Transformer diagnosis method and device based on knowledge graph, electronic equipment and storage medium

Abstract

The invention discloses a transformer diagnosis method and device based on a knowledge graph, electronic equipment and a storage medium, and belongs to the field of power equipment faults, and the method comprises the steps: obtaining the test parameters of a transformer and a knowledge graph containing the incidence relation between the parameters and a diagnosis rule; screening is carried out according to the incidence relation between the parameters and a threshold value, and whether a certain parameter and relevant parameters are abnormal exists or not is judged so as to determine to-be-confirmed abnormity; if the to-be-confirmed exception exists, generating a plurality of candidate exception types according to a diagnosis rule of the knowledge graph; extracting an adjacent matrix and a feature matrix of the parameters from the knowledge graph, and inputting the adjacent matrix and the feature matrix into a prediction model to generate a prediction probability of each abnormal type; and comparing the prediction probability of the candidate fault with a confidence coefficient threshold, and judging the specific abnormity of the transformer. By implementing the method and the device, the problem of low diagnosis sensitivity and accuracy caused by neglecting the internal relevance between different physical parameters in the prior art can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment failure, and in particular to a transformer diagnosis method, device, electronic device and storage medium based on a knowledge graph. Background Art

[0002] Power transformers are key components of the power system, and their operating status directly impacts the safety and stability of the entire power grid. Accurate and timely transformer status diagnosis and early detection of potential insulation defects or structural failures are crucial for ensuring power supply reliability and preventing major power outages and economic losses. Therefore, effectively improving the accuracy and intelligence of transformer fault diagnosis has long been a research priority in the power industry.

[0003] However, existing transformer diagnostic methods still lack accuracy and reliability. Existing methods often analyze each test parameter in isolation, ignoring the inherent correlations between different physical parameters. Many early or complex fault symptoms manifest as coordinated anomalies in multiple related parameters. Isolating fluctuations in individual parameters makes it difficult to effectively identify such faults, resulting in low diagnostic sensitivity and accuracy. Summary of the Invention

[0004] The embodiments of the present invention provide a transformer diagnosis method, device, electronic device and storage medium based on a knowledge graph, which can solve the problem in the prior art of low diagnostic sensitivity and accuracy due to ignoring the intrinsic correlation between different physical parameters.

[0005] An embodiment of the present invention provides a transformer diagnosis method based on a knowledge graph, comprising:

[0006] Obtaining test parameters and a transformer knowledge graph at the time of transformer testing; the test parameters include: DC resistance, insulation resistance, low voltage short-circuit impedance, winding deformation test data, dielectric loss factor, capacitance, and frequency domain dielectric spectrum; the transformer knowledge graph includes correlations between test parameters and diagnostic rules for fault diagnosis;

[0007] Based on the correlation between the test parameters and the preset abnormality threshold, it is determined whether each test parameter is abnormal. If it is determined that a test parameter and the rest of the test parameters associated with it are in an abnormal state, it is determined that the transformer has an abnormality to be confirmed; otherwise, it is determined that the transformer does not have an abnormality;

[0008] In the case that the transformer has an abnormality to be confirmed, fault diagnosis is performed on each test parameter according to the diagnostic rules to generate several candidate abnormality types;

[0009] Extracting an adjacency matrix and a feature matrix of test parameters from the transformer knowledge graph, and inputting the adjacency matrix and the feature matrix into a preset abnormality prediction model, so that the abnormality prediction model generates a prediction probability of each abnormality type in a preset abnormality type set through the adjacency matrix and the feature matrix;

[0010] Determine the prediction probability of the corresponding candidate abnormality type based on the prediction probability of each abnormality type in the preset abnormality type set;

[0011] For each candidate anomaly type, if the corresponding prediction probability is greater than the preset confidence threshold, it is determined that the transformer has the corresponding anomaly.

[0012] Furthermore, the transformer knowledge graph is constructed in the following ways, including:

[0013] Access transformer ontology libraries, structured diagnostic rules in transformer industry standards, and unstructured text from transformer historical maintenance reports;

[0014] Based on the transformer ontology library and the preset physical mechanism model, the equipment entities, parameter entities, fault entities, and the relationships between entities in the transformer knowledge graph are defined to generate a knowledge graph for instantiated data to be imported;

[0015] Using a preset natural language processing model, entity relationship triples are extracted from the unstructured text in the transformer historical maintenance report. The extracted entity relationship triples and the structured diagnostic rules in the transformer industry standard are imported into the knowledge graph of the instantiated data to be imported to generate a transformer knowledge graph.

[0016] Furthermore, the training of the anomaly prediction model includes:

[0017] Obtain a historical transformer fault case dataset; wherein the historical transformer fault case dataset includes several graph-structured training samples and corresponding fault type labels; the graph-structured training samples include an adjacency matrix extracted from the transformer knowledge graph and a feature matrix embedded with state feature parameters;

[0018] Dividing the transformer historical fault case dataset into several batches of training samples according to a preset batch size;

[0019] Each batch of training samples is sequentially input into the anomaly prediction model for iterative training until a preset training termination condition is reached; wherein, when the anomaly prediction model receives each batch of training samples, it outputs the predicted probability of each anomaly type in the preset anomaly type set based on the adjacency matrix and feature matrix in the current batch; a preset loss function is used to calculate and generate a loss function value based on the predicted probability and the corresponding fault type label; and a preset optimizer is used to update the parameters in the anomaly prediction model based on the loss function value.

[0020] Furthermore, the training of the natural language processing model includes:

[0021] Obtain a text corpus dataset; wherein the text corpus dataset includes a number of unstructured texts from historical transformer maintenance reports and corresponding entity relationship triple labels;

[0022] Divide the text corpus dataset into several batches of training samples according to the preset batch size;

[0023] Each batch of training samples is input into the natural language processing model in sequence for iterative training until the preset training termination condition is reached; wherein, when the natural language processing model receives each batch of training samples, it outputs the predicted entity relationship triples based on the unstructured text in the current batch; through a preset loss function, the loss function value is calculated based on the predicted entity relationship triples and the corresponding entity relationship triple labels; and a preset optimizer is used to update the parameters in the natural language processing model according to the loss function value.

[0024] Furthermore, the transformer diagnosis method based on knowledge graph further includes:

[0025] After each new transformer fault type is obtained, the corresponding maintenance report is extracted, and the unstructured text in the maintenance report is input into the natural language processing model, so that the natural language processing model generates corresponding entity relationship triples based on the unstructured text in the maintenance report;

[0026] According to the transformer knowledge graph, the generated entity relationship triples are tested for knowledge conflicts. If no conflict is detected, the generated entity relationship triples are merged into the transformer knowledge graph; if a conflict is detected, the generated entity relationship triples are discarded.

