Transformer state evaluation method and system based on data augmentation and dynamic weighting cooperation

Abstract

This invention provides a transformer condition assessment method and system that combines data augmentation and dynamic weighting. The method includes: constructing a raw historical monitoring dataset by collecting historical transformer operation data; forming a state-labeled classification dataset; generating supplementary minority class samples to obtain a class-balanced training sample set; fusing and normalizing subjective and objective weights to obtain a static comprehensive weight; calculating the correlation strength between indicators to obtain a state-variable weighting factor; and nonlinearly fusing the state-variable weighting factor with the static comprehensive weight to obtain the final dynamic comprehensive weight. This application, by aligning with engineering experience corresponding to transformer fault mechanisms, fully explores the objective statistical patterns of historical monitoring data, avoiding the experience bias of single subjective weighting and reducing the limitations of single objective weighting being divorced from actual operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of power transformer condition monitoring and assessment technology, and in particular to a transformer condition assessment method and system that combines data augmentation and dynamic weighting. Background Technology

[0002] As the core hub of the power system, the reliability of transformer operation is directly related to the safety and stability of the power grid. With the deepening of smart grid construction, the assessment of transformer status has shifted from traditional periodic maintenance to status assessment based on multi-source monitoring data.

[0003] In practical engineering applications, transformer condition assessment still faces challenges. The determination of assessment index weights often relies on static experience, making it difficult to respond to real-time equipment degradation. Specifically, existing transformer condition assessment methods mostly use the Analytic Hierarchy Process (AHP) or entropy weight method to determine index weights. While these methods take into account expert experience and data objectivity to some extent, they are essentially static weighting methods. The operating state of a transformer is a dynamic evolution process. When a certain index (such as the dissolved gas content in oil) suddenly deteriorates, the static weights cannot be dynamically adjusted accordingly, resulting in a sluggish response of the assessment model to abnormal states and difficulty in accurately capturing fault signs.

[0004] In addition, the scarcity of key fault samples in transformer condition assessment leads to insufficient training of the assessment model, affecting the accuracy of the assessment results. Specifically, the existing assessment system based on historical monitoring data is often limited by the lack of sample size, resulting in the model training being in an "underfitting" state. This lack of data support makes it impossible for the assessment model to accurately characterize the high-dimensional feature distribution of fault samples, resulting in a low recognition rate of abnormal states. At the same time, the lack of sufficient extreme working condition samples makes the model prone to misjudgment when facing sudden faults.

[0005] Currently, there are numerous studies on existing assessment methods and related data enhancement. For example: Li Bo, Zhang Wenqian, Liu Wei, et al. Evaluation of insulation status of oil-immersed transformers based on entropy weight hierarchical method [J]. Electric Power Engineering Technology, 2023, 42(04): 223-230. The entropy weight AHP method proposed in this paper introduces the degree of index degradation to achieve dynamic correction, but the adjustment range of the bottom weight is limited by the subjective judgment of the top layer, which weakens the dominant role of objective data; Zhang Shaofeng, Ma Deying, Wu Xibo, et al. Research on combined diagnostic method of transformer key parameter fusion [J]. Journal of Chongqing University of Technology (Natural Science), 2021, 35(11): 198-206. The paper improves the fault identification performance through multi-criteria fusion, reflecting the potential of integrated diagnosis, but its criterion weights still mostly rely on static settings.

[0006] To address the aforementioned issues, a transformer condition assessment method and system based on the synergy of data augmentation and dynamic weighting is now designed. Summary of the Invention

[0007] This application provides a transformer condition assessment method and system that combines data augmentation and dynamic weighting to address the problems in related technologies where static weights cannot adapt to real-time degradation of indicators, and the scarcity of small fault samples leads to data imbalance, resulting in low fault sensitivity and insufficient generalization of assessment results.

[0008] To achieve the above objectives, the technical solution adopted by this invention is: a transformer condition assessment method combining data augmentation and dynamic weighting, comprising the following steps: S1: Construct a three-layer evaluation index system including target layer, criterion layer and indicator layer to obtain evaluation parameters. Determine normalization processing rules based on the positive and negative attributes of each indicator, collect historical monitoring data of transformer operation, and obtain the original historical monitoring dataset. S2: Define the four health status levels of the transformer and the corresponding health index ranges, obtain the status classification judgment criteria, and label and classify the historical monitoring data one by one in combination with the attention value threshold of each indicator to form a classified dataset with status labels. S3: Input the state-labeled classification dataset into the table denoising diffusion probability model, fit the mapping relationship between the original data and the Gaussian noise distribution based on the forward diffusion process, and generate minority class supplementary samples by training the neural network to learn the inverse denoising Markov chain to obtain the training sample set. S4: Solve the linear programming model based on the expert index preference scores to obtain the subjective weights that reflect the expert experience. Calculate the objective weights based on the comparative strength and conflict characteristics of the indicators extracted from historical data. After fusing and normalizing the subjective and objective weights, obtain the static comprehensive weights. S5: Input the class-balanced training sample set into the attention network, generate query vector and key value vector through linear projection, calculate the correlation strength between indicators based on scaling dot product operation, and obtain the state change weighting factor. S6: Nonlinearly fuse the state-variable weighting factor with the static comprehensive weight, introduce a sensitivity adjustment factor to control the response amplitude, and normalize the fused weight vector to obtain the final dynamic comprehensive weight. S7: After standardizing the measured values ​​of the evaluation indicators to eliminate dimensional differences, the dynamic comprehensive weight is input into the cloud model to calculate the comprehensive health index of the transformer. The corresponding evaluation conclusion is output based on the defined four-level health status interval.

[0009] Preferably, step S1 includes the following steps: S11: Based on three dimensions—dissolved gas in oil, oil chemical characteristics, and electrical tests—evaluation parameters are selected to construct a three-layer evaluation index system comprising a target layer, a criterion layer, and an index layer, thereby obtaining a multi-dimensional set of evaluation parameters covering the transformer's operating status. S12: Based on the positive and negative attributes of each indicator, the collected raw monitoring data are normalized and preprocessed by using positive range transformation and inverse range transformation respectively, and all indicator data are uniformly mapped to the [0,1] interval to obtain standardized indicator data with unified dimensions and horizontal comparability. S13: Based on the completed evaluation index system, the system collects historical monitoring data of 10 evaluation parameters during transformer operation, records the initial value, attention value and measured value of each parameter at each moment, and forms the original historical monitoring dataset.

