A transformer multi-modal coupling health assessment method based on FAHP-MEREC game empowerment

Abstract

This invention discloses a transformer multimodal coupled health assessment method based on FAHP-MEREC game-theoretic weighting, relating to the field of power transformer technology. The method includes: synchronously collecting multimodal data of the transformer and using a dedicated standardized algorithm to perform submodal adaptive preprocessing on various types of multimodal data to construct a multimodal feature dataset; extracting evaluation indicators for each mode based on the multimodal feature dataset, calculating subjective weights using fuzzy hierarchical analysis and objective weights using the criterion removal effect method; fusing multi-source evidence using improved evidence theory to obtain global confidence based on the multimodal feature dataset and optimal combined weights; and calculating the critical failure time and remaining safe service life of the transformer based on a preset insulation failure threshold. This invention upgrades the identification of single explicit faults to comprehensive identification of all-dimensional latent and explicit faults, significantly improving the ability to identify early minor defects and latent aging.

Description

Technical Field

[0001] This invention relates to the field of power transformer technology, and in particular to a multimodal coupled health assessment method for transformers based on FAHP-MEREC game-theoretic weighting. Background Technology

[0002] Oil-immersed power transformers are core and critical equipment for energy conversion and transmission in power systems. The health of their insulation system directly affects the safe and stable operation of the power grid and the reliability of power supply. Under long-term and complex operating conditions, the transformer's oil-paper insulation system is continuously subjected to multiple coupled effects such as high electric fields, alternating temperature rise, load fluctuations, moisture aging, and impurity corrosion. This can easily lead to defects such as insulation aging, partial discharge, overheating, and oil cracking and gas production. These defects can overlap and act as mutual inducements, easily forming thermoelectric and electrochemical composite coupled faults, ultimately causing winding breakdown, tank failure, or even unplanned outages.

[0003] Therefore, carrying out multi-dimensional condition monitoring, accurate health assessment and remaining life prediction of transformers has important engineering value and practical significance for realizing the transformation of equipment from periodic maintenance to condition-based maintenance and intelligent operation and maintenance.

[0004] Current transformer condition assessment technologies mainly rely on single monitoring parameters or simple fusion of multiple parameters, and are primarily divided into three categories: oil chromatography chemical diagnostic methods, partial discharge electrical diagnostic methods, and temperature and thermal assessment methods. Each of these existing technologies has inherent technical bottlenecks, making it difficult to adapt to real-world fault scenarios with multiple coupled defects under complex operating conditions. Specific shortcomings are as follows:

[0005] First, traditional single-modal diagnostic methods suffer from limited information coverage and weak early fault identification capabilities. Chemical diagnostic methods such as oil chromatography three-ratio method and David's triangle method rely solely on dissolved gas parameters in the oil for analysis, and can only capture the overt fault gas generation characteristics after severe overheating and violent discharge. They have extremely low sensitivity to early weak insulation aging, intermittent partial discharge, and latent defects such as low-temperature temperature rise, and cannot detect hidden degradation problems caused by multi-parameter coupling. Single monitoring methods such as partial discharge electrical diagnosis and temperature thermal assessment also suffer from incomplete information coverage, reflecting only a single dimension of operating status. They cannot take into account the comprehensive degradation mechanism of transformer chemical aging, electrical damage, and thermal degradation, resulting in serious assessment bias and prominent problems of missed or misdiagnosed early faults.

[0006] Second, traditional multimodal fusion techniques have poor adaptability to heterogeneous data and low feature utilization efficiency. Most existing multi-parameter fusion methods directly perform simple weighted fusion of oil chromatography, partial discharge, and temperature data, without considering the inherent heterogeneous characteristics of these three types of monitoring data: frequency differences, dimensional conflicts, different noise sources, and information asymmetry. Partial discharge data is high-frequency pulse abrupt change data, oil chromatography is low-frequency slow-change drift data, and temperature is operating condition coupled fluctuation data. Unified preprocessing and standardization methods easily lead to weak fault features being submerged by baseline noise and pulse fault features being compressed by the mean, resulting in the loss of effective fault information and the inability to equitably fuse multi-source data features, making it unsuitable for adapting to the multi-dimensional and differentiated degradation scenarios of transformers.

[0007] Third, traditional weighting methods suffer from poor weight rationality, leading to severely distorted evaluation results. Existing evaluation models often employ either a single subjective weighting method (AHP, expert scoring) or a single objective weighting method (entropy weighting, coefficient of variation method) to determine indicator weights. Single subjective weighting relies excessively on human experience and subjective cognition, resulting in rigid scoring biases and failing to adapt to the engineering characteristics of transformer fault boundaries being ambiguous and aging gradually, with insufficient mechanistic adaptability. Single objective weighting relies entirely on statistical data patterns, detached from the physical fault mechanisms of transformer oil-paper insulation deterioration, partial discharge development, and thermal aging, easily leading to false weights that are "abnormal in data but lack fault mechanism support," resulting in distorted determination of indicator contribution and significantly reducing the credibility and interpretability of the evaluation results.

[0008] Fourth, traditional evidence fusion algorithms have weak conflict handling capabilities and lack the ability to identify coupled faults. Conventional DS evidence theory is prone to problems such as fusion paradox and result distortion when there are conflicts in multi-source heterogeneous evidence. Moreover, traditional methods assume that unknown and uncertain confidence is evenly distributed to various health states, completely ignoring the latent thermoelectric and electrochemical coupled fault characteristics hidden behind the conflicts in multimodal evidence. They cannot distinguish between single independent faults and multi-parameter coupled composite faults, and lack the ability to identify coupled faults. This is the core technical shortcoming of the current failure to detect early composite faults in transformers.

[0009] Fifth, traditional life prediction models have poor anti-interference capabilities and insufficient accuracy in characterizing degradation. Existing equipment remaining life prediction models mostly rely on single state parameters to construct degradation models, without combining multimodal comprehensive health degradation characteristics, and without eliminating random noise caused by on-site load fluctuations and environmental interference. They are difficult to fit the long-term nonlinear degradation law of transformer insulation and the accelerated aging degradation law. They have problems such as large deviations in judging short-term disturbance interference trends, large deviations in long-term degradation extrapolation, and high remaining life prediction errors, which cannot meet the engineering requirements of refined full life cycle operation and maintenance. Summary of the Invention

[0010] Therefore, it is necessary to provide a transformer multimodal coupling health assessment method based on FAHP-MEREC game weighting to address the above-mentioned technical problems.

[0011] This invention provides a transformer multimodal coupling health assessment method based on FAHP-MEREC game-theoretic weighting, comprising:

[0012] S1. Synchronously acquire multimodal data of the transformer, and use a dedicated standardized algorithm to perform submodal adaptive preprocessing on various types of multimodal data to construct a multimodal feature dataset; the multimodal data includes partial discharge data, oil chromatography chemical data, and temperature and thermal data;

[0013] S2. Based on the multimodal feature dataset, extract the evaluation indexes for each modality, calculate the subjective weights using the fuzzy hierarchical analysis method, calculate the objective weights using the criterion removal effect method, and solve the optimal fusion coefficients of the subjective and objective weights through game theory Nash equilibrium to generate the optimal combined weights.

[0014] S3. Based on the multimodal feature dataset and the optimal combined weight, the improved evidence theory is used to fuse multi-source evidence to obtain the global confidence score. The confidence scores of the basic health level and the coupled fault warning confidence scores are extracted to form a five-dimensional evaluation result, and a quantitative comprehensive health index is calculated.

[0015] S4. Based on the comprehensive health index of continuous monitoring, a long-term health sequence reflecting the long-term deterioration trend of insulation is constructed, and a degradation model that fits the nonlinear aging law of the oil-paper insulation is fitted. Based on the preset insulation failure threshold, the critical failure time is solved and the remaining safe service life of the transformer is calculated.

[0016] The beneficial effects of this invention are as follows:

[0017] 1. This invention abandons the traditional single-parameter diagnostic mode and integrates three core monitoring dimensions: oil chromatography chemical degradation, partial discharge electrical damage, and temperature thermal aging, covering the entire path of transformer insulation aging degradation characteristics. Through multimodal information complementarity, it completely solves the technical shortcomings of single oil chromatography methods being unable to identify weak partial discharges, single electrical methods being unable to characterize the degree of insulation aging, and single temperature methods being unable to quantify latent gas generation defects. It achieves an upgrade from single overt fault diagnosis to comprehensive identification of all-dimensional latent and overt faults, greatly improving the ability to identify early minute defects and latent aging, and is suitable for transformer full-condition fault monitoring scenarios.

[0018] 2. Compared to traditional unified standardization processing methods, this invention innovatively adopts a modal adaptive differentiated preprocessing strategy, matching specific noise reduction and standardization algorithms to the inherent characteristics of different modal data: extreme value normalization is used for partial discharge high-frequency pulse data to retain pulse mutation fault characteristics; Z-score baseline correction standardization is used for oil chromatography slow drift data to eliminate long-term baseline drift interference; range standardization is used for temperature residual data to adapt to thermal fluctuation characteristics; effectively solving the problem of loss of effective information caused by mismatch in frequency, scale, and noise characteristics of multi-source data, achieving equivalence fusion of heterogeneous data, maximizing the retention of weak early fault characteristics, and adapting to multimodal heterogeneous data fusion scenarios under complex operating conditions.

[0019] 3. Compared to traditional single FAHP subjective weighting and single objective weighting methods, this invention constructs a FAHP-MEREC two-layer game-theoretic combined weighting system. It leverages the FAHP fuzzy hierarchical analysis method to align with transformer fault mechanisms and prior knowledge of operation and maintenance, addressing the issues of large biases and strong subjective arbitrariness in rigid manual scoring. It utilizes the MEREC criterion to remove effects, mining objective disturbance differences in monitoring data and quantifying the true data contribution of each indicator, avoiding the shortcomings of pure data statistics detached from physical mechanisms. Simultaneously, it introduces a game-theoretic Nash equilibrium model to adaptively solve for the optimal fusion coefficients, achieving an optimal balance between mechanistic experience and data characteristics. This completely eliminates the weight bias and distortion problems of single weighting, ensuring that indicator weights possess both mechanistic rationality and data objectivity, significantly improving the universality and credibility of the evaluation model.

[0020] 4. Compared to the technical shortcomings of conventional DS evidence theory, which directly distributes unknown and uncertain confidence levels equally and cannot identify coupled faults, this invention constructs a multi-correction improved DS fusion framework. Through initial correction of Shannon entropy confidence and secondary correction of adaptive discount coefficients, it effectively resolves conflicts between multi-source heterogeneous evidence and solves the problem of distortion in high-conflict evidence fusion. It innovatively abandons the traditional equal distribution mechanism and adds a coupled fault threshold discrimination and unknown confidence retention module, transforming previously discarded conflict and uncertainty information into effective features of coupled faults. It also adds early warning capabilities for thermoelectric and electrochemical coupled latent faults, achieving a five-dimensional output of four-level basic health status and coupled fault early warning. This fills the technical gap in traditional technologies that cannot identify multimodal composite latent faults and significantly reduces the false negative rate of coupled faults.

