An oil-immersed transformer insulation aging evaluation system based on multi-parameter monitoring

Abstract

The application discloses an oil-immersed transformer insulation aging evaluation system based on multi-parameter monitoring, relates to the technical field of power equipment state monitoring and intelligent evaluation, and comprises a data acquisition module, which acquires insulation aging data of an oil-immersed transformer and pre-processes the insulation aging data; an aging evaluation module, which dynamically optimizes the weights of various monitoring indexes; obtains high-dimensional multi-modal features, calculates various-dimensional coupling correlation coefficients, corrects the high-dimensional multi-modal features, and outputs an insulation state fusion feature vector; calculates an initial insulation aging index, and corrects errors of the initial insulation aging index to obtain an insulation aging comprehensive index; and an early warning module, which determines an insulation aging grade according to the insulation aging comprehensive index in combination with preset multi-stage early warning thresholds and triggers an abnormal early warning. The application solves the technical problems of early hidden trouble missed detection and missing of dimensional coupling in the prior art, and greatly improves the equipment operation reliability and the intelligent operation and maintenance level.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and intelligent assessment technology, and more specifically, to an insulation aging assessment system for oil-immersed transformers based on multi-parameter monitoring. Background Technology

[0002] Oil-immersed transformers are core and critical equipment in power transmission and distribution systems. Their insulation system, composed of insulating oil, oil-impregnated paper, partitions, and support bars, is subjected to multiple stress coupling effects from electricity, heat, mechanical vibration, moisture, and oxidation over long periods. This makes them highly susceptible to insulation aging and performance degradation, which can lead to faults such as partial discharge, winding breakdown, and equipment burnout, directly impacting the safety, stability, and reliability of the power grid. According to power equipment failure statistics, transformer insulation aging failure is the leading cause of unplanned transformer outages, accounting for over 60% of all equipment failures.

[0003] Shortcomings of existing technology:

[0004] Traditional transformer insulation assessment relies heavily on dissolved gas detection in the oil, which can only identify severe faults such as mid-to-late stage overheating and gas-generating discharges. It cannot detect early aging defects without gas generation, such as slight dampness, weak heat accumulation, and latent weak discharges. However, most early insulation problems in transformers are induced by dampness, leading to widespread problems of missed early faults and delayed assessments in traditional assessment systems. Furthermore, traditional assessments often use simple weighted fusion of various monitoring parameters, assuming that each dimension ages independently, ignoring the nonlinear synergistic degradation mechanism between dampness, thermal damage, gas generation, and electrical defects. This fails to uncover latent coupled aging characteristics, and the feature representation does not accurately reflect the true complex aging patterns of transformers.

[0005] To address the above problems, this invention proposes a solution. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an oil-immersed transformer insulation aging assessment system based on multi-parameter monitoring. By constructing a multi-parameter, multi-dimensional, and mechanism-coupled intelligent assessment system for oil-immersed transformer insulation aging, the problems mentioned in the background art are solved.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] An insulation aging assessment system for oil-immersed transformers based on multi-parameter monitoring includes:

[0009] The data acquisition module collects insulation aging data of oil-immersed transformers and preprocesses the insulation aging data to build a standardized insulation parameter database.

[0010] The aging assessment module includes a weight optimization unit, a multi-parameter fusion unit, and an aging quantification assessment unit.

[0011] The weight optimization unit uses an entropy weight-AHP coupling algorithm to dynamically optimize the weights of each monitoring indicator.

[0012] The multi-parameter fusion unit uses a tensor fusion algorithm to obtain high-dimensional multimodal features, and corrects the high-dimensional multimodal features by calculating the coupling correlation coefficients of each dimension, and outputs an insulation state fusion feature vector.

[0013] The aging quantification assessment unit calculates the initial insulation aging index based on the optimal index weight and the fusion feature of insulation state, and introduces the ResSSAE residual sparse autoencoder network to correct the error of the initial insulation aging index, thereby obtaining the comprehensive insulation aging index.

[0014] The early warning module determines the insulation aging level based on the comprehensive insulation aging index and preset multi-level early warning thresholds, and triggers abnormal early warnings based on the insulation aging level.

[0015] In a preferred embodiment, the insulation aging data includes dissolved gas parameters in oil, characteristic parameters of the insulating medium, temperature operating condition parameters, and defect electrical characteristic parameters.

[0016] In a preferred embodiment, the process for obtaining the characteristic parameters of the insulating medium is as follows:

[0017] An integrated insulating medium detection sensor is installed at the transformer sampling oil circuit interface to collect parameters such as the moisture content, dielectric strength, dielectric loss factor, and frequency domain dielectric spectrum characteristics of the insulating oil.

[0018] Data from sensor calibration and transient fluctuations in the oil circuit are removed, while valid data on the characteristics of the medium under steady-state conditions are retained.

[0019] Environmental compensation corrections were made to the parameters of insulating oil micro-water content, dielectric strength, dielectric loss factor, and frequency domain dielectric spectrum characteristic parameters, taking into account ambient temperature and humidity.

[0020] Based on the threshold values ​​of the insulation evaluation standard for power equipment, the micro-water content, dielectric strength, dielectric loss factor, and frequency domain dielectric spectrum characteristic parameters of insulating oil are normalized in intervals to unify the data dimensions.

[0021] The corrected and normalized dielectric characteristic data are stored in chronological order to form a dataset of moisture-induced aging of insulating dielectric.

[0022] In a preferred embodiment, the weight optimization unit uses an entropy-weighted AHP coupling algorithm to dynamically optimize the weights of each monitoring indicator, as follows:

[0023] Based on the four-dimensional monitoring system, a three-level hierarchical evaluation index system is established. The target layer is the aging state of transformer insulation, the criterion layer includes four dimensions: gas aging, medium moisture, heat accumulation damage, and electrical defects, and the index layer consists of the subdivided monitoring parameters corresponding to each criterion layer.

