Method for screening gas sensitive materials for offshore wind farm transformer fault marker gas detection

Abstract

A method for screening gas-sensitive materials for detecting fault marker gases in offshore wind farm transformers, belonging to the field of condition monitoring technology for power equipment in offshore wind farms, includes: a target gas determination module to determine the corresponding fault marker gas; a candidate material library construction module to construct candidate adsorption systems based on the fault marker gas; a descriptor generation module to construct descriptors for each candidate combination, obtaining a descriptor set; a high-throughput prediction module to predict the adsorption energy of each candidate combination, outputting candidate combinations with adsorption energy values ​​within a set adsorption energy range to a DFT verification module; a DFT verification module to perform DFT verification on the candidate combinations, passing the verification to a material list output module; and a material list output module to score the verification to the qualified candidate combinations using a multi-index scoring function, and then outputting a candidate material list. The effect is that it can quickly screen candidate gas-sensitive materials that possess adsorption response, recoverability, and environmental adaptability.

Description

Technical Field

[0001] This invention relates to the field of condition monitoring technology for power equipment in offshore wind farms, and in particular to a method for screening gas-sensitive materials for detecting fault indicators of transformers in offshore wind farms. Background Technology

[0002] Offshore wind farms are typically located far from land, and transformers and related electrical equipment operate in environments characterized by high humidity, high salt spray, strong wind vibration, temperature fluctuations, and limited maintenance windows. If oil-immersed equipment such as main transformers at offshore substations and wind turbine box-type transformers experiences latent faults such as insulation aging, partial discharge, overheating, or arc discharge, it can cause the transformer oil and solid insulation materials to decompose, generating and releasing [chemical reactions]. , , , Fault marker gases. Currently, fault gas detection in offshore wind farm transformers typically relies on methods such as dissolved gas analysis in oil, chromatographic analysis, or gas sensor arrays. Among these, the detection sensitivity, selectivity, response / recovery capability, and environmental stability of the gas sensor largely depend on the gas-sensitive material itself. Two-dimensional transition metal chalcogenides and their transition metal-doped systems possess advantages such as large specific surface area, tunable surface activity, and easily modulated electronic structure, and are considered potential gas-sensitive materials for fault marker gas detection.

[0003] However, existing gas-sensitive material discovery still suffers from the following shortcomings: First, traditional experimental trial-and-error methods require the preparation, characterization, and gas-sensitivity testing of a large number of candidate materials one by one, which is time-consuming and costly, making it difficult to meet the needs of rapid material iteration for large-scale condition monitoring in offshore wind farms; Second, while relying solely on DFT calculations can reveal gas adsorption mechanisms at the atomic scale, the exhaustive computational load is enormous when faced with multiple substrates, multiple dopant elements, and multiple combinations of fault gases; Third, existing machine learning material screening methods are often not closely integrated with the fault scenarios of transformers in offshore wind farms, lacking a complete closed-loop process from defining fault gas targets, constructing candidate material space, generating material computational data, feature screening, selecting the optimal algorithm, to verifying candidate materials; Fourth, some prediction models only provide numerical results, lacking explanations of the mechanisms of action of dopant elements, gas molecule characteristics, and substrate characteristics, which is not conducive to subsequent material design and engineering applications.

[0004] Current technological shortcomings: The current screening of gas-sensitive materials for fault marker gas detection in offshore wind farm transformers faces problems such as long R&D cycles, large candidate material pool, high cost of exhaustive first-principles calculations (DFT), and lack of physical interpretation and engineering screening closure loop in traditional machine learning models. Summary of the Invention

[0005] This invention provides a method for screening gas-sensitive materials for detecting fault markers in offshore wind farm transformers, which can quickly screen out candidate gas-sensitive materials for detecting fault markers in offshore wind farm transformers.

[0006] To achieve the above objectives, this invention provides a method for screening gas-sensitive materials for detecting fault marker gases in offshore wind farm transformers, the key of which includes the following steps:

[0007] Step 1: Construct a gas-sensitive material screening system for detecting fault marker gases in offshore wind farm transformers based on GBR algorithm and material calculation. This gas-sensitive material screening system for detecting fault marker gases in offshore wind farm transformers is set up with a target gas determination module, a candidate material library construction module, a descriptor generation module, a high-throughput prediction module, a DFT verification module, and a material list output module connected in sequence.

[0008] Step 2: The target gas determination module determines the fault modes that are easily induced in offshore wind farm transformers based on the operating characteristics of the transformers, and determines the corresponding fault marker gases.

[0009] Step 3: The candidate material library construction module constructs a candidate gas-sensitive material space based on the fault marker gas to obtain a candidate adsorption system. The candidate adsorption system contains one candidate combination consisting of "substrate-doped atoms-fault gas".

[0010] Step 4: The descriptor generation module constructs descriptors for each candidate combination from three dimensions: substrate, doped atoms, and fault gas, thereby obtaining a descriptor set;

[0011] Step 5: The high-throughput prediction module predicts the adsorption energy of each candidate combination based on the descriptor set, and then outputs candidate combinations whose adsorption energy values ​​are within the set adsorption energy range and cover various target fault gases to the DFT verification module.

[0012] Step 6: The DFT verification module performs DFT verification on the candidate combinations selected in Step 5. When the adsorption strength term of the candidate combination... Electronic response items Recovery ability item Stable items and environmental adaptation items If all the set requirements are met, the verification is successful, and then the successful candidate combination is passed to the bill of materials output module;

[0013] Step 7: The material list output module scores the qualified candidate combinations using a multi-index scoring function, and then outputs the candidate material list from high to low scores. The substrate-doped atomic material corresponding to the candidate combination in the candidate material list is the gas-sensitive material for detecting fault marker gas in offshore wind farm transformers that was finally discovered.

[0014] Through the above design, this invention achieves high-throughput prediction, rapid screening, and interpretable recommendation of gas-sensitive materials for detecting fault gases in offshore wind farm transformers by constructing a closed-loop process of "offshore wind power fault scenario determination—fault marker gas determination—candidate adsorption system construction—physicochemical descriptor conversion—adsorption energy model prediction—DFT multi-index verification—material list output." This provides an effective basis for selecting sensor materials for online monitoring of offshore wind farm transformers. It significantly shortens material screening time, reduces system computational costs, and ensures the accuracy and effectiveness of material screening while improving material screening efficiency.