[0021] 5. Compared to traditional single-parameter degradation prediction models, this invention constructs an equipment degradation model based on a multimodal fusion of comprehensive health long-term series, which can comprehensively characterize the multi-dimensional cumulative insulation damage of transformers. By using moving average filtering to eliminate load and environmental random disturbances, it restores the true steady-state degradation trend of the equipment. Combined with an exponential degradation model that conforms to the accelerated aging characteristics of paper-insulated transformers and industry-standard failure thresholds, it achieves an organic integration of qualitative classification of health status and quantitative extrapolation of remaining life. This effectively solves the problems of poor adaptability to operating conditions, inaccurate characterization of degradation patterns, and large life prediction errors in traditional models, providing accurate quantitative basis for transformer condition-based maintenance, full life-cycle asset management, and replacement decisions for aging equipment. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0023] Figure 1 This is a flowchart of a transformer multimodal coupling health assessment method based on FAHP-MEREC game weighting according to an embodiment of the present invention;

[0024] Figure 2 This is a flowchart of the method framework according to an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of a health assessment system according to an embodiment of the present invention;

[0026] Figure 4 This is a graph showing the generation of the five-dimensional assessment results and the comprehensive health index according to an embodiment of the present invention;

[0027] Figure 5 This is a flowchart of transformer remaining life prediction according to an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein...

[0029] Examples are used merely to explain the present invention and are not intended to limit the present invention.

[0030] Please see Figures 1-2 A transformer multimodal coupling health assessment method based on FAHP-MEREC game-theoretic weighting, comprising:

[0031] S1. Synchronously acquire multimodal data of the transformer and use a dedicated standardized algorithm to perform submodal adaptive preprocessing on various types of multimodal data to construct a multimodal feature dataset; the multimodal data includes partial discharge data, oil chromatography chemical data and temperature and thermal data.

[0032] In the description of this invention, multimodal data of oil-immersed transformers are collected synchronously, and a dedicated standardization algorithm is used to perform mode-specific adaptive preprocessing on various types of multimodal data to construct a multimodal feature dataset, including:

[0033] S11. For partial discharge data, an adaptive wavelet threshold denoising algorithm is used to remove on-site electromagnetic white noise, power frequency interference and its harmonic interference, retain the effective discharge pulse waveform, and extract partial discharge fault features; and an extreme value normalization algorithm is used to map all partial discharge fault features to a standardized range.

[0034] Specifically, for partial discharge data, an adaptive wavelet threshold denoising algorithm is first used to remove random electromagnetic white noise and periodic interference noise from the field, preserving the complete and effective discharge pulse waveform. Based on this, refined multi-dimensional feature extraction is performed. The specific extraction content and its physical meaning are as follows:

[0035] (1) PRPD Phase-Resolved Spectrum Feature Extraction: The PRPD spectrum is the core feature characterizing the phase distribution law of partial discharge pulses. For the pre-processed discharge phase-amplitude two-dimensional spectrum, the core features such as phase clustering interval, average discharge amplitude, maximum discharge quantity, phase distribution dispersion, and positive and negative half-cycle discharge asymmetry are accurately extracted. These features can effectively distinguish different defect types such as air gap discharge, surface discharge, and metal tip discharge inside the transformer, reflecting the occurrence law and phase dependence characteristics of partial discharge faults. They are key spectral indicators for identifying structural insulation defects.

[0036] (2) PRPS pulse sequence map feature extraction: The PRPS map focuses on the temporal evolution of partial discharge pulses and extracts temporal features such as pulse repetition rate, pulse time interval distribution, pulse amplitude temporal fluctuation coefficient, and continuous discharge duration. These features can characterize the development trend of partial discharge and can effectively identify the initial intermittent weak discharge, the middle continuous stable discharge, and the later dense and violent discharge, accurately reflecting the deterioration process of insulation defects.

[0037] (3) Extraction of higher-order statistical features in the time domain: Based on the original time-domain waveform of the partial discharge after noise reduction, four types of higher-order statistical features are extracted: steepness, skewness, peak factor, and pulse width. Among them, steepness characterizes the sharpness of the sudden change in the partial discharge pulse waveform, which can identify instantaneous and severe discharge defects; skewness reflects the asymmetry of the waveform amplitude distribution and quantifies the degree of distortion of the discharge signal; peak factor reflects the pulse impact intensity, and pulse width characterizes the discharge duration. These higher-order features are highly sensitive to early weak and non-stationary partial discharge signals and can capture latent discharge faults that cannot be identified by conventional features.

[0038] (4) Extraction of numerical features of physical monitoring: Two types of direct physical monitoring indicators are extracted simultaneously: ultrasonic signal amplitude and instantaneous peak value of high-frequency current. Ultrasonic signals can capture the acoustic vibration generated by weak discharge inside the transformer, which is suitable for the identification of latent discharge defects inside the closed cavity; high-frequency current signals can accurately characterize the current impact intensity of partial discharge pulses and intuitively reflect the severity of insulation damage. The two types of physical features complement each other to realize the quantitative characterization of the intensity of partial discharge faults.

[0039] By extracting complementary features (partial discharge fault features) from the above multi-dimensional sources, the system comprehensively covers different defect states of transformers, such as floating discharge, intermittent discharge, continuous discharge, and severe discharge. This completely solves the problems of low identification rate and missed detection of early defects caused by traditional single partial discharge features, and obtains partial discharge fault features, providing rich and accurate electrical fault feature support for subsequent multi-modal fusion assessment.

[0040] Partial discharge faults are characterized by high-frequency pulse-type abrupt data with a very large dynamic range, strong extreme value impact, and a large number of zero-value silent intervals. During normal operation, the data is in a low-amplitude silent state for a long time. When a fault occurs, a large pulse peak will appear instantaneously. The data distribution is extremely uneven. If mean-variance type standardization is used, it will smooth out the weak discharge differences and compress the fault pulse characteristics, leading to the failure of early partial discharge characteristics.

[0041] Therefore, this invention employs a dedicated extreme value normalization algorithm for partial discharge fault characteristics, specifically adapted to pulse-type sparse abrupt change data, to fully preserve the strength differences of the partial discharge signal and early weak fault information. The specific algorithm principle and implementation process are as follows:

[0042] ;

[0043] In the formula, To standardize the partial discharge eigenvalues, normalize them to the interval [0, 1]. These are the raw partial discharge characteristic parameters (discharge amplitude, pulse repetition rate, waveform statistical characteristics, etc.) extracted at the current moment. The maximum value of all partial discharge characteristic samples within the preset time-series sliding window represents the maximum abnormal discharge level of the device during that time period. The minimum value among all partial discharge characteristic samples within the same time period represents the normal silent noise level of the equipment. This algorithm uses the actual fluctuation range of partial discharge data in a single time period as the scaling benchmark and does not rely on the global statistical mean.

[0044] This invention can completely preserve the abrupt differences and weak discharge characteristics of partial discharge pulses, and will not smooth out the early weak partial discharge characteristics due to mean compression, thus adapting to the high-frequency, abrupt, and sparse data characteristics of electrical pulses. At the same time, it adds pulse threshold filtering and noise reduction to the partial discharge signal: silent fluctuations below a fixed weak threshold are directly set to zero, completely eliminating residual electromagnetic noise interference.

[0045] S12. For oil chromatographic chemical data, a time-series moving average baseline correction algorithm is used to separate the slow baseline caused by natural aging of oil and seasonal temperature changes from the fast incremental baseline caused by insulation faults, eliminate long-term baseline drift and seasonal fluctuation interference, and extract oil chromatographic fault features; and a mean-variance standardization algorithm is used to process the oil chromatographic fault features to correct baseline shift.

[0046] Specifically, oil chromatography chemical data are low-frequency, slowly varying time-series data, mainly containing key insulating characteristic gases such as hydrogen, acetylene, and total hydrocarbons. The minute changes in their concentrations, abrupt increases, and ratio relationships are the core evidence reflecting transformer insulation aging, overheating, and discharge faults. Steady-state baseline data can easily mask early, weak fault information. This invention specifically conducts multi-dimensional fault feature extraction from oil chromatography, specifically including:

[0047] (1) Gas concentration time series characteristics: extract the real-time concentration, time series increment and long-term change slope of each characteristic gas, capture the slow accumulation change and sudden abnormal fluctuation of insulating gas during equipment operation, and adapt to the identification of insulation chronic aging and instantaneous discharge overheating faults.

[0048] (2) Gas growth rate characteristics: Calculate the daily / weekly growth rates of hydrogen, acetylene and total hydrocarbons. Compared with static concentration values, the growth rate can effectively avoid the interference of baseline value fluctuations and accurately identify the deterioration trend of early latent faults.

[0049] (3) Key gas ratio characteristics: Extract the core gas ratio, hydrocarbon gas ratio, hydrogen to total hydrocarbon ratio and other characteristics of the three ratio method. Based on the power fault mechanism, distinguish different types of oil-immersed insulation faults such as low temperature overheating, high temperature overheating, arc discharge and partial discharge, so as to achieve preliminary identification of fault types.

[0050] By extracting fault features from multi-dimensional oil chromatography, eliminating steady-state invalid data, deeply mining fault information hidden in slow-changing data, and accurately capturing early slow fault features such as transformer insulation aging and latent discharge overheating.

[0051] Chemical characteristics such as gas concentration, gas growth rate, and gas ratio in oil chromatography are typical low-frequency, slowly varying, long-term baseline drift, and small-amplitude gradual variation time-series data. This type of data lacks partial discharge-type instantaneous pulse abrupt changes; the characteristic values ​​rise slowly during normal equipment aging and are easily affected by seasonal environmental changes, slow oil quality degradation, and long-term baseline shifts in equipment operation, resulting in persistent baseline drift errors. Using an extreme value normalization algorithm would amplify long-term baseline drift noise and weaken the incremental fault characteristics of early trace gas production, leading to early insulation aging and latent discharge faults being misjudged as normal baseline fluctuations.

