Transformer fault prediction method and device based on dissolved gas, terminal equipment and storage medium
By enhancing the feature enhancement and classifier identification of the transformer's dissolved gas table data, the problem of insufficient ability to capture inter-column relationships when processing table data is solved, and the accuracy of transformer fault type identification is improved.
Patent Information
- Application Number
- CN202510296227.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-13
- Publication Date
- 2025-06-17
AI Technical Summary
When processing table data, the Transformer model has weak ability to capture relationships between specific columns, and cannot effectively utilize the structured characteristics of the transformer dissolved gas concentration and its ratio, resulting in insufficient accuracy in identifying the transformer fault type.
By obtaining the dissolved gas table data of the transformer, map it to the target dissolved gas characteristic matrix, and input it into the dissolved gas data feature optimization model for feature enhancement. Then, the feature-enhanced matrix is input into the pre-built fault type classifier to identify the fault type probability distribution and finally determine the fault type of the transformer.
By enhancing the characteristic of dissolved gas table data, the accuracy of the identification of transformer fault type is significantly improved, so that the fault type of transformer can be predicted more accurately and targeted maintenance measures are taken to prevent further deterioration of the fault.
Smart Images
Figure CN120162672A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of transformer fault detection, and in particular to a transformer fault prediction method, device, terminal equipment and storage medium based on dissolved gas. Background Art
[0002] Transformers are core equipment in power systems. During operation, the insulating oil and solid insulating materials inside them will decompose due to thermal and electrical stresses, generating a variety of dissolved gases. The types and concentrations of these dissolved gases can reflect the operating status and potential fault types of the transformer. In the field of deep learning, the Transformer model has attracted widespread attention for its excellent sequence modeling capabilities in dissolved gas feature analysis. Through the multi-head self-attention mechanism, Transformer can dynamically capture the global dependencies between dissolved gas concentrations, thereby comprehensively reflecting fault characteristics. Especially when processing DGA (Dissolved Gas Analysis) data, Transformer can treat gas concentrations and their ratios as implicit feature sequences, and use its global modeling capabilities to mine complex correlation patterns between gases. However, although Transformer is suitable for modeling unstructured or serialized data, when faced with tabular data with clear structured features and inter-column correlations (such as a table recording transformer dissolved gas concentrations and their ratios), it has a weak ability to capture specific inter-column relationships and cannot utilize these key features. Summary of the invention
[0003] The present invention provides a transformer fault prediction method, device, terminal equipment and storage medium based on dissolved gas, and the method can effectively improve the accuracy of transformer fault type identification.
[0004] An embodiment of the present invention provides a transformer fault prediction method based on dissolved gas, comprising:
[0005] Obtaining dissolved gas table data of the target transformer to be tested; wherein the dissolved gas table data includes concentration data of each dissolved gas;
[0006] Mapping the dissolved gas table data into a target dissolved gas characteristic matrix;
[0007] Inputting the target dissolved gas characteristic matrix into a dissolved gas data characteristic optimization model, so that the dissolved gas data characteristic optimization model generates a characteristic-enhanced target dissolved gas characteristic matrix according to the target dissolved gas characteristic matrix;
[0008] Input the target dissolved gas feature matrix after the feature enhancement into the pre-constructed fault type classifier, so that the fault type classifier can identify the probability distribution of the fault types of the target transformer to be detected according to the target dissolved gas feature matrix after the feature enhancement;
[0009] Determine the fault type of the target transformer to be detected according to the probability distribution of the fault types.
[0010] Further, the construction process of the dissolved gas data feature optimization model includes:
[0011] Obtain sample dissolved gas tabular data; wherein, each row in the sample dissolved gas tabular data corresponds to a dissolved gas sample, and each dissolved gas sample includes the sample concentration data of each dissolved gas and a fault type label; each column in the sample dissolved gas tabular data corresponds to a type of dissolved gas;
[0012] Map the sample dissolved gas tabular data into an original feature matrix; wherein, the original feature matrix contains a number of feature vectors; each feature vector corresponds to a dissolved gas sample;
[0013] Perform a random masking operation on the eigenvalues in the original feature matrix to obtain the original masked feature matrix after random masking;
[0014] Construct a feature extraction model based on the attention mechanism;
[0015] Initialize the feature extraction model, and input the original masked feature matrix into the initialized feature extraction model to perform iterative training on the initialized feature extraction model;
[0016] Take the feature extraction model after the training is completed as the dissolved gas data feature optimization model.
