifying the degree to which the model output satisfies the laws of embedded electromagnetic fields, and is the core guarantee of the model's reliable mechanism. The electromagnetic field equations include:

[0037] ;

[0038] ;

[0039] In the formula: φ is the vector magnetic potential; Js is the source current density; μ is the magnetic permeability of the field medium. This is the eddy current region, where a conductive medium exists but there is no passive current. This is the non-eddy current region, which contains a given source current;

[0040] The construction process of the Physical Information Neural Network (PINN) in the above steps is as follows: Simulation data of transformer normal operation and different magnetic circuit defect states are selected, and on-site measured data are added to enrich the training samples. The Adam adaptive optimization algorithm is used to dynamically adjust the learning rate. The dual objective loss function that integrates physical constraints and data matching is used as the optimization direction. The model is continuously iterated until the loss function is stably converged. The resulting high-precision PINN magnetic circuit forward model can achieve accurate mapping from magnetic circuit defect parameters (magnetic permeability degradation coefficient, air gap thickness, magnetic saturation coefficient, etc.) to transformer electromagnetic response parameters (electrical parameters, partial discharge characteristics, magnetic field distribution), providing a reliable forward modeling foundation for subsequent KSE-PSO-based magnetic circuit defect parameter inversion calculation.

[0041] The objective function described in the above scheme is constructed as follows: Skewness S characterizes the asymmetry of the residual sequence distribution. During normal transformer operation, the residuals approximately follow a symmetrical distribution, and the skewness value approaches 0. When magnetic circuit defects such as core loosening, increased air gap, or winding asymmetry occur, the residual distribution shows a significant shift, and the absolute value of the skewness increases significantly, enabling sensitive detection of asymmetric defects. Kurtosis K characterizes the steepness of the residual sequence distribution. The kurtosis value of the standard normal distribution is 3, which is used as a benchmark for correction. During normal transformer operation, the residual distribution is smooth, and the kurtosis is close to the benchmark value. When defects such as core permeability degradation, magnetic saturation distortion, or local short circuits occur, the signal exhibits sharp spikes, and the kurtosis value deviates significantly from the benchmark, enabling accurate identification of magnetic circuit distortion defects.

[0042] ;

[0043] Among them, e i Let i be the i-th value in the residual sequence. Let N be the mean of the residual sequence and N be the length of the residual sequence. Since the transformer operates at a 50Hz power frequency, the fundamental component will inherently interfere with kurtosis calculations and mask weak defect signals. Therefore, a power frequency harmonic correction term is introduced. To eliminate power frequency interference and improve the function's ability to identify early minor defects, the correction term ranges from 0 to 0.5, fine-tuned according to on-site operating conditions to adapt to transformers of various voltage levels. To balance the identification capabilities of skewness and kurtosis for different magnetic circuit defects, the two indicators are fused and their absolute values ​​are taken to eliminate the influence of positive and negative directions, unify the quantification of defect distortion, and perform baseline and power frequency corrections for kurtosis to align with actual transformer operating conditions. Finally, the weighted adaptive joint error function KSE is obtained. In the formula, These are skewness weighting coefficients. The kurtosis weighting coefficient satisfies Engineering standard values It balances and takes into account various magnetic circuit defects; Correction for the normal distribution baseline. As a power frequency harmonic correction term, KSE, as a function constructed from the skewness and kurtosis of the residual sequence, acts as the fitness function of the particle swarm optimization algorithm during the inversion process. It evaluates the quality of defect parameters by quantifying the distribution difference between simulated and measured features. The smaller the value, the closer the parameter is to the true state. Compared with the traditional mean square error, KSE is more sensitive to signal distortion, asymmetry and impact features caused by transformer magnetic circuit defects. It can effectively identify early weak faults and has stronger anti-noise and anti-power frequency interference capabilities. It can take into account the identification needs of asymmetric and abrupt defects. It can not only provide a clear guide for particle swarm iteration optimization and improve search accuracy and convergence speed, but also intuitively reflect the severity of defects, providing a reliable basis for subsequent quantitative fault diagnosis and classification.