[0024] Construct the criterion layer judgment matrix and solve the dimension weights, and construct the index layer domain judgment matrix according to the dimension and solve the subdivision relative weights;

[0025] Based on the dimensional weights of the criteria layer and the relative weights of the indicator layers under that dimension, the global subjective weights of all monitoring parameters are obtained through a double-layer nested multiplication calculation.

[0026] Obtain the objective entropy weight, combine it with the global subjective weight to construct the optimal combined weight, and obtain the optimal combined weight. The calculation formula is as follows:

[0027] ,

[0028] In the formula, For the optimal combination of weights, For objective entropy weight, This represents the global subjective weight.

[0029] In a preferred embodiment, the process of constructing the criterion layer judgment matrix and solving the dimension weights is as follows:

[0030] For the four dimensions of aging, a pairwise comparison judgment matrix of the fourth-order criterion layer is constructed based on the aging mechanism of electrical insulation using the scaling method;

[0031] The maximum eigenvalue and corresponding eigenvector of the matrix are solved. After normalizing the eigenvector, the weights of the four criteria layers are obtained, namely the weights of gas aging, medium moisture, thermal accumulation damage, and electrical defects. Consistency is then checked based on the maximum eigenvalue.

[0032] In a preferred embodiment, the process of obtaining high-dimensional multimodal features using a tensor fusion algorithm is as follows:

[0033] Construct a four-dimensional insulation aging feature tensor based on four-dimensional normalized time series data. Where T is the number of time-series sampling steps, 4 is the four major monitoring dimensions, K is the number of features in each dimension, and N is the number of samples;

[0034] By introducing an optimal weight vector W to weight and correct the indices of each dimension of the tensor, a weighted tensor is obtained. ,in This refers to tensor multiplication operations;

[0035] The Tucker decomposition algorithm is used to perform third-order modal decomposition on the weighted tensor, decomposing the high-order tensor into a core tensor and factor matrices of each dimension, thus obtaining high-dimensional multimodal features. The decomposition formula is as follows: In the formula For core tensors, For the time-series mode factor matrix, For the dimension modal factor matrix, Here is the characteristic mode factor matrix, where This is used as an identifier for tensor modal products, with numbers 1, 2, and 3 corresponding to temporal modality, dimensional modality, and feature modality, respectively.

[0036] In a preferred embodiment, the high-dimensional multimodal features are corrected by calculating the coupling correlation coefficients of each dimension, and the insulation state fusion feature vector is output as follows:

[0037] A dimensional coupling correlation coefficient matrix is ​​constructed through factor matrix operations, and the co-coupling degree between different monitoring dimensions is calculated using the following formula: ,in The coupling correlation coefficient between the two aging monitoring dimensions, a and b; Covariance operator; This represents the feature vector of the dimensional modality factor corresponding to the a-th aging dimension. Let a and b be the feature vector of the dimensional modality factor corresponding to the b-th aging dimension, where a and b are the dimension indices; This is the variance operator;

[0038] Based on the aging condition mechanism of transformer oil-paper insulation coupling, we distinguish between strongly coupled synergistic degradation dimensions and weakly coupled unrelated dimensions. We set a coupling threshold δ=0.1. If the dimension coupling coefficient Corr(a,b)≥0.1, it is determined to be a strongly coupled dimension combination. If Corr(a,b)<0.1, it is determined to be a weakly coupled dimension combination.

[0039] Constructing the basic gain coefficient Then, an exponential decay suppression function is introduced to construct the final adaptive correction coefficient. , The coupling threshold constant is used; the original dimensional modality factor matrix is ​​corrected element-by-element by the Hadamard operation, as shown in the formula: , This is the corrected dimensional modal feature matrix;

[0040] The high-dimensional multimodal features, after being decomposed, corrected, pooled, compressed, and purified, are input into the Sigmoid nonlinear activation function for normalization mapping, generating a normalized insulation state fusion feature vector. .

[0041] In a preferred embodiment, the aging quantification assessment unit is implemented as follows:

[0042] A weighted comprehensive aging index model is constructed based on the optimal coupling weight W and the fused feature vector F to obtain the initial insulation aging index, as shown in the formula. In the formula This is the initial insulation aging index. The optimal weight for the j-th indicator is... For the corresponding fusion feature quantity;

[0043] A ResSSAE residual sparse autoencoder network is introduced to correct the error of the initial insulation aging index.

[0044] The fused feature values ​​are input into the network, and deep feature extraction is performed through the encoding layer to obtain the theoretical aging value fitted by the deep learning of the network.

[0045] Based on the initial insulation aging index and the theoretical aging value fitted by deep learning, the aging prediction residual is calculated through residual branching. ,in This is the theoretical aging value fitted to the deep learning network. This is the residual correction amount;

[0046] A high-precision comprehensive insulation aging index is calculated based on the aging prediction residual and an adaptive correction coefficient. , This is the adaptive correction coefficient.

[0047] In a preferred embodiment, the aging warning module is implemented as follows:

[0048] Obtain the comprehensive insulation aging index, simultaneously collect the aging state evolution characteristics under continuous time sequence, and construct a dual early warning basic data that combines static aging state and dynamic aging trend.

[0049] Based on the pre-set multi-level insulation aging warning threshold range, the real-time insulation aging comprehensive index is matched and compared. The corresponding insulation aging level is divided according to the threshold range in which the index is located, and four insulation states are distinguished step by step: healthy equipment operation, slight latent aging, moderate abnormal aging, and serious fault risk.

[0050] Based on different insulation aging levels, a corresponding early warning mechanism is matched. No early warning is triggered for equipment in a healthy state; for slightly aging state, a low-level early warning is triggered to record early hidden deterioration characteristics of the equipment; for moderate aging state, a medium-level early warning is triggered to determine that the equipment has a potential for continuous deterioration; for severely aging state, a high-level emergency early warning is triggered to determine that the insulation performance of the equipment has significantly deteriorated and there is a high risk of failure, and an emergency operation and maintenance alarm is pushed out in a timely manner.

