High voltage test data evaluation method and device based on cross attention mechanism

Abstract

The embodiment of the specification provides a high-voltage test data evaluation method and device based on a cross-attention mechanism, wherein the method comprises the following steps: acquiring multi-dimensional measured data output by a transformer high-voltage testing instrument to obtain a standardized test data set; constructing a multi-algorithm comparison library; inputting the standardized test data set into each algorithm respectively and performing parallel fusion calculation, wherein a cross-attention mechanism fusion algorithm divides indexes into query vectors, key vectors and value vectors, generates attention weights through similarity calculation and performs weighted summation, and outputs evaluation scores and confidence; quantitatively scoring the outputs of each algorithm according to an evaluation index system, and screening an optimal fusion algorithm; performing final fusion calculation by using the optimal fusion algorithm to obtain a quantitative score of transformer overall performance, determining a performance grade and identifying a potential defect type. The application realizes multi-index deep fusion and adaptive weight distribution through the cross-attention mechanism, combines a multi-algorithm optimization system, and significantly improves evaluation accuracy and reliability.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of power technology, and in particular to a method for evaluating high-voltage test data based on a cross-attention mechanism. Background Technology

[0002] Transformers are core equipment in the power transmission and distribution system, and their operating status directly affects the safe and stable operation of the power grid. High-voltage testing is a core technical means to determine the performance of transformers and to investigate insulation defects and potential faults. During the test, specialized testing instruments directly output multi-dimensional ready-made data such as winding insulation resistance, DC resistance, dielectric loss factor tanδ of windings and bushings, capacitance, DC leakage current, and insulation resistance of core clamps, without the need for additional algorithms to process and generate raw data.

[0003] Current evaluation of transformer high-voltage test data primarily employs traditional methods such as single-index comparison with standard threshold judgment, conventional weighted fusion, grey relational analysis, fuzzy comprehensive evaluation, or DS evidence theory. While subsequent introduction of machine learning algorithms like support vector machines, random forests, and backpropagation neural networks has achieved preliminary intelligent data fusion evaluation, all existing methods suffer from inherent limitations: First, single-index judgment only reflects local performance and cannot account for the inherent correlations and coupling characteristics between multiple indices, leading to misjudgments and omissions. Second, traditional weighting algorithms such as entropy weighting and analytic hierarchy process (AHP) use fixed weight allocation models, failing to dynamically adjust based on the actual characteristics of the data. The adjustments have several drawbacks. First, they neglect the varying degrees of impact of different test indicators on the overall performance of transformers, resulting in low evaluation accuracy. Second, methods such as grey relational analysis and fuzzy evaluation can only achieve shallow data correlation, failing to uncover deep coupling relationships between multi-source heterogeneous indicators and lacking sufficient ability to identify latent faults such as insulation dampness and aging. Third, conventional machine learning and shallow neural network algorithms can only complete single-dimensional feature extraction, lacking the ability to model the interaction and correlation between multiple indicators, resulting in limited generalization and evaluation accuracy. Fourth, existing technologies generally lack a standardized process for parallel comparison and selection of multiple algorithms, making them unsuitable for data scenarios of different voltage levels and test types.

[0004] Therefore, a better solution is urgently needed. Summary of the Invention

[0005] In view of this, embodiments of this specification provide a method for evaluating high-pressure test data based on a cross-attention mechanism. One or more embodiments of this specification also relate to a device for evaluating high-pressure test data based on a cross-attention mechanism, a computing device, a computer-readable storage medium, and a computer program, to address the technical deficiencies existing in the prior art.

[0006] According to a first aspect of the embodiments of this specification, a method for evaluating high-pressure experimental data based on a cross-attention mechanism is provided, comprising: Acquire multi-dimensional measured data output from transformer high-voltage testing instruments, and perform temperature correction, outlier removal, missing value filling, and dimension normalization on the multi-dimensional measured data to obtain a standardized test dataset. Construct a multi-algorithm comparison library that includes entropy weighted fusion method, grey relational analysis method, fuzzy comprehensive evaluation method, random forest, support vector machine and cross-attention mechanism fusion algorithm; The standardized experimental dataset is input into each algorithm in the multi-algorithm comparison library, and the data fusion calculation is performed in parallel. The cross-attention mechanism fusion algorithm divides the indicators in the standardized experimental dataset into query vectors, key vectors and value vectors. It generates attention weights by calculating the similarity between query vectors and key vectors, and uses the attention weights to perform weighted summation on the value vectors to output the evaluation score and confidence level. The fusion evaluation results of each algorithm are quantitatively scored according to the preset evaluation index system, and the algorithm with the highest score is selected as the optimal fusion algorithm. The optimal fusion algorithm is used to perform fusion calculations on the standardized test dataset to obtain a quantitative score of the overall performance of the transformer. The performance level of the transformer is determined based on the quantitative score, and the potential defect types are identified by combining the attention weight distribution of each index in the cross-attention mechanism fusion algorithm.

[0007] In one possible implementation, the cross-attention mechanism fusion algorithm employs a multi-head attention mechanism, which maps the input to different subspaces through multiple linear projection matrices to compute attention in parallel, and then splices and fuses the attention computation results of each subspace.

[0008] In one possible implementation, outlier removal, missing value imputation, and dimensional normalization are performed on the multidimensional measured data. Specifically, this includes: removing outliers using the three-standard-deviation criterion, imputing missing values ​​using nearest neighbor interpolation, and performing Z-score standardization based on the mean and standard deviation of the data.

[0009] In one possible implementation, the pre-set evaluation index system includes evaluation accuracy, data fit, result stability, and three-phase consistency deviation. When quantifying the fusion evaluation results output by each algorithm, the results are weighted and summed according to the weight coefficients corresponding to each evaluation index.

[0010] In one possible implementation, the cross-attention mechanism fusion algorithm uses winding insulation indices from the standardized test dataset as query vectors and dielectric loss factor and leakage current indices as key and value vectors, respectively, before performing data fusion computation.

[0011] In one possible implementation, the weighting factor for accuracy is 0.55, the weighting factor for data fit is 0.2, the weighting factor for result stability is 0.15, and the weighting factor for three-phase consistency deviation is 0.1.

[0012] In one possible implementation, when determining the performance level of a transformer based on a quantitative score, 90 to 100 points correspond to an excellent level, 80 to 89 points correspond to a good level, 60 to 79 points correspond to a warning level, and below 60 points correspond to an abnormal or severe level. Potential defect types include insulation dampness, insulation aging, local defects, and overheating hazards.