[0044] In step three, the particle swarm optimization (PSO) involves using the KSE error characteristic change rate λ and gradient g as feedback parameters to adjust the PSO algorithm parameters, enabling intelligent switching between global and local search. λ reflects the convergence speed, determining whether the algorithm is currently in a global or local search phase, and dynamically adjusting the inertia weight to balance search breadth and accuracy. g determines whether the algorithm has fallen into a local optimum trap; when the gradient approaches zero, particle mutation is triggered to escape local extrema, preventing premature convergence and obtaining incorrect parameters. A smaller KSE value indicates a higher degree of fit between simulated and measured features, higher particle fitness, and parameters that are closer to the actual defect state. Conversely, a larger KSE value indicates poorer particle quality and parameters that deviate from the actual fault. Through the KSE value, the PSO algorithm can clearly distinguish between high-quality and low-quality particles, establish the optimization direction, avoid blind searching, and allow the particle swarm to iterate towards the actual defect parameters. The KSE error change rate is used to determine the optimal particle size distribution. With gradient magnitude g, the inertia weight and learning factor are dynamically adjusted, and global search and local search modes are intelligently switched to balance the breadth and accuracy of optimization. In the global search stage ( > 0.1): The KSE error decreases rapidly, indicating the algorithm is in a large-scale optimization phase and has not yet locked onto the optimal interval. At this point, increasing the inertia weight is appropriate. To a value of 0.8~0.9, the particle search range is broadened, global search capability is improved, the entire parameter space is traversed quickly, and the globally optimal solution is avoided being missed. This allows for a more refined local search phase (0.01 < ...). ≤0.1): The KSE error decreases slowly, the algorithm has locked the optimal parameter range, and enters the fine-tuning optimization stage. At this time, reduce the inertia weight. Increase the cognitive learning factor c1 to 2.0-2.5 to 0.4-0.6, reduce the particle flight step size, focus on the optimal region to fine-tune parameters, improve the accuracy of defect inversion, and jump out of the stage ( ≤0.01 and g<0.001): The KSE error tends to be stable and the gradient approaches zero. The algorithm is trapped in a local optimum. At this time, the particle local mutation operation is triggered to apply a small range of random perturbation to the global optimum particle, break the extreme value constraint, let the algorithm jump out of the local optimum, and restart the effective optimization.

[0045] In step three, the particle swarm optimization algorithm includes determining the optimization variables: each particle represents a set of magnetic circuit defect parameters to be inverted, denoted as... It includes four core defect parameters: core magnetic permeability degradation coefficient, core air gap thickness, magnetic saturation coefficient, and core loosening displacement. The particle swarm size is set: the number of particles Np is 20-30. Too few particles can easily lead to local optima, while too many will increase computation time. This range balances optimization stability and computational efficiency. The particle swarm velocity update formula... In the formula: Let be the velocity of the i-th particle in the (k+1)-th generation; For adaptive inertia weights; Use random numbers between 0 and 1 to increase the randomness of the optimization process; Let i be the position of the i-th particle in the k-th generation.

[0046] Particle position update formula The algorithm iteration terminates under any of the following conditions: the fitness value KSE(θ) of the dynamic adaptive statistical inversion function corresponding to the globally optimal particle is less than or equal to the preset convergence threshold T (T=0.005 for light load conditions, T=0.01 for medium / heavy load conditions), or the number of iterations reaches the preset maximum number of iterations T. max If, after 500 iterations or 100 consecutive iterations, the fitness value of the globally optimal particle fluctuates by less than 0.0001 and no local optimum correction mechanism is triggered, the iteration terminates, and the set of magnetic circuit defect parameters corresponding to the current globally optimal particle is output. Complete the inversion calculation;

[0047] In step four, the qualitative diagnosis of defects includes obtaining benchmark parameters based on the equipment's factory parameters, measured values ​​of healthy transformers of the same model, and industry standard values. The core benchmark parameter is the core magnetic permeability degradation coefficient benchmark value. Core air gap thickness reference value Magnetic saturation coefficient reference value Reference value of core loosening displacement Diagnosis is achieved based on inversion parameters and baseline parameters. Excessive core magnetic permeability degradation coefficient indicates aging of the core silicon steel sheets, insulation damage, and decreased magnetic properties. Excessive core air gap thickness indicates core loosening, loose clamping bolts, and increased lamination gap. Excessive magnetic saturation coefficient indicates local magnetic saturation, over-excitation, and local short circuit in the core. Excessive core loosening displacement indicates core structural displacement and failure of the clamping device. Based on the corresponding rules of the above parameters and defect types, and combined with the specific type of the excessive parameters, the specific name, cause, and attribution of the transformer magnetic circuit defect are determined, completing the qualitative diagnosis of the defect. Based on qualitative diagnosis, quantitative defect diagnosis is carried out simultaneously. By calculating the relative deviation between the actual inversion value of the exceeding parameter and the benchmark value, the severity of the defect is quantified. The formula for calculating the relative deviation is: (parameter inversion value - parameter benchmark value) / parameter benchmark value × 100%. According to different ranges of deviation, the severity level of the defect is divided, and the development stage and impact of the defect are accurately defined. Finally, the qualitative diagnosis results are combined with the quantitative data of the quantitative diagnosis to form a complete diagnostic result that includes defect type, fault cause, severity level, core exceeding parameters and deviation range. This provides accurate, comprehensive and practical judgment basis for the subsequent operation and maintenance decision-making, maintenance plan formulation and risk prevention of the transformer.

