Intelligent diagnosis method for electrical equipment primitive state recognition and rule fusion

Abstract

The invention discloses an intelligent diagnosis method for power equipment primitive state recognition and rule fusion, and relates to the technical field of power system dispatching automation. The method is used for substation dispatching master station monitoring picture equipment primitive state recognition and abnormity diagnosis, video and graph model library data acquisition, preprocessing denoising, frame synchronization and ROI extraction. Using the improved YOLOv8-Tiny to extract a multi-modal feature to identify a primitive state; establishing an expert experience rule base, a graph-model association rule base and an anomaly judgment rule base; performing primary diagnosis on the forward chain reasoning fusion rule; performing cross checking on telecommand consistency, telemetering relevance and graph model library integrity; and outputting a result coexistence log. The problems that manual checking is low in efficiency and prone to omission are solved, the reliability of primitive recognition and correlation checking is improved, a D5000 system and a domestic operating system are adapted, complex monitoring scenes are coped, the operation and maintenance burden is relieved, it is guaranteed that monitoring pictures of a dispatching master station are accurate, and safe operation and maintenance of a power grid are facilitated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system dispatch automation, and in particular to an intelligent diagnosis method for power equipment graphic element state recognition and rule fusion. BACKGROUND

[0002] With the large-scale popularization of unattended substations in the power system, real-time monitoring of the jurisdictional substations by the dispatching centers at all levels has become a key prerequisite for the safe and stable operation of the power grid, and the accuracy of the dispatching master station monitoring screen directly determines the monitoring effect. The current signal checking and monitoring screen acceptance work on the dispatching master station side still takes manual correction as the core method - the operation and maintenance personnel need to compare the real-time power graphic screen (including the main wiring diagram, interval diagram, and light board diagram) of the dispatching master station with the monitoring information table forwarded by the station side remote, and manually confirm the state ("open" or "closed") of the switch, switch, circuit breaker and other device graphics and the correctness of the associated signals. This method not only brings a great workload to the dispatching automation operation and maintenance personnel, but also is inefficient, especially in scenarios with a large number of substations and dense device graphics, it is easy to cause data deviation due to human oversight, such as light board placed in the wrong interval, inconsistent graphics description and protection signal table, etc., which poses a hidden danger to the safety of the power grid. At the same time, the different states of the remote signal in the power graphic screen correspond to different graphic features, with the "open" state graphics mostly in red and the "closed" state mostly in green, and the fault state graphics also have high-frequency flickering, which further increases the difficulty of manual identification, leading to frequent remote signal checking failures in the checking test.

[0003] Although the existing power system has carried out research and application of automatic monitoring information acceptance technology, the related technical solutions have obvious limitations and cannot cover the monitoring screen checking requirements. Some domestic research focuses on the correctness verification of the master station preposition data, such as through the cooperation of the substation side automatic acceptance device and the master station side function module to realize the automatic checking of remote signal / telemetry signal, or developing an integrated acceptance tool to verify the accuracy of the master station preposition information points, but none of them involves the automatic identification and associated verification of the graphics in the monitoring screen, and manual confirmation of whether the graphics are correctly drawn and whether the graphics and signals are correctly associated is still needed. Although foreign related research has introduced substation test software (such as Siemens IEDSOUT software) and test instruments (such as ABB substation test instrument), such products mostly require manual operation and have single functions, only capable of simulating message sending, receiving or simple function testing, and cannot realize the automatic identification and dynamic verification of the graphics, nor can they adapt to the domestic D5000 dispatching master station system and IEC104 communication protocol, making it difficult to meet the needs of the domestic power system for fast checking of the monitoring screen.

[0004] In actual monitoring scenarios, image element recognition and correlation verification still face multiple technical bottlenecks. On the one hand, the industrial environment of substations is complex, and monitoring videos are easily affected by electromagnetic interference, light fluctuations (such as low illumination at night and strong light reflection during the day), and camera shake (such as wind and equipment vibration), resulting in blurred image elements and unclear features. Traditional image processing technologies are unable to effectively extract key features such as image element contours and colors, further increasing the difficulty of recognition. On the other hand, the correlation between images and remote signaling / telemetry signals in the monitoring screen depends on the image model library configuration. If the image element ID does not match the signal ID in the standard monitoring information point table, or if the interval where the image element is located does not match the actual interval of the equipment, it will lead to the monitoring failure problem of "the signal is correct but the image element is displayed incorrectly." Existing technologies lack automated verification methods for such correlation errors and can only rely on manual investigation, further exacerbating the risk to the accuracy of the monitoring screen and making it difficult to adapt to the intelligent and efficient operation and maintenance requirements of unattended substations for monitoring systems. Summary of the Invention

[0005] The present invention proposes an intelligent diagnostic method for power equipment element status recognition and rule fusion to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent diagnostic method for power equipment element status recognition and rule fusion, comprising:

[0007] Step 1: Data Acquisition and Preprocessing: Acquire data associated with the main station monitoring screen, including monitoring video data; acquire main wiring diagrams, interval diagrams, and light bar diagrams by zone, obtaining device elements marked with yellow boxes; simultaneously, back up the D5000 system output screen via HDMI video capture card, adding timestamps, camera numbers, and monitoring area identifiers during acquisition; during preprocessing, video data undergoes adaptive Gaussian filtering to remove noise, and video frames are aligned using frame synchronization technology to extract the ROI regions of the elements; remote signaling / telemetry data undergoes Kalman filtering to eliminate jitter, and the image model library data is categorized and organized by device type, interval, and element ID;

[0008] Step 2: Dynamic Status Recognition of Power Equipment Elements: Recognizing element status based on deep learning model and feature analysis: The YOLOv8-Tiny model is used to train the dataset containing common substation element samples; during model inference, the shape features, color features, and dynamic features of the elements are extracted; the element recognition results are output, including element ID, equipment name, current status, and recognition confidence.

[0009] Step 3: Construction of the diagnostic rule base: Construct three types of diagnostic rules to cover element status, association, and anomaly determination, including expert experience rules, element association rules, and anomaly determination rules. All rules are stored in the format of rule ID-triggering condition-diagnostic conclusion-processing suggestion, and rule query and modification are supported.

[0010] Step four, diagnostic reasoning of rule fusion: adopt forward chain reasoning mechanism, combine rules to realize abnormal diagnosis, first match the graph element state recognition result of step two with the rule base of step three, trigger the rules that meet the conditions, then adopt weighted fusion strategy for the triggered scene, allocate weights according to rule priority and historical accuracy; finally generate preliminary diagnosis results, including abnormal type, involved equipment, triggered rule ID, abnormal occurrence time;

[0011] Step five, consistency check of diagnostic results and telesignaling / telemetering: cross check the preliminary diagnosis results with data, including telesignaling consistency check, telemetering correlation check, graph model library integrity check; mark the diagnosis results that pass the check as valid, and mark the ones that do not pass the check as to be reviewed and return to step four for re-reasoning;

[0012] Step six, output of diagnostic results and abnormal alarm: present the diagnostic results and push the abnormal alarm according to user demand, superimpose the display of graph element ROI area on the system interface, label abnormal type, triggered rule and confidence, mark the effective results with green and the to-be-reviewed results with yellow; generate diagnosis reports according to day / week / month, including abnormal statistics, high-frequency abnormal equipment list and rectification suggestions.

[0013] Further, it also includes:

[0014] Diagnosis rule confidence optimization step: fuse the confidences of different rules through weighted calculation, the formula is: In the formula, C is the diagnosis confidence after rule fusion, w1 is the rule weight of expert experience, w2 is the rule weight of graph model association, and w3 is the rule weight of abnormal judgment; C1 is the confidence of triggered expert experience rule, C2 is the confidence of triggered graph model association rule, and C3 is the confidence of triggered abnormal judgment rule.

[0015] Abnormal risk level evaluation step: the risk level is calculated by each factor, the formula is: In the formula, R is the abnormal risk value, a is the abnormal occurrence probability weight, b is the abnormal duration weight, and c is the abnormal influence range weight; P is the abnormal occurrence probability score, T is the abnormal duration score, and S is the abnormal influence range score.

[0016] Further, the step three further comprises a dynamic rule updating sub-step for updating the optimization rule library according to the field operation feedback and new scene; the updating process is: collecting the review feedback of the operation personnel on the diagnosis result every quarter, screening out the rules that need to be optimized; collecting the state characteristics of the newly added equipment graph elements, adding corresponding expert experience rules and abnormality judgment rules; adjusting the confidence benchmark of the original rules according to the diagnosis accuracy rate in the past six months; the updated rules need to be verified through the test environment, and the rule triggering frequency and accuracy rate are recorded after going online, which provides the basis for the next update.