[0013] According to a second aspect of the embodiments of this specification, a high-pressure test data evaluation device based on a cross-attention mechanism is provided, comprising: The data acquisition module is configured to acquire multi-dimensional measured data output by the transformer high-voltage test instrument, and perform temperature correction, outlier removal, missing value filling and dimension normalization on the multi-dimensional measured data to obtain a standardized test dataset. The fusion algorithm module is configured to build a multi-algorithm comparison library that includes entropy weighted fusion method, grey relational analysis method, fuzzy comprehensive evaluation method, random forest, support vector machine and cross attention mechanism fusion algorithm; The fusion computing module is configured to input the standardized experimental dataset into each algorithm in the multi-algorithm comparison library and perform data fusion computing in parallel. The cross-attention mechanism fusion algorithm divides the indicators in the standardized experimental dataset into query vectors, key vectors and value vectors. It generates attention weights by calculating the similarity between the query vector and the key vector, and uses the attention weights to perform weighted summation on the value vectors to output the evaluation score and confidence level. The algorithm determination module is configured to quantitatively score the fusion evaluation results output by each algorithm according to a preset evaluation index system, and select the algorithm with the highest score as the optimal fusion algorithm. The performance determination module is configured to use the optimal fusion algorithm to perform fusion calculations on the standardized test dataset to obtain a quantitative score of the overall performance of the transformer. Based on the quantitative score, the performance level of the transformer is determined, and the potential defect types are identified by combining the attention weight distribution of each index in the cross-attention mechanism fusion algorithm.

[0014] According to a third aspect of the embodiments of this specification, a computing device is provided, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the high-pressure test data evaluation method based on the cross-attention mechanism described above.

[0015] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the above-described high-pressure test data evaluation method based on a cross-attention mechanism.

[0016] According to a fifth aspect of the embodiments of this specification, a computer program is provided, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the high-pressure test data evaluation method based on the cross-attention mechanism described above.

[0017] This specification provides a method and apparatus for evaluating high-voltage test data based on a cross-attention mechanism. The method includes: acquiring multi-dimensional measured data output from a transformer high-voltage testing instrument; performing temperature correction, outlier removal, missing value completion, and dimension normalization to obtain a standardized test dataset; constructing a multi-algorithm comparison library including entropy-weighted fusion, grey relational analysis, fuzzy comprehensive evaluation, random forest, support vector machine, and cross-attention mechanism fusion algorithms; inputting the standardized test dataset into each algorithm for parallel fusion calculation, wherein the cross-attention mechanism fusion algorithm divides indicators into query vectors, key vectors, and value vectors, generates attention weights through similarity calculation, performs weighted summation, and outputs evaluation scores and confidence levels; quantifying the outputs of each algorithm according to the evaluation index system, and selecting the optimal fusion algorithm; using the optimal fusion algorithm for final fusion calculation to obtain a quantitative score for the overall transformer performance, determine the performance level, and identify potential defect types. This application achieves deep fusion of multiple indicators and adaptive weight allocation through a cross-attention mechanism, combined with a multi-algorithm optimization system, significantly improving the evaluation accuracy and reliability. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a high-pressure test data evaluation method based on a cross-attention mechanism, provided in one embodiment of this specification. Figure 2 This is a schematic diagram of a multi-head cross-attention fusion architecture for a high-pressure test data evaluation method based on a cross-attention mechanism, provided in one embodiment of this specification. Figure 3 This is an overall flowchart of a high-pressure test data evaluation method based on a cross-attention mechanism, provided in one embodiment of this specification; Figure 4 This is a schematic diagram of a high-pressure test data evaluation device based on a cross-attention mechanism, provided in one embodiment of this specification. Figure 5 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0019] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0020] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0021] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0022] This specification provides a method for evaluating high-pressure test data based on a cross-attention mechanism. This specification also relates to a device for evaluating high-pressure test data based on a cross-attention mechanism, a computing device, and a computer-readable storage medium, which will be described in detail in the following embodiments.

[0023] See Figure 1 , Figure 1 A flowchart is shown of a high-pressure test data evaluation method based on a cross-attention mechanism according to an embodiment of this specification, which specifically includes the following steps.

[0024] Step 101: Obtain multi-dimensional measured data output by the transformer high-voltage test instrument, and perform temperature correction, outlier removal, missing value filling and dimension normalization on the multi-dimensional measured data to obtain a standardized test dataset. Step 102: Construct a multi-algorithm comparison library containing entropy weighted fusion method, grey relational analysis method, fuzzy comprehensive evaluation method, random forest, support vector machine and cross-attention mechanism fusion algorithm; Step 103: Input the standardized experimental dataset into each algorithm in the multi-algorithm comparison library and perform data fusion calculation in parallel. The cross-attention mechanism fusion algorithm divides the indicators in the standardized experimental dataset into query vectors, key vectors and value vectors. It generates attention weights by calculating the similarity between query vectors and key vectors, and uses the attention weights to perform weighted summation on the value vectors to output the evaluation score and confidence level. Step 104: Quantitatively score the fusion evaluation results output by each algorithm according to the preset evaluation index system, and select the algorithm with the highest score as the optimal fusion algorithm; Step 105: Use the optimal fusion algorithm to perform fusion calculation on the standardized test dataset to obtain a quantitative score of the overall performance of the transformer. Determine the performance level of the transformer based on the quantitative score and identify potential defect types.

[0025] Multi-dimensional measured data refers to the raw measurement values ​​directly output by various specialized testing instruments during transformer high-voltage testing, without secondary processing. Examples include winding insulation resistance, dielectric loss factor, and DC leakage current. These data serve as the basic input for evaluation. Temperature correction refers to the process of uniformly converting high-voltage test data to standard temperatures according to power testing regulations, used to eliminate the influence of ambient temperature on insulation performance parameters. Outlier removal refers to using statistical methods to identify and remove invalid data points that deviate from the normal range due to measurement errors, equipment disturbances, etc., such as using the three-standard-deviation criterion for screening. Missing value imputation refers to the reasonable filling of gaps caused by missing data acquisition or outlier removal, such as using nearest-neighbor interpolation based on adjacent samples. Dimensional normalization refers to the standardization process that eliminates the inconsistency in data magnitude caused by unit differences between different physical quantities, such as using the Z-score method to convert each indicator into a dimensionless form with a mean of zero and a standard deviation of the same standard value. Standardized test datasets refer to a set of standardized data that has consistent dimensions and can be directly used for multi-algorithm fusion computation after the aforementioned preprocessing. Multi-algorithm comparison libraries refer to a collection of different types of fusion algorithms integrated within the same framework, used for parallel computation and horizontal comparison of multiple algorithms. Cross-attention mechanism fusion algorithms refer to a feature interaction method based on the attention mechanism, which constructs three types of vectors: query, key, and value. It calculates attention weights using the similarity between queries and keys to achieve weighted fusion of value vectors, used to uncover deep correlations between multiple indicators. Evaluation scores and confidence levels respectively indicate the quantitative evaluation results output by the fusion algorithm and their reliability, with confidence level representing the certainty of the evaluation result. Pre-defined evaluation index systems refer to a set of multi-dimensional evaluation standards used for horizontal comparison of the performance of different algorithms, used to select the optimal fusion algorithm. Quantitative scoring refers to the result of representing the overall performance of the transformer in numerical form, used to intuitively reflect the equipment status. Potential defect types refer to the categories of potential equipment hazards identified by analyzing the fusion computation results and the distribution of correlation weights between indicators.