[0010] Preferably, step S2 includes the following steps: S21: Divided into four levels: normal, attention, abnormal and severe, corresponding to four health index intervals: [0.8, 1.0], [0.5, 0.8), [0.2, 0.5) and [0, 0.2) to obtain a unified standard for judging the classification of health status; S22: Combining the attention value thresholds of each indicator, the historical monitoring data collected in step S1 are labeled one by one, and each sample is classified into the corresponding level according to its overall health status, forming a classification dataset with status labels. S23: Count the number of samples at each level in the categorical dataset with state labels, assess the degree of class imbalance in the dataset, and calculate the number of minority class supplementary samples required for each minority class in the subsequent TabDDPM augmentation.

[0011] Preferably, step S3 includes the following steps: S31: Given initial sample data, noise is gradually added to the original data through a forward diffusion process. Based on Gaussian distribution and reparameterization techniques, the noise at any time step is derived. The sample expression is used to obtain the forward mapping law from the original data to the standard Gaussian noise distribution; S32: Construct a noise prediction neural network to learn an inverse Markov chain. Starting from the standard Gaussian noise distribution, gradually remove noise to restore the original data distribution. Complete the model training by minimizing the simplified variational lower bound loss function to obtain an inverse generative model that can generate data that conforms to the characteristics of the original data distribution. S33: Based on the minority class supplementary sample size determined in step S23, call the trained TabDDPM model to generate a corresponding number of minority class samples, merge the generated samples with the original dataset to obtain a class-balanced training sample set, and provide balanced data support for the subsequent training of the self-attention network.

[0012] Preferably, step S4 includes the following steps: S41: Invite industry experts to select the best and worst indicators from the evaluation indicators, construct the best to other preference vectors and other to worst preference vectors respectively, obtain the BWM subjective weights by solving the linear programming model, and obtain the optimal objective function value that can measure the consistency of the scoring. S42: Based on normalized historical monitoring data, the comparative strength of each indicator is measured by standard deviation. The conflict index of each indicator is calculated by Pearson correlation coefficient. The comprehensive information quantity of each indicator is quantified to obtain the information independence of each indicator. After normalization of the comprehensive information quantity, the CRITIC objective weight is obtained. S43: Multiply the BWM subjective weights and CRITIC objective weights element by element, and normalize the resulting weight vector to obtain a static comprehensive weight that takes into account both subjective and objective information.

[0013] Preferably, step S5 includes the following steps: S51: Input the standardized monitoring vector augmented by TabDDPM into the self-attention network, perform linear projection through the learnable feature extraction matrix, and generate the query vector of the corresponding dimension. With key value vector This provides input for subsequent correlation calculations; S52: Calculate the correlation strength between indicators using scaling dot product operation, normalize the calculation results using the Softmax function, and generate state-variable weighting factors that reflect the real-time deterioration significance of each indicator, thereby achieving adaptive amplification of the weights of key deterioration indicators.

[0014] Preferably, step S6 includes the following steps: S61: The state-variable weighting factor and the static comprehensive weight are nonlinearly fused using the exponential weighting method. A sensitivity adjustment factor is introduced to control the magnitude of dynamic adjustment. The preliminary fused dynamic weight vector is obtained through the exponential weighting calculation formula. S62: Normalize the fused weight vector to ensure that the sum of all weight elements is 1, and obtain the final dynamic comprehensive weight to achieve adaptive amplification of the weight of key deterioration indicators that are seriously deviating from the threshold.

[0015] Preferably, S71: The measured values ​​of each evaluation index are standardized according to the normalization rules determined in step S1 to eliminate the dimensional differences between different indexes and obtain a standardized index vector that can be used for weighted calculation. S72: Input the standardized index vector and the dynamic comprehensive weight output in step S6 into the cloud model, and obtain the comprehensive health index of the transformer through weighted fusion calculation; S73: Based on the four-level health status intervals defined in step S2, map the calculated comprehensive health index to the corresponding health level and output the final health status assessment conclusion.

[0016] A transformer condition assessment system that combines data augmentation and dynamic weighting includes: The indicator construction module is used to construct a three-layer evaluation indicator system that includes a target layer, a criterion layer, and an indicator layer, and to determine the normalization processing rules based on the positive and negative attributes of each indicator. The data acquisition module is used to collect historical monitoring data of transformer operation and output the raw historical monitoring dataset. The status classification and labeling module is used to define the four levels of transformer health status and corresponding health index ranges, obtain the status classification judgment criteria, and label and classify historical monitoring data one by one in combination with the attention value threshold of each indicator, and output a classified dataset with status labels. The table denoising diffusion probability augmentation module is used to input a state-labeled classification dataset into the table denoising diffusion probability model and output a class-balanced training sample set. The static weight calculation module is used to solve the linear programming model based on expert index preference scores to obtain subjective weights, extract the index comparison intensity and conflict characteristics based on historical data to calculate objective weights, and output static comprehensive weights after merging and normalizing the subjective and objective weights. The self-attention weighting module is used to input a class-balanced training sample set into the self-attention network, generate query vectors and key-value vectors through linear projection, and output a state weighting factor that is adaptively adjusted according to the degree of index degradation. The dynamic weight fusion module is used to nonlinearly fuse the state-variable weight factor with the static comprehensive weight, introduce a sensitivity adjustment factor to control the response amplitude, normalize the fused weight vector, and output the final dynamic comprehensive weight of the response index in real time deterioration state. The cloud model evaluation module is used to standardize the measured values ​​of the evaluation indicators to eliminate dimensional differences, and then combine them with dynamic comprehensive weights to input the cloud model to calculate the comprehensive health index of the transformer. Based on the defined four-level health status intervals, the module outputs the corresponding evaluation conclusions.

