A method, device, equipment, and storage medium for multi-source condition analysis and early warning of high-voltage switchgear.
By using a multi-source status analysis and early warning method at the edge terminal, the problem of continuous misjudgment of abnormalities in high-voltage switchgear was solved. Through data deduplication and verification tasks, the accuracy and continuity of monitoring were ensured, and the reliability of the model was improved.
Patent Information
- Application Number
- CN202611602418.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-16
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies for multi-source condition monitoring of high-voltage switchgear suffer from problems such as persistent misjudgment of anomalies and inaccurate diagnostic outputs. In particular, it is difficult to effectively verify and update the model under energized operation conditions, leading to erroneous alarms or missed persistent anomalies.
By executing a multi-source state analysis and early warning method through edge terminals, monitoring data is acquired and a binding diagnostic configuration set is established for verification tasks. Only newly added observation data is included, isolated samples are removed, and data deduplication and verification are performed based on the operation impact range and the energized state to ensure the reliability and accuracy of model parameters.
It improves the reliability of verifying the continuity of abnormalities in high-voltage switchgear, reduces misjudgments and omissions, maintains the continuity and accuracy of monitoring, and avoids the impact of duplicate data and historical data contamination on judgment.
Smart Images

Figure CN123116103A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power equipment monitoring technology, and in particular to a method, device, equipment and storage medium for multi-source status analysis and early warning of high-voltage switchgear. Background Technology
[0002] When abnormalities occur in the conductive connections, insulation structures, and operating mechanisms within high-voltage switchgear, they can manifest as increased temperature, partial discharge, and vibration changes. In substations and power distribution sites, online monitoring typically acquires relevant signals through sensors, extracts features, compares thresholds, or classifies them to form status results, and combines multiple monitoring results to provide a basis for abnormal alarms and maintenance decisions.
[0003] Existing multi-source parameter monitoring methods typically collect monitoring data such as infrared thermal imaging, ultrasound, transient ground potential partial discharge, switching coil current, and vibration. After classifying different physical quantities, classifiers are used to form preliminary evidence sources. Then, weight correction, similarity calculation, and evidence fusion are performed to obtain the switchgear condition evaluation results. This type of method mainly addresses the consistency and conflict between classification results; however, mutual consistency does not directly prove that the anomaly persists over time.
[0004] In continuous monitoring scenarios, transient vibrations caused by circuit breaker opening and closing may overlap with fault symptoms. Furthermore, the sampling periods and transmission delays of different channels mean that the aggregated results may not correspond to the same operating period. When the monitoring chamber is no longer energized, the decline in partial discharge signals does not directly prove the elimination of insulation defects. In addition, for monitoring applications that update models using historical data, unresolved anomalies may alter the normal characteristics learned by the model, causing diagnostic outputs to appear normal while actual defects still exist. These situations may lead to false alarms, premature alarm cancellation, or missed persistent anomalies, affecting maintenance judgment.
[0005] Therefore, how to reduce the misjudgment of the continuous abnormality of high-voltage switchgear under the condition of changes in operating status and diagnostic benchmarks is a technical problem that needs to be solved. Summary of the Invention
[0006] This invention aims to address the shortcomings of existing technologies by providing a method, device, equipment, and storage medium for multi-source status analysis and early warning of high-voltage switchgear. It can improve the reliability of continuous anomaly verification and model updates based on real new observation data, operational impact range, energized operation stage, and isolated sample management.
[0007] As one aspect of the present invention, a multi-source status analysis and early warning method for high-voltage switchgear is provided, executed by an edge terminal, comprising: Acquire monitoring data, including physical sampling source, original sampling identifier, and observation range, as well as operational and energized status; For monitoring data that exceeds the effective limit, an alarm is output. A review task is established that is bound to the diagnostic configuration set. The samples in the backtracking interval before the anomaly and during the review period are marked as isolated samples. The diagnostic configuration set includes the threshold and model parameters and remains unchanged during the task. The operational impact range is determined based on the operational status. Deduplication is performed according to the original sampling identifier. Only valid new observation data whose observation range is after the operational impact range and belongs to the same energized operation phase are included in this task; others are excluded. Specifically, new observation data is included in the review task only if it simultaneously meets the following conditions: the new observation data has genuine new sampling proof, i.e., its original sampling identifier is different from the most recent original sampling identifier saved in this task, and the original sampling identifier is associated with the device identifier, physical measurement channel, sampling session, and device sampling sequence number updated with actual conversion; the observation range of the new observation data is after the operational impact range; the new observation data belongs to the same energized operation phase, i.e., the target chamber remains energized and the operation phase remains unchanged within its complete observation range; otherwise, it is not included in this task, and it is not included due to increasing message sequence number, increased gateway forwarding count, or consistent classification results. If the required number of complete and valid windows are obtained within the task deadline, the number of abnormal windows is counted according to the diagnostic configuration set and a review conclusion is output; otherwise, insufficient evidence is output by the deadline. When updating model parameters, remove training windows containing isolated samples from the target and all historical inputs.
[0008] Furthermore, the edge terminal subtracts the corresponding upper bound of the time error from the start time of the observation interval and adds the upper bound to the end time to form a conservative observation interval; the sum of the operation end time, the upper bound of the operation time error, and the stability margin after the operation is determined as the end boundary of the operation influence interval. Wherein, the conservative observation interval satisfies:
[0009] In the formula, The index for observation records is a positive integer and dimensionless. and These are the actual start and end times of the observation recorded, respectively, satisfying... ; This is the upper bound of the non-negative time error of this record; This represents a conservative observation range that the actual sampling might fall into; The operation's influence range satisfies:
[0010] In the formula, A positive integer index for the operation event; and These are the start and end times of the operation, respectively, satisfying... ; This is the upper bound for non-negative operation time error; and These are the non-negative pre-operation impact margin and the post-operation stability margin, respectively. The area affected by the operation; The newly added observation data will only be included in the verification task if the start time of the conservative observation interval is later than the boundary, and the valid records of the operating status and energized status cover the interval and indicate that the target chamber is continuously energized and the operating phase has not changed. Otherwise, the data will not be included and the alarm will be retained.
[0011] Furthermore, the review task is a cross-sampling source review task; the edge terminal establishes a set of sampling sources for each of the two candidate pieces of evidence based on the physical sampling source; If the sampling source sets are all known and non-empty, mutually disjoint, and the corresponding physical quantities satisfy the preset pairing relationship, and the sum of the absolute value of the time difference at the center of the corresponding window and the upper bound of their respective time errors is not greater than the preset time tolerance at the center of the window, then the candidate evidence will be used as paired evidence for continuous review according to the diagnostic configuration set; otherwise, the judgment results of the candidate evidence will be retained and will not be included in the cross-sampling source review. Wherein, the set of sampling sources satisfies:
[0012] In the formula, , These are the sets of physical sampling sources cited by candidate evidence A and B, respectively; To represent the intersection of sets, Represents the empty set; The window center time condition is satisfied as follows:
[0013] In the formula, candidate evidence A and B correspond to the two observation windows to be paired in this study; and These are the start and end times of observation at window A. and These are the start and end times of observation in window B; and This refers to the center time of the corresponding window; and This serves as the upper bound for the non-negative time error of each window; This is the pre-registered positive window center time tolerance.
[0014] Furthermore, the edge terminal establishes a training sample isolation index that records isolated samples and the reasons for isolation, and saves this index and the version reference of the diagnostic configuration set in the same task record; the relevant model update is postponed until the review task is completed, and the candidate model is trained with the training window that has passed the isolation check after the task is completed. Before release, the training sample index of the candidate model is reviewed. If newly labeled isolated samples are found, the release is canceled and retraining is performed. Otherwise, the candidate model is released if the check based on the fixed health validation set and the confirmed abnormal replay set passes. If the check fails, the diagnostic configuration set is maintained.
[0015] Furthermore, the edge terminal sends a sampling request associated with the verification task to the acquisition front end that supports request-triggered sampling, and saves the request acceptance receipt and measurement completion record respectively; it performs a validity check and a verification task inclusion condition check on the measurement completion record. If the check is passed, the corresponding new observation data is included in the task; otherwise, it is not included. The inclusion conditions include:
[0016] In the formula, For the first The start time of the newly added observation data. This is the upper bound of the time error of the observation. For the reviewed operation event At the end of the day, Let its time error be the upper bound. To provide a stable margin after operation; The receipt for acceptance of the request does not replace the measurement completion record. The observation end time corresponding to the measurement completion record shall not be later than the time when the record is received and used.