[0026] The present application will be further described below through a detailed embodiment: In this embodiment, the aforementioned high-voltage test data evaluation method based on the cross-attention mechanism is applied to the high-voltage handover test scenario of a 110kV oil-immersed transformer.

[0027] First, multi-dimensional measured data from the high-voltage testing instrument of the transformer were acquired, including the insulation resistance to ground of the high-voltage winding, the insulation resistance to ground of the medium-voltage winding, the insulation resistance to ground of the low-voltage winding, the absorption ratio, the polarization index, the dielectric loss factor of the winding and bushing, the winding capacitance, the DC leakage current, the core insulation resistance, the clamp insulation resistance, and the bushing dielectric loss factor. Three sets of parallel test data were collected. Temperature correction was applied to the multi-dimensional measured data, and all data were uniformly converted to equivalent values ​​at a standard temperature of 20 degrees Celsius. Then, the three-standard-deviation criterion was used to identify outliers, and one set of abnormal DC leakage current data caused by instantaneous interference during the test was removed. The missing bushing dielectric loss factor data was supplemented using the nearest neighbor interpolation method. Finally, based on the Z-score method, the mean and standard deviation of each index were used to normalize the dimensions of all data to obtain a standardized test dataset.

[0028] A multi-algorithm comparison library is constructed, which includes three weighted fusion and shallow association analysis methods: entropy weighted fusion method, grey relational analysis method, and fuzzy comprehensive evaluation method; two conventional machine learning classification methods: random forest and support vector machine; and a cross-attention mechanism fusion algorithm as the core innovative method.

[0029] The standardized test dataset is input into various algorithms in a multi-algorithm comparison library, and data fusion calculations are performed in parallel. Taking the cross-attention mechanism fusion algorithm as an example, this algorithm divides the indicators belonging to the winding insulation category in the standardized test dataset into query vectors, and divides the indicators of the dielectric loss factor and leakage current categories into key vectors and value vectors. By calculating the dot product similarity between the query vector and the key vector and performing scaling and normalization processing, attention weights are generated, which reflect the correlation strength between different indicator categories. Subsequently, the value vectors are weighted and summed using these weights to output the evaluation score and corresponding confidence level of the transformer.

[0030] The fusion evaluation results of each algorithm are quantitatively scored according to a pre-defined evaluation index system. This evaluation index system includes four dimensions: evaluation accuracy, data fit, result stability, and three-phase consistency deviation, each assigned a different weight coefficient and then summed. Calculations show that the comprehensive score of the cross-attention mechanism fusion algorithm is higher than that of the entropy-weighted fusion method, grey relational analysis method, fuzzy comprehensive evaluation method, random forest, and support vector machine. Therefore, the cross-attention mechanism fusion algorithm is selected as the optimal fusion algorithm.

[0031] The optimal fusion algorithm is used to perform final fusion calculations on the standardized test dataset to obtain a quantitative score for the overall performance of the transformer. Based on the quantitative score and a pre-defined grading standard, the performance level of the transformer is determined. Simultaneously, by combining the attention weight distribution of each indicator in the cross-attention mechanism, potential defect types are identified, and finally, a complete high-voltage test performance evaluation result for the transformer is output.

[0032] The beneficial effects of this embodiment include at least the following: by constructing a comparison library containing multiple types of algorithms and performing fusion calculations in parallel, horizontal comparison and selection of the best algorithm are realized, avoiding the problem of limited applicability of a single algorithm; the cross-attention mechanism fusion algorithm generates attention weights by calculating the similarity between query vectors and key vectors, realizing adaptive dynamic weighting among multiple indicators, which can better reflect the real correlation between indicators compared with fixed weight methods, thus improving the evaluation accuracy; and by using an evaluation index system to quantitatively score and select the optimal fusion algorithm, the reliability and adaptability of the evaluation results under different experimental scenarios are ensured.

[0033] See Figure 2 , Figure 2 The diagram illustrates a multi-head cross-attention fusion architecture provided by some embodiments of this specification. Based on the aforementioned steps, the cross-attention mechanism fusion algorithm adopts a multi-head attention mechanism, which maps the input to different subspaces through multiple linear projection matrices to compute attention in parallel, and then splices and fuses the attention computation results of each subspace.

[0034] Multi-head attention mechanisms can be seen as an extension of cross-attention mechanisms. This involves setting up multiple parallel attention heads, each using an independent linear projection matrix to map the original input to different feature subspaces. Attention weights are independently calculated and weighted summed within each subspace. Finally, the outputs of all heads are concatenated and fused using a single output projection matrix. This enhances the model's ability to capture interactions between features from different representation dimensions. A linear projection matrix is ​​a parameterized matrix used to linearly transform the input vector. Multiplying by this matrix maps the original vector to a feature space of a specific dimension. Each attention head has its own independent projection matrix, enabling multi-view feature extraction. A subspace can be a low-dimensional projection space within the original high-dimensional feature space. Different subspaces focus on different semantic levels or feature patterns of the input data. Parallel computation helps to comprehensively uncover complex coupling relationships between metrics. Concatenation and fusion refer to connecting the outputs of multiple attention heads along the feature dimension and performing dimensional transformation and information integration through a learnable linear transformation to synthesize the interactive features extracted from each subspace.

[0035] The present application will be further described below through a detailed embodiment: In the aforementioned steps, the cross-attention mechanism fusion algorithm specifically employs a multi-head attention mechanism for deep fusion when performing data fusion calculations.

[0036] This multi-head attention mechanism employs four parallel attention heads. For each attention head, three independent linear projection matrices are first used to linearly transform the query vector, key vector, and value vector in the input data, mapping them to different feature subspaces. Each subspace has an independent projection dimension, used to capture the interaction relationships between metrics from different perspectives.

[0037] Within each subspace, cross-attention computation is performed independently: the similarity between the transformed query vector and the key vector is calculated, and the attention weights for that subspace are generated after scaling and normalization. The transformed value vectors are then weighted and summed using these weights to output the attention results for that subspace.

[0038] After parallel computation of the four attention heads, the attention results output from each subspace are concatenated along the feature dimension to form a comprehensive vector that integrates features from multiple perspectives. Finally, this concatenated comprehensive vector is linearly transformed through an output projection matrix to obtain the final deep fusion feature, which is used for subsequent evaluation score output.

