A transformer insulation state diagnosis method, system, device and medium based on a knowledge graph

By using a knowledge graph-based dual-path long short-term memory encoder and a weighted fusion of membership features and boundary proximity features, the problem of identifying precursors to accelerated transformer insulation degradation was solved, thus improving the accuracy and reliability of the prediction model.

CN123069245APending Publication Date: 2026-11-10STATE GRID FUJIAN ELECTRIC POWER RES INST
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610981876.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-02
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies cannot effectively identify the degradation trend in the critical stage of accelerated transformer insulation degradation. Long Short-Term Memory Networks cannot learn gradient signals because their membership functions are always zero within the normal range. Furthermore, the mechanism diagnosis and prediction models lack real-time data channels, resulting in decreased prediction reliability.

Method used

A knowledge graph-based dual-path long short-term memory encoder is adopted. By calculating the contribution feature vector and neighbor feature vector of membership feature and boundary proximity feature, and combining the posterior probability of mechanism for weighted fusion, a cross-path gated connection is established, and the model is trained to identify the insulation degradation index.

Benefits of technology

It enables the identification of accelerated degradation trends before insulation condition indicators exceed the normal range, reduces the false alarm rate under conditions where multiple degradation mechanisms coexist, and ensures that the model can still accurately identify precursors of accelerated degradation after the insulation condition assessment parameters are updated.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069245A_ABST
    Figure CN123069245A_ABST
Patent Text Reader

Abstract

The application discloses a transformer insulation state diagnosis method, system, device and medium based on a knowledge graph, and belongs to the technical field of power equipment operation and maintenance, and comprises the following steps: respectively calculating membership degree features and boundary proximity features according to interval lower limit values to form contribution feature vectors and proximity feature vectors; extracting pre-activation event labels from the knowledge graph; constructing a double-channel long short-term memory encoder, and introducing a second encoding channel previous hidden state in first encoding channel input gate calculation; determining a fusion weight to perform double-channel hidden state weighted fusion according to a mechanism posterior probability maximum value; and training the encoder by taking an insulation deterioration index prediction loss and a pre-activation prediction loss as a joint objective function. The application separates and encodes the boundary proximity features through the double channels, and solves the problem that the gradient in the zero contribution interval is always zero, thereby making it impossible to learn the deterioration acceleration precursor.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of power equipment operation and maintenance technology, specifically to a method, system, equipment, and medium for diagnosing transformer insulation status based on knowledge graphs. Background Technology

[0002] Quantitative assessment and trend prediction of transformer insulation status rely on comprehensive analysis of multi-source heterogeneous data, including oil sample testing, equipment operation, and environmental data. Existing technologies generally employ knowledge graphs to semantically fuse this multi-source data and combine them with long short-term memory (LSM) networks to model historical time-series data of insulation degradation indices to predict future insulation degradation trends. The insulation degradation index maps each detection indicator to a membership function. When an indicator is within the normal range, the membership function output is always zero, and the partial derivative of the indicator with respect to the insulation degradation index within this range is always zero. This causes the LSM network to fail to obtain gradient signals from the historical dynamics of indicators within the normal range during backpropagation, resulting in a systematic deficiency in its ability to perceive precursors of accelerated insulation degradation. Simultaneously, there is a lack of real-time data pathways between the mechanism diagnosis results in the knowledge graph and the feature fusion process of the prediction model. The uncertainty of the mechanism diagnosis is not effectively incorporated into the prediction model, leading to decreased prediction reliability under complex operating conditions with multiple degradation mechanisms coexisting. Summary of the Invention

[0003] In view of the above-mentioned problems, the present invention provides a method, system, device and medium for diagnosing the insulation status of transformers based on knowledge graphs.

[0004] Therefore, the technical problem solved by this invention is: when the transformer insulation condition is in the critical stage of accelerated deterioration, the various insulation condition indicators have not yet exceeded the normal range threshold, but have been continuously approaching the threshold within the normal range. At this time, the insulation deterioration index is completely insensitive to the changes in indicators during this stage because the membership function output is always zero within the normal range. Therefore, the long short-term memory network that uses the insulation deterioration index as the training target cannot learn any effective patterns from the historical data of this stage.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a transformer insulation condition diagnosis method based on knowledge graphs, comprising, Membership features are calculated based on the lower bound of the interval in the insulation condition assessment parameters, and contribution feature vectors are constructed using the membership features; boundary proximity features are calculated based on the lower bound of the interval, and proximity feature vectors are constructed using the boundary proximity features. Based on the lower bound of the interval, extract pre-activated event tags from the knowledge graph that indicate the occurrence of interval crossings. A dual-path long short-term memory encoder is constructed, comprising a first encoding path with the contribution feature vector as input and a second encoding path with the neighboring feature vector as input; the preorder hidden state of the second encoding path is introduced into the input gate calculation of the first encoding path to establish a cross-path gating connection; The mechanism posterior probability is obtained from the knowledge graph, and the fusion weight is determined based on the maximum value of the mechanism posterior probability. The hidden state sequences of the first coding path and the second coding path are weighted and fused step by step to generate a fused coding sequence. Using the fused encoding sequence as input, the dual-path long short-term memory encoder is trained by using the insulation degradation index prediction loss and the pre-activation prediction loss calculated based on the pre-activation event label as a joint objective function, and outputs the insulation degradation index prediction result.

[0006] As a preferred embodiment of the knowledge graph-based transformer insulation state diagnosis method of the present invention, the step of calculating the membership feature based on the lower bound of the interval includes: dividing the insulation state assessment range into a zero contribution interval, an effective contribution interval, and a full contribution interval based on the lower bound of the interval and the upper bound of the interval in the insulation state assessment parameters; when in the zero contribution interval, the membership feature takes a zero value; when in the effective contribution interval, the membership feature takes a linear interpolation between the lower bound of the interval and the upper bound of the interval; and when in the full contribution interval, the membership feature takes a unit value. The calculation of the boundary proximity feature based on the lower bound of the interval includes: when the boundary proximity feature is in the zero contribution interval, the boundary proximity feature is taken as a linear interpolation between the lower bound of the interval and the zero value; when the boundary proximity feature is in the effective contribution interval or the full contribution interval, the boundary proximity feature is taken as zero.

[0007] As a preferred embodiment of the transformer insulation state diagnosis method based on knowledge graphs according to the present invention, the step of extracting pre-activation event tags of index interval crossings from the knowledge graph based on the lower bound value of the interval includes: Using each time step within the prediction time window as a detection unit, in response to the normalized value of the insulation status index at the current time step exceeding the lower bound of the interval and the normalized value of the insulation status index at the previous time step not exceeding the lower bound of the interval, it is determined that the insulation status index has crossed the interval. When the insulation status index is determined to have crossed a range within the prediction time window, the pre-activation event label takes a positive value; when the insulation status index is determined not to have crossed a range within the prediction time window, the pre-activation event label takes a negative value.

