A power distribution room operation and maintenance full-link knowledge graph cross-modal alignment method

Abstract

This invention discloses a cross-modal alignment method for knowledge graphs across the entire operation and maintenance chain of a power distribution room, relating to the fields of intelligent operation and maintenance of power distribution rooms and knowledge graph data processing. The method acquires multimodal feature vectors of the power distribution room's operation and maintenance equipment, and obtains equipment degradation index and real-time electrical topology mask; it suppresses aging background disturbances based on the equipment degradation index, obtaining aging-suppressed multimodal feature vectors; it obtains candidate knowledge sets and candidate fault entity sets through a multimodal knowledge base, and generates candidate discreteness and aging compatibility residuals; it obtains initial fault inference control quantities based on the equipment degradation index, aging compatibility residuals, and candidate discreteness; it performs hop-by-hop inference by combining knowledge graph link integrity verification and real-time electrical topology connectivity conditions, and outputs refined alignment results based on the comparison between the remaining inference quantity of the current link and the inference truncation threshold.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance of power distribution rooms and knowledge graph data processing technology. More specifically, this invention relates to a cross-modal alignment method for knowledge graphs across the entire operation and maintenance chain of power distribution rooms. Background Technology

[0002] With the improvement of intelligent operation and maintenance levels in power distribution substations, multi-source data, including images, videos, audio, equipment ledgers, operating parameters, fault records, handling procedures, and standard clauses, has gradually been generated during on-site inspections, online monitoring, and operation and maintenance management. Existing power distribution substation operation and maintenance systems typically collect, store, and retrieve this data through equipment monitoring platforms, power distribution automation systems, inspection terminals, and operation and maintenance knowledge bases. In scenarios such as fault analysis, operation and maintenance Q&A, and handling suggestion generation, similarity retrieval, knowledge base matching, or knowledge graph association methods are used to associate the abnormal characteristics collected on-site with existing fault cases, equipment information, handling steps, and standard references to help operation and maintenance personnel quickly locate fault types and form handling references.

[0003] However, the current problem is that the normal operating background of power distribution room equipment, such as aging, heavy load, slight vibration, and local temperature rise during long-term operation, is easily intertwined with actual fault symptoms in multimodal data. Existing methods based on similarity or single knowledge matching are difficult to accurately distinguish between normal degradation backgrounds and abnormal fault characteristics, which can easily lead to scattered or mismatched candidate fault results. At the same time, existing knowledge base or knowledge graph alignment methods often focus on semantic associations, making it difficult to fully verify whether candidate results can form a complete operation and maintenance business link from equipment, fault, handling to standard basis. They also fail to combine the real-time electrical connection status of the power distribution room to limit the expansion range of the knowledge link, which may lead to problems such as incomplete handling basis, excessive expansion range, crossing actual isolation boundaries, or inconsistency with the on-site electrical topology in the alignment results. This affects the accuracy, interpretability, and on-site applicability of the power distribution room operation and maintenance results.

[0004] To address the above problems, this invention proposes a solution. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a cross-modal alignment method for knowledge graphs across the entire operation and maintenance chain of a power distribution room, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for cross-modal alignment of knowledge graphs across the entire operation and maintenance chain of a power distribution room includes the following steps: obtaining multimodal feature vectors of power distribution room operation and maintenance equipment, obtaining equipment degradation index based on equipment ledger, and obtaining real-time electrical topology mask based on switch status; suppressing aging background disturbances of multimodal feature vectors based on equipment degradation index to obtain aging-suppressed multimodal feature vectors;

[0008] The aging suppression multimodal feature vector is input into the power distribution room operation and maintenance multimodal knowledge base. A candidate knowledge set is obtained according to the preset similarity threshold, and a candidate fault entity set is extracted. The candidate discreteness is obtained by the inverse measure of the dispersion of the similarity scores of each candidate fault entity. The aging compatibility residual is obtained by the deviation intensity of the abnormal residual features determined in the aging background disturbance suppression process relative to the normal aging law of the target equipment.

[0009] The initial fault simulation control quantity is obtained based on the equipment degradation index, aging compatibility residual, and candidate dispersion. The link integrity of the candidate knowledge set is verified according to the preset link rules of the power distribution room operation and maintenance-specific knowledge graph. Based on the preset binding relationship between the equipment entity nodes and the corresponding equipment access branches in the candidate knowledge, and combined with the real-time electrical topology mask connectivity conditions, candidate links are obtained. Starting with the initial fault simulation control quantity, the simulation is performed hop by hop along the candidate links. The branches along the real-time electrical topology mask are marked as valid connected branches. The remaining simulation quantity of the current link is updated according to the link attenuation of the branch. The expansion stops when the remaining simulation quantity of the current link is not greater than the simulation truncation threshold, and the current number of hops is determined as the maximum number of valid hops. The fine alignment result is output within the link range limited by the maximum number of valid hops.

[0010] In a preferred embodiment, the equipment deterioration index is obtained as follows: the operational duration and average annual heavy load frequency of the target equipment are read from the equipment ledger data, and normalized according to the upper limit of the service life and the upper limit of the heavy load frequency of similar equipment, respectively. The two normalized results are linearly combined according to the contribution weight of operational duration and the contribution weight of heavy load frequency, and the smaller value of the result is taken as the output of the equipment deterioration index. The sum of the contribution weight of operational duration and the contribution weight of heavy load frequency is equal to 1, and the value range of the equipment deterioration index is [0,1].

[0011] In a preferred embodiment, the real-time electrical topology mask is obtained by reading the real-time opening and closing status of circuit breakers, disconnectors, load switches, and bus tie switches from the SCADA system, distribution automation system, or station monitoring, and generating a Boolean adjacency matrix as the real-time electrical topology mask in combination with the primary wiring relationship of the distribution room; for currently closed and electrically connectable branches, they are marked as 1 in the adjacency matrix and used as valid connected branches when expanding candidate links; for currently disconnected, under maintenance isolation, or blocked isolation branches, they are marked as 0 in the adjacency matrix and used as uncrossable electrical topology boundaries.

[0012] In a preferred embodiment, the aging suppression multimodal feature vector is generated as follows: the equipment degradation state of the target equipment is determined according to the equipment degradation index; based on the historical normal operation samples of similar equipment corresponding to the equipment degradation state, the normal fluctuation boundary of each feature in the equipment degradation state is determined; and the interval defined by the normal fluctuation boundary is used as the degradation compatibility reference range of the corresponding dimension.

[0013] For a certain dimension of the multimodal feature vector, the feature value of that dimension in the current sampling window is normalized to obtain the normalized feature energy. The normalized feature energy is then compared with the degradation compatibility reference range of the corresponding dimension. When the feature falls within the degradation compatibility reference range, it is classified as a degradation compatible feature and assigned a low retention weight. When the feature exceeds the degradation compatibility reference range but the extent of the exceedance does not reach the preset abnormal residual threshold, it is classified as a critical deviation feature and assigned an intermediate retention weight. When the feature exceeds the degradation compatibility reference range and the extent of the exceedance reaches the preset abnormal residual threshold, it is classified as an abnormal residual feature and assigned a high retention weight. After dimension-wise weighting, an aging suppression multimodal feature vector with the same dimensions as the multimodal feature vector is obtained. The preset abnormal residual threshold is determined based on the normal fluctuation boundary width of the feature dimension.

[0014] In a preferred embodiment, the aging compatibility residual is obtained by: based on the abnormal residual features obtained from the division, the aging compatibility residual is obtained by normalization according to the number of abnormal residual features, the extent to which the abnormal residual features exceed the degradation compatibility reference range, and the reliability weight corresponding to the mode to which the abnormal residual features belong; wherein, the value range of the aging compatibility residual is [0,1], which is the quantification result of the deviation intensity of the multimodal feature vector from the normal aging law of the target device.