[0017] Further, the step two further comprises a small sample graph element state compensation identification sub-step for dealing with the problem of insufficient new or rare graph element samples in the substation; the compensation identification process is: generating 1000 expanded samples for the small number of samples of the newly added graph elements by using data enhancement technology; extracting the core features of the expanded samples to establish a feature mapping table, which is associated with the rules of similar devices in the existing rule library; using transfer learning to fine-tune the graph element identification model, only updating the output layer and the last two convolutional layer parameters; the compensation identification result needs to be compared with the manual annotation, and actual running samples are continuously collected after application, and the feature mapping and rule association are gradually optimized.

[0018] Further, the step four further comprises a cross-interval rule checking sub-step for diagnosing abnormalities involving multiple interval associations; the checking process is to construct an interval association relationship table; when the interval graph element triggers abnormal diagnosis, the graph element state and telesign data of the associated interval are automatically retrieved; the cross-interval special rule is called to verify the diagnosis conclusion in combination with the interval data.

[0019] Further, it further comprises:

[0020] Diagnosis result tracing step: for tracking the data source and rule triggering path in the diagnosis process; the tracing content includes data tracing, rule tracing and reasoning tracing; the tracing information is stored through the diagnosis ID-tracing internal format, and is bound with the diagnosis result, supporting one-key query of the complete tracing chain through the diagnosis ID, and generating a tracing report.

[0021] Further, the step four further comprises a rule conflict resolution sub-step for solving the problem of contradictory conclusions between rules; the conflict resolution adopts priority weighting calculation, and the formula is: , wherein W is the rule priority weight, k1 is the rule usage frequency weight, k2 is the rule historical accuracy rate weight, and k3 is the expert rating weight; U is the rule usage frequency score in the past three months, F is the rule historical accuracy rate score, and H is the expert rating score.

[0022] Further, the step two further comprises a sub-step of pixel state compensation identification in a low-illumination environment, for solving the identification difficulty caused by poor video quality of the substation at night or in a dark corner; the compensation identification process is as follows: the Retinex algorithm is used to enhance the low-illumination video frame, the brightness of the enhanced image is increased by 3-5 times, and the contrast is increased by 2 times; for the color distortion problem in low-illumination, an HSV color space compensation model is established; the dynamic feature weight is increased; after compensation identification, the identification result is compared with that of the adjacent normal-illumination frame.

[0023] Further, it further comprises:

[0024] A user-defined rule interface step: for meeting the individualized diagnosis needs of different substations; the interface functions include rule editing, rule testing, rule online, and rule management; through the interface, the local device characteristics and operation and maintenance habits of the substation are adapted.

[0025] Compared with the existing technology, the application has the following advantages:

[0026] Firstly, the application replaces manual correction with automatic technology, greatly reduces the work burden of operation and maintenance personnel, avoids the correction deviation caused by manual omission, does not require operation and maintenance personnel to compare pixels and signals one by one, the system can automatically complete pixel identification, state determination and correlation verification, and covers various monitoring pictures such as main wiring diagram, interval diagram and light card diagram, greatly improves the correction efficiency, and ensures that the monitoring picture acceptance work can be quickly promoted.

[0027] Secondly, the application fills the gap of the existing monitoring information automatic acceptance technology, and realizes the change of monitoring picture correction from "manual" to "automation". The existing technology can only verify the correctness of the front-end data of the main station, and cannot judge the correlation between the pixel state and the correlation in the monitoring picture. The application combines the pixel identification algorithm and multi-dimensional rules, can not only accurately identify the pixel type and state, but also verify the correlation correctness of the pixel and the remote signal / telemetry signal, identify the problems such as pixel loss, correlation error and state mismatch, and ensure that the monitoring picture is highly consistent with the actual device state and signal data, providing a real and reliable monitoring basis for the dispatching center.

[0028] Thirdly, the application has strong environmental adaptability and can effectively cope with the technical challenges in the complex monitoring scene of the substation. For the problem of unclear pixel features caused by electromagnetic interference, light fluctuation and camera jitter in the substation, the system optimizes the image quality through video preprocessing technology (denoising, enhancement, and blur correction), combines multi-modal features (shape, color, and dynamic) to improve the accuracy of pixel identification, and avoids identification deviation caused by environmental factors; for the problem of insufficient new or rare pixel samples, small sample learning and feature mapping strategies are used to quickly adapt to new device pixels and ensure that the correction range is fully covered.

[0029] Finally, the application has good compatibility and practical value, can seamlessly adapt to the existing power system architecture. It can be directly connected to the existing substation monitoring network, without the need for large-scale modification of existing equipment, reducing deployment costs; at the same time, the system has the functions of diagnosis result tracing, operation and maintenance closed-loop management and user-defined rules, which can not only provide basis for operation and maintenance personnel to locate the root cause of the problem, but also meet the individual operation and maintenance needs of different substations, promote the development of dispatching master station monitoring system operation and maintenance towards standardization, precision and intelligence, and provide strong guarantee for the safe and stable operation of power grid. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 A schematic block diagram of the intelligent diagnosis method for power equipment graph element state recognition and rule fusion proposed by the application;

[0031] Figure 2 A comparison diagram before and after confidence optimization of graph element recognition of the intelligent diagnosis method for power equipment graph element state recognition and rule fusion proposed by the application;

[0032] Figure 3 A monitoring picture checking error type distribution diagram of the intelligent diagnosis method for power equipment graph element state recognition and rule fusion proposed by the application. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.

[0034] In the description of the application, it should be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the application.

[0035] Moreover, the terms "first", "second", etc. are used only for the purpose of description and do not imply or imply the relative importance or implicitly indicate the number of the technical features indicated. Therefore, the features defined as "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise specifically limited. In addition, the terms "mounting", "connection", "connection" should be broadly understood, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be connected inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances, and the present application will be further described in detail below with reference to the accompanying drawings.

[0036] Reference Figures 1 to 3 An intelligent diagnosis method for power equipment graphic element state recognition and rule fusion is used for state recognition and abnormal diagnosis of switch, switch, circuit breaker, light board and other equipment graphic elements in the dispatching master station monitoring screen of the substation. Relying on multi-source data acquisition, graphic element feature analysis and multi-dimensional rule fusion, it supports the rapid checking of the dispatching master station monitoring screen. The method includes six steps:

[0037] Step 1: Multi-source data acquisition and preprocessing. Collecting related multi-source data of the dispatching master station monitoring screen, including: monitoring video data (collecting the main wiring diagram, interval diagram, light board diagram video through high-definition industrial camera (resolution 1920x1080, frame rate 25fps, low illumination ≤0.01Lux), focusing on capturing key equipment graphic elements marked with a yellow frame, and simultaneously collecting the D5000 system output screen through the HDMI video capture card, collecting the millisecond level time stamp, camera number and monitoring area identifier simultaneously); remote signaling / telemetry data (receiving real-time remote signaling values (split = 0, combined = 1) and telemetry values (such as current, voltage) uploaded by the substation through IEC104 protocol, data sampling frequency 1 / time / second, ensuring time synchronization with video data); graphic model library configuration data (exporting standard monitoring information point table (including remote signaling / telemetry point number, device description, associated graphic element ID) from D5000 system, graphic element and power grid model mapping relationship data). During preprocessing, video data uses adaptive Gaussian filter (σ value 0.8-1.5, dynamically adjusted according to light) to remove noise, and multi-source video frames are aligned through frame synchronization technology (based on NTP protocol, synchronization error ≤10ms) to extract graphic element ROI area (according to yellow frame marker coordinates, positioning accuracy ≤1 pixel); remote signaling / telemetry data uses Kalman filter to eliminate jitter, and graphic model library data is classified and arranged according to "device type-interval-graphic element ID" to ensure uniform data format.