[0051] By combining the continuous change pattern of the time-series insulation aging comprehensive index with the predicted aging trend, the static threshold judgment result is dynamically corrected.

[0052] Once an early warning is triggered, the abnormal causes are traced back to their source by combining multi-dimensional aging coupling and correlation characteristics, and standardized early warning information including aging level, degradation risk, abnormal source tracing results and targeted operation and maintenance strategies is generated.

[0053] The technical effects and advantages of the present invention, a multi-parameter monitoring-based insulation aging assessment system for oil-immersed transformers, are as follows:

[0054] 1. This invention utilizes online monitoring of multiple characteristics of the insulating medium to specifically capture early-stage slight dampness, weak heat accumulation, and latent discharge aging defects that traditional oil chromatography cannot identify. It covers the entire lifecycle aging characteristics from early latent degradation to mid-to-late-stage severe faults, completely solving the problem of missing early-stage hidden dangers. Through joint compensation and correction of environmental temperature and humidity, and the removal of invalid data, it eliminates spurious parameter fluctuations caused by external operating conditions, accurately restoring the true aging state of the transformer's internal oil-paper insulation. This constructs a high-precision standardized time-series database, providing a reliable data foundation for subsequent evaluation. Employing entropy weighting-AHP dual-layer coupling weighting, it balances the subjective rationality of the insulation aging mechanism with the objective differences in monitoring data, avoiding the subjectivity and one-sidedness of a single weighting method. It dynamically adapts the weight distribution of each indicator under different aging conditions, significantly reducing the bias in multi-parameter fusion evaluation.

[0055] 2. This invention achieves structured fusion of multimodal data through high-order tensor Tucker decomposition. It distinguishes between strong and weak coupling conditions using dimensional coupling coefficients, enhances the gain of real collaborative aging features, and suppresses the weight of spurious interference features. This effectively uncovers multi-dimensional nonlinear coupling degradation patterns, resulting in feature representations that better align with the composite aging mechanism of transformers. A ResSSAE residual sparse autoencoder network is used to correct residual errors in the initial aging index, compensating for the systematic bias of traditional linear evaluation models, effectively eliminating data noise and model fitting errors, and outputting a high-precision comprehensive insulation aging index. The quantification results better reflect the actual aging level of the equipment. A dual early warning mechanism combining static threshold grading and dynamic trend correction can predict accelerated degradation trends in advance and provide early warning intervention, avoiding critical missed detections and delayed alarms. Simultaneously, it supports reverse tracing of abnormal causes, outputting targeted operation and maintenance strategies, achieving proactive prevention and control of transformer insulation status, and tiered operation and maintenance, significantly improving equipment reliability and intelligent operation and maintenance levels. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the structure of an oil-immersed transformer insulation aging assessment system based on multi-parameter monitoring according to the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0058] Example 1, Figure 1 This invention presents an insulation aging assessment system for oil-immersed transformers based on multi-parameter monitoring.

[0059] The data acquisition module collects insulation aging data of oil-immersed transformers and preprocesses the insulation aging data to build a standardized insulation parameter database.

[0060] The insulation aging data includes parameters of dissolved gases in oil, characteristic parameters of the insulating medium, temperature operating conditions, and electrical characteristic parameters of defects.

[0061] High-temperature pyrolysis, arc discharge, and localized overheating of transformer oil-paper insulation can all break down cellulose and oxidize the insulating oil, generating characteristic gases. Obtaining dissolved gas parameters in the oil is used to determine whether the insulation has suffered substantial damage or severe deterioration. The process is as follows: An online oil chromatography sensor is installed in the transformer's oil circulation pipeline. Utilizing the principle of vacuum degassing, dissolved gases in the insulating oil are continuously extracted. Real-time raw data on the volume concentrations of seven characteristic gases—hydrogen, methane, ethane, ethylene, acetylene, carbon monoxide, and carbon dioxide—are collected. The sampling period is set to 5 minutes / sampling, and the sampling time sequence label is recorded simultaneously. Based on... The normal operating gas threshold range of the transformer is used to eliminate out-of-range sudden data and invalid extreme value data caused by instantaneous sensor jitter, as well as interference data corresponding to equipment start-up and shutdown and instantaneous load impact. A moving average filtering algorithm is used to smooth the effective gas concentration time series data to eliminate high-frequency noise caused by substation electromagnetic interference. An extreme value normalization method is used to map various gas concentration data to the [0,1] interval to eliminate the differences in the dimensions and numerical spans of different gas parameters. The processed gas characteristic data is bound with time series labels and entered into a standardized insulation parameter database to form an oil chromatography aging time series dataset.

[0062] Dissolved gas parameters in oil can only detect mid-to-late stage aging. Slight moisture absorption, slight heat accumulation, and weak discharge do not produce gas. Using gas parameters alone will miss early-stage problems. Furthermore, 80% of early aging in transformer insulation is caused by moisture. Therefore, it is necessary to obtain characteristic parameters of the insulating medium to specifically capture early moisture-induced aging that traditional gas monitoring cannot fully identify. The process is as follows: Install an integrated insulating medium detection sensor at the transformer sampling oil circuit interface to collect online parameters such as the moisture content of the insulating oil, dielectric strength, dielectric loss factor, and frequency domain dielectric spectrum characteristics. The sampling period is set to 10 minutes / time, and the device operating condition tag is simultaneously bound. Data during sensor calibration and jump data caused by brief fluctuations in the oil circuit are discarded, retaining effective medium characteristic data under steady-state conditions. Environmental compensation correction is applied to the medium parameters based on ambient temperature and humidity. The core compensation formula is: In the formula These are the true values ​​of the standard operating condition medium parameters after temperature and humidity compensation; The raw medium parameters (trace water content, dielectric loss, dielectric strength, frequency domain dielectric spectrum) are collected in real time by the sensor. , These are the real-time ambient temperature and the real-time ambient relative humidity, respectively. , These are the preset standard reference temperature and standard reference humidity, respectively. This is a temperature compensation coefficient to adapt to the temperature-sensitive characteristics of oil-paper insulation parameters; The humidity interference compensation coefficient is used to offset the pseudo-fluctuations in parameters introduced by the external environmental humidity; to eliminate the interference errors of external temperature and humidity fluctuations on micro-water and dielectric parameters; to normalize parameters such as micro-water content, dielectric loss, and withstand voltage value according to the threshold of the power equipment insulation evaluation standard, and to unify the data dimensions; and to store the corrected and normalized dielectric characteristic data in time sequence to form a dataset of moisture aging of insulating medium.

