A transformer gas-sensitive material screening method and device and a storage medium

Abstract

This invention discloses a method, apparatus, and storage medium for screening gas-sensitive materials for transformers, belonging to the technical field of power equipment condition monitoring and gas sensing material design. The method includes: acquiring raw feature data of candidate gas-sensitive materials to be screened; extracting corresponding feature values ​​from the raw feature data according to a pre-determined feature subset, wherein the feature subset is obtained based on feature contribution analysis using a pre-built global prediction model; inputting the extracted feature values ​​into a pre-trained engineering sub-model, and obtaining a comprehensive score for the candidate materials by combining engineering rule constraints; classifying the candidate materials into engineering levels based on the comprehensive score, and outputting a list of materials that meet engineering admission criteria. This invention achieves engineering control of the material exploration process, improving the reliability and engineering applicability of the material screening process.

Description

Technical Field

[0001] This invention relates to a method, apparatus, and storage medium for screening gas-sensitive materials for transformers, belonging to the technical field of power equipment condition monitoring and gas sensing material design. Background Technology

[0002] As a critical piece of equipment in the power system, the operating status of transformers directly affects the safety and stability of the power grid. During long-term operation, faults such as insulation aging, partial discharge, overheating, or arc discharge can gradually generate and accumulate various characteristic gases in transformer oil, such as hydrogen, carbon monoxide, methane, and acetylene. Detection and analysis of these characteristic gases can enable early identification of the transformer's operating status and potential fault types; therefore, transformer fault gas detection is an important technical means in the field of online monitoring of power equipment. Gas sensors are widely used in transformer fault gas detection scenarios due to their advantages such as fast response speed, compact structure, and ease of integration. Among them, gas-sensitive materials, as the core of gas sensors, directly determine the detection performance through their adsorption characteristics, electrical response capabilities, and long-term stability. With the development of materials science, various materials, including metal oxides, transition metal sulfides, carbon-based materials, and two-dimensional metal-doped systems, have been introduced into gas-sensitive material research, significantly expanding the available material options.

[0003] However, the screening of gas-sensitive materials for transformer fault gas detection is significantly complex. On the one hand, the number of combinations of degrees of freedom, such as substrate type, dopant element type, doping method, and structural parameters, is enormous. On the other hand, different fault gases differ in their adsorption mechanisms and response characteristics, resulting in material properties exhibiting obvious nonlinear and multi-factor coupling characteristics. Exhaustively verifying all possible material combinations through experiments or first-principles calculations would consume substantial computational and experimental resources, making it difficult to meet the efficiency and cost requirements of engineering applications.

[0004] Existing research employs machine learning-based methods for selecting gas-sensitive materials. These methods typically acquire the composition, structure, and related physical properties of multiple candidate gas-sensitive materials, establish a material-performance prediction model, and generate a large number of theoretical material combinations within a pre-defined material composition range through computer traversal, performing predictions and evaluations. Then, performance ranking and cluster analysis are combined to select representative materials for experimental verification. While this improves the automation and efficiency of material screening to some extent, it still relies on a single material-performance prediction model as the core of the screening process. When the candidate material space is large or the material system itself is biased, it is difficult to simultaneously consider both the overall characteristics of the material space and the specific engineering application requirements. This can easily lead to path dependence in the early stages of screening, causing the material exploration process to prematurely converge to a local material space. Furthermore, the model prediction results directly drive the material screening and verification process, lacking a mechanism to differentiate between the material's engineering relevance, prediction reliability, and verification priority, making it difficult to finely control the material screening process under limited computational and experimental resources.

[0005] Therefore, there is an urgent need for a gas-sensitive material screening method that can meet the engineering needs of transformer fault gas detection, take into account both the overall spatial understanding of materials and the specificity of engineering applications in the material screening process, effectively schedule the material exploration and verification process by introducing hierarchical modeling and engineering decision-making mechanisms, and combine interpretable analysis and physical verification methods to achieve efficient and reliable engineering material screening. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, device and storage medium for screening gas-sensitive materials for transformers. By constructing a dual-model screening architecture that combines a global prediction model covering the material space with an engineering sub-model for transformer fault gas detection tasks, and by introducing feature contribution analysis and engineering rule constraint mechanisms, the invention achieves engineering control of the material exploration process and improves the reliability and engineering applicability of the material screening process.

[0007] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0008] In a first aspect, the present invention provides a method for screening gas-sensitive materials for transformers, comprising:

[0009] Obtain the original characteristic data of the candidate gas-sensitive materials to be screened;

[0010] Based on a predetermined subset of features, corresponding feature values ​​are extracted from the original feature data, wherein the subset of features is obtained by feature contribution analysis based on a pre-built global prediction model;

[0011] The extracted feature values ​​are input into a pre-trained engineering sub-model, and the comprehensive score of the candidate materials is obtained by combining engineering rule constraints.