[0017] Further, the iterative training of the initialized feature extraction model includes:
[0018] For each iterative training, project the original masked feature matrix into a query matrix, a key matrix, and a value matrix respectively based on the current projection weight matrix of the feature extraction model;
[0019] Calculate the dot product of the query matrix and the key matrix to obtain a similarity matrix;
[0020] Adjust the similarity matrix according to the dimension information of the key matrix to obtain an adjusted similarity matrix;
[0021] Perform normalization calculation on the adjusted similarity matrix to obtain an attention weight matrix;
[0022] Perform a weighted sum on the value matrix according to the attention weight matrix to obtain a masked feature matrix with enhanced features.
[0023] Compare the masked feature matrix with enhanced features with the original masked feature matrix, and calculate the value of the loss function according to the comparison result.
[0024] Determine whether the value of the loss function converges.
[0025] If so, stop training to obtain the feature extraction model after training is completed.
[0026] If not, adjust the projection weight matrix of the feature extraction model to obtain the projection weight matrix for the next iterative training.
[0027] Further, the construction process of the fault type classifier includes:
[0028] Input the original masked feature matrix into the dissolved gas data feature optimization model so that the dissolved gas data feature optimization model outputs a first masked feature matrix with enhanced features.
[0029] Input the first masked feature matrix into a pre-constructed occlusion data restoration model so that the occlusion data restoration model predicts the restored feature values at the occluded positions of the first masked feature matrix according to the input data, and generates a second masked feature matrix according to the restored feature values and the first masked feature matrix.
[0030] Construct a classifier, use the second masked feature matrix as the input, and use the fault probabilities of each predicted fault type as the output, and perform iterative training on the classifier.
[0031] Use the classifier after training is completed as the fault type classifier.
[0032] Among them, in each iterative training process, compare the fault probabilities of each predicted fault type by the current classifier with the corresponding fault type labels in the sample dissolved gas table data, and calculate the cross-entropy loss value according to the comparison result; determine whether the cross-entropy loss value converges, if so, stop training to obtain the classifier after training is completed, if not, adjust the model parameters of the classifier to obtain the classifier for the next iterative training.
[0033] Further, the construction process of the occlusion data restoration model includes:
[0034] Construct a regression head; among them, the regression head is composed of several fully connected layers.
[0035] Using the first masking feature matrix as the input and the predicted feature values at the masked positions in the original masking feature matrix as the output, iteratively train the regression head;
[0036] Use the trained classifier as the occlusion data restoration model;
[0037] Among them, in each iterative training process, compare the predicted feature values predicted by the current regression head with the true feature values corresponding to the masked positions in the original feature matrix, and calculate the mean square error value according to the comparison result; determine whether the mean square error value reaches convergence. If so, stop training and obtain the trained regression head. If not, adjust the model parameters of the regression head to obtain the regression head for the next iterative training.
[0038] Further, the fault types include partial discharge fault, arc discharge fault, winding overheating fault, low-energy discharge, high-energy discharge, and high-temperature overheating.
[0039] Further, the dissolved gases include hydrogen, methane, ethane, ethylene, and acetylene.
[0040] An embodiment of the present invention further provides a transformer fault prediction device based on dissolved gases, including: a table acquisition module, a matrix mapping module, a feature enhancement module, a fault probability determination module, and a fault type determination module;
[0041] The table acquisition module is used to acquire the dissolved gas table data of the target transformer to be detected; wherein, the dissolved gas table data includes the concentration data of each dissolved gas;
[0042] The matrix mapping module is used to map the dissolved gas table data into a target dissolved gas feature matrix;
[0043] The feature enhancement module is used to input the target dissolved gas feature matrix into the dissolved gas data feature optimization model, so that the dissolved gas data feature optimization model generates a feature-enhanced target dissolved gas feature matrix according to the target dissolved gas feature matrix;
[0044] The fault probability determination module is used to input the feature-enhanced target dissolved gas feature matrix into a pre-constructed fault type classifier, so that the fault type classifier identifies the fault type probability distribution of the target transformer to be detected according to the feature-enhanced target dissolved gas feature matrix;
[0045] The fault type determination module is used to determine the fault type of the target transformer to be detected according to the fault type probability distribution.