[0017] Preferably, the static weight calculation module includes: The BWM subjective weight calculation unit is used to construct a preference vector based on the selected optimal and worst indicators, and solve the linear programming model to obtain the BWM subjective weights. The CRITIC objective weight calculation unit is used to measure the strength of index comparison using standard deviation, calculate the index conflict index using Pearson correlation coefficient, and quantify the CRITIC objective weight.

[0018] This invention provides a transformer condition assessment method and system that combines data augmentation and dynamic weighting, with the following advantages: By selecting the optimal and worst indicators from the evaluation metrics by industry experts, constructing a preference vector, and solving a linear programming model to obtain subjective weights, the weight results have a quantifiable consistency measure that aligns with engineering experience corresponding to transformer fault mechanisms. The CRITIC method quantifies the comparative strength of indicators based on the standard deviation of normalized monitoring data and calculates the conflict of indicators through the Pearson correlation coefficient, fully exploring the objective statistical patterns of historical monitoring data. After element-wise multiplication and normalization fusion, the two methods avoid the experience bias of single subjective weighting and reduce the limitations of single objective weighting that is divorced from the actual operation and maintenance, effectively improving the rationality and objectivity of the evaluation weights.

[0019] By introducing a table-based denoising diffusion probability model, small sample fault data of abnormal and severity levels are targeted for expansion. Through a two-stage process of progressively adding noise via forward diffusion and progressively denoising via reverse generation, high-quality fault samples that conform to the distribution characteristics of real monitoring data are generated. After merging the original data, a class-balanced training dataset is constructed. This reduces the class imbalance problem in the dataset caused by the scarcity of transformer fault samples, avoids underfitting of the evaluation model due to sample distribution bias, improves the shortcomings of traditional methods such as low recognition rate of abnormal states and easy misjudgment of sudden faults, and significantly enhances the generalization ability and fault identification robustness of the subsequent self-attention evaluation model.

[0020] By extracting the nonlinear coupling features among multiple evaluation indicators through a self-attention network, and quantifying the correlation strength between indicators through scaling dot product operations, a state-variable weighting factor reflecting the real-time deterioration significance of the indicators is generated through Softmax normalization. The state-variable weighting factor is nonlinearly fused with the static comprehensive weight by combining the exponential weighting method. A sensitivity adjustment factor is introduced to control the response amplitude of the dynamic adjustment of the weights, reducing the limitations of static weighting in failing to respond to dynamic deterioration of indicators and being slow to respond to early fault signs. The calculated comprehensive health index is more in line with the actual operating condition of the equipment. At the same time, it can simultaneously identify key deterioration indicators that deviate from the threshold, providing accurate and reliable quantitative basis for the formulation of transformer condition-based maintenance strategies and differentiated operation and maintenance decisions. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart for transformer health status assessment provided in this application embodiment. Figure 2 A diagram illustrating the transformer condition assessment index system provided in this application embodiment. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] This application provides a transformer condition assessment method and system that combines data augmentation and dynamic weighting. It can solve the problems in related technologies where static weights cannot adapt to real-time degradation of indicators and the scarcity of small fault samples leads to imbalance in the dataset, resulting in low fault sensitivity and insufficient generalization of the assessment results.

[0025] like Figure 1-2 As shown, a transformer condition assessment method combining data augmentation and dynamic weighting includes the following steps: S1: Construct a three-layer evaluation index system including target layer, criterion layer and indicator layer to obtain evaluation parameters. Determine normalization processing rules based on the positive and negative attributes of each indicator, collect historical monitoring data of transformer operation, and obtain the original historical monitoring dataset. S2: Define the four health status levels of the transformer and the corresponding health index ranges, obtain the status classification judgment criteria, and label and classify the historical monitoring data one by one in combination with the attention value threshold of each indicator to form a classified dataset with status labels. S3: Input the state-labeled classification dataset into the table denoising diffusion probability model, fit the mapping relationship between the original data and the Gaussian noise distribution based on the forward diffusion process, and generate minority class supplementary samples by training the neural network to learn the inverse denoising Markov chain to obtain the training sample set. S4: Solve the linear programming model based on the expert index preference scores to obtain the subjective weights that reflect the expert experience. Calculate the objective weights based on the comparative strength and conflict characteristics of the indicators extracted from historical data. After fusing and normalizing the subjective and objective weights, obtain the static comprehensive weights. S5: Input the class-balanced training sample set into the attention network, generate query vector and key value vector through linear projection, calculate the correlation strength between indicators based on scaling dot product operation, and obtain the state change weighting factor. S6: Nonlinearly fuse the state-variable weighting factor with the static comprehensive weight, introduce a sensitivity adjustment factor to control the response amplitude, and normalize the fused weight vector to obtain the final dynamic comprehensive weight. S7: After standardizing the measured values ​​of the evaluation indicators to eliminate dimensional differences, the dynamic comprehensive weight is input into the cloud model to calculate the comprehensive health index of the transformer. The corresponding evaluation conclusion is output based on the defined four-level health status interval.

[0026] In one embodiment, step S1 includes the following steps: S11: Based on three dimensions—dissolved gas in oil, oil chemical characteristics, and electrical tests—evaluation parameters are selected to construct a three-layer evaluation index system comprising a target layer, a criterion layer, and an index layer, thereby obtaining a multi-dimensional set of evaluation parameters covering the transformer's operating status. S12: Based on the positive and negative attributes of each indicator, the collected raw monitoring data are normalized and preprocessed by using positive range transformation and inverse range transformation respectively, and all indicator data are uniformly mapped to the [0,1] interval to obtain standardized indicator data with unified dimensions and horizontal comparability. S13: Based on the completed evaluation index system, the system collects historical monitoring data of 10 evaluation parameters during transformer operation, records the initial value, attention value and measured value of each parameter at each moment, and forms the original historical monitoring dataset.