[0017] Furthermore, when establishing a review task, the edge terminal saves the most recent original sampling identifier of the physical sampling source and waits for the collection front-end to update the record naturally in a periodic manner, without sending a request to trigger a sampling command; If the original sampling identifier in the periodic natural update record is different from the saved original sampling identifier and meets the conditions for inclusion in the review task, it will be counted; otherwise, it will be rejected and the alarm will be retained. The original sampling identifier is associated with the device identifier, physical measurement channel, sampling session, and device sampling sequence number updated with actual conversion. New sampling is not identified by incrementing the message sequence number.
[0018] Furthermore, the edge terminal acquires the cumulative peak value from the acquisition front end; when the acquisition front end has a separate real-time monitoring window and a verification cumulative window, it saves the cumulative peak value before clearing and only sends a clearing command to the verification cumulative window; otherwise, it does not clear the value and the output capability is insufficient; after obtaining the clearing completion feedback, it verifies the new cumulative start and end time; if the verification task inclusion condition is met, it counts; otherwise, it does not count; if the clearing completion feedback is not obtained, the return value is saved as the historical cumulative value. Each counting window corresponds to a different zeroing completion feedback and observation interval, and overlapping accumulation windows are not counted repeatedly.
[0019] Furthermore, the sampling request includes the task identity of the verification task, the target channel, the planned start and end times, the window length, and the request number; the measurement completion record includes the device sampling number, the observation interval, and the characteristic value. The edge terminal associates and saves the sampling request, request acceptance receipt and measurement completion record based on the task identity and request sequence number, and associates the diagnostic configuration version identifier, original sampling identifier and review conclusion of the diagnostic configuration set with the same task record; the request sequence number and the device sampling sequence number are saved separately.
[0020] Furthermore, the edge terminal saves the basic alarm level, review conclusion, data availability and confirmation status respectively, and displays the trigger value, threshold and review conclusion on the local display; When a confirmation input is received from the alarm acknowledgment device, only the confirmation status is updated; otherwise, the confirmation status is maintained. The diagnostic configuration set also includes recovery rules. When the criteria for triggering an alarm meet the recovery rules, the corresponding current alarm status is lifted; otherwise, the status is retained. Unavailable data is not recorded as normal, and historical events are retained at each stage of occurrence, review, and recovery.
[0021] Furthermore, the edge terminal stores the event identity, record version, and review conclusion corresponding to the alarm in non-volatile memory; When uplink communication is interrupted, data acquisition, diagnosis, verification, and local display continue, and unsent records are marked as pending; when uplink communication is normal, pending records are sent to the remote platform. When a receipt confirmation corresponding to the event identity and record version is received, the corresponding record is marked as sent; otherwise, it is kept pending. Retransmission does not change the event identity and record version. The review conclusion is updated by appending the higher record version, and the original record is retained.
[0022] As another aspect of the present invention, a multi-source status analysis and early warning device for high-voltage switchgear is also provided, which is installed at an edge terminal and includes at least: a model and training management module, a data access module, a branch diagnosis module, an operation phase management module, a review and scheduling module, and a continuous review module; wherein: The model and training management module is used to store the diagnostic configuration set containing thresholds and model parameters, and to provide the diagnostic configuration set to the branch diagnostic module and the review scheduling module; The data access module is used to send monitoring data containing physical sampling sources, original sampling identifiers, and observation intervals to the quality and source management module for validity verification and deduplication based on the original sampling identifiers; The branch diagnostic module is used to diagnose the monitoring data verified by the quality and source management module according to the diagnostic configuration set, and send the valid limit violation results to the alarm and storage module and the review and scheduling module respectively. The alarm and storage module directly outputs alarms based on the valid limit violation results. The operation phase management module is used to determine the operation impact range and operation phase based on the received operation status and energized status, and provides this information to the review and scheduling module. The review scheduling module is used to establish a review task bound to the diagnostic configuration set based on the valid over-limit results, and send the status information of the task to the model and training management module. The model and training management module keeps the bound diagnostic configuration set unchanged during the task, and marks the samples in the backtracking interval before the anomaly and during the review period as isolated samples. The continuous verification module receives verification tasks, operational impact intervals, and operational phases from the verification scheduling module, and receives new observation data from the quality and source management module. Only valid new observation data whose observation interval is after the operational impact interval and belongs to the same energized operational phase corresponding to the task are included in the task; others are excluded. Specifically, new observation data is included in the task only if it simultaneously meets the following conditions: it has genuine new sampling proof, meaning its original sampling identifier is different from the most recent original sampling identifier saved by the task; the observation interval is after the operational impact interval; and it belongs to the same energized operational phase, meaning the target chamber remains energized and the operational phase remains unchanged within its complete observation interval; otherwise, it is excluded. The continuous review module is used to count the number of abnormal windows according to the diagnostic configuration set and send the review conclusion to the alarm and storage module when a specified number of complete and valid windows are obtained within the task period. Otherwise, it sends the result of insufficient evidence when the deadline is reached. The alarm and storage module saves and outputs the received result. When updating model parameters, the model and training management module removes training windows containing isolated samples from the target and all historical inputs. These isolated samples are recorded by the training sample isolation index, which is stored in the same task record as the version reference of the diagnostic configuration set. When any historical input sample hits the isolation index, the entire candidate training window is not allowed to enter.
[0023] As another aspect of the present invention, an electronic device is provided, characterized in that it includes a processor and a memory; the memory stores a computer program, and when the computer program is executed by the processor, the electronic device, as an edge terminal, enables the high-voltage switchgear multi-source status analysis and early warning method as described above.
[0024] As another aspect of the present invention, a computer-readable storage medium is provided, characterized in that the computer-readable storage medium stores a computer program; when the computer program is executed by the processor of an edge terminal, it implements the aforementioned multi-source status analysis and early warning method for high-voltage switchgear.
[0025] Implementing the embodiments of the present invention has the following beneficial effects: This invention provides a method, device, equipment, and storage medium for multi-source status analysis and early warning of high-voltage switchgear. By executing real-time alarms and evidence verification in parallel, it enables monitoring data that exceeds the effective limit to directly trigger alarms, and forms a verification conclusion only based on a specified number of complete and valid windows. If the deadline is insufficient, it outputs insufficient evidence, avoiding waiting for verification to block alarms and misclassifying missing windows as normal. In this invention, the original sampling identifier is used to deduplicate and the sampling source set and preset pairing relationship are verified so that message retransmission and derived results from the same source do not increase evidence of non-duplicate sampling sources, thus avoiding duplicate data from increasing the number of anomalies and the number of sources. In this invention, the conservative observation interval is compared with the operation influence interval and combined with the operation state and energized state records covering the entire conservative observation interval to verify the energized operation stage of newly added observation data, thereby reducing the risk that operation transients are counted as continuous abnormalities and signal fallback when not energized is judged as defect recovery. In this invention, by binding the diagnostic configuration set to the review task and isolating the pre-abnormal backtesting interval and the review period samples, while keeping the criteria unchanged, the training window involving the target and all historical inputs involving isolated samples is excluded, thus synergistically limiting the impact of criterion changes and historical input contamination on the abnormality judgment. In this invention, when uplink communication is interrupted, the edge terminal continues to collect, verify and display data locally, and retransmits the data according to the event identity and record version after communication is restored. This ensures that the same event is not counted repeatedly due to retransmission, retains the old and new verification conclusions, and maintains the continuity of on-site monitoring and anomaly recording. Attached Figure Description
[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, obtaining other drawings based on these drawings without creative effort still falls within the scope of the present invention. Figure 1 A schematic diagram illustrating the implementation environment of the multi-source status analysis and early warning method for high-voltage switchgear provided by this invention; Figure 2 This is a schematic diagram of the main flow of an embodiment of a multi-source status analysis and early warning method for high-voltage switchgear provided by the present invention; Figure 3 This is a flowchart illustrating the parallel processing of the real-time alarm path and the evidence review path in the method provided by the present invention. Figure 4 This is a flowchart of the branch processing of temperature, vibration, partial discharge related signals and trend diagnosis involved in the method provided by the present invention. Figure 5 The flowchart of the review task criterion freezing and training sample isolation, admission and model update involved in the method provided by the present invention; Figure 6 This is a time-series diagram showing the operational impact range and the newly added observation data after the operation involved in the method provided by this invention. Figure 7 This is a flowchart illustrating the process of verifying the validity of newly added observation data and determining admission criteria in the method provided by this invention. Figure 8 This is a schematic diagram illustrating the alarm, verification, and abnormal degradation state transitions involved in the method provided by the present invention. Figure 9 This is a schematic diagram illustrating the correlation of physical sampling sources and the inspection of non-repeating sampling sources of monitoring data involved in the method provided by the present invention; Figure 10 A schematic diagram of the functional module structure of an embodiment of the analysis and early warning device provided by the present invention; Figure 11 A hardware topology diagram of an embodiment of an electronic device provided by the present invention; Detailed Implementation 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.