[0027] Furthermore, the physical mechanism model includes a leakage current model, an AC impedance model, a mechanical resonance frequency model, a Debye relaxation model, a plate capacitance model, and a Cole-Cole model:

[0028] The leakage current model is specifically:

[0029]

[0030] Where, I leak is the leakage current; V is the voltage applied across the insulating medium; R ins is the insulation resistance;

[0031] The AC impedance model is specifically:

[0032]

[0033] Where Z is the impedance; R is the resistance of the circuit; f is the frequency of the AC power supply; L is the inductance of the circuit;

[0034] The mechanical resonance frequency model is specifically:

[0035]

[0036] Where, f n is the natural resonant frequency of the system; k is the equivalent stiffness coefficient of the system; m is the equivalent mass of the system;

[0037] The Debye relaxation model is specifically:

[0038]

[0039] Where tanδ is the dielectric loss factor; ∈″ is the imaginary part of the complex dielectric constant; ∈′ is the real part of the complex dielectric constant;

[0040] The flat plate capacitor model is specifically:

[0041]

[0042] Where C is the capacitance; r is the relative dielectric constant of the insulating medium; ∈0 is the dielectric constant of vacuum; A is the equivalent area of ​​the electrode; d is the distance between the electrodes;

[0043] The Cole-Cole model is specifically:

[0044]

[0045] Where,∈ * (ω) is the complex dielectric constant that varies with angular frequency; ∞ is the high-frequency limiting dielectric constant; Δ∈ is the dielectric strength increment; j is the imaginary unit; ω is the angular frequency; τ is the characteristic relaxation time; α is the empirical parameter describing the width of the relaxation time distribution.

[0046] Furthermore, the abnormality threshold is calculated in the following manner:

[0047] Obtain the operating temperature and relative humidity of the transformer at the time of detection;

[0048] For each test parameter, a corresponding abnormal threshold is calculated according to the operating temperature and the ambient relative humidity.

[0049] Based on the above method embodiments, the present invention provides corresponding device embodiments.

[0050] An embodiment of the present invention provides a transformer diagnosis device based on a knowledge graph, comprising: a transformer data acquisition module, a transformer preliminary state diagnosis module, a candidate abnormality type generation module, and a transformer final diagnosis module;

[0051] The transformer data acquisition module is used to obtain the test parameters of the transformer at the time of detection and the transformer knowledge map; the test parameters include: DC resistance value, insulation resistance value, low voltage short-circuit impedance value, winding deformation test data, dielectric loss factor, capacitance, and frequency domain dielectric spectrum; the transformer knowledge map includes the correlation between test parameters and diagnostic rules for fault diagnosis;

[0052] The transformer preliminary state diagnosis module is used to determine whether each test parameter is abnormal based on the correlation between the test parameters and the preset abnormality threshold. If it is determined that a test parameter and the other test parameters associated with it are all in an abnormal state, it is determined that the transformer has an abnormality to be confirmed; otherwise, it is determined that the transformer does not have an abnormality.

[0053] The candidate abnormality type generating module is used to perform fault diagnosis on each test parameter according to the diagnostic rules and generate several candidate abnormality types when there is an abnormality to be confirmed in the transformer;

[0054] The transformer final diagnosis module is used to extract the adjacency matrix and feature matrix of the test parameters from the transformer knowledge graph, and input the adjacency matrix and the feature matrix into a preset abnormality prediction model, so that the abnormality prediction model generates a prediction probability of each abnormality type in a preset abnormality type set through the adjacency matrix and the feature matrix; based on the prediction probability of each abnormality type in the preset abnormality type set, the prediction probability of the corresponding candidate abnormality type is determined; for each candidate abnormality type, if the corresponding prediction probability is greater than a preset confidence threshold, it is determined that the transformer has a corresponding abnormality.

[0055] Based on the above method embodiment, the present invention provides a corresponding electronic device embodiment.

[0056] An embodiment of the present invention provides an electronic device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the knowledge graph-based transformer diagnosis method described in any one of the above method embodiments is implemented.

[0057] Based on the above method embodiment, the present invention provides a corresponding storage medium embodiment.

[0058] An embodiment of the present invention provides a storage medium having a computer program stored thereon, wherein when the computer program is running, the device where the storage medium is located is controlled to execute the transformer diagnosis method based on the knowledge graph described in any one of the above method embodiments.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] The embodiment of the present invention provides a transformer diagnosis method, device, electronic device and storage medium based on a knowledge graph. The method first obtains the test parameters of the transformer and a knowledge graph containing the correlation relationship between the parameters and the diagnosis rules. Then, a preliminary screening is performed based on the correlation relationship between the parameters and the preset threshold to determine whether a certain parameter and its associated parameters are all abnormal to determine the unconfirmed abnormality. If there is an unconfirmed abnormality, a candidate list containing several possible fault types is generated according to the diagnosis rules in the knowledge graph. Finally, in order to achieve accurate prediction, the method extracts the adjacency matrix and feature matrix of the parameters from the knowledge graph and inputs them into the prediction model. The model generates a predicted probability for each abnormality type; by comparing the predicted probability of the candidate fault with the confidence threshold, the specific abnormality of the transformer is finally determined.

[0061] The present invention utilizes a knowledge graph that includes the correlation between test parameters. During diagnosis, it can make a comprehensive judgment on multiple parameters based on the correlation, avoiding the defect of the existing technology of isolating a single parameter and ignoring its internal connection. It can more effectively identify complex faults manifested by the coordinated anomalies of multiple parameters, thereby improving the accuracy of diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flow chart of a transformer diagnosis method based on a knowledge graph provided in one embodiment of the present invention.