[0027] Specifically, based on the principles of scientific rigor, completeness, and minimal redundancy, and taking DL / T1685-2017 "Guidelines for Condition Evaluation of Oil-Immersed Transformers (Reactors)" as the core basis, key characteristic parameters are selected from three dimensions: dissolved gas in oil, oil chemical characteristics, and electrical tests. A three-layer evaluation index system is constructed, comprising a target layer, a criterion layer, and an indicator layer. The target layer represents the transformer's operating state; the criterion layer includes three dimensions: dissolved gas analysis (U1), insulating oil characteristic indicators (U2), and electrical characteristic indicators (U3); and the indicator layer includes H2 content (U...). 11 CH4 content (U) 12 ), C2H6 content (U 13 ), C2H4 content (U 14 ), C2H2 content (U 15 ), Insulating oil breakdown voltage (U 21 ), Insulating oil moisture content (U 22 furfural content (U) 23 ), core grounding current (U) 31 ) and winding dielectric loss (U 32 A total of 10 evaluation parameters; Based on the positive and negative attributes of each indicator, the collected raw monitoring data are preprocessed using normalization: positive indicators (higher is better, such as the breakdown voltage of insulating oil) are processed using positive range transformation, while negative indicators (lower is more dangerous, such as the content of various gases) are processed using inverse mapping range transformation. All indicator data are uniformly mapped to the [0,1] interval to achieve dimensional uniformity. The normalization formulas for positive and negative indicators are as follows: ; Historical monitoring data of the above 10 evaluation parameters were collected during the operation of the transformer. The initial value, attention value and measured value of each parameter at each time were recorded to form the original historical monitoring dataset, which serves as the data basis for subsequent data augmentation and weight modeling.

[0028] In one embodiment, step S2 includes the following steps: S21: Divided into four levels: normal, attention, abnormal and severe, corresponding to four health index intervals: [0.8, 1.0], [0.5, 0.8), [0.2, 0.5) and [0, 0.2) to obtain a unified standard for judging the classification of health status; S22: Combining the attention value thresholds of each indicator, the historical monitoring data collected in step S1 are labeled one by one, and each sample is classified into the corresponding level according to its overall health status, forming a classification dataset with status labels. S23: Count the number of samples at each level in the categorical dataset with state labels, assess the degree of class imbalance in the dataset, and calculate the number of minority class supplementary samples required for each minority class in the subsequent TabDDPM augmentation.

[0029] Specifically, according to DL / T1685-2017, the health status of transformers is divided into four levels: normal, attention, abnormal, and severe, with corresponding health index ranges of [0.8, 1.0], [0.5, 0.8), [0.2, 0.5), and [0.0, 0.2], respectively. Combining the attention threshold of each indicator, the historical monitoring data collected in step S1 are labeled one by one. Based on the comprehensive health status of each sample, they are assigned to the corresponding level to form a classification dataset with status labels. Finally, the number of samples in each level is counted, the degree of class imbalance is assessed, and the supplementary sample size required for each minority class in the subsequent TabDDPM augmentation is determined, providing a quantitative basis for the targeted augmentation in step S3.

[0030] It should be noted that the steps for annotating each historical monitoring data in step S22 are as follows: Historical monitoring data are organized according to equipment number and monitoring time. Each evaluation parameter of the same equipment at the same time is combined into a status sample, and the data quality is checked.

[0031] The measured values ​​of each indicator are compared with the initial values, warning values, limit values ​​and related status criteria to determine the degree of degradation and individual status of each indicator.

[0032] By combining the results of individual indicator assessments with operation, testing, defect, and maintenance records, the overall health status of each sample is determined; samples that are critical or whose assessments are inconsistent are manually reviewed.

[0033] The samples are categorized into normal, attention, abnormal, or severe levels and assigned corresponding status labels to form a status-labeled classification dataset, which can be used for subsequent inter-class imbalance analysis and TabDDPM data augmentation.

[0034] The implementation steps for inter-class imbalance assessment and sample size determination in step S23 are as follows: Count the number of samples in each health status category (normal, attentive, abnormal, and severe) to determine the maximum and minimum sample sizes for each category.

[0035] Calculate the ratio of the largest class sample size to the smallest class sample size to assess the degree of class imbalance in the dataset.

[0036] The target sample size for each class is determined based on the sample size of the majority class or the size of the preset training set.

[0037] Calculate the difference between the existing sample size and the target sample size for each category, and use this difference as the number of additional samples that TabDDPM needs to generate.

[0038] In one embodiment, step S3 includes the following steps: S31: Given initial sample data, noise is gradually added to the original data through a forward diffusion process. Based on Gaussian distribution and reparameterization techniques, the noise at any time step is derived. The sample expression is used to obtain the forward mapping law from the original data to the standard Gaussian noise distribution; S32: Construct a noise prediction neural network to learn an inverse Markov chain. Starting from the standard Gaussian noise distribution, gradually remove noise to restore the original data distribution. Complete the model training by minimizing the simplified variational lower bound loss function to obtain an inverse generative model that can generate data that conforms to the characteristics of the original data distribution. S33: Based on the minority class supplementary sample size determined in step S23, call the trained Tabular Denoising Diffusion Probabilistic Models (TabDDPM) to generate a corresponding number of minority class samples. Merge the generated samples with the original dataset to obtain a class-balanced training sample set, providing balanced data support for the subsequent training of the self-attention network.

[0039] Specifically, given initial sample data Noise is gradually added to the original data through a forward diffusion process. The conditional probability distribution of the step satisfies: ; in, Indicates the sample number t The diffusion state of the step, t =1,2,…, T , T The total number of diffusion steps, It follows a Gaussian distribution. It is the identity matrix. For noise number The noise addition factor of the step; Using reparameterization techniques, at any time The sample is: ; in, The cumulative diffusion coefficient from step 1 to step t is... , , Standard Gaussian noise; The reverse generation process aims to utilize neural networks Learning inverse Markov chains from noise distribution Starting from this point, we gradually remove noise to the original data distribution.

[0040] The transition probability of the reverse generation process is defined as follows: ; in, The mean of the backward distribution predicted by the neural network The covariance matrix is ​​usually set to a fixed value in practice. This is a preset or learnable variance related to noise scheduling. Denoising Mean Parameterization The calculation is as follows: ; in, For parameters The noise prediction network for The prediction results for the noise contained in it; These are the trainable parameters for the noise prediction network.