[0038] Step two: dynamic state recognition of power equipment graphics. Based on deep learning model and multi-modal feature analysis to recognize graphic state: improved YOLOv8-Tiny model (lightweight design to ensure real-time performance) is used, and the training data set contains common graphic samples in substations (12000 frames of switches, 10000 frames of knife switches, 8000 frames of circuit breakers, 5000 frames of light boards, each type of graphic contains three states of "on / off / fault"); when the model is inferred, the multi-modal features of the graphics are extracted: shape features (contour is extracted by Canny operator, such as rectangular structure of switch and folded line structure of knife switch), color features (HSV color space modeling, "on" state red (H=0-10°, S=70%-100%, V=50%-100%), "off" state green (H=110-130°, S=70%-100%, V=50%-100%)), dynamic features (fault graphic flicker frequency 5-10Hz, normal state without flicker); output graphic recognition result, including graphic ID, device name, current state (on / off / fault), recognition confidence (initial confidence ≥90%), if confidence <90%, trigger secondary recognition (re-extract ROI features and strengthen dynamic feature weight).

[0039] Step three: construction of multi-dimensional diagnostic rule base. Three types of diagnostic rules are constructed to cover graphic state, correlation and abnormality judgment: expert experience rules (based on experience of substation operation and maintenance experts, such as "circuit breaker should be opened before opening two side knife switches", "light board flicker frequency >8Hz is judged as emergency fault", "when remote current exceeds rated value 120%, associated switch graphic should trigger alarm"); graphic module correlation rules (based on standard monitoring information point table, such as "graphic ID and remote signaling point number one-to-one correspondence", "graphic interval should be consistent with actual interval", "light board description should be completely matched with protection signal table"); abnormality judgment rules (set state threshold and correlation threshold, such as "remote signaling value and graphic state inconsistent for more than 3s is judged as state mismatch abnormality", "graphic is missing corresponding remote signaling point is judged as correlation missing abnormality", "light board does not flicker according to fault level is judged as display abnormality"), all rules are stored in the format of "rule ID-triggering condition-diagnostic conclusion-treatment suggestion", supporting rule query and modification.

[0040] Step four: intelligent diagnostic reasoning of rule fusion. Adopt forward chain reasoning mechanism, and realize abnormal diagnosis in combination with multi-dimensional rules: firstly, the graph element state recognition result in step two is matched with the rule library in step three, a rule meeting the condition is triggered (for example, the "operation logic abnormality" rule in the expert experience rule is triggered when it is recognized that "the circuit breaker is tripped but the switch is not tripped"); then, for a multi-rule triggering scene (for example, the "state mismatch" rule and the "description inconsistency" rule are triggered at the same time for the same graph element), a weighted fusion strategy is adopted, the weight is distributed according to the rule priority (expert experience rule > graph model association rule > abnormal judgment rule) and the historical accuracy rate (the rule diagnosis accuracy rate in the last three months), and the diagnosis conclusion of the rule with high priority and high accuracy rate is preferentially adopted; finally, the preliminary diagnosis result is generated, including the abnormal type (such as operation logic abnormality, association error, display abnormality), the involved device, the triggered rule ID and the abnormal occurrence time.

[0041] Step five: consistency check of the diagnosis result and the remote signaling / telemetry. The preliminary diagnosis result is cross-checked with multi-source data: remote signaling consistency check (compare the IEC104 remote signaling real-time value collected in step one with the graph element recognition state, if the remote signaling value is 1 (closed) and the graph element displays "open", it is confirmed that "state mismatch abnormality" occurs); telemetry association check (if the diagnosis result is "overcurrent fault", it is checked whether the corresponding telemetry current value exceeds the threshold, if yes, the fault diagnosis conclusion is strengthened, and if no, it is marked as "suspected misjudgment" and re-reasoned); graph model library integrity check (traverse the standard monitoring information point table, check whether the "graph element missing" abnormality in the diagnosis result is indeed missing the corresponding graph element, so as to avoid misjudgment caused by recognition omission). The diagnosis result that passes the check is marked as "valid", and the diagnosis result that does not pass the check is marked as "to be reviewed" and returned to step four for re-reasoning.

[0042] Step six: output of the diagnosis result and abnormal alarm. The diagnosis result is presented and the abnormal alarm is pushed according to the user demand: real-time display (the ROI area of the graph element is displayed on the system interface, the abnormal type, the triggered rule and the confidence are marked, the valid result is marked with green color, and the result to be reviewed is marked with yellow color); report generation (diagnosis report is generated on a daily / weekly / monthly basis, including abnormality statistics (the number and proportion of each type of abnormality), the list of high-frequency abnormal devices, rectification suggestions, and the support for PDF / Excel format export); abnormal alarm (for valid abnormalities, the alarm is notified through sound and light alarm (alarm sound + interface red flashing), short message push (to the mobile phone of the operation and maintenance personnel), the alarm response time is less than or equal to 1s); log storage (diagnosis records, triggered rules and check results are stored, the log contains time stamp, operator and processing result, the storage period is greater than or equal to 6 months, and the log can be searched according to the graph element ID and the abnormal type).

[0043] In the present application, the following are also included:

[0044] Diagnosis rule confidence optimization step: for improving the accuracy of diagnosis when multiple rules are triggered. The confidence of different rules is fused by weighted calculation, and the formula is: In the formula, C is the diagnosis confidence after rule fusion (0-1, C≥0.9 is determined as high reliable diagnosis, 0.7≤C<0.9 is determined as medium reliable, and C<0.7 is determined as low reliable), w1 is the weight of expert experience rule (0.45, the expert rule is based on field experience, and the reliability is the highest), w2 is the weight of graph model association rule (0.35, the association rule is based on the standard point table, and the accuracy is strong), and w3 is the weight of abnormality determination rule (0.2, the determination rule is a basic threshold judgment, and the priority is low); C1 is the confidence of triggered expert experience rule (0-1, C1=1 when the historical diagnosis accuracy rate of the rule is greater than or equal to 95%), C2 is the confidence of triggered graph model association rule (0-1, C2=1 when the rule matching degree is greater than or equal to 98%), and C3 is the confidence of triggered abnormality determination rule (0-1, C3=1 when the threshold deviation is greater than 50%). Through the calculation, the conclusion with high confidence can be selected preferentially when multiple rules conflict, and the misjudgment caused by single rule priority can be reduced.

[0045] Abnormal risk level evaluation step: for quantifying the influence of abnormality on power grid operation, and processing high-risk abnormality preferentially. The risk level is calculated by multiple factors, and the formula is: In the formula, R is the abnormal risk value (0-10, R≥8 is extremely high risk, 6≤R<8 is high risk, 4≤R<6 is medium risk, and R<4 is low risk), a is the abnormal occurrence probability weight (0.35, the probability is based on the frequency of similar abnormality in the past year, and P=10 when the frequency is greater than 10 times / year), b is the abnormal duration weight (0.3, T=10 when the duration is greater than 10 min, and T=2 when the duration is less than 1 min), and c is the abnormal influence range weight (0.35, S=10 when the influence is on core equipment such as transformer and bus, and S=3 when the influence is only on a single switch); P is the abnormal occurrence probability score (0-10, the higher the frequency, the higher the score), T is the abnormal duration score (0-10, the longer the time, the higher the score), and S is the abnormal influence range score (0-10, the wider the range, the higher the score). Extremely high-risk abnormality triggers emergency alarm (automatically upgraded to superior operation and maintenance within 5 minutes if not handled), and low-risk abnormality only records log, thereby improving the efficiency of operation and maintenance resource allocation.

[0046] In the application, the step three further comprises a dynamic rule updating sub-step for updating the optimization rule library according to the on-site operation feedback and new scene. The updating process is as follows: the review feedback of the operation personnel on the diagnosis result (such as "rule misjudgment" and "newly-added exception not covered") is collected every quarter, and the rules that need to be optimized (rules with a feedback misjudgment rate greater than 5%) are screened out; the state characteristics (such as a unique flicker frequency of 12-15 Hz) of the newly-added device graph elements (such as a new type of intelligent knife switch) are collected, and the corresponding expert experience rules and abnormal judgment rules are added; for the original rules, the confidence benchmark is adjusted according to the diagnosis accuracy rate in the past six months (for example, if the rule accuracy rate decreases from 95% to 90%, the C1 benchmark value is lowered from 1 to 0.95); the updated rules need to be verified in a test environment (simulate 1000 abnormal scenes, and the diagnosis accuracy rate is greater than or equal to 97% before going online), and the rule triggering frequency and accuracy rate are recorded after going online to provide a basis for the next update, and the substation device iteration and the change of the running scene are adapted.