[0008] As a preferred embodiment of the knowledge graph-based transformer insulation state diagnosis method of the present invention, the step of introducing the pre-order hidden state of the second encoding path into the input gate calculation of the first encoding path to establish a cross-path gating connection includes: The preceding hidden state of the second coding path is linearly transformed and superimposed on the calculation term of the input gate of the first coding path. The weight matrix of the linear transformation is used as a learnable parameter in the training. The cross-path gating connection acts on the input gate of the first coding path, and modulates the proportion of the contribution feature vector received by the memory unit of the first coding path using the boundary proximity feature.

[0009] As a preferred embodiment of the knowledge graph-based transformer insulation state diagnosis method of the present invention, the step of determining the fusion weight based on the maximum value of the posterior probability of the mechanism, and performing weighted fusion on the hidden state sequences of the first encoding path and the second encoding path step by step to generate a fused encoding sequence includes: The maximum value among the posterior probabilities of thermal aging mechanism, electrical aging mechanism, and chemical aging mechanism at the predicted start time is determined as the first fusion weight, and the second fusion weight is determined by the constraint that the sum of the first fusion weight and the second fusion weight is one. The hidden state sequence of the first coding path is weighted according to the first fusion weight, and the hidden state sequence of the second coding path is weighted according to the second fusion weight. The weighted results of each coding step are summed to generate the fused coding sequence.

[0010] As a preferred embodiment of the knowledge graph-based transformer insulation state diagnosis method of the present invention, wherein: the pre-activation prediction loss calculated based on the pre-activation event tags includes: The hidden state of the last encoding step in the second encoding path is input to the learnable linear layer, and the predicted probability of the insulation state index being crossed within the predicted time window is output. The predicted probability and the pre-activated event label are used to calculate the binary cross-entropy, which provides a supervisory gradient for the second coding path through backpropagation.

[0011] As a preferred embodiment of the knowledge graph-based transformer insulation status diagnosis method of the present invention, wherein: in response to the number of newly added operation and maintenance confirmation records in the knowledge graph reaching a preset number, the lower limit value of the interval, the upper limit value of the interval, and the weight parameters in the insulation status assessment parameters are recalculated; The contribution feature vector and the neighboring feature vector are recalculated based on the recalculated insulation state assessment parameters. The pre-activated event label is re-extracted from the knowledge graph based on the recalculated lower bound of the interval. The dual-path long short-term memory encoder is updated and trained using the recalculated contribution feature vector, the neighboring feature vector, and the re-extracted pre-activated event label.

[0012] This invention provides a transformer insulation condition diagnosis system based on knowledge graphs.

[0013] To solve the above technical problems, the present invention provides the following technical solution: a transformer insulation state diagnosis system based on knowledge graph, comprising: a feature construction module, used to calculate membership features based on the lower bound of the interval in the insulation state assessment parameters to form a contribution feature vector, and to calculate boundary proximity features based on the lower bound of the interval to form a proximity feature vector; The tag extraction module is used to extract pre-activated event tags from the knowledge graph based on the lower bound of the interval; A dual-path encoding module is used to construct a dual-path long short-term memory encoder. The dual-path long short-term memory encoder includes a first encoding path with the contribution feature vector as input and a second encoding path with the neighboring feature vector as input. In the input gate calculation of the first encoding path, the preorder hidden state of the second encoding path is introduced to establish a cross-path gating connection. The fusion coding module is used to obtain the posterior probability of the mechanism from the knowledge graph, determine the fusion weight based on the maximum value of the posterior probability of the mechanism, and perform weighted fusion on the hidden state sequences of the first coding path and the second coding path step by step to generate a fusion coding sequence. The training prediction module is used to train the dual-path long short-term memory encoder by taking the fused coding sequence as input, using the insulation degradation index prediction loss and the pre-activation prediction loss calculated based on the pre-activation event label as a joint objective function, and outputting the insulation degradation index prediction result.

[0014] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the knowledge graph-based transformer insulation state diagnosis method.

[0015] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the knowledge graph-based transformer insulation state diagnosis method.

[0016] The beneficial effects of this invention are as follows: This invention separates the membership feature and boundary proximity feature of the insulation state index into two independent input paths. The second encoding path can still obtain a non-zero gradient signal when the insulation state index is in the zero contribution range. This solves the problem that the precursory patterns cannot be learned in the existing methods because the gradient of the insulation degradation index is always zero in the zero contribution range. This allows the prediction model to identify the deterioration acceleration trend before the insulation state index enters the effective contribution range.

[0017] The maximum posterior probability of the mechanism is used as the first fusion weight. When the posterior probabilities of the three mechanisms tend to be equal, the fusion weight of the second coding path is twice that of the first coding path. This ensures that when the mechanism identification result is uncertain, the boundary proximity features occupy an effective proportion in the fusion coding sequence, and reduces the false alarm rate caused by the suppression of precursor signals under the condition of multiple deterioration mechanisms coexisting.

[0018] A linkage mechanism was established between the update of insulation condition assessment parameters, the recalculation of pre-activated event tags, and the update of model parameters. This eliminated the inconsistency between the lower bound of the interval and the judgment threshold of the pre-activated event tags after the update of the insulation condition assessment parameters. This enabled the model to accurately identify the accelerated degradation precursors corresponding to the lower bound of the current version interval even when the insulation condition assessment parameters are continuously adjusted with the accumulation of operation and maintenance data. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a knowledge graph-based method for diagnosing the insulation status of a transformer, as provided in one embodiment of the present invention.

[0021] Figure 2 The flowchart illustrates the construction of a dual-path long short-term memory encoder for a knowledge graph-based transformer insulation state diagnosis method, as provided in one embodiment of the present invention.

[0022] Figure 3 The flowchart illustrates the generation of a fusion coding sequence for a knowledge graph-based transformer insulation state diagnosis method, as provided in one embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of the basic structure of a knowledge graph triplet for a transformer insulation state diagnosis method based on knowledge graph, provided in one embodiment of the present invention.

[0024] Figure 5This is a schematic diagram illustrating the insulation degradation index calculation and grade classification process of a knowledge graph-based transformer insulation condition diagnosis method, as provided in one embodiment of the present invention. Detailed Implementation

[0025] To make the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0026] Example 1, referring to Figures 1-5 This is one embodiment of the present invention, which provides a transformer insulation condition diagnosis method based on knowledge graphs, including: S1. Calculate the membership feature based on the lower bound of the interval in the insulation condition assessment parameters, and construct the contribution feature vector using the membership feature; calculate the boundary proximity feature based on the lower bound of the interval, and construct the proximity feature vector using the boundary proximity feature.