[0041] The training objective is to minimize the simplified variational lower bound loss: ; in, To simplify noise prediction loss, This represents the expectation of the original sample x0, random noise ε, random diffusion step t, and class condition c.

[0042] The steps for training the TabDDPM model in step S33 are as follows: The TabDDPM model uses a multilayer perceptron as the noise prediction network to analyze the noisy samples. The diffusion time-step embedding vector and the state category embedding vector are used as input, and the output is prediction noise with the same dimension as the evaluation metric. The multilayer perceptron is configured with 2–6 hidden layers, each with 128–512 neurons. ReLU or SiLU activation functions are used, and a dropout rate of 0–0.3 is set to reduce model overfitting. The total number of diffusion steps... Set to 100-1000 steps, noise addition factor The number of diffusion steps is gradually increased using linear or cosine scheduling. In this embodiment, the total number of diffusion steps is preferably set to 500, and the initial noise figure is set to 10. -4 The termination noise figure is set to 0.02; Randomly select original samples from the training sample set with state labels, and add Gaussian noise at random diffusion time. Input the noisy samples into the noise prediction network, and update the network parameters according to the loss between the real noise and the predicted noise until the loss tends to stabilize or the preset training rounds are reached. After the model training is completed, for the first c A minority class, using the corresponding state category label as a condition, from standard Gaussian noise. Depart, according to t = T , T The reverse denoising process is performed sequentially from 1 to 1 to obtain the generated samples for that category. Based on the supplementary sample quantity determined in step S23, the corresponding number of normal, attentive, abnormal, or severe state samples are generated respectively. The generated samples are denormalized, and each evaluation parameter is verified to ensure it is within a reasonable range, meets the threshold criteria for the corresponding health status, and establishes the basic correlation between the indicators. Unsatisfactory generated samples are discarded and regenerated. The verified generated samples are then merged with the original training samples to obtain a training sample set with a balanced class distribution.

[0043] In one embodiment, step S4 includes the following steps: S41: Invite industry experts to select the best and worst indicators from the evaluation indicators, construct the best to other preference vectors and other to worst preference vectors respectively, obtain the BWM subjective weights by solving the linear programming model, and obtain the optimal objective function value that can measure the consistency of the scoring. S42: Based on normalized historical monitoring data, the comparative strength of each indicator is measured by standard deviation. The conflict index of each indicator is calculated by Pearson correlation coefficient. The comprehensive information quantity of each indicator is quantified to obtain the information independence of each indicator. After normalization of the comprehensive information quantity, the CRITIC objective weight is obtained. S43: Multiply the BWM subjective weights and CRITIC objective weights element by element, and normalize the resulting weight vector to obtain a static comprehensive weight that takes into account both subjective and objective information.

[0044] Specifically, subjective weights are determined based on the Best-Worst Method (BWM): industry experts are invited to select the optimal indicator from 10 evaluation indicators. With worst indicators We construct the optimal to other preference vectors and the other to worst preference vectors respectively, and obtain the Brown-Warren subjective weights by solving the following linear programming model. The calculation is as follows: ; in, For the first j Subjective weighting of each evaluation indicator; The weight of the optimal indicator; The weight of the worst-case indicator; The optimal index given by the experts relative to the first j Importance preference values ​​for each indicator; For the first j The importance preference value of each indicator relative to the worst indicator; n The total number of evaluation indicators; This represents the maximum deviation between the actual weighting ratio and the expert preference value. The smaller the value, the higher the consistency between the obtained weight and expert preference; constraints This means that the sum of the weights of all indicators is 1. This indicates that the weights of each indicator are non-negative.

[0045] in, The smaller the value of the objective function, the higher the consistency, and it can be directly used as a consistency measure. Objective weights are determined based on the CRITIC method (Criteria Importance Through Intercriteria Correlation), and the standard deviation of normalized data is used to measure the comparative strength of each indicator. Conflict indices were calculated using the Pearson correlation coefficient. : ; in, Let be the standard deviation of the j-th evaluation indicator across all samples. The Pearson correlation coefficient between the j-th and k-th indicators; used to quantify the independence of information from each indicator and the overall information content. After normalization, the CRITIC objective weights are obtained. ; After multiplying the subjective weights of the BMW and the objective weights of the CRITIC and normalizing the result, the static composite weights are obtained. The calculation formula is as follows: ; In one embodiment, step S5 includes the following steps: S51: Input the standardized monitoring vector augmented by TabDDPM into the self-attention network, perform linear projection through the learnable feature extraction matrix, and generate the query vector of the corresponding dimension. With key value vector This provides input for subsequent correlation calculations; S52: Calculate the correlation strength between indicators using scaling dot product operation, normalize the calculation results using the Softmax function, and generate state-variable weighting factors that reflect the real-time deterioration significance of each indicator, thereby achieving adaptive amplification of the weights of key deterioration indicators.

[0046] Specifically, the standardized monitoring vectors augmented by TabDDPM will be... Generate query vectors through linear projection. With key value vector The calculation formula is as follows: ; in, and These are the trainable linear projection matrices corresponding to the query vector and the key vector, respectively; The correlation strength between various indicators is calculated using the scaled dot product, and then the state-variable weighting factor is generated by Softmax normalization. The calculation formula is: ; in, The feature dimension of the key vector This is a scaling factor used to prevent the gradient of the Softmax function from becoming too small due to excessively large dot product values. Softmax( ) is a normalization function used to convert each relevance score into attention weights that sum to 1.

[0047] In one embodiment, step S6 includes the following steps: S61: The state-variable weighting factor and the static comprehensive weight are nonlinearly fused using the exponential weighting method. A sensitivity adjustment factor is introduced to control the magnitude of dynamic adjustment. The preliminary fused dynamic weight vector is obtained through the exponential weighting calculation formula. S62: Normalize the fused weight vector to ensure that the sum of all weight elements is 1, and obtain the final dynamic comprehensive weight to achieve adaptive amplification of the weight of key deterioration indicators that are seriously deviating from the threshold.

[0048] The specific steps include: using an exponential weighting method to nonlinearly fuse the state-varying weights with the static comprehensive weights; the dynamic weight calculation formula is as follows: ; in, As a sensitivity adjustment factor, Let i be the state-variable weighting factor corresponding to the i-th index. Let i be the static comprehensive weight of the i-th evaluation index. is the final dynamic comprehensive weight of the i-th evaluation indicator.