[0046] The present application also provides a terminal device, including:
[0047] one or more processors;
[0048] A memory, coupled to the processor, for storing one or more programs;
[0049] When the one or more programs are executed by the one or more processors, the one or more processors implement the transformer fault prediction method based on dissolved gas as described in the above-mentioned embodiment of the invention.
[0050] The present application also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the transformer fault prediction method based on dissolved gas as described in the above-mentioned embodiment of the invention is implemented.
[0051] The following beneficial effects are achieved by implementing the present invention:
[0052] The present invention provides a transformer fault prediction method, apparatus, terminal equipment and storage medium based on dissolved gas. The method maps the acquired dissolved gas table data of the target transformer to be detected into a target dissolved gas feature matrix; and inputs the target dissolved gas feature matrix into a dissolved gas data feature optimization model, so that the dissolved gas data feature optimization model generates a target dissolved gas feature matrix after feature enhancement according to the target dissolved gas feature matrix; inputs the target dissolved gas feature matrix after feature enhancement into a pre-built fault type classifier, so that the fault type classifier identifies the fault type probability distribution of the target transformer to be detected according to the target dissolved gas feature matrix after feature enhancement; and determines the fault type of the target transformer to be detected according to the fault type probability distribution;
[0053] Therefore, by performing feature enhancement on the dissolved gas table data, when the target dissolved gas feature matrix after feature enhancement is input into the fault type classifier, the fault type probability distribution of the target transformer to be detected can be more accurately identified, thereby improving the accuracy of transformer fault type identification, and then taking corresponding maintenance measures in a targeted manner to prevent the fault from further deteriorating and ensure the safe and stable operation of the transformer. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] In order to more clearly illustrate the technical solution of the present application, the drawings required for use in the implementation manner will be briefly introduced below. Obviously, the drawings described below are only some implementation manners of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0055] Figure 1 It is a schematic flowchart of a transformer fault prediction method based on dissolved gas provided by an embodiment of the present application;
[0056] Figure 2 It is a schematic structural diagram of a transformer fault prediction device based on dissolved gas provided by an embodiment of the present application;
[0057] Figure 3 It is a schematic structural diagram of a terminal device provided by an embodiment of the present application. Detailed implementation manners
[0058] To make the objectives, technical solutions and advantages of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments in the present application belong to the scope of protection of the present application.
[0059] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs; the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above drawings are intended to cover non-exclusive inclusion.
[0060] In the description of the embodiments of the present application, technical terms such as "first" and "second" are only used to distinguish different objects and cannot be understood as indicating or implying relative importance or implicitly indicating the quantity, specific order or primary-secondary relationship of the indicated technical features. In the description of the embodiments of the present application, the meaning of "a plurality of" is two or more unless otherwise specifically defined.
[0061] Referring to "embodiments" herein means that specific features, structures or characteristics described in connection with the embodiments can be included in at least one embodiment of the present application. The phrase appears in various places in the specification and does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.
[0062] In the description of the embodiments of the present application, the term "and / or" is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this article generally represents an "or" relationship between the associated objects before and after.
[0063] In the description of the embodiments of the present application, the term "a plurality of" means two or more (including two). Similarly, "a plurality of groups" means two or more groups (including two groups), and "a plurality of sheets" means two or more sheets (including two sheets).
[0064] In the description of the embodiments of the present application, unless otherwise clearly specified and limited, technical terms such as "installation", "connection", "connection", "fixation", etc. should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can also be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two components or the interaction relationship between two components. For those of ordinary skill in the art, the specific meanings of the above terms in the embodiments of the present application can be understood according to specific situations.
[0065] See Figure 1 , which is a schematic flowchart of a transformer fault prediction method based on dissolved gas provided by an embodiment of the present invention, including:
[0066] S1. Obtain the dissolved gas tabular data of the target transformer to be detected; wherein, the dissolved gas tabular data includes the concentration data of each dissolved gas;
[0067] In a preferred embodiment, the dissolved gas includes hydrogen, methane, ethane, ethylene, and acetylene;
[0068] Schematically, in order to determine whether there is a fault in the target transformer to be detected and the specific fault type, it is first necessary to obtain the dissolved gas tabular data of the target transformer to be detected;
[0069] Specifically, the dissolved gas analysis (DGA) detection data of the target transformer to be detected is used as input data, including the concentration data of hydrogen, the concentration data of methane, the concentration data of ethane, the concentration data of ethylene, and the concentration data of acetylene;
[0070] To better illustrate the present application, the following provides the specific form of the dissolved gas table:
[0071]
[0072] It should be noted that the dissolved gas analysis detection data of the present application includes but is not limited to the above gases, and the types of gases are not limited herein.