[0027] Before performing step S1, some embodiments further include a knowledge graph construction step: based on the transformer equipment's oil sample testing data, operating status data, and maintenance record data, construct equipment entity nodes, oil sample testing entity nodes, maintenance entity nodes, and formula version nodes in the knowledge graph. Equipment entity nodes store the transformer equipment's nameplate parameters and operating status attributes; oil sample testing entity nodes are connected to equipment entity nodes through association relationships, permanently storing the measured values ​​of seven insulation status indicators for each oil sample test; maintenance entity nodes store the timestamps and operation types of maintenance operations; formula version nodes store insulation status evaluation parameters, including interval lower bound values, interval upper bound values, and weight parameters, and record the historical evolution of the parameters through version chain relationships.

[0028] like Figure 4 As shown, the information in the knowledge graph is organized in the form of entity-relation-entity triples. Each entity node stores corresponding attribute information, and entity nodes are connected through relations. In this invention, equipment entity nodes, oil sample testing entity nodes, operation and maintenance entity nodes, and formula version nodes are four types of entity nodes, which are interconnected through association relations and version chain relations to form the basic knowledge graph structure that supports contribution feature vector calculation, pre-activation event label extraction, and mechanism posterior probability reading.

[0029] It should be noted that the seven insulation condition indicators are acetylene content, moisture content, breakdown voltage, dielectric loss, top oil temperature, load fluctuation coefficient, and ambient temperature. The load fluctuation coefficient is the ratio of the difference between the maximum and minimum loads within the statistical period to the average load. The lower limit of the interval is derived from the normalized upper limit of the normal range for the seven insulation condition indicators, and the upper limit is derived from the normalized lower limit of the severely deteriorated range for the seven insulation condition indicators. The normalization method is maximum-minimum normalization. It should also be noted that the breakdown voltage deteriorates in the opposite direction to the other six insulation condition indicators; a lower breakdown voltage indicates a more severe degree of insulation deterioration. Therefore, the lower limit of the breakdown voltage interval corresponds to the normalized upper limit of the severely deteriorated range, and the upper limit corresponds to the normalized lower limit of the normal range.

[0030] The weight parameters are calculated using the PCA-entropy weighting method. Understandably, the PCA-entropy weighting method first constructs a covariance matrix from the historical normalized data of the seven insulation state indicators, performs eigenvalue decomposition on the covariance matrix, extracts the loading coefficients of each insulation state indicator along the principal component directions, and calculates the initial objective weights of each insulation state indicator by multiplying the variance contribution rate of each principal component by the corresponding loading coefficient. Further, information entropy is calculated for the loading coefficients of the seven insulation state indicators, expressed as: in, Indicators of insulation status Information entropy; Indicators of insulation status In the The normalized probability values ​​in each observation sample are derived from the insulation state index. The Each observed value divided by the insulation condition index It is calculated by summing all observations; Represents the natural logarithm; Indicates the number of observed samples; Indicates the observation sample number; Indicates the insulation condition index number, .

[0031] Furthermore, with As an indicator of insulation status The information utility value is normalized by multiplying the information utility value and the principal component loading coefficients to obtain the weight parameters. ,satisfy ,in Indicators of insulation status The weight parameters, Indicates the insulation condition index number, By using the PCA-entropy weighting method, both the objective information content of the insulation condition indicators and the variance contribution in the principal component direction are taken into account. This ensures that the determination of the weight parameters reflects the information richness of each insulation condition indicator in historical data and satisfies the ability to distinguish insulation degradation states.

[0032] In some embodiments, step S1 may include steps S11 to S13, wherein: S11. Based on the lower limit of the interval and the upper limit of the interval in the insulation condition assessment parameters, the insulation condition assessment range is divided into a zero contribution interval, an effective contribution interval, and a full contribution interval.

[0033] Understandably, insulation condition indicators normalized value and the lower bound of the interval and the upper bound of the interval The numerical comparison relationship determines the interval assignment: Insulation status index In the zero contribution range, Insulation status index Within the effective contribution range, Insulation status index Within the full contribution range, and The insulation status evaluation parameters from the current version of the formula version node meet the requirements. .in, Indicators of insulation status The normalized value; Indicators of insulation status The lower bound of the interval; Indicators of insulation status The upper bound of the interval; Indicates the insulation condition index number, .

[0034] It should be noted that the breakdown voltage The direction of degradation is opposite to that of the other six insulation condition indicators, and the direction of the interval assignment criteria is also opposite: When the breakdown voltage is in the zero contribution range, The breakdown voltage is within the effective contribution range. The breakdown voltage is in the full contribution range, where The normalized value corresponding to the upper limit of the severely degraded breakdown voltage range. The normalized value corresponding to the lower limit of the normal range of breakdown voltage.

[0035] S12: When the membership feature is in the zero contribution interval, it takes a zero value; when the membership feature is in the effective contribution interval, it takes a linear interpolation between the lower and upper bounds of the interval; and when the membership feature is in the full contribution interval, it takes a unit value.

[0036] It is understandable that the interval assignment results based on step S11 are used to quantitatively characterize insulation state indicators. The insulation condition index is defined to reflect the current contribution of the insulation degradation index. Membership characteristics For acetylene content, moisture content, dielectric loss value, top oil temperature, load fluctuation coefficient, and ambient temperature, The membership degree feature is calculated using a monotonically increasing type, and is expressed as follows: in, Indicators of insulation status The membership feature, with values ​​ranging from 1 to 10. ; Indicators of insulation status The normalized value; Indicators of insulation status The lower bound of the interval; Indicators of insulation status The upper bound of the interval; Indicates the insulation condition index number.

[0037] Furthermore, regarding the breakdown voltage Based on the explanation that the breakdown voltage range attribution criteria in step S11 are in the opposite direction, the membership feature of the monotonically decreasing type is used for calculation, which is expressed as: in, The membership characteristic representing the breakdown voltage, with a value range of [value range missing]. ; This represents the normalized value of the breakdown voltage. This represents the lower bound of the breakdown voltage range, corresponding to the normalized upper bound of the severely degraded range. This represents the upper limit of the breakdown voltage range, corresponding to the normalized lower limit of the normal range.

[0038] by The contribution feature vector is formed by arranging the insulation state index numbers. ,in This indicates transpose.

[0039] Furthermore, using the contributing feature vector The insulation degradation index is obtained by weighting and summing the components with weight parameters, and is expressed as: in, This represents the insulation degradation index, with a value range of [value range missing]. ; Indicators of insulation status The weight parameters are derived from the insulation state evaluation parameters of the current version of the formula version node; Indicators of insulation status Membership degree characteristic; Indicates the insulation condition index number, The insulation degradation index is calculated by the formula version node based on the current version's insulation status assessment parameters and stored in the device entity node. IDI The insulation condition is normal at this time. At this time, it is only a slight deterioration. At that time, it was at a moderate level of degradation. At that time, it was severely degraded.