[0027] I. Brief Explanation of Technical Terminology Physical sampling source refers to the measurement entry point jointly determined by the sensor identifier and the physical measurement channel, not the communication address or algorithm name. The sampling source set refers to all physical sampling sources referenced in forming a certain feature or diagnostic result. For example, the absolute value of the temperature of the same phase and the absolute prediction error contain the same source, and the temperature difference between two phases simultaneously contains sources from both phases.
[0028] The original sampling identifier is a record identifier formed by associating the device identifier, physical measurement channel, sampling session, and device sampling sequence number or other verified sampling completion information; its function is to identify the original measurement, and not to determine whether to resample based on whether the measurement value has changed.
[0029] The observation interval refers to the actual sampling start and end time corresponding to the feature, not the time from message transmission to reception. Numerical validity indicates that message parsing, range, self-checking, and sample integrity meet the requirements; temporal correlation validity indicates that the observation interval and its upper bound of time error are sufficient to support comparisons at the specified granularity.
[0030] The operation impact range refers to the conservative range formed after considering the start and end times of the operation and the margin of influence before and after. It only indicates that the observation may be affected by the operation, and does not pre-determine that all anomalies in it are interferences.
[0031] The operation phase refers to the continuous phase in which the energized state of the target chamber and the relevant open or closed positions remain consistent, and no new operations occur.
[0032] The diagnostic configuration set is a read-only configuration object consisting of basic thresholds, recovery rules, model parameters, absolute prediction error thresholds, feature order, calibration mapping, and time configuration. Different configuration objects are distinguished by diagnostic configuration version identifiers.
[0033] Criterion freeze means that the task always references the diagnostic configuration set bound at the time of creation, but does not mean that sensor measurements are frozen or immediate alarms are paused.
[0034] The training sample isolation index is a logical index that records the identifiers of data that are temporarily not allowed to participate in training and the reasons for isolation. It is not a separate physical storage area, nor does it delete the original data.
[0035] II. Method Examples (a) Hardware environment for implementing this method Reference Figure 1 This diagram illustrates the implementation environment of a multi-source status analysis and early warning method for high-voltage switchgear provided by the present invention. It shows a multi-source status analysis and early warning system for high-voltage switchgear, specifically including a temperature sensor 11, a transient ground voltage sensor 12, an acoustic emission sensor 13, and a triaxial vibration sensor 14 installed at corresponding measuring points on the high-voltage switchgear 10. These sensors are connected to an edge terminal 100 via a data acquisition front-end 20. An operation status interface 21 and a live status interface 22 provide isolated status information to the edge terminal 100. The edge terminal 100 is connected to a local display 200, an alarm acknowledgment device 210, and a remote platform 300, respectively completing sampling requests, on-site display, confirmation input, and evidence uploading. The remote platform 300 is not a necessary computational node for local alarms and verification.
[0036] Combination Figure 1 In this embodiment, the sampling request only changes the data acquisition process of the sensor or acquisition front-end 20, and does not drive the high-voltage switchgear 10 to perform opening and closing, nor does it replace the relay protection and the original interlocking. The following values are used to illustrate a set of configurable processing conditions and calculation relationships, and do not represent the general safety limits for all cabinet types, nor are they considered as the test performance that has been achieved; when the corresponding sampling, calibration and state conditions are not met, the edge terminal 100 executes the corresponding degraded branch.
[0037] Temperature sensors 11 correspond to three similar cable connectors or conductive connection points. Edge terminal 100 stores the mapping between cabinet number, bay, chamber, installation location, and phase; only measurement points with comparable installation locations can form a phase-to-phase temperature difference group. Temperature sensors 11 can employ a contact-type temperature measurement structure with digital output or a fiber optic temperature measurement structure with matching demodulation; changes in the communication link do not alter the identity of the measurement point, avoiding the interpretation of temperature differences between different components as phase-to-phase anomalies.
[0038] The transient ground voltage sensor 12 acquires the voltage-related signal of the cabinet measuring point through an adapted coupling structure. The acoustic emission sensor 13 adopts a coupling surface or acoustic orientation adapted to its detection method. The triaxial vibration sensor 14 is fixed to a designated mechanical measuring point of the mechanism or cabinet. The triaxial direction is pre-registered relative to the mounting surface. When the sensor is reinstalled or its fastening state is changed, the edge terminal 100 disables the relevant comparison configuration under the original installation conditions to prevent changes in installation transmission characteristics from being mistaken for changes in equipment defects.
[0039] Combined Figure 11 As shown, the acquisition front-end 20 can be composed of a signal conditioning circuit, an analog-to-digital conversion circuit, and a processing circuit for performing window calculations. High-frequency partial discharge related signals are detected and feature extracted at the acquisition front-end 20, and then the window peak value, observation range, and update identifier are output through the sensor communication interface 104. It is not required that the entire high-frequency waveform be transmitted via a low-speed link. For sensors with built-in digital processing capabilities, the acquisition front-end 20 can be integrated inside the sensor, but the physical measurement channel identifier and window definition still need to be retained.
[0040] The operation status interface 21 connects to the auxiliary contact for the open or closed position or to operation feedback with a reliable time stamp. The energized status interface 22 connects to the voltage presence signal bound to the monitoring chamber, which enters the edge terminal 100 via the isolated status input interface 105. The interface can be implemented by an isolated input circuit and an edge capture circuit. It is prohibited to directly connect ordinary low-voltage inputs to the primary high-voltage conductor. When the circuit breaker is open but the bus side is still energized, the bus side measuring point continues to be processed according to its own status, and the open signal cannot replace the conclusion that the chamber is not energized.
[0041] (II) Overall Method and Steps Reference Figure 2 As shown, the present invention provides a multi-source status analysis and early warning method for high-voltage switchgear, which consists of... Figure 1 The edge terminal 100 executes, combined with Figures 3 to 9 As shown, in this embodiment, the method includes the following steps: Step S1: Obtain monitoring data, operating status, and energized status, including physical sampling source, original sampling identifier, and observation range; Step S2: Output alarms for valid out-of-limit monitoring data, establish a review task bound to the diagnostic configuration set, and mark the samples in the pre-abnormal backtesting interval and during the review period as isolated samples. The diagnostic configuration set includes thresholds and model parameters and remains unchanged during the task. Step S3: Determine the operational impact range based on the operational status. Remove duplicates according to the original sampling identifier. Only valid new observation data whose observation range is after the operational impact range and belongs to the same energized operation phase are included in this task; the rest are excluded. Specifically, new observation data is included in the review task only if it simultaneously meets the following conditions: the new observation data has genuine new sampling proof, i.e., its original sampling identifier is different from the most recent original sampling identifier saved in this task, and the original sampling identifier is associated with the device identifier, physical measurement channel, sampling session, and device sampling sequence number updated with actual conversion; the observation range of the new observation data is after the operational impact range; the new observation data belongs to the same energized operation phase, i.e., the target chamber remains energized and the operation phase remains unchanged within its complete observation range; otherwise, it is not included in this task, and it is not included due to increasing message sequence number, increased gateway forwarding count, or consistent classification results. Step S4: When the specified number of complete and valid windows are obtained within the task deadline, count the number of abnormal windows according to the diagnostic configuration set and output the review conclusion; otherwise, output insufficient evidence by the deadline. Step S5: When updating the model parameters, remove training windows containing isolated samples from the target and all historical inputs.
[0042] Wherein, step S1 corresponds to Figure 3 Steps S101 to S103; step S2 corresponds to Figure 3 Steps S104 to S108 in the text, and corresponding to Figure 4 The branch diagnosis process is shown; step S3 corresponds to Figure 3 Steps S109 to S110 in the text, and corresponding to Figure 7 The review and approval process is shown; step S4 corresponds to Figure 3 Step S111 in the middle, and corresponding to Figure 7 The complete valid window count determination and insufficient evidence of expiration are shown; step S5 corresponds to... Figure 3 The model update and training admission steps are described, and corresponding Figure 5The diagram illustrates the training sample isolation and model deployment process.
[0043] The steps involved in the method of the present invention will be described in detail below.
[0044] (III) Step S1: Data Acquisition and Configuration Loading This step mainly corresponds to Figure 3 Steps S101 to S103 in the process involve the edge terminal completing interface configuration, message parsing, validity checks, and deduplication of the original sampling identifier, providing a unified data object for subsequent branch diagnosis and review tasks.
[0045] Combination Figure 3 In step S101, edge terminal 100 loads the measurement point mapping, protocol parsing, units, range, calibration version, and sampling capabilities. Capability registration records whether it supports request-triggered sampling, clearing the cumulative window, returning the device sampling sequence number, providing the observation timescale, or providing the maximum data age. For unregistered commands, edge terminal 100 prohibits their transmission; for devices with only periodic output, it waits for natural updates to avoid creating false verification receipts using operations not supported by the protocol.
