Power transformer dissociated gas detection method and system based on material screening by machine learning

Abstract

This invention provides a method and system for detecting ionized gases in power transformers based on machine learning-based material screening. The method includes: sampling the compound space; calculating a material property dataset of the adsorption performance of different doping configurations for target gases using high-throughput first-principles calculations; obtaining a pre-trained material screening model based on the material property dataset by fusing CGCNN convolutional kernels with a GNN framework, used to predict the adsorption performance of different doping configurations for target gases; using the pre-trained material screening model to screen for the optimal doping configuration that meets a preset performance threshold; obtaining the optimal gas-sensitive material based on the optimal doping configuration; acquiring the response signal of a sensor array to dissolved gases released from transformer oil, identifying the type of target gas, and estimating the corresponding gas concentration; wherein the sensor array is prepared using the optimal gas-sensitive material. This invention enables rapid, accurate, and online detection of multiple key ionized gases in transformer oil.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and fault diagnosis technology, specifically to a method and system for detecting ionized gas in power transformers based on machine learning for material sieving. Background Technology

[0002] Power transformers are core equipment in the power grid, and their operating status directly affects the grid's safety and stability. Dissolved free gases in transformer oil (such as H2, CH4, C2H2, C2H4, C2H6, CO, CO2, etc.) are important characteristic quantities reflecting latent faults within the transformer. Traditional gas detection methods mainly rely on offline gas chromatography (GC), which, while highly accurate, suffers from drawbacks such as long detection cycles, expensive equipment, complex operation, and inability to achieve online monitoring.

[0003] In recent years, online monitoring technology based on gas sensors has received widespread attention due to its advantages such as real-time performance and convenience. However, existing sensor technologies face the following key challenges:

[0004] 1. Poor gas selectivity: Transformer oil contains a variety of dissolved gases with low concentrations that interfere with each other, making it difficult for a single sensor to accurately distinguish specific gases.

[0005] 2. Insufficient material sensitivity: For low concentrations (ppm level) of fault characteristic gases (especially early faults), the response signals of existing sensitive materials are weak, and the detection limits are difficult to meet the requirements.

[0006] 3. Poor environmental adaptability: Transformers operate in complex environments (temperature and humidity changes, interference from complex components in the oil), making sensors prone to drift and exhibiting poor stability. For example, in extremely cold regions (<-20℃), the activity of conventional sensor materials decreases, leading to a sharp drop in detection sensitivity.

[0007] 4. Lack of efficient material screening methods: The development of new high-performance sensitive materials is time-consuming and costly, and traditional trial-and-error methods are inefficient.

[0008] Therefore, there is an urgent need for an online detection method for ionized gases in power transformers that can efficiently screen high-performance gas-sensitive materials and achieve high sensitivity, high selectivity, and high stability. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a method and system for detecting ionized gases in power transformers based on machine learning material screening. The method utilizes machine learning models to efficiently screen gas-sensitive materials with excellent sensitivity and selectivity to specific transformer fault gases. A sensor array is then constructed based on the screened materials. Combined with optimized detection strategies and data analysis algorithms, this enables rapid, accurate, and online detection of various key ionized gases in transformer oil, solving the three major pain points of traditional sensors in terms of accuracy, lifespan, and environmental adaptability.

[0010] The present invention adopts the following technical solution.

[0011] According to a first aspect of the present invention, a method for detecting ionized gas in power transformers based on machine learning for material screening is provided. The method includes the following steps:

[0012] Different doping configurations of candidate gas-sensitive materials are obtained by sampling the compound space, and material property datasets of the adsorption performance of the target gas by the different doping configurations are calculated by high-throughput first-principles calculation.

[0013] Based on the aforementioned material property dataset, a pre-trained material screening model is obtained by fusing CGCNN convolutional kernels through the GNN framework, which is used to predict the adsorption performance of different doping configurations for target gases.

[0014] In the compound space, the optimal doping configuration that meets the preset performance threshold is obtained by using a Bayesian optimization framework, with Gaussian process regression as the surrogate model and expected improvement as the acquisition function, and by using a pre-trained material screening model.

[0015] The response performance of sensor arrays made from gas-sensitive materials with the optimal doping configuration but different doping concentrations was tested, and the optimal gas-sensitive material was selected.

[0016] The sensor array is used to obtain the response signal of dissolved gas released from transformer oil. By combining pattern recognition and regression algorithms, the type of target gas is identified and the corresponding gas concentration is estimated. The sensor array is prepared using the optimal gas-sensitive material.

[0017] Furthermore, the candidate gas-sensitive material is metal-doped enhanced molybdenum diselenide.

[0018] Furthermore, the material property dataset includes multiple parameters such as bond length, adsorption energy, density of states, Malicken charge population, deformation charge density, and band structure.

[0019] Further, based on the material property dataset, a material screening model is trained; the material screening model is used to evaluate the adsorption performance of different doping configurations for the target gas, including:

[0020] Based on the aforementioned material property dataset, extract the material property vector corresponding to each doping configuration. xi And transform it into a graph structure Gi , i The doping configuration number is used;

[0021] use xi Calculate the overall adsorption performance score for each doping configuration. Si Combined with the aforementioned graph structure Gi Construct a graph structure set {( Gi , Si )};

[0022] Using the graph structure set {( Gi , Si Using these as training samples, a neural network with CGCNN convolutional kernels fused through the GNN framework is trained to obtain a pre-trained material screening model.

[0023] Furthermore, the comprehensive adsorption performance score Si Calculated using the following formula:

[0024]

[0025] in, For the first i Adsorption performance rating of different doped configurations. The comprehensive energy loss function characterizes the amount of charge energy transferred between gas molecules and the material substrate after gas adsorption. The energy of the adsorption state. The difference in the work function, This is the activation energy barrier for the desorption of gas molecules from the material surface. , , , They are respectively , , , The corresponding weighting coefficients.