[0063] Figure 2 It is a structural schematic diagram of a transformer diagnosis device based on a knowledge graph provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0064] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0065] like Figure 1 As shown, in order to solve the problem in the prior art of low diagnostic sensitivity and accuracy due to ignoring the intrinsic correlation between different physical parameters, an embodiment of the present invention provides a transformer diagnosis method based on a knowledge graph, which includes at least the following steps:

[0066] Step S1: Obtaining test parameters and a transformer knowledge graph at the time of transformer detection; the test parameters include: DC resistance value, insulation resistance value, low voltage short-circuit impedance value, winding deformation test data, dielectric loss factor, capacitance, and frequency domain dielectric spectrum; the transformer knowledge graph includes correlations between test parameters and diagnostic rules for fault diagnosis;

[0067] Specifically, the system first acquires several transformer test parameters at a specific testing moment, along with a pre-built transformer knowledge graph. These test parameters are key indicators for assessing transformer health and may include DC resistance, insulation resistance, low-voltage short-circuit impedance, winding deformation test data for determining winding mechanical condition, dielectric loss factor, capacitance, and frequency-domain dielectric spectrum. It should be clarified that the test parameters in the specific implementation refer to a data set obtained by a series of specific measurements of specific components or circuits of the transformer according to the industry standard test procedures. Specifically: the DC resistance value includes the resistance value set measured for each phase winding at each tap position; the insulation resistance value, dielectric loss factor and capacitance all include the numerical value set measured for the insulation system of each high-voltage, medium-voltage and low-voltage winding of the transformer to the ground and between each winding; the low-voltage short-circuit impedance value includes the impedance value set measured for each phase and each winding pair; the winding deformation test data refers to the response curve data set obtained for each phase winding under different test connection methods through the frequency domain response analysis (FRA) method; the frequency domain dielectric spectrum refers to the dielectric characteristic data set measured for each main insulation system in a wide frequency range. These parameter data can be obtained by connecting to various online monitoring systems or importing them into offline test data management systems.

[0068] In parallel with acquiring the test parameters, this method also acquires the transformer knowledge graph, which serves as the knowledge foundation for all subsequent intelligent analysis and diagnosis. The relationships between the test parameters contained in the knowledge graph are not simply data correlations, but are deeply integrated with the transformer's operating mechanism. For example, it structuredly stores the cross-validation relationships between different test parameters, as well as the physical mechanism models underlying these relationships, such as Ohm's law for characterizing DC resistance and the Debye relaxation model for characterizing dielectric loss, thereby effectively linking the measured data with physical laws. In addition, the knowledge graph also includes diagnostic rules for fault diagnosis. These rules can be derived from authoritative criteria in industry standards such as IEC 60599, or from the mature experience of domain experts, and are converted into machine-readable and executable logical forms. In this way, the knowledge graph organically networks the transformer's device entities, parameter states, fault modes, physical mechanisms, and diagnostic rules, forming a comprehensive domain knowledge base.

[0069] By obtaining the above-mentioned parameter data and rich knowledge graphs, a solid foundation has been laid for the subsequent realization of intelligent data cleaning and accurate fault diagnosis based on the collaboration of data and mechanisms.

[0070] In a preferred embodiment, the transformer knowledge graph is constructed in the following manner, including:

[0071] Access transformer ontology libraries, structured diagnostic rules in transformer industry standards, and unstructured text from transformer historical maintenance reports;

[0072] Based on the transformer ontology library and the preset physical mechanism model, the equipment entities, parameter entities, fault entities, and the relationships between entities in the transformer knowledge graph are defined to generate a knowledge graph for instantiated data to be imported;

[0073] Using a preset natural language processing model, entity relationship triples are extracted from the unstructured text in the transformer historical maintenance report. The extracted entity relationship triples and the structured diagnostic rules in the transformer industry standard are imported into the knowledge graph of the instantiated data to be imported to generate a transformer knowledge graph.

[0074] In a specific embodiment of the present invention, the step of constructing a transformer knowledge graph aims to create a comprehensive, accurate, and multi-source knowledge-integrated diagnostic foundation. This process begins by acquiring the knowledge sources required for construction. These knowledge sources are primarily divided into three categories: the first is a pre-defined transformer ontology library, which defines the basic conceptual framework and entity hierarchy for the knowledge graph; the second is structured diagnostic rules derived from transformer industry standards such as IEC 60599, which are authoritative and directly applicable to fault judgment; and the third is unstructured text derived from a large number of historical transformer maintenance reports, which contain rich implicit knowledge reflecting real-world fault scenarios.

[0075] After obtaining the above-mentioned knowledge sources, the next step is to accurately define the schema layer (Schema) of the knowledge graph based on the transformer ontology library and the preset physical mechanism model. In this step, three types of core entities will be clearly defined: equipment entities (such as windings, tap changers, cores, etc.), parameter entities (such as DC resistance, dielectric loss factor and other state characteristic parameters) and fault entities (such as moisture, deformation, partial discharge, etc.). More importantly, this step will establish the association relationship between entities. These relationships are not simple connections, but are strictly constructed based on physical mechanisms. After this step is completed, a "knowledge graph of instantiated data to be imported" with a rigorous structure and clear logic but not yet filled with specific case data is generated, which is equivalent to completing the "blueprint design" of the diagnostic knowledge system.

[0076] Finally, in the stage of filling and generating the knowledge graph, this method adopts corresponding processing paths for different types of knowledge sources. On the one hand, a preset natural language processing model (such as a BERT-based relational extraction model) is used to deeply analyze the unstructured text in the historical maintenance report, and entity relationship triples that can describe the specific relationship between equipment, faults, and parameter states are automatically and batch-extracted. On the other hand, the structured diagnostic rules obtained from industry standards are formatted to make them compatible with the knowledge graph model. Finally, the automatically extracted entity relationship triples and the formatted structured diagnostic rules are imported into the "knowledge graph of instantiated data to be imported" template generated in the previous step. By storing and linking in a graph database (such as Neo4j or GraphDB), the knowledge is instantiated, thereby generating a transformer knowledge graph with rich content that can be used for subsequent intelligent diagnosis.

[0077] The knowledge graph constructed through the above steps provides a comprehensive and structured knowledge foundation for the present invention to achieve intelligent data cleaning and accurate fault diagnosis based on the collaboration of data and mechanism.

[0078] In a preferred embodiment, the physical mechanism model includes a leakage current model, an AC impedance model, a mechanical resonance frequency model, a Debye relaxation model, a plate capacitance model, and a Cole-Cole model:

[0079] The leakage current model is specifically:

[0080]

[0081] Where, I leak is the leakage current; V is the voltage applied across the insulating medium; R ins is the insulation resistance;

[0082] The AC impedance model is specifically:

[0083]

[0084] Where Z is the impedance; R is the resistance of the circuit; f is the frequency of the AC power supply; L is the inductance of the circuit;

[0085] The mechanical resonance frequency model is specifically:

[0086]

[0087] Where, f n is the natural resonant frequency of the system; k is the equivalent stiffness coefficient of the system; m is the equivalent mass of the system;

[0088] The Debye relaxation model is specifically:

[0089]

[0090] Where tanδ is the dielectric loss factor; ∈″ is the imaginary part of the complex dielectric constant; ∈′ is the real part of the complex dielectric constant;

[0091] The flat plate capacitor model is specifically:

[0092]

[0093] Where C is the capacitance; r is the relative dielectric constant of the insulating medium; ∈0 is the dielectric constant of vacuum; A is the equivalent area of ​​the electrode; d is the distance between the electrodes;

[0094] The Cole-Cole model is specifically:

[0095]

[0096] Where,∈ * (ω) is the complex dielectric constant that varies with angular frequency; ∞is the high-frequency limiting dielectric constant; Δ∈ is the dielectric strength increment; j is the imaginary unit; ω is the angular frequency; τ is the characteristic relaxation time; α is the empirical parameter describing the width of the relaxation time distribution.