[0046] Combination Figure 3 Step S101 further loads the trigger threshold, recovery threshold, data freshness, window integrity rate, operation impact margin, and number of reviews for each channel to form an initial diagnostic configuration set. For example, temperature is sampled once per second, the maximum allowable age for the current value is 3 seconds, and three valid windows are used for review. Fixed thresholds are saved based on the corresponding equipment technical requirements or approved tuning records and are not temporarily increased based on values after an anomaly occurs; configuration changes create new versions, while historical records remain associated with the original versions.
[0047] Combination Figure 3 The basic limit violation criterion refers to the threshold or allowable frequency band criterion that has been enabled in the temperature, partial discharge related signals, and vibration branches. The trend criterion refers to the absolute prediction error criterion between the measured value and the predicted value; the two form results respectively. "Data freshness" is verified according to the maximum possible time interval between the current judgment time and the actual sampling time corresponding to the data; the window feature must also verify its actual cumulative start and end time, and the message reception time cannot be used to replace the sampling time.
[0048] like Figure 3As shown, in step S102, the edge terminal 100 parses and collects messages from the front-end 20, first saving the original bytes, and then establishing a unified data object. This data object is a measurement record organized according to fixed fields, including equipment and physical measurement channel identification, measurement point ownership, original value, engineering value, unit, calibration version, observation start and end time, reception time, upper limit of time error, equipment sampling sequence number, message sequence number, and quality information. Unit conversion is performed according to the registered proportional coefficient and offset or calibration lookup table, and the original value and conversion relationship are retained; when the meaning of the register, calibration relationship, or decibel reference value cannot be identified, the physical meaning is not filled in based on the numerical appearance.
[0049] like Figure 3 As shown, the edge terminal 100 checks communication verification, field length, sensor self-test, range, saturation, and sample integrity. When a message is corrupted, its diagnostic input qualification is discarded and the reason is recorded; when a channel is saturated, the output range is abnormal, instead of treating the truncated waveform as a low-amplitude normal signal. An invalid channel does not block other channels, and exceeding the effective temperature limit still immediately triggers the alarm path, allowing sensor faults and equipment malfunctions to be expressed separately.
[0050] like Figure 3 As shown, the edge terminal 100 identifies duplicate records based on the device identifier, physical measurement channel, and original sampling identifier. An exemplary interface refinement method also stores the device startup batch; after a device restart, a new sampling session is established first, and then the recalculated sequence number is interpreted. The device sampling sequence number is only used as proof of update when the device protocol explicitly specifies that the sequence number is updated with actual conversion. When the gateway continuously sends three frames with the device sampling sequence number 2005, the three frames are associated with only one observation, preventing duplicate reports from being accumulated as continuous anomaly counts.
[0051] (iv) Step S2: Establishment of real-time alarm, branch diagnosis and review tasks This step mainly corresponds to Figure 3 Steps S104 to S108 in the process, and combined in a concentrated manner Figure 4 Explain the diagnostic methods for temperature, vibration, partial discharge-related signals and trend branches. Figure 5 The establishment of the verification task, the binding of the diagnostic configuration set, and the isolation of the training samples are also explained in this step.
[0052] 1. General Description of Branch Diagnosis Combination Figure 3 and Figure 4Edge terminal 100 performs temperature branch, vibration branch, partial discharge related branch, and trend branch diagnostics on the monitoring data verified by quality and source management module 402. Each branch is executed separately according to its own data conditions, and it is not a serial chain waiting for all branches to complete. Valid baseline exceeding the limit directly enters the alarm output; anomalies requiring continuous confirmation are established as separate verification tasks. An invalid channel does not block other channels, and an effective temperature exceeding the limit can still immediately enter the alarm path, allowing sensor faults and equipment anomalies to be expressed separately.
[0053] 2. Temperature branch Combination Figure 4 The temperature branch includes absolute temperature criteria, interphase temperature difference criteria, and historical temperature rise criteria. For example... Figure 3 As shown, step S103 first performs an absolute temperature determination, calculating for each effective sample:
[0054] In the formula, It is a phase identifier for one of phases A, B, and C. A, B, and C are electrical phase identifiers, not English technical terms. This is the integer index of the phase sample; This is the set of sample indices for which the phase value is valid and fresh within the current temperature determination window; For the corresponding temperature sample; This represents the maximum temperature within the window, where "maximum" is a descriptive subscript. All temperatures are in degrees Celsius (°C), and the value is limited to the effective range of the corresponding sensor and is not required to be non-negative. This indicates that the maximum value is taken from a non-empty set; if the set is empty, the output is unavailable and 0℃ is not padded. The example trigger threshold is 60℃; an alarm is triggered when the maximum temperature of any phase reaches this threshold. When there is only one valid sample, the maximum temperature equals the temperature of that sample. like Figure 3 As shown, edge terminal 100 takes two comparable previous values with comparable sampling times and calculates:
[0055] In the formula, and This is used to identify two different phases among phases A, B, and C. and The current effective temperature is the current paired value in the aforementioned temperature sample, and is comparable to the sampling time and location of the same type. This represents the non-negative temperature difference between phases, expressed in °C. Vertical lines indicate the absolute value of the real number. This indicates the difference. The example threshold is 35℃; reaching this threshold triggers an interphase temperature difference alarm. Edge terminal 100 checks each pair of observation time differences and time error upper limits to ensure they meet the tolerance registered in the temperature sampling configuration, without subtracting the maximum values occurring at different times in the two windows. Even with one phase missing, the remaining valid phases can still be identified, but a complete three-phase comparison cannot be declared. When load information is missing, only a temperature difference exceeding the limit is output, without specifying a contact resistance fault.
[0056] like Figure 3 As shown, edge terminal 100 further calculates the temperature rise of the current temperature relative to the previous 24 full hours:
[0057] In the formula, The index is the index of the 24 consecutive full hours preceding the current hour, and can be an integer from 1 to 24; For the sake of the prime minister The effective average temperature for the corresponding hour is obtained by taking the arithmetic mean of the effective equally spaced samples after deduplication for that hour. Use 24-hour temperature baseline; This is equivalent to the previous fresh temperature; This represents the temperature rise relative to the baseline. The horizontal line indicates the average value. This indicates a summation based on the given index; the subscript 24 represents 24 full hours, and "current" is a descriptive subscript. All temperature values and temperature rises are expressed in °C; temperature rises can be negative, indicating that the current temperature is below the baseline. An example temperature rise threshold is 40 °C, at which point an alarm is triggered. This value is not divided by time and is therefore not referred to as the rate of temperature rise.
[0058] like Figure 3 As shown, in the example sampling per second, the effective hourly sample ratio must be at least 90%, and the maximum continuous gap must not exceed 60 seconds to form a valid hourly average. If any necessary hour becomes invalid, the historical temperature rise criterion is paused, while the absolute temperature and available interphase temperature difference criteria continue to execute. The three temperature criteria obtain the required inputs and generate results respectively. The absolute temperature result is not an input to the interphase temperature difference, nor is the interphase temperature difference result an input to the historical temperature rise. The unavailability of any criterion only affects that criterion.
[0059] 3. Vibration branch like Figure 3 As shown, step S104 receives equally spaced acceleration samples from the triaxial vibration sensor 14. The example uses 12800 samples / second per axis and 12800 points per window, corresponding to a 1-second window; the sampling and anti-aliasing capabilities of the acquisition front-end 20 should be adapted to this configuration. The original mean and the result after removing the DC component are formed by the following formula:
[0060] In the formula, The axial markings of the triaxial vibration sensor 14 correspond to three pre-registered, mutually distinct axes; A positive integer representing the number of samples in the current vibration window; From 0 to Integer sampling index; , and These represent the original acceleration, the window-averaged acceleration, and the acceleration after removing the DC component, respectively, all in meters per second squared (m² / s), and are real numbers within their respective ranges. The sampling index and sample number are dimensionless, and summation is performed within the same complete window on that axis. The original mean is used for checking for installation offset or changes in gravity direction; the average value after removing the DC component is close to 0 and is not repeated as an independent defect criterion.
[0061] like Figure 3 As shown, the edge terminal 100 calculates the effective value, peak value, and peak-to-peak value of the AC acceleration for each axis:
[0062] In the formula, , and respectively the axis The effective value, absolute peak value, and peak-to-peak value of the AC acceleration are all non-negative, and the unit is meters per second squared; "effective", "peak", and "peak-to-peak" are descriptive subscripts. This means taking the minimum value within the current full window, and the square root taking the non-negative value. The summation, maximum value, and absolute value operations follow the aforementioned definitions. For example, if the effective value of a certain axis is 0.8 meters per second squared and the corresponding tuning threshold is 0.6 meters per second squared, the output will show that the effective value of that axis exceeds the limit, without canceling the result based on the lower amplitude of other axes.