[0047] In the application, the step two further comprises a small sample graph element state compensation identification sub-step for solving the problem of insufficient samples of newly-added or rare graph elements (such as special light cards and new type of ground knife) in the substation. The compensation identification process is as follows: for a small amount of samples (≥20 frames) of the newly-added graph elements, 1000 frames of expanded samples are generated by using data enhancement technology (rotation ±15°, scaling 0.8-1.2 times, brightness ±20%, and adding simulated industrial noise); the core features (such as the character shape of the special light card and the opening angle of the new type of ground knife) of the expanded samples are extracted, a feature mapping table is established, and the features are associated with the rules of similar devices in the existing rule library (such as associating the new type of ground knife with the "open / close" judgment rule of the traditional ground knife and adjusting the color feature threshold); the graph element recognition model is fine-tuned by using transfer learning, only the output layer and the last two convolution layer parameters are updated, and the original graph element recognition ability is retained; the compensation identification result needs to be compared with the artificial labeling (the accuracy rate is greater than or equal to 95% before being applied), and the actual running samples are continuously collected after the application, the feature mapping and rule association are gradually optimized, and the recognition missed detection in the small sample scene is reduced.

[0048] In the application, the step four further comprises a cross-bay rule checking sub-step for diagnosing complex abnormalities involving multi-bay correlation (such as bus fault leading to multi-bay switch correlation abnormality). The checking procedure is as follows: constructing a bay correlation table (based on the primary wiring diagram of the substation, marking the electrical connection relationship of each bay, such as 110kV I section bus correlation #1 main transformer bay, #1 outgoing line bay); when a certain bay element triggers abnormal diagnosis (such as #1 outgoing line bay switch abnormal opening), automatically retrieving the element state and telemetering data of the associated bay (such as checking 110kV I section bus voltage telemetering value, #2 outgoing line bay switch state); calling cross-bay special rules (such as "when the bus voltage is less than 80% of the rated value, the associated bay switch should be interlocked to open, and if it is not opened, it is determined as a linkage abnormality"), and verifying the diagnosis conclusion in combination with multi-bay data; if the associated bay data supports the diagnosis conclusion (such as the bus voltage is indeed low), the abnormality is determined; if it does not support (such as the bus voltage is normal), it is marked as "bay correlation abnormality" and the search range is expanded (such as checking the adjacent bus bay), avoiding misjudgment due to single bay data, and improving the diagnosis integrity in complex power grid scenarios.

[0049] In the application, it further comprises:

[0050] The diagnosis result tracing step is used for tracing the data source and rule triggering path in the diagnosis process, so as to facilitate the operation and maintenance personnel to locate the problem source. The tracing content includes: data tracing (recording the camera number, timestamp, preprocessing parameter of the video frame used for element state identification, substation IP, transmission protocol frame number of telemetering data); rule tracing (recording the rule ID, rule content, confidence calculation process called in the diagnosis process in the triggering order, and marking the reason for finally adopting the rule); reasoning tracing (recording each step of logical judgment of the forward chain reasoning, such as "triggering rule A first → calling rule B due to insufficient confidence → obtaining the conclusion after fusion"); the tracing information is stored in the format of "diagnosis ID-tracing content", is bound with the diagnosis result, supports one-key query of the complete tracing chain through the diagnosis ID, and generates a tracing report (including data flow chart, rule triggering time sequence chart), thereby providing a basis for the operation and maintenance personnel to analyze the misjudgment reason (such as improper data preprocessing parameter, unreasonable rule threshold), and improving the maintainability of the diagnosis system.

[0051] In the application, the step four further comprises a rule conflict resolution sub-step for solving the problem of conclusion contradiction when multiple rules are triggered (such as rule A determining "state mismatch" and rule B determining "element identification error"). The conflict resolution adopts priority weighted calculation, and the formula is: In the formula, W is a rule priority weight (value 0-100, the higher the weight, the higher the priority), k1 is a rule usage frequency weight (value 0.3, reflecting rule practicability), k2 is a rule history accuracy weight (value 0.4, reflecting rule reliability), and k3 is an expert rating weight (value 0.3, reflecting rule authority); U is a rule usage frequency score in the last 3 months (value 0-100, 500 times of usage is 100 points), F is a rule history accuracy score (value 0-100, accuracy rate ≥98% is 100 points), and H is an expert rating score (value 0-100, expert rating “excellent” is 100 points). The rule conclusion with the highest priority weight is preferentially adopted, and if the weight difference is less than 5, the consistency checking result of step five is combined to further determine, to ensure the rationality of the diagnosis conclusion in the case of rule conflict.

[0052] In the present application, the step two further comprises a sub-step of pixel state compensation identification in a low-illumination environment, for solving the identification difficulty caused by poor video quality of the substation at night or in a dark corner. The compensation identification process is as follows: the Retinex algorithm is used to enhance the low-illumination video frame (the image is decomposed into a reflection component and an illumination component, the illumination component is adjusted to improve the brightness, and the pixel details in the reflection component are retained), the brightness of the enhanced image is improved by 3-5 times, and the contrast is improved by 2 times; for the color distortion problem under low illumination, an HSV color space compensation model is established (for example, if the H value of the red pixel at night deviates by 5-10°, the identification threshold is adjusted to H=0-20°); the dynamic feature weight is increased (for example, the flicker frequency of the fault pixel is more easily identified under low illumination, and the dynamic feature weight is increased from 0.25 to 0.4); after compensation identification, the identification results of adjacent normal illumination frames are compared (if the results of 5 consecutive frames are consistent, the state is confirmed), to avoid misjudgment caused by poor single-frame enhancement effect, and to maintain the pixel identification accuracy rate at above 96% under low-illumination scenes.

[0053] In the present application, further comprising:

[0054] User-defined rule interface step: to meet the personalized diagnosis needs of different substations (such as special operation rules of specific equipment). The interface functions include: rule editing (providing a visual editing interface, supporting users to add trigger conditions (such as figure state, telemetry threshold), diagnostic conclusions, processing suggestions, supporting logical operators (and / or / not) to combine conditions); rule testing (after the user adds the rule, historical video and remote signaling data can be imported to simulate the trigger scene to test the effectiveness of the rule, and the test accuracy and misjudgment cases are output); rule online (after the test is passed, the rule is reviewed by the system administrator, added to the rule library, and assigned a weight (default 0.2, adjustable) and a confidence benchmark (default 0.9)); rule management (supporting users to query, modify, and disable custom rules, and recording rule usage logs). Through the interface, the substation can quickly adapt to the local equipment characteristics and operation and maintenance habits, such as the rule of "if there is still a weak current after the switch is opened, it is determined to be poor contact" for old switches, to improve the flexibility and applicability of the diagnosis method.

[0055] DETAILED DESCRIPTION OF THE INVENTION OF THE INTELLIGENT DIAGNOSIS METHOD BASED ON POWER EQUIPMENT FIGURE STATE RECOGNITION AND RULE FUSION

[0056] Example One: 110kV Conventional Substation Power Equipment Figure Intelligent Diagnosis Application

[0057] This example is aimed at a 110kV conventional substation (including 3 main transformers, 6 outgoing line intervals, and 2 capacitor banks), focusing on the state recognition and abnormal diagnosis of conventional equipment figures such as switches, switches, circuit breakers, and light boards, adapting to the D5000 dispatching master station system (version V2.0) of the substation, solving the problem of low efficiency and easy to miss of traditional manual checking.

[0058] Step One: Multi-source Data Acquisition and Preprocessing

[0059] The acquisition device is deployed according to the scene: 2 Hikvision DS-2CD5A26EFWD-A high-definition cameras are deployed in the main wiring diagram area (covering 110 kV, 35 kV, 10 kV voltage levels), with a resolution of 1920x1080, a frame rate of 25 fps, a low illumination of 0.005 Lux, a support of 120 dB wide dynamic, an installation height of 6 m, and a horizontal field of view of 90°, ensuring complete capture of busbars and main transformer associated graphics; 1 Dahua DH-IPC-HFW5831E close-up camera is deployed in each interval (6 outgoing line intervals + 3 main transformer intervals), installed 2 m above the interval control cabinet, focusing on key graphics such as circuit breakers (such as #1 main transformer 110 kV side circuit breaker) and knife switches (such as 10 kV I segment busbar knife switch); 1 Lingyun LNH-IPC-MC200 macro camera is deployed in the light board area (main control room light board screen), with a resolution of 2560x1440 and a lens focal length of 16 mm, ensuring clear and distinguishable text on the light board such as "#1 main transformer overcurrent protection action". All cameras are connected to the system host (Advantech IPC-610L, configured with Intel Core i7-12700K CPU and NVIDIA RTX 3060 GPU) through Huawei S5720-28X-PWR-LI-AC gigabit industrial switches, and are connected to Tianchuang Hengda TC-737N4 HDMI video capture card at the same time, which backs up the D5000 system output picture. When collecting, millisecond timestamps are added through the time module of Python, and camera numbers are named according to "area-number" (such as main wiring diagram-1, outgoing line interval-2), and monitoring area identifiers are associated with the device location code of the GIS system of the substation (such as "110 kV I segment busbar-001").