[0012] Based on the comprehensive score, the candidate materials are classified into engineering levels, and a list of materials that meet the engineering admission criteria is output.

[0013] Furthermore, the method for constructing the global prediction model includes:

[0014] A basic dataset for gas-sensitive materials is obtained. The basic dataset includes the chemical composition information, structural parameters, and electronic structure parameters of each material sample as input feature parameters, and the gas adsorption performance and electrical response performance data of each material sample as output performance indicators.

[0015] The input feature parameters in the basic dataset are normalized or standardized to obtain the preprocessed basic dataset.

[0016] The preprocessed basic dataset is divided into a training set and a test set, wherein each sample in the training set and the test set contains preprocessed input feature parameters and corresponding output performance metrics.

[0017] A multi-task learning framework is adopted, with the preprocessed input feature parameters of each sample in the training set forming a feature vector as input, and the output performance index corresponding to each sample as output. The training tree model is used as the initial prediction model. The output performance index includes the response intensity, selectivity coefficient, response recovery time, long-term stability score and detection limit of the candidate gas-sensitive material to the target fault gas.

[0018] The initial prediction model is optimized using K-fold cross-validation, with weighted mean squared error as the loss function, to obtain the trained global prediction model. The loss function formula is as follows:

[0019] ;

[0020] in, For loss function, For the number of performance metrics, The total number of samples, For the i-th sample, the value of the m-th indicator is... Let m be the predicted value of the i-th sample for the m-th indicator. The weight of the m-th indicator;

[0021] The trained global prediction model is evaluated using the test set, and the evaluation metric includes the coefficient of determination R. 2 The root mean square error (RMSE) is calculated using the following formula:

[0022] ;

[0023] in, Let m be the mean of the m-th indicator.

[0024] Furthermore, the feature subset is obtained based on feature contribution analysis of a pre-built global prediction model, including:

[0025] Obtain the trained global prediction model;

[0026] Input the sample data into the global prediction model to obtain the prediction results for each sample;

[0027] Based on the sample data, the global prediction model and its prediction results, feature contribution analysis is performed to quantify the contribution of each material feature to key detection performance indicators and select a set of high-contribution features.

[0028] The high-contribution feature set is dimensionality reduced to obtain a feature subset.

[0029] Furthermore, the feature contribution analysis of the global prediction model quantifies the contribution of each material feature to the key detection performance indicators. The calculation formula is as follows: ;

[0030] in, Features Global contribution For the first In the nth sample The contribution value of each feature to the model's prediction results. The total number of samples.

[0031] Furthermore, the dimensionality reduction processing of the high-contribution feature set to obtain a feature subset includes:

[0032] Principal component analysis is used to perform linear dimensionality reduction on the high-contribution feature set. The covariance matrix is ​​constructed, and the eigenvalues ​​and eigenvectors of the principal components are solved, as shown in the following formula: ;

[0033] in, The standardized high-contribution feature matrix, Let covariance matrix be the variance matrix. , The first The eigenvalues ​​and eigenvectors of each principal component The total number of samples;

[0034] Select the top K principal components whose cumulative explained variance reaches a preset threshold to construct a low-dimensional linear feature space;

[0035] Using the low-dimensional linear feature space output by principal component analysis as input, the t-SNE nonlinear dimensionality reduction method is introduced to mine the distribution structure of material samples in the low-dimensional space.

[0036] By combining the high-contribution feature set, the low-dimensional linear feature space, and the distribution structure of material samples in the low-dimensional space, a feature subset is constructed.

[0037] Furthermore, the training method for the engineering sub-model includes:

[0038] Obtain the subset of features as training input features;

[0039] Obtain engineering judgment labels and performance index labels corresponding to the training input features as training supervision data. The engineering judgment labels are used to identify whether the material meets the basic engineering requirements. The performance index labels include measured values ​​or highly reliable calculated values ​​of response intensity, selectivity coefficient, response recovery time, detection limit, and long-term stability score.

[0040] Based on the training input features and the training supervision data, a classification-regression dual-branch coupled model is constructed and trained, wherein:

[0041] The classification branch adopts a lightweight classification algorithm, using the training input features as input and the engineering judgment label as supervision. It is trained using the cross-entropy loss function to obtain the trained classification branch, which is used to output the admission probability value of candidate materials.

[0042] The regression branch employs a lightweight regression algorithm, using the training input features as input, the performance index labels as supervision, and a weighted mean squared error loss function for training.