[0097] In a specific embodiment of the present invention, the preset physical mechanism model included in the knowledge graph provides a solid theoretical basis for the relationships between different experimental parameters, ensuring that the knowledge graph is constructed based on a deep collaboration between data and mechanisms. Specifically, the physical mechanism model not only defines the physical nature of a single experimental parameter but also reveals the inherent connections between multiple experimental parameters due to common physical causes.

[0098] The leakage current model explains that insulation resistance (IR) is a physical quantity that characterizes an insulator's ability to prevent leakage current. When the transformer's overall insulation deteriorates (such as due to moisture or aging), not only does the insulation resistance (IR) decrease due to the principles illustrated by the model, but this degradation, as a common physical root cause, also causes an increase in dielectric polarization loss, which is directly reflected in an increase in the dielectric loss factor (tanδ) and an abnormal low-frequency response in the frequency-domain dielectric spectrum. Therefore, the model links the three test parameters—insulation resistance, dielectric loss factor, and frequency-domain dielectric spectrum—through the state of overall insulation degradation.

[0099] Similarly, the AC impedance model defines the relationship between the low-voltage short-circuit impedance (Z) and the winding's resistance (R) and inductance (L). When a winding deforms or shifts, the geometric changes cause changes in the inductance L, which in turn changes the low-voltage impedance. This mechanical deformation is the core target of frequency response analysis (FRA) and is correlated with the winding deformation test data. Furthermore, if the fault involves a conductor connection problem, it will change the resistance R, or the measured DC resistance (DCR), which will also affect the low-voltage impedance. Therefore, this electromagnetic model links the low-voltage impedance, winding deformation test data, and DC resistance through the common foundation of the winding's mechanical and electrical integrity. Furthermore, the mechanical resonant frequency model explains the principle of frequency response analysis (FRA) from a structural dynamics perspective: mechanical deformation of the winding changes its equivalent stiffness (k) and mass (m), resulting in a shift in the resonant frequency. This is also consistent with the aforementioned change in the low-voltage short-circuit impedance (Z).

[0100] The Debye relaxation model, the flat plate capacitor model, and the Cole-Cole model together form a complete theoretical framework for evaluating the dielectric properties of insulating systems. The basic definition of the Debye relaxation model reveals that the dielectric loss factor (tanδ), capacitance (C) (positively correlated with ∈′), and the frequency domain dielectric spectrum (FDS) (directly measuring the changes in ∈′ and ∈″ with frequency) are essentially mathematical descriptions of different aspects of the state of the same insulating medium, and they are inevitably interrelated. The Cole-Cole model provides a more detailed physical explanation for the complete spectrum of the frequency domain dielectric spectrum (FDS), and its analysis results can be cross-validated with the frequency response characteristics of the dielectric loss factor (tanδ) and the temperature coefficient of capacitance.

[0101] By integrating these physical mechanism models into the knowledge graph, the correlation between different test parameters is clearly explained, providing a solid and reliable knowledge foundation for subsequent fault diagnosis.

[0102] In a preferred embodiment, the training of the natural language processing model includes:

[0103] Obtain a text corpus dataset; wherein the text corpus dataset includes a number of unstructured texts from historical transformer maintenance reports and corresponding entity relationship triple labels;

[0104] Divide the text corpus dataset into several batches of training samples according to the preset batch size;

[0105] Each batch of training samples is input into the natural language processing model in sequence for iterative training until the preset training termination condition is reached; wherein, when the natural language processing model receives each batch of training samples, it outputs the predicted entity relationship triples based on the unstructured text in the current batch; through a preset loss function, the loss function value is calculated based on the predicted entity relationship triples and the corresponding entity relationship triple labels; and a preset optimizer is used to update the parameters in the natural language processing model according to the loss function value.

[0106] Specifically, to automatically extract massive amounts of unstructured textual knowledge, a dedicated natural language processing model needs to be pre-trained. This model is a key technical tool for the automated construction and incremental updating of knowledge graphs. The model training process begins by acquiring a high-quality, manually annotated text corpus dataset. This dataset, which serves as the foundation for model learning, primarily derives from a large number of historical transformer maintenance reports and related technical documentation. Each data sample in the dataset consists of two parts: one is the original unstructured text fragment, such as the sentence describing "DCR value increases due to poor contact of the tap changer"; the other is the corresponding pre-annotated entity relationship triplet label, such as (<tap changer, status, poor contact>, <poor contact, resulting in, DCR increase>).

[0107] Before training begins, to improve training efficiency and stability, the entire text corpus dataset is randomly divided into several batches of training samples according to a preset batch size. The system then sequentially inputs these batches of training samples into the natural language processing model (e.g., a BERT-based relation extraction model) for iterative training until the model's performance converges or a preset training termination condition (e.g., a preset number of training rounds or accuracy requirements) is met. In each iteration, the natural language processing model analyzes and outputs its predicted entity relationship triples based on the unstructured text in each batch of training samples. The system then uses a preset loss function (e.g., a cross-entropy loss function) to compare the difference between the model's predicted triples and the true entity relationship triple labels corresponding to the batch of samples, and calculates a loss function value that quantifies this difference. Finally, a preset optimizer (e.g., the Adam optimizer) fine-tunes and updates the parameters within the natural language processing model using a backpropagation algorithm based on the calculated loss function value, with the goal of making the predicted results in subsequent iterations increasingly close to the true labels.

[0108] Through the above training process, we finally obtained a natural language processing model that can automatically and efficiently identify entities and extract structured relationship triplets from new, unseen transformer-related texts, providing key technical support for the automated construction and continuous incremental updating of knowledge graphs.

[0109] Step S2: determining whether each test parameter is abnormal based on the correlation between the test parameters and a preset abnormality threshold; if it is determined that a test parameter and the rest of the test parameters associated with it are abnormal, determining that the transformer has an abnormality to be confirmed; otherwise, determining that the transformer has no abnormality;

[0110] In a specific embodiment of the present invention, after intelligent data cleaning of the acquired data, the system performs a two-stage collaborative determination of whether each test parameter is abnormal based on the correlation between the test parameters and the preset abnormality threshold. This determination process first compares the measured value of each cleaned test parameter with the dynamically calculated abnormality threshold. This step can initially identify which parameter readings deviate from their normal range under the current operating conditions and temporarily mark these parameters as preliminarily abnormal.