[0060] Remote signaling / telemetry data is uploaded by the substation through IEC104 protocol (port 2404), with a remote signaling value sampling frequency of 1 / s (split = 0, combined = 1) and a remote measurement value (current, voltage) sampling frequency of 5 / s, and time synchronization with video data is achieved through NTP time server (accuracy ±1 ms). The configuration data of the diagram library is exported from the D5000 system (Oracle 12c database), including 238 remote signaling / telemetry points (such as "1101" corresponding to #1 main transformer 110 kV side circuit breaker), device description, associated graphic ID (such as "CB-001"), and mapping relationship between graphics and power grid model, classified and sorted according to "device type-interval-graphic ID" (such as "circuit breaker-#1 main transformer interval-CB-001").

[0061] Preprocessing stage: The video data is filtered with adaptive Gaussian filter, with σ=0.8 for stable light during the day (8:00-18:00) and σ=1.5 for large light fluctuation at night (18:00-8:00 the next day), combined with db4 wavelet basis (3 decomposition levels) for denoising, implemented through the wden function of MATLAB; frame synchronization aligns multiple camera video streams based on the NTP protocol, with a synchronization error controlled within 8 ms; ROI extraction refers to the key element coordinates marked by the yellow frame in the customer chat record, such as setting the circuit breaker element ROI as (x=320, y=240, w=80, h=60), automatically framed through the cv2.rectangle function of OpenCV, with a positioning accuracy of ≤1 pixel; Retinex algorithm (single scale, σ=80) is used to improve the brightness in low-light scenes (such as at night), and CLAHE histogram equalization (clipLimit=2.0, tileGridSize=(8,8)) is used in the light board area to enhance the contrast between the light board text and the background by 3 times. Kalman filter is used to eliminate jitter in the remote signaling / telemetry data, with filter coefficients Q=0.01 and R=0.1, and the graphic model library data is cleaned through the pandas library of Python, eliminating null values and duplicates.

[0062] Step two: dynamic state recognition of power equipment elements

[0063] An improved YOLOv8-Tiny model is used, and the training data set is taken from 3-month video clips of the substation. Switch (12000 frames, including "open / close" state), switch (10000 frames), circuit breaker (8000 frames), and light board (5000 frames, including fault flicker) samples are selected. The LabelImg labeling tool is used to label the element type, state, coordinates, and feature parameters (color, flicker frequency). Model training is performed under the PyTorch framework, with batchsize=16, epoch=50, initial learning rate 0.001 (cosine annealing decay to 0.0001), and AdamW optimizer (weight decay 0.001).

[0064] During model inference, the ROI region is first extracted to extract multi-modal features: shape features are extracted using the Canny operator (threshold 50-150) to extract contours, such as the rectangular structure of the switch (width-height ratio 2:1) and the polyline structure of the knife switch (included angle 60°-90°); color features are calculated after converting RGB to HSV, with the "on" state being red H=5°±5°, S=85%±15%, V=70%±30%, and the "off" state being green H=120°±10°, S=85%±15%, V=70%±30%; dynamic features are calculated by frame difference method, with the frame difference absolute value being greater than 20 for more than 80% of the 10 consecutive frames, and the flicker frequency being 5-10 Hz (fault state). During the inference process, the SE attention mechanism is added to strengthen the graph element features (channel attention weight 0.6, spatial attention weight 0.4), and the output includes graph element ID, device name, current state, and recognition confidence (initial confidence ≥90%), with secondary recognition triggered when the confidence is less than 90% (re-extracting ROI features, and the dynamic feature weight is increased from 0.25 to 0.4).

[0065] Step three: multi-dimensional diagnostic rule base construction

[0066] The expert experience rules are based on the experience of 50 substation operation and maintenance experts (service length ≥10 years), with a total of 32 rules, such as "the breaker must be split before the two side knife switches are split (after the interval knife switch is in the'split' state, the breaker can be'split')" "the emergency fault is determined when the light board flicker frequency is greater than 8Hz, and the response must be within 5 minutes" "the associated switch element should trigger an alarm when the remote measurement current exceeds the rated value by 120% (110kV outgoing line rated current 1200A, exceeding 1440A)".

[0067] The graph model association rules are based on the standard monitoring information point table, with a total of 28 rules, such as "the graph element ID corresponds one-to-one with the remote signaling point number (such as 'CB-001' corresponding to point number '1101')" "the interval where the graph element is located must be consistent with the actual interval of the device (such as the #1 main transformer interval graph element cannot appear in the #2 main transformer interval)" "the light board description must completely match the protection signal table ('#1 main transformer overcurrent protection action' cannot be simplified as '#1 main transformer overcurrent')".

[0068] The abnormality determination rules have a total of 25 rules, with threshold values such as "the remote signaling value and the graph element state are inconsistent for more than 3s, which is determined as a state mismatch anomaly" "the graph element is missing the corresponding remote signaling point (the standard point table has a record but no graph element is identified, which is determined as an association missing anomaly)" "the light board does not flicker according to the fault level (emergency fault 5-10Hz, general fault 3-5Hz), which is determined as a display anomaly". All rules are stored in the MySQL database in the format "rule ID-triggering condition-diagnostic conclusion-treatment suggestion", and the rule ID is named according to "type-sequence number" (such as expert experience rule-E01, graph model association rule-T01).

[0069] Step four: intelligent diagnostic reasoning of rule fusion

[0070] With the forward chain reasoning mechanism, the reasoning engine is developed based on the PyKE framework of Python. First, the graph element recognition result (graph element ID, state, confidence) is matched with the rule base, such as recognizing "#1 main transformer 110kV side circuit breaker is open but the two side switches are not open", triggering the expert experience rule E01 ("circuit breaker needs to be open before the two side switches are open"), and the diagnostic conclusion is "operation logic abnormality". When multiple rules are triggered (such as E01 and T02 ("the interval where the graph element is located is consistent with the actual interval of the device") are triggered at the same time), a weighted fusion strategy is adopted, and the diagnostic rule confidence optimization formula is called: . Wherein w1=0.45 (expert experience rule weight), w2=0.35 (graph model association rule weight), w3=0.2 (abnormal judgment rule weight); the historical diagnostic accuracy rate of E01 is 96%, so C1=0.96; the matching degree of T02 is 99%, so C2=0.99; no abnormal judgment rule is triggered, so C3=0. Calculate C=0.45x0.96+0.35x0.99+0.2x0=0.8865, determine as medium reliable diagnosis, combined with operation and maintenance experience, prefer to adopt E01 conclusion, and mark the matching result of T02.

[0071] Generate preliminary diagnostic results, including abnormal type (operation logic abnormality), involved device (#1 main transformer 110kV side circuit breaker and two side switches), triggered rule ID (E01, T02), abnormal occurrence time (accurate to seconds), confidence (0.8865).

[0072] Step five: consistency check of diagnostic results and remote signaling / telemetry

[0073] Remote signaling consistency check: compare IEC104 remote signaling real-time value and graph element recognition state every 5 seconds, such as remote signaling value is 1 (close) and graph element shows "open" (red), confirm "state mismatch abnormality", record remote signaling frame number (such as "20240520103005-Frame123") and graph element frame number (such as "main wiring diagram-120240520103005-Frame45").

[0074] Telemetry correlation check: if the diagnostic result is "overcurrent fault", check the corresponding telemetry current value, such as #1 outgoing interval telemetry current 1500A (exceeding rated 1440A), then strengthen the "overcurrent fault" conclusion; if the current is 1300A (not exceeding rated), mark "suspected misjudgment" and return to step four to re-match rules (such as triggering abnormal judgment rule A03 "telemetry value does not exceed threshold but light board alarm, determine as light board mis-triggering").

[0075] Figure module library integrity check: traverse the standard point table 238 remote signaling / telemetry points, check if there is an unidentified figure, such as "35kV capacitor ground knife" in the point table (point number "3508") but no corresponding figure is identified, mark "figure missing exception", generate missing figure device description, interval position and associated point number. The diagnostic results of the check are marked "valid", and those that do not pass are marked "to be reviewed". The review process is confirmed by two operation and maintenance personnel, and the diagnostic results are updated after confirmation.