[0015] Preferably, in step 2, the target gas determination module determines the fault marker gas based on the operating characteristics of the offshore wind farm transformer as follows:

[0016] When the transformer in an offshore wind farm is an oil-immersed transformer or a box-type transformer, its operating characteristics include, but are not limited to, load fluctuations, compact space, and limited heat dissipation. Fault modes include, but are not limited to, low-temperature overheating, high-temperature overheating, and insulation aging. Fault indicator gases include, but are not limited to, those mentioned above. , , ;

[0017] When the transformer in an offshore wind farm serves as the main transformer for an offshore substation, its operating characteristics include, but are not limited to, large capacity, high voltage, and complex oil-paper insulation. Fault modes include, but are not limited to, partial discharge, arc discharge, and high-temperature overheating. Fault indicator gases include, but are not limited to, [missing information]. , , ;

[0018] When the marine environment is characterized by one or more of the following combinations: high humidity, high salt spray, strong wind vibration, and temperature difference cycling, the failure modes of offshore wind farm transformers include, but are not limited to, insulation damping, local defects, and increased contact resistance. Fault indicator gases include, but are not limited to, those mentioned above. , , , .

[0019] Based on different marine environments and the operating characteristics of different transformers, this invention determines the corresponding easily induced fault modes, and then determines the corresponding fault indicator gases, thereby improving the accuracy of fault indicator gas determination.

[0020] Preferably, in step 3, a two-dimensional transition metal chalcogenide and its Janus or alloyed derivatives are selected as the substrate, including but not limited to tungsten disulfide. Molybdenum diselenide tungsten diselenide and tungsten, sulfur, selenium ;

[0021] Transition metals are selected as doping atoms, including but not limited to scandium (Sc), titanium (Ti), vanadium (V), chromium (Cr), manganese (Mn), iron (Fe), cobalt (Co), nickel (Ni), copper (Cu), zinc (Zn), yttrium (Y), zirconium (Zr), niobium (Nb), molybdenum (Mo), technetium (Tc), ruthenium (Ru), rhodium (Rh), palladium (Pd), silver (Ag), cadmium (Cd), hafnium (Hf), tantalum (Ta), tungsten (W), rhenium (Re), osmium (Os), iridium (Ir), platinum (Pt), and gold (Au).

[0022] Preferably, in step 5, the high-throughput prediction module uses a gradient boosting regression model to predict the adsorption energy of the candidate combinations, as expressed by:

[0023] ;

[0024] in, Let be the predicted adsorption energy value for the i-th candidate combination. For gradient boosting regression models, For the m-th regression tree, For the total number of regression trees, Let be the model parameters for the m-th regression tree. Let be the descriptor vector for the i-th candidate combination.

[0025] Preferably, in step 6, the DFT verification module performs DFT verification on the candidate combinations, and the specific process is as follows:

[0026] Step 6.1: The DFT verification module establishes corresponding clean substrate, doped substrate, isolated gas and adsorption system models according to the selected candidate combinations, and performs geometric optimization.

[0027] Step 6.2: Perform DFT calculations on each candidate combination to obtain the corresponding candidate combination parameters, including but not limited to the adsorption energy of the candidate combination. Adsorption distance Charge transfer amount band gap band gap variation Change in density of states Change in work function Theoretical sensitivity and recovery time ;

[0028] Step 6.3: Calculate the adsorption strength term based on the candidate combination parameters. Electronic response items Recovery ability item Stable items and environmental adaptation items ;

[0029] Step 6.4: Adsorption strength term for candidate combinations Electronic response items Recovery ability item Stable items and environmental adaptation items Make a judgment and select the adsorption strength item. Electronic response items Recovery ability item Stable items and environmental adaptation items Candidate combinations that meet all the set requirements are entered into the bill of materials output module.

[0030] This invention uses a DFT verification module to verify candidate combinations that meet the adsorption energy requirements, and finally selects combinations that simultaneously meet the adsorption strength requirements. Electronic response items Recovery ability item Stable items and environmental adaptation items High-performance candidate gas-sensitive materials that meet multiple performance requirements.

[0031] Preferably, the gradient boosting regression model is trained using the following steps:

[0032] Step S1: The target gas determination module determines the fault mode of the offshore wind farm transformer based on the operating characteristics of the transformer, and determines the corresponding fault marker gas.

[0033] Step S2: The candidate material library construction module constructs a candidate gas-sensitive material space based on the fault marker gas to obtain a candidate adsorption system. The candidate adsorption system contains I candidate combinations consisting of "substrate-doped atoms-fault gas".

[0034] Step S3a: The DFT verification module performs DFT calculations on each of the I candidate combinations to obtain the adsorption energy vector of the candidate combinations;

[0035] Step S3b: The descriptor generation module constructs an initial descriptor from three dimensions: substrate, doped atoms, and fault gas, and then performs feature filtering on the initial descriptor set to obtain a descriptor set.

[0036] Step S4: The high-throughput prediction module takes the descriptor set as input and the adsorption energy as output to construct a gradient boosting regression model. The descriptor set and the adsorption energy vector are divided into a training set and an independent test set. Then, the gradient boosting regression model is trained and its parameters are tuned on the training set to finally obtain a gradient boosting regression model for adsorption energy prediction.

[0037] This invention takes the fault diagnosis of transformers in offshore wind farms as its application scenario, and... , , , Using fault marker gases as the detection target and transition metal-doped two-dimensional materials as candidate gas-sensitive materials, representative adsorption performance data were obtained through DFT calculation. A rapid prediction model for gas adsorption energy was established using the gradient boosting regression (GBR) algorithm. Candidate materials were then screened based on indicators such as adsorption energy, charge transfer, band structure change, recovery ability, and stability to obtain the optimal combination of gas-sensitive materials for fault marker gas detection. The selected gas-sensitive materials were then used as the materials for fabricating gas sensors in the gas detection module of the corresponding offshore wind farm transformer fault diagnosis system, thereby improving the fault detection accuracy of the offshore wind farm transformer fault diagnosis system.