[0073] S2. Map the dissolved gas tabular data to a target dissolved gas feature matrix;
[0074] Specifically, the input dissolved gas tabular data is mapped into a continuous vector representation, InputEmbedding, and integrated into a target dissolved gas feature matrix with the shape of (B, D), where B (Batch Size) represents the number of input samples, and D (Dimension) represents the dimension of the input features.
[0075] S3. Input the target dissolved gas feature matrix into the dissolved gas data feature optimization model, so that the dissolved gas data feature optimization model generates a target dissolved gas feature matrix with enhanced features according to the target dissolved gas feature matrix.
[0076] Schematically, to improve the accuracy of fault prediction, the present application introduces a dissolved gas data feature optimization model. By using the trained dissolved gas data feature optimization model to enhance the association strength between tabular data, a target dissolved gas feature matrix with enhanced features is obtained.
[0077] In a preferred embodiment, the construction process of the dissolved gas data feature optimization model includes:
[0078] Obtain sample dissolved gas tabular data; wherein, each row in the sample dissolved gas tabular data corresponds to a dissolved gas sample, and each dissolved gas sample includes sample concentration data of each dissolved gas and a fault type label; each column in the sample dissolved gas tabular data corresponds to a type of dissolved gas.
[0079] Map the sample dissolved gas tabular data into an original feature matrix; wherein, the original feature matrix contains several feature vectors; each feature vector corresponds to a dissolved gas sample.
[0080] Perform a random masking operation on the eigenvalues in the original feature matrix to obtain an original masked feature matrix after random masking.
[0081] Construct a feature extraction model based on the attention mechanism.
[0082] Initialize the feature extraction model, and input the original masked feature matrix into the initialized feature extraction model, and perform iterative training on the initialized feature extraction model.
[0083] Use the feature extraction model after training as the dissolved gas data feature optimization model.
[0084] Specifically, to better illustrate the present application, the following provides the specific form of the sample dissolved gas tabular data:
[0085]
[0086] Schematically, during the process of training the dissolved gas data feature optimization model, in order to enhance the model's understanding ability of the structured information of tabular data and effectively improve the modeling performance, it is necessary to perform random masking on the sample dissolved gas tabular data, and then perform feature prediction to reconstruct the masked tabular data features;
[0087] Specifically, a random masking operation is performed on the feature values in the original feature matrix to obtain the original masked feature matrix X after random masking mask ; then a feature extraction model based on the attention mechanism is constructed, the feature extraction model is initialized, and the original masked feature matrix X mask is input into the initialized feature extraction model, the initialized feature extraction model is iteratively trained, and the trained feature extraction model is used as the dissolved gas data feature optimization model.