[0015] In a preferred embodiment, the candidate dispersion is determined as follows: the top K candidate fault entities are selected from the candidate fault entity set in descending order of similarity to form a Top-K candidate fault entity set, and their similarity score set is obtained; the ratio between the sum of the standard deviation of the Top-K similarity score set and the stability constant, and the sum of the mean of the Top-K similarity score set and the stability constant are used as the stabilization coefficient of variation, and the negative logarithm of the stabilization coefficient of variation is taken to obtain the candidate dispersion; when the candidate dispersion is less than 0, it is truncated to 0, and when the candidate dispersion exceeds the preset upper limit, it is truncated to the preset upper limit; wherein, the stability constant is a positive number, used to prevent the denominator of the coefficient of variation from being zero or the logarithmic operation from being abnormal when the mean or standard deviation of the Top-K similarity score set approaches 0.

[0016] In a preferred embodiment, the preset link rules include association rules between device entities, fault entities, handling step entities, and standard basis entities; when performing link integrity verification, candidate knowledge that can form a complete business link of device-fault-handling-standard is screened, and candidate knowledge that cannot be bound to handling steps or standard basis is filtered.

[0017] In a preferred embodiment, the link attenuation is determined as follows: For a valid connected branch marked as 1 in the real-time electrical topology mask, the equivalent electrical impedance of the branch is first normalized to a normalized equivalent electrical impedance according to the reference impedance, and the protection time difference between adjacent protection devices of the branch is normalized to a normalized protection time difference according to the reference time. The sum of the products of impedance contribution weight multiplied by the normalized equivalent electrical impedance and protection time difference contribution weight multiplied by the normalized protection time difference is used as the link attenuation of the branch, wherein the sum of impedance contribution weight and protection time difference contribution weight is equal to 1. For isolated branches marked as 0 in the real-time electrical topology mask, they are treated as uncrossable electrical topology boundaries and do not participate in the update process of the current link remaining projection.

[0018] In a preferred embodiment, during the hop-by-hop simulation using the initial fault simulation control quantity, a maximum allowable number of search hops is set as a fallback constraint; when the number of simulation hops reaches the maximum allowable number of search hops, regardless of whether the remaining simulation quantity of the current link is greater than the simulation truncation threshold, the expansion along that direction is stopped.

[0019] In a preferred embodiment, after outputting the fine alignment result, the topology attenuation model is rewritten and calibrated based on the on-site review results, wherein the topology attenuation model includes the synthetic parameters of link attenuation.

[0020] Obtain the actual fault impact level confirmed by on-site review, and compare the actual fault impact level with the maximum effective number of hops to obtain the level deviation. When the actual fault impact level is greater than the maximum effective number of hops, the model is prematurely truncated. Based on the contribution ratio of each attenuation component to the link attenuation in this batch of simulation records, the attenuation component to be adjusted is determined, and the weight corresponding to the attenuation component to be adjusted is reduced. When the actual fault impact level is less than the maximum effective number of hops, the model is deemed to have expanded too far. Based on the contribution ratio of each attenuation component to the link attenuation in this batch of simulation records, the attenuation component to be adjusted is determined, and the weight corresponding to the attenuation component to be adjusted is increased. The attenuation component includes the normalized equivalent electrical impedance component and the normalized protection time difference component. The sum of the adjusted impedance contribution weight and the protection time difference contribution weight remains equal to 1.

[0021] The technical effects and advantages of this invention's knowledge graph cross-modal alignment method for the entire operation and maintenance chain of a power distribution room are as follows: By introducing equipment degradation index and aging background disturbance suppression processing before multimodal knowledge alignment, this invention can distinguish normal degradation backgrounds such as slight temperature rises, vibrations, and insulation aging traces generated by long-term operation of power distribution room equipment from real fault anomalies, reducing the interference of non-fault factors on candidate fault matching results; by generating aging-compatible residuals based on abnormal residual characteristics and combining the similarity dispersion of candidate fault entities to obtain candidate discreteness, an initial fault inference control quantity is further formed, enabling the degree of equipment degradation, the intensity of abnormal deviation, and the semantic uncertainty of candidates to jointly participate in subsequent fine alignment judgment; through knowledge graph link integrity verification, a complete business link that can cover equipment, faults, handling, and standard basis can be screened out, avoiding the output of candidate results lacking handling basis or standard support. By performing hop-by-hop deduction of candidate links using real-time electrical topology masks and link attenuation, and determining the maximum effective number of hops based on the remaining deduction amount of the current link and the deduction truncation threshold, this invention can prevent knowledge links from crossing disconnected, isolated, or blocked branches and suppress unnecessary over-expansion. Thus, this invention improves the accuracy, completeness, interpretability, and field applicability of cross-modal knowledge alignment results for power distribution room operation and maintenance. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a knowledge graph cross-modal alignment method for the entire operation and maintenance chain of a power distribution room, as described in this invention.

[0023] Figure 2 This is a schematic diagram illustrating the aging background disturbance suppression and aging-compatible residual generation of the present invention;

[0024] Figure 3 This is a schematic diagram of the real-time electrical topology mask constraint and topology attenuation truncation of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example: Please refer to Figure 1 As shown, this invention discloses a cross-modal alignment method for knowledge graphs across the entire operation and maintenance chain of a power distribution room, comprising the following steps:

[0027] Step 1: Collect multimodal data such as images, audio, video, and text from the power distribution room maintenance equipment. Preprocess and extract features from various types of data to obtain multimodal feature vectors. Simultaneously, obtain the equipment degradation index based on equipment ledger data, and obtain the real-time electrical topology mask based on the real-time opening and closing status of switchgear and primary wiring relationships. This step provides the basic input for subsequent aging background disturbance suppression, knowledge base coarse alignment, and topology constraint inference.

[0028] Specifically, for target equipment in the power distribution room, at least two types of status data are collected from image data, audio data, video data, and text data. For image data, anti-glare, anti-occlusion restoration, noise suppression, size unification, and resolution normalization are performed. For audio data, adaptive noise reduction, sampling rate unification, effective segment extraction, and amplitude normalization are performed. For video data, key frame sequences that reflect the equipment's operating status, indicator light status, instrument readings, or critical fault locations are extracted. For text data, terminology segmentation, semantic structure annotation, and terminology unification are performed. This approach enables a structured representation of fault descriptions, handling procedures, standard clauses, and equipment parameters. Based on fault ID, equipment ID, timestamp, data collection location, and maintenance task number, images, audio, video, text specifications, equipment parameters, and handling records corresponding to the same equipment or the same fault event are bound together to establish a unified data association mapping table. Furthermore, a multimodal knowledge base for power distribution room operation and maintenance is constructed based on the data association mapping table. The multimodal knowledge base includes a multimodal knowledge fragment table, a fault entity index table, an equipment entity index table, a handling procedure index table, and a standard basis index table.

[0029] Each multimodal knowledge fragment in the multimodal knowledge fragment table includes at least a knowledge fragment ID, a device ID, a device type, a fault entity ID, a fault description text, a handling step ID, a standard basis ID, an image evidence ID, an audio evidence ID, a video keyframe ID, a text standard ID, and a corresponding multimodal vector index. The multimodal vector index is used to store the knowledge fragment vector obtained by fusing text features, image features, audio features, and video keyframe features.

[0030] When constructing the multimodal knowledge base, historical fault cases, inspection records, equipment ledgers, defect records, handling tickets, operation and maintenance procedures, and standard clauses are uniformly numbered according to equipment ID, fault entity ID, handling step ID, and standard basis ID. For image, audio, video, and text data under the same fault event, the same knowledge fragment ID is generated, and the knowledge fragment ID is bound to the corresponding multimodal vector index. Thus, in the subsequent coarse alignment stage, the aging suppression multimodal feature vector can perform similarity calculation with the knowledge fragment vector and return a candidate knowledge set containing equipment information, fault entity, handling steps, and standard basis.