[0052] Therefore, this invention, targeting the characteristics of oil chromatography chemical modal data, employs the Z-score mean-variance standardization algorithm to adapt to slow-drift and gradually varying time-series data, accurately removing baseline interference and amplifying weak fault increments. The specific algorithm principle and implementation process are as follows:

[0053] ;

[0054] In the formula, The standardized oil chromatographic characteristic values ​​were then digitized and scaled. The raw oil chromatographic features extracted at the current moment include parameters such as the concentration of each feature gas, time increment, daily / weekly growth rate, and key gas ratio; The mean value of all oil chromatographic feature samples within a preset time-series sliding window is used to characterize the baseline level of normal equipment operation during that period. This represents the standard deviation of all oil chromatographic characteristic samples within the same time period, used to characterize the normal fluctuation range and dispersion of the characteristic data. The algorithm uses the overall data distribution pattern as a benchmark to correct for baseline shifts and adapt to the drift characteristics of slowly varying data.

[0055] This invention effectively eliminates numerical shifts caused by long-term baseline drift and minor seasonal fluctuations in oil chromatography, amplifies subtle gas increment changes, retains early latent aging and trace gas production fault characteristics, and adapts to the characteristics of slow-variable chemical data. Simultaneously, a baseline drift correction is added to eliminate fault-free interference from long-term slow baseline rises in oil chromatography data, accurately retaining effective abrupt increments induced by faults. Specifically, it employs a time-series moving average baseline correction method, setting a fixed-duration sliding window to smooth and filter the original gas characteristic sequence, and calculating the dynamic baseline value at each moment. This dynamic baseline can adaptively fit the slow drift components caused by oil aging, seasonal temperature changes, and long-term static operation of the equipment. The original characteristic time-series data is subtracted from the corresponding sliding baseline value to completely eliminate slow baseline shifts, leaving only the true effective changes caused by internal insulation faults, discharges, and overheating. The specific calculation formula is as follows:

[0056] ;

[0057] ;

[0058] In the formula, is the adaptive sliding baseline value at time t; N is the sliding window length, which is set to a fixed time step size according to the slow variation characteristics of oil chromatography; x(k) is the original oil chromatography characteristic value at time k; This is the pure fault characteristic value after baseline correction, which completely eliminates slow baseline drift and retains only the information on abnormal fault changes.

[0059] Normal aging in oil chromatography exhibits a slow, continuous, and low-slope baseline rise, while faulty gas production shows abrupt and rapidly increasing abnormal fluctuations. Sliding baseline correction can eliminate slowly changing baselines and retain rapidly changing faults, completely resolving the technical shortcomings of traditional algorithms that misjudge baseline drift as early faults or submerge weak faults within baseline drift.

[0060] S13. For temperature and thermal data, a theoretical benchmark model between temperature, real-time load, and ambient temperature is established using the multiple linear regression method to extract thermal fault characteristics; and the residual range standardization algorithm is used to normalize the thermal fault characteristics.

[0061] Specifically, winding temperature and oil temperature are continuous, steady-state, slowly varying data, easily affected by external operating conditions such as ambient temperature and load fluctuations. The raw temperature data cannot directly reflect thermal faults in the equipment itself. This invention, based on adaptive calibration under operating conditions and the elimination of spurious temperature rise errors, accurately extracts the core characteristics of thermal faults. The specific implementation methods of adaptive calibration under operating conditions and the elimination of spurious temperature rise errors are as follows:

[0062] This invention establishes a transformer temperature reference model based on historical operating data, decomposing the temperature data into: ;

[0063] in, For actual measured temperature, For load-based and ambient temperature The theoretical reference temperature, This reflects the actual temperature rise residual component that reflects the internal heating of the equipment.

[0064] In the description of this invention, for temperature thermal data, a theoretical benchmark model between temperature and real-time load and ambient temperature is established using a multiple linear regression method to extract thermal fault characteristics; and a residual range standardization algorithm is used to normalize the actual temperature rise residual, including:

[0065] S131. Select historical data of normal operation of the transformer without faults or maintenance, take the real-time load and ambient temperature of the transformer as independent variables, and the measured value of winding or oil temperature as dependent variables, establish a theoretical reference model, and calculate the current theoretical reference temperature of the transformer.

[0066] Specifically, this step uses a multiple linear regression method to establish a statistical relationship between transformer temperature and load and ambient temperature, and obtains a theoretical temperature baseline curve under healthy conditions, which is used to eliminate temperature drift caused by external operating conditions.

[0067] The modeling input consists of historical data from normal operation of the transformer without faults or maintenance, excluding sensor malfunctions, missing data, and extreme operating conditions, with the transformer load as the input. and ambient temperature The independent variable is the measured value of the winding or oil temperature. It is the dependent variable.

[0068] Theoretical benchmark model form: ;

[0069] in, The theoretical reference temperature, These are the regression coefficients obtained by fitting using the least squares method.

[0070] S132. The difference between the measured temperature under real-time operating conditions and the theoretical reference temperature is used to obtain the true temperature rise residual of the equipment body under the condition of eliminating external operating interference.

[0071] Specifically, the difference between the measured temperature under real-time operating conditions and the theoretical reference temperature is used to obtain the true temperature rise residual of the equipment body after eliminating external operating condition interference: The residual eliminates the normal temperature rise caused by load and ambient temperature, and only reflects the changes in the transformer's internal thermal state such as heating and abnormal heat dissipation, thus effectively eliminating false temperature rise errors.

[0072] S133. Based on the calculation results of the theoretical reference temperature and the actual temperature rise residual, extract the current thermal fault characteristics of the transformer to realize the quantification of the thermal anomaly state of the transformer; the thermal fault characteristics include relative temperature rise characteristics, steady-state temperature deviation characteristics and short-term temperature change characteristics.

[0073] Specifically, based on the data preprocessing, the following core features of thermal faults are extracted:

[0074] (1) Relative temperature rise characteristics: Based on the load and ambient temperature correction baseline, the relative reference temperature rise values ​​of winding and oil are extracted to eliminate external working condition interference and truly reflect the degree of abnormal heating of the equipment body.

[0075] (2) Steady-state temperature deviation characteristics: Extract the temperature deviation under long-term steady-state operation to characterize chronic thermal defects such as aging of the transformer heat dissipation system and abnormal core loss.

[0076] (3) Short-term temperature change characteristics: capture the amount and rate of temperature change caused by short-term load fluctuations, internal short circuits, and local overheating, and identify sudden thermal failures and local heating defects.

[0077] By extracting multi-dimensional thermal features, the accurate quantification of transformer thermal anomalies can be achieved, effectively avoiding misjudgment caused by operating condition disturbances and improving the multi-modal fault feature system.

[0078] After extracting differentiated features for the electrical, chemical, and thermal modes, this invention employs a mode-specific standardized algorithm to unify the scale of all feature dimensions, eliminate the influence of amplitude differences between high-frequency and low-frequency data, and construct an equivalent, regular, and highly recognizable multimodal feature dataset. This completely solves the problems of information asymmetry and feature contribution imbalance in multi-source heterogeneous data, providing high-quality and comprehensive fault feature input for subsequent FAHP-MEREC combined weighting and DS evidence fusion.

[0079] S134. Subtract the minimum value of all real temperature rise residuals within the preset time window from the real temperature rise residual of the current transformer equipment body to obtain the actual difference value. Subtract the minimum value of all real temperature rise residuals within the preset time window from the maximum value of all real temperature rise residuals within the preset time window to obtain the benchmark difference value. Calculate the ratio of the actual difference value to the benchmark difference value to obtain the standardized thermal mode characteristic value.

[0080] Specifically, temperature data is strongly coupled with interference from load and ambient temperature, so the absolute temperature value has no evaluation significance. Only the actual temperature rise residual after preprocessing has fault characterization value.

[0081] Therefore, this invention features a unique residual range standardization for thermal modes, normalizing only the actual temperature rise residual of the substrate:

[0082] ;

[0083] In the formula, These are the standardized thermal modal eigenvalues, with their value range uniformly mapped to the interval [0, 1]. The residual temperature rise of the transformer windings and oil at the current moment is the actual temperature rise of the transformer windings and oil after the false errors have been eliminated by operating condition calibration. This is the minimum value of all actual temperature rise residuals within the preset timing window; This represents the maximum value of all actual temperature rise residuals within the preset timing window.

[0084] This invention completely avoids the numerical fluctuations of the original temperature caused by operating conditions, mapping only the changes in thermal defects of the equipment body to a unified range, thus thoroughly solving the technical problem of temperature data being obscured by operating condition interference. Simultaneously, a steady-state anomaly screening mechanism is added to eliminate spurious temperature rises caused by short-term load fluctuations, retaining the true thermal defect offset and abrupt change characteristics.

[0085] Through the novel technical architecture of modal independent error elimination and modal-specific standardization, this invention configures dedicated processing logic for the inherent interference characteristics, data distribution patterns, and change mechanisms of electrical, chemical, and thermal data, avoiding the over-compression or noise retention defects of traditional unified standardization algorithms for different modal data. Ultimately, the three completely heterogeneous monitoring data with different scales and noise types are uniformly mapped to the [0, 1] equivalent feature interval, constructing a highly consistent, highly discriminative, and modal-bias-free multimodal feature dataset. This provides accurate, balanced, and distortion-free pre-input for subsequent FAHP-MEREC game weighting and DS evidence fusion, significantly improving the accuracy and robustness of multimodal coupled fault assessment.

[0086] S14. Integrate the fault features of various multimodal data after dedicated standardization processing to form a multimodal feature dataset without modal bias.

[0087] S2. Based on the multimodal feature dataset, extract the evaluation indexes for each modality, calculate the subjective weights using the fuzzy hierarchical analysis method, calculate the objective weights using the criterion removal effect method, and solve the optimal fusion coefficients of the subjective and objective weights through game theory Nash equilibrium to generate the optimal combined weights.

[0088] Specifically, to address the issues of excessive subjectivity or data being detached from the fault mechanism in single weighting methods, this invention constructs a dual-layer weighting solution system consisting of FAHP fuzzy subjective weighting and MEREC criterion removal effect objective weighting. The initial weights of the indicators are calculated from two dimensions: prior knowledge of equipment fault mechanisms and objective disturbance characteristics of multimodal monitoring data, providing a dual-weight basis for subsequent optimal game fusion.

[0089] Based on the preprocessed standardized multimodal feature dataset, the MEREC objective weighting method is adopted to quantify the objective contribution of each indicator by evaluating the overall fluctuation range after removing the indicators. It is completely data-driven, without human intervention, and accurately adapts to the problem of information asymmetry in multimodal heterogeneous data.

[0090] The core innovation of this invention is to use the Nash equilibrium algorithm of game theory to solve the optimal fusion coefficient of subjective weights and objective weights, thereby avoiding the subjective defects of traditional fixed coefficient weighted fusion and realizing the dynamic equilibrium fusion of prior knowledge of fault mechanism and objective characteristics of monitoring data.

[0091] In the description of this invention, the calculation of subjective weights using fuzzy hierarchical analysis includes:

[0092] S21. Compare the importance of each evaluation index extracted from the multimodal feature data pairwise, and construct a fuzzy complementary judgment matrix based on the importance comparison results; and the sum of the importance of any two evaluation indices in the fuzzy complementary judgment matrix is ​​equal to one, and the importance of each evaluation index itself is 0.5.