[0049] Normalize the fused weight vector to ensure The final dynamic comprehensive weight is obtained, and the weight of key indicators that deviate significantly from the threshold can be amplified by more than 100% compared with the static weight.

[0050] In one embodiment, step S7 includes the following steps: S71: Standardize the measured values ​​of each evaluation index according to the normalization rules determined in step S1 to eliminate the dimensional differences between different indicators and obtain a standardized index vector that can be used for weighted calculation. S72: Input the standardized index vector and the dynamic comprehensive weight output in step S6 into the cloud model, and obtain the comprehensive health index of the transformer through weighted fusion calculation; S73: Based on the four-level health status intervals defined in step S2, map the calculated comprehensive health index to the corresponding health level and output the final health status assessment result.

[0051] In one embodiment, step S7 uses a cloud model to complete the comprehensive health status assessment of the transformer. The cloud model is an uncertainty evaluation model used to convert between qualitative health levels and quantitative monitoring data, employing the expected value... E x ,entropy E n and hyperentropy H eThree numerical features describe each health status level. Expectation represents the central value of the status level, entropy represents the range and ambiguity of the status level, and hyperentropy represents the uncertainty of the entropy. The overall process is as follows: a four-level health status cloud is constructed; the certainty of each indicator for different status levels is calculated; and a weighted fusion is performed using dynamic comprehensive weights to ultimately obtain a comprehensive health index and a health status assessment conclusion.

[0052] The implementation steps for cloud model computing are as follows: The measured values ​​of each indicator are processed according to the normalization rules determined in step S1 to obtain the standardized indicator vector: ; in, x i For the first i The standardized value of each indicator, with a larger value indicating a better state.

[0053] Based on the four levels of health status defined by S21: normal, attentive, abnormal, and severe. L j , U j Construct the corresponding standard state cloud: ; in, ρ This is the preset hyperentropy adjustment coefficient; Calculate the first i The indicator belongs to the first j Certainty of a health status: ; The overall certainty and overall health index of each health status are calculated by combining dynamic comprehensive weights: ; in, For the first i The dynamic comprehensive weight of each indicator For the first j The normalized overall certainty of a health status, HI It is a comprehensive health index.

[0054] In one embodiment, a 220kV transformer (SFPS9-150000 / 220) with a service life of 10 years is used as an example for illustration. The initial values, precautionary values, and measured values ​​of various indicators of the transformer are shown in Table 1 below:

[0055] Table 1 Transformer Data The training dataset mainly comes from historical monitoring data of a substation and related literature, totaling 377 records. After classification in step S2, the sample size for each category is: normal 156 records, attention 126 records, abnormal 70 records, and severe 25 records, showing a significant class imbalance. This data is input into TabDDPM for training, with targeted augmentation performed on the minority classes such as abnormal and severe, ultimately generating a balanced dataset of approximately 2000 records. This provides sufficient and balanced sample support for subsequent training of the self-attention mechanism network. The static comprehensive weights were calculated using the BWM-CRITIC combined weighting method. Industry experts were invited to determine the optimal index (winding insulation dielectric loss, most sensitive to insulation degradation) and the worst-case index (insulating oil moisture content, which varies relatively little under normal operating conditions) based on transformer fault mechanisms. Vectors from the optimal to other indices and from other indices to the worst-case indices were constructed, and a linear programming model was solved to obtain the BWM subjective weights. Simultaneously, the CRITIC objective weights were calculated based on historical monitoring data. The combined static comprehensive weights are as follows:

[0056] The corresponding values ​​are H2, CH4, C2H6, C2H4, C2H, insulating oil breakdown voltage, moisture content, furfural content, core grounding current, and winding dielectric loss, respectively.

[0057] After training the self-attention mechanism network with the balanced dataset input, the measured values ​​in the table above are used as the guideline threshold constraints for calculating each index: the measured H2 content of 140 μL / L has reached 93.3% of the attention value of 150 μL / L, exceeding the limit index. However, it is close to the trigger boundary; the measured value of winding insulation dielectric loss is 0.79%, reaching 98.8% of the attention value of 0.8%, and the over-limit index is close to the critical value; the measured value of core grounding current is 0.06A, reaching 60% of the attention value of 0.1A. The self-attention mechanism network integrates the self-attention distribution and the guideline threshold, and outputs the final dynamic weight after dynamic fusion, as follows:

[0058] Substituting the dynamic weights and measured values ​​into the cloud model, the calculated comprehensive health index is 0.615. Based on the level classification standard in step S2, the transformer is determined to be in a state of alert. This result is highly consistent with the actual situation—a short circuit occurred at the equipment's operating outlet, and the windings deformed after the short-circuit current impact. Key indicators such as H2 content and winding insulation dielectric loss are close to or slightly exceed the alert value, thus the transformer is comprehensively judged to be in a state of alert. The evaluation results compared with those of other comparative methods are shown in Table 2 below.

[0059]

[0060] Table 2 Evaluation Results As can be seen from the table, the static weighting method of the cloud model is too conservative in its evaluation results (judged as abnormal) because it cannot perceive the critical deterioration state of H2 content and winding dielectric loss. In contrast, the health index of 0.615 obtained by the dynamic weighting method of this invention is closest to the original method of 0.618. While correctly determining the state level, the health index is 8.8% lower than that of the constant weighting method, which more sensitively reflects the potential risks of the equipment and the evaluation results are closer to the actual operating conditions of the transformer.