[0039] By employing a multi-head attention mechanism, this embodiment can uncover the deep coupling relationships between different test indicators from multiple characterization dimensions. For example, one head may focus on capturing the correlation between insulation resistance and dielectric loss factor, while another head may focus on capturing the coupling between DC leakage current and absorption ratio. Combining the outputs of multiple heads makes the fusion results more comprehensive and robust.

[0040] The beneficial effects of this embodiment include at least the following: by adopting a multi-head attention mechanism, the input is mapped to different subspaces using multiple linear projection matrices to compute attention in parallel, which can mine the coupling relationship between experimental indicators from multiple feature dimensions and has a stronger feature interaction mining capability compared with the single-head attention mechanism; each attention head independently learns different association patterns, and the final splicing and fusion result integrates the interactive features from multiple perspectives, improving the fusion accuracy and generalization ability.

[0041] Based on the aforementioned steps, outlier removal, missing value filling, and dimensional normalization are performed on the multi-dimensional measured data. Specifically, outliers are removed using the three-standard-deviation criterion, missing values ​​are filled using the nearest neighbor interpolation method, and Z-score standardization is performed based on the mean and standard deviation of the data.

[0042] Among these, the three-standard-deviation criterion refers to an outlier identification method based on the assumption of a normal distribution. It considers values ​​in a dataset that exceed the mean plus or minus three standard deviations to be low-probability events and should be removed as outliers. Neighbor interpolation is a missing value imputation method that estimates and fills in missing values ​​by using valid data points adjacent to them, such as linear interpolation or spline interpolation. Z-score standardization is a dimensionless processing method that reduces the original data by its mean and then divides by its standard deviation, making the processed data conform to a distribution with a mean of zero and a standard deviation of one, thus eliminating the influence of differences in dimensions and magnitudes between different indicators.

[0043] The present application will be further described below through a detailed embodiment: In the aforementioned steps, the acquired multi-dimensional measured data are preprocessed.

[0044] For outlier removal, a three-standard-deviation criterion is used. Specifically, for each test index, the mean and standard deviation of all sample data under that index are calculated. The mean plus or minus three standard deviations constitutes a reference interval for normal data. Sample points falling outside this interval are identified as outliers and removed. For example, for the DC leakage current index, if a set of measured data significantly deviates from the normal range, it is removed from the dataset.

[0045] For missing value imputation, the nearest neighbor interpolation method is used. For data gaps caused by outlier removal or data collection issues, linear interpolation is performed using the two valid data points adjacent to the gap, and the interpolation result is used to fill the gap, ensuring the integrity and continuity of the dataset.

[0046] For dimensional normalization, Z-score standardization is employed. For each experimental index, the mean and standard deviation of all sample data under that index are calculated. The mean of the index is subtracted from each original data point, and then the result is divided by the standard deviation to obtain the standardized data value. Through this process, the data of all indices are converted into a dimensionless form with a mean of zero and a standard deviation of one, eliminating the order-of-magnitude differences between different physical quantities such as insulation resistance, dielectric loss factor, and DC leakage current. This ensures that subsequent cross-attention mechanism fusion calculations can treat all indices equally.

[0047] The beneficial effects of this embodiment include at least the following: outliers are eliminated by the three-standard-deviation criterion, effectively avoiding interference from invalid data caused by measurement errors or experimental disturbances on the evaluation results; missing values ​​are filled in by the nearest neighbor interpolation method, ensuring the integrity of the dataset and providing a complete data foundation for subsequent parallel fusion computing; and dimensional normalization is achieved by Z-score standardization, eliminating the magnitude differences between different physical quantities, enabling the cross-attention mechanism to fairly learn the correlation weights between various indicators, and improving the accuracy and stability of fusion computing.

[0048] Based on the aforementioned steps, the pre-set evaluation index system includes evaluation accuracy, data fit, result stability, and three-phase consistency deviation. When quantifying the fusion evaluation results output by each algorithm, the results are weighted and summed according to the weight coefficients corresponding to each evaluation index.

[0049] Among these, evaluation accuracy refers to the degree of consistency between the algorithm's output evaluation result and the known state or expert judgment result, used to measure the correctness of the algorithm's judgment. Data fit refers to the degree of matching between the algorithm's output result and the actual measured data, used to evaluate the algorithm's efficiency in retaining and utilizing original data information. Result stability refers to the degree of fluctuation in the output results of the same algorithm in multiple runs or on different data subsets, used to measure the robustness of the algorithm under different input conditions. Three-phase consistency deviation refers to the degree of difference between the evaluation results of each phase when evaluating the three-phase windings of the transformer separately, used to reflect the algorithm's sensitivity to three-phase imbalance states, indirectly reflecting the rationality of the evaluation. Weighted summation refers to the method of multiplying each evaluation index by its corresponding weight coefficient and then summing them to obtain a comprehensive score, used to achieve a comprehensive consideration of multi-dimensional evaluation indicators.

[0050] The present application will be further described below through a detailed embodiment: In the aforementioned steps, the fusion evaluation results output by each algorithm are quantitatively scored according to a preset evaluation index system in order to select the optimal fusion algorithm.

[0051] The evaluation index system includes four core evaluation indicators: evaluation accuracy, data fit, result stability, and three-phase consistency deviation.

[0052] The accuracy of the assessment is calculated by comparing the assessment results of each algorithm with the actual operating status of the transformer or the reference value determined by the expert group, reflecting the correctness of the algorithm's judgment.

[0053] Data fit is measured by calculating the correlation coefficient or goodness of fit between the evaluation results of each algorithm and the original standardized experimental dataset, reflecting the algorithm's ability to retain the original data information during the fusion process.

[0054] The stability of the results is characterized by randomly sampling the standardized experimental dataset multiple times, inputting it into each algorithm for multiple independent fusion calculations, and calculating the fluctuation variance of the output results of each algorithm.

[0055] The three-phase consistency deviation is used to independently evaluate the performance of the three-phase windings of the transformer, and the maximum deviation between the three-phase evaluation results is calculated to reflect the algorithm's ability to identify and process three-phase unbalanced states.

[0056] When quantifying the fusion evaluation results of each algorithm, a weighted sum is calculated based on the weight coefficients corresponding to evaluation accuracy, data fit, result stability, and three-phase consistency deviation, to obtain the comprehensive score of each algorithm. By comparing the comprehensive scores of each algorithm, the algorithm with the highest score is selected as the optimal fusion algorithm.