[0063] Moisture does not necessarily mean gas production. Individual medium parameters cannot determine whether overheating or discharge faults have occurred. Oil-paper insulation follows a 10°C halving rule: for every 10°C increase in temperature, the insulation life is directly halved. Therefore, obtaining temperature operating parameters quantifies the accumulated hidden lifespan loss. The process is as follows: High-precision temperature sensors are deployed at transformer winding hotspots, the top oil chamber, the equipment casing, and the station environment to collect real-time data on winding hotspot temperature, top oil temperature, and ambient temperature around the clock. The sampling period is set to 1 minute / time to capture the heat accumulation process with high density. Instantaneous temperature mutations caused by external interference such as short-term equipment overload, start-up and shutdown of air-cooled equipment, and direct sunlight are removed, retaining long-term steady-state temperature time-series data. Based on the continuous temperature time-series data, the temperature mean, temperature fluctuation variance, and over-temperature accumulation duration are calculated to construct derived characteristic parameters that can characterize insulation thermal aging damage. The original temperature data and the heat accumulation derived characteristic data are uniformly normalized. The processed temperature operating data and thermal damage characteristic data are then stored in a database to form an insulation thermal aging time-series dataset.

[0064] High temperature only creates stress; it doesn't necessarily produce gas or cause moisture, making it impossible to determine the degree of insulation damage on its own. Tiny cracks, bubbles, and breaks within the insulation can cause partial discharge: this discharge doesn't break down equipment or produce large amounts of gas, but it continuously erodes the insulating paper / paperboard, acting as a precursor to sudden short-circuit faults. Therefore, capturing the nascent stage of electrical aging defects by acquiring their electrical characteristic parameters is the earliest indicator of fault warning among all parameters. The acquisition process is as follows: High-frequency partial discharge sensors and grounding current sensors are installed at the transformer winding ends and in the core grounding circuit to continuously collect partial discharge quantity, discharge frequency, and discharge phase characteristics. The sampling period for the core grounding current was set to 5 minutes per sampling. A wavelet denoising algorithm was used to filter out external pulse noise caused by substation switch operations and line interference, separating the actual partial discharge signal from the interference signal. The number of effective discharges, average discharge amount, and maximum discharge amount per unit time were statistically analyzed to extract the steady-state amplitude and fluctuation characteristics of the core grounding current, eliminating instantaneous interference pulse data. The electrical defect characteristic parameters were normalized to match the data scale with gas, medium, and temperature parameters. The processed electrical defect characteristic data were then stored in the database according to time-series labels, forming an insulation latent defect characteristic dataset.

[0065] A standardized insulation parameter database was constructed based on oil chromatography aging time series datasets, insulation medium moisture aging datasets, insulation thermal aging time series datasets, and insulation latent defect feature datasets.

[0066] The aging assessment module includes a weight optimization unit, a multi-parameter fusion unit, and an aging quantification assessment unit. It integrates entropy weighted hierarchical analysis, multi-source data fusion algorithm, and deep learning model to calculate the comprehensive aging index of transformer insulation based on insulation aging data and determine the aging level according to the preset grading standard.

[0067] The weight optimization unit uses an entropy weight-AHP coupling algorithm to dynamically optimize the weights of each monitoring indicator.

[0068] Since the insulation aging data includes parameters of dissolved gases in oil, characteristic parameters of insulating medium, temperature operating conditions, and electrical characteristic parameters of defects, the system includes a four-dimensional monitoring system. Based on the four-dimensional monitoring system, a three-level hierarchical evaluation index system is built. The target layer is the aging state of transformer insulation, the criterion layer includes four dimensions: gas aging, medium moisture, thermal accumulation damage, and electrical defects, and the index layer consists of the subdivided monitoring parameters corresponding to each criterion layer. These are the original characteristic variables that directly participate in the data calculation and include all monitoring indicators such as seven types of gas concentration, trace moisture, dielectric loss, temperature, partial discharge, and grounding current.

[0069] First, the criterion-level judgment matrix and dimension weights are constructed. The process is as follows: For the four aging dimensions, a 4th-order criterion-level pairwise comparison judgment matrix is ​​constructed based on the aging mechanism of power insulation using the 1-9 scaling method. The matrix satisfies the reciprocity rule (AHP mandatory requirement) to quantify the contribution importance of different aging dimensions to the overall insulation failure of the transformer. The maximum eigenvalue and corresponding eigenvector of the matrix are solved. After normalizing the eigenvector, the weights of the four criterion-level dimensions are obtained, namely the weights of gas aging dimension, medium moisture dimension, thermal accumulation damage dimension, and electrical defect dimension. A consistency check CR < 0.1 is performed based on the maximum eigenvalue to ensure that the dimension weights are reasonable and effective.

[0070] It should be noted that, for the judgment matrix constructed using the scaling method, gas aging is the core characteristic of substantial insulation cracking failure and has the highest importance; medium moisture and electrical defects are the core causes of early latent aging and have the second highest importance; temperature heat accumulation is a continuous auxiliary damage and has the lowest relative weight, which is completely consistent with the aging failure law of transformer oil-paper insulation.