[0076] Step six: diagnostic result output and abnormal alarm

[0077] Real-time display uses Qt framework to develop interface, ROI area is marked with red font figure ID (such as "CB-001"), abnormal type (such as "operation logic exception"), trigger rule (such as "E01"), confidence (such as "0.89"), effective result is marked with green, and to be reviewed is marked with yellow; The report is generated daily, including abnormal statistics (such as 3 times of operation logic exception and 2 times of state mismatch exception), high frequency abnormal device list (such as #1 main transformer interval exception accounts for 40%), rectification suggestion (such as "check #1 main transformer 110kV side knife switch operation logic configuration"), support PDF / Excel format export (through Python reportlab and openpyxl library).

[0078] Abnormal alarm grading push: valid exception through industrial siren (decibel 85dB) + interface red flashing (frequency 2Hz) to realize sound and light alarm, at the same time through Aliyun SMS API push to operation and maintenance personnel mobile phone (including exception type, device location, processing suggestion), alarm response time ≤1s; To be reviewed exception only interface yellow flashing, not trigger sound and light and SMS alarm. Log storage in MySQL database, including diagnostic record (diagnostic ID, figure ID, exception type), trigger rule (rule ID, content), check result (check type, whether passed), operator (operation and maintenance personnel account), processing result (has been rectified / to be rectified), storage period ≥6 months, support according to figure ID, exception type, time range retrieval (through Python SQLAlchemy framework).

[0079] Abnormal risk level evaluation and rule conflict resolution

[0080] Call abnormal risk level evaluation formula: As " #1 main transformer 110kV side circuit breaker operation logic exception", similar abnormalities occurred 12 times (P=10) in the past year (T=8), with a duration of 5 minutes (T=8), affecting #1 main transformer (core equipment, S=10); a=0.35, b=0.3, c=0.35, R=0.35x10+0.3x8+0.35x10=9.4, determined as high risk, triggered emergency alarm (5 minutes without processing automatically upgraded to city-level operation and maintenance center).

[0081] When there is a rule conflict (e.g., rule A determines "state mismatch" and rule B determines "graph element recognition error"), the rule conflict resolution formula is called: Rule A (E05) has been used 600 times in the past 3 months (U=100), with an accuracy rate of 98% (F=100), and an expert rating of "excellent" (H=100); Rule B (A02) has been used 300 times (U=60), with an accuracy rate of 92% (F=92), and an expert rating of "good" (H=80); k1=0.3, k2=0.4, k3=0.3, WA=0.3x100+0.4x100+0.3x100=100, WB=0.3x60+0.4x92+0.3x80=80.8, and the conclusion of rule A is adopted.

[0082] Table 1: Comparison of diagnostic effects before and after implementation of 110kV conventional substation

[0083] Evaluation index Before implementation (manual) After implementation (the method) Change trend Single review time 4.5 hours 0.8 hours Greatly shortened Primitive identification accuracy 82.3% 98.7% Significantly improved Associated error identification rate 65.2% 99.1% Significantly improved State misjudgment rate 9.8% 0.4% Greatly reduced High-risk abnormal response time 15 minutes 1 minute Greatly shortened

[0084] Table 1 shows that after implementing this method, the monitoring picture checking efficiency and accuracy of 110kV conventional substation are significantly improved. The single checking time is shortened from 4.5 hours to 0.8 hours, the graph element recognition and rule matching are completed automatically instead of manually, and the time-consuming of manual checking is reduced; the graph element recognition accuracy is improved from 82.3% to 98.7%, benefiting from the multi-modal feature fusion and attention mechanism, effectively solving the recognition deviation caused by light fluctuation and equipment shielding; the associated error recognition rate and state misjudgment rate are significantly reduced, reflecting the value of graph module library association verification and multi-rule fusion; the response time of high-risk abnormalities is shortened from 15 minutes to 1 minute, thanks to the risk level evaluation and emergency alarm mechanism, ensuring quick disposal of high-risk problems and avoiding delay caused by manual troubleshooting. Overall, this method completely solves the problems of low efficiency and easy to miss in traditional manual checking, and adapts to the operation and maintenance needs of 110kV conventional substation.

[0085] Example Two: Application of 220kV intelligent substation power equipment graph element intelligent diagnosis

[0086] This embodiment is aimed at 220 kV intelligent substation (including 2 main transformers, 4 outgoing line intervals, and 3 groups of SVG devices), focusing on solving the identification of new intelligent devices (such as ABBSF6 intelligent switch and electronic current transformer icon), cross-interval complex abnormal diagnosis, and user-defined rule adaptation problems.

[0087] Step one: Multi-source data collection and preprocessing

[0088] Upgrade configuration of collection equipment: 2 Huawei HUAWEI-IPC-6325E high-definition cameras are deployed in the main wiring diagram area, with a resolution of 2560x1440, a frame rate of 30 fps, a low illumination of 0.001 Lux, a support of 140 dB wide dynamic, an installation height of 8 m, a horizontal field of view angle of 120°, covering 220 kV, 110 kV, and 35 kV voltage levels; 1 Yuxi UNV-IPC3022ER3-DP camera is deployed in each interval of the interval diagram area, supporting AI intelligent coding (code rate 4-8 Mbps adaptive), focusing on intelligent switch (such as #1 main transformer 220 kV side intelligent switch) and circuit breaker icon; a Hikvision DS-2CD6A26FWD-IZS macro lens camera is deployed in the light board area, with a resolution of 3840x2160, supporting 2.7-13.5 mm zoom, and recognizing intelligent light boards (with LED display and touch control). All cameras are connected to the system through ONVIF protocol, supporting real-time monitoring of device status (online / offline, code rate, frame rate).

[0089] Remote signaling / telemetry data is received through IEC104 protocol (port 2404) and MMS protocol (IEC61850) dual protocols, with a remote signaling sampling frequency of 2 times / second and a remote telemetry sampling frequency of 10 times / second, and IEEE1588 PTP protocol (accuracy ±1 ms) is used for synchronization. In addition to the standard point table of D5000 system, the model data of intelligent devices (such as electronic nameplate information and state feedback interface parameters of intelligent switch) are added, which are classified by "intelligent device type-interval-icon ID-electronic nameplate code" (such as "intelligent switch-#1 main transformer 220 kV interval-DS-001-202405AB").

[0090] A new motion blur correction is added in the preprocessing stage: the DeblurGAN-v2 model is used, the training dataset is the blurred video segments (1000 segments, 5s each) caused by the operation of the main transformer cooling fan in the substation, the blind deconvolution algorithm (initial kernel size 15x15) is used for blur kernel estimation, the generator uses the U-Net structure (4 layers of downsampling, 4 layers of upsampling), the discriminator uses PatchGAN, the training epoch=100, batchsize=8, the optimizer is Adam (learning rate 0.0002), and the correction is qualified when PSNR≥30dB. The clarity of the intelligent knife gap graph element profile is improved by 60% after correction. Other preprocessing parameters: σ=0.9 during the day, σ=1.2 at night, ROI extraction combined with the device coordinates of the substation BIM model (Autodesk Revit modeling), the intelligent knife gap ROI is set to (x=800, y=600, w=100, h=80), and the coordinates are automatically framed.

[0091] Step two: dynamic state recognition of power equipment graph elements

[0092] For the new intelligent knife gap (sample only 20 frames), small sample compensation recognition is used: 1000 frames of expanded samples are generated through data enhancement techniques (rotation ±15°, scaling 0.8-1.2 times, brightness ±20%, adding simulated industrial noise (signal-to-noise ratio 30dB)); extract the core features of the expanded samples (the electronic nameplate position of the intelligent knife gap (upper right corner, accounting for 1 / 5 of the graph element area), the opening angle (0°-90°, accuracy ±1°), the state indicator light color (split=red, closed=green, fault=yellow)), establish a feature mapping table, and associate it to the "split / closed" determination rule of the traditional knife gap (adjust the color feature threshold, yellow H=60°±5° is judged as fault).

[0093] Model fine-tuning is based on the pre-trained YOLOv8-Tiny, only the output layer (the number of categories is increased from 4 to 5, and "intelligent knife gap" is added) and the last 2 layers of convolutional layers (Conv2d-10, Conv2d-11) are updated, the first 10 layers of parameters are frozen, the training batchsize=8, epoch=30, learning rate 0.0005, and the cosine annealing learning rate scheduling is used. Incremental training fuses the weights of the fine-tuned model and the original model (original model weight 0.7, new model weight 0.3), retains the traditional graph element recognition ability, and the intelligent knife gap recognition confidence is ≥92%.