[0031] Simultaneously, a dedicated knowledge graph for power distribution room operation and maintenance is constructed based on the aforementioned equipment entity index table, fault entity index table, handling step index table, and standard basis index table. The nodes in the dedicated knowledge graph for power distribution room operation and maintenance include at least equipment entity nodes, fault entity nodes, handling step entity nodes, standard basis entity nodes, and electrical topology nodes. The edge relationships include at least the following: equipment entity-fault entity relationship, image feature-fault entity relationship, audio feature-fault entity relationship, video status-fault entity relationship, fault entity-handling step relationship, handling step-standard basis relationship, and equipment entity-electrical topology node relationship.

[0032] The device entity-device access branch relationship is used to record the corresponding device access branch of the device entity in a primary wiring relationship; the fault entity-handling step relationship is used to record the handling actions corresponding to different fault types; the handling step-standard basis relationship is used to record the operation and maintenance procedures or standard clauses corresponding to the handling actions; through the above entity and relationship settings, the multimodal knowledge base is used for coarse similarity retrieval, and the power distribution room operation and maintenance-specific knowledge graph is used for business link integrity verification and topology constraint deduction of candidate results;

[0033] Subsequently, features for each modality were extracted. For text data, text features such as fault type, handling steps, standard clauses, and equipment parameters were extracted. For image data, visual features such as equipment fault area, instrument readings, localized heating, component deformation, and surface discoloration were extracted. For audio data, abnormal noise spectrum features, partial discharge acoustic signature features, and abnormal frequency distribution features were extracted. For video data, equipment runtime sequence features were extracted based on keyframe sequences. After normalization, each modal feature was associated according to equipment ID, fault ID, or maintenance task number to form a unified multimodal feature vector. Among them, multimodal feature vectors This represents a multimodal feature vector obtained by splicing or fusing text features, image features, audio features, and video keyframe features. Its dimension can be preset according to the output dimension of each modality encoder. The initial dimension can be set to 512 to 1024 dimensions. The specific dimension is based on the knowledge base index structure and similarity calculation efficiency.

[0034] For example, for a 10kV feeder cabinet, infrared thermal images, partial discharge acoustic fingerprint clips, key frames of inspection videos, and the corresponding ledgers and maintenance specifications for the equipment can be collected simultaneously. Among them, the infrared images reflect the local heating state, the acoustic fingerprint clips reflect abnormal discharge or mechanical vibration, the key frames of the videos reflect the switch position and indicator light status, and the text specifications are used to provide a basis for handling. The above data are bound and normalized into multimodal feature vectors to provide input for subsequent coarse alignment.

[0035] Furthermore, parameters such as the target equipment's operational duration, average annual heavy load frequency, historical defect records, or maintenance intervals are retrieved from the equipment ledger data to determine the equipment degradation index. The equipment degradation index is used to represent the current degradation state of the target equipment, and its value ranges from [0,1], where 0 represents a low degradation state and 1 represents a high degradation state. This index is used to distinguish between background disturbances caused by normal aging and abnormal signs that deviate from the normal aging pattern. Its value is determined based on the service life, heavy load frequency, historical defect records, and aging level samples confirmed by maintenance experts for similar equipment. In one embodiment, the equipment degradation index is calculated according to the following formula: ; In the formula, Indicates the duration the equipment has been in operation. This indicates the upper limit of the rated service life of similar equipment. This indicates the average annual frequency of heavy loads on the equipment. This indicates the upper limit of the rated heavy-load frequency for similar equipment; for a 10kV feeder cabinet, The initial value can be taken as 20 years. The initial value can be taken as 50 times / year; for other equipment types, and Calibration is performed based on the equipment's design life, operation and maintenance procedures, asset ledger statistics, or historical operating samples of similar equipment. Indicates the contribution weight of the commissioning time. Indicates the contribution weight of overload frequency;

[0036] During initial implementation, the duration of operation contributes weight. A value of 0.6 can be used, representing the contribution weight of overload frequency. A value of 0.4 can be taken, as both conditions are satisfied. The initial value is based on the fact that the aging of power distribution room equipment is usually mainly affected by service life and the frequency of heavy loads as accelerated deterioration factors, and can be adjusted according to historical failure samples of similar equipment. If the equipment ledger also includes the number of historical defects or maintenance intervals, these can also be used as normalized auxiliary parameters in the equipment deterioration index, but the final output is still limited to... Within the range;

[0037] Simultaneously, the real-time opening and closing status of switching equipment such as circuit breakers, disconnectors, load switches, and bus tie switches is read from SCADA, distribution automation, or station monitoring, and a real-time electrical topology mask is generated by combining it with the primary wiring relationship of the power distribution room. Among them, real-time electrical topology mask This represents a Boolean adjacency matrix generated by the primary wiring relationships of the power distribution room and the real-time status of the switches. A matrix element of 1 indicates that the corresponding branch is currently electrically connected, while a matrix element of 0 indicates that the corresponding branch is currently disconnected, isolated, or blocked, and is not allowed to continue to expand as a normal fine-aligned link. For branches that are currently closed and electrically connected, they are marked as 1 in the real-time electrical topology mask; for branches that are currently disconnected, under maintenance, or blocked, they are marked as 0 in the real-time electrical topology mask. This real-time electrical topology mask serves as the physical boundary constraint for subsequent knowledge graph link fine-alignment.

[0038] The output of this step includes: multimodal feature vectors. Equipment deterioration index Data association mapping table and real-time electrical topology mask .

[0039] Step 2: Determine the equipment degradation state based on the equipment degradation index, establish the corresponding degradation compatibility reference range, and divide the multimodal features into degradation compatibility features, critical deviation features, and abnormal residual features. After weighted processing, obtain the aging suppression multimodal feature vector; obtain the aging compatibility residual based on the abnormal residual features; input the aging suppression multimodal feature vector into the power distribution room operation and maintenance multimodal knowledge base to obtain the candidate knowledge set and candidate fault entity set, and obtain the candidate dispersion based on the similarity dispersion of the candidate fault entities. Finally, obtain the initial fault inference control quantity based on the equipment degradation index, aging compatibility residual, and candidate dispersion.

[0040] Specifically, in the long-term operation scenario of power distribution rooms, some multimodal anomalies are not caused by immediate faults, but by aging background disturbances formed by natural aging of equipment, long-term heavy load, slight deterioration of insulation materials, or loosening of mechanical structures. For example, old equipment may have mechanical vibrations within a certain stable range, there may be slight uniform temperature rise near the joints after long-term operation, and aging marks that do not constitute immediate faults may appear on the surface of insulation components. If these aging background disturbances are directly input into the coarse alignment model, the candidate knowledge set is prone to excessive divergence.

[0041] Please see Figure 2 As shown, based on the equipment degradation index output in step one The current state of equipment degradation of the target equipment is determined, and a degradation compatibility reference range corresponding to the current state of equipment degradation is established based on historical normal operation samples of similar equipment. The degradation compatibility reference range is used to represent the normal range of variation of the multi-modal characteristics of each dimension of the target equipment under the current state of equipment degradation.