[0071] Secondly, the index layer domain-specific judgment matrix and the subdivision relative weights are independently constructed according to the dimension and domain. Homogeneous parameters within the same dimension are compared pairwise, while parameters from different dimensions are not compared across domains. The process is as follows:

[0072] Gas aging dimension index layer: A 7-order judgment matrix is ​​constructed for seven characteristic gases: hydrogen, methane, ethane, ethylene, acetylene, carbon monoxide, and carbon dioxide. The gas is assigned a value according to its lethality to the oil-paper insulation. Acetylene and CO have the highest priority, corresponding to solid insulation damage and arc discharge aging; ethylene and hydrogen are next, corresponding to high temperature overheating and early discharge aging; alkane gases have the lowest priority, corresponding to slight thermal aging. The relative weights of each sub-gas parameter within the gas dimension are obtained by solving the problem.

[0073] The indicator layer for the dielectric moisture dimension is as follows: a fourth-order judgment matrix is ​​constructed for the parameters of trace moisture content, dielectric loss, dielectric strength, and frequency domain dielectric spectrum. The values ​​are assigned according to the moisture aging mechanism. Trace moisture is the core cause of insulation aging and has the highest weight, followed by dielectric loss and dielectric strength. The frequency domain dielectric spectrum is used as an auxiliary feature parameter. The relative weights of each parameter within the dielectric dimension are obtained by solving the matrix.

[0074] Thermal cumulative damage dimension index layer: A third-order judgment matrix is ​​constructed for winding hot spot temperature, top oil temperature, and over-temperature accumulation time. The winding hot spot temperature directly determines the aging rate of the insulation paper and has the highest weight. The top oil temperature and over-temperature accumulation time are secondary. The relative weights of the internal parameters of the thermal dimension are obtained by solving.

[0075] Electrical defect dimension index layer: A third-order judgment matrix is ​​constructed for partial discharge quantity, discharge frequency, and core grounding current. Partial discharge directly erodes the insulation structure and has the highest weight, while discharge frequency and grounding current are secondary. The relative weights of the internal parameters of the electrical dimension are obtained by solving.

[0076] Each indicator layer matrix independently completes the eigenvector calculation and consistency check to ensure the rationality and rigor of the layer weights.

[0077] Based on the weights of the four criteria layers—gas aging, medium moisture absorption, thermal accumulation damage, and electrical defects—and the relative weights of each monitoring indicator, and through a double-layer nested multiplication, the importance of macroscopic dimensions is deeply bound to the contribution of microscopic parameters. This yields the global subjective weights of all monitoring parameters, resulting in a subjective weight vector. The calculation formula is: In the formula, For the corresponding criterion layer dimension weights, To further subdivide the relative weights of the indicators under this dimension, Global subjective weight;

[0078] It should be noted that traditional AHP algorithms for transformer insulation assessment all adopt a single-layer flat weighting structure, directly mixing all monitoring parameters into one layer for pairwise comparisons, ignoring the hierarchical relationship of insulation aging. This results in inherent defects such as chaotic indicator hierarchy, inability to distinguish between major category differences and sub-parameter differences, coarse weight accuracy, and poor adaptability. This invention adopts a two-layer nested hierarchical AHP architecture of target layer-criteria layer-indicator layer, constructing a judgment matrix layer by layer, verifying consistency layer by layer, and synthesizing global weights layer by layer. It accurately distinguishes between the "macro-importance of the four aging dimensions" and the "micro-contribution of sub-parameters under the same dimension," and the hierarchical logic closely aligns with the actual aging mechanism of oil-paper insulation.

[0079] Based on the standardized insulation parameter database, a standardized evaluation matrix of n indicators for m samples is constructed. ,in Let be the normalized value of the j-th indicator in the i-th sample group;

[0080] First, calculate the proportion of the i-th sample under the j-th indicator.

[0081] Then, calculate the information entropy of the j-th indicator based on the proportion of the i-th group of samples under the j-th indicator. ,like , then let

[0082] Finally, the objective entropy weight of the j-th indicator is calculated based on the information entropy of the j-th indicator. , thus obtaining the objective weight vector .

[0083] The weights are optimized using a nonlinear coupling formula to avoid the linearity defects of simple weighted averaging, thus constructing an optimal combination of weights and obtaining the optimal weight vector. Calculation formula:

[0084] In the formula, For the optimal combination of weights, For objective entropy weight, This represents the global subjective weight.

[0085] This formula achieves deep coupling between subjective experience weights and objective data weights, automatically increasing the weights of indicators with high data discriminative power and significant aging contribution, and automatically reducing the weights of redundant and interfering indicators, ultimately obtaining the optimal indicator weight vector that adapts to the real-time operating conditions of the transformer.

[0086] The multi-parameter fusion unit uses a tensor fusion algorithm to obtain high-dimensional multimodal features, and corrects the high-dimensional multimodal features by calculating the coupling correlation coefficients of each dimension, and outputs an insulation state fusion feature vector.

[0087] Based on four types of standardized time-series data from the standardized insulation parameter database, a four-dimensional insulation aging characteristic tensor is constructed. Where T is the number of time-series sampling steps, 4 is the four major monitoring dimensions (gas, medium, temperature, and electrical), K is the number of features in each dimension, and N is the number of samples; the one-dimensional discrete monitoring data is reconstructed into a high-order tensor that retains time-series features and dimensional correlation features, thereby achieving multi-parameter structured integration;

[0088] Introducing an optimal weight vector W to weight and correct the indices of each dimension of the tensor eliminates the fusion bias caused by differences in the importance of the indices, resulting in a weighted tensor. ,in For tensor multiplication operations, the Tucker decomposition algorithm is used to perform third-order modal decomposition on the weighted tensor, decomposing the high-order tensor into a core tensor and factor matrices of each dimension, thus obtaining high-dimensional multimodal features. The decomposition formula is as follows: In the formula For core tensors, For the time-series mode factor matrix, For the dimension modal factor matrix, Here is the characteristic mode factor matrix, where The tensor modal product identifiers, with numbers 1, 2, and 3 corresponding to temporal modality, dimensional modality, and feature modality, respectively, are a core creative design that distinguishes them from traditional two-dimensional data fusion, accurately separating three types of interference information: temporal fluctuations, dimensional differences, and feature coupling.