[0040] like Figure 5 As shown, the calculation and grading process of the insulation degradation index is as follows: Historical measured values ​​of seven insulation status indicators are read from the knowledge graph oil sample detection entity node and equipment entity node. The membership characteristics of each insulation status indicator are calculated according to steps S11 to S12. The insulation degradation index is obtained by weighted summation of the components of the contributing feature vector and the weight parameters. Grading is performed based on the insulation degradation index value. The grading results are calculated by the formula version node based on the current version insulation status assessment parameters and stored in the equipment entity node for use in the mechanism posterior probability fusion in step S4 and the training label reading in step S52.

[0041] S13, when in the zero contribution interval, the boundary-adjacent feature takes the linear interpolation between the lower bound of the interval and the zero value; when in the effective contribution interval or the full contribution interval, the boundary-adjacent feature takes the zero value.

[0042] It is understandable that, based on the interval assignment results of step S11, the insulation status index of the zero contribution interval meets the requirements. .when equal At that time, insulation status index It is located at the critical position between the zero contribution interval and the effective contribution interval, and its membership characteristic is... The value changes from zero to non-zero. This is for quantitative characterization. Within the zero contribution interval The degree of closeness is based on the lower bound of the interval in step S12. Definition of insulation state index The boundary proximity feature is represented as: in, Indicators of insulation status The boundary proximity feature, with values ​​ranging from 1 to 10. ; Indicators of insulation status The lower bound of the interval; Indicators of insulation status The normalized value; Represents a positive number, with values ​​ranging from 1 to 2. ,prevent When the denominator is zero, the fraction is zero. This indicates the insulation condition index number. When... hour This indicates the insulation condition index. It is at the critical position between the zero contribution interval and the effective contribution interval; when hour, Take a positive value, and and The larger the difference, The larger; when hour A value of zero indicates the insulation condition index. It has entered the effective contribution range or the full contribution range.

[0043] Furthermore, regarding the breakdown voltage ( Based on the explanation that the breakdown voltage range attribution criteria in step S11 are in the opposite direction, its zero contribution range is... The boundary proximity feature is calculated and represented as follows: in, The boundary proximity characteristic of the breakdown voltage is represented, and its value range is: ; This represents the upper limit of the breakdown voltage range, corresponding to the normalized lower limit of the normal range. This represents the normalized value of the breakdown voltage. Represents the normalized range Inside The distance to the upper bound of 1 is used as the normalization benchmark for the boundary proximity feature, and is compared with the positive index boundary proximity feature formula. The computational logic for the normalized benchmark is symmetrical; Represents a positive number, with values ​​ranging from 1 to 2. .when hour This indicates that the breakdown voltage is at the critical position between the zero contribution region and the effective contribution region; when hour, , Take a positive value, and and The larger the difference, The larger; when hour A value of zero indicates that the breakdown voltage has entered the effective contribution range or the full contribution range.

[0044] by Arranged by insulation condition index number to form a nearest neighbor feature vector ,in This indicates transpose.

[0045] It should be noted that for each time step within the historical training window Based on the membership feature calculation method described above, the seven insulation state indices were calculated at time step [number missing]. Membership characteristics ,by Time steps are arranged according to the insulation condition index sequence number. Contribution feature vector ,in Indicates time step Insulation status indicators Membership degree feature Indicates the month sequence number of the historical time step, superscript This indicates transposition. Similarly, based on the boundary proximity feature calculation method described above, the seven insulation state indices are calculated at time step [number missing]. Boundary proximity features ,by Time steps are arranged according to the insulation condition index sequence number. Neighboring feature vectors ,in Indicates time step Insulation status indicators The boundary proximity feature.

[0046] In step S1, the current state of the insulation state index in the zero contribution interval is extracted independently in the form of boundary proximity features, thus realizing the separate expression of contribution feature vector and proximity feature vector.

[0047] S2, based on the lower bound of the interval, extracts pre-activated event tags from the knowledge graph when the index crosses the interval.

[0048] In some embodiments, step S2 includes steps S21 to S22, wherein: S21, taking each time step within the prediction time window as a detection unit, in response to the insulation status index normalized value of the current time step exceeding the lower limit of the interval and the insulation status index normalized value of the previous time step not exceeding the lower limit of the interval, it is determined that the insulation status index has crossed the interval.

[0049] It is understandable that the prediction time window is the start time of the prediction. The following series Each time step, of which Indicates the predicted start time. This indicates the number of steps in the prediction time window, with a value of 3, in months. Insulation status index At time step normalized value With time step normalized value The numerical comparison relationship determines the interval crossing: when and At that time, determine the insulation status index At time step A section crossing occurred.

[0050] in, Indicators of insulation status At time step The normalized value; Indicators of insulation status At time step The normalized value; Indicators of insulation status The lower bound of the interval is derived from the insulation state evaluation parameters of the current version of the formula version node; Indicates the prediction step number. ; Indicates the insulation condition index number, ; Indicates the start time of the prediction.

[0051] It should be noted that the normalized value and After accurately retrieving historical measured values ​​for the corresponding time steps from the oil sample detection entity nodes and equipment entity nodes in the knowledge graph, the results are calculated using the maximum-minimum normalization method in step S11. The training set requires all training samples to satisfy... ,in This indicates the latest historical observation month, ensuring that the normalized values ​​for all time steps within the prediction time window are accurately readable within the historical observation period.

[0052] It should be noted that the training samples are constructed by sliding month by month, that is... Take the satisfaction All month serial numbers, total number of training samples It equals the number of month numbers that satisfy the constraints.

[0053] Furthermore, regarding the breakdown voltage ( Based on the explanation in step S11 that the determination conditions for the breakdown voltage range are in opposite directions, the determination conditions for the range crossing are in opposite directions: when and At that time, the breakdown voltage is determined at time step An interval crossing occurred, in which This represents the upper limit of the breakdown voltage range.

[0054] S22, when the insulation status index is determined to have crossed intervals within the prediction time window, the pre-activation event label takes a positive value; when the insulation status index is determined not to have crossed intervals within the prediction time window, the pre-activation event label takes a negative value.

[0055] It is understandable that insulation condition indicators are defined. At the predicted start time Corresponding pre-activation event tag As follows: When the prediction time window Memory at time step When the interval crossing determination condition in step S21 is met, Take a positive example value, with a positive example value of 1; if there is no time step within the prediction time window that meets the interval crossing judgment condition, Take negative examples, and take 0 for negative examples.