[0042] Specifically, according to the equipment deterioration index Within the specified interval, samples with the same or adjacent degradation state as the current equipment are selected from historical normal operating samples of similar equipment. The normal fluctuation boundaries of each multi-modal feature under this degradation state are statistically analyzed to obtain the degradation compatibility reference range corresponding to each feature. The normal fluctuation boundaries can be determined based on the quantile values, median absolute deviation, or normal operating boundaries confirmed by maintenance experts for the corresponding features in normal samples of similar equipment. The initial values ​​for the quantile value interval can be from the 10th to the 90th percentile, or calibrated based on the stable intervals of historical samples of similar equipment. In a preferred embodiment, based on historical normal operating samples of similar equipment corresponding to the current equipment degradation state, the normal fluctuation boundaries of each feature under this degradation state are determined, and the interval defined by the normal fluctuation boundaries is used as the degradation compatibility reference range for the corresponding dimension.

[0043] For example, the equipment degradation index of a certain feeder cabinet If the value is 0.506, the system will no longer compare the feeder cabinet with brand new equipment, but will instead select... Normal samples of similar feeder cabinets in a moderately deteriorated state are used as a reference. If slight uniform temperature rise and small mechanical vibration are common phenomena in such samples, these characteristics will fall into the corresponding deterioration compatibility reference range and will not be directly used as strong fault characteristics in subsequent coarse alignment.

[0044] For a certain dimension of a multimodal feature vector, first determine the normalized feature energy of that dimension in the current sampling window, and then compare the normalized feature energy with the degradation compatibility reference range of the corresponding dimension; the normalized feature energy is the value obtained after normalizing the feature value of that dimension in the current sampling window, and is used to compare with the degradation compatibility reference range of the corresponding dimension.

[0045] When a feature falls within the degradation compatibility reference range, it is classified as a degradation compatibility feature; when a feature exceeds the degradation compatibility reference range but the extent of the exceedance does not reach the preset abnormal residual threshold, it is classified as a critical deviation feature; when a feature exceeds the degradation compatibility reference range and the extent of the exceedance reaches the preset abnormal residual threshold, it is classified as an abnormal residual feature. The preset abnormal residual threshold is used to distinguish between critical deviation features and abnormal residual features. Its initial value can be determined based on the statistical distribution of temperature rise amplitude, acoustic energy, image abnormal response intensity, or video status change amplitude in historical fault samples of similar equipment. In the initial implementation, the amplitude exceeding the degradation compatibility reference range can reach 50% of the normal fluctuation boundary width of the feature as the preset abnormal residual threshold, or it can be adjusted according to on-site operation and maintenance samples.

[0046] For deterioration compatibility features, low retention weights are assigned to reduce their impact on the coarse alignment results; for critical deviation features, intermediate retention weights are assigned based on the extent to which they exceed the deterioration compatibility reference range, allowing them to participate in subsequent coarse alignment in a weaker form; for anomalous residual features, high retention weights are assigned to ensure their full participation in subsequent coarse alignment; after dimension-wise weighting, the aging suppression multimodal feature vector is obtained. Among them, the aging suppression multimodal feature vector In the multimodal feature vector Based on this, the feature vector obtained by performing low-weight suppression on degraded compatibility features, medium-weight preservation on critical deviation features, and high-weight preservation on abnormal residual features has the same dimension as the multimodal feature vector. Maintain consistency;

[0047] The initial value of the low retention weight can be 0.1, the initial value of the intermediate retention weight can be 0.5, and the initial value of the high retention weight can be 1.0. The low retention weight is used to suppress the aging background disturbance that conforms to the current equipment degradation state, the intermediate retention weight is used to retain the critical deviation characteristics that have a certain abnormal tendency but have not yet reached the point of strong fault indication, and the high retention weight is used to fully retain the abnormal residual characteristics that deviate significantly from the current equipment degradation state. The above weights can be adjusted according to the separability of normal aging characteristics, critical deviation characteristics, and fault characteristics in historical samples.

[0048] Simultaneously, based on the number of abnormal residual features, the extent to which the abnormal residual features exceed the degradation compatibility reference range, and the reliability weight of the mode to which the abnormal residual features belong, the aging compatibility residual is obtained. The reliability weight is a preset weight corresponding to the mode to which the abnormal residual feature belongs, used to represent the relative reliability of different modes for fault anomalies in the current power distribution room operation and maintenance scenario; the aging compatible residual... Used to represent the deviation strength of the current multimodal characteristics relative to the normal aging pattern of the target device, and Limited to the [0,1] interval; when the number of abnormal residual features is greater and the magnitude of the excess is larger, and when they are concentrated in modes with strong fault directionality such as infrared hotspots, partial discharge acoustic patterns, and abnormal keyframe switch states, The higher the value, the better; when most features are degraded compatibility features. The lower;

[0049] For example, a feeder cabinet that has been in operation for 11 years exhibits a slight, uniform temperature rise. This temperature rise falls within the degradation compatibility reference range of samples in the same degradation state. Therefore, this temperature rise characteristic is classified as a degradation compatibility characteristic and suppressed. If the same equipment simultaneously exhibits localized spike-like hot spots and obvious partial discharge acoustic signatures, and both significantly exceed the normal fluctuation boundaries under the same degradation state, then this type of characteristic is classified as an abnormal residual characteristic and is included in the aging suppression multimodal feature vector. High weights are retained in the middle, and make Therefore, the input for subsequent coarse alignment in step two is not the undifferentiated multimodal features, but rather the aging suppression multimodal feature vector after equipment degradation compatibility residual analysis. ;

[0050] Subsequently, the aging suppression multimodal feature vector Input the multimodal knowledge base for power distribution room operation and maintenance, calculate the similarity between it and each multimodal knowledge fragment in the knowledge base, and filter out a candidate knowledge set according to a preset similarity threshold; the candidate knowledge set is the preliminary retrieval result of the coarse alignment stage, which may include equipment information, fault description, historical cases, handling steps, standard clauses, and associated image, audio or video clips related to the current input anomaly; the preset similarity threshold is used to filter knowledge fragments that are not sufficiently relevant to the input anomaly data, and its initial value can be set to 0.75, and can be adjusted according to the recall and precision in the historical retrieval results;

[0051] After obtaining the candidate knowledge set, a candidate fault entity set is further extracted or mapped from the candidate knowledge set; the candidate fault entity set is used to represent the fault type that the current input anomaly may correspond to, such as partial discharge of the feeder cabinet, busbar insulation breakdown, CT secondary open circuit, joint overheating, etc.; the candidate fault entity set can be obtained by mapping the fault type field bound in the candidate knowledge set, the fault class entity ID in the knowledge graph, or the device-fault relationship edge.

[0052] It should be noted that the candidate knowledge set is a broad knowledge result that includes information such as equipment, faults, handling, and specifications, while the candidate fault entity set is the set of entities that directly correspond to the fault type; the subsequent calculation of the candidate discreteness Div is based on the candidate fault entity set, not the entire candidate knowledge set.

[0053] To determine whether the current coarse alignment result clearly points to a specific fault type, this invention introduces a candidate dispersion (Div). The top K candidate fault entities are selected based on similarity from highest to lowest, forming a Top-K candidate fault entity set, and their similarity scores are obtained. The initial value of K can be 5, which is used to control the amount of subsequent fine alignment calculation while ensuring the candidate recall capability. The candidate dispersion Div is calculated based on the similarity score dispersion of the Top-K candidate fault entities to quantify whether multiple candidate fault entities are difficult to distinguish.

[0054] When the similarity scores of the Top-K candidate fault entities are close to each other, it indicates that the current multimodal anomaly is close to multiple fault types at the same time, and it is difficult to determine the unique fault entity by relying solely on coarse alignment similarity. In this case, the candidate dispersion Div is high. When the similarity scores of the Top-K candidate fault entities differ greatly, it indicates that the candidate fault entity with the highest similarity is relatively clear, and the coarse alignment result is relatively stable. In this case, the candidate dispersion Div is low. Therefore, the candidate dispersion Div is used to characterize the semantic uncertainty of the fault in the coarse alignment stage and serves as one of the input parameters for generating the initial fault inference control quantity.