[0093] Specifically, based on the fault mechanism of power transformers and industry operation and maintenance experience, pairwise importance comparisons are performed on three modal indicators—oil chromatography data, partial discharge, and temperature-thermal performance—and their subordinate characteristic indicators to construct an m-order fuzzy complementary judgment matrix. :

[0094] ;

[0095] The matrix satisfies the fuzzy complementarity constraint:

[0096] ;

[0097] In the formula, Let i be the fuzzy importance scale relative to ij. This indicates that the indicators themselves are equally important; A value greater than 0.5 indicates that indicator i is more important than indicator j. <0.5 indicates that indicator i is weaker than indicator j.

[0098] S22. Perform consistency correction on the initially constructed fuzzy complementary judgment matrix. By calculating the average difference between each row element and other row elements, correct the judgment logic contradictions existing in the matrix elements, and obtain a fuzzy consistent matrix that meets the fuzzy consistency requirements.

[0099] Specifically, to avoid confusion in the judgment logic, the initial fuzzy matrix is ​​corrected for consistency to obtain a fuzzy consistent matrix. The corrected formula is:

[0100] ;

[0101] In the formula, The elements are the elements of the corrected matrix; n is the total number of evaluation indicators; the corrected matrix satisfies fuzzy consistency logic and has no contradictory judgments.

[0102] S23. Calculate the initial fuzzy subjective weights for each evaluation indicator using the row sum method, and normalize all the initial fuzzy subjective weights to obtain the subjective weights that reflect the importance of the indicator mechanism.

[0103] Specifically, the initial fuzzy subjective weights of each indicator are calculated using the row summation method:

[0104] ;

[0105] In the formula, Let the initial fuzzy subjective weights be the i-th index; normalize the initial fuzzy subjective weights to obtain the final FAHP subjective weights. With subjective weight set :

[0106] ;

[0107] ;

[0108] In the description of this invention, the calculation of objective weights using the criterion removal effect method includes:

[0109] S24. Construct a normalized decision matrix based on the multimodal feature dataset, and uniformly map the feature values ​​of all evaluation indicators to the interval [0, 1] to obtain standardized indicator feature values.

[0110] Specifically, construct the normalized decision matrix:

[0111] ;

[0112] In the formula, For the i-th sample and the j-th indicator, the original value after preprocessing is...

[0113] Data standardization (linear normalization, dimensionless): unifies the characteristics of PRPD amplitude, time-domain pulse number, ultrasound mV, and high-frequency current mA of different modes and magnitudes to the [0, 1] interval.

[0114] ;

[0115] In the formula, For the i-th sample, the standardized value of the j-th index (0≤ ≤1); B is a set of benefit-type indicators.

[0116] S25. For each sample in the normalized decision matrix, calculate the average of the absolute logarithms of the characteristic values ​​of all standardized indicators for that sample, and substitute them into the quantitative formula to obtain the comprehensive performance value of all indicators for that sample.

[0117] Specifically, the quantitative calculation of the overall performance value of the standard state sample:

[0118] ;

[0119] In the formula, This represents the comprehensive performance value of all indicators for the i-th sample. To preprocess the normalized index feature values; to amplify the differences in weak fault features through logarithmic transformation, adapting to early latent fault identification scenarios.

[0120] S26. Remove each evaluation indicator individually in turn, keeping all other evaluation indicators unchanged, and recalculate the overall performance value of each sample to obtain the reconstructed performance value after removing the evaluation indicator.

[0121] Specifically, after removing individual indicators, the performance values ​​are reconstructed. The j-th indicator is removed one by one, while the remaining indicators remain unchanged, and the sample performance values ​​are recalculated. The formula represents the overall evaluation level after the absence of this indicator:

[0122] ;

[0123] In the formula, The reconstruction performance of the i-th sample after removing the j-th metric; Exclude the j-th indicator and calculate using only the remaining indicators, where i=1,2....,m; j=1,2.....,n.

[0124] S27. Calculate the sum of the absolute differences between the reconstructed performance values ​​of all samples and the comprehensive performance values ​​of all indicators to obtain the removal effect value of the corresponding evaluation indicator.

[0125] Specifically, the calculation of the indicator removes the disturbance effect value:

[0126] ;

[0127] In the formula, Let be the removal effect value of the i-th indicator; The larger the value, the more drastic the fluctuation in the evaluation results after the missing indicator, indicating that the indicator has a stronger objective distinguishing ability and higher data sensitivity in assessing the health status of transformers.

[0128] S28. Normalize the removal effect values ​​of all indicators so that the sum of all weights is 1, thus obtaining the objective weights that reflect the objective contribution of the indicator data.

[0129] Specifically, objective weight normalization:

[0130] ;

[0131] ;

[0132] In the formula, Let be the MEREC objective weight of the j-th indicator; It is a set of objective weights.

[0133] In the description of this invention, the optimal combination weights are generated by solving the optimal fusion coefficients of subjective and objective weights through game theory Nash equilibrium, including:

[0134] S29. Based on the calculation results of subjective weights and objective weights, a linear combination weighting model is pre-constructed, and a fusion coefficient is assigned to the subjective weights and objective weights.

[0135] Specifically, a linear combination weighting model is established based on FAHP subjective weights and MEREC objective weights:

[0136] ;

[0137] In the formula, Let be the optimal game fusion coefficient to be solved, with a fixed value range of . This is used to balance the interplay between the weight of subjective experience and the weight of objective data. The FAHP fuzzy subjective weight corresponding to the j-th evaluation index is obtained by fuzzy evaluation of transformer fault mechanism and operation and maintenance experience, and represents the importance of the theoretical mechanism of the index. The objective weight of the j-th evaluation indicator is calculated from the perturbation difference of the multimodal monitoring data and represents the objective data discrimination ability of the indicator. The final optimized combination weights of the j-th indicator are obtained after game equilibrium fusion, taking into account both the rationality of the mechanism and the objectivity of the data.

[0138] S210. Treat subjective weights and objective weights as two game participants, construct an objective function with the goal of minimizing the overall deviation between subjective weights and objective weights, and use the L2 norm operation to quantify the deviation distance between subjective weights and objective weights in the objective function.

[0139] Specifically, the core logic of game theory integration is as follows: using subjective weight sets and objective weight sets as two game participants, iterative optimization minimizes the deviation between the two types of weights and maximizes their fit, thereby achieving a stable Nash equilibrium. The objective function for minimizing weight deviation is constructed as follows:

[0140] ;

[0141] In the formula, It is a 2-norm operation used to quantify the overall deviation distance of the weight vector, which can accurately characterize the overall fit between the combined weights and the single weight set; This is the set of FAHP subjective weight vectors composed of all indicators; The set of objective weight vectors for MEREC, consisting of all indicators; To use the fusion coefficient Let the total deviation of the variable be the objective function. The physical meaning of the objective function is: to solve for the optimal... This allows the game combination weight vector to simultaneously and infinitely approximate both the subjective weight vector of the mechanism and the objective weight vector of the data, achieving the minimum conflict and optimal equilibrium between the two types of weights.

[0142] S211. Set constraints for the fusion of subjective and objective weights, including limiting the fusion coefficient to be in the range of 0 to 1, ensuring that the sum of the optimal combination weights of all evaluation indicators is 1, and that the optimal combination weight of a single item cannot be negative.

[0143] Specifically, to ensure the effectiveness and physical rationality of the weights, the following constraints are set:

[0144] ;

[0145] In the formula, the first row of constraints To constrain the values ​​of the fusion coefficients, ensuring the weight fusion ratio has physical validity and avoiding negative or excessive weights; the second row of constraints... The third row of constraints is for weight normalization, ensuring that the final combined weights of all evaluation indicators sum to 1, satisfying the basic weight rules of the comprehensive evaluation system; The weights are non-negative to ensure that all indicator weights are positive and effective contribution values, and there are no negative and invalid weights.

[0146] The constraints ensure that the fusion coefficients are valid and that the final combined weights meet the requirements of normalization and non-negativity, conforming to the basic physical rules of evaluation index weights.

[0147] S212. Take the first derivative of the objective function and set the derivative value to zero to find the extreme points of the objective function. Then, obtain the unique optimal solution of the fusion coefficient through analytical calculation, which is taken as the optimal fusion coefficient.

[0148] Specifically, by taking the first derivative of the objective function and setting it to zero, the extreme points are solved to obtain the unique optimal analytical solution for the fusion coefficients.

[0149] ;

[0150] In the formula, The squared L2 norm of the difference between the subjective and objective weight vectors is used to quantify the overall difference between the two sets of weights. By taking the first derivative of the objective function with respect to the bias and setting it to zero to find the extreme value, a unique analytical optimal solution can be obtained. This value can achieve a balanced game between subjective and objective weights, without favoring human operation and maintenance experience or single data statistical characteristics, and fully satisfies the minimum conflict stability condition of Nash equilibrium.

[0151] Through iterative engineering corrections, the final optimal fusion coefficients satisfy the following: minimizing the conflict between subjective and objective weights, ensuring that the combined weights are neither biased towards purely human experience nor purely data features, thus achieving a stable Nash equilibrium.

[0152] S213. Substitute the optimal fusion coefficient into the linear combination weighting model, calculate the initial optimal combination weights of all evaluation indicators, and perform global normalization on the optimal combination weights to ensure that the sum of all optimal combination weights is 1 and all are non-negative, thus obtaining the optimal combination weight matrix.

[0153] Specifically, the optimal fusion coefficients obtained from the solution... Substituting into the linear combination model, we calculate the initial combination weights of all indicators, and then perform global normalization to satisfy the constraint that the sum of the weights is 1, thus obtaining the final optimal game-theoretic combination weight matrix for health assessment:

[0154] ;

[0155] In the formula, This is the final solution yielding the optimal combination weight vector matrix for the game. These are the final optimized weights corresponding to the 1st to the mth evaluation indicators, respectively; The global normalization constraint ensures that the weights of all indicators are reasonably allocated and the system is closed, and can be directly used for subsequent multimodal evidence fusion and quantitative assessment of health status.

[0156] It is evident that traditional weighting methods often employ manually set fixed fusion coefficients, resulting in high randomness and poor adaptability, failing to accommodate the differentiated characteristics of multimodal heterogeneous transformer data. This invention utilizes game theory Nash equilibrium modeling, treating subjective mechanism weights and objective data weights as the two sides in a game. With minimizing weight deviation as the equilibrium objective, it adaptively solves for the optimal fusion coefficients without manual intervention. This approach retains the FAHP algorithm's prior knowledge of transformer fault mechanisms and maintenance experience while fully leveraging the MEREC algorithm's objective ability to uncover multimodal data perturbation characteristics and subtle fault differences. It completely overcomes the limitations of single weighting methods, eliminates the game conflict between subjective and objective weights, and ensures that the final weights possess both mechanistic rationality and data objectivity, significantly improving the accuracy and robustness of transformer health assessment.