[0088] In a preferred embodiment, the iterative training of the initialized feature extraction model includes:
[0089] For each iterative training, based on the current projection weight matrix of the feature extraction model, the original masked feature matrix is projected into a query matrix, a key matrix, and a value matrix respectively;
[0090] Calculate the dot product of the query matrix and the key matrix to obtain a similarity matrix;
[0091] According to the dimension information of the key matrix, the similarity matrix is adjusted to obtain an adjusted similarity matrix;
[0092] Perform normalization calculation on the adjusted similarity matrix to obtain an attention weight matrix;
[0093] According to the attention weight matrix, perform weighted summation on the value matrix to obtain a masked feature matrix with enhanced features;
[0094] Compare the masked feature matrix with enhanced features with the original masked feature matrix, and calculate the loss function value according to the comparison result;
[0095] Judge whether the loss function value converges,
[0096] If so, stop training and obtain the trained feature extraction model,
[0097] If not, adjust the projection weight matrix of the feature extraction model to obtain the projection weight matrix for the next iterative training;
[0098] Specifically, the original masked feature matrix X maskInput into the initialized feature extraction model. For each iterative training, based on the current projection weight matrices \(W\) Q 、\(W\) K and \(W\) V of the feature extraction model, project the original masked feature matrix into a query matrix. The specific formula is as follows:
[0099] \(Q = X\) mask ·\(W\) Q ;
[0100] Project the original masked feature matrix into a key matrix. The specific formula is as follows:
[0101] \(K = X\) mask ·\(W\) K ;
[0102] Project the original masked feature matrix into a value matrix. The specific formula is as follows:
[0103] \(V = X\) mask ·\(W\) V ;
[0104] Calculate the dot product of the query matrix and the key matrix to obtain a similarity matrix, and then adjust the similarity matrix according to the dimension information of the key matrix to obtain an adjusted similarity matrix where \(d_k\) is the dimension of the key matrix \(K\);
[0105] Then perform a normalization calculation on the adjusted similarity matrix to obtain an attention weight matrix, and then perform a weighted sum on the value matrix according to the attention weight matrix to obtain a masked feature matrix with enhanced features. The specific calculation formula is as follows:
[0106]
[0107] Compare the masked feature matrix with enhanced features with the original masked feature matrix, and calculate the loss function value according to the comparison result; determine whether the loss function value converges. If so, stop training to obtain the feature extraction model after training is completed. If not, adjust the projection weight matrix of the feature extraction model to obtain the projection weight matrix for the next iterative training.
[0108] Preferably, in order to improve the correlation degree between samples, before inputting the original masked feature matrix \(X\) mask into the initialized feature extraction model, different weights can also be assigned to the input data to emphasize the part that contributes the most to the result and weaken the secondary part, so as to enhance the model's ability to capture the global feature correlation between different dissolved gas samples.
[0109] S4. Input the target dissolved gas feature matrix after feature enhancement into the pre-constructed fault type classifier, so that the fault type classifier can identify the probability distribution of the fault types of the target transformer to be detected according to the target dissolved gas feature matrix after feature enhancement;
[0110] In a preferred embodiment, the construction process of the fault type classifier includes:
[0111] Input the original masked feature matrix into the dissolved gas data feature optimization model, so that the dissolved gas data feature optimization model outputs the first masked feature matrix after feature enhancement;
[0112] Input the first masked feature matrix into the pre-constructed occlusion data restoration model, so that the occlusion data restoration model predicts the restored feature values of the occluded positions of the first masked feature matrix according to the input data, and generates a second masked feature matrix according to the restored feature values and the first masked feature matrix;
[0113] Construct a classifier, use the second masked feature matrix as the input, and the fault probabilities of each fault type predicted as the output, and perform iterative training on the classifier;
[0114] Use the classifier after training as the fault type classifier;
[0115] Among them, in each iterative training process, compare the fault probabilities of each fault type predicted by the current classifier with the corresponding fault type labels in the sample dissolved gas table data, and calculate the cross-entropy loss value according to the comparison result; judge whether the cross-entropy loss value converges. If so, stop training to obtain the classifier after training. If not, adjust the model parameters of the classifier to obtain the classifier for the next iterative training;
[0116] Specifically, input the original masked feature matrix X mask into the trained dissolved gas data feature optimization model, so that the dissolved gas data feature optimization model outputs the first masked feature matrix after feature enhancement
[0117] Then input the first masked feature matrix into the pre-constructed occlusion data restoration model, predict the restored feature values of the occluded positions of the first masked feature matrix, that is, map the high-dimensional features to the specific values of the occluded positions in the table data, and generate the second masked feature matrix with the occluded position features restored according to the restored feature values and the first masked feature matrix
[0118] Construct a classifier with the second masking feature matrix as the input and output a probability distribution with the same number as the number of fault types, representing the predicted probability of each fault type;
[0119] During this process, calculate the cross-entropy loss value according to the comparison result, and its specific expression is as follows:
[0120]
[0121] wherein, represents the probability that the predicted i-th sample belongs to the j-th class;
[0122] In a preferred embodiment, the fault types include partial discharge fault, arc discharge fault, winding overheating fault, low-energy discharge, high-energy discharge and high-temperature overheating;
[0123] Specifically, when the electric field distribution inside the transformer is uneven, partial discharge may occur at the weak insulation part. This partial discharge will decompose the insulating oil to generate hydrogen and a small amount of methane; from the perspective of gas concentration, generally the content of hydrogen is relatively higher than that of other hydrocarbon gases, and the total hydrocarbon (the sum of hydrocarbon gases such as CH4, C2H6, C2H4, C2H2, etc.) content is relatively lower. At the same time, in the initial stage of partial discharge, the hydrogen concentration may rise slowly.