[0057] The beneficial effects of this embodiment include at least the following: by constructing a multi-dimensional evaluation index system that includes evaluation accuracy, data fit, result stability, and three-phase consistency deviation, it is possible to comprehensively and objectively measure the overall performance of different algorithms in the transformer high-voltage test data fusion evaluation task; by using a weighted summation method for quantitative scoring, a comprehensive consideration of the evaluation index is achieved, ensuring that the selected optimal fusion algorithm has superior performance in terms of accuracy, robustness, and rationality.

[0058] Based on the aforementioned steps, before performing data fusion calculation, the cross-attention mechanism fusion algorithm uses winding insulation indicators in the standardized test dataset as query vectors, and dielectric loss factor and leakage current indicators as key and value vectors, respectively.

[0059] Among them, winding insulation indicators refer to test parameters used to characterize the insulation performance of transformer windings, such as winding insulation resistance, absorption ratio, and polarization index. These indicators mainly reflect the overall resistive characteristics of the insulating medium. Dielectric loss factor indicators refer to test parameters used to characterize the active power loss of the insulating medium under AC voltage, such as the dielectric loss factor of the winding and bushing, mainly reflecting the polarization loss characteristics of the insulating medium. Leakage current indicators refer to test parameters used to characterize the conductivity of the insulating medium under DC voltage, such as DC leakage current, mainly reflecting the conductivity of the insulating medium and the presence of local defects. Query vectors refer to a type of feature sequence used as the query target in the cross-attention mechanism, used to calculate similarity with other features to determine the focus of attention. Key vectors and value vectors refer to a type of feature sequence used as the query and weighted objects in the cross-attention mechanism, where key vectors are used to calculate similarity with query vectors, and value vectors are used to carry the information for the final weighted sum.

[0060] The present application will be further described below through a detailed embodiment: In the aforementioned steps, before performing data fusion computation, the cross-attention mechanism fusion algorithm classifies the indicators in the standardized experimental dataset to construct query vectors, key vectors, and value vectors.

[0061] Specifically, winding insulation indicators are used as query vectors. These indicators include the insulation resistance to ground of the high-voltage winding, the insulation resistance to ground of the medium-voltage winding, the insulation resistance to ground of the low-voltage winding, the absorption ratio, and the polarization index. These indicators collectively characterize the resistive insulation properties of the transformer windings.

[0062] The dielectric loss factor and leakage current indices are used as key and value vectors, respectively. The dielectric loss factor indices include the dielectric loss factor of the winding and bushing together, as well as the bushing dielectric loss factor, which are used to characterize the polarization loss characteristics of the insulating medium. The leakage current indices include DC leakage current, which are used to characterize the conductivity characteristics of the insulating medium.

[0063] In the cross-attention calculation process, winding insulation indicators are used as query targets, and their similarity with dielectric loss factor and leakage current indicators is calculated to generate attention weights. These weights reflect the coupling relationships between different insulation characteristics: for example, when the winding insulation resistance is low and the dielectric loss factor is high, the attention weights adaptively increase, indicating a possible coupling characteristic of insulation moisture. Through this classification method, the cross-attention mechanism can effectively uncover the deep correlation between resistive and dielectric characteristics.

[0064] The beneficial effects of this embodiment include at least the following: by using winding insulation indicators as query vectors and dielectric loss factor and leakage current indicators as key and value vectors, the coupling relationship between different physical characteristic indicators of transformers is realized through cross-attention mechanism. This can effectively capture the synergistic change law between resistive and dielectric characteristics in typical fault modes such as insulation dampness and aging, and improve the pertinence of defect identification.

[0065] Based on the above embodiments, the weighting coefficient for accuracy is 0.55, the weighting coefficient for data fit is 0.2, the weighting coefficient for result stability is 0.15, and the weighting coefficient for three-phase consistency deviation is 0.1.

[0066] The weighting coefficients for assessment accuracy, data fit, result stability, and three-phase consistency deviation are respectively assigned to the three-phase consistency deviation in the weighted summation calculation.

[0067] The present application will be further described below through a detailed embodiment: In the aforementioned steps, when quantifying the fusion evaluation results of each algorithm output according to the preset evaluation index system, a weighted summation method is used to calculate the comprehensive score. The weight coefficients of each evaluation index are as follows: the weight coefficient of evaluation accuracy is 0.55, the weight coefficient of data fit is 0.2, the weight coefficient of result stability is 0.15, and the weight coefficient of three-phase consistency deviation is 0.1.

[0068] In the specific calculation, the original scores of each algorithm under each evaluation index are first normalized to make them of the same order of magnitude. Then, the evaluation accuracy score is multiplied by 0.55, the data fit score by 0.2, the result stability score by 0.15, and the three-phase consistency deviation score by 0.1. The four products are then added together to obtain the comprehensive score of the algorithm.

[0069] With the aforementioned weighting coefficients, evaluation accuracy carries the highest weight in the overall score, reflecting the paramount importance of algorithmic judgment correctness in the optimization process. Data fit is secondary, indicating that the algorithm's ability to retain information from the original data is also highly valued. Result stability and three-phase consistency deviation are assigned lower weighting coefficients as auxiliary evaluation dimensions. Through this weighting configuration, the selected optimal fusion algorithm first ensures the accuracy of the evaluation results, while also considering information utilization efficiency, robustness, and three-phase balance.

[0070] The beneficial effects of this embodiment include at least the following: by clarifying the weight coefficients of each evaluation index, a quantifiable algorithm selection standard is established, in which the evaluation accuracy rate, as the core index, occupies the highest weight, ensuring the advantage of the selected optimal fusion algorithm in terms of judgment correctness; the reasonable allocation of weight coefficients achieves a comprehensive balance of multi-dimensional performance, highlighting both the core evaluation dimensions and taking into account the auxiliary evaluation dimensions, thereby improving the scientificity and rationality of algorithm selection.

[0071] Based on the aforementioned steps, when determining the performance level of a transformer according to the quantitative score, 90 to 100 points correspond to the excellent level, 80 to 89 points correspond to the good level, 60 to 79 points correspond to the attention level, and below 60 points correspond to the abnormal or serious level. Potential defect types include insulation dampness, insulation aging, local defects, and overheating hazards.

[0072] Among them, the "Excellent" level indicates that the transformer's performance is excellent, all test indicators meet the requirements of the regulations, and there is sufficient safety margin. The "Good" level indicates that the transformer's performance is normal, all test indicators meet the requirements of the regulations, and there are no obvious abnormalities. The "Attention" level indicates that the transformer's performance is somewhat abnormal, some test indicators are close to or slightly exceed the regulations' attention values, and enhanced monitoring is required. The "Abnormal or Severe" level indicates that the transformer's performance is significantly abnormal, test indicators exceed the regulations' limits, and power outage maintenance or other corrective measures are required. Insulation dampness refers to a defect type in the transformer insulation system where moisture intrusion leads to a decline in insulation performance, typically characterized by reduced insulation resistance and increased dielectric loss factor. Insulation aging refers to a defect type in the transformer insulation material where performance deteriorates due to the combined effects of heat, electricity, and mechanical factors during long-term operation, typically characterized by increased dielectric loss factor and abnormal polarization index. Local defects refer to discontinuous defects in the transformer insulation system such as partial discharge, air gaps, and cracks, typically characterized by abnormal DC leakage current and changes in capacitance. Overheating hazards refer to the potential risk of abnormal local temperature rise inside the transformer, typically manifested as accelerated insulation aging and abnormal dielectric loss factor.