[0157] S3. Based on the multimodal feature dataset and the optimal combined weight, the improved evidence theory is used to fuse multi-source evidence to obtain the global confidence level. The confidence level of the basic health level and the confidence level of the coupled fault warning are extracted to form a five-dimensional evaluation result, and the quantitative comprehensive health index is calculated.

[0158] In the description of this invention, as Figure 4As shown, based on a multimodal feature dataset and optimal combined weights, an improved evidence theory is used to fuse multi-source evidence to obtain global confidence. The basic health level confidence and coupled fault warning confidence are extracted to form a five-dimensional evaluation result, and a quantified comprehensive health index is calculated, including:

[0159] S31. Based on the multimodal feature dataset and the optimal combination weight, the feature data of each modality are weighted and summed to calculate the initial probability distribution of the three independent evidence sources (partial discharge mode, oil chromatography chemical mode, and temperature thermal mode) for the four basic health levels. The confidence level of each set of evidence is calculated using Shannon entropy. The confidence level is multiplied by the corresponding initial probability distribution to generate the basic probability assignment of each set of evidence. An initial unknown confidence level representing uncertainty is assigned to each set of evidence.

[0160] Specifically, the confidence score of the q-th group of state information is calculated based on the Shannon entropy definition:

[0161] ;

[0162] In the formula, Let be the probability (or normalized statistic) of the k-th initial state category in the q-th data group, where k represents the four insulation state levels.

[0163] The basic probability distribution in the transformed DS evidence theory:

[0164] ;

[0165] In the formula, The kth insulation health level; For the qth group of evidence, the level is... The basic probability distribution; For the qth group of evidence, assign it to the unknown BPA.

[0166] S32. Based on the data noise level and fluctuation stability of each set of evidence, dynamically calculate the adaptive discount coefficient, and use the discount coefficient to make a secondary correction to the basic probability allocation of each evidence source.

[0167] In the description of this invention, the adaptive discount coefficient is dynamically calculated based on the data noise level and fluctuation stability of each set of evidence, and the basic probability allocation of each evidence source is then corrected using the discount coefficient, including:

[0168] S321. Calculate the standard deviation and outlier percentage of the feature data corresponding to the multi-source evidence respectively, and dynamically determine the adaptive discount coefficient for each group of evidence based on the statistical results.

[0169] S322. Multiply the certainty reliability of each basic health level in the basic probability allocation of each evidence group by the adaptive discount coefficient corresponding to that evidence to obtain the corrected certainty reliability of each basic health level for each evidence group.

[0170] S323. Update the unknown reliability of each set of evidence by subtracting (with a value of 1) the difference between the baseline value and the sum of all corrected basic health level certainty reliability of the evidence, and use the difference as the corrected unknown reliability of the evidence, thus completing the second correction of the three types of modal evidence.

[0171] Specifically, in multi-source information fusion, the reliability of different evidence sources (such as different sensors or different projects) varies, requiring the use of discount factors. Further refine the BPA by reducing the weight of unreliable evidence, and calculate the basic probability distribution after discounting:

[0172] ;

[0173] In the formula, is the adaptive discount coefficient for the qth modal evidence. It is dynamically selected based on the noise level and abnormal fluctuation of each group of data. The value range is [0, 1]. The greater the data interference, the worse the stability, and the more obvious the evidence conflict, the smaller the discount coefficient value. This can accurately weaken the interference weight of unreliable evidence on the fusion result. After a second discount correction, the final effective basic probability assignment value of the qth modal evidence to the kth insulation health level is given, and the effective deterministic reliability for subsequent fusion is given. As the probability of uncertainty in unknown states after discount correction, this correction method can effectively distinguish between valid fault information and interference noise, solve the technical problem of inconsistent reliability of multi-source heterogeneous evidence, and greatly improve the anti-interference ability and result stability of multi-source evidence fusion.

[0174] S33. Using the Dempster combination rule, the three types of modal evidence after the basic probability allocation is modified are cross-fused in pairs to obtain the global basic probability allocation; the global basic probability allocation includes four types of basic health level reliability and global unknown uncertainty reliability.

[0175] Specifically, the inter-group fusion is performed using Dempster's combination rule:

[0176] ;

[0177] In the formula, The Dempster evidence combination operator represents the multi-source evidence cross-fusion calculation rule, which can realize the superposition and complementary update of probability information of multiple independent evidences; These are the probability distributions of independent evidence bodies after correction for three modes: electrical partial discharge, oil chromatography, and temperature-thermal engineering, as well as the comprehensive benchmark mode. The global comprehensive basic probability allocation result obtained after cross-fusion of all modal evidence fully integrates multi-dimensional fault feature information, providing core data support for subsequent health status discrimination and coupled fault identification.

[0178] S34. Perform coupled fault threshold discrimination on the global unknown and uncertain confidence, and perform differentiated confidence allocation based on the discrimination results to obtain the final confidence of four basic health levels and coupled fault warning confidence, thus forming a five-dimensional evaluation result.

[0179] Traditional DS evidence fusion methods often employ an average allocation of unknown confidence levels m(γ), which has significant technical drawbacks. During transformer multimodal monitoring, multiple sources of evidence, such as electrical partial discharge, gas generation from oil chromatography, and temperature rise, are prone to local conflicts and feature coupling superposition. The unknown uncertainty confidence level m(γ) generated by these conflicts is not meaningless noise, but rather likely corresponds to coupled latent faults (thermoelectric coupling faults, electrochemical coupling faults) that cannot be identified by a single sensor and are induced by the superposition of multiple features. Simply allocating the unknown confidence level would completely erase the implicit information of coupled faults, leading to missed or misjudged early coupled defects and an inability to distinguish between single faults and complex coupled faults.

[0180] To address the pain points of this technology, this invention abandons the traditional equal distribution strategy and designs an adaptive threshold discrimination and coupled fault confidence interception mechanism. It introduces a brand-new coupled fault latent state and realizes a five-dimensional output of four basic health states and coupled fault early warning states, accurately identifying multimodal coupled latent faults.

[0181] In the description of this invention, coupled fault threshold discrimination is performed on the globally unknown and uncertain confidence level. Based on the discrimination result, differentiated confidence level allocation is performed to obtain the final confidence levels of four basic health levels and coupled fault warning confidence levels, constituting a five-dimensional evaluation result including:

[0182] S341. Based on the preset coupling fault judgment threshold, determine the relationship between the numerical values ​​of the global unknown uncertainty confidence.

[0183] Specifically, a preset coupling fault detection threshold is set. Based on transformer engineering operation and maintenance experience and the fluctuation characteristics of multimodal data, a fixed judgment threshold is set. Unknown reliability after global fusion Threshold discrimination is performed.

[0184] S342. If the global unknown uncertainty confidence is less than the preset coupled fault judgment threshold, then the uncoupled latent fault is judged, and the global unknown uncertainty confidence is evenly distributed to the four basic health levels to obtain the final confidence of each basic level.

[0185] Specifically, when When there is no obvious conflict of evidence, and the uncertain component is caused by minor noise in the data, a minor equalization correction is adopted to ensure the accuracy of the basic level.

[0186] S343. If the global unknown uncertainty confidence level is greater than the preset coupling fault judgment threshold, then a coupling latent fault is determined to exist. The part exceeding the threshold is retained as the coupling fault warning confidence level, and only the threshold part is evenly distributed to the four basic health levels.

[0187] Specifically, when When it is determined that there is significant conflict and feature coupling superposition in the multimodal evidence, and there is a potential for latent coupling failure, the coupling failure determination module is immediately activated, and most of the unknown confidence is retained as the exclusive confidence of coupling failure, and no longer all of them are allocated to the four basic states.

[0188] S344. Integrate the final confidence scores of the four basic health levels with the coupled fault warning confidence scores to output a five-dimensional evaluation result.

[0189] Specifically, for conventional uncoupled operating conditions ( At this point, the unknown confidence level is generated by random noise and has no coupled fault characteristics. The trace amount of unknown confidence level is evenly distributed to the four basic health levels to ensure the accuracy of the basic assessment.

[0190] ;

[0191] In the formula, This represents the final overall confidence level for the basic insulation health class of type k. This is the certainty reliability for direct assignment to each level after evidence fusion; This refers to the unknown and uncertain confidence components after global fusion.

[0192] For coupled fault latent conditions ( At this point, the unknown confidence level contains a large amount of implicit information about multimodal coupled faults. To avoid the loss of coupled features, this invention adopts a primary-secondary separation confidence allocation mechanism: most of the unknown confidence level is retained as the confidence level for coupled faults, and only the remaining small amount of confidence level is evenly distributed to the basic state. The specific formula is as follows:

[0193] ;

[0194] In the formula, To add a confidence level for coupled fault early warning, specifically characterizing the severity of composite latent faults such as thermoelectric coupling and electrochemical coupling in transformers; The preset coupling determination threshold is used; The confidence levels for the four basic health levels are as follows.

[0195] Based on the above calculation results, this invention breaks through the limitations of the traditional four-state output and constructs a five-dimensional evaluation output system consisting of four basic states and coupled early warning states:

[0196] Basic states: Normal state, Attention state, Abnormal state, Critical fault state;

[0197] New hidden state: Coupled fault warning state (corresponding to multi-mode composite latent faults such as partial discharge-overheating, discharge-aging).

[0198] when The higher the value, the stronger the conflict of the current transformer multimodal data and the higher the degree of feature coupling. The greater the probability of the equipment having unmanifested coupled latent faults, the more likely it is to achieve early composite fault prediction that traditional methods cannot cover.

[0199] S35. Multiply the final confidence scores of the four basic health levels by the corresponding hazard gradient weights and sum them to calculate the quantified comprehensive health index at the current moment.

[0200] Specifically, the formula for calculating the insulation health level index is as follows:

[0201] ;

[0202] In the formula, The transformer insulation health level index is the final normalized index. The value range is close to the health evaluation range. The smaller the value, the worse the insulation health status of the equipment, and the higher the degree of deterioration and failure. The weighting coefficient for the k-th insulation health level is set according to the severity of transformer faults and the gradient of maintenance priorities. The higher the fault level and the greater the harm, the larger the corresponding weighting coefficient, which can effectively amplify the assessment weight of serious faults and meet the actual needs of power equipment operation and maintenance. The final comprehensive confidence level corresponding to the k-th basic insulation health level is obtained by weighted summation, which transforms the discrete level confidence levels into a continuous quantitative health index, thereby achieving a refined and quantitative assessment of the equipment health status.