[0094] In low-illumination environments (such as dark corners of substations), the Retinex algorithm is used for enhancement (multi-scale, σ = 20, 40, 80), the brightness is increased by 3-5 times, the HSV color space is compensated (night red H value offset 5-10°, adjust the recognition threshold to H = 0-20°), increase the dynamic feature weight (from 0.25 to 0.4), and compensate the recognition after comparison with adjacent normal illumination frames (5 consecutive frames), the result is consistent, the state is confirmed, and the pixel recognition accuracy under low illumination is maintained at 96% or above.

[0095] Step three: multi-dimensional diagnostic rule base construction and dynamic update

[0096] The rule base adds new rules for intelligent devices based on Example One: expert experience rules E33 (“intelligent knife switch opening angle > 85° is determined as ‘closed’, < 5° is determined as ‘split’, 5°-85° is determined as ‘half open abnormal’”), E34 (“when the intelligent light board LED display fault code is ‘E01’, the associated electronic current transformer telemetry value needs to be checked synchronously”); graph association rule T30 (“intelligent device graph ID needs to be bound to electronic nameplate code (such as ‘DS-001’ bound to ‘202405AB’)”); abnormal judgment rule A28 (“if the color of the intelligent knife switch status indicator light is inconsistent with the remote signaling value for more than 2s, it is determined as ‘state feedback abnormal’”).

[0097] Dynamic rule update is performed every quarter: collect review feedback from operation and maintenance personnel (such as “rule E33 misjudges half open abnormal”), filter rules with misjudgment rate > 5% (such as E33 misjudgment rate 7%), adjust the opening angle threshold (8°-82° is determined as half open abnormal); adjust the confidence benchmark according to the diagnostic accuracy rate in the past 6 months (E33 accuracy rate from 92% to 95%, C1 benchmark value from 0.92 to 0.95); the updated rule is verified in the test environment (simulate 1000 times of intelligent device abnormal scenarios), the diagnostic accuracy rate ≥ 97% can be online, and the rule trigger frequency (such as E33 triggers 25 times per month) and accuracy rate (95%) are recorded by Python's logging library after online, which provides the basis for the next update.

[0098] Step four: intelligent diagnostic reasoning and cross-interval verification of rule fusion

[0099] Positive chain reasoning adds cross-interval rule verification: build interval correlation table (based on substation primary wiring diagram), such as 220kV I section bus correlation #1 main transformer 220kV interval, #1 outgoing line interval, #2 outgoing line interval; when #1 outgoing line interval switch abnormally opens (graph element shows "open", remote signal value = 0), automatically retrieve associated interval data (220kV I section bus voltage telemetry value 90kV < 80% (176kV) of rated 220kV, #2 outgoing line interval switch state "open"), call cross-interval special rule E35 ("when bus voltage < 80% rated value, associated interval switch should be interlocked to open, if not, it is determined to be a linkage abnormality"), combined with multi-interval data verification: #1, #2 outgoing line interval switches have opened, consistent with rule E35, strengthening the "bus fault causes switch interlocking to open" conclusion; if #2 outgoing line interval switch has not opened, mark "interval correlation abnormality" and expand the search range to adjacent 220kV II section bus interval.

[0100] Rule conflict resolution formula , such as rule E35 (usage frequency 400 times, U = 80; accuracy rate 97%, F = 97; expert rating "excellent", H = 100) and rule T30 (usage frequency 200 times, U = 40; accuracy rate 99%, F = 99; expert rating "good", H = 80) conflict, calculate WE35 = 0.3x80 + 0.4x97 + 0.3x100 = 92.8, WT30 = 0.3x40 + 0.4x99 + 0.3x80 = 77.6, prefer to adopt E35 conclusion, and record T30 matching results for operation and maintenance reference.

[0101] Step five: diagnosis result and remote signal / telemetry consistency verification

[0102] Newly added intelligent device telemetry correlation verification: such as diagnosis "intelligent switch status feedback abnormality", in addition to comparing remote signal value and graph element state, it also needs to verify the electronic nameplate feedback data of intelligent switch (such as opening angle, motor current), if motor current > rated value 110% (intelligent switch motor rated current 5A, exceeding 5.5A), then strengthen the "status feedback abnormality" conclusion, and determine it as "motor jam causes status feedback deviation"; if motor current is normal, mark "sensor fault suspected", and suggest operation and maintenance personnel to check status sensor.

[0103] Newly added intelligent device graph element missing determination: traverse 128 intelligent device remote signal points in standard point table, such as "#2 main transformer 35kV side intelligent ground switch" (point number "3512") does not identify corresponding graph element, mark "intelligent graph element missing abnormality", generate electronic nameplate code of missing graph element (such as "202406CD"), installation location (#2 main transformer 35kV interval C phase) and associated interface parameters (RS485 communication address 0x0A), to facilitate operation and maintenance personnel to quickly locate installation.

[0104] Step six: Diagnosis result output and user-defined rule

[0105] Result output adds diagnosis result traceability function: query data traceability (smart knife gap graph video frame camera number "interval-1", timestamp "20240520143000", preprocessing parameter σ=1.0), rule traceability (trigger rule E35, T30, confidence calculation process), inference traceability ("first match E35→call T30→calculate weight→adopt E35"), generate traceability report (including data flow chart, rule trigger timing chart), and visualize through Python's matplotlib library.

[0106] User-defined rule interface supports personalized configuration: provides a web-based visual editing interface (based on the Vue framework), users can add trigger conditions (such as "#1 main transformer 220kV side smart knife gap opens, remote motor current > 3A for > 5s"), diagnosis conclusions ("smart knife gap opens after motor does not stop turning abnormally"), and processing suggestions ("emergency shutdown of motor, check control loop"), and supports logical operators (and / or / not) to combine conditions (such as "current > 3A and open state and time > 5s"). When testing rules, import historical data (such as smart knife gap operation data in April 2024), simulate trigger scenarios, output test accuracy (such as 96%) and misjudgment cases (such as 1 misjudgment due to motor start-up current fluctuation), test through system administrator review after joining the rule library, assign weight 0.2, confidence benchmark 0.9, support query, modification, disable (update rule status through MySQL UPDATE statement).

[0107] Table 2: Comparison of diagnosis effects before and after the implementation of 220kV smart substation

[0108] Evaluation index Before implementation (manual + basic system) After implementation (the method) Change trend Intelligent primitive identification accuracy 65.8% 95.3% Significantly improved Cross-interval abnormal diagnosis accuracy 58.2% 96.7% Significantly improved Rule update cycle 3 months (manual) 1 month (automatic + manual) Greatly shortened Custom rule adaptation efficiency 7 days (development customization) 2 hours (user configuration) Greatly improved Intelligent device abnormal rectification rate 72.3% 95.1% Significantly improved

[0109] The data in Table 2 shows that the method of the application effectively solves the problems of difficult identification of intelligent equipment, complex cross-bay diagnosis and slow rule adaptation in the application of 220kV intelligent substation. The intelligent graph element recognition accuracy is improved from 65.8% to 95.3%, which breaks through the limitation of insufficient samples of new equipment due to small sample learning and feature mapping strategy; the cross-bay abnormal diagnosis accuracy is improved from 58.2% to 96.7%, which avoids misjudgment caused by one-sided single bay data due to bay correlation table and cross-bay rule checking; the rule updating period is shortened from 3 months to 1 month, and the dynamic updating mechanism reduces the cost of manual sorting; the self-defined rule adaptation efficiency is improved from 7 days to 2 hours, and the visual interface meets the individual needs of the substation; the intelligent equipment abnormal rectification rate is improved from 72.3% to 95.1%, and the traceability function and rectification suggestion help the operation and maintenance personnel to quickly locate the problem. Overall, the method adapts to the technical characteristics of 220kV intelligent substation, and provides a reliable solution for intelligent power equipment monitoring and checking.

[0110] Figure 2 The figure shows that before optimization, the confidence of low-illumination, camera shaking and other scenes is less than 85%, which is lower than the qualified threshold, and is easy to cause misjudgment; after optimization, the confidence of all scenes is improved to more than 95%, among which the normal light scene reaches 99.2%, mainly due to the multi-modal feature fusion and targeted correction technology - such as multi-frame feature fusion to supplement the missing information in the device occlusion scene, and the combination of OCR auxiliary verification (customer chat records mention the need for character recognition of the light card) in the light card scene, solving the problem of "single frame dependence and weak anti-interference" in traditional recognition. This optimization ensures that even in complex scenes, the graph element recognition result is still reliable, laying a foundation for subsequent correlation checking and abnormal diagnosis.