[0124] Specifically, the arc discharge fault is caused by the breakdown of the insulation inside the transformer, resulting in the formation of an arc channel between the electrodes; for example, when the insulation between the windings of the transformer is damaged, or the insulation between the winding and the iron core fails, arc discharge may occur; this fault will generate a large amount of acetylene, because acetylene is the product of the violent breaking and recombination of carbon-hydrogen bonds during high-energy discharge; its gas characteristics are that the acetylene content is very high, and the contents of hydrogen and other hydrocarbon gases (such as CH4, C2H4, etc.) will also increase significantly; the total hydrocarbon content rises rapidly, and the proportion of acetylene in the total hydrocarbon is very large, which is an important feature distinguishing it from other fault types;
[0125] Specifically, when the current passing through the transformer winding is too large and exceeds its rated current, it will cause overheating faults in the winding; this situation may be caused by reasons such as overload operation and short circuit inside the winding. For example, when the transformer is overloaded for a long time or the cooling system fails, the winding temperature will continue to rise. During this process, the insulating oil and solid insulating materials will decompose due to high temperature. The main gases generated during the fault process are methane (CH4) and ethylene (C2H4). From the perspective of gas concentration changes, as the temperature rises, the concentrations of methane and ethylene will gradually increase, and to a certain extent, the ratio of ethylene to methane concentration can reflect the severity of overheating; if the temperature is higher, the growth rate of ethylene will be faster than that of methane, and their ratio will increase.
[0126] In a preferred embodiment, the construction process of the occlusion data restoration model includes:
[0127] Construct a regression head; wherein, the regression head is composed of several fully connected layers;
[0128] Using the first masked feature matrix as the input and the predicted feature values of the masked positions in the original masked feature matrix as the output, perform iterative training on the regression head;
[0129] Take the trained classifier as the occlusion data restoration model;
[0130] Among them, in each iterative training process, compare the predicted feature values obtained by the current regression head with the true feature values corresponding to the masked positions in the original feature matrix, calculate the mean square error value according to the comparison result; determine whether the mean square error value reaches convergence. If so, stop training and obtain the trained regression head. If not, adjust the model parameters of the regression head to obtain the regression head for the next iterative training.
[0131] Schematically, in order to improve the prediction accuracy of the masked positions, it is necessary to train the regression head so that the occlusion data restoration model can more accurately predict the restored feature values of the masked positions;
[0132] Specifically, using the first masked feature matrix as the input and the predicted feature values of the masked positions in the original masked feature matrix as the output, perform iterative training on the regression head,
[0133] and in each training process, use the mean square error (MSE) to calculate the mean square error value between the predicted value and the true value. The specific expression is as follows:
[0134]
[0135] S5. Determine the fault type of the target transformer to be detected according to the probability distribution of the fault types.
[0136] In a preferred embodiment, the fault types include partial discharge fault, arc discharge fault, winding overheating fault, low energy discharge, high energy discharge, and high temperature overheating.
[0137] Specifically, assume the probability distribution of the fault types is as follows:
[0138] The probability of partial discharge fault is 0.1; the probability of arc discharge fault is 0.05; the probability of winding overheating fault is 0.2; the probability of low energy discharge is 0.15; the probability of high energy discharge is 0.08; the probability of high temperature overheating is 0.42.
[0139] Then, according to the above probability values, it is analyzed that the probability of high temperature overheating fault is the highest; thus, the fault type of the target transformer to be detected is determined to be high temperature overheating.
[0140] Refer to Figure 2 , which is a transformer fault prediction device based on dissolved gas provided by an embodiment of the present invention, including: a table acquisition module, a matrix mapping module, a feature enhancement module, a fault probability determination module, and a fault type determination module.
[0141] The table acquisition module is used to acquire the dissolved gas table data of the target transformer to be detected; wherein, the dissolved gas table data includes the concentration data of each dissolved gas.