[0026] Furthermore, the comprehensive energy loss function Represented as:

[0027]

[0028]

[0029]

[0030] in, and These are atomic energy prediction functions. and atomic force prediction function The normalized atomic energy prediction function and the normalized atomic force prediction function obtained by dividing by the corresponding scaling factor are dimensionless quantities; These are the weighting coefficients; The number of doped configurations; The model predicts the first i Total energy of each doped configuration; The first DFT calculated i Standard energy for each doped configuration; This is the set of three-dimensional coordinates of all atoms in all doped configurations; To predict energy right The gradient, in physical terms, represents the force on the atom predicted by the model. ; The true values ​​of atomic forces obtained through DFT calculation, shape and same; For hyperparameters; This represents the total number of atoms contained in all doped configurations in the training batch; For the first i The number of atoms in each doped configuration; The first one directly predicted by the model i In the doping configuration, the first j The force vector of each atom; For the first i In the doping configuration, the first j DFT reference force vector of each atom; This means multiplying the force vectors of all atoms by the same random rotation matrix; These are the weighting coefficients for the rotational isovariability constraint term.

[0031] Furthermore, in the compound space, using a Bayesian optimization framework, with Gaussian process regression as a surrogate model and expected improvement as the acquisition function, a pre-trained material screening model is used to screen for the optimal doping configuration that meets a preset performance threshold, including:

[0032] Screening comprehensive adsorption performance score Si Doping configurations exceeding a preset threshold are designated as high-potential subsets, and their optimal adsorption performance scores are assigned. A labeled dataset is constructed using the material property vectors of the high-potential subsets and their corresponding comprehensive adsorption performance scores.

[0033] Based on the labeled dataset, a well-trained Gaussian process surrogate model is obtained by constructing a kernel function and maximizing marginal likelihood.

[0034] Based on the current optimal adsorption performance score and the trained Gaussian process surrogate model, the desired improvement function is obtained;

[0035] Maximizing the desired improvement function in the compound space yields the optimal candidate doping configuration.

[0036] Verify whether the overall adsorption performance score of the best candidate doping configuration exceeds the preset threshold, and if it exceeds the preset threshold, add the best candidate doping configuration to the high-potential subset, and update the labeled dataset at the same time;

[0037] Based on the updated labeled dataset, the Gaussian process proxy model, the desired improvement function, and the optimal candidate doping configuration are iteratively updated until a preset iteration stopping condition is reached.

[0038] From the final updated high-potential subset, the doping configuration with the highest overall adsorption performance score and verified kinetic stability was selected as the optimal doping configuration.

[0039] According to a second aspect of the present invention, a free gas detection system for power transformers using machine learning-based material screening, employing the method described in the first aspect of the present invention, is provided. The system includes:

[0040] The dataset construction module is used to sample the compound space to obtain different doping configurations of candidate gas-sensitive materials, and to calculate the material property datasets of the adsorption performance of the target gas by the different doping configurations through high-throughput first-principles calculations.

[0041] The model building module is used to obtain a pre-trained material screening model based on the material property dataset by fusing CGCNN convolutional kernels through the GNN framework, which is used to predict the adsorption performance of different doping configurations for target gases.

[0042] The configuration screening module is used to select the optimal doped configuration that meets the preset performance threshold in the compound space by using a Bayesian optimization framework, a Gaussian process regression as a surrogate model, and the expected improvement as the acquisition function, and by using a pre-trained material screening model.

[0043] The material screening module is used to detect the response performance of sensor arrays made of gas-sensitive materials with the optimal doping configuration but different doping concentrations, and to screen out the optimal gas-sensitive materials.

[0044] The identification module is used to acquire the response signal of the sensor array to the dissolved gas released from the transformer oil, and to identify the type of target gas and estimate the corresponding gas concentration by combining pattern recognition and regression algorithms; wherein, the sensor array is prepared using the optimal gas-sensitive material.

[0045] According to a third aspect of the present invention, a terminal is provided. The terminal includes a processor and a storage medium;

[0046] The storage medium is used to store instructions;

[0047] The processor is configured to operate according to the instructions to perform the steps of the method according to the first aspect of the invention.

[0048] According to a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps of the method according to the first aspect of the present invention.

[0049] The beneficial effects of this invention are as follows: Compared with existing technologies, by sampling the compound space and combining high-throughput first-principles calculations of material property datasets corresponding to different doping configurations regarding the adsorption performance of target gases, a pre-trained material screening model is constructed based on this data set. This model is then used to screen for the optimal doping configuration in the compound space, thereby identifying the optimal gas-sensitive material. Finally, based on the response signal of the sensor array made from the optimal gas-sensitive material, the type and concentration of the target gas are identified. This enables rapid, accurate, and intelligent localization of high-performance gas-sensitive materials from a vast pool of candidate materials. Simultaneously, by combining the sensor array made from the optimal gas-sensitive material with signal recognition, the efficiency of material development and sensor performance are significantly improved. It enables simultaneous identification of multiple characteristic gases (H2, CO, C2H2, etc.) by a single sensor, breaking the limitation of traditional sensors that require only one probe per gas, and exhibits high sensitivity and good environmental adaptability. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the free gas detection method for power transformers based on machine learning for material screening, as described in this invention.

[0051] Figure 2 This is a schematic diagram of the doping of modified molybdenum diselenide as a candidate gas-sensitive material;

[0052] Figure 3 This is a flowchart illustrating the process of selecting the optimal doping configuration that meets the preset performance threshold in the power transformer free gas detection method based on machine learning in this invention.

[0053] Figure 4 This is a schematic diagram of the preparation process of intrinsic molybdenum diselenide;

[0054] Figure 5 This is a schematic diagram of the preparation process of modified molybdenum diselenide;

[0055] Figure 6 This is a schematic diagram of the fabrication process of a gas sensor;

[0056] Figure 7 This is a schematic diagram illustrating the response performance test of the mixed free dissolved gases extracted from transformer oil. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, other embodiments obtained by those skilled in the art without creative effort are all within the protection scope of this invention.

[0058] According to a first aspect of the present invention, a method for detecting ionized gas in power transformers based on machine learning for material screening is provided.