[0111] However, the present invention does not simply treat any preliminary abnormality as a fault signal directly, but further utilizes the knowledge graph for collaborative verification to improve the accuracy of judgment and reduce false alarms. Specifically, when it is determined that there is a test parameter in a preliminary abnormal state, the system will query the pre-built knowledge graph to obtain other test parameters that have a physical mechanism association with the abnormal parameter. Subsequently, the system will check whether these associated test parameters are also in their respective abnormal states. If it is determined that there is an initially abnormal test parameter, and at least one of the other test parameters associated with it is also in an abnormal state in a collaborative manner, the system will jointly determine this pattern composed of multiple collaborative abnormal parameters as a transformer that has a pending abnormality that requires subsequent in-depth diagnosis. On the contrary, if a parameter is initially abnormal, but it has a strongly associated parameter defined in the knowledge graph that behaves normally, the system will judge it as measurement noise or an isolated, less important fluctuation, and thus determine that there is no abnormality in the transformer.

[0112] This two-stage judgment method, which is based on collaborative verification of knowledge graph association relationships, can effectively filter out instantaneous fluctuations or measurement noise of a single parameter, significantly improving the reliability and accuracy of abnormal state identification.

[0113] In a preferred embodiment, the abnormality threshold is calculated in the following manner:

[0114] Obtain the operating temperature and relative humidity of the transformer at the time of detection;

[0115] For each test parameter, a corresponding abnormal threshold is calculated according to the operating temperature and the ambient relative humidity.

[0116] In a specific embodiment of the present invention, the threshold used to determine whether a test parameter is abnormal is not fixed but rather dynamically calculated. This calculation process first obtains the transformer's operating parameters at the time of testing, primarily including operating temperature and ambient relative humidity. The system then calculates the corresponding abnormality threshold for each test parameter based on these real-time operating parameters and a preset dynamic adjustment formula corresponding to the physical mechanism of the test parameter.

[0117] Specifically, for different test parameters, the dynamic adjustment formula varies due to different physical mechanisms, mainly including:

[0118] For DC resistance (DCR), its dynamic threshold is mainly affected by the linear relationship between temperature and the operating temperature. When the deviation between the operating temperature and a reference temperature exceeds a preset temperature change threshold, the following linear temperature correction formula can be used to calculate:

[0119] Th DCR，dyn =Th DCR，static ×[1+α(T meas -T ref )]

[0120] Where Th DCR，dyn is the calculated DC resistance dynamic threshold; Th DCR，static is the preset static threshold value of DC resistance at the reference temperature; α is the resistance temperature coefficient of the preset conductor material, with a typical value of 0.00393 / °C; T meas is the transformer operating temperature obtained in real time; T ref is the reference temperature corresponding to the static threshold.

[0121] For insulation resistance (IR), its dynamic threshold is mainly affected by the exponential relationship between humidity. When the relative humidity of the environment deviates from its baseline value and reaches a preset humidity change threshold, it can be calculated using the following exponential humidity compensation formula:

[0122]

[0123] Where TH IR，dyn is the calculated dynamic threshold value of insulation resistance; TH IR，static RH is the preset static threshold value of insulation resistance at reference humidity; ref Is the reference relative humidity; RH mean The relative humidity of the environment is obtained in real time; C RII is an empirical constant for the effect of humidity on the insulation structure and surface condition (e.g., a value of 200).

[0124] The dynamic threshold of dielectric loss factor (tan δ) is affected by both temperature and humidity. When the operating temperature deviates from a reference temperature by more than a preset temperature change threshold, and the relative humidity deviates from its reference value by a preset humidity change threshold, the following composite weighted correction formula can be used for calculation:

[0125] Th tanδ，dyn =Th tanδ，static ×[1+K T (T meas -Tref )+K RH (RH meas -RH ref )]

[0126] Where Th tanδ，dyn The calculated dynamic threshold value of dielectric loss factor; Th tanδ，static K is the preset static threshold value of dielectric loss factor; T is the weight coefficient of temperature influence; K RH is the weight coefficient of humidity influence;

[0127] For capacitance (C), its dynamic reference value is affected by temperature. When the deviation between the operating temperature and a reference temperature exceeds a preset temperature change threshold, it can be adjusted using the following formula:

[0128] C dym_ref =C static_ref ×[1-k c (T meas -T ref )]

[0129] Where C dym_ref is the calculated capacitance dynamic threshold; C static_ref is the preset static capacitance threshold; k c is the comprehensive correction coefficient of capacitance;

[0130] The low voltage short-circuit impedance (Z) is also affected by temperature. When the deviation between the operating temperature and a reference temperature exceeds a preset temperature change threshold, its dynamic reference value can be corrected using the following formula:

[0131] Z dyn =Z static ×[1+βΔT]

[0132] Where Z dyn is the dynamic threshold of low voltage short circuit impedance; Z static is the static threshold of low voltage short-circuit impedance; β is the temperature coefficient of resistance; ΔT is the difference between the measured temperature and the reference temperature;

[0133] For the winding deformation test data, the judgment threshold is also affected by temperature. When the deviation between the operating temperature and a reference temperature exceeds a preset temperature change threshold, it can be adjusted using the following formula:

[0134] Th CCF，dyn =Th CCF，static -k w ΔT

[0135] Where Th CCF，dyn is the calculated dynamic threshold of the winding deformation test data; ThCCF，static k is the static threshold value of the preset winding deformation test data; w is the winding deformation temperature coefficient;

[0136] For frequency domain dielectric spectroscopy (FDS), its dynamic threshold is also related to temperature. When the deviation between the operating temperature and a reference temperature exceeds a preset temperature change threshold, it can be calculated using the following model:

[0137]

[0138] Where Th FDS，dyn (T) is the dynamic threshold of the frequency domain dielectric spectrum at temperature T; Th FDS，static (T) is the static threshold of the frequency domain dielectric spectrum at temperature T; k T is the preset temperature coefficient.

[0139] Step S3: When the transformer has an abnormality to be confirmed, perform fault diagnosis on each test parameter according to the diagnostic rules to generate several candidate abnormality types;

[0140] In a specific embodiment of the present invention, once the system determines, through the aforementioned collaborative verification steps, that a transformer has a pending anomaly consisting of multiple collaboratively abnormal test parameters, it initiates a knowledge graph-based fault diagnosis process. The first step in this process is to perform preliminary symbolic logic-based fault diagnosis on the pending anomaly, based on the diagnostic rules stored in the knowledge graph. The goal is to quickly select several candidate anomaly types that are most relevant to the current anomaly from among numerous possible faults.