[0043] During the training of the regression branch, the admission probability value output by the classification branch is used as the sample weight, and the training samples with high admission probability are given higher weights to obtain the trained regression branch.

[0044] Based on the completed classification and regression branches, a comprehensive scoring mechanism is constructed by combining engineering rule constraints. The formula for calculating the comprehensive score is as follows: ;

[0045] in, For comprehensive scoring, For the engineering sub-model to the first The predicted values ​​of each indicator, For the first The weight of each performance indicator For index mapping function, The number of engineering constraints, For candidate materials The predicted value of the constraint index, For the first Engineering thresholds for constraint indicators, Let the degree of violation be a function. For the first The penalty weight of the constraint.

[0046] Furthermore, the training of the engineering sub-model also includes a model iterative correction step, specifically:

[0047] All candidate materials are pre-screened through comprehensive scoring, and materials that meet the pre-screening threshold and key performance standards are selected to form a high-potential candidate subset.

[0048] First-principles calculations were performed on samples in the high-potential candidate subset to obtain core physical quantities, including adsorption energy, charge transfer, work function change, and density of states distribution characteristics.

[0049] Experimental verification was carried out on the samples verified by first-principles calculations, including material preparation, device fabrication, structural characterization, gas response performance testing and stability testing, to obtain experimental verification data;

[0050] Core physical quantities and experimental verification data are used as high-confidence samples to backfill the basic dataset for iterative training of engineering sub-models and global prediction models.

[0051] Furthermore, the step of classifying candidate materials into engineering levels based on comprehensive scores includes:

[0052] Materials whose comprehensive scores reach the pre-screening threshold and are verified by first-principles calculations are classified as engineering access materials, while materials whose comprehensive scores are below the pre-screening threshold are classified as exploration and verification materials.

[0053] The output is a list of materials categorized into engineering access categories. The list of materials includes basic material information, core physical quantities calculated using first-principles calculations, experimental verification data, and standardized experimental reports.

[0054] Secondly, the present invention provides a transformer gas-sensitive material screening device for implementing the transformer gas-sensitive material screening method described in any one of the preceding claims, comprising:

[0055] The data acquisition module is used to acquire the original characteristic data of the candidate gas-sensitive materials to be screened;

[0056] The feature extraction module is used to extract corresponding feature values ​​from the original feature data according to a pre-determined feature subset, wherein the feature subset is obtained by feature contribution analysis based on a pre-built global prediction model;

[0057] The scoring module is used to input the extracted feature values ​​into the pre-trained engineering sub-model and obtain the comprehensive score of the candidate materials by combining the engineering rule constraints.

[0058] The project level classification module is used to classify candidate materials into project levels based on comprehensive scores and output a list of materials that meet the project admission criteria.

[0059] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0060] Fourthly, the present invention provides an electronic device, comprising:

[0061] Memory, used to store computer programs / instructions;

[0062] A processor for executing the computer program / instructions to implement the steps of any of the methods described above.

[0063] Fifthly, the present invention provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0064] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0065] (1) Reduce the path dependence risk of material screening and improve the integrity of material space cognition: By constructing a global prediction model covering the material space, the overall correlation between material composition, structural characteristics and transformer fault gas detection performance is characterized, avoiding the screening process from over-reliance on the initial candidate material system or local sample distribution. This reduces the risk of a single model forming path dependence in the early stage of screening from a mechanism perspective, which is conducive to discovering material directions with potential engineering application value.

[0066] (2) Improve the engineering interpretability and decision reliability of material screening results: Introduce a feature contribution analysis mechanism to quantify the influence of different material composition features, structural features and electronic structure features on key detection performance indicators, and provide interpretable engineering decision basis; Combine engineering sub-model and engineering rule constraints, the engineering correlation of model prediction results can be determined, effectively reducing the cascading amplification of model prediction errors in the screening process and improving the reliability of material screening decisions.

[0067] (3) Realize hierarchical scheduling and cost control of engineering verification resources: divide candidate materials into engineering levels through dual-model prediction results, and only perform first-principles calculations or experimental verifications on materials that meet engineering access requirements, avoiding unified verification of a large number of materials with low engineering relevance, thereby effectively controlling the investment of high-cost calculation and experimental verification resources while ensuring material exploration capabilities. Attached Figure Description

[0068] Figure 1 This is a flowchart of a method for screening gas-sensitive materials for transformers provided in an embodiment of the present invention;

[0069] Figure 2 This is a schematic diagram of the gas testing platform provided in an embodiment of the present invention. Detailed Implementation

[0070] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0071] Example 1: This example describes a method for screening gas-sensitive materials for transformers, including:

[0072] Obtain the original characteristic data of the candidate gas-sensitive materials to be screened;

[0073] Based on a predetermined subset of features, corresponding feature values ​​are extracted from the original feature data, wherein the subset of features is obtained by feature contribution analysis based on a pre-built global prediction model;

[0074] The extracted feature values ​​are input into a pre-trained engineering sub-model, and the comprehensive score of the candidate materials is obtained by combining engineering rule constraints.