[0059] like Figure 1 In one embodiment, the method includes the following steps:

[0060] S101. Sampling is performed on the compound space to obtain different doping configurations of the candidate gas-sensitive materials. The material property dataset of the adsorption performance of the target gas by the different doping configurations is calculated by high-throughput first-principles calculation.

[0061] This step mainly includes: based on the multi-component content generation patterns of ionized gases under different discharge faults in transformers, combining machine learning algorithms and first-principles high-throughput calculation methods, and incorporating parameters such as bond length, adsorption energy, density of states, Malicen charge population, deformation charge density, and band structure, a data set of material properties for the adsorption performance of target gases is constructed for different doping configurations of candidate gas-sensitive materials. The data set table structure is shown in Tables 1-3. Among them: Table 1 is the main table, mainly involving basic numerical features and identifiers; Table 2 is the spectral-derived feature sub-table; Table 3 is the spatial feature sub-table.

[0062] Table 1

[0063]

[0064] Table 2

[0065]

[0066] Table 3

[0067]

[0068] Among the candidate gas-sensitive materials, metal-doped enhanced molybdenum diselenide (MoSl) is proposed. Studies have shown that MoSl itself is oleophobic, reducing oil adsorption, and surface modification layers (such as graphene composites) can form a physical barrier, blocking oil molecule penetration. Furthermore, in extremely cold regions (<-20℃), the activity of conventional sensor materials decreases, leading to a sharp drop in detection sensitivity. Modified MoSl, by doping with metal oxides, enhances electron mobility, maintaining gas adsorption-desorption kinetic efficiency at low temperatures. Therefore, using modified MoSl as a candidate gas-sensitive material can effectively improve the sensor's accuracy, lifespan, and environmental adaptability. Modified MoSl can be obtained by doping intrinsic MoSl with noble metals such as Pb, Ag, and Pt. Figure 2 As shown.

[0069] S102. Based on the material property dataset, a pre-trained material screening model is obtained by fusing CGCNN convolutional kernels through the GNN framework, which is used to predict the adsorption performance of different doping configurations on the target gas.

[0070] Graph neural networks (GNNs) can process material crystal structure diagrams and predict interatomic interactions.

[0071] Step S102 specifically includes the following steps:

[0072] S1021: Based on the aforementioned material property dataset, extract the material property vector corresponding to each doping configuration. xi And transform it into a graph structure Gi , i This is the sequence number of the doped configuration.

[0073] Among them, the diagram structure Gi include:

[0074] Node: Each atom acts as a node, carrying the atom's feature vector.

[0075] Edges: Connecting nearest-neighbor atoms based on the cutoff radius (e.g., 8 Å), encoding distance and bond.

[0076] By employing a lightweight Graph Neural Network (GNN) and using a Convolutional Graph Neural Network (CGCNN) as the core architecture, this approach further combines a model selection mechanism, data augmentation techniques for imbalanced data structures, and a series of efficient optimization strategies. This comprehensive method aims to significantly improve computational efficiency while maintaining model performance, thus achieving an ideal balance between performance and computational efficiency. Specifically, the model selection mechanism ensures that the most suitable GNN model is selected for different tasks and scenarios. The imbalanced data augmentation technique enhances the model's ability to handle complex graph structures by increasing the number of minority class samples. In the feature space, for each minority class sample xi, its k nearest neighbors are found, and neighbors xj are randomly selected to generate new samples.

[0077]

[0078] Where λ∈[0,1] are random coefficients, and the new sample is located on the line connecting xi and xj.

[0079] S1022: Utilization xi Calculate the overall adsorption performance score for each doping configuration. Si Combined with the aforementioned graph structure Gi Construct a graph structure set {( Gi , Si )}.

[0080] The efficient optimization strategy further reduces computing costs by optimizing computing resources and algorithm processes, enabling the overall system to maintain high performance output while having higher operating efficiency and practicality.

[0081] In this step, the calculation of key indicators includes:

[0082] 1) Adsorption energy

[0083]

[0084] in:

[0085] This represents the adsorption energy of gas molecules on the material surface (unit: eV). Negative values ​​indicate exothermic adsorption processes, and larger absolute values ​​indicate more stable adsorption.

[0086] The total energy of the entire system after gas molecules are adsorbed on the surface of the material (unit: eV).

[0087] Total energy (in eV) for a pure (or doped) MoSe2 substrate model.

[0088] The total energy of isolated gas molecules (unit: eV);

[0089] 2) Work function

[0090]

[0091] in:

[0092] The work function represents the minimum energy required to move an electron from the Fermi level to a vacuum outside the material surface;

[0093] This represents the electrostatic potential energy (in eV) in the vacuum region near the material surface. It is typically calculated by taking the plane-averaged electrostatic potential at a sufficiently high position (generally greater than 10 Å) above the model surface.

[0094] The Fermi level of the system (unit: eV);

[0095] 3) Desorption barrier

[0096]

[0097] in:

[0098] This is the activation energy barrier for the desorption of gas molecules from the material surface;

[0099] The energy is the transition state energy.

[0100] Energy of the adsorption state;

[0101] 4) Energy loss function:

[0102] 1. Atomic energy prediction:

[0103] The core of atomic energy prediction is dual-objective joint optimization, which aims to ensure the model accurately predicts the total energy of the crystal to guarantee thermodynamic accuracy, and to match the gradient of the predicted energy with respect to the atomic coordinates (i.e., the atomic forces) with the true value, thereby ensuring the rationality of structural dynamics and ultimately improving the reliability of material performance prediction.

[0104] Constructing an atomic energy prediction function The ideas include:

[0105] Basic objective: Use mean absolute error (MAE) to measure the deviation between the predicted value and the actual DFT calculation value to ensure the accuracy of energy prediction;

[0106] Constraint Supplement: Add a force constraint term, because the force on an atom is the gradient of energy with respect to coordinates, and the two are physically related;

[0107] Weight balancing: through hyperparameters The importance of adjusting force constraints, balancing the accuracy of energy prediction with the rationality of force, and adapting to the structural stability requirements of material selection scenarios.