[0073] The present application will be further described below through a detailed embodiment: In the aforementioned steps, the standardized test dataset is finally fused and calculated based on the optimal fusion algorithm to obtain a quantitative score of the overall performance of the transformer. The performance level of the transformer is then determined based on the quantitative score, and potential defect types are identified.

[0074] Performance levels are divided into four grades: A quantitative score between 90 and 100 points corresponds to an excellent grade, indicating that all test indicators of the transformer are excellent, insulation performance is good, and it can be safely put into operation or continue to operate; a quantitative score between 80 and 89 points corresponds to a good grade, indicating that all test indicators of the transformer meet the requirements of the regulations, there are no obvious abnormalities, and it can be put into normal operation or routine tests can be arranged according to the cycle; a quantitative score between 60 and 79 points corresponds to a warning grade, indicating that some test indicators of the transformer are abnormal, and it is recommended to strengthen monitoring or shorten the test cycle; a quantitative score below 60 points corresponds to an abnormal or serious grade, indicating that the transformer has obvious defects, and it is recommended to immediately shut down for maintenance or take appropriate measures.

[0075] While determining the performance level, potential defect types are identified by combining the attention weight distribution of each indicator in the cross-attention mechanism fusion algorithm. Specifically, by analyzing the attention weight matrix generated by the cross-attention mechanism, the combination of indicators that contributes significantly to the evaluation results and their correlation patterns are identified. For example, when the weight of winding insulation resistance indicators decreases while the weight of dielectric loss factor indicators increases significantly, it suggests a possible insulation moisture defect; when polarization index and dielectric loss factor indicators show abnormal weight distribution simultaneously, it suggests a possible insulation aging defect; when DC leakage current indicator receives an abnormally high weight, it suggests a possible local defect; and when dielectric loss factor and capacitance indicators show abnormal synergistic weighting, it suggests a possible overheating hazard.

[0076] Through the above methods, this embodiment not only outputs a quantitative score and grade of the overall performance of the transformer, but also provides a preliminary assessment of the defect type by combining the attention weight distribution, providing more instructive reference information for operation and maintenance.

[0077] The beneficial effects of this embodiment include at least the following: by setting four performance levels, the quantitative score is transformed into an intuitive description of the equipment status, which makes it easier for maintenance personnel to quickly grasp the overall condition of the transformer; by combining the attention weight distribution of the cross-attention mechanism to identify potential defect types, the extension from "overall assessment" to "defect localization" is realized, which enhances the practical value of the assessment results; and by combining the performance level classification with defect type identification, data support is provided for the differentiated operation and maintenance and targeted repair of transformers.

[0078] The following is in conjunction with the appendix Figure 3 Taking the application of the high-voltage test data evaluation method based on the cross-attention mechanism provided in this specification in the evaluation of high-voltage handover test data of a 110kV oil-immersed transformer as an example, the high-voltage test data evaluation method based on the cross-attention mechanism will be further explained. Among them, Figure 3 The present specification illustrates a process flowchart of a high-pressure test data evaluation method based on a cross-attention mechanism, according to one embodiment of this specification, which specifically includes the following steps.

[0079] S1. Acquisition and preprocessing of readily available data from high-voltage handover tests. High-voltage commissioning tests were conducted on a newly commissioned 110kV oil-immersed power transformer. Twelve core measured parameters were directly extracted from the testing instruments: high / medium / low voltage winding insulation resistance to ground, absorption ratio, polarization index, winding dielectric loss factor, winding capacitance, DC leakage current, core insulation resistance, clamp insulation resistance, and bushing dielectric loss factor. Three sets of parallel test data were collected, with a sample size of [missing data]. Indicator Number .

[0080] Temperature correction was performed according to the preventive testing procedures for power equipment, adjusting all indicators to the standard temperature of 20℃; one set of abnormal DC leakage current data was eliminated using the 3σ criterion; one set of missing bushing dielectric loss factor data was supplemented using the nearest neighbor interpolation method; and the Z-score standardization formula was used. Complete the dimensional normalization and calculate the mean of each indicator. with standard deviation Finally, a standardized dataset is obtained. There is no data distortion, and the results are completely consistent with the actual on-site measurement results.

[0081] S2. Construct a multi-algorithm fusion comparison library Four types of parallel comparison libraries were built: traditional fusion algorithms (entropy weighted fusion method, grey relational analysis method), conventional machine learning algorithms (random forest RF, support vector machine SVM), and core innovative algorithms (cross-attention mechanism fusion algorithm). The algorithm library is configured with a unified running environment, supports independent parallel computing, and ensures that the comparison results are fair and objective.

[0082] S3, Multi-algorithm Parallel Data Fusion Computation The standardized dataset X from step S1 is simultaneously input into the four types of algorithms, and fusion computation is performed in parallel: (1) Entropy weighting method: Calculate the information entropy and weight of each indicator. The core formula is: The weighted summation outputs an evaluation score of 82.3 points. (2) Grey relational analysis method: Calculate the correlation between each indicator and the ideal healthy sample. The shallow fusion output evaluation score was 80.7 points; (3) Random Forest and Support Vector Machine: Extract shallow features for classification and evaluation, and output scores of 83.5 and 82.9 respectively; (4) Cross-attention mechanism fusion algorithm: using winding insulation indexes as query vectors Dielectric loss and leakage current parameters are used as the key vector K and value vector V, and the vector dimensions are set. Attention count Through the core formula Calculate single-head attention, then apply the multi-head fusion formula:

[0083]

[0084] The system achieves deep integration, automatically learns the correlation weights between indicators, weakens the weights of redundant indicators, strengthens the characteristics of key defects, and finally outputs an evaluation score of 91.2 points with a confidence level of 96.7%.

[0085] S4, Comprehensive Optimization of Fusion Algorithm Construct an algorithm optimization evaluation system to assess accuracy. Data fit Result stability Three-phase consistency deviation The weighting percentages are 55%, 20%, 15%, and 10%, respectively, using the weighted scoring formula:

[0086] The comprehensive scores of each algorithm are calculated as follows: Entropy Weight Method 76.5 points, Grey Relational Analysis 74.2 points, Random Forest 80.3 points, Support Vector Machine 79.6 points, and Cross-Attention Mechanism 94.8 points.