[0056] in, Indicates the predicted sample Corresponding insulation condition index The pre-activation event tag has a value of 1 or 0; Indicates the predicted sample number; Indicates the insulation condition index number, ; This indicates the number of steps within the prediction time window. .

[0057] by Pre-activation event tag vectors are arranged according to insulation status index numbers. ,in This indicates transpose.

[0058] The pre-activated event label vector is calculated by accurately reading historical measured values ​​from the knowledge graph's oil sample detection entity node and equipment entity node, without the need for additional manual annotation.

[0059] It should be noted that when the lower bound of the interval in the formula version node... When an update occurs, steps S21 to S22 are re-executed for all training samples using the updated lower bound of the interval, and the pre-activated event label vector is recalculated for the entire dataset to ensure that the pre-activated event label vector is calculated based on the insulation state evaluation parameters of the current version.

[0060] S3. Construct a dual-path long short-term memory encoder, which includes a first encoding path with contribution feature vectors as input and a second encoding path with neighboring feature vectors as input. In the input gate calculation of the first encoding path, the preorder hidden state of the second encoding path is introduced to establish a cross-path gating connection.

[0061] It is understandable that the coding step in step S3 Corresponding historical month number Contribution feature vector at location and neighboring feature vectors Take the values ​​defined in step S1 respectively and exist The value at which, Indicates coding step The corresponding historical month number, ; Indicates the predicted start time; Indicates the encoding step number. .

[0062] In some embodiments, step S3 specifically includes steps S31 to S32, wherein: S31, the preceding hidden state of the second coding path is linearly transformed and superimposed on the calculation term of the input gate of the first coding path, and the weight matrix of the linear transformation is used as a learnable parameter in the training.

[0063] The first encoding pathway is a standard long short-term memory network structure, using time steps. Contribution feature vector With time step Derivative time series features spliced ​​vector For input, ,in Indicates time step The contribution feature vector; Indicates time step The derived time-series features consist of 13 items; Indicates transpose; Indicates coding step The corresponding historical month number; Indicates the encoding step number. .

[0064] It should be noted that derived time series features Normalized values ​​of insulation condition index For computational foundation, six types of time-series statistical features were calculated for the normalized value sequences of seven insulation state indicators within the historical training window: first-order difference, moving average, standard deviation, coefficient of variation, maximum increase, and number of mutations, forming the initial feature pool. The first-order difference is the difference between normalized values ​​of adjacent time steps; the moving average is the arithmetic mean of the normalized values ​​within the sliding window; the standard deviation is the standard deviation of the normalized values ​​within the sliding window; the coefficient of variation is the ratio of the standard deviation to the moving average; the maximum increase is the maximum single-step increment of the normalized value within the sliding window; and the number of mutations is the number of times the single-step change of the normalized value within the sliding window exceeds a preset threshold, which is set to 0.05. Using root mean square error as the evaluation metric, the initial feature pool was sorted by feature importance according to a recursive feature elimination method. Features with the lowest importance were repeatedly eliminated until the number of features was reduced to 13. The 13 selected derived time-series features are denoted as follows: Recursive feature elimination is performed once before initial training, and the 13 derived temporal features selected remain unchanged during the training update in step S53.

[0065] Furthermore, the input gate calculation for the first coding path is expressed as: in, Indicates coding step The input gate activation vector of the first coding path; This represents the sigmoid activation function; This represents the weight matrix of the input gate of the first coding path, with dimensions 64×84, where 84 is the weight of the input gate. Dimension 64 and The sum of 20 dimensions represents the learnable parameters; Indicates coding step The hidden state vector of the first coding path has a dimension of 64 and an initial value of... It should be noted that the number of units in the Long Short-Term Memory network in the first coding pathway is 64. Indicates coding step The input vector of the first encoding path; The bias vector representing the input gate of the first coding path is a learnable parameter. represents the cross-path pre-activation gating weight matrix, which is a learnable parameter initialized using Xavier uniform initialization; Indicates coding step The hidden state vector of the second coding path, initial value ; This represents a vector concatenation operation; Indicates the encoding step number. .

[0066] Furthermore, the remaining gating calculations for the first coding path are expressed as follows: in, Indicates coding step First encoding path forget gate activation vector; Indicates coding step First encoding path candidate memory unit vector; Indicates coding step First encoding path memory unit vector; Indicates coding step First encoding path memory unit vector, initial value ; Indicates coding step The first encoding path outputs the gate activation vector; Indicates coding step First encoding path hidden state vector; Let represent the weight matrices of the forget gate, candidate memory, and output gate in the first encoding pathway, respectively. Each matrix has a dimension of 64×84, where 84 represents the maximum value. Dimension 64 and The sum of 20 dimensions consists of learnable parameters; These represent the bias vectors of the forget gate, candidate memory, and output gate of the first coding path, respectively, and are all learnable parameters; Represents the hyperbolic tangent activation function; This represents element-wise product.

[0067] S32, the cross-path gating connection acts on the input gate of the first coding path, and modulates the proportion of the contribution feature vector received by the memory unit of the first coding path based on the boundary proximity feature.

[0068] Understandable, The second encoding path is in the encoding step The accumulated temporal information of boundary proximity features is introduced into the calculation of the input gate of the first coding path, making the activation vector of the input gate of the first coding path... Subject to the hidden state of the second coding path Modulation. When hour, Cross-path pre-activation gating item As a zero vector, the input gate computation of the first encoding path degenerates into the form of a standard long short-term memory network; with each encoding step... Incrementing from 1 to 12, the hidden state of the second coding path Continuously accumulate neighboring feature vectors Time sequence information, The modulation effect on the input gate of the first coding path is then established. It should be noted that the second coding path uses neighboring feature vectors... As input, standard long short-term memory network gating computation is performed, represented as: in, Indicates coding step The activation vector of the forget gate in the second encoding path; Indicates coding step The second encoding path input gate activation vector; Indicates coding step Second coding path candidate memory unit vector; Indicates coding step Second encoding path memory unit vector; Indicates coding step Second encoding path memory unit vector, initial value ; Indicates coding step The second encoding path outputs the gate activation vector; Indicates coding step The hidden state vector of the second coding path has a dimension of 64; , Let represent the weight matrices of the forget gate, input gate, candidate memory, and output gate of the second encoding pathway, respectively. Each matrix has a dimension of 64×71, where 71 represents the minimum value of a given value. Dimension 64 and The sum of the seven dimensions consists of learnable parameters; These represent the bias vectors of the forget gate, input gate, candidate memory, and output gate of the second coding path, respectively, and are all learnable parameters; Indicates time step The nearest feature vectors; Indicates the encoding step number. The number of units in the long short-term memory network in the second coding pathway is 64.