[0141] Specifically, the system takes the parameter combination that constitutes the anomaly to be confirmed as input and performs pattern matching in the diagnostic rule library of the knowledge graph to generate several candidate anomaly types. This rule library contains a library of rule templates derived from industry standards and expert experience.

[0142] This rule-based preliminary diagnostic step enables us to quickly narrow the scope of troubleshooting by leveraging deterministic domain knowledge, providing a clear focus and input for subsequent, more refined probabilistic model analysis.

[0143] Step S4: extracting an adjacency matrix and a feature matrix of test parameters from the transformer knowledge graph, and inputting the adjacency matrix and the feature matrix into a preset abnormality prediction model, so that the abnormality prediction model generates a prediction probability for each abnormality type in a preset abnormality type set through the adjacency matrix and the feature matrix;

[0144] In a specific embodiment of the present invention, after generating candidate abnormality types through diagnostic rules, in order to quantitatively evaluate these possibilities and draw more accurate diagnostic conclusions, the method will further use a preset abnormality prediction model for in-depth analysis. The input of this model is not traditional flat data, but graph-structured data that can reflect the complex relationship between parameters. Specifically, the system will extract a relevant subgraph from the transformer knowledge graph for the current diagnostic scenario and represent it as an adjacency matrix and a feature matrix of test parameters. The adjacency matrix describes the connection relationship between the entity nodes such as each test parameter, equipment component, etc. in the subgraph, that is, the local topological structure of the knowledge graph; and the feature matrix contains the quantitative attributes of each node. For example, for a parameter node, its features may include the measurement value after cleaning, the deviation rate from the benchmark, etc.

[0145] After preparing the adjacency matrix and feature matrix, the system inputs the adjacency matrix and the feature matrix together into the preset anomaly prediction model. This anomaly prediction model is preferably a graph neural network (GNN) model, which can propagate and aggregate the information in the feature matrix at multiple levels along the graph structure path defined by the adjacency matrix, thereby learning the deep nonlinear dependencies between parameters. Through this graph-based calculation, the anomaly prediction model ultimately generates a prediction probability distribution covering each anomaly type in the preset anomaly type set. This probability distribution provides a quantitative confidence score for each candidate anomaly type and other potential fault types, for example, outputting "winding deformation: 85% confidence" or "tap changer contact failure: 32% confidence."

[0146] Through this deep analysis step based on the graph neural network model, complex data patterns that are difficult to discover with traditional rules can be discovered, thereby significantly improving the accuracy of fault diagnosis and the ability to identify complex faults.

[0147] In a preferred embodiment, the training of the abnormality prediction model includes:

[0148] Obtain a historical transformer fault case dataset; wherein the historical transformer fault case dataset includes several graph-structured training samples and corresponding fault type labels; the graph-structured training samples include an adjacency matrix extracted from the transformer knowledge graph and a feature matrix embedded with state feature parameters;

[0149] Dividing the transformer historical fault case dataset into several batches of training samples according to a preset batch size;

[0150] Each batch of training samples is sequentially input into the anomaly prediction model for iterative training until a preset training termination condition is reached; wherein, when the anomaly prediction model receives each batch of training samples, it outputs the predicted probability of each anomaly type in the preset anomaly type set based on the adjacency matrix and feature matrix in the current batch; a preset loss function is used to calculate and generate a loss function value based on the predicted probability and the corresponding fault type label; and a preset optimizer is used to update the parameters in the anomaly prediction model based on the loss function value.

[0151] In a specific embodiment of the present invention, in order to enable the abnormality prediction model to have the ability to identify fault modes from complex parameter associations, it is necessary to perform supervised learning training on it. The training process first needs to obtain a dataset of historical transformer fault cases. This dataset is the source of model learning experience, and each case is a verified real event, which includes a complete parameter snapshot at the time of the event and the exact fault type label finally obtained by manual confirmation or experimental report. In order to enable the model to understand the physical and logical associations between parameters, each case sample is not a flat vector, but is constructed into a training sample with a graph structure. Specifically, each sample includes an adjacency matrix related to the case extracted from the main knowledge graph to describe the topological relationship between nodes such as parameters and components; and a feature matrix, in which the quantized values ​​of the state feature parameters corresponding to each node are embedded.

[0152] During the training execution phase, in order to achieve efficient and stable learning, the transformer historical fault case dataset is first divided into several batches of training samples according to a preset batch size. Subsequently, the system sequentially inputs the training samples of each batch into the anomaly prediction model (preferably a graph neural network model) for iterative training until the loss function value of the model converges or reaches the preset training rounds and other training termination conditions. In a single iteration, the model receives a batch of graph structure training samples and calculates the adjacency matrix and feature matrix in the current batch based on its internal graph convolution or message passing mechanism, thereby outputting a predicted probability distribution covering all preset anomaly type sets. The system then uses a preset loss function (such as the cross entropy loss function) to quantify the gap between the predicted probability output by the model and the actual fault type label corresponding to the batch sample, and calculates a loss function value. Finally, a preset optimizer (such as the Adam optimizer) will update the trainable parameters in the anomaly prediction model through the backpropagation algorithm based on the loss function value, with the goal of making the prediction made by the model in the next iteration closer to the actual label.

[0153] Through the above-mentioned supervised training process, we finally obtained an anomaly prediction model that can deeply understand the topological correlation between parameters and accurately infer the probability of occurrence of different fault types from complex data patterns, greatly improving the intelligence level and accuracy of the diagnostic system.

[0154] Step S5: Determine the prediction probability of the corresponding candidate abnormality type based on the prediction probability of each abnormality type in the preset abnormality type set;

[0155] Step S6: For each candidate abnormality type, if the corresponding prediction probability is greater than a preset confidence threshold, it is determined that the transformer has the corresponding abnormality.

[0156] In a specific embodiment of the present invention, after the diagnostic rules have generated preliminary candidate anomaly types and the anomaly prediction model has output a probability distribution covering all pre-set fault types, the system performs a final fusion decision and judgment step. The core of this step is to use the deterministic conclusions of rule reasoning (i.e., the list of "candidate anomaly types") to guide and focus the interpretation of the probabilistic results output by the model, thereby avoiding the evaluation of a large number of low-probability, irrelevant faults and realizing the complementary advantages of the two diagnostic methods.