[0203] Traditional DS evidence fusion algorithms suffer from two inherent technical flaws: First, multi-source heterogeneous modal data are prone to high evidence conflict, leading to result distortion and evaluation failure in conventional fusion methods. Second, uniformly distributing unknown and uncertain confidence levels completely loses the implicit features of multi-modal coupled faults, only identifying single explicit faults and failing to predict early latent compound faults, resulting in high rates of missed and false positives. This invention constructs a triple-improved architecture of entropy-weighted confidence initial correction, adaptive discount secondary correction, and unknown confidence coupling discrimination. First, it eliminates noise interference from single-modal evidence itself through Shannon entropy; second, it balances the reliability differences of multi-source evidence through adaptive discount coefficients; and finally, it abandons the traditional uniform distribution strategy and innovatively introduces threshold discrimination and coupled fault interception mechanisms to transform invalid uncertain noise into effective coupled fault early warning features. Compared with existing technologies, this invention adds an independent dimension for identifying coupled faults, realizing a dual-output mode of basic fault level determination and compound coupled fault early warning. It solves the technical problems of existing technologies being unable to identify multimodal coupled latent faults in transformers and missing early compound faults, greatly improving the comprehensiveness, foresight and accuracy of multimodal fault assessment, and has significant technological progress and inventiveness.

[0204] S4. Based on the comprehensive health index of continuous monitoring, a long-term health sequence reflecting the long-term deterioration trend of insulation is constructed, and a degradation model that fits the nonlinear aging law of the oil-paper insulation is fitted. Based on the preset insulation failure threshold, the critical failure time is solved and the remaining safe service life of the transformer is calculated.

[0205] Specifically, this invention utilizes modal adaptive preprocessing, FAHP-MEREC fuzzy game-based weighting, and improved DS evidence theory multi-source fusion technology to achieve real-time static health status assessment of transformers. It can accurately output the current explicit health level, multi-dimensional aging confidence, and coupled latent fault warning results. To achieve a closed-loop assessment process from current state diagnosis and long-term degradation projection to remaining service life prediction, this invention uses the comprehensive health status fused with continuous time-series DS evidence as the core characterization quantity for insulation degradation, depicting the long-term performance degradation and damage accumulation trend of transformers. Combining the insulation attenuation degradation model and engineering failure threshold criteria, it constructs a quantitative prediction model for the remaining service life (RUL) of transformers, achieving refined quantification of the transformer's entire lifecycle status.

[0206] This RUL prediction model uses a time-series continuous comprehensive health index sequence as the basis for degradation, and sequentially completes the construction of time-series health, noise smoothing and denoising, degradation model fitting, failure threshold determination, critical time solution, and remaining lifetime calculation.

[0207] In the description of this invention, as Figure 5As shown, a long-term health sequence reflecting the long-term deterioration trend of insulation is constructed based on a comprehensive health index from continuous monitoring, and a degradation model suitable for the nonlinear aging law of oil-paper insulation is fitted. Based on a preset insulation failure threshold, the critical failure time is solved and the remaining safe service life of the transformer is calculated, including:

[0208] S41. Collect comprehensive health indices at multiple monitoring times continuously to construct a long-term health sequence that reflects the long-term deterioration trend of transformer insulation.

[0209] Specifically, after improving the DS evidence theory multimodal fusion decision, each real-time monitoring moment can output four independent confidence levels of basic insulation health status, namely, the confidence level of the normal state. Pay attention to the confidence level of the state. Confidence of abnormal states Confidence level of severe fault conditions Based on four types of state confidence, a weighted comprehensive health evaluation index for the transformer at time t is constructed. To achieve multi-confidence fusion quantization, the formula is as follows:

[0210] After DS evidence theory decision fusion, confidence levels for four health states can be obtained at each monitoring time point, namely: normal, Notice, abnormal, serious.

[0211] At this point, a comprehensive health evaluation index at time t is constructed. The formula is as follows:

[0212] ;

[0213] In the formula, The overall health status of the transformer at time t, with a value range of [0, 1]; , , , Quantify the weighting coefficients for each health status level; - The confidence level is configured for the basic probability scores of each state after DS evidence fusion.

[0214] Continuous collection of time-series data over multiple days / months creates a degradation time-series dataset. This series visually reflects the year-on-year / month-on-month decline in transformer health and the cumulative degradation pattern of damage. The formula is as follows:

[0215] ;

[0216] S42. The moving average filtering algorithm is used to smooth and denoise the long time series health status sequence, and the original value at the current time is replaced by the average health status within a time window before and after the current time.

[0217] Specifically, during transformer field monitoring, accidental factors such as short-term load fluctuations, sudden changes in ambient temperature, minor sensor interference, and electromagnetic disturbances in the power grid can cause irregular spikes and random noise in the health time series curve, failing to accurately reflect the long-term aging trend of the insulation. To eliminate instantaneous operating condition interference and retain inherent degradation patterns, this invention employs a moving average filtering algorithm to smooth and correct the original health sequence, as shown in the following formula:

[0218] ;

[0219] In the formula, The current health status is obtained after smoothing and correction by sliding filter, eliminating random disturbances and retaining the true aging trend; To preset the sliding window length, a fixed timing step size is set based on the slow aging characteristics of the transformer, balancing sequence smoothness and trend sensitivity; Before the current moment The raw, comprehensive health data at any given moment. By replacing the instantaneous values ​​with a windowed mean, high-frequency random interference can be effectively filtered out, restoring the long-term steady-state degradation trend of the insulation.

[0220] S43. A first-order exponential decay model is used to fit the smoothed long-term health sequence. The constant parameters and aging rate coefficients in the first-order exponential decay model are solved by the least squares method to obtain a degradation model that adapts to the nonlinear accelerated aging law of transformer oil-paper insulation.

[0221] Specifically, transformer oil-paper insulation aging, dielectric loss accumulation, partial discharge damage, and thermal aging deterioration all conform to a degradation law of slow initial decay followed by rapid, nonlinear accelerated decay. The exponential degradation model can accurately adapt to the physical mechanism of insulation deterioration in power equipment. This invention uses a first-order exponential decay model to construct a transformer health degradation prediction model, as shown in the following formula:

[0222] ;

[0223] In the formula, t represents the health status after smoothing; A and B are the constant fitting parameters of the degradation model, obtained by least squares fitting of historical data; This is the aging rate coefficient; This represents the cumulative operating time of the transformer.

[0224] S44. Set the critical health threshold for insulation failure according to the power industry standard, substitute the critical health threshold for insulation failure into the fitted degradation model, and solve the critical failure time required for the transformer health to decay from the current level to the failure threshold through logarithmic inverse transformation.

[0225] Specifically, such as Figure 3 As shown, the present invention defines the health level classification threshold and the failure threshold. It is strictly formulated in accordance with the "DL / T596-2021 Preventive Test Procedure for Power Equipment", "Guidelines for Condition-Based Maintenance of Transformers" and the on-site operation and maintenance evaluation standards for power grid transformers. Combining the multimodal electrical, chemical and thermal coupling degradation characteristics, it divides the insulation health status into four levels, realizes the accurate mapping from the continuous health index H(t) to the discrete qualitative status, and completely solves the public defect of having only quantitative values ​​and no qualitative judgment.

[0226] Quantitative classification criteria for transformer insulation health levels:

[0227] (1) Normal state (Level 1 health): 0.85 < H(t) ≤ 1.0, the equipment has excellent insulation performance, no partial discharge, no abnormal gas generation, normal temperature and operating conditions, no aging and deterioration trend, and all monitoring indicators are within the standard allowable range. It can operate stably for a long time without maintenance.

[0228] (2) Attention status (Level 2 health): 0.70 < H(t) ≤ 0.85, the equipment has slight insulation aging, trace gas production, occasional weak partial discharge or small temperature deviation, no obvious fault, slow deterioration rate, which belongs to early latent deterioration, and the monitoring frequency needs to be increased to track the status changes.

[0229] (3) Abnormal state (Level 3 health): 0.50 < H(t) ≤ 0.70, the equipment insulation has deteriorated significantly, there is continuous partial discharge, stable growth of characteristic gas or abnormal temperature deviation, the insulation damage continues to accumulate, there is a risk of hidden faults, and planned maintenance and targeted defect elimination need to be arranged.

[0230] (4) Severe fault state (Level 4 health): 0≤H(t)≤0.50, the equipment insulation is severely aged, the multi-modal data is significantly abnormal, there is discharge and overheating coupling fault, the insulation damage is approaching the critical failure, the operation risk is extremely high, and it is necessary to immediately stop the machine for maintenance or scrap and replace it.

[0231] Based on the power industry's maintenance standards and the bottom line for safe equipment operation, a critical value for severe fault conditions is selected as the failure threshold, and a critical health threshold for failure is set. =0.5. When the overall health of the equipment continuously drops to 0.5 or below, the transformer insulation performance is deemed to have reached a critical failure state, and it no longer meets the conditions for safe continuous operation. It must be taken out of service for maintenance or replaced. The threshold constraints are as follows:

[0232] ;

[0233] In the formula, The critical health threshold for transformer insulation failure is set at 0.5, which has clear industry standards and engineering operation and maintenance basis.

[0234] S45. Obtain the current cumulative operating time of the transformer, subtract the current cumulative operating time from the critical failure time, and obtain the remaining safe service life of the transformer.

[0235] Specifically, by substituting the failure threshold into the degradation model and using logarithmic identity transformation, the critical failure time corresponding to the decline of the transformer's health to the failure threshold is solved. The formula is as follows:

[0236] ;

[0237] ;

[0238] In the formula, The critical operating time at which a transformer reaches the critical state of insulation failure and requires maintenance or scrapping; A is a fixed failure health threshold; A and B are the individual degradation parameters of a single device obtained by model fitting; the accurate solution of the nonlinear degradation model is achieved through logarithmic transformation, avoiding iterative solution errors.

[0239] Let the current time be The final remaining service life calculation formula is as follows:

[0240] ;

[0241] This method organically combines instantaneous health diagnosis with long-term degradation evolution, fully considering the multi-dimensional coupled aging effects of partial discharge, oil chromatography, and temperature. It can objectively reflect the cumulative damage and performance degradation law of transformer insulation, and provide a quantitative theoretical basis for transformer condition-based maintenance, full life cycle management, and aging equipment replacement decisions.

[0242] According to another embodiment of the present invention, a transformer multimodal coupling health assessment system based on FAHP-MEREC game-theoretic weighting is also provided, the system comprising:

[0243] The acquisition and preprocessing module is used to synchronously acquire multimodal data of the transformer and use a dedicated standardized algorithm to perform adaptive preprocessing of various types of multimodal data to construct a multimodal feature dataset. The multimodal data includes partial discharge data, oil chromatography chemical data, and temperature and thermal data.