[0108] Based on the above ideas, the atomic energy prediction function Specifically, it is constructed as follows:

[0109]

[0110] in:

[0111] The number of crystal structures;

[0112] The model predicts the first i Total energy of each doped configuration;

[0113] The first DFT calculated i The standard energy of a doped configuration can be obtained based on density functional theory (DFT) using the first-principles calculation software VASP. VASP is a computer package for atomic-scale material simulation. Its core is to obtain the electronic states and energies of the system by approximating the Schrödinger equation. The specific implementation is divided into two paths: one is to solve the Kohn-Sham equation (which has integrated a hybrid functional calculation module) in the DFT framework and use electronic density functional approximation to handle exchange correlations; the other is to obtain the electronic structure information by solving the Roothaan equation under the Hartree-Fock (HF) approximation to achieve the basis set expansion of the wavefunction.

[0114] This is the set of three-dimensional coordinates of all atoms in all doped configurations;

[0115] To predict energy right The gradient (tensor) of the model represents the physical meaning of the forces on the atoms predicted by the model. ;

[0116] The true values ​​of atomic forces (tensors) obtained through DFT calculation have the same shape as R. The VASP software calculates the rate of energy change by making small perturbations to the coordinates of each atom, obtaining the force components of the atom in the x, y, and z directions, forming a 3D vector. .

[0117] For hyperparameters; The value can be determined in the following way: first, obtain a batch of doped configurations through DFT calculation. and This constitutes the training dataset; cross-validation is used to test different... Energy and stress predictions of the model under given values; selecting values ​​that optimize the overall performance of the model (e.g., achieving the required energy accuracy and minimizing stress deviation). Value. It should be noted that... The dimensional conflict between different weighting terms has been taken into account when determining the weighting, and it implicitly contains a dimensional conversion factor, the dimension of which is the square of the length divided by the energy (e.g., Ų / eV).

[0118] 2. Prediction of atomic forces:

[0119] The core of atomic force prediction is based on the physical principle of "energy-coordinate gradient relationship". By jointly learning energy and force, the predicted force can not only conform to the laws of thermodynamics, but also meet the accuracy requirements of gas-sensitive material screening.

[0120] With the construction of atomic energy prediction functions The approach is similar, atomic force prediction function Specifically, it is constructed as follows:

[0121]

[0122] in:

[0123] The number of doped configurations;

[0124] This represents the total number of atoms contained in all doped configurations in the training batch;

[0125] The number of atoms in the i-th doped configuration;

[0126] : The force vector (3D vector) of the j-th atom in the i-th doped configuration directly predicted by the model;

[0127] : The DFT reference force vector (3D vector) of the j-th atom in the i-th doped configuration;

[0128] This indicates that the force vectors of all atoms are multiplied by the same random rotation matrix;

[0129] The weighting coefficient (scalar, hyperparameter) of the rotational isovariability constraint term is typically 0.1.

[0130] 3. Overall energy loss:

[0131] Based on atomic energy prediction function and atomic force prediction function The comprehensive energy loss function is obtained. :

[0132]

[0133] Weighting coefficient (usually 0.4–0.6), adjusted according to task importance;

[0134] : Predicting atomic energy function The normalized atomic energy prediction function obtained by dividing by the corresponding "scaling factor" (such as the maximum value of the mean absolute error (MAE) in the training set) is a dimensionless quantity;

[0135] : Predicting the force on atoms The normalized atomic force prediction function obtained by dividing by the corresponding "scaling factor" (such as the maximum value of the mean square error MSE in the training set) is a dimensionless quantity.

[0136] This normalization method eliminates and The difference in dimensions.

[0137] 5) Comprehensive evaluation indicators

[0138]

[0139] in:

[0140] For the first i The adsorption performance score of each doping configuration is used to compare the performance of different doping configurations. The higher the score, the better the gas sensing performance.

[0141] This is a comprehensive energy loss function used to characterize the amount of charge energy transferred between gas molecules and the material substrate after gas adsorption.

[0142] This is the difference in work function, specifically the difference in work function before and after gas adsorption;

[0143] They are respectively , , , The corresponding weighting coefficients (dimensionless, because) , , , All of these are energy-related terms, and the dimensions of these terms can be, for example, eV. They are used to adjust the contribution of each indicator to the overall score. These weighting coefficients can be determined by combining the analytic hierarchy process (AHP) with cross-validation.

[0144] S1023, using the graph structure set {( Gi , Si Using these as training samples, a neural network with CGCNN convolutional kernels fused through the GNN framework is trained to obtain a pre-trained material screening model.

[0145] Specifically, the atomic-bond feature extraction capability of CGCNN can be integrated through a message passing mechanism. The pre-trained material selection model includes graph convolutional layers, pooling layers, and fully connected layers.

[0146] The graph convolutional layer uses the convolutional kernel design of CGCNN to process the input graph structure. Gi The atomic features (such as element type and valence state) and bond features (such as bond length and bond angle) are concatenated or multiplied by elements, and the node features are updated through fully connected layers (FC) and activation functions (such as Sigmoid / Softplus).

[0147] The pooling layer uses average pooling or attention mechanisms (such as multi-head attention) to aggregate node features, reduce dimensionality, and retain key information.

[0148] The adsorption performance score of the fully connected layer with corresponding doped configuration is predicted by linear transformation.

[0149] Meanwhile, molecular descriptors are embedded in the message passing and readout stages, and feature fusion is optimized through attention mechanisms to improve the model's learning efficiency for the driving force of eutectic formation.

[0150] Furthermore, a loss function needs to be constructed to determine the termination of training. Specifically, this loss function can be constructed based on the mean squared error (MSE), expressed as:

[0151]

[0152] in, This represents the true value of the adsorption performance score. The model predicted value for the adsorption performance score. n This represents the number of samples.

[0153] When the mean square error (MSE) is less than a preset threshold, training is terminated, and a pre-trained material screening model is obtained.

[0154] S103. In the compound space, using a Bayesian optimization framework, with Gaussian process regression as the surrogate model and expected improvement as the acquisition function, the optimal doping configuration that meets the preset performance threshold is obtained by using a pre-trained material screening model.