[0075] Based on the comprehensive score, the candidate materials are classified into engineering levels, and a list of materials that meet the engineering admission criteria is output.

[0076] like Figure 1 As shown in this embodiment, the method for screening transformer gas-sensitive materials involves the following steps in its application:

[0077] Step 1 (Data Acquisition and Preprocessing): Specific Implementation Techniques

[0078] We constructed a basic dataset of "materials-gas-performance" and normalized it to ensure that data from different sources are comparable and trainable, providing input for subsequent model training and engineering judgment.

[0079] (1) Data collection, the specific process is as follows:

[0080] The data acquisition and storage module obtains basic data on candidate gas-sensitive materials. Data sources may include published papers / patents, public databases, existing simulation or calculation results, and a small number of experimental or first-principles (DFT) high-confidence samples. Candidate material systems can cover metal oxides, transition metal sulfides, two-dimensional materials and their metal-doped systems, carbon-based composite materials, heterojunction materials, etc. Specific acquisition fields include:

[0081] Substrate / material characteristics: material composition (elements, doping type and proportion), structural parameters (lattice constant, defect density, surface active site density, specific surface area / porosity, etc.), electronic structure related characteristics (band gap, work function, d-band center, density of states statistics, and other available indicators).

[0082] Gas molecule characteristics: molecular size / van der Waals radius, relative molecular mass, polarity / dipole moment, polarizability, first ionization energy or electron affinity related indicators, molecular complexity / branching index, etc.

[0083] Performance label / target: Response strength, selectivity, response / recovery time, stability, detection limit, etc. to the target fault gas (such as H2, CO, etc.);

[0084] (2) Normalization process, the specific process is as follows:

[0085] Z-score normalization is applied to feature j:

[0086] ;

[0087] in: These are the original eigenvalues; Let j be the mean of feature j on the training set; Let j be the standard deviation of feature j on the training set; Let be the standardized eigenvalue of feature j.

[0088] Step Two (Building a Global Prediction Model): Specific Implementation Techniques

[0089] A global prediction model for "overall laws of material space" is trained to learn the mapping relationship between material characteristics and multiple performance indicators, gain an overall understanding of the material space (trends, boundaries, transferable features), and provide a foundation for subsequent "feature contribution analysis + engineering sub-model construction". The specific steps are as follows:

[0090] (1) Multi-indicator output, the specific process is as follows:

[0091] The objective is defined as multi-task learning: simultaneously predicting multiple performance metrics such as sensitivity / response strength, selectivity, response time, stability, and detection limit. A regression head is used for regression metrics; a classification head can be used for grade or pass / fail determinations.

[0092] (2) Model form, the specific process is as follows:

[0093] Global prediction models can employ gradient boosting trees, random forests, multilayer perceptrons, or graph neural networks (if crystal / graph structure encoding is available). This example uses a "tree model + multi-task regression / classification" structure, with the input being the feature vector obtained in step one.

[0094] (3) Dataset partitioning and cross-validation, the specific process is as follows:

[0095] The dataset is divided into training and testing sets (e.g., 8:2) based on the material system or gas type, and K-fold cross-validation is used. The loss function can be the weighted mean square error.

[0096] ;

[0097] in, For loss function, For the number of performance metrics, The total number of samples, For the i-th sample, the value of the m-th indicator is... Let m be the predicted value of the i-th sample for the m-th indicator. The weight of the m-th indicator;

[0098] The evaluation metrics for model prediction performance m include the coefficient of determination (R²) and the root mean square error (RMSE).

[0099] ;

[0100] in, Let m be the mean of the m-th indicator.

[0101] Step 3 (Feature Contribution Analysis): Specific Implementation Techniques

[0102] The strategy of "feature contribution analysis + linear dimensionality reduction + nonlinear structure exploration" is adopted. First, the contribution weight of features to each performance index is quantified. Then, redundancy removal, noise reduction, and structural analysis are performed on high-dimensional material features to form a stable, interpretable feature subset suitable for engineering decision-making. This provides high-quality input for subsequent engineering sub-model training. The specific steps are as follows:

[0103] (1) Feature contribution analysis, the specific process is as follows:

[0104] A feature contribution analysis method is introduced into the global prediction model trained in step two to quantify the marginal contribution of the original material features to each detection performance index (sensitivity, selectivity, response time, stability, detection limit, etc.). For each index output by the global model, the contribution value of feature j to the prediction result of sample i is calculated, and the global contribution of feature j is statistically analyzed.