[0244] The game theory combination weighting module is used to extract evaluation indicators for each modality based on a multimodal feature dataset. It calculates subjective weights using fuzzy hierarchical analysis and objective weights using the criterion removal effect method. Finally, it solves for the optimal fusion coefficient between subjective and objective weights through game theory Nash equilibrium to generate the optimal combination weights.

[0245] The evidence fusion evaluation module is used to fuse multi-source evidence based on multimodal feature datasets and optimal combined weights, and adopts improved evidence theory to obtain global confidence. It also extracts the basic health level confidence and coupled fault warning confidence to form a five-dimensional evaluation result, and calculates a quantitative comprehensive health index.

[0246] The remaining life prediction module is used to construct a long-term health sequence reflecting the long-term deterioration trend of insulation based on the comprehensive health index of continuous monitoring, and to fit a degradation model that fits the nonlinear aging law of the oil-paper insulation. Based on the preset insulation failure threshold, it solves the critical failure time and calculates the remaining safe service life of the transformer.

[0247] The present invention will be further described below with reference to specific embodiments.

[0248] A 110kV oil-immersed power transformer in operation of a power grid was selected as the test object. The equipment has been in operation for 10.5 years and exhibits slight insulation aging, intermittent weak partial discharge, and trace gas generation, suggesting potential early thermoelectric coupling latent degradation. Multimodal monitoring time-series data of the equipment over the past 12 months were continuously collected, including raw monitoring data for partial discharge electrical modes, oil chromatography chemical modes, and temperature thermal modes, which served as the input sample set for this technical solution.

[0249] I. Key Fixed Parameter Engineering Values: (1) Health Level Weighting Coefficient: Assigned based on the differential values ​​of fault hazard gradient, (2) Sliding filter window length: Based on the slow aging characteristics of the transformer, set the sliding filter window length. (3) Coupling fault determination threshold: (4) Insulation failure critical health threshold: set according to industry standards (5) Time series unit: The basic time series unit is the month, and the cumulative runtime is calculated. moon.

[0250] II. Modal Preprocessing and Standardization Intermediate Results: The original multimodal data were processed using the proprietary modal preprocessing strategy of this invention: extreme value normalization was used for partial discharge gas mode, Z-score baseline correction standardization was used for oil chromatography chemical mode, and residual range standardization was used for temperature thermal mode. After removing the specific interference of each mode, a regular and equivalent multimodal feature dataset was obtained, eliminating the differences in dimensions and heterogeneity, and completing the preprocessing of the data.

[0251] III. FAHP-MEREC Game Combinatorial Weighting Results: Mechanism weights were obtained through FAHP fuzzy subjective weighting, and data weights were obtained through MEREC data objective weighting. The optimal fusion coefficients were then adaptively solved using game theory Nash equilibrium. The final index weights are obtained through balanced integration, with the following weights: temperature thermal mode weight 0.38, oil chromatography chemical mode weight 0.34, and partial discharge gas mode weight 0.28. This aligns with the fault mechanism of transformer insulation aging, which is characterized by thermal aging and chemical degradation as the core, and electrical discharge as a derivative defect.

[0252] IV. Improved DS Evidence Fusion State Assessment Results: After multi-source evidence confidence correction, discount-based secondary correction, and adaptive allocation of conflict uncertainty confidence, the confidence levels of each health state at the current time are obtained: (normal), (Notice), (abnormal), (Severe); Fusion of unknown reliability This triggers the coupling fault determination module.

[0253] Confidence of intercepted coupling faults It determines that the equipment has an early thermoelectric coupling latent degradation trend, and outputs a five-dimensional evaluation result of four basic states and coupling early warning, which makes up for the deficiency of traditional methods in the ability to identify coupling faults.

[0254] V. Comprehensive Health Index Calculation Results: Substitute the results into the health weighting formula to calculate the current comprehensive health index:

[0255] ;

[0256] According to the Level 4 health rating criteria: 0.50 < 0.678 ≤ 0.70, the equipment is currently in an abnormal state, with significant cumulative insulation degradation. Combined with the coupling warning results, it is determined that there is a risk of latent thermoelectric coupling fault, which is consistent with the actual operation and maintenance status of the equipment.

[0257] VI. Results of Time-Series Smoothing and Degradation Model Fitting: A k=5 moving average filter was applied to the 12-month time-series health data to remove random spikes caused by load fluctuations and environmental disturbances, restoring the steady-state degradation trend. An exponential degradation model was fitted using the least squares method to obtain the specific degradation parameters for each device. The final degradation model is:

[0258] ;

[0259] VII. Critical Failure Time and Final RUL Prediction Results: The failure threshold... Substituting into the model, we can solve for the critical failure time:

[0260] moon;

[0261] Combined with the current runtime Monthly calculation of remaining useful life:

[0262] Months ≈ 5.53 years;

[0263] VIII. Final Conclusion of the Example: The current insulation condition of this 110kV transformer is abnormal, with early thermoelectric coupling latent faults. The insulation deterioration is accelerating, and the remaining safe service life is predicted to be about 5.53 years. It is recommended that the operation and maintenance unit increase the monitoring frequency and formulate planned maintenance and insulation defect elimination plans in advance to avoid sudden operation risks caused by coupling faults.

[0264] To verify the superiority and unexpected technical effects of the present invention, a comparative experiment was conducted between the method of the present invention and three traditional mainstream assessment and prediction methods: the traditional single oil chromatography three ratio method, the conventional AHP-DS fusion method, and the traditional index degradation RUL prediction method. The same transformer, the same time series dataset, and the same failure threshold standard were used. The accuracy of state identification, the rate of missed detection of coupled faults, and the mean absolute error of lifetime prediction (MAE) were used as quantitative evaluation indicators. The comparison results are as follows.

[0265] I. Definition of comparative experimental indicators: (1) Status recognition accuracy: the degree of consistency between the equipment health status judgment and the actual operation and maintenance conclusion. The higher the accuracy, the more accurate the status assessment. (2) Coupled fault omission rate: the probability of missing identification of thermoelectric and electrochemical latent coupled faults. The lower the value, the stronger the early fault perception capability. (3) Life prediction MAE: the absolute error between the predicted life and the actual decommissioning / maintenance life. The smaller the value, the higher the prediction accuracy and stability.

[0266] II. Quantitative comparison results of multiple methods: (1) Traditional single oil chromatography three ratio method: state identification accuracy 78.2%, coupling fault false negative rate 42.5%, lifetime prediction MAE = 11.4 months; (2) Conventional AHP-DS fusion method: state identification accuracy 85.6%, coupling fault false negative rate 28.3%, lifetime prediction MAE = 7.8 months; (3) Traditional non-game weighted and average reliability DS method: state identification accuracy 88.1%, coupling fault false negative rate 25.1%, lifetime prediction MAE = 6.5 months; (4) The FAHP-MEREC game weighted and improved coupling discrimination DS-RUL method of this invention: state identification accuracy 96.7%, coupling fault false negative rate 3.8%, lifetime prediction MAE = 2.1 months.

[0267] III. Comparative Experimental Conclusions and Creative Technical Effects: (1) Compared with the traditional single-parameter method: The multimodal fusion system of this invention fully covers electrical, chemical, and thermal multi-dimensional degradation information, completely solves the problem of one-sided information and poor anti-interference of single index, significantly improves the accuracy of state recognition by 18.5%, and reduces the life prediction error by 81.6%; (2) Compared with the conventional fixed-weight fusion method: The game theory Nash equilibrium adaptive weighting achieves the optimal balance of subjective and objective weights, avoids the bias of manual weights, and abandons the traditional unknown confidence equalization mechanism. A new coupled fault interception and discrimination module is added, which for the first time realizes the effective identification of multimodal coupled latent faults in transformers and eliminates the missed judgment of coupled faults. The rate is reduced by more than 21.3%, which solves the long-standing problem of early composite fault omission in existing technologies; (3) The overall technical effect has the advantages of non-obviousness and unexpectedness: This method achieves multiple technical gains such as improved state diagnosis accuracy, identification of latent faults, and significant reduction of life prediction error through a multi-layer innovative architecture of modal-specific preprocessing, game equilibrium weighting, resource utilization of coupling conflict information, and accurate fitting of time-series degradation. It breaks through the inherent defects of traditional methods that can only identify obvious faults, cannot predict coupling degradation, and have poor prediction stability. It has significant technical advancement and creativity in the field of power transformer insulation health assessment and life prediction.

[0268] In summary, by utilizing the above-mentioned technical solution of this invention, early insulation aging, weak partial discharge, latent thermal anomalies, and multimodal coupled latent faults in transformers can be effectively identified. Compared with traditional single-parameter evaluation, fixed-weight fusion, and conventional evidence fusion methods, this method has higher condition assessment accuracy, stronger anti-interference capability under operating conditions, and better degradation fitting performance. It can accurately characterize the cumulative damage and long-term nonlinear degradation law of transformer insulation, and provide high-precision and highly reliable theoretical support and engineering application basis for early fault warning, differentiated condition maintenance, hidden danger coupled defect management, and full life cycle health management of power transformers.

[0269] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated in this invention, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying drawings may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

Claims

1. A transformer multimodal coupled health assessment method based on FAHP-MEREC game-theoretic weighting, characterized in that, include: Multimodal data of transformers are collected synchronously, and a dedicated standardized algorithm is used to perform adaptive preprocessing of various types of multimodal data to construct a multimodal feature dataset; Multimodal data includes partial discharge data, oil chromatography chemical data, and temperature and thermal data; Evaluation indicators for each modality are extracted based on a multimodal feature dataset. Subjective weights are calculated using fuzzy hierarchical analysis and objective weights are calculated using the criterion removal effect method. The optimal fusion coefficients of subjective and objective weights are then obtained by solving the game theory Nash equilibrium to generate the optimal combined weights. Based on a multimodal feature dataset and optimal combined weights, an improved evidence theory is used to fuse multi-source evidence to obtain global confidence. The confidence of basic health level and coupled fault warning confidence are extracted to form a five-dimensional evaluation result, and a quantitative comprehensive health index is calculated. Based on the comprehensive health index of continuous monitoring, a long-term health sequence reflecting the long-term deterioration trend of insulation is constructed, and a degradation model that fits the nonlinear aging law of the oil-paper insulation is fitted. Based on the preset insulation failure threshold, the critical failure time is solved and the remaining safe service life of the transformer is calculated.