[0155] The general idea of ​​this step is to use machine learning methods to accelerate the design of the configuration of free characteristic gas gas-sensitive materials, analyze the mechanism of the gas-sensitive response performance of candidate gas-sensitive materials (e.g., metal-doped enhanced molybdenum diselenide) from a molecular perspective, and obtain the optimal configuration of the characteristic gas adsorption gas-sensitive material.

[0156] Specifically, such as Figure 3 As shown, this step specifically includes:

[0157] S1031, Screening Comprehensive Adsorption Performance Score Si Doping configurations exceeding a preset threshold are designated as high-potential subsets, and their optimal adsorption performance scores are assigned. A labeled dataset is constructed using the material property vectors of these high-potential subsets and their corresponding comprehensive adsorption performance scores.

[0158] S1032. Based on the labeled dataset, a well-trained Gaussian process surrogate model is obtained by constructing a kernel function and maximizing marginal likelihood.

[0159] In this step, the marginal likelihood function used for marginal likelihood maximization is expressed as:

[0160]

[0161] Where X is the feature matrix of the high-potential subset, and S is the comprehensive adsorption performance score vector corresponding to the high-potential subset. K The kernel function matrix, To observe the noise variance, l For length scale hyperparameters, is a unit diagonal matrix, and C is a constant term. The kernel function matrix is... K For example, the RBF kernel function can be used. Logarithmic function. The base is assumed to be the natural constant e.

[0162] By maximizing the marginal likelihood function mentioned above, the Gaussian process surrogate model can be trained to obtain a well-trained Gaussian process surrogate model, which can be used to predict the score distribution of the new configuration.

[0163] S1033. Based on the current optimal adsorption performance score and the trained Gaussian process surrogate model, the desired improvement function is obtained.

[0164] In this step, the Gaussian process surrogate model is represented as: ,in To predict the mean, we represent the effect of the Gaussian process surrogate model on the doping configuration. Overall score The standard deviation of the forecast represents the quantification of the uncertainty of the model regarding the predicted values.

[0165] The desired improvement function is expressed as:

[0166]

[0167]

[0168] Where Z represents the standardization improvement amount. and These are the standard normal cumulative distribution function and the probability density function, respectively. For the current doping configuration; function The value represents in The expected gain from sampling is given by the value. The larger this value is, the more likely the current doping configuration is to significantly improve the existing best performance.

[0169] S1034. Maximize the desired improvement function in the compound space to obtain the optimal candidate doping configuration.

[0170] The optimal candidate doping configuration can be calculated using a global optimization algorithm, such as the L-BFGS (Broyden-Fletcher-Goldfarb-Shanno, a quasi-Newton method) algorithm.

[0171] S1035. Verify whether the comprehensive adsorption performance score of the best candidate doping configuration exceeds the preset threshold, and if it exceeds the preset threshold, add the best candidate doping configuration to the high-potential subset, and update the labeled dataset.

[0172] S1036. Based on the updated labeled dataset, iteratively update the Gaussian process surrogate model, the desired improvement function, and the optimal candidate doping configuration until a preset iteration stopping condition is reached.

[0173] The iteration stopping condition can be, for example, the number of iterations reaching a threshold or the quantized value of the function to be improved falling below a preset threshold.

[0174] S1037. From the final updated high-potential subset, select the doping configuration with the highest comprehensive adsorption performance score and verified kinetic stability as the optimal doping configuration.

[0175] S104. The response performance of the sensor array made of gas-sensitive materials with the optimal doping configuration but different doping concentrations is tested, and the optimal gas-sensitive material is selected.

[0176] like Figure 4 and5 Intrinsic and metal-doped molybdenum diselenide materials can be prepared using a hydrothermal method. After the gas-sensitive materials are prepared, their properties can be characterized using SEM (scanning electron microscopy), EDS (energy dispersive spectroscopy), XRD (X-ray diffraction), and XPS (X-ray photoelectron spectroscopy). The effects of different concentrations of metal doping on the structure and gas-sensitive properties of intrinsic molybdenum diselenide materials can be studied from multiple perspectives, including morphological characteristics and crystal analysis.

[0177] This step specifically includes the following four aspects:

[0178] I. Gas Sample Collection and Pretreatment

[0179] 1) Automatic gas intake and transmission

[0180] Triggering mechanism: The system continuously monitors the signal from the transformer gas relay or the online degassing device. When a sufficient amount of gas accumulation is detected, the control unit automatically opens the solenoid valve group, and the gas is drawn to the detection unit through a fully sealed stainless steel or polytetrafluoroethylene (PTFE) pipeline under the drive of a micro-pressure differential or a micro-pump.

[0181] Oil-gas separation and purification: Gas samples may carry trace amounts of oil mist or moisture. The system will pass them through:

[0182] 1. Filter membrane / permeation membrane: For example, PDMS (polydimethylsiloxane) membranes can be used to selectively enrich target gases (such as H2, CO, C2H2) while blocking large molecular oil vapors to prevent sensor contamination and poisoning.

[0183] 2. Drying tube: Built-in molecular sieve or desiccant to adsorb water vapor and control the sample humidity within the sensor's optimal operating range (usually <40% RH), avoiding interference from water vapor on sensitive materials.

[0184] 2) Environmental parameter control and flow regulation

[0185] Temperature control module: The heater or Peltier element controlled by the PID (proportional-integral-derivative) algorithm stabilizes the working temperature of the gas sample and sensor at 25±2℃, eliminating sensor signal drift caused by ambient temperature fluctuations.

[0186] Flow control: The gas flow rate is stabilized at 100±10 ml / min using a mass flow controller (MFC) or a precision needle valve. A constant flow rate is crucial for ensuring the stability and repeatability of the sensor response. It is a minimum value to avoid X being undefined.