[0105] ;

[0106] in: The global contribution of feature j (used for ranking and filtering); This represents the contribution of the j-th feature in the i-th sample to the model's prediction result.

[0107] Based on the global contribution ranking, a set of original features with high contributions that have a significant impact on key performance indicators is selected, and redundant features with low contribution and no physical meaning are removed to avoid interference from noisy features in the subsequent dimensionality reduction process.

[0108] (2) Principal component analysis for linear dimensionality reduction, the specific process is as follows:

[0109] Using the original set of high-contribution features as input, a feature covariance matrix is ​​constructed and orthogonal transformation by principal component analysis (PCA) is performed to map the original correlated features into several independent principal components, thereby reducing feature redundancy and suppressing noise propagation.

[0110] ;

[0111] in: This is the standardized high-contribution feature matrix; It is the covariance matrix; , These are the eigenvalues ​​and their corresponding eigenvectors of the kth principal component;

[0112] Calculate the loading coefficients of the original high-contribution features to each principal component, and identify the core original features that dominate each principal component based on the absolute value of the loading coefficients. Select the top K principal components whose cumulative explained variance reaches a preset threshold (e.g., 90%–95%), and construct a low-dimensional linear feature space to achieve feature dimensionality reduction while preserving core information.

[0113] (3) t-distributed random neighborhood embedding (t-SNE) nonlinear dimensionality reduction, the specific process is as follows:

[0114] Using the low-dimensional linear feature space output by PCA as input, the t-SNE nonlinear dimensionality reduction method is introduced to further explore the distribution structure of material samples in the low-dimensional space. This is used to reveal the potential clustering characteristics and nonlinear correlations between different material systems, doping methods, or gas response behaviors, and can identify material clusters with similar performance but large differences in composition.

[0115] (4) Construct a feature subset, the specific process is as follows:

[0116] By combining the Shapley Additive Explanation (SHAP) feature contribution screening results, PCA dimensionality reduction results, and t-SNE structure exploration results, a feature subset is formed. This subspace retains key features that contribute significantly to prediction performance and have physical significance, while also possessing low dimensionality and strong stability, making it suitable as input for subsequent engineering sub-models.

[0117] Step Four (Constructing the Engineering Sub-Model): Specific Implementation Techniques

[0118] A classification-regression dual-branch engineering sub-model is constructed, which integrates explicit engineering rule constraints to form a quantitative scoring system. The model is then corrected using high-confidence samples (DFT / experiment), thereby transforming the "broad-coverage cognition" of the global model into "engineering-usable decision-making." The specific steps are as follows:

[0119] (1) Construction of the dual-branch engineering sub-model, the specific process is as follows:

[0120] The feature subset obtained in step three Using this as input, a classification-regression dual-branch coupled model is constructed to balance engineering admission criteria and quantitative prediction of performance indicators:

[0121] 1. Engineering Access Classification Branches;

[0122] Model selection: Use lightweight classification algorithms, such as logistic regression, lightweight gradient boosting trees, or shallow neural networks.

[0123] Model objective: Determine whether candidate materials initially meet the basic engineering requirements for transformer fault gas detection, and output the admission probability value.

[0124] Training constraints: The training set is a subset of features. The corresponding samples should cover mainstream material systems such as metal oxides and two-dimensional materials. The loss function is cross-entropy loss, and the evaluation indicators are accuracy, precision, and recall. Priority is given to ensuring recall to avoid missing potential high-performance materials.

[0125] 2. Performance index regression branch;

[0126] Model selection: Use lightweight regression algorithms that match the classification branch, such as linear regression and lightweight gradient boosting tree regressors.

[0127] Model objective: To quantitatively predict the core performance indicators of candidate materials, including response intensity, selectivity coefficient, response recovery time, detection limit, and long-term stability score to target gases such as H2 and CO.

[0128] Training constraints: The performance labels after preprocessing in step one are used as supervised data. The loss function is weighted mean square error, and the evaluation metrics are the coefficient of determination R2 and the root mean square error RMSE.

[0129] 3. Two-branch coupling;

[0130] The "admission probability" output by the classification branch is used as the weight coefficient of the regression branch. For samples with high admission probability, the regression prediction accuracy is enhanced; for samples with low admission probability, their weight in subsequent decisions is reduced.