[0063] like Figure 3 As shown, the edge terminal 100 uses the Hanning window and discrete Fourier transform to obtain the spectrum:
[0064] In the formula, is a dimensionless Hanning window coefficient, with a value between 0 and 1; For discrete frequency indexes, the range is 0 to 1000 in the complete transform. Integers; as axis The complex spectral coefficients, defined by discrete summation according to this formula, have the same units as acceleration; The imaginary unit satisfies ; Pi; and Let represent the cosine function and the natural exponential function, respectively, with dimensionless inputs. The frequency domain branching requirement for using the Hanning window is... and ;when When both window coefficients are 0, subsequent window gain normalization is not performed; instead, the output shows that the frequency domain branch is unavailable. This condition is the computable boundary of the formula and does not change the 12800-point example. The program can perform the above operations via Fast Fourier Transform. Windowing is used to reduce spectral leakage from finite window truncation, but it cannot recover frequency information lost due to insufficient sampling.
[0065] like Figure 3 As shown, for non-DC, non-Nyquist frequency points, edge terminal 100 calculates the single-sided amplitude estimate and frequency:
[0066] In the formula, This is a non-negative one-sided magnitude estimate after window gain correction, in meters per second squared. Represents the modulus of the complex spectral coefficients; For positive sampling frequency, The corresponding frequency points are all in Hertz (Hz), and "sample" is a descriptive subscript for the sampling frequency. Single-sided spectrum preservation. frequency points, This indicates rounding down; the above amplitude formula is only applicable to non-DC, non-Nyquist frequencies. DC frequency and even numbers Nyquist frequency at time Do not multiply by 2; odd numbers There is no frequency point that is exactly at the Nyquist frequency.
[0067] The example analysis frequency band is 10~5000 Hz, and the Nyquist frequency is 6400 Hz when the sampling frequency is 12800 Hz.
[0068] like Figure 3 As shown, the edge terminal 100 only searches for the main frequency within the valid frequency band after the amplitude exceeds the configured lower limit of the spectrum amplitude analysis, and compares the normal operating frequency band, RMS value, and peak value respectively. The example allows a main frequency range of 90~110 Hz; when 150 Hz is detected and the spectrum amplitude exceeds this analysis lower limit, a frequency band deviation record is generated. The RMS value, peak value, and peak-to-peak value of acceleration are calculated from windowed samples after removing the DC component; the windowed samples are used for spectrum calculation. Amplitude exceeding the limit and frequency band deviation are recorded separately; both are not required to trigger an alarm simultaneously.
[0069] 4. Branches related to partial discharge like Figure 3As shown, step S105 obtains the effective window peak values of the transient ground voltage sensor 12 and the acoustic emission sensor 13, respectively. For the example channel using voltage amplitude scaling, the scaling relationship is:
[0070] In the formula, To convert the transient ground voltage amplitude to a specified reference surface, Both are positive values within their effective range, calculated according to the acoustic emission front-end calibration relationship. The symbol for millivolt is . The unit symbol is microvolt. After converting the numerator and denominator to the same voltage unit, a dimensionless ratio is formed. and These are the logarithmic amplitudes with reference voltages of 1 millivolt and 1 microvolt, respectively, both expressed in decibels and can be negative; "electric" and "sound" are descriptive subscripts for the transient ground voltage channel and acoustic emission channel, respectively. The coefficient 20 corresponds to the voltage amplitude ratio, not the power ratio, used in this example. Data below the effective detection range are recorded according to the range rules, and the logarithm of zero amplitude is not taken. The acoustic emission equivalent voltage is not the uncalibrated actual sound pressure; other instrument scales must use their own definitions.
[0071] like Figure 3 As shown, in examples where the reference value, gain, and detection method are consistent with the above definitions, an alarm is triggered if the transient ground voltage exceeds 55 dB or the acoustic emission exceeds 30 dB. The two types of dB values are not added, subtracted, or directly averaged; they are not converted to picocoulombs when there is no discharge calibration relationship. When two channels simultaneously exceed their limits, candidate co-occurrences are first identified, and then it is checked whether the observation was added after the operation. Each of the two peak values is compared to the threshold of its respective channel; the merged result is the judgment result with channel identity, not just the two dB values; if one channel is invalid, the valid limit exceedance of the other channel is still output.
[0072] 5. Trend Branches like Figure 3 and Figure 5 As shown, step S106 establishes an extreme gradient boosting regression model for the continuous features of each measuring point. Taking the minute-average temperature as an example, a target value is formed only when the sampling integrity condition is met in that minute, and the next value is predicted using the previous 12 consecutive valid minute values; the history length can also be set to no less than 7. One-step prediction uses regression tree ensemble:
[0073] In the formula, It is an integer time index that increments with the continuous sampling period, not a physical moment in seconds; The length of the historical input should be an integer not less than 7, for example, 12; For the current measured target, For this reason Measured targets for a continuous period; For having The historical input vector of each component; This is a predicted value; Model version identifier The corresponding regression model. The number of trees is a positive integer. From 1 to Integer tree index; As the initial prediction constant, The learning rate is a dimensionless value between 0 and 1. For the first A regression tree maps the input vector to the output of the leaf nodes. The initial prediction constant, leaf node outputs, measured targets, and predicted values are all real numbers with the same dimensions as the target; in this temperature example, they are measured in °C. The model version identifier, index, and number of trees are dimensionless. Current measured values do not participate in predicting themselves.
[0074] like Figure 5 As shown, the edge terminal 100 stores each regression tree according to its feature number, split threshold, and left and right child node indices. During inference, it reaches the leaf node node by node based on the comparison result between the corresponding input component and the split threshold, and accumulates the output of the leaf node. Training uses a squared error objective, combined with a regularization term for the number of leaf nodes and the output value of the leaf nodes. The example trees are 100, the maximum tree depth is 3, the learning rate is 0.05, the training sample sampling ratio is 0.8, the feature sampling ratio of each tree is 1, and the squared regularization coefficient of the leaf node output value is 1.
[0075] like Figure 5 As shown, healthy samples are determined based on operation, inspection, or testing records, and cannot be labeled as healthy solely based on the absence of alarms in older models. The example divides 30 consecutive days of available data into 20 days of training, 5 days of parameter validation, and 5 days of independent error calibration, discarding input windows that cross segment boundaries or isolation zones. Trend branches are enabled after obtaining at least 2000 training windows and 200 calibration windows; if these are not reached, an insufficient sample count is displayed.
[0076] like Figure 5 As shown, the edge terminal 100 forms an absolute prediction error criterion based on independent calibration segments:
[0077] In the formula, The sample index for the independent error calibration segment is 1 to... Integers; The number of calibration samples for an effective error greater than 1 remains the same as the aforementioned minimum of 200 calibration windows. For the first The measured target of each error calibration sample Predictions generated for the same model using a previous valid historical window of the sample; This represents the absolute prediction error between the two. and These are the mean and sample standard deviation of the absolute prediction error, respectively. The absolute prediction error threshold. This represents the non-negative minimum threshold value, tuned based on measurement resolution and repeatability; both "error" and "minimum" are descriptive subscripts. The aforementioned error statistics are in the same unit as the target and are non-negative. Correction for the degrees of freedom calculated for the sample standard deviation. Current measured target. and current forecast value The absolute difference is greater than The output trend is abnormal; if it equals the threshold, the trend criterion is not triggered. Coefficient 2 is an existing empirical coefficient in this example, and it is not assumed that the error follows a normal distribution or has a fixed false alarm rate.
[0078] 6. Establishment of verification tasks and isolation of training samples. like Figure 5 As shown, when basic alarms or abnormal trends require continuous confirmation, or when the main evidence comes from the same sampling source within the operational impact range or when time comparability is insufficient, step S108 establishes a review task. The edge terminal 100 records the event identity, target chamber, source and pairing type, operational stage, diagnostic configuration set, new sampling requirements for each channel, window arrangement, and cutoff time. A task references only one set of defined criteria; regular version updates cannot change the established comparison benchmark for that task.
[0079] like Figure 5 As shown, when the task is established, the edge terminal 100 writes the data indexes of the operation impact interval, the backtesting interval before the anomaly, and the review period into the training sample isolation index. The backtesting length is configured according to the historical input span of the trigger feature; for example, a 12-minute temperature input should backtest for at least 12 minutes to avoid isolating only data after the trigger moment. If the write fails, an immediate alarm is maintained and the review is marked as not ready, without initiating a full review that might compete with model updates.
[0080] (v) Step S3: Verification of newly added observation data This step mainly corresponds to Figure 3 Steps S109 to S110 in the process, and concentrated combination Figure 6 Explain the conservative observation interval, the operational impact interval, the post-operation inclusion conditions, and the window center time condition; combined with Figure 7 Explain the validity and inclusion criteria of the newly added observation data.