[0187] The key parameters are controlled as shown in Table 4 below:

[0188] Table 4

[0189]

[0190] II. Sensor Response Signal Acquisition

[0191] The prepared gas-sensitive materials with optimal doping configurations but different doping concentrations were mixed with deionized water, anhydrous ethanol, and a binder, respectively. These mixtures were then uniformly coated onto the electrode sheet of the gas sensor, and subsequently calcined to fabricate the gas sensor. Figure 6 As shown. Next, the performance of the gas sensor is tested using a pre-mixed free mixed characteristic gas. Finally, the practicality of the gas sensor is tested using free gas under different discharge fault conditions, such as... Figure 7 As shown.

[0192] Furthermore, signal acquisition and digitization are performed, including:

[0193] 1. Signal conditioning circuit: The weak electrical signal from the sensor is first amplified and reduced in noise by a preamplifier and filter circuit.

[0194] 2. High-precision ADC conversion: The conditioned analog signal is sampled by a 24-bit high-precision analog-to-digital converter and converted into a digital signal. High resolution ensures the detection of minute signal changes caused by low concentrations of gas.

[0195] 3. Dynamic response recording: The system not only records steady-state signal values, but also records the time-series curve of the entire response-recovery process, providing rich information for subsequent feature extraction.

[0196] III. Signal Processing and Feature Extraction

[0197] 1) Signal preprocessing and compensation

[0198] Noise reduction: Digital filters are used to further smooth the signal and eliminate random noise.

[0199] Temperature / humidity compensation: The signal is corrected using a pre-established compensation model to eliminate residual environmental interference.

[0200] 1. Temperature compensation model:

[0201]

[0202] in:

[0203] To compensate for the sensor resistance value after reaching the reference temperature;

[0204] The original resistance value was measured at the actual temperature (Tactual).

[0205] Activation energy is a key property of materials, representing the energy required for electron transitions.

[0206] k: Boltzmann constant;

[0207] The actual temperature of the gas (measured by a thermistor such as a PT1000 integrated near the sensor).

[0208] Reference temperature (the optimal operating temperature calibrated by the sensor, such as 300°C = 573K);

[0209] 2. Humidity compensation

[0210] The core idea behind constructing the humidity compensation formula is to quantify the coupling interference of "humidity – relative humidity RH – gas concentration" to achieve signal correction, specifically expressed as:

[0211]

[0212] in:

[0213] Normalized resistance change under dry conditions;

[0214] Relative humidity;

[0215] The actual concentration of the target gas;

[0216] , , , , Model coefficients; these coefficients have already taken into account the dimensional conflicts between different weight terms, and implicitly contain dimensional conversion factors; for example, The dimension of the quantity is " " The dimension of the quantity is " These coefficients can be obtained through experimental calibration combined with data fitting; furthermore, for the detection of transformer fault gases by the modified molybdenum diselenide sensor, the coefficient range after experimental calibration (dimensionless, slightly different due to gas type) is as follows:

[0217] (Constant term model coefficients): 0.001~0.005, corresponding to baseline shift at dry and zero concentration;

[0218] (RH linear term model coefficients): -0.002 to -0.0005, where the negative sign indicates that increased humidity leads to a decrease in response;

[0219] (RH quadratic model coefficients): 1e -5 ~5e -5 The positive sign indicates that the nonlinear attenuation is aggravated under high humidity;

[0220] (Cgas linear term model coefficients): 0.0001~0.001, where a positive sign indicates an enhanced response with increasing concentration, and the slope represents sensitivity;

[0221] (Interaction term model coefficients): -1e -7 ~-5e -7 The negative sign indicates that concentration and humidity work together to suppress the response.

[0222] 2) Multidimensional feature extraction

[0223] The following feature vectors were extracted from the time-series response curves of each sensor, as shown in Table 5:

[0224] Table 5

[0225]

[0226] IV. Screening of Optimal Gas-Sensitive Materials

[0227] Once multiple materials (such as MoSe2 nanosheets with different Pd doping concentrations) are prepared, the optimal gas-sensitive material is evaluated and screened based on the following key performance indicators. Specific experimental indicators are exemplified below:

[0228] 1. Core gas-sensitive performance indicators, including:

[0229] Sensitivity: For example, ΔR / R0, where R0 is the initial value of the signal and ΔR is the change in the signal value.

[0230] Selectivity: The response to a target gas (e.g., C2H2) in a mixed gas background (e.g., H2, CO, CH4 coexisting).

[0231] Detection limit: The lowest concentration that can be reliably detected, usually defined as the concentration at which the signal-to-noise ratio is ≥3.

[0232] Response / Recovery Time: The rate at which a gas reacts and recovers; this is crucial for online monitoring.

[0233] 2. Stability and Reliability Indicators

[0234] Short-term repeatability: The reproducibility of the signal when testing the same concentration of gas multiple times.

[0235] Long-term stability: After continuous operation for several days or weeks, whether the baseline signal and sensitivity drift are within an acceptable range (e.g., <5%).

[0236] Environmental resistance: performance stability under different humidity and temperature fluctuations.

[0237] 3. Manufacturability and Integration

[0238] Process repeatability: The stability of the process for synthesizing the material, and the difference in material properties between different batches.

[0239] Compatibility with MEMS processes: Whether the material is easily integrated onto micro / nano-scale sensor chips.

[0240] By comprehensively evaluating the above test indicators, the optimal gas-sensitive material was selected.

[0241] S105. Obtain the response signal of the sensor array to the dissolved gas released from the transformer oil, and combine the pattern recognition and regression algorithm to identify the type of target gas and estimate the corresponding gas concentration; wherein, the sensor array is prepared using the optimal gas-sensitive material.

[0242] The following explains this step from two aspects: identification of the target gas type and estimation of gas concentration:

[0243] 1) Gas type identification

[0244] Gas type identification can be achieved, for example, using a one-dimensional convolutional neural network (1D-CNN). In a 1D-CNN model, the time-series response curve of each sensor can be used as input. The 1D-CNN model automatically learns the local patterns and dependencies within these curves, making it well-suited for processing such time-series signals. It can effectively identify the unique response patterns produced by different gases. The model outputs a probability distribution, such as [C2H2: 92%, H2: 5%, CH4: 3%], thereby determining the predominant fault gas.