[0131] (2) Comprehensive scoring mechanism, the specific process is as follows:

[0132] Engineering rules are used to generate admission scores / constraints, which, together with the predictive performance of the engineering sub-models, determine the overall score.

[0133] ;

[0134] in, For comprehensive scoring; This is the prediction of the m-th indicator by the engineering sub-model; Let m be the weight of the m-th performance indicator; For index mapping functions; The number of project constraints; Let r be the predicted value of the constraint index of the candidate material; Let r be the engineering threshold of the r-th constraint index; Let the degree of violation be a function. For the first Penalty weights for each constraint;

[0135] (3) Iterative model correction, the specific process is as follows:

[0136] The model parameters are iteratively corrected through DFT verification and experimental validation data backfilling, and the decision threshold is dynamically generated as the model iterates. The specific steps are as follows:

[0137] 1. Model scoring pre-screening;

[0138] For all candidate materials, a comprehensive score is first calculated. Filter out those that meet the requirements ≥Pre-screening threshold Furthermore, materials that meet the key performance standards constitute a high-potential candidate subset. ;

[0139] 2. DFT theoretical verification;

[0140] high-potential candidate subsets The samples in the sample are subjected to DFT calculations to obtain core physical quantities, including the adsorption energy of gas molecules on the material surface, charge transfer amount, work function change value, and density of states distribution characteristics.

[0141] 3. Experimental verification;

[0142] Experimental verification was carried out on the samples that passed the DFT verification and screening. The testing process included material preparation and device fabrication, structural composition characterization, gas response performance testing, and stability testing.

[0143] 4. Model correction;

[0144] The physical quantity data calculated by DFT and the performance data verified by experiments are used together as high-confidence samples to fill the basic dataset for iterative training of engineering sub-models and global prediction models.

[0145] Step 5 (Output of Stratified and Filtered Results): Specific Implementation Techniques

[0146] (1) Layered division, the specific process is as follows:

[0147] The overall score output from step four And hierarchical determination, implementing differentiated processing strategies for different categories of materials to avoid resource waste:

[0148] Engineering Access Category ( ≥ (And verified via DFT): Prioritize the allocation of experimental resources to conduct comprehensive performance verification. Simultaneously record DFT physical quantity data and experimental test data as high-confidence samples for subsequent model iterations;

[0149] Exploration and verification (low-frequency verification): < We do not perform high-cost verification, but keep records and make predictions and evaluations after the model is updated.

[0150] (2) The output content is as follows:

[0151] Output a material list that is "dual-qualified in DFT and experiment". The list includes: ① basic material information (composition, structural parameters); ② core physical quantities of DFT (adsorption energy, charge transfer, etc.); ③ experimental performance data (detection limit, response recovery time, etc.); ④ standardized experimental report (including characterization spectrum and response curve) to support subsequent device development.

[0152] like Figure 2 The diagram shows a schematic of a gas testing platform, including a hydrogen gas source, a carbon monoxide gas source, a mass flow controller, a testing chamber, an analysis system, and a computer. The hydrogen and carbon monoxide gas sources are connected to the mass flow controller via pipelines, which regulates the gas flow rate entering the testing chamber. The gas-sensitive material to be tested is placed inside the testing chamber, and the output of the testing chamber is connected to the analysis system. The analysis system collects the response signal generated by the gas-sensitive material under the action of the target gas and transmits the response signal to the computer for display and recording.

[0153] This embodiment has the following beneficial effects:

[0154] (1) Reduce the path dependence risk of material screening and improve the integrity of material space cognition: By constructing a global prediction model covering the material space, the overall correlation between material composition, structural characteristics and transformer fault gas detection performance is characterized, avoiding the screening process from over-reliance on the initial candidate material system or local sample distribution. This reduces the risk of a single model forming path dependence in the early stage of screening from a mechanism perspective, which is conducive to discovering material directions with potential engineering application value.

[0155] (2) Improve the engineering interpretability and decision reliability of material screening results: Introduce a feature contribution analysis mechanism to quantify the influence of different material composition features, structural features and electronic structure features on key detection performance indicators, and provide interpretable engineering decision basis; Combine engineering sub-model and engineering rule constraints, the engineering correlation of model prediction results can be determined, effectively reducing the cascading amplification of model prediction errors in the screening process and improving the reliability of material screening decisions.

[0156] (3) Realize hierarchical scheduling and cost control of engineering verification resources: divide candidate materials into engineering levels through dual-model prediction results, and only perform first-principles calculations or experimental verifications on materials that meet engineering access requirements, avoiding unified verification of a large number of materials with low engineering relevance, thereby effectively controlling the investment of high-cost calculation and experimental verification resources while ensuring material exploration capabilities.