[0081] 1. Conservative observation interval and operational impact interval like Figure 6 As shown, for the jj-th observation, a conservative observation interval is established for edge terminal 100:
[0082] In the formula, The index for observation records is a positive integer and dimensionless. and These are the actual start and end times of the observation recorded, respectively, satisfying... ; This is the upper bound of the non-negative time error of this record; This represents a conservative observation interval that the actual sampling might fall into. All times and time errors mentioned above use the same verified time base, and the unit is seconds; square brackets indicate intervals including endpoints. The upper bound of the time error covers clock deviation, time stamp marking, and sampling time error, and not only uses the network average delay; if a reliable upper bound cannot be obtained, this data must not be used for fine-grained cross-sampling source verification.
[0083] like Figure 3 and Figure 6 As shown, devices with only the receiving time can construct a conservative observation interval using the maximum data age limited by the protocol and sampling mechanism; where the maximum data age is the upper limit of the time interval between the receiving time and the earliest possible sampling time of the record. Without this upper limit guarantee, successful polling does not necessarily mean that the data is fresh.
[0084] like Figure 6 As shown, the edge terminal 100 establishes the operation influence range based on the actual feedback from the operation status interface 21:
[0085] In the formula, A positive integer index for the operation event; and These are the start and end times of the operation, respectively, satisfying... ; This is the upper bound for non-negative operation time error; and These are the non-negative pre-operation impact margin and the post-operation stability margin, respectively. This represents the area affected by the operation. All times and time intervals are in seconds. Dimensionless, "before" and "after" are descriptive subscripts, and square brackets indicate time intervals including boundaries. The margin is determined based on the sequence of actions and verified waveforms; for example, 0.5 seconds before the operation and 2 seconds after the operation. The receipt of the operation command cannot be directly used as feedback for the end of the mechanical action.
[0086] 2. Request Triggering Method like Figure 1 , Figure 3 and Figure 6 As shown, the main embodiment adopts a request-triggered method. After confirming that the acquisition front-end 20 has the capability, the edge terminal 100 sends a sampling request containing the task identity, target channel, planned start and end times, window length, and request sequence number through the sensor communication interface 104. After receiving the request, the acquisition front-end 20 starts a new conversion or window accumulation, and returns the device sampling sequence number, observation interval, and feature value upon completion. The edge terminal 100 saves the request acceptance receipt and the measurement completion record respectively, and only the latter is included in the verification if the observation conditions are met. The observation completion corresponding to the returned record must not be later than the time when the record can be received and used, and the request acceptance receipt cannot replace the measurement completion.
[0087] like Figure 6 As shown, the j-th newly added observation data used for post-operation review must satisfy the following:
[0088] In the formula, For the first The start time of the newly added observation data. This is the upper bound of the time error of the observation. For the reviewed operation event At the end of the day, Let its time error be the upper bound. To ensure stability after operation, all time parameters are measured in seconds. and These are dimensionless indices for observation and operation, respectively. The left side represents the earliest possible start time of the new observation, and the right side represents the latest end time of the operation's impact; if both sides are equal, the strict post-operation condition is not met. Simultaneously, the observation must be within the sampling plan established after the mission is set up, the original sampling identifier must not be reused in the mission, numerical validity and temporal correlation validity must both meet the requirements, and the target chamber's charged state information must be valid within its complete conservative observation interval, actually remaining charged, and unchanged during the operational phase. The fact that the window ends later than the operation does not preclude the possibility that the initial portion of the window still contains operational transients.
[0089] 3. Figure 7 The review and approval process is shown below. Combination Figure 7 After obtaining candidate new observation data in step S108, the edge terminal 100 performs the following judgments in sequence: First, determine whether the original sampling identifier has genuine new sampling proof and whether it has not been reused in this task. If not, it will not be included in this review, the original record and valid basic alarms will be retained, and the process will continue to wait according to the task deadline; if yes, proceed to the next step.
[0090] Second, determine whether the complete conservative observation interval is located after the end of the operational impact interval. If not, it will not be included in the post-operation review, and the process will continue to wait according to the task deadline; if so, proceed to the next step.
[0091] Third, determine whether the records of valid operating states and energized states cover the entire conservative observation interval. If not, output insufficient evidence and end the task; if yes, proceed to the next step.
[0092] Fourth, determine whether the target chamber is continuously energized and consistent with the operational phase corresponding to the task. If not, the operational phase changes, and the current verification task is interrupted; if yes, proceed to the next step.
[0093] Fifth, determine whether both the numerical validity and time correlation validity meet the task conditions. If not, the window is invalid, not counted, and continues to wait according to the task deadline; if yes, proceed to the next step.
[0094] Sixth, determine whether it is a cross-sampling source verification task. If not, include it in the complete and valid window for a single sampling source; if so, proceed to the next step.
[0095] Seventh, determine whether all source sets are known, non-empty, disjoint, and satisfy a preset pairing relationship. If not, retain the determination results for each sampling source and do not include cross-sampling source verification; if yes, proceed to the next step.
[0096] Eighth, determine whether the paired window meets the preset window center time tolerance and window length. If not, retain the judgment results of each sampling source and do not include them in the cross-sampling source verification; if yes, include them in the complete and valid cross-sampling source window.
[0097] Ninth, determine whether the task has obtained the required number of complete and valid windows within the deadline. If yes, proceed to step S109 to count the number of abnormal windows and output the review conclusion; otherwise, determine whether the task deadline has been reached. If it has not been reached, wait for the next new observation data; if it has been reached, output insufficient evidence and end this review task.
[0098] 4. Window Center Time Conditions like Figure 6 As shown, step S109 calculates the center time condition for the paired windows:
[0099] In the formula, candidate evidence A and B correspond to the two observation windows to be paired in this study; and These are the start and end times of observation at window A. and These are the start and end times of observation in window B; and This refers to the center time of the corresponding window; and This serves as the upper bound for the non-negative time error of each window; For the pre-registered positive window center time tolerance, all time quantities are in seconds. The end time of each of the two windows is no earlier than its start time; the absolute value of the time difference equals the tolerance, satisfying this check. This formula is used in conjunction with the same task, the same chamber, the same operating stage, and the allowable window length. Example: With a 1-second window and a tolerance of 0.1 seconds, when the windows of the two channels are aligned and the upper limit of the time error is 0.01 seconds for each, the most unfavorable center time difference is 0.02 seconds, which meets the requirements.
[0100] like Figure 6 As shown, the edge terminal 100 selects windows according to a preset pairing relationship, and does not move the window afterward to obtain co-occurrence due to a certain limit exceeding. The same original sample is not counted repeatedly across multiple verification windows; window-level co-occurrence is not equivalent to confirming the same microsecond-level discharge pulse.
[0101] (vi) Step S4: Continuity assessment and review conclusion This step mainly corresponds to Figure 3 Step S111 in the middle, and combined in a concentrated manner Figure 8 This section describes the basic alarm level, review conclusion, data availability, confirmation status, and recovery rules; it also incorporates... Figure 7 Explain the determination of the number of complete and valid windows and the output of insufficient evidence of expiration.
[0102] like Figure 3 and Figure 6 As shown, after the edge terminal 100 obtains a specified number of complete windows, it performs a persistence determination:
[0103] In the formula, The number of complete and valid review windows required by the task, rounded to a positive integer; From 1 to Integer window index; For the first An anomaly co-occurrence indicator for a complete and valid window is set to 1 when the source is non-repeating, the time is comparable, and both target criteria are triggered under the frozen diagnostic configuration; and to 0 when the window is complete and valid but not triggered together. The number of times the abnormal co-occurrence occurs is 0 to 10. Integers; To determine the number of times required for continuous co-occurrence, take the number that satisfies... Integers; "total" is a descriptive subscript. All counts and indications are dimensionless. Incorrect values will occur if the window is missing or of invalid quality. Assigning 0 will not include this window. Three complete and valid windows; in this example, at least two co-occurrences are considered in three complete and valid windows, i.e. , .
[0104] Combination Figure 7 After the edge terminal 100 obtains the required number of complete valid windows, it counts the number of abnormal windows according to the diagnostic configuration set and outputs the review conclusion. If the required number of complete valid windows is still insufficient by the task deadline, it outputs insufficient evidence, does not change the missing windows to windows that have not triggered the criterion, and does not recalculate the completion ratio based on the windows already obtained. For cross-sampling source review tasks, the source pairing and window time conditions must also be checked separately; if the required number of valid windows is not reached, it simply continues to wait within the task period, and outputs insufficient evidence upon expiration.
[0105] like Figure 3 and Figure 8 As shown, the deadline for the 1-second feature review task in this example is 30 seconds, starting from when the task is created and the diagnostic configuration binding is completed. If less than 3 complete and valid windows are completed by the deadline, the output "insufficient evidence" will be displayed. "Not reproduced" will only be displayed if none of the target criteria are triggered within all complete and valid windows; if an anomaly of a certain target is reproduced but does not reach the corresponding duration, "partially reproduced" will be displayed.