[0111] Figure 3 When manually checking, "graph element description mismatch" is often misjudged as "bay error" due to lack of system classification, and the rectification efficiency is low; the system not only can sort error types according to the proportion ("graph element description mismatch" has the highest proportion, which is the primary rectification direction), but also can give practical suggestions combined with technical standards in the file (such as D5000 point table format, GIS coordinates) - such as avoiding misplacement of graph elements by binding bay codes, and solving state mismatch by synchronizing IEC104 frame data. This accurate classification and targeted suggestions greatly reduce the difficulty of troubleshooting for operation and maintenance personnel, meet the needs of "improving checking efficiency and accuracy", and avoid the drawbacks of traditional manual "blind rectification".

[0112] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent substitutions or changes within the technical scope disclosed by the present application according to the technical solutions and inventive concepts of the present application, which should be covered within the protection scope of the present application.

Claims

1. An intelligent diagnosis method for power equipment graph element state recognition and rule fusion, characterized in that, Comprise: Step one, data collection and pretreatment: collect the dispatching master station monitoring picture associated data, including monitoring video data; Collect the main wiring diagram, interval diagram, and light board diagram video in the partition, obtain the device graph element marked with a yellow frame, and simultaneously backup collect the D5000 system output picture through the HDMI video collection card; add the timestamp, camera number, and monitoring area identification synchronously during collection; adopt the adaptive Gaussian filter to remove the noise during pretreatment, align the video frame through the frame synchronization technology, and extract the ROI area of the graph element; adopt the Kalman filter to eliminate the jitter of the telesignaling / telemetering data; and sort the graph model library data according to the device type-interval-graph element ID; Step two, dynamic state recognition of the power equipment graph element: recognize the graph element state based on the deep learning model and feature analysis: adopt the YOLOv8-Tiny model to train the data set containing the common graph element samples of the transformer substation; extract the shape feature, color feature, and dynamic feature of the graph element during model inference; Output the graph element recognition result, including the graph element ID, device name, current state, and recognition confidence; Step three, construction of the diagnosis rule library: construct three types of diagnosis rules, covering the graph element state, associated relationship, and abnormality judgment, including the expert experience rule, graph model associated rule, and abnormality judgment rule; store all the rules in the format of rule ID-triggering condition-diagnosis conclusion-treatment suggestion, support the rule query and modification; Step four, diagnosis reasoning of the rule fusion: adopt the forward chain reasoning mechanism, realize the abnormality diagnosis in combination with the rules; first, match the graph element state recognition result of step two with the rule library of step three, trigger the rules meeting the conditions; then, adopt the weighted fusion strategy for the triggering scene, and distribute the weight according to the rule priority and historical accuracy rate; finally, generate the preliminary diagnosis result, including the abnormality type, involved device, triggering rule ID, and abnormality occurrence time; Step five, consistency check of the diagnosis result and the telesignaling / telemetering: cross check the preliminary diagnosis result with the data, including the telesignaling consistency check, telemetering associatedness check, and graph model library integrity check; mark the diagnosis result as valid after the check, and mark the result as to be reviewed and return to step four for re-reasoning after the check fails; Step six, output of the diagnosis result and abnormality alarm: present the diagnosis result and push the abnormality alarm according to the user demand, superimpose the display of the graph element ROI area on the system interface, mark the abnormality type, triggering rule, and confidence; mark the effective result with green color, and mark the result to be reviewed with yellow color; generate the diagnosis report according to the day / week / month, including the abnormality statistics, high-frequency abnormality device list, and rectification suggestion. 2.The intelligent diagnosis method of power equipment graph element state recognition and rule fusion according to claim 1, characterized in that, Further comprise: Diagnosis rule confidence optimization step: fuse the confidence of different rules through weighted calculation, and the formula is: In the formula, C is the diagnostic confidence after rule fusion, w1 is the weight of the expert experience rule, w2 is the weight of the graph model association rule, and w3 is the weight of the abnormality determination rule; C1 is the triggered expert experience rule confidence, C2 is the triggered graph model association rule confidence, and C3 is the triggered abnormality determination rule confidence. 3.The intelligent diagnosis method of power equipment graph element state recognition and rule fusion according to claim 1, characterized in that, Further comprise: Abnormality risk level evaluation step: the risk level is calculated through various factors, and the formula is: where R is the abnormal risk value, a is the abnormal probability weight, b is the abnormal duration weight, and c is the abnormal impact range weight; P is the abnormal probability score, T is the abnormal duration score, and S is the abnormal impact range score.

4. The intelligent diagnosis method for power equipment graph element state recognition and rule fusion according to claim 1, characterized in that, The third step further comprises a dynamic rule updating sub-step for adapting and optimizing the rule library according to on-site operation feedback and new scenarios; the updating process is as follows: the review feedback of the operation personnel on the diagnostic results is collected every quarter, and the rules that need to be optimized are screened out; for the newly added equipment graph elements, the state characteristics thereof are collected, and corresponding expert experience rules and abnormality judgment rules are added; the original rules are adjusted according to the diagnostic accuracy rate in the past six months to adjust the confidence benchmark; the updated rules need to be verified through a test environment, and the rule triggering frequency and accuracy rate are recorded after going online, thereby providing a basis for the next update.

5. The intelligent diagnosis method for power equipment graph element state recognition and rule fusion according to claim 1, characterized in that, The second step further comprises a small sample graph element state compensation identification sub-step for dealing with the problem of insufficient new or rare graph element samples in the substation; the compensation identification process is as follows: for the samples of the newly added graph elements, 1000 expanded samples are generated by using a data enhancement technology; the core features of the expanded samples are extracted to establish a feature mapping table, which is associated with the rules of the equipment in the existing rule library; the graph element identification model is fine-tuned by using transfer learning, and only the output layer and the last two convolutional layer parameters are updated; the compensation identification result needs to be compared with the manual annotation, and actual operation samples are continuously collected after the application, so as to gradually optimize the feature mapping and rule association.

6. The intelligent diagnosis method for power equipment graph element state recognition and rule fusion according to claim 1, characterized in that, The fourth step further comprises a cross-interval rule checking sub-step for diagnosing abnormalities involving multiple interval correlations; the checking process is as follows: an interval correlation relationship table is constructed; when the graph elements of the interval trigger abnormal diagnosis, the graph element states and telemetering data of the correlated intervals are automatically searched; the cross-interval special rules are called to verify the diagnostic conclusion in combination with the interval data.

7. The intelligent diagnosis method for power equipment graph element state recognition and rule fusion according to claim 1, characterized in that, Further comprising: A diagnostic result tracing step for tracing the data sources and rule triggering paths in the diagnostic process; the tracing content includes data tracing, rule tracing and reasoning tracing; The tracing information is stored in the format of diagnostic ID-tracing internal format and is bound to the diagnostic result, supports one-key query of the complete tracing chain through the diagnostic ID, and generates a tracing report. 8.The intelligent diagnosis method of power equipment graph element state recognition and rule fusion according to claim 1, characterized in that, The step four further comprises a rule conflict resolution sub-step for resolving the contradiction of conclusions between rules when triggered; the conflict resolution uses priority weighting calculation, with the formula being: , wherein W is the rule priority weight, k1 is the rule usage frequency weight, k2 is the rule historical accuracy rate weight, and k3 is the expert rating weight. U is the rule usage frequency score in the past three months, F is the rule historical accuracy rate score, and H is the expert rating score. 9.The intelligent diagnosis method of power equipment graph element state recognition and rule fusion according to claim 1, characterized in that, The second step further comprises a graph element state compensation identification sub-step in a low-illumination environment for solving the identification difficulty caused by poor monitoring video quality in the substation at night or in dark corners; the compensation identification process is as follows: the Retinex algorithm is used to enhance the low-illumination video frames; for the color distortion problem in low-illumination, an HSV color space compensation model is established; the dynamic feature weight is increased; After the compensation identification, the identification results of the adjacent normal illumination frames are compared.

10. The intelligent diagnosis method for power equipment graph element state recognition and rule fusion according to claim 1, characterized in that, Further comprising: A user-defined rule interface step for meeting the individualized diagnostic needs of different substations; The interface functions include rule editing, rule testing, rule going online, and rule management; Through the interface, the local equipment characteristics and operation habits of the substation are adapted.

Images