[0142] The matrix mapping module is used to map the dissolved gas table data into a target dissolved gas feature matrix.
[0143] The feature enhancement module is used to input the target dissolved gas feature matrix into the dissolved gas data feature optimization model, so that the dissolved gas data feature optimization model generates a feature-enhanced target dissolved gas feature matrix according to the target dissolved gas feature matrix.
[0144] The fault probability determination module is used to input the feature-enhanced target dissolved gas feature matrix into a pre-constructed fault type classifier, so that the fault type classifier identifies the probability distribution of the fault types of the target transformer to be detected according to the feature-enhanced target dissolved gas feature matrix.
[0145] The fault type determination module is used to determine the fault type of the target transformer to be detected according to the probability distribution of the fault types.
[0146] See Figure 3 , and an embodiment of the present application further provides a terminal device, including:
[0147] One or more processors;
[0148] A memory, coupled to the processor, for storing one or more programs;
[0149] When the one or more programs are executed by the one or more processors, the one or more processors implement the dissolved gas-based transformer fault prediction method as described above.
[0150] The processor is used to control the overall operation of the terminal device to complete all or part of the steps of the above-mentioned dissolved gas-based transformer fault prediction method. The memory is used to store various types of data to support the operation of the terminal device. These data may include, for example, instructions for any application program or method operating on the terminal device, as well as application-related data. The memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic memory, flash memory, a magnetic disk, or an optical disk.
[0151] In an exemplary embodiment, the terminal device can be implemented by one or more Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), Digital Signal Processing Devices (DSPDs), Programmable Logic Devices (PLDs), Field Programmable Gate Arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components, and is used to execute the dissolved gas-based transformer fault prediction method as described in any of the above embodiments, and achieve the same technical effects as the above method.
[0152] In another exemplary embodiment, a computer-readable storage medium including a computer program is further provided. When the computer program is executed by a processor, the steps of the transformer fault prediction method based on dissolved gas described in any of the above embodiments are implemented. For example, the computer-readable storage medium may be the memory including the computer program described above, and the computer program may be executed by the processor of the terminal device to complete the transformer fault prediction method based on dissolved gas described in any of the above embodiments and achieve the same technical effects as the above method.
[0153] The above are the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.
Claims
1. A transformer fault prediction method based on dissolved gas, characterized in that: include: Obtaining dissolved gas table data of the target transformer to be tested; wherein the dissolved gas table data includes concentration data of each dissolved gas; Mapping the dissolved gas table data into a target dissolved gas characteristic matrix; Inputting the target dissolved gas characteristic matrix into a dissolved gas data characteristic optimization model, so that the dissolved gas data characteristic optimization model generates a characteristic-enhanced target dissolved gas characteristic matrix according to the target dissolved gas characteristic matrix; Inputting the target dissolved gas feature matrix after feature enhancement into a pre-built fault type classifier, so that the fault type classifier can identify the fault type probability distribution of the target transformer to be detected according to the target dissolved gas feature matrix after feature enhancement; The fault type of the target transformer to be detected is determined according to the fault type probability distribution.
2. The transformer fault prediction method based on dissolved gas according to claim 1, characterized in that: The process of constructing the dissolved gas data characteristic optimization model includes: Obtaining sample dissolved gas table data; wherein each row in the sample dissolved gas table data corresponds to a dissolved gas sample, and each dissolved gas sample includes sample concentration data of each dissolved gas and a fault type label; each column in the sample dissolved gas table data corresponds to a type of dissolved gas; Mapping the sample dissolved gas table data into an original feature matrix; wherein the original feature matrix includes a plurality of feature vectors; each of the feature vectors corresponds to one of the dissolved gas samples; Performing a random masking operation on the eigenvalues in the original feature matrix to obtain an original masked feature matrix after random masking; Construct a feature extraction model based on attention mechanism; Initializing the feature extraction model, inputting the original masked feature matrix into the initialized feature extraction model, and iteratively training the initialized feature extraction model; The trained feature extraction model is used as the dissolved gas data feature optimization model.