[0089] By exploiting factor matrix operations to uncover nonlinear correlations among parameters, a dimensional coupling correlation coefficient matrix is ​​constructed, and the co-coupling degree between different monitoring dimensions is calculated using the following formula: ,in , which is the coupling correlation coefficient between the two aging monitoring dimensions a and b, used to quantify the degree of mutual influence between different insulation aging dimensions; The covariance operator represents the overall collaborative change trend of two types of dimensional feature data. This represents the feature vector of the dimensional modality factor corresponding to the a-th aging dimension. , where a and b are the dimension modal factor feature vectors corresponding to the b-th aging dimension, and the value ranges correspond to the four major monitoring dimensions: gas aging, medium moisture absorption, thermal accumulation damage, and electrical defects. This is the variance operator, representing the degree of discrete fluctuation in single-dimensional feature data. It can accurately quantify the degree of interaction between gas aging, medium moisture absorption, thermal damage, and electrical defects, capturing coupled aging characteristics that cannot be reflected by a single parameter, such as implicit correlations like high temperature accelerating gas production from damp paper and partial discharge exacerbating medium degradation.

[0090] Based on the Tucker third-order modal decomposition results and dimensional coupling quantification results presented earlier, multimodal aging feature purification, redundancy removal, and dimensional unification are completed, ultimately generating a standardized insulation state fusion feature vector. The complete process is as follows: Based on the transformer oil-paper insulation coupled aging condition mechanism, strong-coupled collaborative degradation dimensions and weak-coupled irrelevant dimensions are distinguished. A differential adaptive correction strategy of "strong-coupled feature gain and weak-coupled feature weight suppression" is adopted to overcome the inherent defects of traditional fusion algorithms, such as equal dimensional weighting and inability to distinguish physical relationships. The specific correction process is as follows: Set the coupling threshold δ=0.1. If the dimensional coupling coefficient Corr(a,b)≥0.1, it is determined to be a strong-coupled dimensional combination; if Corr(a,b)<0.1, it is determined to be a weak-coupled dimensional combination. Construct the basic gain coefficient. Then, an exponential decay suppression function is introduced to construct the final adaptive correction coefficient. , The coupling threshold constant is used; the original dimensional modality factor matrix is ​​corrected element-by-element by the Hadamard operation, as shown in the formula: , This is the corrected dimensional modal feature matrix. For strongly coupled combinations, Corr(a,b) is large, and the correction coefficient is significantly greater than 1, amplifying the coupled aging feature gain and strengthening the core coupled aging features such as high temperature-induced moisture degradation, partial discharge-accelerated medium damage, and thermal stress-promoted oil cracking and gas production. For weakly coupled combinations, Corr(a,b) approaches 0, and the correction coefficient, after exponential decay, is less than 1, suppressing invalid interference feature weights, suppressing spurious data associations without mechanistic support, and avoiding noise features interfering with the fusion results. After correction, the static independent features of the original dimensional modal factor matrix are upgraded to a coupled aging dimensional modal feature matrix with differentiated weighting that conforms to the actual physical aging conditions of transformers. This achieves precise purification by "strengthening effective coupling features and weakening ineffective interference features";

[0091] For disassembly and modification The high-dimensional multimodal features constructed by the high-dimensional tensor have problems such as dimensional redundancy, feature sparsity and information disorder. The high-order tensor cross-dimensional joint fusion strategy is adopted to adaptively retain the core trend of temporal evolution, strong coupled dimensional correlation features and core aging features of highly sensitive parameters, while removing sensor random jitter, steady-state invalid data, repetitive and redundant features and weakly coupled false correlation features, so as to achieve low-dimensional accurate and robust representation of high-dimensional tensor features.

[0092] The pooled and purified multimodal coupled features are input into the Sigmoid nonlinear activation function to complete the global feature generation. Interval standardization and normalization mapping unifies the dimensions and scales of three types of features: time series, dimensionality, and subdivision parameters. Finally, it concatenates these features to generate a standardized, insulating state fusion feature vector with unified dimensions, no information loss, and no dimensional interference. The corrected output fusion feature vector eliminates the shortcomings of traditional equal fusion, fully conforms to the complex operating conditions of transformers, and provides high-fidelity and highly targeted fusion data support for the accurate calculation of the comprehensive insulation aging index.

[0093] The aging quantification assessment unit calculates the initial insulation aging index based on the optimal index weight and the fusion feature of insulation state, and introduces the ResSSAE residual sparse autoencoder network to correct the error of the initial insulation aging index, thereby obtaining the comprehensive insulation aging index.

[0094] A weighted comprehensive aging index model is constructed based on the optimal coupling weight W and the fused feature vector F. The initial aging index calculation formula is as follows: In the formula This is the initial insulation aging index. The optimal weight for the j-th indicator is... To correspond to the fusion feature quantity, this formula realizes the weighted integration of multi-dimensional aging features, and completes the preliminary quantification of the degree of insulation aging;

[0095] To address the issue that traditional linear calculations cannot adapt to the nonlinearity, hysteresis, and cumulative nature of insulation aging, a trained ResSSAE residual sparse autoencoder network is introduced to correct the error of the initial exponent. The fused feature values ​​are input into the network, and deep feature extraction is performed through the encoding layer to obtain the theoretical aging value fitted by the deep learning of the network.

[0096] Aging prediction residuals are calculated using residual branching. ,in This is the theoretical aging value fitted to the deep learning network. This is the residual correction amount; ultimately, a high-precision insulation aging comprehensive index is obtained. , The adaptive correction coefficient is dynamically determined based on the transformer's operating conditions, effectively offsetting the calculation errors of the linear model and accurately matching the actual aging patterns of the insulation.