[0038] Preferably, in step S3b, the descriptor generation module filters the descriptors in the initial descriptor set through the following steps:

[0039] Step 1: Sample creation: Construct a sample set by combining the initial descriptor set and the adsorption energy vector, where each sample contains a candidate combination of initial descriptor vectors and its corresponding adsorption energy;

[0040] Step 2: Calculate the Spearman rank correlation coefficient between each descriptor and adsorption energy in the sample set, expressed as:

[0041] ;

[0042] in, The Spearman rank correlation coefficient. For the first The rank difference of each sample in the descriptor and adsorption energy ranking The total number of samples, ;

[0043] Select descriptors with Spearman rank correlation coefficients greater than the set threshold to proceed to the next step;

[0044] Step 3: Calculate the Pearson correlation coefficient between any two descriptors, expressed as:

[0045] ;

[0046] in, For the first The descriptor and the first Pearson correlation coefficient between descriptors , For the i-th sample, the first... One descriptor, For the i-th sample, the first... One descriptor, For the first of all samples The mean of the descriptors, For the first of all samples The mean of the descriptors;

[0047] when When the value is less than the set threshold, both descriptors are retained;

[0048] when Greater than the set threshold, such as At that time, only the descriptor with the larger Spearman rank correlation coefficient among the two descriptors is retained;

[0049] Step 4: Perform Min-Max normalization on the final set of retained descriptors, using the following expression:

[0050] ;

[0051] in, Let c be the normalized descriptor of the i-th sample. The minimum value of the c-th descriptor. This represents the maximum value of the c-th descriptor. The test set and the samples to be predicted are scaled using the same methods as the training set statistics.

[0052] Through the above design, by calculating the Spearman correlation coefficient between each descriptor and the adsorption energy, descriptors with effective correlation to the target variable, i.e., the adsorption energy, are retained. By calculating the Pearson correlation coefficient between pairwise descriptors, highly redundant features can be effectively removed, prioritizing the retention of descriptors with stronger correlation to the adsorption energy and clearer physical meaning. This invention, by screening descriptors, effectively retains features strongly correlated with the target variable, removes highly redundant features, improves the quality of input features for the prediction model, significantly reduces the computational load of the prediction model, improves model computational efficiency, and reduces system computational costs.

[0053] As a preferred option: in the initial descriptor set, the descriptors of the substrate include, but are not limited to, substrate composition, lattice constant, band gap, surface energy, and work function; the descriptors of the doped atoms include, but are not limited to, atomic radius, atomic mass, number of valence electrons, electronegativity, ionization energy, d orbital energy level, density, melting point, and cohesive energy; and the descriptors of the fault gas include, but are not limited to, molecular radius, molecular mass, number of atoms, molecular energy, molecular volume, bond index, and topological / local structure.

[0054] The final set of descriptors obtained through screening includes, but is not limited to, the melting point, d orbital energy level, density, and cohesive energy of doped atoms, and the molecular mass, molecular configuration, molecular radius, number of atoms, and local structure of fault gases.

[0055] The beneficial effects of this invention are as follows: By constructing a closed-loop process of "offshore wind power fault scenario determination—fault marker gas determination—candidate adsorption system construction—physicochemical descriptor conversion—adsorption energy model prediction—DFT multi-index verification—material list output," this invention achieves high-throughput prediction, rapid screening, and interpretable recommendation of gas-sensitive materials for fault gas detection in offshore wind farm transformers, providing an effective basis for the selection of sensor materials for online monitoring of offshore wind farm transformers. It significantly shortens material screening time, reduces system computational costs, and ensures the accuracy and effectiveness of material screening while improving material screening efficiency. Attached Figure Description

[0056] Figure 1 This is a flowchart of the method of the present invention;

[0057] Figure 2 This is a block diagram of the gas-sensitive material screening system in the embodiment;

[0058] Figure 3 This is a schematic diagram illustrating the application scenario of gas detection for transformer fault indicators in offshore wind farms, as shown in the example.

[0059] Figure 4 This is a flowchart of the gas-sensitive material discovery process based on the GBR algorithm and material calculation in the example.

[0060] Figure 5 This is a schematic diagram of the spatial construction of candidate gas-sensitive materials in the embodiments;

[0061] Figure 6 This is a diagram illustrating the performance comparison between the GBR prediction model and other machine learning models in this embodiment.

[0062] Figure 7 This is a schematic diagram showing the results of multi-index screening and ranking of candidate gas-sensitive materials in the embodiment. Detailed Implementation

[0063] The present invention will be further described in detail below with reference to the accompanying drawings and specific examples. The following embodiments or drawings are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0064] like Figure 1 As shown, a method for screening gas-sensitive materials for detecting fault marker gases in offshore wind farm transformers includes the following steps:

[0065] Step 1: Construct a gas-sensitive material screening system for fault marker gas detection in offshore wind farm transformers based on the GBR algorithm and material calculation. This system includes a target gas determination module, a candidate material library construction module, a descriptor generation module, a high-throughput prediction module, a DFT verification module, and a material list output module, all connected in sequence. Figure 2 As shown;

[0066] Step 2: The target gas determination module determines the fault modes that are easily induced in offshore wind farm transformers based on the operating characteristics of the transformers, and determines the corresponding fault marker gases.

[0067] Step 3: The candidate material library construction module constructs a candidate gas-sensitive material space based on the fault marker gas to obtain a candidate adsorption system. The candidate adsorption system contains one candidate combination consisting of "substrate-doped atoms-fault gas".

[0068] Step 4: The descriptor generation module constructs descriptors for each candidate combination from three dimensions: substrate, doped atoms, and fault gas, thereby constructing a descriptor set;

[0069] Step 5: The high-throughput prediction module predicts the adsorption energy of each candidate combination based on the descriptor set, and then outputs candidate combinations whose adsorption energy values ​​are within the set adsorption energy range and cover various target fault gases to the DFT verification module.

[0070] Step 6: The DFT verification module performs DFT verification on the candidate combinations selected in Step 5. When the adsorption strength term of the candidate combination... Electronic response items Recovery ability item Stable items and environmental adaptation items If all the set requirements are met, the verification is successful, and then the successful candidate combination is passed to the bill of materials output module;

[0071] Step 7: The material list output module scores the qualified candidate combinations using a multi-index scoring function, and then outputs the candidate material list from high to low scores. The substrate-doped atomic material corresponding to the candidate combination in the candidate material list is the gas-sensitive material for detecting fault marker gas in offshore wind farm transformers that was finally discovered.

[0072] Application scenarios of gas detection for transformer fault signs in offshore wind farms, such as Figure 3 As shown, Figure 3 The gas-sensitive detection material used in the gas detection module is one or more combinations of the candidate gas-sensitive materials finally screened by this invention.

[0073] In step 2, the target gas determination module determines the fault marker gas based on the operating characteristics of the offshore wind farm transformer as follows:

[0074] When the transformer in an offshore wind farm is an oil-immersed transformer or a box-type transformer, its operating characteristics include, but are not limited to, load fluctuations, compact space, and limited heat dissipation. Fault modes include, but are not limited to, low-temperature overheating, high-temperature overheating, and insulation aging. Fault indicator gases include, but are not limited to, those mentioned above. , , ;

[0075] When the transformer in an offshore wind farm serves as the main transformer for an offshore substation, its operating characteristics include, but are not limited to, large capacity, high voltage, and complex oil-paper insulation. Fault modes include, but are not limited to, partial discharge, arc discharge, and high-temperature overheating. Fault indicator gases include, but are not limited to, [missing information]. , , ;

[0076] When the marine environment is characterized by one or more of the following combinations: high humidity, high salt spray, strong wind vibration, and temperature difference cycling, the failure modes of offshore wind farm transformers include, but are not limited to, insulation damping, local defects, and increased contact resistance. Fault indicator gases include, but are not limited to, those mentioned above. , , , .