[0087] The cross-attention mechanism fusion algorithm scored significantly higher than the other algorithms, and was therefore selected as the optimal algorithm for evaluating the high-voltage handover test data of the 110kV transformer in this station.

[0088] S5. Comprehensive Performance Evaluation of Transformer Based on the optimal cross-attention mechanism fusion algorithm, the standardized dataset was deeply fused and calculated again, and the final output performance quantification score was 91.2 points. According to the performance level standard, the high-voltage handover test performance level of the 110kV oil-immersed transformer was determined to be excellent. There were no potential hidden dangers such as insulation dampness, aging, or local defects. The fusion evaluation results of all indicators fully met the commissioning standards, and a complete high-voltage handover test performance evaluation report was generated, which was directly used as the basis for transformer commissioning approval.

[0089] The beneficial effects of this application include at least the following.

[0090] This application directly uses readily available measured data output from high-voltage testing instruments, eliminating the need for algorithmic generation or secondary processing of the original data, thus avoiding the risk of data distortion. It strictly conforms to the actual application scenarios of power testing sites and covers all types of high-voltage tests, including routine tests, handover tests, and diagnostic tests, making it widely applicable and highly practical.

[0091] This application explicitly selects the cross-attention mechanism as the core fusion algorithm, which possesses unparalleled core advantages over existing algorithms and completely solves the pain points of traditional technologies: Compared with traditional fixed-weight algorithms such as entropy weight method and grey relational analysis, the cross-attention mechanism can adaptively and dynamically allocate indicator weights without manual intervention, avoiding evaluation bias caused by weight solidification; Compared with conventional machine learning algorithms such as random forest and support vector machine, this algorithm can deeply explore the coupling correlation and interaction relationship between multiple experimental indicators, rather than simply extracting shallow features, resulting in higher accuracy in identifying latent insulation defects; Compared with shallow neural networks, this algorithm has stronger generalization ability through multi-feature cross-interaction and is adaptable to data features of different voltage levels and different experimental scenarios, which is also the core reason for selecting this algorithm in this application.

[0092] This application introduces the cross-attention mechanism into the field of transformer high-voltage test data fusion for the first time. By leveraging its advantages of adaptive feature interaction and dynamic weight allocation, it deeply explores the coupling correlation between multiple indicators, overcomes the shortcomings of traditional fusion algorithms such as fixed weights and one-sided evaluation, and significantly improves the accuracy of data fusion and performance evaluation.

[0093] This application constructs a multi-algorithm comparison and optimization system, which compares the innovative cross-attention mechanism with traditional fusion algorithms and conventional machine learning algorithms in the power field to achieve adaptive selection of the optimal fusion algorithm, taking into account both innovation and engineering practicality. This fills the application gap of emerging algorithms in this field and optimizes the existing experimental data evaluation process.

[0094] The method described in this application has a standardized and highly automated operation process, requiring no manual intervention in weight allocation. It can be adapted to oil-immersed transformers of different voltage levels. The evaluation results are quantitative, intuitive, and in line with power regulations. It can provide accurate data support for transformer operation and maintenance and condition determination, which is in line with the development trend of intelligent operation and maintenance of power grids.

[0095] Corresponding to the above method embodiments, this specification also provides embodiments of a high-pressure test data evaluation device based on a cross-attention mechanism. Figure 4 A schematic diagram of a high-pressure test data evaluation device based on a cross-attention mechanism, according to one embodiment of this specification, is shown. Figure 4 As shown, the device includes: The data acquisition module 401 is configured to acquire multi-dimensional measured data output by the transformer high-voltage test instrument, and perform temperature correction, outlier removal, missing value filling and dimension normalization on the multi-dimensional measured data to obtain a standardized test dataset. The fusion algorithm module 402 is configured to build a multi-algorithm comparison library containing entropy weighted fusion method, grey relational analysis method, fuzzy comprehensive evaluation method, random forest, support vector machine and cross attention mechanism fusion algorithm; The fusion computing module 403 is configured to input the standardized experimental dataset into each algorithm in the multi-algorithm comparison library and perform data fusion computing in parallel. The cross-attention mechanism fusion algorithm divides the indicators in the standardized experimental dataset into query vectors, key vectors and value vectors. It generates attention weights by calculating the similarity between the query vectors and key vectors, and uses the attention weights to perform weighted summation on the value vectors to output the evaluation score and confidence level. The algorithm determination module 404 is configured to quantitatively score the fusion evaluation results output by each algorithm according to a preset evaluation index system, and select the algorithm with the highest score as the optimal fusion algorithm. The performance determination module 405 is configured to use the optimal fusion algorithm to perform fusion calculation on the standardized test dataset to obtain a quantitative score of the overall performance of the transformer, determine the performance level of the transformer based on the quantitative score, and identify potential defect types by combining the attention weight distribution of each index in the cross-attention mechanism fusion algorithm.

[0096] In one possible implementation, the cross-attention mechanism fusion algorithm employs a multi-head attention mechanism, which maps the input to different subspaces through multiple linear projection matrices to compute attention in parallel, and then splices and fuses the attention computation results of each subspace.

[0097] In one possible implementation, outlier removal, missing value imputation, and dimensional normalization are performed on the multidimensional measured data. Specifically, this includes: removing outliers using the three-standard-deviation criterion, imputing missing values ​​using nearest neighbor interpolation, and performing Z-score standardization based on the mean and standard deviation of the data.

[0098] In one possible implementation, the pre-set evaluation index system includes evaluation accuracy, data fit, result stability, and three-phase consistency deviation. When quantifying the fusion evaluation results output by each algorithm, the results are weighted and summed according to the weight coefficients corresponding to each evaluation index.

[0099] In one possible implementation, the cross-attention mechanism fusion algorithm uses winding insulation indices from the standardized test dataset as query vectors and dielectric loss factor and leakage current indices as key and value vectors, respectively, before performing data fusion computation.

[0100] In one possible implementation, the weighting factor for accuracy is 0.55, the weighting factor for data fit is 0.2, the weighting factor for result stability is 0.15, and the weighting factor for three-phase consistency deviation is 0.1.

[0101] In one possible implementation, when determining the performance level of a transformer based on a quantitative score, 90 to 100 points correspond to an excellent level, 80 to 89 points correspond to a good level, 60 to 79 points correspond to a warning level, and below 60 points correspond to an abnormal or severe level. Potential defect types include insulation dampness, insulation aging, local defects, and overheating hazards.