[0055] For example, if the Top-5 candidate fault entities corresponding to the input anomaly are partial discharge of the feeder cabinet, busbar insulation breakdown, CT secondary open circuit, joint overheating, and insulator flashover, and their similarity is concentrated between 0.81 and 0.91, it indicates that these fault entities are difficult to distinguish under the current multimodal features, and the candidate dispersion Div is high; if the similarity of partial discharge of the feeder cabinet is 0.93, while the similarity of other candidate fault entities is all below 0.60, it indicates that the coarse alignment result has clearly pointed to partial discharge of the feeder cabinet, and the candidate dispersion Div is low.

[0056] Candidate Discreteness Calculate using the following formula: ; In the formula, This represents the standard deviation of the Top-K similarity score set. This represents the mean of the Top-K similarity score set. It represents a stability constant to prevent the denominator from being zero or the logarithm from being abnormal;

[0057] When the calculated result of the candidate dispersion Div is less than 0, it can be truncated to 0; when the candidate dispersion Div exceeds the preset upper limit... At that time, it can be truncated to a preset upper limit. , The initial value can be 5 to avoid abnormal amplification of the initial fault deduction control quantity when the candidate similarity is extremely close; The calibration is based on the correspondence between candidate ambiguities and actual handling extension levels in historical fault samples;

[0058] The stability constant The initial value can be taken as When the system uses low-precision floating-point calculations or the mean of candidate similarity is small, Adjustable to to To ensure the stability of candidate dispersion calculation;

[0059] After completing the candidate dispersion calculation, the equipment degradation index is... , Aging compatible residual amount and candidate dispersion They are converted into initial fault simulation control quantities. The initial fault inference control variable is used to indicate the degree of necessity for the current coarse alignment result to continue entering the knowledge graph link fine alignment under the combined effects of equipment degradation state, aging compatibility residual and fault semantic uncertainty. The more degraded the equipment, the higher the aging compatibility residual, and the more difficult it is to distinguish candidate fault entities, the higher the risk of relying solely on coarse alignment similarity for output, and the more necessary it is to further verify through knowledge graph links and electrical topology boundaries.

[0060] The initial fault simulation control quantity is calculated according to the following formula: ; in, This represents the basic failure intensity constant, used to indicate the basic risk level for different equipment types or different anomaly types. The initial value can be taken as 1.0; for high-risk equipment or high-risk anomaly types, the value can be adjusted based on historical fault severity, equipment voltage level, and risk classification by maintenance experts. Increase the risk level; for low-risk anomalies, the risk level can be raised. Lower; Indicates the equipment degradation index, Indicates the aging compatibility residual amount, Indicates the candidate dispersion; Before proceeding to step three, the fault propagation driving force is considered to be dimensionless, and its numerical scale can be uniformly calibrated based on historical fault samples or expert calibration. It is not actual physical energy, but a virtual quantity used to represent the degree of necessity for the coarse alignment result to continue into the knowledge graph link fine alignment and topology boundary inference. Its value needs to be on the same dimensionless comparison scale as the link attenuation in step three.

[0061] For example, for the same local voiceprint anomaly, if the equipment is relatively new, has low aging compatibility residuals, and the Top-1 candidate fault entities are prominent, the initial fault inference control quantity is low, indicating that the coarse alignment result is relatively clear, and the subsequent knowledge graph link fine alignment does not need to be over-extended; if the equipment has been running for many years, has high aging compatibility residuals, and the Top-K candidate fault entities are close to multiple fault semantics at the same time, the initial fault inference control quantity is high, indicating that the candidate knowledge set needs to enter more rigorous link fine alignment and electrical topology boundary verification.

[0062] Step 3: Call the power distribution room operation and maintenance-specific knowledge graph to perform link integrity verification on the candidate knowledge set to obtain candidate links that meet the preset link rules; then combine the real-time electrical topology mask to perform electrical topology boundary constraints on the candidate links; for effective connected branches, obtain the link attenuation based on the normalized equivalent electrical impedance and the normalized protection time limit difference, and use the initial fault simulation control quantity as the starting quantity to simulate hops along the candidate links, updating the remaining simulation quantity of the current link; when the remaining simulation quantity of the current link is not greater than the simulation truncation threshold, stop the expansion and determine the maximum number of effective hops;

[0063] First, the dedicated knowledge graph for power distribution room operation and maintenance is invoked to verify the link integrity of the candidate knowledge set output in step two. This dedicated knowledge graph includes at least equipment entities, fault entities, operation entities, standard entities, and parameter entities, and contains core relationships such as equipment type—fault type, image feature—fault type, audio feature—fault type, fault type—handling steps, and handling steps—standard basis. Candidate knowledge that can form a complete business link from equipment to fault to handling to standard is selected, while candidate knowledge that cannot be bound to handling steps or standard basis is filtered out. To enable joint constraints between the knowledge graph link and the real-time electrical topology mask, this embodiment pre-builds... Establish the binding relationship between the physical nodes of the equipment and the corresponding equipment access branches in the real-time electrical topology mask; the equipment access branches are the electrical connection branches corresponding to the target equipment accessing the bus, feeder, switch or protection device in the primary wiring relationship; after the candidate link passes the link integrity verification, first query the equipment access branch bound to it according to the equipment ID corresponding to the physical node of the equipment, and then perform hop-by-hop deduction along the valid connected branches marked as 1 in the real-time electrical topology mask; if the expansion direction corresponding to the candidate link involves the disconnected branch, maintenance isolation branch or blocking isolation branch marked as 0 in the real-time electrical topology mask, then the branch is regarded as an uncrossable electrical topology boundary, and the candidate link is prohibited from continuing to expand along this direction;

[0064] Therefore, the jump in the knowledge graph is not directly equivalent to arbitrary semantic jump. Instead, it is first mapped to the effective connected branch in the real-time electrical topology mask through the preset binding relationship between the device entity node and the corresponding device access branch. Then, it is determined whether to continue the deduction based on the branch connectivity status and link attenuation.

[0065] Among them, the link completeness is used to indicate whether the candidate knowledge simultaneously includes the device entity, the fault entity, the handling step entity, and the standard basis entity. When the link completeness meets the preset requirements, that is, when the candidate link at least simultaneously includes the device entity, the fault entity, the handling step entity, and the standard basis entity, and when there are preset effective associations between the above entities in the knowledge graph, the candidate knowledge will enter the subsequent topology decay deduction.

[0066] For example, if a candidate knowledge can only match the description of a partial discharge fault, but cannot be further matched with the corresponding handling steps and standard basis in the knowledge graph, then the candidate knowledge cannot be used as the final fine alignment result; if another candidate knowledge can form a complete link of feeder cabinet - partial discharge - local isolation and verification - corresponding operation and maintenance standard clauses, then the candidate knowledge will enter the subsequent electrical topology boundary verification.

[0067] Please see Figure 3 As shown, the real-time electrical topology mask output in step one is then used. As an electrical topology boundary constraint, topology boundary verification is performed on candidate links that have passed the link integrity check; for For the valid connected branches marked as 1, read the corresponding line length, impedance per unit length, switch status, protection device setting value, and time difference between adjacent protection devices; for Isolated branches marked as 0 are treated as uncrossable boundaries and not as valid connected branches. When numerical calculations are required, the impedance corresponding to the branch can be assigned a preset impedance cutoff constant to express the blocking effect of physical isolation on link expansion.