[0077] In step 3, a two-dimensional transition metal chalcogenide and its Janus or alloyed derivatives are selected as the substrate, including but not limited to tungsten disulfide. Molybdenum diselenide tungsten diselenide and tungsten, sulfur, selenium ;

[0078] Transition metals are selected as doping atoms, including but not limited to scandium (Sc), titanium (Ti), vanadium (V), chromium (Cr), manganese (Mn), iron (Fe), cobalt (Co), nickel (Ni), copper (Cu), zinc (Zn), yttrium (Y), zirconium (Zr), niobium (Nb), molybdenum (Mo), technetium (Tc), ruthenium (Ru), rhodium (Rh), palladium (Pd), silver (Ag), cadmium (Cd), hafnium (Hf), tantalum (Ta), tungsten (W), rhenium (Re), osmium (Os), iridium (Ir), platinum (Pt), and gold (Au).

[0079] In step 5, the high-throughput prediction module uses a gradient boosting regression model to predict the adsorption energy of the candidate combinations, expressed as:

[0080] ;

[0081] in, Let be the predicted adsorption energy value for the i-th candidate combination. For gradient boosting regression models, For the m-th regression tree, For the total number of regression trees, Let be the model parameters for the m-th regression tree. Let be the descriptor vector for the i-th candidate combination.

[0082] In step 6, the DFT verification module performs DFT verification on the candidate combinations, and the specific process is as follows:

[0083] Step 6.1: The DFT verification module establishes corresponding clean substrate, doped substrate, isolated gas and adsorption system models according to the selected candidate combinations, and performs geometric optimization.

[0084] Step 6.2: Perform DFT calculations on each candidate combination to obtain the corresponding candidate combination parameters, including but not limited to the adsorption energy of the candidate combination. Adsorption distance Charge transfer amount band gap band gap variation Change in density of states Change in work function Theoretical sensitivity and recovery time ;

[0085] Step 6.3: Calculate the adsorption strength term based on the candidate combination parameters. Electronic response items Recovery ability item Stable items and environmental adaptation items ;

[0086] Step 6.4: Adsorption strength term for candidate combinations Electronic response items Recovery ability item Stable items and environmental adaptation items Make a judgment and select the adsorption strength item. Electronic response items Recovery ability item Stable items and environmental adaptation items Candidate combinations that meet all the set requirements are entered into the bill of materials output module.

[0087] The optimal screening criteria are: the adsorption energy is neither too weak nor too strong, to avoid insufficient physical adsorption response or irreversible chemical adsorption; the amount of interfacial charge transfer, band gap or work function changes significantly after gas adsorption, which can be converted into a measurable electrical signal; the recovery time is within an acceptable range to ensure the sensor can be reused; the structure remains intact after AIMD simulation, and the doped atoms do not show obvious desorption or aggregation.

[0088] According to this rule, such as Figure 7 As shown, in this embodiment, systems such as Rh-WSSe that exhibit a balance between adsorption strength and recovery ability can be used as preferred candidate materials, while Fe-WSSe, Cu-WSSe, etc. can be used as alternative materials for different fault gas response scenarios.

[0089] In step 6.2, the adsorption energy The calculation expression is:

[0090] ;

[0091] in, This represents the total energy of the gas adsorbed on the surface of the doped atom-substrate. To clean the total energy of the doped atom-substrate, Energy of isolated gas molecules; The smaller the value, the stronger the adsorption effect;

[0092] Adsorption distance The calculation expression is:

[0093] ;

[0094] in, Let represent the coordinates of the a-th atom in the gas molecule. Represents the coordinates of the b-th atom in the substrate or active site of the doped atom;

[0095] Charge transfer amount The calculation expression is:

[0096] ;

[0097] in, Let d be the charge of the gas atom d in the adsorption system. The charge of the corresponding gas atom d in the isolated gas; The larger the absolute value, the more obvious the interfacial electron exchange.

[0098] band gap The calculation expression is:

[0099] ;

[0100] in, For the bottom energy level of the conduction band, This is the top energy level of the valence band;

[0101] Change in band gap before and after gas adsorption The calculation expression is:

[0102] ;

[0103] in, The band gap is the energy gap after gas adsorption. The band gap before gas adsorption;

[0104] Density of states and density of states change The calculation expression is:

[0105] ;

[0106] ;

[0107] in, For density of states, As an energy variable, it is usually expressed relative to the Fermi level EF with zero point; δ is the Dirac delta function, used to characterize the distribution of electronic energy levels along the energy axis; Let K be the energy of the nth energy band at point K. To be able to include serial numbers, Let K be the point. The change in density of states For orbital projection coefficients, The projected atomic orbitals or types of atoms;

[0108] Work function and the change in work function The calculation expression is:

[0109] ;

[0110] ;

[0111] in, For vacuum energy level, It is the Fermi level. Let be the work function of the system after gas adsorption. Let be the work function of cleaning the doped atom-substrate before gas adsorption. This represents the change in work function before and after adsorption; the work function reflects the minimum energy required for an electron to escape from the Fermi level of the material to the vacuum level. The larger the absolute value, the more significant the modulation of the material surface electron escape capability and interfacial potential by gas adsorption.

[0112] Theoretical sensitivity The calculation expression is:

[0113] ;

[0114] in, Represents Boltzmann's constant. Indicates the operating temperature. Represents an exponential function;

[0115] Recovery time The calculation expression is:

[0116] ;

[0117] in, Indicates the frequency of attempts; recovery time. The smaller the value, the easier the material is to desorb and recover;

[0118] In step 6.3, the adsorption strength term The calculation expression is:

[0119] ;

[0120] in, For the truncation function, For reference adsorption energy, This is the i-th candidate combination;

[0121] Electronic response item The calculation expression is:

[0122] ;

[0123] in, The weighting coefficient for charge transfer. The weighting coefficient for the bandgap change. The weighting coefficients for the change in the work function. The weighting coefficients for theoretical sensitivity. This is the normalization function;

[0124] Recovery ability item The calculation expression is:

[0125] ;

[0126] in, For reference recovery time;

[0127] Stable terms The expression for calculating is:

[0128] ;

[0129] in, For normalization function, The weighting factor is the doping binding energy. The weighting coefficients for AIMD energy fluctuations. This represents the weighting coefficient for the root mean square displacement of the structure. For doping binding energy, For AIMD energy fluctuations, This represents the root mean square displacement of the structure.