[0069] It should be noted that the cross-path gating connection applies to the input gate of the first encoding path, but not to the forget gate and output gate of the first encoding path. The reason is that the input gate... Control coding steps The proportion of new input information written into the memory unit, and the timing information of the adjacent feature vector accumulated by the second coding path are modulated through the input gate, so that the memory unit of the first coding path increases the proportion of receiving the contributing feature vector in advance when the normalized value of the insulation state index approaches the lower limit of the interval, thus simulating the physical prior relationship between the precursor signal and the main deterioration signal of transformer insulation degradation.

[0070] S4. Obtain the posterior probability of the mechanism from the knowledge graph, determine the fusion weight based on the maximum value of the posterior probability of the mechanism, and perform weighted fusion on the hidden state sequences of the first coding path and the second coding path step by step to generate the fused coding sequence.

[0071] In some embodiments, step S4 includes steps S41 to S42, wherein: S41, the maximum value among the posterior probabilities of thermal aging mechanism, electrical aging mechanism and chemical aging mechanism at the predicted start time is determined as the first fusion weight, and the second fusion weight is determined by the constraint that the sum of the first fusion weight and the second fusion weight is one.

[0072] Understandably, the posterior probability of the mechanism is stored in the device entity node by the mechanism identification module in the knowledge graph after each diagnosis, including the posterior probability of the thermal aging mechanism. Posterior probability of electro-aging mechanism and the posterior probability of chemical aging mechanism ,satisfy ,in This represents the posterior probability of the thermal aging mechanism. This represents the posterior probability of the electro-aging mechanism. This represents the posterior probability of the chemical aging mechanism.

[0073] Furthermore, to predict the start time The maximum value among the posterior probabilities of thermal aging mechanism, electrical aging mechanism, and chemical aging mechanism is determined as the first fusion weight, expressed as: in, Indicates the first fusion weight; Indicates the start time of the prediction Posterior probability of thermal aging mechanism; Indicates the start time of the prediction Posterior probability of the electro-aging mechanism; Indicates the start time of the prediction Posterior probability of chemical aging mechanism; Indicates the predicted start time; Indicates the predicted sample number; This indicates the operation of retrieving the maximum value.

[0074] It should be noted that, The range of values ​​is :when hour, Take the minimum value This indicates that the posterior probabilities of the three mechanisms are equal; when the posterior probability of a certain mechanism is 1, Taking the maximum value of 1 indicates that the mechanism identification result is completely determined. The second fusion weight is determined by the constraint that the sum of the first fusion weight and the second fusion weight is one, that is, the second fusion weight is... .

[0075] S42, the hidden state sequence of the first coding path is weighted according to the first fusion weight, the hidden state sequence of the second coding path is weighted according to the second fusion weight, and the weighted results of each coding step are summed to generate a fused coding sequence.

[0076] Understandably, with the first fusion weight Second fusion weight Weighted fusion is performed stepwise on the hidden states of all coding steps in the first and second coding paths, defining the coding steps. The fusion-encoded hidden state is represented as: in, This represents the fused encoded hidden state vector, which has a dimension of 32. Indicates the first fusion weight; This represents the projection matrix from the first encoding path to the fusion layer, with a dimension of 32×64, and is a learnable parameter. Represents the hidden state vector of the first coding path; This represents the projection matrix from the second coding path to the fusion layer, with a dimension of 32×64, and is a learnable parameter. Represents the hidden state vector of the second coding path; Indicates the encoding step number. ; This indicates the predicted sample number.

[0077] It should be noted that, Only depends on the predicted start time The posterior probability of the mechanism for predicting samples All 12 coding steps take the same value. At that time, the fusion coding hidden state The fusion result is entirely determined by the hidden state of the first coding path; when At that time, the fusion coding hidden state The fusion weight of the hidden state in the second coding path is twice that of the first coding path. According to the coding step number The arrangement forms a fused coding sequence.

[0078] S5 takes the fused coding sequence as input, and uses the insulation degradation index prediction loss and the pre-activation prediction loss calculated based on the pre-activation event label as the joint objective function to train the dual-path long short-term memory encoder, and outputs the insulation degradation index prediction result.

[0079] In some embodiments, step S5 includes steps S51 to S52, wherein: S51, input the hidden state of the last coding step of the coding sequence in the second coding path to the learnable linear layer, and output the predicted probability of the insulation state index crossing the interval within the prediction time window.

[0080] It is understandable that the fused coding sequence Input to an additive attention layer to fuse encoded hidden states. With decoder hidden state Calculate decoding steps For the coding step Attention score: in, Indicates attention score; This represents the attention query vector, which has a dimension of 16 and consists of learnable parameters. Represents the hyperbolic tangent activation function; The projection matrix representing the fused encoded hidden state has a dimension of 16×32 and is a learnable parameter; This represents the fused encoded hidden state vector; The projection matrix represents the hidden state of the decoder, with dimensions of 16×32, and consists of learnable parameters; This represents the decoder's hidden state vector; Indicates the decoding step number. ; Indicates the encoding step number. .

[0081] Furthermore, the attention scores are normalized to obtain the attention weights, which are expressed as: in, Represents the normalized attention weights, satisfying ; Indicates attention score; This indicates a dummy element in the summation. ; Indicates the decoding step number; This indicates the encoding step number.

[0082] Furthermore, the decoding step is obtained by fusing the encoded sequences using attention weights. The context vector is represented as: in, Represents the context vector; This represents the normalized attention weights; This represents the fused encoded hidden state vector; Indicates the decoding step number. ; Indicates the encoding step number. .

[0083] The decoder uses context vectors The input is a single-layer long short-term memory network, which outputs the hidden state of the decoder. The decoder has 32 units in its Long Short-Term Memory network. The dimension is 32, and the decoding step is obtained through a linear output layer. The corresponding predicted insulation degradation index is expressed as: in, This represents the predicted value of the insulation degradation index; This represents the output layer weight matrix, which has a dimension of 1×32 and consists of learnable parameters. This represents the decoder's hidden state vector; This represents the output layer bias, which is a learnable parameter. Indicates the decoding step number. ; Indicates the predicted sample number; Indicates the predicted start time; This represents the initial value of the decoder's hidden state. .

[0084] Furthermore, the hidden state of the last coding step in the second coding path Input to a learnable linear layer, output insulation state index The predicted probability of an interval crossing occurring within the prediction time window is expressed as: in, Indicates the predicted probability; This represents the sigmoid activation function; This represents the weight vector of the learnable linear layer, with a dimension of 64, and consists of learnable parameters. This represents the hidden state vector of the last coding step in the coding sequence of the second coding path; This represents the bias of the learnable linear layer, which is a learnable parameter; Indicates the insulation condition index number, ; This indicates the predicted sample number.