[0068] Because the equivalent electrical impedance and the protection time difference have different dimensions, they are not directly added together. Instead, the equivalent electrical impedance is first normalized to... Normalize the protection time limit difference to ; It can be determined based on the upper limit of the reference impedance of the same voltage level, the same type of line, or the same type of equipment. It can be determined based on the upper limit of the reference protection time limit difference for similar protection coordination scenarios; among which, It represents the dimensionless impedance after the equivalent electrical resistance is normalized by the reference impedance. The initial value of the reference impedance can be taken as the 95th percentile value of the historical sample of the line impedance of the same voltage level. This represents the dimensionless protection time limit difference after normalization to a reference time limit. The initial value of the reference time limit can be taken as the 95th percentile of historical samples of time limit differences for similar protection scenarios. and All are dimensionless quantities and are on a comparable scale;

[0069] Link attenuation Calculate using the following formula: ; in, Indicates the first Level node to the first Link attenuation between nodes; This represents the normalized equivalent electrical impedance of the corresponding branch. This indicates the normalized protection time limit difference for the corresponding branch. Indicates the impedance contribution weight. This represents the contribution weight of the normalized protection time limit difference, and and Satisfy normalization constraints; during initial implementation, 0.6 is acceptable. A value of 0.4 can be taken, as both conditions are satisfied. The initial value is based on the fact that the extension of fault links in the power distribution room is usually primarily affected by electrical connectivity and line impedance, and the protection time limit difference is used as an auxiliary isolation constraint. It can also be calibrated based on the results of on-site review.

[0070] For example, if a candidate link extends from the feeder cabinet to the protection node of the current busbar, and this branch is in a closed connection state with a certain normalized equivalent electrical impedance and normalized protection time limit difference, then its link attenuation can be calculated. If the candidate link continues to extend to the direction of the next higher bus tie, but the current bus tie switch is in an open state, then this direction is marked as an uncrossable boundary by the real-time electrical topology mask, and even if there is a semantic association in the knowledge graph, it will not continue to be used as an effective fine-aligned link extension.

[0071] After completing the link attenuation calculation, the initial fault simulation control quantity generated in step two is used. As the starting point for topology attenuation deduction, the deduction is performed hop-by-hop along candidate links that have passed the knowledge graph link integrity check and meet the real-time topology connectivity conditions; where... Includes aging compatibility residues The intensity of the abnormal deviation characterized when At higher levels, The corresponding improvement allows the candidate links to obtain more sufficient topology attenuation verification in step three; when At lower levels, The corresponding reduction makes it easier for candidate links corresponding to ordinary aging disturbances to be truncated at reasonable electrical topology boundaries; for each effective connected branch crossed, the link attenuation corresponding to that branch is deducted. The remaining inference quantity of the current link is the remaining inference quantity obtained by subtracting the link attenuation amount that has crossed effective connected branches from the initial fault inference control quantity when inferring the candidate link hop by hop to the current node; at the k-th level node, the remaining inference quantity of the current link can be expressed as the remaining inference quantity of the k-th level node, and the remaining inference quantity of the k-th level node is obtained. ;

[0072] When the deduction reaches the [number]th When the node is at level 1, the remaining inference quantity is calculated according to the following formula: ; in, Indicates the deduction up to the 1st The remaining inference amount at the level node. This represents the initial fault simulation control variable. Indicates the first Level node to the first Link attenuation between nodes;

[0073] It should be noted that, It is not the actual resistance value, but a dimensionless link attenuation formed by the normalized impedance factor and the normalized protection time factor, which is used to be on the same comparison scale as the initial fault simulation control quantity and the remaining simulation quantity.

[0074] After each hop expansion is completed, immediately compare the remaining inference quantities of the k-th level node. With the deduced cutoff threshold Among them, the deduced cutoff threshold This represents the threshold of remaining inference quantity for determining whether the knowledge graph link should continue to expand; its initial value can be 0. Less than or equal to When the link attenuation occurs, it indicates that the necessity for further link expansion has been offset by the link attenuation, and the initial fault prediction control quantity formed by the equipment degradation state, aging compatibility residual, and candidate dispersion in step two has been attenuated within the current electrical topology boundary. Therefore, expansion along this direction is stopped, and the current hop count is determined as the maximum effective hop count in this direction. ;when Greater than If this occurs, it indicates that the initial fault simulation control quantity formed in step two has not been completely offset by the link attenuation on the current path, and the link still needs to be further extended for verification.

[0075] Deducing the cutoff threshold Calibration can also be performed based on the actual impact level and the distribution of remaining inference quantities in historical fault samples;

[0076] For example, using partial discharge of the feeder cabinet as a candidate fault entity, a link is first formed in the knowledge graph: feeder cabinet—partial discharge—local isolation and verification—corresponding specification clauses; if aging compatibility residual A higher level indicates that the partial discharge sound pattern or localized hot spot deviates significantly from the normal aging pattern of similarly degraded equipment. The initial jump to the protection node of this busbar is relatively high. It may still be greater than The protection node of this level busbar remains in the effective link; if A lower value indicates that the current anomaly is closer to normal aging perturbations. Lower It is easier to drop to the current level node. This avoids extending ordinary aging disturbances into excessively long fine-alignment service links;

[0077] To prevent unbounded searches caused by abnormal parameters or graph relationships, a maximum allowed number of search hops is also set. Even if the remaining projection quantity The deduction truncation threshold has not yet been reached; when the deduction hop count reaches the maximum allowed search hop count... At that time, further expansion will also cease; maximum allowed search hops. It can be preset according to the equipment level, protection configuration level or operation and maintenance boundary of the power distribution room; during the initial implementation, the maximum allowable number of search hops can be 3 to 5, preferably 5; this initial value is based on the fact that the common handling links in the power distribution room usually do not exceed the limited levels of this equipment, this level of protection, the upper level protection and adjacent related equipment;

[0078] When there are multiple candidate fault entities or multiple candidate links, link integrity verification, real-time topology constraint verification, topology attenuation inference and dynamic boundary truncation are performed on each candidate link respectively, and the final fine alignment result is determined by first hard filtering and then comprehensive sorting.

[0079] The hard filtering includes: if a candidate link cannot simultaneously cover the device entity, fault entity, handling procedure entity, and standard basis entity, then the candidate link is eliminated; if the device entity node corresponding to the candidate link cannot be mapped to the corresponding device access branch in the real-time electrical topology mask, then the candidate link is eliminated; if the candidate link crosses a disconnected branch, maintenance isolation branch, or blocking isolation branch marked as 0 in the real-time electrical topology mask during the simulation, then the candidate link is eliminated; if the candidate link cannot be bound to a valid standard basis, then the candidate link is eliminated.

[0080] For candidate links that pass the hard filtering, their fine alignment score is calculated. The fine alignment score is obtained by weighting feature similarity score, link integrity score, handling specification matching score, topology consistency score, and remaining inference quantity decay score. Among them, the feature similarity score is determined based on the similarity between the candidate knowledge fragment corresponding to the candidate link and the aging suppression multimodal feature vector; the link integrity score is determined based on whether the candidate link completely covers the device entity, fault entity, handling step entity, and specification basis entity; the handling specification matching score is determined based on whether the handling step is bound to the corresponding specification basis and whether the specification basis is applicable to the target device type; the topology consistency score is determined based on whether the candidate links all extend along the effective connected branches marked as 1 in the real-time electrical topology mask; the remaining inference quantity decay score is determined based on whether the remaining inference quantity of the current link decreases hop by hop with the effective electrical topology level and triggers truncation near the inference truncation threshold.