[0130] Doping binding energy The total energy is calculated using DFT. For substitutional doping, models of the doped substrate, the substrate containing vacancies, and isolated doped atoms can be established and optimized separately, with the following expressions:

[0131] ;

[0132] in, The total energy of the system after doping. To remove the substrate energy after replacing the atoms, Energy for isolated doped atoms;

[0133] AIMD energy fluctuations This was obtained through first-principles molecular dynamics simulations. Specifically, AIMD simulations were performed at the target operating temperature or extreme temperatures, and the total energy E(t) at each time step during the simulation was recorded. The expression is:

[0134] ;

[0135] Energy standard deviation can also be used as an auxiliary evaluation index. The smaller the value, the more stable the material's energy is under thermal disturbance.

[0136] Root mean square displacement of structure It can be calculated from the changes in atomic coordinates before and after AIMD simulation, and the expression is:

[0137] ;

[0138] in, The total number of atoms in the system. Let be the coordinates of the u-th atom at the initial time. Let be the coordinates of the i-th atom at time t; The smaller the value, the more stable the material structure is during the AIMD simulation.

[0139] Environmental adaptation items satisfy: Environmental adaptation items This can be obtained through an evaluation of the operating conditions of offshore wind farms, with a value range of 0 ≤ ≤1. This item can be obtained by weighting sub-indicators such as resistance to damp heat, resistance to salt spray corrosion, temperature cycling stability, vibration stability, chemical compatibility in oil, and feasibility of device fabrication.

[0140] In step 7, the calculation expression for the multi-index scoring function is:

[0141] ;

[0142] in, The score for the i-th candidate combination is... , , , , These are the weight parameters for each item.

[0143] according to Output a list of candidate materials from high to low; when a material simultaneously meets the conditions of moderate adsorption energy, significant charge transfer or band gap change, short recovery time, stable doping sites and stable AIMD structure, it is determined to be suitable as a candidate gas-sensitive material for fault marker gas detection in offshore wind farm transformers.

[0144] The gradient boosting regression model is trained through the following steps:

[0145] Step S1: The target gas determination module determines the fault mode of the offshore wind farm transformer based on the operating characteristics of the transformer, and determines the corresponding fault marker gas.

[0146] Step S2: The candidate material library construction module constructs a candidate gas-sensitive material space based on the fault marker gas, obtaining a candidate adsorption system. This candidate adsorption system contains I candidate combinations consisting of "substrate-doped atoms-fault gas", such as... Figure 5 As shown;

[0147] Step S3a: The DFT verification module performs DFT calculations on each of the I candidate combinations to obtain the adsorption energy vector of the candidate combinations;

[0148] Step S3b: The descriptor generation module constructs an initial descriptor from three dimensions: substrate, doped atoms, and fault gas, and then performs feature filtering on the initial descriptor set to obtain a descriptor set.

[0149] Step S4: The high-throughput prediction module takes the descriptor set as input and the adsorption energy as output to construct a gradient boosting regression model. The descriptor set and the adsorption energy vector are divided into a training set and an independent test set according to a preset ratio, preferably an 8:2 ratio. Then, five-fold cross-validation is used in the training set to train and tune the gradient boosting regression model.

[0150] This paper compares the Gradient Boosting Regression (GBR) algorithm with Support Vector Regression (SVR), Random Forest Regression (RF), Extremely Random Tree Regression (ExtraTree), Lightweight Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting Tree (XGBoost), Category Feature Gradient Boosting (CatBoost), and Multilayer Perceptron Regression (MLP) algorithms to determine the coefficients of determination. The mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE) were used as evaluation indicators. The GBR model with the best overall prediction performance was selected as the core prediction model, and finally, a gradient boosting regression model for adsorption energy prediction was obtained.

[0151] The comparison results of the Gradient Boosting Regression (GBR) algorithm with other comparative algorithms are shown in Table 1 and... Figure 6 As shown.

[0152] Table 1. Comparison of the method of the present invention, comparative methods, and regression prediction performance.

[0153]

[0154] The gradient boosting regression model expression is as follows:

[0155] ;

[0156] ;

[0157] This represents the initial prediction function of the gradient boosting regression model, which is usually taken as the average value of the adsorption energy labels in the training set under the squared loss function. This represents the initial constant prediction value that minimizes the sum of the loss functions for all training samples. This represents the ensemble prediction function after the m-th round of improvements; This represents the ensemble prediction function before the m-th update. This indicates increasing the number of iterations or the regression tree index. A descriptor vector representing candidate combinations;

[0158] The negative gradient residual in the m-th round is:

[0159] ;

[0160] The m-th regression tree is obtained by fitting the negative gradient residual:

[0161] ;

[0162] in, This represents the candidate regression tree function to be selected;

[0163] The step size factor is:

[0164] ;

[0165] in, I is the descriptor vector of the i-th sample. Let be the adsorption energy of the i-th sample calculated by DFT, where i is the i-th candidate combination sample. For the loss function, For the m-th regression tree, This is the step size coefficient. This represents the variable to be determined during the optimization process. For learning rate, The total number of regression trees;

[0166] Under squared loss, negative gradient residual Equivalent to:

[0167] ;

[0168] Therefore, the expression for predicting adsorption energy is:

[0169] ;

[0170] Its simplified expression is: That is, the initial model Learning rate and step size coefficient Functions absorbed into each tree or parameter middle.

[0171] The five-fold cross-validation process is as follows: Divide the training set into five equal subsets; select four subsets to train the model and one subset to validate the model each time; repeat the training for the candidate hyperparameters and calculate the validation error. The error of the k-th fold cross-validation is denoted as:

[0172] ;

[0173] by As the target for parameter tuning, select:

[0174] ;

[0175] ;

[0176] in, This represents the root mean square error calculated on the k-th fold validation set. This represents the combination of hyperparameters to be optimized. This represents the number of samples in the k-th fold validation set. This represents the true DFT adsorption energy of the i-th validation sample. This represents the model-predicted adsorption energy of the i-th validation sample. This represents the average error of the five-fold cross-validation. Indicates to make The smallest optimal combination of hyperparameters. This represents the total number of regression trees. Indicates the learning rate. This represents the maximum depth of a single regression tree. This represents the minimum number of samples for a leaf node. This indicates the sampling ratio of samples used in each round of training. Indicates the type of loss function.