[0106] like Figure 8 As shown, the edge terminal 100 outputs the basic alarm level, review conclusion, data availability, and operational stage. The basic alarm level is 2 when a valid basic limit is exceeded; it is 1 when no valid basic limit is exceeded but a trend anomaly or other concerning state occurs; and it is 0 only when the required immediate branch is complete, no basic limit is triggered, and no enabled trend branch anomaly occurs. When there is no valid anomaly but the required immediate branch is missing, the basic comprehensive result is undetermined rather than 0.
[0107] like Figure 8 As shown, the local display 200 shows the abnormal interval, trigger value, threshold, occurrence time, missing channel, and review status. The alarm acknowledgment device 210 inputs personnel confirmation into the edge terminal 100. Assuming the absolute temperature trigger value is 60°C, the example recovery rule is that the temperature is below 55°C and the conditions are met for three consecutive fresh windows, while the necessary status information is valid. The confirmation input of the alarm acknowledgment device 210 only changes the read status and cannot exceed the recovery conditions.
[0108] like Figure 8 As shown, basic alarms, review conclusions, data availability, and confirmation status are stored separately and can coexist. Recovery must independently meet recovery rules and cannot be replaced by confirmation, non-reproducibility, or insufficient evidence; historical events are retained at each stage of occurrence, review, and recovery.
[0109] (vii) Step S5: Model update and training sample isolation This step mainly corresponds to Figure 5The training sample isolation, admission, candidate model training, pre-release review, and configuration set switching are all part of the process.
[0110] like Figure 5 As shown, step S110 causes new records to first enter the original storage and then delays admission; the example delay is 24 hours. The target and all historical inputs in the training window must be numerically and temporally valid, without isolation flags, be in the model's applicable phase, and have no unclosed related anomalies; reaching the delay does not necessarily indicate health. Failure to reproduce the anomaly during review does not automatically release the original operation segment or the abnormal segment; only subsequent stable segments are allowed to be reviewed according to the same rules. Disputed historical segments are released from isolation through independent verification of records.
[0111] Combination Figure 5 The isolation and inspection of the target and all historical inputs are the admission conditions. If any one of them is not met, the candidate will not be allowed to enter this training. The expiration of the delay period and the completion of the review task are not equivalent to the automatic release of isolation.
[0112] Combination Figure 5 The admission delay is calculated from the end time of observation of the target sample in the candidate training window and is not restarted due to message retransmission; late operations or abnormal records can still update the isolation index. Fixed health validation sets, independent error calibration segments, and confirmed abnormal playback records are not allowed to enter the current candidate training set, and historical input windows that cross these boundaries are also excluded to avoid using the same window for training and validation.
[0113] like Figure 5 As shown, in an example, a channel is isolated from minutes 100 to 110. If the model uses the predictions from the previous 12 minutes to determine the current value, then when the target is used from minutes 111 to 122, the isolated input will still be referenced, and the entire training window will be closed. The input for minute 123 is from minutes 111 to 122, and it will only be allowed if there are no other reasons for isolation of these inputs and the target, the runtime is compatible, the relevant anomalies have been closed, and the latency is met.
[0114] like Figure 5 As shown, edge terminal 100 checks the number of newly added qualified windows daily during low-load periods. A candidate model is generated after 500 windows are reached; model updates are postponed if relevant verifications are not completed. The candidate model is checked on a fixed health validation set and an anomaly replay set with clearly labeled anomalies. The example allows the mean absolute error of health to not exceed 1.05 times that of the old model, while requiring that the detection rate of confirmed anomalies does not decrease and the base threshold remains unchanged. If there is no anomaly replay evidence, it is run in bypass mode and automatically released without relying on a decrease in health error.
[0115] like Figure 5As shown, in the exemplary deployment refinement method, the edge terminal 100 saves the candidate training sample index and the isolation index version, and checks again before deployment whether there are any newly added operation markers or late arrival exception records. If the input or target used by the candidate is newly isolated, the deployment is canceled and retraining is performed. It is not permissible to simply delete samples from the index table and then reuse the model that they have already participated in generating. When the conditions are met, the processor 101 updates and switches to the complete diagnostic configuration set containing the new model and its corresponding error threshold once through indivisible configuration references, while retaining the old diagnostic configuration set and task references; if deployment fails, the old diagnostic configuration continues to be used, and the saved predictions and alarms are not written back.
[0116] (viii) Verification across sampling sources This section combines... Figure 9 Explain the physical sampling sources, the set of sampling sources, evidence of non-repeating sampling sources, the pre-defined pairing relationships, and the conditions under which the source sets do not intersect. Figure 9 This will be explained in detail here.
[0117] Combination Figure 9 Physical sampling sources refer to the measurement entry point jointly determined by the sensor identifier and the physical measurement channel, not the communication address or algorithm name. The sampling source set refers to all physical sampling sources referenced in forming a feature or diagnostic result. For example, the absolute value of the temperature of the same phase and the absolute prediction error contain the same source, and the temperature difference between two phases simultaneously contains sources from both phases. Evidence of non-repeating sampling sources requires that the selected source set be disjoint and satisfy the configured physical quantity pairing relationship, but does not mean that these sources are statistically independent.
[0118] like Figure 9 As shown, step S107 first checks the source set of candidate evidence:
[0119] In the formula, , These are the sets of physical sampling sources cited by candidate evidence A and candidate evidence B, respectively; "A" and "B" are fixed identifiers used to distinguish the two candidate pieces of evidence and do not indicate the electrical phase. To represent the intersection of sets, The empty set is represented by a dimensionless set and its identifier. Both source sets must be known and non-empty; unknown or missing sources cannot be represented as an empty set to satisfy the no-intersection condition. Satisfying this condition is only a necessary condition for forming evidence of non-resampled sources. Edge terminal 100 also checks preset pairing relationships: transient ground voltage and acoustic emission support anomalous co-occurrence of partial discharge-related signals; temperature and vibration only support anomalous co-occurrence of different physical quantities; acoustic and vibration results derived from the same piezoelectric element are treated as related sources by default.
[0120] like Figure 9As shown, when the source is unknown or there is overlap, edge terminal 100 retains all algorithm results but does not increase the number of non-repeated sampling sources. Existing time pairing or multi-source fusion cannot replace this channel-level relationship check; this embodiment further requires that the observations used for enhanced confirmation come from subsequent review stages and are linked to fixed versions and training admissions. Disjoint source sets are only used to screen candidate evidence that allows pairing and cannot replace real new observations, time comparability, and persistence checks; anomalies with overlapping sources can still be subject to persistent review from a single sampling source.
[0121] III. Device Examples like Figure 10 As shown, the present invention also provides a multi-source status analysis and early warning device 400 for high-voltage switchgear, which is installed in... Figure 1 The edge terminal 100 shown includes: a model and training management module 407, a data access module 401, a quality and source management module 402, a branch diagnosis module 403, an operation phase management module 404, a review and scheduling module 405, a continuous review module 406, an alarm and storage module 408, and a self-test module 409.
[0122] The model and training management module 407 stores a diagnostic configuration set containing thresholds and model parameters, and provides the diagnostic configuration set to the branch diagnostic module 403 and the review scheduling module 405. The model and training management module 407 implements the freezing of step S108 and the admission and updating of step S110 through read-only version objects, isolation indexes, training queues, and release pointers. The model and training management module 407 provides applicable criteria to the branch diagnostic module 403, provides task-bound versions to the review scheduling module 405 and the continuous review module 406, and receives task and isolation status.
[0123] The data access module 401 sends monitoring data containing the physical sampling source, original sampling identifier, and observation interval to the quality and source management module 402 for validity verification and deduplication based on the original sampling identifier. The data access module 401 executes the interface configuration in step S101 and the message parsing in step S102. The underlying structure consists of a serial communication receiving task or a network receiving task and a field parsing function. The input is the original message and measurement point configuration, and the output is a unified data object to the quality and source management module 402. If parsing fails, an error object is returned; falsified measurement values are not output.
[0124] The quality and source management module 402 can be composed of a verification function, an original sampling identifier index, and a source set operation function to complete the validity processing in step S102 and the source check in step S107.
[0125] The branch diagnostic module 403 diagnoses the monitoring data verified by the quality and source management module 402 according to the diagnostic configuration set, and sends valid out-of-limit results to the alarm and storage module 408 and the review and scheduling module 405 respectively. The alarm and storage module 408 directly outputs an alarm based on the valid out-of-limit result. The branch diagnostic module 403 includes a temperature comparison function, a vibration transformation function, a partial discharge related peak comparison function, and a tree model inference function, and executes steps S103 to S106. After receiving a valid object, the branch diagnostic module 403 outputs a result with source and version to the alarm and storage module 408, and hands over anomalies that need continuous confirmation to the review and scheduling module 405; invalid branches explicitly return unavailable.