[0245] 2) Gas concentration estimation

[0246] Gas concentration estimation can be performed using the Support Vector Regression (SVR) algorithm. This algorithm is robust to noise, effective with small samples, and suitable for general scenarios, especially laboratory calibration with a small sample size. Specifically, Support Vector Regression (SVR) constructs a supervised learning model for predicting continuous variables. Its core objective is to find an optimal regression function f(x) that minimizes the deviation between the predicted and true values ​​while tolerating a certain level of error.

[0247] In the Support Vector Regression (SVR) algorithm, the supervised learning model can use the extracted steady-state features as the main input, and sometimes transient features are added as auxiliary features.

[0248] The supervised learning model requires training with standard gases of different concentrations to establish a “feature-concentration” mapping relationship.

[0249] The supervised learning model directly outputs the predicted concentration values ​​of various gases (e.g., C2H2: 1.8 ppm, H2: 120 ppm).

[0250] As can be seen, in this embodiment, by sampling the compound space and combining high-throughput first-principles calculations of material property datasets corresponding to different doping configurations regarding the adsorption performance of the target gas, a pre-trained material screening model is constructed based on this data set. This model is then used to screen for the optimal doping configuration in the compound space, thereby identifying the optimal gas-sensitive material. Finally, based on the response signal of the sensor array made from the optimal gas-sensitive material, the type and concentration of the target gas are identified. This achieves rapid, accurate, and intelligent localization of high-performance gas-sensitive materials from a vast pool of candidate materials. Simultaneously, by combining the sensor array made from the optimal gas-sensitive material with signal recognition, the efficiency of material development and sensor performance are significantly improved. It enables simultaneous identification of multiple characteristic gases (H2, CO, CH4, C2H2, etc.) by a single sensor, breaking the traditional sensor limitation of "one gas, one probe," and exhibits high sensitivity and good environmental adaptability.

[0251] Furthermore, the advantages of this embodiment are reflected in the following aspects:

[0252] 1. Conventional automatic measuring devices mostly rely on infrared spectroscopy or general electrochemical sensors, which are highly sensitive to multiple fault gases (such as H2, CO, CH4, C2H2, etc.), and are particularly susceptible to interference in mixed gas environments, leading to false judgments. This invention, however, through surface modification (such as noble metal doping and nanostructure modulation), can significantly improve the response sensitivity to specific fault gases and reduce false judgments.

[0253] 2. Sensors in direct contact with oil, gas, or humid environments are prone to signal drift due to oil film adhesion and chemical reactions, requiring frequent calibration or replacement. Molybdenum diselenide itself is oleophobic, reducing oil adsorption. The surface-modified layer of molybdenum diselenide forms a physical barrier, blocking oil molecule penetration.

[0254] 3. In extremely cold regions (<-20℃), the activity of conventional sensor materials decreases, leading to a sharp drop in detection sensitivity. Modified molybdenum diselenide, through doping with metal oxides, enhances electron mobility and maintains efficient gas adsorption-desorption kinetics at low temperatures. The "modified molybdenum diselenide detection method" solves the three major pain points of traditional devices—accuracy, lifespan, and environmental adaptability—through material innovation.

[0255] According to a second aspect of the present invention, a free gas detection system for power transformers using machine learning-based material screening, employing the method described in the first aspect of the present invention, is provided. The system includes:

[0256] The dataset construction module is used to sample the compound space to obtain different doping configurations of candidate gas-sensitive materials, and to calculate the material property datasets of the adsorption performance of the target gas by the different doping configurations through high-throughput first-principles calculations.

[0257] The model building module is used to obtain a pre-trained material screening model based on the material property dataset by fusing CGCNN convolutional kernels through the GNN framework, which is used to predict the adsorption performance of different doping configurations for target gases.

[0258] The configuration screening module is used to select the optimal doped configuration that meets the preset performance threshold in the compound space by using a Bayesian optimization framework, a Gaussian process regression as a surrogate model, and the expected improvement as the acquisition function, and by using a pre-trained material screening model.

[0259] The material screening module is used to detect the response performance of sensor arrays made of gas-sensitive materials with the optimal doping configuration but different doping concentrations, and to screen out the optimal gas-sensitive materials.

[0260] The identification module is used to acquire the response signal of the sensor array to the dissolved gas released from the transformer oil, and to identify the type of target gas and estimate the corresponding gas concentration by combining pattern recognition and regression algorithms; wherein, the sensor array is prepared using the optimal gas-sensitive material.

[0261] According to a third aspect of the present invention, a terminal is provided, comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to perform the steps of the method according to the first aspect of the present invention.

[0262] The detailed steps are the same as those of the data-driven transformer area line loss quantification method provided in the first aspect of this invention, and will not be repeated here.

[0263] According to a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon. When executed by a processor, the program implements the steps of the method described in the first aspect of the present invention.

[0264] The detailed steps are the same as those of the data-driven transformer area line loss quantification method provided in the first aspect of this invention, and will not be repeated here.

[0265] The beneficial effects of this invention are as follows: Compared with existing technologies, by sampling the compound space and combining high-throughput first-principles calculations of material property datasets corresponding to different doping configurations regarding the adsorption performance of target gases, a pre-trained material screening model is constructed based on this data set. This model is then used to screen for the optimal doping configuration in the compound space, thereby identifying the optimal gas-sensitive material. Finally, based on the response signal of the sensor array made from the optimal gas-sensitive material, the type and concentration of the target gas are identified. This enables rapid, accurate, and intelligent localization of high-performance gas-sensitive materials from a vast pool of candidate materials. Simultaneously, by combining the sensor array made from the optimal gas-sensitive material with signal recognition, the efficiency of material development and sensor performance are significantly improved. It enables simultaneous identification of multiple characteristic gases (H2, CO, C2H2, etc.) by a single sensor, breaking the limitation of traditional sensors that require only one probe per gas, and exhibits high sensitivity and good environmental adaptability.