[0177] Finally use The GBR model is retrained on the complete training set, and its generalization performance is evaluated using an independent test set. The internal parameter update of GBR is the process of residual fitting and model addition update described above. The hyperparameters are determined within the cross-validation framework through grid search, random search, or Bayesian optimization.

[0178] Mechanism Explanation and Secondary Validation: In this embodiment, feature importance analysis, permutation importance, or SHAP analysis are used to determine the key descriptors affecting the adsorption energy, and the top-ranked candidate materials are selected for DFT secondary validation, including geometry optimization, adsorption energy calculation, Mulliken / Hirshfeld charge analysis, band structure, TDOS, and AIMD stability verification.

[0179] Feature importance analysis: Feature importance analysis is obtained by cumulatively calculating the loss reduction caused by splitting nodes in the tree model.

[0180] ;

[0181] in, For the tree model feature importance of the j-th descriptor, Let represent the set of nodes split using the j-th feature. The proportion of samples reaching node t. This represents the decrease in loss before and after node t splits. The larger the value, the higher the contribution of the descriptor to the model's prediction.

[0182] Permutation importance analysis: Permutation importance is obtained by randomly shuffling the j-th feature and observing the changes in model error.

[0183] ;

[0184] in, For the importance of the permutation of the j-th descriptor, This represents the error of the original model. The model error is the result of randomly shuffling the j-th feature. The larger the value, the more important this feature is for predicting adsorption energy.

[0185] SHAP analysis: SHAP analysis decomposes the predicted value of a single sample into:

[0186] ;

[0187] in, Let be the predicted value for the i-th sample. As the model baseline value, Let the contribution of the j-th descriptor to the predicted value of the i-th sample be denoted as . This represents the number of descriptors.

[0188] Global SHAP importance of the j-th descriptor for:

[0189] ;

[0190] in, This represents the number of samples. The larger the value, the more significant the impact of the descriptor on the overall prediction results of the model.

[0191] Comprehensive importance calculation and key descriptor determination:

[0192] To synthesize the results of different interpretation methods, the three types of importance indicators were normalized and weighted and fused:

[0193]

[0194] in, For the overall importance of the j-th descriptor, Normalized function from 0 to 1 , , Let be the weighting coefficients, and satisfy:

[0195]

[0196]

[0197] The steps for determining key descriptors based on the above analysis are as follows: First, train the optimal GBR model; second, calculate the tree model feature importance, permutation importance, and mean absolute SHAP value for each descriptor; third, normalize each importance index and calculate the comprehensive importance (steps one through four); finally, select the descriptors with the highest comprehensive importance ranking and clear physical meaning as key descriptors.

[0198] In step S3b, the descriptor generation module filters the descriptors in the initial descriptor set through the following steps:

[0199] Step 1: Sample creation: Construct a sample set by combining the initial descriptor set and the adsorption energy vector, where each sample contains a candidate combination of initial descriptor vectors and its corresponding adsorption energy;

[0200] Step 2: Calculate the Spearman rank correlation coefficient between each descriptor and adsorption energy in the sample set, expressed as:

[0201] ;

[0202] in, The Spearman rank correlation coefficient. For the first The rank difference of each sample in the descriptor and adsorption energy ranking The total number of samples, ;

[0203] Descriptors with Spearman rank correlation coefficients greater than a set threshold are selected for the next step; descriptors with effective correlation to the target variable are retained; when there are unequal ranks, the Pearson correlation form of the ordinal variable is used for calculation.

[0204] When there are congruent ranks, first take the value of the original descriptor. and adsorption energy Convert them to average rank variables respectively and The average of the ranks of the parallel samples is taken, and then the Spearman correlation coefficient is calculated using the Pearson correlation form, expressed as:

[0205] ;

[0206] in, This represents the average rank of the i-th sample on the c-th descriptor. The average rank of the adsorption energy of the i-th sample is represented. These are the means of the corresponding rank variables.

[0207] Step 3: Calculate the Pearson correlation coefficient between any two descriptors, expressed as:

[0208] ;

[0209] in, For the first The descriptor and the first Pearson correlation coefficient between descriptors , For the i-th sample, the first... One descriptor, For the i-th sample, the first... One descriptor, For the first of all samples The mean of the descriptors, For the first of all samples The mean of the descriptors;

[0210] when When the value is less than the set threshold, both descriptors are retained;

[0211] when Greater than the set threshold, such as At that time, only the descriptor with the larger Spearman rank correlation coefficient among the two descriptors is retained;

[0212] For highly redundant features with |r|>0.85, descriptors that are more strongly correlated with adsorption energy and have clearer physical meaning are preferred to be retained.

[0213] Step 4: Perform Min-Max normalization on the final set of retained descriptors, using the following expression:

[0214] ;

[0215] in, Let c be the normalized descriptor of the i-th sample. The minimum value of the c-th descriptor. This represents the maximum value of the c-th descriptor. The test set and the samples to be predicted are scaled using the same methods as the training set statistics.

[0216] In the initial descriptor set, the descriptors for the substrate include, but are not limited to, substrate composition, lattice constant, band gap, surface energy, and work function; the descriptors for doped atoms include, but are not limited to, atomic radius, atomic mass, number of valence electrons, electronegativity, ionization energy, d orbital energy level, density, melting point, and cohesive energy; and the descriptors for fault gases include, but are not limited to, molecular radius, molecular mass, number of atoms, molecular energy, molecular volume, bond index, and topological / local structure.

[0217] The final set of descriptors includes, but is not limited to, the melting point, d-orbital energy level, density, and cohesive energy of the doped atoms, and the molecular mass, molecular configuration, molecular radius, number of atoms, and local structure of the fault gas. Among these, the characteristics related to the doped metal contribute significantly to the prediction of adsorption energy, indicating that the properties of the doped atoms are a key factor in regulating gas adsorption behavior.

[0218] The beneficial effects of this invention are as follows:

[0219] (1) This invention combines the fault diagnosis needs of offshore wind farm transformers with the gas-sensitive material discovery process, and can address the needs of offshore wind farm transformer fault diagnosis. , , , The method of rapidly screening candidate materials using fault indicator gases is suitable for remote online monitoring scenarios in offshore wind farms.

[0220] (2) The present invention uses the GBR model as a fast proxy model for DFT calculation, which can predict the adsorption energy of a large number of candidate systems based on limited high-confidence material calculation data, significantly reducing the time cost of DFT calculation and experimental trial and error.