[0102] The above is a schematic scheme of a high-pressure test data evaluation device based on a cross-attention mechanism according to this embodiment. It should be noted that the technical solution of this high-pressure test data evaluation device based on a cross-attention mechanism belongs to the same concept as the technical solution of the high-pressure test data evaluation method based on a cross-attention mechanism described above. For details not described in detail in the technical solution of the high-pressure test data evaluation device based on a cross-attention mechanism, please refer to the description of the technical solution of the high-pressure test data evaluation method based on a cross-attention mechanism described above.

[0103] Figure 5 A structural block diagram of a computing device 500 according to one embodiment of this specification is shown. The components of the computing device 500 include, but are not limited to, a memory 510 and a processor 520. The processor 520 is connected to the memory 510 via a bus 530, and a database 550 is used to store data.

[0104] The computing device 500 also includes an access device 540, which enables the computing device 500 to communicate via one or more networks 560. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 540 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0105] In one embodiment of this specification, the above-described components of the computing device 500 and Figure 5 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 5 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0106] The computing device 500 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 500 can also be a mobile or stationary server.

[0107] The processor 520 executes the following computer-executable instructions, which, when executed by the processor, implement the steps of the high-pressure test data evaluation method based on the cross-attention mechanism described above. The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the high-pressure test data evaluation method based on the cross-attention mechanism described above belong to the same concept. Details not described in detail in the technical solution of the computing device can be found in the description of the technical solution of the high-pressure test data evaluation method based on the cross-attention mechanism described above.

[0108] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the above-described high-pressure test data evaluation method based on a cross-attention mechanism.

[0109] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the technical solution of the high-pressure test data evaluation method based on the cross-attention mechanism described above. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the high-pressure test data evaluation method based on the cross-attention mechanism described above.

[0110] An embodiment of this specification also provides a computer program, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described high-pressure test data evaluation method based on cross-attention mechanism.

[0111] The above is an illustrative example of a computer program according to this embodiment. It should be noted that the technical solution of this computer program belongs to the same concept as the technical solution of the high-pressure test data evaluation method based on the cross-attention mechanism described above. Details not described in detail in the computer program's technical solution can be found in the description of the technical solution of the high-pressure test data evaluation method based on the cross-attention mechanism described above.

[0112] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0113] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0114] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0115] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0116] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A method for evaluating high-voltage test data based on a cross-attention mechanism, characterized in that, include: Acquire multi-dimensional measured data output by transformer high-voltage testing instrument, and perform temperature correction, outlier removal, missing value filling and dimension normalization on the multi-dimensional measured data to obtain a standardized test dataset. Construct a multi-algorithm comparison library that includes entropy weighted fusion method, grey relational analysis method, fuzzy comprehensive evaluation method, random forest, support vector machine and cross-attention mechanism fusion algorithm; The standardized test dataset is input into each algorithm in the multi-algorithm comparison library, and the data fusion calculation is performed in parallel. The cross-attention mechanism fusion algorithm divides the indicators in the standardized test dataset into query vectors, key vectors and value vectors. It generates attention weights by calculating the similarity between the query vectors and the key vectors, and uses the attention weights to perform a weighted summation on the value vectors to output the evaluation score and confidence level. The fusion evaluation results of each algorithm are quantitatively scored according to the preset evaluation index system, and the algorithm with the highest score is selected as the optimal fusion algorithm. The standardized test dataset is fused using the optimal fusion algorithm to obtain a quantitative score of the overall performance of the transformer. The performance level of the transformer is determined based on the quantitative score, and potential defect types are identified by combining the attention weight distribution of each index in the cross-attention mechanism fusion algorithm.

2. The method of claim 1, wherein, The cross-attention mechanism fusion algorithm adopts a multi-head attention mechanism, which maps the input to different subspaces through multiple linear projection matrices to compute attention in parallel, and then splices and fuses the attention computation results of each subspace.

3. The method according to claim 1, characterized in that, The multi-dimensional measured data is subjected to outlier removal, missing value filling and dimension normalization processing, specifically including: removing outliers using the three-standard-deviation criterion, filling missing values ​​using the nearest neighbor interpolation method, and performing Z-score standardization based on the mean and standard deviation of the data.

4. The method according to claim 1, characterized in that, The preset evaluation index system includes evaluation accuracy, data fit, result stability, and three-phase consistency deviation. When quantifying the fusion evaluation results output by each algorithm, the results are weighted and summed according to the weight coefficients corresponding to each evaluation index.

5. The method according to claim 1, characterized in that, Before performing data fusion calculation, the cross-attention mechanism fusion algorithm uses the winding insulation index in the standardized test dataset as the query vector, and the dielectric loss factor and leakage current index as the key vector and value vector, respectively.

6. The method according to claim 4, characterized in that, The weighting coefficient for the accuracy of the assessment is 0.55, the weighting coefficient for the data fit is 0.2, the weighting coefficient for the stability of the results is 0.15, and the weighting coefficient for the three-phase consistency deviation is 0.

1.

7. The method according to claim 1, characterized in that, When determining the performance level of a transformer based on the quantitative score, 90 to 100 points correspond to an excellent level, 80 to 89 points correspond to a good level, 60 to 79 points correspond to a warning level, and below 60 points correspond to an abnormal or serious level. The potential defect types include insulation dampness, insulation aging, local defects, and overheating hazards.

8. A high-pressure test data evaluation device based on a cross-attention mechanism, characterized in that, include: The data acquisition module is configured to acquire multi-dimensional measured data output by the transformer high-voltage test instrument, and perform temperature correction, outlier removal, missing value filling and dimension normalization on the multi-dimensional measured data to obtain a standardized test dataset. The fusion algorithm module is configured to build a multi-algorithm comparison library that includes entropy weighted fusion method, grey relational analysis method, fuzzy comprehensive evaluation method, random forest, support vector machine and cross attention mechanism fusion algorithm; The fusion computing module is configured to input the standardized experimental dataset into each algorithm in the multi-algorithm comparison library and perform data fusion computing in parallel. The cross-attention mechanism fusion algorithm divides the indicators in the standardized experimental dataset into query vectors, key vectors and value vectors, generates attention weights by calculating the similarity between the query vectors and the key vectors, and uses the attention weights to perform weighted summation on the value vectors to output the evaluation score and confidence level. The algorithm determination module is configured to quantitatively score the fusion evaluation results output by each algorithm according to a preset evaluation index system, and select the algorithm with the highest score as the optimal fusion algorithm. The performance determination module is configured to perform fusion calculation on the standardized test dataset using the optimal fusion algorithm to obtain a quantitative score of the overall performance of the transformer, determine the performance level of the transformer based on the quantitative score, and identify potential defect types by combining the attention weight distribution of each index in the cross-attention mechanism fusion algorithm.

9. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the high-pressure test data evaluation method based on the cross-attention mechanism as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the high-pressure test data evaluation method based on the cross-attention mechanism as described in any one of claims 1 to 7.

Images