[0157] During specific execution, the system will first determine the specific predicted probability corresponding to each candidate abnormality type based on the predicted probability of each abnormality type in the preset abnormality type set. Then, for each candidate abnormality type, the system will compare the predicted probability obtained with a preset confidence threshold that may be different for different fault types. If the predicted probability of a candidate abnormality type is greater than its corresponding confidence threshold, the system will ultimately determine that the transformer has the corresponding abnormality, and can encapsulate this diagnostic conclusion, specific probability value, and related explainable fault chain information together as the final diagnostic result output. If the predicted probabilities of all candidate abnormality types fail to reach their respective confidence thresholds, the system will determine that the current evidence is insufficient to confirm any specific fault, and can mark the status as "to be observed" or output a corresponding prompt message.

[0158] This decision-making mechanism, which combines preliminary screening of rules with quantitative verification of models, effectively prevents possible misjudgments of a single model and ensures that the final diagnostic conclusion is based on the dual verification of logical reasoning and data patterns, thereby significantly improving the reliability and confidence of the diagnosis.

[0159] In a preferred embodiment, the transformer diagnosis method based on knowledge graph further includes:

[0160] After each new transformer fault type is obtained, the corresponding maintenance report is extracted, and the unstructured text in the maintenance report is input into the natural language processing model, so that the natural language processing model generates corresponding entity relationship triples based on the unstructured text in the maintenance report;

[0161] According to the transformer knowledge graph, the generated entity relationship triples are tested for knowledge conflicts. If no conflict is detected, the generated entity relationship triples are merged into the transformer knowledge graph; if a conflict is detected, the generated entity relationship triples are discarded.

[0162] Based on the above method embodiments, the present invention provides corresponding device embodiments.

[0163] like Figure 2 As shown, an embodiment of the present invention provides a transformer diagnosis device based on a knowledge graph, comprising: a transformer data acquisition module, a transformer preliminary state diagnosis module, a candidate abnormality type generation module, and a transformer final diagnosis module;

[0164] The transformer data acquisition module is used to obtain the test parameters of the transformer at the time of detection and the transformer knowledge map; the test parameters include: DC resistance value, insulation resistance value, low voltage short-circuit impedance value, winding deformation test data, dielectric loss factor, capacitance, and frequency domain dielectric spectrum; the transformer knowledge map includes the correlation between test parameters and diagnostic rules for fault diagnosis;

[0165] The transformer preliminary state diagnosis module is used to determine whether each test parameter is abnormal based on the correlation between the test parameters and the preset abnormality threshold. If it is determined that a test parameter and the other test parameters associated with it are all in an abnormal state, it is determined that the transformer has an abnormality to be confirmed; otherwise, it is determined that the transformer does not have an abnormality.

[0166] The candidate abnormality type generating module is used to perform fault diagnosis on each test parameter according to the diagnostic rules and generate several candidate abnormality types when there is an abnormality to be confirmed in the transformer;

[0167] The transformer final diagnosis module is used to extract the adjacency matrix and feature matrix of the test parameters from the transformer knowledge graph, and input the adjacency matrix and the feature matrix into a preset abnormality prediction model, so that the abnormality prediction model generates a prediction probability of each abnormality type in a preset abnormality type set through the adjacency matrix and the feature matrix; based on the prediction probability of each abnormality type in the preset abnormality type set, the prediction probability of the corresponding candidate abnormality type is determined; for each candidate abnormality type, if the corresponding prediction probability is greater than a preset confidence threshold, it is determined that the transformer has a corresponding abnormality.

[0168] It should be noted that the embodiment of the device described above corresponds to the above-mentioned embodiment of the present invention, and it can implement any of the transformer diagnosis methods based on knowledge graphs described above in the present invention. In addition, the embodiment of the above-mentioned device is merely illustrative, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the drawings of the embodiment of the device provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which can be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art can understand and implement it without making any creative effort.

[0169] Based on the above method embodiment of the present invention, a corresponding electronic device embodiment is provided.

[0170] An embodiment of the present invention provides an electronic device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the knowledge graph-based transformer diagnosis method according to any one of the present inventions is implemented, or when the processor executes the computer program, the functions of the modules in the above-mentioned device embodiments are implemented.

[0171] Exemplarily, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to implement the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program in the terminal device.

[0172] The terminal device may be a computing device such as a desktop computer, a notebook computer, a PDA, a cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0173] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the terminal device, connecting various parts of the entire terminal device using various interfaces and lines.

[0174] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function, etc.; the data storage area can store data created based on the use of the mobile phone, etc. In addition, the memory can include a high-speed random access memory and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0175] Based on the above method embodiment, the present invention provides a corresponding storage medium embodiment;

[0176] Another embodiment of the present invention provides a storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the storage medium is located is controlled to execute any one of the above-mentioned knowledge graph-based transformer diagnosis methods of the present invention.

[0177] The storage medium is a computer-readable storage medium, and the computer program includes computer program code, which may be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium may include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a mobile hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunications signal, and a software distribution medium.

[0178] In the description of this specification, the reference terms "one embodiment," "some embodiments," "example," "specific example," or "some examples" mean that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials, or characteristics described may be combined in any appropriate manner in any one or more embodiments or examples. In addition, those skilled in the art may combine and integrate different embodiments or examples described in this specification, as well as features of different embodiments or examples, unless they are mutually inconsistent.

[0179] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A transformer diagnosis method based on knowledge graph, characterized in that: include: Obtain the test parameters and transformer knowledge graph at the time of transformer detection; The test parameters include: DC resistance, insulation resistance, low voltage short-circuit impedance, winding deformation test data, dielectric loss factor, capacitance, and frequency domain dielectric spectrum; the transformer knowledge graph includes the correlation between test parameters and diagnostic rules for fault diagnosis; Based on the correlation between the test parameters and the preset abnormality threshold, it is determined whether each test parameter is abnormal. If it is determined that a test parameter and the rest of the test parameters associated with it are in an abnormal state, it is determined that the transformer has an abnormality to be confirmed; otherwise, it is determined that the transformer does not have an abnormality; In the case that the transformer has an abnormality to be confirmed, fault diagnosis is performed on each test parameter according to the diagnostic rules to generate several candidate abnormality types; Extracting an adjacency matrix and a feature matrix of test parameters from the transformer knowledge graph, and inputting the adjacency matrix and the feature matrix into a preset abnormality prediction model, so that the abnormality prediction model generates a prediction probability of each abnormality type in a preset abnormality type set through the adjacency matrix and the feature matrix; Determine the prediction probability of the corresponding candidate abnormality type based on the prediction probability of each abnormality type in the preset abnormality type set; For each candidate anomaly type, if the corresponding prediction probability is greater than the preset confidence threshold, it is determined that the transformer has the corresponding anomaly.