[0085] S52 calculates the binary cross-entropy using the predicted probability and the pre-activated event label. The binary cross-entropy provides a supervised gradient for the second coding path through backpropagation.

[0086] It is understandable that the joint objective function, composed of the insulation degradation index prediction loss and the pre-activation prediction loss, is expressed as: in, Describe the joint objective function; This indicates the loss predicted by the insulation degradation index; This represents the weighting coefficient for the pre-activation prediction loss, with a value of 0.3. This represents the pre-activation prediction loss.

[0087] It should be noted that, The basis for this setting is: the insulation degradation index grading interval is 20 points, and 25% of the grading interval of 20 points is used as the upper limit of the estimation error that will not affect the grading conclusion, i.e., 5 points. The contribution of the pre-activation prediction loss is made comparable to that of the insulation degradation index prediction loss at the grading judgment boundary, ensuring that the pre-activation auxiliary task does not suppress the optimization of the insulation degradation index prediction accuracy.

[0088] Furthermore, the insulation degradation index predicts the loss, expressed as: in, Indicates the insulation degradation index as a predictive loss; Indicates the total number of training samples; Indicates the training sample number; Indicates the prediction step number. ; This represents the predicted value of the insulation degradation index; This represents the true insulation degradation index value, calculated by the formula version node based on the current version's insulation status assessment parameters and time steps. The measured values ​​of the seven insulation status indicators are calculated and stored in the equipment entity node; This indicates the prediction start time; the denominator 100 is the upper bound of the insulation degradation index range, used to normalize the root mean square error.

[0089] Furthermore, the pre-activation prediction loss is expressed as: in, This represents the pre-activation prediction loss; Indicates the total number of training samples; Indicates the training sample number; Indicates the insulation condition index number, ; Indicates a pre-activation event label; Indicates the predicted probability; It represents the natural logarithm.

[0090] All learnable parameters are based on the joint objective function. Perform backpropagation, including the long short-term memory network parameters in the first encoding pathway, the long short-term memory network parameters in the second encoding pathway, and the cross-pathway pre-activation gating weight matrix. Projection matrix from the first encoding path to the fusion layer Projection matrix from the second encoding path to the fusion layer Attention query vector Projection matrix of fused encoding hidden states The projection matrix of the decoder's hidden state Long Short-Term Memory Network Parameters and Output Layer Weight Matrix in the Decoder Output layer bias The weight vector of a learnable linear layer and the bias of learnable linear layers .

[0091] The training parameters are set as follows: Adam optimizer, initial learning rate 0.001, batch size 32, maximum training epochs 200, training stops if the root mean square error of the validation set does not decrease for 10 consecutive epochs.

[0092] The inference phase only performs forward computation from the fused coded sequence to the predicted insulation degradation index, predicting the probability. and pre-activation prediction loss It only participates in computation during the training phase and is not included in the computation process during the inference phase.

[0093] After completing step S5, in some embodiments, step S53 is also included: in response to the number of newly added operation and maintenance confirmation records in the knowledge graph reaching a preset number, the lower bound value, upper bound value, and weight parameters of the insulation status assessment parameters are recalculated; the contribution feature vector and the neighboring feature vector are recalculated based on the recalculated insulation status assessment parameters; the pre-activated event labels are re-extracted from the knowledge graph based on the recalculated lower bound value of the interval; and the dual-path long short-term memory encoder is updated and trained using the recalculated contribution feature vector, the neighboring feature vector, and the re-extracted pre-activated event labels.

[0094] Understandably, the preset quantity is 10, meaning that when 10 new maintenance confirmation records are added to the knowledge graph, the insulation status assessment parameters are updated and the dual-path long short-term memory encoder is updated and trained.

[0095] Update training based on joint objective function Perform backpropagation, covering all learnable parameters in step S52, where the weight vector and bias of the learnable linear layer participate in gradient calculation through pre-activation prediction loss.

[0096] It should be noted that the index-sample pair whose values ​​changed before and after the update is used. Divide by the product of the total number of indicators and the total number of training samples to obtain the label inversion ratio. ,in This indicates the number of sample pairs whose values ​​changed in the pre-activated event label vector before and after the update; 7 represents the total number of insulation status indicators. This represents the total number of training samples.

[0097] The label inversion ratio threshold is set to 0.20, based on the following: when the label inversion ratio is not less than 0.20, the parameter update amount of 5 rounds of update training is approximately 2.5% of the total initial training update amount, which is insufficient for the dual-path long short-term memory encoder to fully adapt to the new pre-activated event label distribution, requiring a complete retraining. When the label inversion ratio is less than 0.20, the dual-path long short-term memory encoder is trained for 5 rounds using the updated contribution feature vector, neighboring feature vector, and pre-activated event labels, with a learning rate of 0.0001; when the label inversion ratio is not less than 0.20, the dual-path long short-term memory encoder is fully retrained using the updated contribution feature vector, neighboring feature vector, and pre-activated event labels, with training parameters consistent with step S5. After the update training is completed, if the root mean square error of the validation set increases by not less than 5% compared to before the update, the parameter snapshot of the dual-path long short-term memory encoder before the update is reverted.

[0098] Example 2 is an embodiment of the present invention, which provides a transformer insulation condition diagnosis system based on knowledge graphs, including: The feature construction module is used to calculate membership features based on the lower bound of the interval in the insulation state assessment parameters to form a contribution feature vector, and to calculate boundary proximity features based on the lower bound of the interval to form a proximity feature vector. The tag extraction module is used to extract pre-activated event tags from the knowledge graph based on the lower bound of the interval; A dual-path coding module is used to construct a dual-path long short-term memory encoder. The dual-path long short-term memory encoder includes a first coding path with the contribution feature vector as input and a second coding path with the neighboring feature vector as input. The preorder hidden state of the second coding path is introduced into the input gate calculation of the first coding path to establish a cross-path gating connection. The fusion coding module is used to obtain the posterior probability of the mechanism from the knowledge graph, determine the fusion weight based on the maximum value of the posterior probability of the mechanism, and perform weighted fusion on the hidden state sequences of the first coding path and the second coding path step by step to generate the fusion coding sequence. The training prediction module is used to train a dual-path long short-term memory encoder by taking the fused encoded sequence as input, using the insulation degradation index prediction loss and the pre-activation prediction loss calculated based on the pre-activation event label as the joint objective function, and outputting the insulation degradation index prediction result.

[0099] This embodiment also provides an electronic device applicable to a knowledge graph-based transformer insulation condition diagnosis method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the knowledge graph-based transformer insulation condition diagnosis method proposed in the above embodiment.

[0100] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a knowledge graph-based transformer insulation state diagnosis method as proposed in the above embodiments.