2. The transformer multimodal coupling health assessment method based on FAHP-MEREC game-theoretic weighting as described in claim 1, characterized in that, The synchronous acquisition of multimodal data from oil-immersed transformers, and the use of a dedicated standardization algorithm to perform adaptive preprocessing of various types of multimodal data to construct a multimodal feature dataset, includes: For partial discharge data, an adaptive wavelet threshold denoising algorithm is used to remove on-site electromagnetic white noise, power frequency interference and its harmonic interference, retain the effective discharge pulse waveform, and extract partial discharge fault features; and an extreme value normalization algorithm is used to map all partial discharge fault features to a standardized range. For oil chromatographic chemical data, a time-series moving average baseline correction algorithm is used to separate the slow-changing baseline caused by natural oil aging and seasonal temperature changes from the fast-changing increment caused by insulation faults, eliminate long-term baseline drift and seasonal fluctuation interference, and extract oil chromatographic fault features; and a mean-variance standardization algorithm is used to process the oil chromatographic fault features to correct baseline shift. For temperature and thermal data, a theoretical benchmark model between temperature, real-time load, and ambient temperature is established using multiple linear regression to extract thermal fault characteristics; and the residual range standardization algorithm is used to normalize the thermal fault characteristics. Fault features from various multimodal data are integrated and standardized to form a multimodal feature dataset without modal bias.

3. The transformer multimodal coupling health assessment method based on FAHP-MEREC game weighting as described in claim 2, characterized in that, For temperature and thermal data, a theoretical benchmark model between temperature, real-time load, and ambient temperature is established using a multiple linear regression method to extract thermal fault characteristics. The residual range normalization algorithm is used to normalize the true temperature rise residual, including: Historical data of normal operation of the transformer without faults or maintenance were selected. The real-time load and ambient temperature of the transformer were used as independent variables, and the measured winding or oil temperature was used as the dependent variable. A theoretical reference model was established to calculate the current theoretical reference temperature of the transformer. By subtracting the measured temperature under real-time operating conditions from the theoretical reference temperature, the true temperature rise residual of the equipment body under the condition of eliminating external operating interference is obtained. Based on the calculation results of the theoretical reference temperature and the actual temperature rise residual, the current thermal fault characteristics of the transformer are extracted to realize the quantification of the transformer's thermal anomaly state; the thermal fault characteristics include relative temperature rise characteristics, steady-state temperature deviation characteristics, and short-term temperature change characteristics. The actual difference is obtained by subtracting the minimum value of all actual temperature rise residuals within the preset time window from the actual temperature rise residual of the current transformer equipment body. The reference difference is obtained by subtracting the minimum value of all actual temperature rise residuals within the preset time window from the maximum value of all actual temperature rise residuals within the preset time window. The ratio of the actual difference to the reference difference is calculated to obtain the standardized thermal mode characteristic value.

4. The transformer multimodal coupling health assessment method based on FAHP-MEREC game weighting as described in claim 1, characterized in that, The calculation of subjective weights using fuzzy hierarchical analysis includes: The importance of each evaluation index extracted from the multimodal feature data is compared pairwise, and a fuzzy complementary judgment matrix is ​​constructed based on the importance comparison results. Consistency correction is performed on the initially constructed fuzzy complementary judgment matrix. By calculating the average difference between the elements of each row and the elements of other rows, the logical contradictions in the judgment of the matrix elements are corrected, and a fuzzy consistent matrix that meets the fuzzy consistency requirements is obtained. The initial fuzzy subjective weights of each evaluation index are calculated using the row sum method. All initial fuzzy subjective weights are then normalized to obtain the subjective weights that reflect the importance of the index mechanism.

5. The transformer multimodal coupling health assessment method based on FAHP-MEREC game weighting as described in claim 1, characterized in that, The calculation of objective weights using the criterion removal effect method includes: A normalized decision matrix is ​​constructed based on a multimodal feature dataset, and the feature values ​​of all evaluation indicators are uniformly mapped to the interval [0, 1] to obtain standardized indicator feature values. For each sample in the normalized decision matrix, calculate the average of the absolute logarithms of the eigenvalues ​​of all standardized indicators for that sample, and substitute them into the quantitative formula to obtain the comprehensive performance value of all indicators for that sample. Remove each evaluation indicator individually in turn, keeping all other evaluation indicators unchanged, and recalculate the overall performance value of each sample to obtain the reconstructed performance value after removing the evaluation indicator. The sum of the absolute differences between the reconstructed performance values ​​of all samples and the comprehensive performance values ​​of all indicators is used to obtain the removal effect value of the corresponding evaluation indicator. The removal effect values ​​of all indicators are normalized so that the sum of all weights is one, thus obtaining the objective weights that reflect the objective contribution of the indicator data.

6. The transformer multimodal coupling health assessment method based on FAHP-MEREC game-theoretic weighting as described in claim 1, characterized in that, The process of obtaining the optimal fusion coefficients of subjective and objective weights through game theory Nash equilibrium to generate the optimal combined weights includes: Based on the calculation results of subjective weights and objective weights, a linear combination weighting model is pre-constructed, and a fusion coefficient is assigned to the subjective weights and objective weights; Subjective weights and objective weights are regarded as two game participants. An objective function is constructed with the goal of minimizing the overall deviation between subjective weights and objective weights. The objective function uses L2 norm operation to quantify the deviation distance between subjective weights and objective weights. Set constraints for the fusion of subjective and objective weights, including limiting the fusion coefficient to the range of 0 to 1, ensuring that the sum of the optimal combination weights of all evaluation indicators is 1, and prohibiting any single optimal combination weight from being negative. Find the first derivative of the objective function and set the derivative value to zero to find the extreme points of the objective function. Then, obtain the unique optimal solution of the fusion coefficient through analytical calculation, which is taken as the optimal fusion coefficient. Substitute the optimal fusion coefficients into the linear combination weighting model to calculate the initial optimal combination weights for all evaluation indicators. Then, perform global normalization on the optimal combination weights to ensure that the sum of all optimal combination weights is one and all are non-negative, thus obtaining the optimal combination weight matrix.

7. The transformer multimodal coupling health assessment method based on FAHP-MEREC game weighting as described in claim 1, characterized in that, The method, based on a multimodal feature dataset and optimal combined weights, employs improved evidence theory to fuse multi-source evidence to obtain a global confidence score. It then extracts the basic health level confidence score and the coupled fault warning confidence score to form a five-dimensional evaluation result, and calculates a quantified comprehensive health index, including: Based on the multimodal feature dataset and the optimal combination weights, the feature data of each modality are weighted and summed to calculate the initial probability distribution of the four basic health levels for three independent evidence sources: partial discharge mode, oil chromatography chemical mode, and temperature thermodynamic mode. The confidence level of each set of evidence is calculated using Shannon entropy. The confidence level is multiplied by the corresponding initial probability distribution to generate the basic probability assignment of each set of evidence. An initial unknown confidence level representing uncertainty is assigned to each set of evidence. Based on the data noise level and fluctuation stability of each set of evidence, an adaptive discount coefficient is dynamically calculated, and the basic probability allocation of each evidence source is then corrected using the discount coefficient. Using the Dempster combination rule, the three types of modal evidence after the basic probability assignment is corrected are cross-fused in pairs to obtain the global basic probability assignment; the global basic probability assignment includes four types of basic health level reliability and global unknown uncertainty reliability. The global unknown and uncertain confidence is coupled with the fault threshold discrimination. Based on the discrimination result, the differential confidence allocation is performed to obtain the final confidence of four basic health levels and the confidence of coupled fault warning, which constitutes the five-dimensional evaluation result. The final confidence scores of the four basic health levels are multiplied by their corresponding hazard gradient weights and then summed to obtain the quantified comprehensive health index at the current moment.

8. The transformer multimodal coupling health assessment method based on FAHP-MEREC game weighting according to claim 7, characterized in that, The step of dynamically calculating an adaptive discount coefficient based on the data noise level and fluctuation stability of each set of evidence, and then using the discount coefficient to perform a secondary correction on the basic probability allocation of each evidence source, includes: The standard deviation and outlier ratio of the feature data corresponding to the multi-source evidence are calculated separately, and the adaptive discount coefficient of each group of evidence is dynamically determined based on the statistical results. Multiply the certainty reliability of each basic health level in the basic probability allocation of each evidence group by the adaptive discount coefficient corresponding to that evidence to obtain the corrected certainty reliability of each basic health level for each evidence group. The unknown reliability of each set of evidence is updated by subtracting the sum of all corrected basic health level certainty reliability of that evidence from the baseline value. This difference is used as the corrected unknown reliability of that evidence, thus completing the secondary correction of the three types of modal evidence.

9. The transformer multimodal coupling health assessment method based on FAHP-MEREC game weighting according to claim 7, characterized in that, The process involves performing coupled fault threshold discrimination on the globally unknown and uncertain confidence level, and then assigning differentiated confidence levels based on the discrimination results. This yields four categories of basic health level final confidence levels and coupled fault warning confidence levels, constituting a five-dimensional evaluation result, including: Based on a preset coupling fault judgment threshold, the relationship between the numerical values ​​of global unknown and uncertain confidence is determined. If the global unknown uncertainty confidence is less than the preset coupled fault judgment threshold, then the uncoupled latent fault is judged, and the global unknown uncertainty confidence is evenly distributed to the four basic health levels to obtain the final confidence of each basic level. If the global unknown uncertainty confidence level is greater than the preset coupling fault judgment threshold, then a coupling latent fault is judged to exist. The part exceeding the threshold is retained as the coupling fault warning confidence level, and only the threshold part is evenly distributed to the four basic health levels. The final confidence scores of the four basic health levels are integrated with the confidence scores of coupled fault warnings to output a five-dimensional assessment result.

10. The transformer multimodal coupling health assessment method based on FAHP-MEREC game weighting according to claim 1, characterized in that, The comprehensive health index based on continuous monitoring is used to construct a long-term health sequence that reflects the long-term deterioration trend of insulation, and a degradation model suitable for the nonlinear aging law of oil-impregnated paper insulation is proposed. Based on the preset insulation failure threshold, the critical failure time is solved and the remaining safe service life of the transformer is calculated, including: By continuously collecting comprehensive health indices at multiple monitoring points, a long-term time-series health sequence reflecting the long-term deterioration trend of transformer insulation is constructed. A moving average filtering algorithm is used to smooth and denoise the long-term health status sequence, and the original value at the current time is replaced by the average health status within a time window before and after the current time. A first-order exponential decay model was used to fit the smoothed long-term health sequence. The constant parameters and aging rate coefficients in the first-order exponential decay model were solved by the least squares method to obtain a degradation model that adapts to the nonlinear accelerated aging law of transformer oil-paper insulation. Based on power industry standards, a critical health threshold for insulation failure is set. The critical health threshold for insulation failure is then substituted into the fitted degradation model, and the critical failure time required for the transformer's health to decay from the current level to the failure threshold is solved by inverse logarithmic transformation. The current cumulative operating time of the transformer is obtained, and the remaining safe service life of the transformer is obtained by subtracting the current cumulative operating time from the critical failure time.

Images