3. The transformer fault prediction method based on dissolved gas according to claim 2, characterized in that: The iterative training of the initialized feature extraction model comprises: For each iteration of training, based on the current projection weight matrix of the feature extraction model, the original masked feature matrix is projected into a query matrix, a key matrix and a value matrix respectively; Calculating the dot product of the query matrix and the key matrix to obtain a similarity matrix; According to the dimension information of the key matrix, the similarity matrix is adjusted to obtain an adjusted similarity matrix; Performing normalization calculation on the adjusted similarity matrix to obtain an attention weight matrix; According to the attention weight matrix, weighted summing is performed on the value matrix to obtain a masked feature matrix after feature enhancement; Compare the masked feature matrix after feature enhancement with the original masked feature matrix, and calculate the loss function value according to the comparison result; Determine whether the loss function value converges, If so, stop training and get the feature extraction model after training. If not, the projection weight matrix of the feature extraction model is adjusted to obtain the projection weight matrix for the next iterative training.
4. The transformer fault prediction method based on dissolved gas according to claim 3, characterized in that: The construction process of the fault type classifier includes: Inputting the original masking feature matrix into the dissolved gas data feature optimization model, so that the dissolved gas data feature optimization model outputs a first masking feature matrix after feature enhancement; Inputting the first masking feature matrix into a pre-built masking data restoration model, so that the masking data restoration model predicts the restored feature values of the masked positions of the first masking feature matrix according to the input data, and generates a second masking feature matrix according to the restored feature values and the first masking feature matrix; Constructing a classifier, taking the second masked feature matrix as input and taking the predicted fault probabilities of each fault type as output, and iteratively training the classifier; Using the trained classifier as the fault type classifier; Among them, in each iterative training process, the fault probability of each fault type predicted by the current classifier is compared with the corresponding fault type label in the sample dissolved gas table data, and the cross entropy loss value is calculated according to the comparison result; it is judged whether the cross entropy loss value converges, if so, the training is stopped to obtain the classifier after training, if not, the model parameters of the classifier are adjusted to obtain the classifier for the next iterative training.
5. The transformer fault prediction method based on dissolved gas according to claim 4, characterized in that: The process of constructing the occlusion data restoration model includes: Constructing a regression head; wherein the regression head is composed of a plurality of fully connected layers; Taking the first masked feature matrix as input and taking the predicted feature value of the masked position in the original masked feature matrix as output, iteratively training the regression head; Using the trained classifier as the occlusion data restoration model; Among them, in each iterative training process, the predicted eigenvalues predicted by the current regression head are compared with the real eigenvalues corresponding to the masked position in the original feature matrix, and the mean square error value is calculated according to the comparison result; it is judged whether the mean square error value has reached convergence, if so, the training is stopped to obtain the regression head after the training is completed, if not, the model parameters of the regression head are adjusted to obtain the regression head for the next iterative training.
6. The transformer fault prediction method based on dissolved gas according to claim 5, characterized in that: The fault types include partial discharge fault, arc discharge fault, winding overheat fault, low energy discharge, high energy discharge and high temperature overheat.
7. The transformer fault prediction method based on dissolved gas according to any one of claims 1 to 6, characterized in that: The dissolved gases include hydrogen, methane, ethane, ethylene and acetylene.
8. A transformer fault prediction device based on dissolved gas, characterized in that: include: Table acquisition module, matrix mapping module, feature enhancement module, fault probability determination module and fault type determination module; The table acquisition module is used to acquire dissolved gas table data of the target transformer to be detected; wherein the dissolved gas table data includes concentration data of each dissolved gas; The matrix mapping module is used to map the dissolved gas table data into a target dissolved gas characteristic matrix; The feature enhancement module is used to input the target dissolved gas feature matrix into the dissolved gas data feature optimization model, so that the dissolved gas data feature optimization model generates a target dissolved gas feature matrix after feature enhancement according to the target dissolved gas feature matrix; The fault probability determination module is used to input the target dissolved gas feature matrix after feature enhancement into a pre-built fault type classifier, so that the fault type classifier can identify the fault type probability distribution of the target transformer to be detected according to the target dissolved gas feature matrix after feature enhancement; The fault type determination module is used to determine the fault type of the target transformer to be detected according to the fault type probability distribution.
9. A terminal device, characterized in that: include: one or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the transformer fault prediction method based on dissolved gas as described in any one of claims 1 to 7.
10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the transformer fault prediction method based on dissolved gas as described in any one of claims 1 to 7 is implemented.