[0048] The above description is merely an illustration of the present invention. Those skilled in the art can make various modifications or additions to the specific structure described, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, and all such modifications or additions should fall within the protection scope of the present invention.

Claims

1. A transformer defect identification method based on statistical feature inversion and mechanism constraints, characterized in that, Includes the following steps: Step 1: Collect transformer operating data, process the raw data, and construct a feature set; Step 2: Embed electromagnetic field physical equation constraints into the physical information neural network to construct the forward model of the transformer's magnetic circuit; Step 3: Construct the objective function KSE for inversion using kurtosis and skewness, and use the function KSE to drive the particle swarm optimization algorithm to invert the magnetic circuit defect parameters; Step 4: Compare the magnetic circuit parameters obtained from the inversion with the reference parameters to complete the diagnosis of transformer magnetic circuit defects.

2. The method according to claim 1, characterized in that, In step one, real-time operating data is collected through a transformer online monitoring sensor array. The operating data includes electrical and fault characteristics such as three-phase voltage, three-phase current, active power, reactive power, and partial discharge. The original collected data is then processed sequentially with time alignment, noise reduction, feature extraction, and normalization to construct a feature set.

3. The method according to claim 1, characterized in that, In step two, the physical information neural network is used to construct the mapping relationship between the transformer magnetic circuit parameters and the collected data. The input layer of the network consists of the core magnetic permeability degradation coefficient, the air gap magnetic reluctance coefficient, the magnetic saturation nonlinearity coefficient, and the core loosening coefficient, while the output layer consists of the theoretical values ​​of the collected data.

4. The method according to claim 1 or 3, characterized in that, In step two, the loss function of the physical information neural network is composed of the sum of three parts: data-driven loss, physical constraint loss, and boundary condition loss. The training process of the physical information neural network involves selecting simulation data of transformers under normal operation and different defect states, and iteratively training until the loss function converges to obtain the final magnetic circuit forward model.

5. The method according to claim 1, characterized in that, The inversion objective function KSE, constructed from skewness and kurtosis, is as follows: Skewness S characterizes the asymmetry of the residual sequence distribution, and kurtosis K characterizes the steepness of the residual sequence distribution. Kurtosis is corrected for both normal distribution and power frequency harmonics. The absolute values ​​of the skewness and corrected kurtosis are then combined, with the power frequency harmonic correction term ranging from 0 to 0.

5. The sum of the skewness weighting coefficient and the kurtosis weighting coefficient is 1; and the sum of the two weighting coefficients is also 1, resulting in the KSE weighted adaptive joint error function. , .

6. The method according to claim 1, characterized in that, In step three, the error characteristic change rate and gradient of KSE are incorporated into the particle swarm algorithm as feedback parameters to realize the intelligent switching between global search and local search. At the same time, the particle fitness is judged based on the value of KSE, and the smaller the KSE value, the higher the corresponding particle fitness.

7. The method according to claim 6, characterized in that, The rules for switching search modes in the Particle Swarm Optimization (PSO) algorithm combined with KSE are as follows: When the error feature change rate of KSE is greater than 0.1, the algorithm enters the global search phase and increases the inertia weight to broaden the particle search range; when the error feature change rate of KSE is between 0.01 and 0.1, the algorithm enters the local fine search phase, decreases the inertia weight and increases the cognitive learning factor to reduce the particle flight step size for refined optimization; when the error feature change rate of KSE is less than or equal to 0.01 and the gradient is less than 0.001, the local mutation operation of the particles is triggered, and a small-scale random perturbation is applied to the globally optimal particle to help the algorithm escape the local optimum trap.

8. The method according to claim 1, characterized in that, The transformer defect diagnosis is based on the equipment's factory parameters, measured values ​​of healthy equipment of the same model, and industry standard values. The benchmark values ​​of core magnetic permeability degradation coefficient, core air gap thickness, magnetic saturation coefficient, and core loosening displacement are obtained. The inversion parameters are compared with the benchmark values, and the type and cause of transformer magnetic circuit defects are determined according to the correspondence. The deviation of the inversion parameters from the benchmark values ​​is calculated to quantify the severity of the defects and classify them into levels, forming a complete diagnostic conclusion.

Images