[0157] This embodiment has the following socio-economic benefits:

[0158] (1) Improve the R&D efficiency of transformer fault gas detection materials: By using an engineered material screening and verification scheduling mechanism, unnecessary high-precision calculations and experimental verifications can be reduced, which can significantly shorten the R&D cycle of transformer fault gas detection sensing materials, reduce R&D costs, and improve material screening efficiency.

[0159] (2) Enhance the reliability and engineering applicability of the online transformer monitoring system: By screening, the sensing materials with higher engineering relevance and better performance stability are obtained, which helps to improve the accuracy of transformer fault gas detection and long-term operational reliability, and provides more reliable technical support for power equipment condition monitoring and fault early warning, which has good engineering application and promotion value.

[0160] Example 2: This example provides a transformer gas-sensitive material screening device, comprising:

[0161] The data acquisition module is used to acquire the original characteristic data of the candidate gas-sensitive materials to be screened;

[0162] The feature extraction module is used to extract corresponding feature values ​​from the original feature data according to a pre-determined feature subset, wherein the feature subset is obtained by feature contribution analysis based on a pre-built global prediction model;

[0163] The scoring module is used to input the extracted feature values ​​into the pre-trained engineering sub-model and obtain the comprehensive score of the candidate materials by combining the engineering rule constraints.

[0164] The project level classification module is used to classify candidate materials into project levels based on comprehensive scores and output a list of materials that meet the project admission criteria.

[0165] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0166] Example 3: This example provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described in Example 1.

[0167] Example 4: This example provides an electronic device, including:

[0168] Memory, used to store computer programs / instructions;

[0169] A processor for executing the computer program / instructions to implement the steps of any of the methods described in Embodiment 1.

[0170] Example 5: This example provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described in any one of Examples 1.

[0171] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0172] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0173] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0174] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0175] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0176] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit its protection scope. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading this disclosure, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the pending claims.

Claims

1. A method for screening gas-sensitive materials for transformers, characterized in that, include: Obtain the original characteristic data of the candidate gas-sensitive materials to be screened; Based on a predetermined subset of features, corresponding feature values ​​are extracted from the original feature data. This subset is obtained through feature contribution analysis using a pre-built global prediction model, specifically including: A basic dataset for gas-sensitive materials is obtained. The basic dataset includes the chemical composition information, structural parameters, and electronic structure parameters of each material sample as input feature parameters, and the gas adsorption performance and electrical response performance data of each material sample as output performance indicators. The input feature parameters in the basic dataset are normalized or standardized to obtain the preprocessed basic dataset. The preprocessed basic dataset is divided into a training set and a test set, wherein each sample in the training set and the test set contains preprocessed input feature parameters and corresponding output performance metrics. A multi-task learning framework is adopted, with the preprocessed input feature parameters of each sample in the training set forming a feature vector as input, and the output performance index corresponding to each sample as output. The training tree model is used as the initial prediction model. The output performance index includes the response intensity, selectivity coefficient, response recovery time, long-term stability score and detection limit of the candidate gas-sensitive material to the target fault gas. The initial prediction model is optimized by K-fold cross-validation, and the weighted mean square error is used as the loss function to obtain the trained global prediction model. Input the sample data into the global prediction model to obtain the prediction results for each sample; Based on the sample data, the global prediction model and its prediction results, feature contribution analysis is performed to quantify the contribution of each material feature to key detection performance indicators and select a set of high-contribution features. Dimensionality reduction is performed on the high-contribution feature set to obtain a feature subset; The extracted feature values ​​are input into a pre-trained engineering sub-model, and a comprehensive score for candidate materials is obtained by combining engineering rule constraints, including: Obtain the subset of features as training input features; Obtain engineering judgment labels and performance index labels corresponding to the training input features as training supervision data. The engineering judgment labels are used to identify whether the material meets the basic engineering requirements. The performance index labels include measured values ​​or highly reliable calculated values ​​of response intensity, selectivity coefficient, response recovery time, detection limit, and long-term stability score. Based on the training input features and the training supervision data, a classification-regression dual-branch coupled model is constructed and trained, wherein: The classification branch adopts a lightweight classification algorithm, using the training input features as input and the engineering judgment label as supervision. It is trained using the cross-entropy loss function to obtain the trained classification branch, which is used to output the admission probability value of candidate materials. The regression branch employs a lightweight regression algorithm, using the training input features as input, the performance index labels as supervision, and a weighted mean squared error loss function for training. During the training of the regression branch, the admission probability value output by the classification branch is used as the sample weight. Training samples with high admission probabilities are given higher weights to obtain the trained regression branch. Based on the completed classification and regression branches, and combined with engineering rule constraints, a comprehensive scoring mechanism is constructed. The formula for calculating the comprehensive score is as follows: ; in, For comprehensive scoring, For the engineering sub-model to the first The predicted values ​​of each indicator, For the first The weight of each performance indicator For index mapping function, The number of engineering constraints, For candidate materials The predicted value of the constraint index, For the first Engineering thresholds for constraint indicators, Let the degree of violation be a function. For the first Penalty weights for constraints; Based on the comprehensive score, the candidate materials are classified into engineering levels, and a list of materials that meet the engineering admission criteria is output.