[0081] In one implementation, the fine alignment score is determined according to the following rules: Fine Alignment Score = a1 × Feature Similarity Score + a2 × Link Integrity Score + a3 × Handling Specification Matching Score + a4 × Topology Consistency Score + a5 × Remaining Calculation Attenuation Score; where a1, a2, a3, a4, and a5 are score weights, and a1 + a2 + a3 + a4 + a5 = 1. Initially, a1 can be 0.3, a2 can be 0.2, a3 can be 0.2, a4 can be 0.2, and a5 can be 0.1; the above weights can be adjusted based on the results of on-site review; the link integrity score, handling specification matching score, and topology consistency score can be assigned a value of 1 or 0 depending on whether the corresponding conditions are met; the feature similarity score is obtained by normalizing the similarity calculation results; the remaining calculation attenuation score is determined based on the proximity of the position where the remaining calculation decreases hop by hop and triggers truncation to the preset reasonable truncation range;

[0082] All candidate links that pass the hard filter are sorted from high to low according to their fine alignment scores, and the candidate link with the highest score is selected as the final fine alignment result. When the difference between the fine alignment scores of two candidate links is less than the preset score difference threshold, the candidate link with the higher handling specification matching score is selected first. If the handling specification matching scores are still the same, the candidate link with the higher topology consistency score and the smaller maximum effective hop count is selected first, so as to avoid outputting an overly extended handling range when there is insufficient fault evidence.

[0083] The final fine alignment result should include at least the target device, faulty entity, handling steps, standard basis, associated multimodal evidence, link node sequence, maximum effective hop count, effective handling scope, prohibited extension nodes, fine alignment score, and knowledge source traceability information.

[0084] Step 4: Output and trace the alignment results determined in Step 3, and after on-site review, write back and calibrate the topology attenuation model based on the comparison results of the actual fault impact level and the maximum effective hop count to adjust the relevant parameters of link attenuation.

[0085] Specifically, the alignment results will be output in a collaborative format of text and associated multimodal data of power distribution room operation and maintenance. The output content should include at least the target equipment, the fault entity corresponding to the multimodal data of power distribution room operation and maintenance, the handling steps, the standard basis, associated images, associated audio, associated video keyframes, link node sequences, and the effective handling scope and prohibited expansion nodes. For operation and maintenance Q&A scenarios, the above content will be sent to the Q&A generation module to generate structured Q&A results. For on-site handling scenarios, the above content can be organized into standardized handling suggestions or operation ticket auxiliary information.

[0086] For example, for the partial discharge fine alignment result of the feeder cabinet, the output results may include: the fault diagnosis is partial discharge of the feeder cabinet, and the associated evidence includes infrared heat map, partial discharge acoustic print fragments and key frames of on-site video; the handling steps include on-site isolation, verification of discharge location, and inspection of the protection device of this level bus; the standard basis includes the corresponding operation and maintenance procedure clauses; the prohibited expansion nodes include branch associated nodes that are in the disconnected or isolated state in the current real-time electrical topology mask, as well as peripheral nodes beyond the maximum effective number of hops;

[0087] Each alignment result is assigned a unique traceability ID and bound to the knowledge graph link ID, knowledge base data fragment ID, input power distribution room operation and maintenance multimodal data ID, and candidate knowledge set. , , sequence, And the final output content; through the above binding relationship, operation and maintenance personnel or auditors can reverse verify which image, audio, video, text specifications and knowledge graph links the result comes from, thereby meeting the requirements of power distribution room operation and maintenance for the authenticity, compliance and traceability of the result; for example, the unique traceability ID can be generated in the format of date stamp + equipment asset number + task sequence number, such as 20260509-FXG001-003, where the date stamp represents the output date, the equipment asset number represents the target equipment, and the task sequence number represents the alignment result sequence number under the same equipment on the same day;

[0088] Furthermore, the topology attenuation model can be rewritten and calibrated based on the on-site review results; if on-site experts or fault review records confirm the actual fault impact level... ,Will The maximum number of valid hops output in step three A comparison was made; among which, the actual fault cascade level was considered. This indicates the furthest level affected by the actual fault confirmed during the on-site review, and its level numbering rule is the same as in step three. Maintain consistency;

[0089] when and When they are consistent, it means that the results of this fine-alignment of the link boundaries are consistent with the field results, and the parameters of the topology attenuation model remain unchanged; when Greater than When this occurs, it indicates that the model prematurely truncates the link, and the current link attenuation may be too high; when Less than If this occurs, it indicates that the model has been extended too far, and the current link attenuation may be too small.

[0090] To avoid model oscillations caused by single field feedback, the topology decay model parameters are not immediately modified due to a single deviation. Instead, a cumulative trigger condition is set. Parameter calibration is only initiated when the cumulative number of simulations for similar equipment or similar fault scenarios reaches a preset number, and the number of hierarchical deviations reaches a preset number of deviations. The initial value of the cumulative simulation threshold can be 10, and the initial value of the hierarchical deviation threshold can be 3. These initial values ​​serve as the calibration basis to avoid frequent model oscillations caused by single false alarms, misoperations, or individual field recording deviations.

[0091] The topology attenuation model includes synthetic parameters for link attenuation. During calibration, the attenuation components to be adjusted are determined based on the contribution ratio of each attenuation component to the link attenuation in this batch of simulation records, and the corresponding weights of the attenuation components to be adjusted are adjusted. The attenuation components include normalized equivalent electrical impedance components and normalized protection time difference components. The sum of the adjusted impedance contribution weight and the protection time difference contribution weight remains equal to 1. The overall attenuation scaling factor is used for overall amplification or reduction. The numerical level is applicable to situations where field feedback indicates that the model is prematurely truncated or excessively extended, but the relative contributions of impedance factors and protection level factors are not significantly different.

[0092] For example, if multiple on-site reviews show that partial discharge faults in the same type of feeder cabinet are cut off in advance, and the error mainly occurs on lines with large normalized protection time limit differences, it indicates that the normalized protection time limit difference component may cause excessive attenuation, and the contribution weight of the corresponding component can be appropriately reduced; if multiple reviews show that it often extends to unnecessary peripheral nodes, the corresponding attenuation component or the overall attenuation scale can be enhanced to make the subsequent link boundary more convergent.

[0093] The output of this step includes: fine alignment results for the question-and-answer generation module or the operation and maintenance module, unique traceability ID, knowledge source traceability information, inference process record, and model parameters calibrated after on-site feedback.

[0094] The above parameters can be calibrated based on historical samples of similar equipment, operation and maintenance procedures, or on-site review results. The preset parameters in the formula should be set by technical personnel in the field according to the actual situation.

[0095] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0096] Those skilled in the art will recognize that the modules and algorithm modules of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0097] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0098] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0099] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for cross-modal alignment of knowledge graphs across the entire operation and maintenance chain of a power distribution room, characterized in that, Includes the following steps: The system obtains multimodal feature vectors of power distribution room maintenance equipment, obtains equipment degradation index based on equipment ledger, and obtains real-time electrical topology mask based on switch status; it then suppresses aging background disturbances of multimodal feature vectors based on equipment degradation index to obtain aging-suppressed multimodal feature vectors. The aging suppression multimodal feature vector is input into the power distribution room operation and maintenance multimodal knowledge base. A candidate knowledge set is obtained according to the preset similarity threshold, and a candidate fault entity set is extracted. The candidate discreteness is obtained by the inverse measure of the dispersion of the similarity scores of each candidate fault entity. The aging compatibility residual is obtained by the deviation intensity of the abnormal residual features determined in the aging background disturbance suppression process relative to the normal aging law of the target equipment. The initial fault simulation control quantity is obtained based on the equipment degradation index, aging compatibility residual, and candidate dispersion. The candidate knowledge set is checked for link integrity according to the preset link rules of the power distribution room operation and maintenance exclusive knowledge graph. Based on the preset binding relationship between the device entity node and the corresponding device access branch in the candidate knowledge, and combined with the real-time electrical topology mask connectivity conditions, the candidate links are obtained. Starting with the initial fault simulation control quantity, the simulation is performed hop-by-hop along the candidate links. The branches are marked as valid connected branches along the real-time electrical topology mask. The remaining simulation quantity of the current link is updated according to the link attenuation of the branch. The expansion stops when the remaining simulation quantity of the current link is not greater than the simulation truncation threshold, and the current number of hops is determined as the maximum number of valid hops. The fine alignment result is output within the link range limited by the maximum number of valid hops.