[0097] The early warning module determines the insulation aging level based on the comprehensive insulation aging index and preset multi-level early warning thresholds, and triggers abnormal early warnings based on the insulation aging level.

[0098] The high-precision insulation aging comprehensive index output by the aging quantification assessment unit is retrieved in real time, and the aging state evolution characteristics under continuous time sequence are collected simultaneously to construct a dual early warning basic data that combines static aging state and dynamic aging trend. This avoids the random bias that exists in the data assessment at a single moment and ensures that the early warning judgment basis is comprehensive and reliable.

[0099] Based on the multi-level insulation aging warning threshold range preset by the system, the real-time insulation aging comprehensive index is matched and compared. According to the threshold range in which the index is located, the corresponding insulation aging level is divided, and four insulation states are distinguished step by step: healthy equipment operation, slight latent aging, moderate abnormal aging, and serious fault risk, so as to achieve a refined classification and judgment of the degree of insulation aging.

[0100] Based on the different insulation aging levels determined, a corresponding early warning mechanism is matched to achieve a shift from passive fault alarms to proactive risk warnings. For equipment in a healthy state, routine monitoring is maintained without triggering an early warning. For slightly aging equipment, a low-level early warning is triggered, recording early hidden deterioration characteristics and prompting increased routine monitoring and inspection. For moderately aging equipment, a medium-level early warning is triggered, indicating a potential for continued deterioration and prompting a shortening of the monitoring cycle and the implementation of a special insulation status check. For severely aging equipment, a high-level emergency early warning is triggered, indicating a significant decrease in insulation performance and a high risk of faults, and timely emergency maintenance alarms are pushed out.

[0101] By combining the continuous change pattern of the time-series aging index with the predicted aging trend, the static threshold judgment result is dynamically corrected. The system continuously collects the comprehensive insulation aging index over a continuous time series and calculates the slope of the time series change to identify the current aging rate of the equipment. If the aging index continues to rise, the slope is positive and continues to increase, it indicates that the insulation is in an accelerated deterioration state. Even if the current aging index has not yet reached a higher level of static warning threshold, the system will dynamically raise the current warning judgment level and weaken the static threshold boundary restriction to complete the adaptive correction of the static judgment result and avoid missing the critical state. For equipment with stable conditions and slow aging, the corresponding basic warning level is maintained to avoid invalid and frequent alarms.

[0102] Once an early warning is triggered, the abnormal causes are traced back to their source by combining multi-dimensional aging coupling and correlation characteristics. Standardized early warning information is generated, which includes aging level, deterioration risk, abnormal source tracing results, and targeted operation and maintenance strategies. This provides accurate early warning data support for intelligent operation and maintenance and hierarchical inspection and control of transformer insulation status.

[0103] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0104] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0105] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0106] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0107] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0108] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A system for evaluating the insulation aging of oil-immersed transformers based on multi-parameter monitoring, characterized in that, include: The data acquisition module collects insulation aging data of oil-immersed transformers and preprocesses the insulation aging data to build a standardized insulation parameter database. The aging assessment module includes a weight optimization unit, a multi-parameter fusion unit, and an aging quantification assessment unit. The weight optimization unit uses an entropy weight-AHP coupling algorithm to dynamically optimize the weights of each monitoring indicator. The multi-parameter fusion unit uses a tensor fusion algorithm to obtain high-dimensional multimodal features, and corrects the high-dimensional multimodal features by calculating the coupling correlation coefficients of each dimension, and outputs an insulation state fusion feature vector. The aging quantification assessment unit calculates the initial insulation aging index based on the optimal index weight and the fusion feature of insulation state, and introduces the ResSSAE residual sparse autoencoder network to correct the error of the initial insulation aging index, thereby obtaining the comprehensive insulation aging index. The early warning module determines the insulation aging level based on the comprehensive insulation aging index and preset multi-level early warning thresholds, and triggers abnormal early warnings based on the insulation aging level.

2. The oil-immersed transformer insulation aging assessment system based on multi-parameter monitoring according to claim 1, characterized in that, The insulation aging data includes parameters of dissolved gases in oil, characteristic parameters of the insulating medium, temperature and operating conditions, and electrical characteristic parameters of defects.

3. The oil-immersed transformer insulation aging assessment system based on multi-parameter monitoring according to claim 2, characterized in that, The process for obtaining the characteristic parameters of the insulating medium is as follows: An integrated insulating medium detection sensor is installed at the transformer sampling oil circuit interface to collect parameters such as the moisture content, dielectric strength, dielectric loss factor, and frequency domain dielectric spectrum characteristics of the insulating oil. Data from sensor calibration and transient fluctuations in the oil circuit are removed, while valid data on the characteristics of the medium under steady-state conditions are retained. Environmental compensation corrections were made to the parameters of insulating oil micro-water content, dielectric strength, dielectric loss factor, and frequency domain dielectric spectrum characteristic parameters, taking into account ambient temperature and humidity. Based on the threshold values ​​of the insulation evaluation standard for power equipment, the micro-water content, dielectric strength, dielectric loss factor, and frequency domain dielectric spectrum characteristic parameters of insulating oil are normalized in intervals to unify the data dimensions. The corrected and normalized dielectric characteristic data are stored in chronological order to form a dataset of moisture-induced aging of insulating dielectric.

4. The oil-immersed transformer insulation aging assessment system based on multi-parameter monitoring according to claim 3, characterized in that, The weight optimization unit uses an entropy-weighted AHP coupled algorithm to dynamically optimize the weights of each monitoring indicator, as follows: Based on the four-dimensional monitoring system, a three-level hierarchical evaluation index system is established. The target layer is the aging state of transformer insulation, the criterion layer includes four dimensions: gas aging, medium moisture, heat accumulation damage, and electrical defects, and the index layer consists of the subdivided monitoring parameters corresponding to each criterion layer. Construct the criterion layer judgment matrix and solve the dimension weights, and construct the index layer domain judgment matrix according to the dimension and solve the subdivision relative weights; Based on the dimensional weights of the criteria layer and the relative weights of the indicator layers under that dimension, the global subjective weights of all monitoring parameters are obtained through a double-layer nested multiplication calculation. Obtain the objective entropy weight, combine it with the global subjective weight to construct the optimal combined weight, and obtain the optimal combined weight.