2. The method for screening transformer gas-sensitive materials according to claim 1, characterized in that, The loss function formula is as follows: ； in, For loss function, For the number of performance metrics, The total number of samples, For the i-th sample, the value of the m-th indicator is... Let m be the predicted value of the i-th sample for the m-th indicator. Let m be the weight of the m-th indicator.

3. The method for screening transformer gas-sensitive materials according to claim 2, characterized in that, The method further includes: evaluating the trained global prediction model using the test set, wherein the evaluation metric includes the coefficient of determination R. 2 The root mean square error (RMSE) is calculated using the following formula: ； in, Let m be the mean of the m-th indicator.

4. The method for screening transformer gas-sensitive materials according to claim 3, characterized in that, The contribution of each material characteristic to the key detection performance indicators is calculated using the following formula: ; in, Features Global contribution For the first In the nth sample The contribution value of each feature to the model's prediction results. The total number of samples.

5. The method for screening transformer gas-sensitive materials according to claim 3, characterized in that, The dimensionality reduction process performed on the high-contribution feature set yields a feature subset, including: Principal component analysis is used to perform linear dimensionality reduction on the high-contribution feature set. The covariance matrix is ​​constructed, and the eigenvalues ​​and eigenvectors of the principal components are solved, as shown in the following formula: ; in, The standardized high-contribution feature matrix, Let covariance matrix be the variance matrix. , The first The eigenvalues ​​and eigenvectors of each principal component The total number of samples; Select the top K principal components whose cumulative explained variance reaches a preset threshold to construct a low-dimensional linear feature space; Using the low-dimensional linear feature space output by principal component analysis as input, the t-SNE nonlinear dimensionality reduction method is introduced to mine the distribution structure of material samples in the low-dimensional space. By combining the high-contribution feature set, the low-dimensional linear feature space, and the distribution structure of material samples in the low-dimensional space, a feature subset is constructed.

6. The method for screening transformer gas-sensitive materials according to claim 1, characterized in that, The training of the engineering sub-model also includes an iterative model correction step, specifically: All candidate materials are pre-screened through comprehensive scoring, and materials that meet the pre-screening threshold and key performance standards are selected to form a high-potential candidate subset. First-principles calculations were performed on samples in the high-potential candidate subset to obtain core physical quantities, including adsorption energy, charge transfer amount, work function change value and density of states distribution characteristics; Experimental verification was carried out on the samples verified by first-principles calculations, including material preparation, device fabrication, structural characterization, gas response performance testing and stability testing, to obtain experimental verification data; Core physical quantities and experimental verification data are used as high-confidence samples to backfill the basic dataset for iterative training of engineering sub-models and global prediction models.

7. The method for screening transformer gas-sensitive materials according to claim 1, characterized in that, The process of classifying candidate materials into engineering levels based on comprehensive scores includes: Materials whose comprehensive scores reach the pre-screening threshold and are verified by first-principles calculations are classified as engineering access materials, while materials whose comprehensive scores are below the pre-screening threshold are classified as exploration and verification materials. The output is a list of materials categorized into engineering access categories. The list of materials includes basic material information, core physical quantities calculated using first-principles calculations, experimental verification data, and standardized experimental reports.

8. A transformer gas-sensitive material screening device, used to implement the transformer gas-sensitive material screening method according to any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire the original characteristic data of the candidate gas-sensitive materials to be screened; The feature extraction module is used to extract corresponding feature values ​​from the original feature data according to a pre-determined feature subset, wherein the feature subset is obtained by feature contribution analysis based on a pre-built global prediction model; The scoring module is used to input the extracted feature values ​​into the pre-trained engineering sub-model and obtain the comprehensive score of the candidate materials by combining the engineering rule constraints. The project level classification module is used to classify candidate materials into project levels based on comprehensive scores and output a list of materials that meet the project admission criteria.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the computer program implements the steps of the method described in any one of claims 1-7.

Images