[0061] A transformer condition assessment system that combines data augmentation and dynamic weighting includes: The indicator construction module is used to construct a three-layer evaluation indicator system, which includes a target layer, a criterion layer, and an indicator layer. Based on the positive and negative attributes of each indicator, it determines the normalization processing rules and outputs the determined evaluation parameter system to the data acquisition module. The data acquisition module is used to collect historical monitoring data of transformer operation, output the raw historical monitoring dataset, and output the raw historical monitoring dataset to the status classification and labeling module; The status classification and labeling module is used to define the four levels of transformer health status and corresponding health index ranges, obtain the status classification judgment criteria, and label and classify historical monitoring data one by one in combination with the attention value threshold of each indicator, and output the classified dataset with status labels to the table denoising and diffusion probability augmentation module. The table denoising diffusion probability augmentation module is used to input the state-labeled classification dataset into the table denoising diffusion probability model, fit the noise mapping relationship through forward diffusion, generate targeted augmentation of minority class fault samples in reverse, and output a class-balanced training sample set after merging the original data to the self-attention variable weight module. The static weight calculation module is used to solve the linear programming model based on expert index preference scores to obtain subjective weights, extract the index comparison intensity and conflict characteristics based on historical data to calculate objective weights, and output the static comprehensive weight to the dynamic weight fusion module after fusing and normalizing the subjective and objective weights. The self-attention weighting module is used to input a class-balanced training sample set into the self-attention network, generate query vectors and key value vectors through linear projection, and output the state weighting factor that is adaptively adjusted according to the degree of index deterioration to the dynamic weight fusion module. The dynamic weight fusion module is used to nonlinearly fuse the state-variable weight factor with the static comprehensive weight, introduce a sensitivity adjustment factor to control the response amplitude, normalize the fused weight vector, and output the final dynamic comprehensive weight of the response index in real time to the cloud model evaluation module. The cloud model evaluation module is used to standardize the measured values ​​of the evaluation indicators to eliminate dimensional differences, and then combine them with dynamic comprehensive weights to input the cloud model to calculate the comprehensive health index of the transformer. Based on the defined four-level health status intervals, the module outputs the corresponding evaluation conclusions.

[0062] In one embodiment, the static weight calculation module includes: The BWM subjective weight calculation unit is used to construct a preference vector based on the selected optimal and worst indicators, and solve the linear programming model to obtain the BWM subjective weights. The CRITIC objective weight calculation unit is used to measure the strength of the comparison of indicators using standard deviation, calculate the conflict index of indicators using Pearson correlation coefficient, and quantify the CRITIC objective weight. The static weight calculation module multiplies the BWM subjective weights and CRITIC objective weights element by element and normalizes them before outputting the static comprehensive weights.

[0063] In one embodiment, the functional modules of this system can be integrated into the transformer online monitoring platform's host computer, substation edge computing terminal, or cloud server in the form of software programs, and real-time and historical monitoring data can be obtained by connecting to the transformer oil chromatography online monitoring device and electrical test acquisition terminal.

[0064] In the description of this application, it should be noted that the terms "upper," "lower," etc., indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Unless otherwise expressly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication between two elements. For those skilled in the art, the specific meaning of the above terms in this application can be understood according to the specific circumstances.

[0065] It should be noted that in this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0066] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A transformer condition assessment method that combines data augmentation and dynamic weighting, characterized in that, Includes the following steps: S1: Construct a three-layer evaluation index system including target layer, criterion layer and indicator layer to obtain evaluation parameters. Determine normalization processing rules based on the positive and negative attributes of each indicator, collect historical monitoring data of transformer operation, and obtain the original historical monitoring dataset. S2: Define the four health status levels of the transformer and the corresponding health index ranges, obtain the status classification judgment criteria, and label and classify the historical monitoring data one by one in combination with the attention value threshold of each indicator to form a classified dataset with status labels. S3: Input the state-labeled classification dataset into the table denoising diffusion probability model, fit the mapping relationship between the original data and the Gaussian noise distribution based on the forward diffusion process, and generate minority class supplementary samples by training the neural network to learn the inverse denoising Markov chain to obtain the training sample set. S4: Solve the linear programming model based on the expert index preference scores to obtain the subjective weights that reflect the expert experience. Calculate the objective weights based on the comparative strength and conflict characteristics of the indicators extracted from historical data. After fusing and normalizing the subjective and objective weights, obtain the static comprehensive weights. S5: Input the class-balanced training sample set into the attention network, generate query vector and key value vector through linear projection, calculate the correlation strength between indicators based on scaling dot product operation, and obtain the state change weighting factor. S6: Nonlinearly fuse the state-variable weighting factor with the static comprehensive weight, introduce a sensitivity adjustment factor to control the response amplitude, and normalize the fused weight vector to obtain the final dynamic comprehensive weight. S7: After standardizing the measured values ​​of the evaluation indicators to eliminate dimensional differences, the dynamic comprehensive weight is input into the cloud model to calculate the comprehensive health index of the transformer. The corresponding evaluation conclusion is output based on the defined four-level health status interval.

2. The transformer condition assessment method based on data augmentation and dynamic weighting as described in claim 1, characterized in that: Step S1 includes the following steps: S11: Based on three dimensions—dissolved gas in oil, oil chemical characteristics, and electrical tests—evaluation parameters are selected to construct a three-layer evaluation index system comprising a target layer, a criterion layer, and an index layer, thereby obtaining a multi-dimensional set of evaluation parameters covering the transformer's operating status. S12: Based on the positive and negative attributes of each indicator, the collected raw monitoring data are normalized and preprocessed by using positive range transformation and inverse range transformation respectively, and all indicator data are uniformly mapped to the [0,1] interval to obtain standardized indicator data with unified dimensions and horizontal comparability. S13: Based on the completed evaluation index system, the system collects historical monitoring data of 10 evaluation parameters during transformer operation, records the initial value, attention value and measured value of each parameter at each moment, and forms the original historical monitoring dataset.

3. The transformer condition assessment method based on data augmentation and dynamic weighting as described in claim 1, characterized in that: Step S2 includes the following steps: S21: Divided into four levels: normal, attention, abnormal and severe, corresponding to four health index intervals: [0.8, 1.0], [0.5, 0.8), [0.2, 0.5) and [0, 0.2) to obtain a unified standard for judging the classification of health status; S22: Combining the attention value thresholds of each indicator, the historical monitoring data collected in step S1 are labeled one by one, and each sample is classified into the corresponding level according to its overall health status, forming a classification dataset with status labels. S23: Count the number of samples at each level in the categorical dataset with state labels, assess the degree of class imbalance in the dataset, and calculate the number of minority class supplementary samples required for each minority class in the subsequent TabDDPM augmentation.