2. The method for cross-modal alignment of knowledge graphs across the entire operation and maintenance chain of a power distribution room according to claim 1, characterized in that, The equipment deterioration index is obtained as follows: the operational duration and average annual heavy load frequency of the target equipment are read from the equipment ledger data. The data are normalized according to the upper limit of the service life of similar equipment and the upper limit of the heavy load frequency of similar equipment. The two normalized results are linearly combined according to the contribution weight of operational duration and the contribution weight of heavy load frequency, and the smaller value of the result is taken as the output of the equipment deterioration index. The sum of the contribution weight of operational duration and the contribution weight of heavy load frequency is equal to 1. The value range of the equipment deterioration index is [0,1].

3. The method for cross-modal alignment of knowledge graphs across the entire operation and maintenance chain of a power distribution room according to claim 1, characterized in that, The real-time electrical topology mask is obtained by reading the real-time opening and closing status of circuit breakers, disconnectors, load switches, and bus tie switches from the SCADA system, distribution automation system, or station monitoring, and generating a Boolean adjacency matrix as the real-time electrical topology mask in combination with the primary wiring relationship of the distribution room. For branches that are currently closed and electrically connected, they are marked as 1 in the adjacency matrix and used as valid connected branches when expanding candidate links. For branches that are currently disconnected, under maintenance isolation, or blocked isolation, they are marked as 0 in the adjacency matrix and used as uncrossable electrical topology boundaries.

4. The method for cross-modal alignment of knowledge graphs across the entire operation and maintenance chain of a power distribution room according to claim 1, characterized in that, The generation method of the aging suppression multimodal feature vector is as follows: the equipment degradation state of the target equipment is determined according to the equipment degradation index, and the normal fluctuation boundary of each feature in the equipment degradation state is determined based on the historical normal operation samples of similar equipment corresponding to the equipment degradation state. The interval defined by the normal fluctuation boundary is used as the degradation compatibility reference range of the corresponding dimension. For a certain dimension of the multimodal feature vector, the feature value of that dimension in the current sampling window is normalized to obtain the normalized feature energy, and the normalized feature energy is compared with the degradation compatibility reference range of the corresponding dimension; when the feature falls into the degradation compatibility reference range, it is classified as a degradation compatibility feature and assigned a low retention weight. When a feature exceeds the degradation compatibility reference range but the extent of the exceedance does not reach the preset abnormal residual threshold, it is classified as a critical deviation feature and assigned an intermediate retention weight. When a feature exceeds the degradation compatibility reference range and the extent of the exceedance reaches the preset abnormal residual threshold, it is classified as an abnormal residual feature and given a high retention weight; after dimensional weighting, an aging suppression multimodal feature vector with the same dimension as the multimodal feature vector is obtained. The preset abnormal residual threshold is determined based on the normal fluctuation boundary width of this feature.

5. The method for cross-modal alignment of knowledge graphs across the entire operation and maintenance chain of a power distribution room according to claim 4, characterized in that, The aging compatibility residual is obtained by normalizing the abnormal residual features obtained from the division, based on the number of abnormal residual features, the extent to which the abnormal residual features exceed the degradation compatibility reference range, and the reliability weight corresponding to the mode to which the abnormal residual features belong. The value range of the aging compatibility residual is [0,1], which is the quantitative result of the deviation intensity of the multimodal feature vector from the normal aging law of the target equipment.

6. The method for cross-modal alignment of knowledge graphs across the entire operation and maintenance chain of a power distribution room according to claim 1, characterized in that, The candidate dispersion is determined as follows: The top K candidate fault entities are selected from the candidate fault entity set in descending order of similarity to form a Top-K candidate fault entity set, and their similarity score set is obtained. The ratio between the sum of the standard deviation and the stability constant of the Top-K similarity score set, and the sum of the mean and the stability constant of the Top-K similarity score set, is used as the stabilization coefficient of variation. The negative logarithm of the stabilization coefficient of variation is taken to obtain the candidate dispersion. When the candidate dispersion is less than 0, it is truncated to 0; when the candidate dispersion exceeds a preset upper limit, it is truncated to the preset upper limit. The stability constant is a positive number, used to prevent the denominator of the coefficient of variation from being zero or the logarithmic operation from being abnormal when the mean or standard deviation of the Top-K similarity score set approaches 0. The candidate dispersion is used to characterize the degree of similarity between the Top-K candidate fault entities and the degree of semantic indistinguishability of the faults.

7. The method for cross-modal alignment of knowledge graphs across the entire operation and maintenance chain of a power distribution room according to claim 1, characterized in that, The preset link rules include association rules between device entities, fault entities, handling step entities, and standard basis entities; when performing link integrity verification, candidate knowledge that can form a complete business link of device-fault-handling-standard is selected, and candidate knowledge that cannot be bound to handling steps or standard basis is filtered out.

8. The method for cross-modal alignment of knowledge graphs across the entire operation and maintenance chain of a power distribution room according to claim 1, characterized in that, The link attenuation is determined as follows: For a valid connected branch marked as 1 in the real-time electrical topology mask, the equivalent electrical impedance of the branch is first normalized to a normalized equivalent electrical impedance according to the reference impedance, and the protection time limit difference between adjacent protection devices of the branch is normalized to a normalized protection time limit difference according to the reference time limit. The link attenuation of the branch is calculated by multiplying the impedance contribution weight by the normalized equivalent electrical impedance and the protection time difference contribution weight by the normalized protection time difference. The sum of the two products is used as the link attenuation of the branch, where the sum of the impedance contribution weight and the protection time difference contribution weight is equal to 1. For isolated branches marked as 0 in the real-time electrical topology mask, they are treated as uncrossable electrical topology boundaries and do not participate in the update process of the remaining inference quantity of the current link.

9. A method for cross-modal alignment of knowledge graphs across the entire operation and maintenance chain of a power distribution room, as described in claim 8, is characterized in that... During the hop-by-hop simulation using the initial fault simulation control quantity, a maximum allowable number of search hops is set as a fallback constraint. When the number of simulation hops reaches the maximum allowable number of search hops, the expansion along that direction is stopped regardless of whether the remaining simulation quantity of the current link is greater than the simulation truncation threshold.

10. A method for cross-modal alignment of knowledge graphs across the entire operation and maintenance chain of a power distribution room, as described in claim 8, is characterized in that... After outputting the fine alignment results, the topology attenuation model is rewritten and calibrated based on the on-site review results. The topology attenuation model includes the synthetic parameters of link attenuation. Obtain the actual fault impact level confirmed by on-site review, and compare the actual fault impact level with the maximum effective number of hops to obtain the level deviation; when the actual fault impact level is greater than the maximum effective number of hops, the model is prematurely truncated, and the attenuation component to be adjusted is determined based on the contribution ratio of each attenuation component to the link attenuation in this batch of simulation records, and the weight of the attenuation component to be adjusted is reduced; when the actual fault impact level is less than the maximum effective number of hops, the model is determined to have expanded too far, and the attenuation component to be adjusted is determined based on the contribution ratio of each attenuation component to the link attenuation in this batch of simulation records, and the weight of the attenuation component to be adjusted is increased; wherein, the attenuation component includes the normalized equivalent electrical impedance component and the normalized protection time difference component; the sum of the adjusted impedance contribution weight and the protection time difference contribution weight is kept equal to 1.

Images