2. The transformer diagnosis method based on knowledge graph according to claim 1, characterized in that: The transformer knowledge graph is constructed through the following methods, including: Access transformer ontology libraries, structured diagnostic rules in transformer industry standards, and unstructured text from transformer historical maintenance reports; Based on the transformer ontology library and the preset physical mechanism model, the equipment entities, parameter entities, fault entities, and the relationships between entities in the transformer knowledge graph are defined to generate a knowledge graph for instantiated data to be imported; Using a preset natural language processing model, entity relationship triples are extracted from the unstructured text in the transformer historical maintenance report. The extracted entity relationship triples and the structured diagnostic rules in the transformer industry standard are imported into the knowledge graph of the instantiated data to be imported to generate a transformer knowledge graph.

3. The transformer diagnosis method based on knowledge graph according to claim 2, characterized in that: The training of the anomaly prediction model includes: Obtain a historical transformer fault case dataset; wherein the historical transformer fault case dataset includes several graph-structured training samples and corresponding fault type labels; the graph-structured training samples include an adjacency matrix extracted from the transformer knowledge graph and a feature matrix embedded with state feature parameters; Dividing the transformer historical fault case dataset into several batches of training samples according to a preset batch size; Each batch of training samples is sequentially input into the anomaly prediction model for iterative training until a preset training termination condition is reached; wherein, when the anomaly prediction model receives each batch of training samples, it outputs the predicted probability of each anomaly type in the preset anomaly type set based on the adjacency matrix and feature matrix in the current batch; a preset loss function is used to calculate and generate a loss function value based on the predicted probability and the corresponding fault type label; and a preset optimizer is used to update the parameters in the anomaly prediction model based on the loss function value.

4. The transformer diagnosis method based on knowledge graph according to claim 3 is characterized in that: The training of the natural language processing model includes: Obtain a text corpus dataset; wherein the text corpus dataset includes a number of unstructured texts from historical transformer maintenance reports and corresponding entity relationship triple labels; Divide the text corpus dataset into several batches of training samples according to the preset batch size; Each batch of training samples is input into the natural language processing model in sequence for iterative training until the preset training termination condition is reached; wherein, when the natural language processing model receives each batch of training samples, it outputs the predicted entity relationship triples based on the unstructured text in the current batch; through a preset loss function, the loss function value is calculated based on the predicted entity relationship triples and the corresponding entity relationship triple labels; and a preset optimizer is used to update the parameters in the natural language processing model according to the loss function value.

5. The transformer diagnosis method based on knowledge graph according to claim 4 is characterized in that: Also includes: After each new transformer fault type is obtained, the corresponding maintenance report is extracted, and the unstructured text in the maintenance report is input into the natural language processing model, so that the natural language processing model generates corresponding entity relationship triples based on the unstructured text in the maintenance report; According to the transformer knowledge graph, the generated entity relationship triples are tested for knowledge conflicts. If no conflicts are detected, the generated entity relationship triples are merged into the transformer knowledge graph. If a conflict is detected, the generated entity-relationship triples are discarded.

6. The transformer diagnosis method based on knowledge graph according to claim 5, characterized in that: The physical mechanism models include leakage current model, AC impedance model, mechanical resonance frequency model, Debye relaxation model, plate capacitance model and Cole-Cole model: The leakage current model is specifically: Where, I leak is the leakage current; V is the voltage applied across the insulating medium; R ins is the insulation resistance; The AC impedance model is specifically: Where Z is the impedance; R is the resistance of the circuit; f is the frequency of the AC power supply; L is the inductance of the circuit; The mechanical resonance frequency model is specifically: Where, f n is the natural resonant frequency of the system; k is the equivalent stiffness coefficient of the system; m is the equivalent mass of the system; The Debye relaxation model is specifically: Where tanδ is the dielectric loss factor; ∈″ is the imaginary part of the complex dielectric constant; ∈′ is the real part of the complex dielectric constant; The flat plate capacitor model is specifically: Where C is the capacitance; r is the relative dielectric constant of the insulating medium; ∈0 is the dielectric constant of vacuum; A is the equivalent area of ​​the electrode; d is the distance between the electrodes; The Cole-Cole model is specifically: Where,∈ * (ω) is the complex dielectric constant that varies with angular frequency; ∞ is the high-frequency limiting dielectric constant; Δ∈ is the dielectric strength increment; j is the imaginary unit; ω is the angular frequency; τ is the characteristic relaxation time; α is the empirical parameter describing the width of the relaxation time distribution.

7. The transformer diagnosis method based on knowledge graph according to claim 6, characterized in that: The anomaly threshold is calculated as follows: Obtain the operating temperature and relative humidity of the transformer at the time of detection; For each test parameter, a corresponding abnormal threshold is calculated according to the operating temperature and the ambient relative humidity.

8. A transformer diagnosis device based on knowledge graph, characterized in that: include: Transformer data acquisition module, transformer preliminary status diagnosis module, candidate anomaly type generation module and transformer final diagnosis module; The transformer data acquisition module is used to obtain the test parameters of the transformer at the time of detection and the transformer knowledge map; the test parameters include: DC resistance value, insulation resistance value, low voltage short-circuit impedance value, winding deformation test data, dielectric loss factor, capacitance, and frequency domain dielectric spectrum; the transformer knowledge map includes the correlation between test parameters and diagnostic rules for fault diagnosis; The transformer preliminary state diagnosis module is used to determine whether each test parameter is abnormal based on the correlation between the test parameters and the preset abnormality threshold. If it is determined that a test parameter and the other test parameters associated with it are all in an abnormal state, it is determined that the transformer has an abnormality to be confirmed; otherwise, it is determined that the transformer does not have an abnormality. The candidate abnormality type generating module is used to perform fault diagnosis on each test parameter according to the diagnostic rules and generate several candidate abnormality types when there is an abnormality to be confirmed in the transformer; The transformer final diagnosis module is used to extract the adjacency matrix and feature matrix of the test parameters from the transformer knowledge graph, and input the adjacency matrix and the feature matrix into a preset abnormality prediction model, so that the abnormality prediction model generates a prediction probability of each abnormality type in a preset abnormality type set through the adjacency matrix and the feature matrix; based on the prediction probability of each abnormality type in the preset abnormality type set, the prediction probability of the corresponding candidate abnormality type is determined; for each candidate abnormality type, if the corresponding prediction probability is greater than a preset confidence threshold, it is determined that the transformer has a corresponding abnormality.

9. An electronic device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the transformer diagnosis method based on the knowledge graph according to any one of claims 1 to 7 is implemented.

10. A storage medium, characterized in that: The storage medium includes a stored computer program, wherein when the computer program is running, the device where the storage medium is located is controlled to execute the transformer diagnosis method based on the knowledge graph according to any one of claims 1 to 7.

Images