4. The transformer condition assessment method based on data augmentation and dynamic weighting as described in claim 1, characterized in that: Step S3 includes the following steps: S31: Given initial sample data, noise is gradually added to the original data through a forward diffusion process. Based on Gaussian distribution and reparameterization techniques, the noise at any time step is derived. The sample expression is used to obtain the forward mapping law from the original data to the standard Gaussian noise distribution; S32: Construct a noise prediction neural network to learn an inverse Markov chain. Starting from the standard Gaussian noise distribution, gradually remove noise to restore the original data distribution. Complete the model training by minimizing the simplified variational lower bound loss function to obtain an inverse generative model that can generate data that conforms to the characteristics of the original data distribution. S33: Based on the minority class supplementary sample size determined in step S23, call the trained TabDDPM model to generate a corresponding number of minority class samples, merge the generated samples with the original dataset to obtain a class-balanced training sample set, and provide balanced data support for the subsequent training of the self-attention network.

5. The transformer condition assessment method based on data augmentation and dynamic weighting as described in claim 1, characterized in that: Step S4 includes the following steps: S41: Invite industry experts to select the best and worst indicators from the evaluation indicators, construct the best to other preference vectors and other to worst preference vectors respectively, obtain the BWM subjective weights by solving the linear programming model, and obtain the optimal objective function value that can measure the consistency of the scoring. S42: Based on normalized historical monitoring data, the comparative strength of each indicator is measured by standard deviation. The conflict index of each indicator is calculated by Pearson correlation coefficient. The comprehensive information quantity of each indicator is quantified to obtain the information independence of each indicator. After normalization of the comprehensive information quantity, the CRITIC objective weight is obtained. S43: Multiply the BWM subjective weights and CRITIC objective weights element by element, and normalize the resulting weight vector to obtain a static comprehensive weight that takes into account both subjective and objective information.

6. The transformer condition assessment method based on data augmentation and dynamic weighting as described in claim 1, characterized in that: Step S5 includes the following steps: S51: Input the standardized monitoring vector augmented by TabDDPM into the self-attention network, perform linear projection through the learnable feature extraction matrix, and generate the query vector of the corresponding dimension. With key value vector This provides input for subsequent correlation calculations; S52: Calculate the correlation strength between indicators using scaling dot product operation, normalize the calculation results using the Softmax function, and generate state-variable weighting factors that reflect the real-time deterioration significance of each indicator, thereby achieving adaptive amplification of the weights of key deterioration indicators.

7. The transformer condition assessment method based on data augmentation and dynamic weighting as described in claim 1, characterized in that: Step S6 includes the following steps: S61: The state-variable weighting factor and the static comprehensive weight are nonlinearly fused using the exponential weighting method. A sensitivity adjustment factor is introduced to control the magnitude of dynamic adjustment. The preliminary fused dynamic weight vector is obtained through the exponential weighting calculation formula. S62: Normalize the fused weight vector to ensure that the sum of all weight elements is 1, and obtain the final dynamic comprehensive weight to achieve adaptive amplification of the weight of key deterioration indicators that are seriously deviating from the threshold.

8. The transformer condition assessment method based on data augmentation and dynamic weighting as described in claim 1, characterized in that: Step S7 includes the following steps: S71: Standardize the measured values ​​of each evaluation index according to the normalization rules determined in step S1 to eliminate the dimensional differences between different indicators and obtain a standardized index vector that can be used for weighted calculation. S72: Input the standardized index vector and the dynamic comprehensive weight output in step S6 into the cloud model, and obtain the comprehensive health index of the transformer through weighted fusion calculation; S73: Based on the four-level health status intervals defined in step S2, map the calculated comprehensive health index to the corresponding health level and output the final health status assessment conclusion.

9. A transformer condition assessment system combining data augmentation and dynamic weighting, used to implement the transformer condition assessment method combining data augmentation and dynamic weighting as described in any one of claims 1-8, characterized in that, include: The indicator construction module is used to construct a three-layer evaluation indicator system that includes a target layer, a criterion layer, and an indicator layer, and to determine the normalization processing rules based on the positive and negative attributes of each indicator. The data acquisition module is used to collect historical monitoring data of transformer operation and output the raw historical monitoring dataset. The status classification and labeling module is used to define the four levels of transformer health status and corresponding health index ranges, obtain the status classification judgment criteria, and label and classify historical monitoring data one by one in combination with the attention value threshold of each indicator, and output a classified dataset with status labels. The table denoising diffusion probability augmentation module is used to input a state-labeled classification dataset into the table denoising diffusion probability model and output a class-balanced training sample set. The static weight calculation module is used to solve the linear programming model based on expert index preference scores to obtain subjective weights, extract the index comparison intensity and conflict characteristics based on historical data to calculate objective weights, and output static comprehensive weights after merging and normalizing the subjective and objective weights. The self-attention weighting module is used to input a class-balanced training sample set into the self-attention network, generate query vectors and key-value vectors through linear projection, and output a state weighting factor that is adaptively adjusted according to the degree of index degradation. The dynamic weight fusion module is used to nonlinearly fuse the state-variable weight factor with the static comprehensive weight, introduce a sensitivity adjustment factor to control the response amplitude, normalize the fused weight vector, and output the final dynamic comprehensive weight of the response index in real time deterioration state. The cloud model evaluation module is used to standardize the measured values ​​of the evaluation indicators to eliminate dimensional differences, and then combine them with dynamic comprehensive weights to input the cloud model to calculate the comprehensive health index of the transformer. Based on the defined four-level health status intervals, the module outputs the corresponding evaluation conclusions.

10. The transformer condition assessment system based on data augmentation and dynamic weighting as described in claim 1, characterized in that: The static weight calculation module includes: The BWM subjective weight calculation unit is used to construct a preference vector based on the selected optimal and worst indicators, and solve the linear programming model to obtain the BWM subjective weights. The CRITIC objective weight calculation unit is used to measure the strength of index comparison using standard deviation, calculate the index conflict index using Pearson correlation coefficient, and quantify the CRITIC objective weight.

Images