[0221] (3) This invention enables the material screening results to have a clear physical interpretation through correlation screening, feature importance analysis and multi-index ranking, and can reveal the contribution of doped metal atoms, gas molecules and substrate features to adsorption behavior.

[0222] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for screening gas-sensitive materials for detecting fault marker gases in offshore wind farm transformers, characterized in that, Includes the following steps: Step 1: Construct a gas-sensitive material screening system for detecting fault marker gases in offshore wind farm transformers based on GBR algorithm and material calculation. This gas-sensitive material screening system for detecting fault marker gases in offshore wind farm transformers is set up with a target gas determination module, a candidate material library construction module, a descriptor generation module, a high-throughput prediction module, a DFT verification module, and a material list output module connected in sequence. Step 2: The target gas determination module determines the fault mode of the offshore wind farm transformer based on its operating characteristics, and obtains the corresponding fault marker gas. Step 3: The candidate material library construction module constructs a candidate gas-sensitive material space based on the fault marker gas to obtain a candidate adsorption system. The candidate adsorption system contains one candidate combination consisting of "substrate-doped atoms-fault gas". Step 4: The descriptor generation module constructs descriptors for each candidate combination from three dimensions: substrate, doped atoms, and fault gas, thereby obtaining a descriptor set; Step 5: The high-throughput prediction module predicts the adsorption energy of each candidate combination based on the descriptor set, and then outputs candidate combinations whose adsorption energy values ​​are within the set adsorption energy range and cover various target fault gases to the DFT verification module. Step 6: The DFT verification module performs DFT verification on the candidate combinations selected in Step 5. When the adsorption strength term of the candidate combination... Electronic response items Recovery ability item Stable items and environmental adaptation items If all the set requirements are met, the verification is successful, and then the successful candidate combination is passed to the bill of materials output module; Step 7: The bill of materials output module scores the qualified candidate combinations using a multi-index scoring function, and then outputs the candidate bill of materials from high to low scores.

2. The method for screening gas-sensitive materials for detecting fault marker gases in offshore wind farm transformers according to claim 1, characterized in that: In step 2, the target gas determination module determines the fault marker gas based on the operating characteristics of the offshore wind farm transformer as follows: When the transformer in an offshore wind farm is an oil-immersed transformer or a box-type transformer, its operating characteristics include, but are not limited to, load fluctuations, compact space, and limited heat dissipation. Fault modes include, but are not limited to, low-temperature overheating, high-temperature overheating, and insulation aging. Fault indicator gases include, but are not limited to, those mentioned above. , , ; When the transformer in an offshore wind farm serves as the main transformer for an offshore substation, its operating characteristics include, but are not limited to, large capacity, high voltage, and complex oil-paper insulation. Fault modes include, but are not limited to, partial discharge, arc discharge, and high-temperature overheating. Fault indicator gases include, but are not limited to, [missing information]. , , ; When the marine environment is characterized by one or more of the following combinations: high humidity, high salt spray, strong wind vibration, and temperature difference cycling, the failure modes of offshore wind farm transformers include, but are not limited to, insulation damping, local defects, and increased contact resistance. Fault indicator gases include, but are not limited to, those mentioned above. , , , .

3. The method for screening gas-sensitive materials for detecting fault marker gases in offshore wind farm transformers according to claim 1, characterized in that: In step 3, two-dimensional transition metal chalcogenides and their alloyed derivatives are selected as the substrate, including but not limited to tungsten disulfide. Molybdenum diselenide tungsten diselenide and tungsten, sulfur, selenium ; Transition metals are selected as doping atoms, including but not limited to scandium (Sc), titanium (Ti), vanadium (V), chromium (Cr), manganese (Mn), iron (Fe), cobalt (Co), nickel (Ni), copper (Cu), zinc (Zn), yttrium (Y), zirconium (Zr), niobium (Nb), molybdenum (Mo), technetium (Tc), ruthenium (Ru), rhodium (Rh), palladium (Pd), silver (Ag), cadmium (Cd), hafnium (Hf), tantalum (Ta), tungsten (W), rhenium (Re), osmium (Os), iridium (Ir), platinum (Pt), and gold (Au).

4. The method for screening gas-sensitive materials for detecting fault marker gases in offshore wind farm transformers according to claim 1, characterized in that: In step 5, the high-throughput prediction module uses a gradient boosting regression model to predict the adsorption energy of the candidate combinations, expressed as: ； in, Let be the predicted adsorption energy value for the i-th candidate combination. For gradient boosting regression models, For the m-th regression tree, For the total number of regression trees, Let be the model parameters for the m-th regression tree. Let be the descriptor vector for the i-th candidate combination.

5. The method for screening gas-sensitive materials for detecting fault marker gases in offshore wind farm transformers according to claim 4, characterized in that: The gradient boosting regression model is trained through the following steps: Step S1: The target gas determination module determines the fault mode of the offshore wind farm transformer based on the operating characteristics of the transformer, and determines the corresponding fault marker gas. Step S2: The candidate material library construction module constructs a candidate gas-sensitive material space based on the fault marker gas to obtain a candidate adsorption system. The candidate adsorption system contains I candidate combinations consisting of "substrate-doped atoms-fault gas". Step S3a: The DFT verification module performs DFT calculations on each of the I candidate combinations to obtain the adsorption energy vector of the candidate combinations; Step S3b: The descriptor generation module constructs an initial descriptor from three dimensions: substrate, doped atoms, and fault gas, and then performs feature filtering on the initial descriptor set to obtain a descriptor set. Step S4: The high-throughput prediction module takes the descriptor set as input and the adsorption energy as output to construct a gradient boosting regression model. It divides the descriptor set and the adsorption energy vector into a training set and an independent test set. Then, it trains and tunes the gradient boosting regression model in the training set to finally obtain a gradient boosting regression model for adsorption energy prediction. The gradient boosting regression model expression is as follows: ； ； This represents the initial prediction function of the gradient boosting regression model. This represents the initial constant prediction value that minimizes the sum of the loss functions for all training samples. This represents the ensemble prediction function after the m-th round of improvements; This represents the ensemble prediction function before the m-th update. This indicates increasing the number of iterations or the regression tree index. A descriptor vector representing candidate combinations; The negative gradient residual in the m-th round is: ； The m-th regression tree is obtained by fitting the negative gradient residual: ； in, This represents the candidate regression tree function to be selected; The step size factor is: ； in, I is the descriptor vector of the i-th sample. Let be the adsorption energy of the i-th sample calculated by DFT, where i is the i-th candidate combination sample. For loss function, For the m-th regression tree, This is the step size coefficient. This represents the variable to be determined during the optimization process. For learning rate, The total number of regression trees; Under squared loss, negative gradient residual Equivalent to: ； Therefore, the expression for predicting adsorption energy is: ； Its simplified expression is: .