5. The oil-immersed transformer insulation aging assessment system based on multi-parameter monitoring according to claim 4, characterized in that, The process of constructing the criterion layer judgment matrix and solving the dimension weights is as follows: For the four dimensions of aging, a pairwise comparison judgment matrix of the fourth-order criterion layer is constructed based on the aging mechanism of electrical insulation using the scaling method; The maximum eigenvalue and corresponding eigenvector of the matrix are solved. After normalizing the eigenvector, the weights of the four criteria layers are obtained, namely the weights of gas aging, medium moisture, thermal accumulation damage, and electrical defects. Consistency is then checked based on the maximum eigenvalue.

6. The oil-immersed transformer insulation aging assessment system based on multi-parameter monitoring according to claim 5, characterized in that, The process of obtaining high-dimensional multimodal features using the tensor fusion algorithm is as follows: Construct a four-dimensional insulation aging feature tensor based on four-dimensional normalized time series data. Where T is the number of time-series sampling steps, 4 is the four major monitoring dimensions, K is the number of features in each dimension, and N is the number of samples; We introduce an optimal weight vector W to perform weighted correction on the indices of each dimension of the tensor, resulting in a weighted tensor. The Tucker decomposition algorithm is used to perform third-order modal decomposition on the weighted tensor, decomposing the high-order tensor into a core tensor and factor matrices of each dimension, thus obtaining high-dimensional multimodal features.

7. The oil-immersed transformer insulation aging assessment system based on multi-parameter monitoring according to claim 6, characterized in that, The process of correcting high-dimensional multimodal features by calculating coupling correlation coefficients in each dimension and outputting insulation state fusion feature vector is as follows: A dimensional coupling correlation coefficient matrix is ​​constructed through factor matrix operations, and the co-coupling degree between different monitoring dimensions is calculated using the following formula: ,in The coupling correlation coefficient between the two aging monitoring dimensions, a and b; Covariance operator; This represents the feature vector of the dimensional modality factor corresponding to the a-th aging dimension. Let a and b be the feature vector of the dimensional modality factor corresponding to the b-th aging dimension, where a and b are the dimension indices; This is the variance operator; Based on the aging condition mechanism of transformer oil-paper insulation coupling, we distinguish between strongly coupled synergistic degradation dimensions and weakly coupled unrelated dimensions. We set a coupling threshold δ=0.

1. If the dimension coupling coefficient Corr(a,b)≥0.1, it is determined to be a strongly coupled dimension combination. If Corr(a,b)<0.1, it is determined to be a weakly coupled dimension combination. Constructing the basic gain coefficient Then, an exponential decay suppression function is introduced to construct the final adaptive correction coefficient. , The coupling threshold constant is used; the original dimensional modality factor matrix is ​​corrected element-by-element by the Hadamard operation, as shown in the formula: , This is the corrected dimensional modal feature matrix; The high-dimensional multimodal features, after being decomposed, corrected, pooled, compressed, and purified, are input into the Sigmoid nonlinear activation function for standardized normalization mapping, generating a standardized insulation state fusion feature vector.

8. The oil-immersed transformer insulation aging assessment system based on multi-parameter monitoring according to claim 7, characterized in that, The specific implementation process of the aging quantification assessment unit is as follows: A weighted comprehensive aging index model is constructed based on the optimal coupling weight W and the fused feature vector F to obtain the initial insulation aging index, as shown in the formula. In the formula This is the initial insulation aging index. The optimal weight for the j-th indicator is... For the corresponding fusion feature quantity; A ResSSAE residual sparse autoencoder network is introduced to correct the error of the initial insulation aging index. The fused feature values ​​are input into the network, and deep feature extraction is performed through the encoding layer to obtain the theoretical aging value fitted by the deep learning of the network. Based on the initial insulation aging index and the theoretical aging value fitted by deep learning, the aging prediction residual is calculated through residual branching. ,in This is the theoretical aging value fitted to the deep learning network. This is the residual correction amount; A high-precision comprehensive insulation aging index is calculated based on the aging prediction residuals combined with adaptive correction coefficients.

9. The insulation aging assessment system for oil-immersed transformers based on multi-parameter monitoring according to claim 8, characterized in that, The aging early warning module is implemented as follows: Obtain the comprehensive insulation aging index, simultaneously collect the aging state evolution characteristics under continuous time sequence, and construct a dual early warning basic data that combines static aging state and dynamic aging trend. Based on the pre-set multi-level insulation aging warning threshold range, the real-time insulation aging comprehensive index is matched and compared. The corresponding insulation aging level is divided according to the threshold range in which the index is located, and four insulation states are distinguished step by step: healthy equipment operation, slight latent aging, moderate abnormal aging, and serious fault risk. The warning mechanism is matched with the corresponding level according to different insulation aging levels. No warning is triggered for equipment in a healthy state; for equipment in a slightly aging state, a low-level warning is triggered to record the early hidden deterioration characteristics of the equipment. For moderate aging conditions, a medium-level warning is triggered, indicating that the equipment has a potential for continued deterioration; for severe aging conditions, a high-level emergency warning is triggered, indicating that the equipment's insulation performance has significantly deteriorated and there is a high risk of failure, and an emergency maintenance alarm is promptly pushed out. By combining the continuous change pattern of the time-series insulation aging comprehensive index with the predicted aging trend, the static threshold judgment result is dynamically corrected. Once an early warning is triggered, the abnormal causes are traced back to their source by combining multi-dimensional aging coupling and correlation characteristics, and standardized early warning information including aging level, degradation risk, abnormal source tracing results and targeted operation and maintenance strategies is generated.

Images