[0266] Furthermore, this invention can also bring significant socio-economic benefits, including:

[0267] (1) Prevent major accidents, monitor fault gases in real time, provide early warning of internal discharge or overheating faults in transformers, avoid serious accidents such as fires and explosions, and ensure the stability of the power grid and public safety.

[0268] (2) Early fault intervention can shorten the power outage repair time and reduce the social chain impact on residents' lives, medical facilities and transportation systems.

[0269] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0270] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0271] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0272] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0273] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for detecting ionized gas in power transformers based on machine learning for material screening, characterized in that, Includes the following steps: Different doping configurations of candidate gas-sensitive materials are obtained by sampling the compound space, and material property datasets of the adsorption performance of the target gas by the different doping configurations are calculated by high-throughput first-principles calculation. Based on the aforementioned material property dataset, the material property vector corresponding to each doping configuration is extracted and transformed into a graph structure; The comprehensive adsorption performance score for each doping configuration is calculated using material property vectors, and a graph structure set is constructed by combining the graph structure. Using the graph structure set as training samples, a pre-trained material screening model is obtained; the pre-trained material screening model includes graph convolutional layers, pooling layers, and fully connected layers; the graph convolutional layers adopt the convolutional kernel design of CGCNN; the pooling layers aggregate node features; The adsorption performance score of the fully connected layer with corresponding doped configuration is predicted by linear transformation; In the compound space, doped configurations with a comprehensive adsorption performance score greater than a preset threshold are selected as high-potential subsets, and the optimal adsorption performance score is labeled. A labeled dataset is constructed using the material property vector of the high-potential subset and the corresponding comprehensive adsorption performance score. Based on the labeled dataset, a well-trained Gaussian process surrogate model is obtained by constructing a kernel function and maximizing marginal likelihood. Based on the current optimal adsorption performance score and the trained Gaussian process surrogate model, the desired improvement function is obtained; Maximizing the desired improvement function in the compound space yields the optimal candidate doping configuration. Verify whether the overall adsorption performance score of the best candidate doping configuration exceeds the preset threshold, and if it exceeds the preset threshold, add the best candidate doping configuration to the high-potential subset, and update the labeled dataset at the same time; Based on the updated labeled dataset, the Gaussian process proxy model, the desired improvement function, and the optimal candidate doping configuration are iteratively updated until a preset iteration stopping condition is reached. From the final updated high-potential subset, the doping configuration with the highest comprehensive adsorption performance score and verified kinetic stability was selected as the optimal doping configuration. The response performance of sensor arrays made from gas-sensitive materials with the optimal doping configuration but different doping concentrations was tested, and the optimal gas-sensitive material was selected. The sensor array is used to obtain the response signal of dissolved gas released from transformer oil. By combining pattern recognition and regression algorithms, the type of target gas is identified and the corresponding gas concentration is estimated. The sensor array is prepared using the optimal gas-sensitive material.

2. The method for detecting ionized gas in power transformers based on machine learning for material screening according to claim 1, characterized in that, The candidate gas-sensitive material is metal-doped enhanced molybdenum diselenide.

3. The method for detecting ionized gas in power transformers based on machine learning for material screening according to claim 1, characterized in that, The material property dataset includes multiple parameters such as bond length, adsorption energy, density of states, Malicken charge population, deformation charge density, and band structure.

4. The method for detecting ionized gas in power transformers based on machine learning for material screening according to claim 1, characterized in that, The comprehensive adsorption performance score Si Calculated using the following formula: in, For the first i Adsorption performance rating of different doped configurations. The comprehensive energy loss function characterizes the amount of charge energy transferred between gas molecules and the material substrate after gas adsorption. The energy of the adsorption state. The difference in the work function, This is the activation energy barrier for the desorption of gas molecules from the material surface. , , , They are respectively , , , The corresponding weighting coefficients.

5. The method for detecting ionized gas in power transformers based on machine learning for material screening according to claim 4, characterized in that, The comprehensive energy loss function Represented as: in, and These are atomic energy prediction functions. and atomic force prediction function The normalized atomic energy prediction function and the normalized atomic force prediction function obtained by dividing by the corresponding scaling factor are dimensionless quantities; These are the weighting coefficients; The number of doped configurations; The model predicts the first i Total energy of each doped configuration; The first DFT calculated i Standard energy for each doped configuration; This is the set of three-dimensional coordinates of all atoms in all doped configurations; To predict energy right The gradient, in physical terms, represents the force on the atom predicted by the model. ; The true values ​​of atomic forces obtained through DFT calculation, shape and same; For hyperparameters; This represents the total number of atoms contained in all doped configurations in the training batch; For the first i The number of atoms in each doped configuration; The first one directly predicted by the model i In the doping configuration, the first j The force vector of each atom; For the first i In the doping configuration, the first j DFT reference force vector of each atom; This means multiplying the force vectors of all atoms by the same random rotation matrix; These are the weighting coefficients for the rotational isovariability constraint term.

6. A power transformer free gas detection system employing the method described in any one of claims 1 to 5, based on machine learning for material screening, characterized in that, include: The dataset construction module is used to sample the compound space to obtain different doping configurations of candidate gas-sensitive materials, and to calculate the material property datasets of the adsorption performance of the target gas by the different doping configurations through high-throughput first-principles calculations. The model building module is used to obtain a pre-trained material screening model based on the material property dataset by fusing CGCNN convolutional kernels through the GNN framework, which is used to predict the adsorption performance of different doping configurations for target gases. The configuration screening module is used to select the optimal doped configuration that meets the preset performance threshold in the compound space by using a Bayesian optimization framework, a Gaussian process regression as a surrogate model, and the expected improvement as the acquisition function. The material screening module is used to detect the response performance of sensor arrays made of gas-sensitive materials with the optimal doping configuration but different doping concentrations, and to screen out the optimal gas-sensitive materials. The identification module is used to acquire the response signal of the sensor array to the dissolved gas released from the transformer oil, and to identify the type of target gas and estimate the corresponding gas concentration by combining pattern recognition and regression algorithms; wherein, the sensor array is prepared using the optimal gas-sensitive material.

7. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-5.

Images