6. The method for screening gas-sensitive materials for detecting fault marker gases in offshore wind farm transformers according to claim 1, characterized in that: In step 6, the DFT verification module performs DFT verification on the candidate combinations, and the specific process is as follows: Step 6.1: The DFT verification module establishes corresponding clean substrate, doped substrate, isolated gas and adsorption system models according to the selected candidate combinations, and performs geometric optimization. Step 6.2: Perform DFT calculations on each candidate combination to obtain the corresponding candidate combination parameters, including but not limited to the adsorption energy of the candidate combination. Adsorption distance Charge transfer amount band gap band gap variation Change in density of states Change in work function Theoretical sensitivity and recovery time ; Step 6.3: Calculate the adsorption strength term based on the candidate combination parameters. Electronic response items Recovery ability item Stable items and environmental adaptation items ; Step 6.4: Adsorption strength term for candidate combinations Electronic response items Recovery ability item Stable items and environmental adaptation items Make a judgment and select the adsorption strength item. Electronic response items Recovery ability item Stable items and environmental adaptation items Candidate combinations that meet all the set requirements are entered into the bill of materials output module.

7. The method for screening gas-sensitive materials for detecting fault marker gases in offshore wind farm transformers according to claim 6, characterized in that: In step 6.2, the adsorption energy The calculation expression is: ； in, This represents the total energy of the gas adsorbed on the surface of the doped atom-substrate. To clean the total energy of the doped atom-substrate, Energy of isolated gas molecules; Adsorption distance The calculation expression is: ； in, Let represent the coordinates of the a-th atom in the gas molecule. Represents the coordinates of the b-th atom in the substrate or active site of the doped atom; Charge transfer amount The calculation expression is: ； in, Let d be the charge of the gas atom d in the adsorption system. The charge of the corresponding gas atom d in the isolated gas; band gap The calculation expression is: ； in, For the bottom energy level of the conduction band, This is the top energy level of the valence band; Change in band gap before and after gas adsorption The calculation expression is: ； in, The band gap is the energy gap after gas adsorption. The band gap before gas adsorption; Density of states and density of states change The calculation expression is: ； ； in, For density of states, For energy variables, For Dirac delta function, Let K be the energy of the nth energy band at point K. To be able to include serial numbers, Let K be the point. The change in density of states For orbital projection coefficients, The projected atomic orbitals or types of atoms; Work function and the change in work function The calculation expression is: ； ； in, For vacuum energy level, It is the Fermi level. Let be the work function of the system after gas adsorption. Let be the work function of cleaning the doped atom-substrate before gas adsorption. This represents the change in work function before and after adsorption; Theoretical sensitivity The calculation expression is: ； in, Represents Boltzmann's constant. Indicates the operating temperature. Represents an exponential function; Recovery time The calculation expression is: ； in, Indicates the frequency of attempts; In step 6.3, the adsorption strength term The calculation expression is: ； in, For the truncation function, For reference adsorption energy, This is the i-th candidate combination; Electronic response item The calculation expression is: ； in, The weighting coefficient for charge transfer. The weighting coefficient for the bandgap change. The weighting coefficients for the change in the work function. The weighting coefficients for theoretical sensitivity. This is the normalization function; Recovery ability item The calculation expression is: ； in, For reference recovery time; Stable terms The expression for calculating is: ； in, For normalization function, The weighting factor is the doping binding energy. The weighting coefficients for AIMD energy fluctuations. This represents the weighting coefficient for the root mean square displacement of the structure. For doping binding energy, For AIMD energy fluctuations, This represents the root mean square displacement of the structure. Doping binding energy The expression is: ； in, The total energy of the system after doping. To remove the substrate energy after replacing the atoms, Energy for isolated doped atoms; AIMD energy fluctuations The expression is: ； Root mean square displacement of structure The expression is: ； in, The total number of atoms in the system. Let be the coordinates of the u-th atom at the initial time. Let be the coordinates of the i-th atom at time t; Environmental adaptation items satisfy: ; In step 7, the calculation expression for the multi-index scoring function is: ； in, The score for the i-th candidate combination is... , , , , These are the weight parameters for each item.

8. The method for screening gas-sensitive materials for detecting fault marker gases in offshore wind farm transformers according to claim 5, characterized in that: In step S3b, the descriptor generation module filters the descriptors in the initial descriptor set through the following steps: Step 1: Sample creation: Construct a sample set by combining the initial descriptor set and the adsorption energy vector, where each sample contains a candidate combination of initial descriptor vectors and its corresponding adsorption energy; Step 2: Calculate the Spearman rank correlation coefficient between each descriptor and adsorption energy in the sample set, expressed as: ； in, The Spearman rank correlation coefficient. For the first The rank difference of each sample in the descriptor and adsorption energy ranking The total number of samples, ; Select descriptors with Spearman rank correlation coefficients greater than the set threshold to proceed to the next step; Step 3: Calculate the Pearson correlation coefficient between any two descriptors, expressed as: ； in, For the first The descriptor and the first Pearson correlation coefficient between descriptors , For the i-th sample, the first... One descriptor, For the i-th sample, the first... One descriptor, For the first of all samples The mean of the descriptors, For the first of all samples The mean of the descriptors; when When the value is less than the set threshold, both descriptors are retained; when Greater than the set threshold, such as At that time, only the descriptor with the larger Spearman rank correlation coefficient among the two descriptors is retained; Step 4: Perform Min-Max normalization on the final set of retained descriptors, using the following expression: ； in, Let c be the normalized descriptor of the i-th sample. Let c be the minimum value of the c-th descriptor. This represents the maximum value of the c-th descriptor.

9. The method for screening gas-sensitive materials for detecting fault marker gases in offshore wind farm transformers according to claim 5 or 8, characterized in that: In the initial descriptor set, the descriptors for the substrate include, but are not limited to, substrate composition, lattice constant, band gap, surface energy, and work function; the descriptors for doped atoms include, but are not limited to, atomic radius, atomic mass, number of valence electrons, electronegativity, ionization energy, d orbital energy level, density, melting point, and cohesive energy; and the descriptors for fault gases include, but are not limited to, molecular radius, molecular mass, number of atoms, molecular energy, molecular volume, bond index, and topological / local structure. The final set of descriptors obtained through screening includes, but is not limited to, the melting point, d orbital energy level, density, and cohesive energy of doped atoms, and the molecular mass, molecular configuration, molecular radius, number of atoms, and local structure of fault gases.

Images