[0101] The storage medium proposed in this embodiment belongs to the same inventive concept as the knowledge graph-based transformer insulation state diagnosis method proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0102] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0103] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for diagnosing the insulation condition of a transformer based on a knowledge graph, characterized in that, include: Membership features are calculated based on the lower bound of the interval in the insulation condition assessment parameters, and contribution feature vectors are constructed using the membership features; boundary proximity features are calculated based on the lower bound of the interval, and proximity feature vectors are constructed using the boundary proximity features. Based on the lower bound of the interval, extract pre-activated event tags from the knowledge graph that indicate the occurrence of interval crossings. A dual-path long short-term memory encoder is constructed, comprising a first encoding path with the contribution feature vector as input and a second encoding path with the neighboring feature vector as input; the preorder hidden state of the second encoding path is introduced into the input gate calculation of the first encoding path to establish a cross-path gating connection; The mechanism posterior probability is obtained from the knowledge graph, and the fusion weight is determined based on the maximum value of the mechanism posterior probability. The hidden state sequences of the first coding path and the second coding path are weighted and fused step by step to generate a fused coding sequence. Using the fused encoding sequence as input, the dual-path long short-term memory encoder is trained by using the insulation degradation index prediction loss and the pre-activation prediction loss calculated based on the pre-activation event label as a joint objective function, and outputs the insulation degradation index prediction result.

2. The transformer insulation condition diagnosis method based on knowledge graph as described in claim 1, characterized in that, The calculation of membership features based on the lower bound of the interval includes: dividing the insulation state assessment range into a zero contribution interval, an effective contribution interval, and a full contribution interval based on the lower bound of the interval and the upper bound of the interval in the insulation state assessment parameters; when in the zero contribution interval, the membership feature takes a zero value; when in the effective contribution interval, the membership feature takes a linear interpolation between the lower bound of the interval and the upper bound of the interval; and when in the full contribution interval, the membership feature takes a unit value. The calculation of the boundary proximity feature based on the lower bound of the interval includes: when the boundary proximity feature is in the zero contribution interval, the boundary proximity feature is taken as a linear interpolation between the lower bound of the interval and the zero value; when the boundary proximity feature is in the effective contribution interval or the full contribution interval, the boundary proximity feature is taken as zero.

3. The transformer insulation condition diagnosis method based on knowledge graph as described in claim 2, characterized in that, The step of extracting pre-activation event tags from the knowledge graph based on the lower bound of the interval includes: Using each time step within the prediction time window as a detection unit, in response to the normalized value of the insulation status index at the current time step exceeding the lower bound of the interval and the normalized value of the insulation status index at the previous time step not exceeding the lower bound of the interval, it is determined that the insulation status index has crossed the interval. When the insulation status index is determined to have crossed a range within the prediction time window, the pre-activation event label takes a positive value; when the insulation status index is determined not to have crossed a range within the prediction time window, the pre-activation event label takes a negative value.

4. The transformer insulation condition diagnosis method based on knowledge graph as described in claim 3, characterized in that, The step of introducing the preorder hidden state of the second encoding path into the input gate calculation of the first encoding path to establish a cross-path gating connection includes: The preceding hidden state of the second coding path is linearly transformed and superimposed on the calculation term of the input gate of the first coding path. The weight matrix of the linear transformation is used as a learnable parameter in the training. The cross-path gating connection acts on the input gate of the first coding path, and modulates the proportion of the contribution feature vector received by the memory unit of the first coding path using the boundary proximity feature.

5. The transformer insulation condition diagnosis method based on knowledge graph as described in claim 4, characterized in that, The step of determining the fusion weight based on the maximum value of the posterior probability of the mechanism, and performing weighted fusion on the hidden state sequences of the first coding path and the second coding path step by step to generate a fused coding sequence includes: The maximum value among the posterior probabilities of thermal aging mechanism, electrical aging mechanism, and chemical aging mechanism at the predicted start time is determined as the first fusion weight, and the second fusion weight is determined by the constraint that the sum of the first fusion weight and the second fusion weight is one. The hidden state sequence of the first coding path is weighted according to the first fusion weight, and the hidden state sequence of the second coding path is weighted according to the second fusion weight. The weighted results of each coding step are summed to generate the fused coding sequence.

6. The transformer insulation condition diagnosis method based on knowledge graph as described in claim 5, characterized in that, The pre-activation prediction loss calculated based on the pre-activation event label includes: The hidden state of the last encoding step in the second encoding path is input to the learnable linear layer, and the predicted probability of the insulation state index being crossed within the predicted time window is output. The predicted probability and the pre-activated event label are used to calculate the binary cross-entropy, which provides a supervisory gradient for the second coding path through backpropagation.

7. The method for diagnosing transformer insulation status based on knowledge graphs as described in claim 6, characterized in that, In response to the number of newly added operation and maintenance confirmation records in the knowledge graph reaching a preset number, the lower limit value of the interval, the upper limit value of the interval, and the weight parameters in the insulation status assessment parameters are recalculated. The contribution feature vector and the neighboring feature vector are recalculated based on the recalculated insulation state assessment parameters. The pre-activated event label is re-extracted from the knowledge graph based on the recalculated lower bound of the interval. The dual-path long short-term memory encoder is updated and trained using the recalculated contribution feature vector, the neighboring feature vector, and the re-extracted pre-activated event label.

8. A knowledge graph-based transformer insulation condition diagnosis system, employing the knowledge graph-based transformer insulation condition diagnosis method as described in any one of claims 1 to 7, characterized in that, include: The feature construction module is used to calculate membership features based on the lower bound of the interval in the insulation state assessment parameters to form a contribution feature vector, and to calculate boundary proximity features based on the lower bound of the interval to form a proximity feature vector. The tag extraction module is used to extract pre-activated event tags from the knowledge graph based on the lower bound of the interval; A dual-path encoding module is used to construct a dual-path long short-term memory encoder. The dual-path long short-term memory encoder includes a first encoding path with the contribution feature vector as input and a second encoding path with the neighboring feature vector as input. In the input gate calculation of the first encoding path, the preorder hidden state of the second encoding path is introduced to establish a cross-path gating connection. The fusion coding module is used to obtain the posterior probability of the mechanism from the knowledge graph, determine the fusion weight based on the maximum value of the posterior probability of the mechanism, and perform weighted fusion on the hidden state sequences of the first coding path and the second coding path step by step to generate a fusion coding sequence. The training prediction module is used to train the dual-path long short-term memory encoder by taking the fused coding sequence as input, using the insulation degradation index prediction loss and the pre-activation prediction loss calculated based on the pre-activation event label as a joint objective function, and outputting the insulation degradation index prediction result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the knowledge graph-based transformer insulation state diagnosis method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the knowledge graph-based transformer insulation state diagnosis method according to any one of claims 1 to 7.