[0126] The operation phase management module 404 determines the operation impact range and operation phase based on the received operation status and energized status, and provides this information to the review and scheduling module 405. The operation phase management module 404 updates the operation phase through a status edge queue and range index, completing the operation impact judgment in step S107. The operation phase management module 404 receives and quality-verifies the status information provided by the isolation status quantity input interface 105.
[0127] The review scheduling module 405 establishes a review task bound to the diagnostic configuration set based on the valid limit violation results, and sends the status information of the task to the model and training management module 407. During the task, the model and training management module 407 maintains the bound diagnostic configuration set unchanged and marks the samples in the backtracking interval before the anomaly and during the review period as isolated samples. The review scheduling module 405 executes step S108 through the task state machine and sampling request queue, receiving anomalies and status conditions, and sending a request to the sensor communication interface 104 or registering for natural update waiting. The review scheduling module 405 obtains new observation data through the sensor communication interface 104. The new observation data is processed by the data access module 401 and the quality and source management module 402 before being handed over to the continuous review module 406.
[0128] The continuous review module 406 receives the review task, operational impact range, and operating phase provided by the review scheduling module 405, and receives new observation data provided by the quality and source management module 402. Only valid new observation data whose observation range is after the operational impact range and belongs to the same energized operating phase corresponding to the task are included in the task; others are excluded. Specifically, new observation data is included in the task only if it simultaneously meets the following conditions: it has genuine new sampling proof, i.e., its original sampling identifier is different from the most recent original sampling identifier saved by the task; the observation range is after the operational impact range; and it belongs to the same energized operating phase, i.e., the target chamber remains energized and the operating phase remains unchanged within its complete observation range; otherwise, it is excluded. The continuous review module 406 executes step S109, verifying the observation identity, interval, freeze criterion, and window count, and sends the results to the alarm and storage module 408. Timeouts, duplicates, or phase changes are output with corresponding reasons, and are not combined into a single "normal" result. When the continuous review module 406 obtains the specified number of complete and valid windows within the task period, it counts the number of abnormal windows according to the diagnostic configuration set and sends the review conclusion to the alarm and storage module 408; otherwise, it sends an insufficient evidence result upon expiration. The alarm and storage module 408 saves and outputs the received results.
[0129] When the model and training management module 407 updates the model parameters, it removes training windows containing isolated samples from the target and all historical inputs. The isolated samples are recorded by the training sample isolation index, which is stored in the same task record as the version reference of the diagnostic configuration set. When any historical input sample hits the isolation index, the entire candidate training window is not allowed to enter.
[0130] The alarm and storage module 408 executes step S111 through the event index, append logs and pending queue, connects the local display 200 and the remote platform 300, and receives input from the alarm acknowledgment device 210 but does not delete the original evidence based on it.
[0131] The self-test module 409 is implemented through periodic check tasks and heartbeat monitoring tasks, sending failure and recovery conditions to each module, thus providing clear call boundaries for exception handling. The self-test module 409 provides failure or recovery conditions but does not act as an operational event or a source of measurement data.
[0132] For more details, please refer to the aforementioned [reference]. Figures 1 to 9 The description of that will not be repeated here.
[0133] IV. Examples of Electronic Devices like Figure 11 As shown, the present invention also provides an electronic device, including a processor 101 and a memory; the memory stores a computer program, and when the computer program is executed by the processor 101, the electronic device acts as an edge terminal 100 to implement the aforementioned high-voltage switchgear multi-source status analysis and early warning method.
[0134] One implementation of the electronic device is an edge terminal 100, whose processor 101 is connected via a communication bus 109 to a running memory 102, a non-volatile memory 103, a sensor communication interface 104, an isolated status input interface 105, a local display interface 106, an alarm confirmation input interface 107, and an uplink communication interface 108. The processor 101 may be an industrial processor with integer, floating-point, and interrupt processing capabilities; the sensor communication interface 104 may be an isolated differential transceiver circuit; the local display interface 106 is connected to a local display 200; the alarm confirmation input interface 107 is connected to an alarm acknowledgment device 210; and the uplink communication interface 108 is connected to a remote platform 300.
[0135] Combination Figure 11 The displayed information is sent to the local display 200 via the local display interface 106, and personnel confirm the alarm by inputting the alarm confirmation input interface 107 via the alarm confirmation device 210; the two do not add touch feedback or alarm clearing commands due to the indicated connection.
[0136] like Figure 11 As shown, processor 101 loads the program and version configuration from non-volatile memory 103 into runtime memory 102, processes the real sensor input according to steps S101 to S111, controls the new sampling process through sensor communication interface 104, and forms the field visible status through local display interface 106. Version objects use read-only references, and task logs use ordered submission; sampling interrupt handling takes precedence over training tasks. When the acquisition front-end 20 shares the spectrum or peak calculation, processor 101 still verifies its source and observation range, and does not omit the verification qualification check because the algorithm is moved to external hardware.
[0137] For more details, please refer to the aforementioned [reference]. Figures 1 to 9 The description of that will not be repeated here.
[0138] V. Examples of Computer-Readable Storage Media The present invention also provides a computer-readable storage medium storing a computer program; when the computer program is executed by the processor 101 of the edge terminal 100, it implements the multi-source status analysis and early warning method for high-voltage switchgear.
[0139] like Figure 11 As shown, the computer-readable storage medium can be flash memory, electrically erasable programmable read-only memory, or solid-state memory to store program instructions, protocol mappings, diagnostic configuration sets, and task configurations. It can also be random access memory to store windows and indexes during program execution. The medium is a physical medium capable of storing computer-readable information, not the signal itself during transmission. When the program is executed by processor 101, it implements the aforementioned on-site sampling, branch alarm, fixed version verification, sample isolation, and recovery processing. The volatile working cache does not replace the non-volatile storage required for evidence retention after power failure.
[0140] For more details, please refer to the aforementioned [reference]. Figures 1 to 9 The description of that will not be repeated here.
[0141] Implementing the embodiments of the present invention has the following beneficial effects: This invention provides a method, device, equipment, and storage medium for multi-source status analysis and early warning of high-voltage switchgear. By executing real-time alarms and evidence verification in parallel, it enables monitoring data that exceeds the effective limit to directly trigger alarms, and forms a verification conclusion only based on a specified number of complete and valid windows. If the deadline is insufficient, it outputs insufficient evidence, avoiding waiting for verification to block alarms and misclassifying missing windows as normal. In this invention, the original sampling identifier is used to deduplicate and the sampling source set and preset pairing relationship are verified so that message retransmission and derived results from the same source do not increase evidence of non-duplicate sampling sources, thus avoiding duplicate data from increasing the number of anomalies and the number of sources. In this invention, the conservative observation interval is compared with the operation influence interval and combined with the operation state and energized state records covering the entire conservative observation interval to verify the energized operation stage of newly added observation data, thereby reducing the risk that operation transients are counted as continuous abnormalities and signal fallback when not energized is judged as defect recovery. In this invention, by binding the diagnostic configuration set to the review task and isolating the pre-abnormal backtesting interval and the review period samples, while keeping the criteria unchanged, the training window involving the target and all historical inputs involving isolated samples is excluded, thus synergistically limiting the impact of criterion changes and historical input contamination on the abnormality judgment. In this invention, when uplink communication is interrupted, the edge terminal continues to collect, verify and display data locally, and retransmits the data according to the event identity and record version after communication is restored. This ensures that the same event is not counted repeatedly due to retransmission, retains the old and new verification conclusions, and maintains the continuity of on-site monitoring and anomaly recording.
[0142] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 Units that specify functions within one or more boxes.
[0143] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.
Claims
1. A multi-source status analysis and early warning method for high-voltage switchgear, executed by an edge terminal, characterized in that, include: Acquire monitoring data, including physical sampling source, original sampling identifier, and observation range, as well as operational and energized status; For monitoring data that exceeds the effective limit, an alarm is output. A review task is established that is bound to the diagnostic configuration set. The samples in the backtracking interval before the anomaly and during the review period are marked as isolated samples. The diagnostic configuration set includes the threshold and model parameters and remains unchanged during the task. Based on the operational status, the operational impact range is determined. Duplicate data are removed according to the original sampling identifier. Only valid new observation data whose observation range is after the operational impact range and belongs to the same energized operation phase are included in this task, and the rest are not included. If the required number of complete and valid windows are obtained within the task deadline, the number of abnormal windows is counted according to the diagnostic configuration set and a review conclusion is output; otherwise, insufficient evidence is output by the deadline. When updating model parameters, remove training windows containing isolated samples from the target and all historical inputs.
2. The method according to claim 1, characterized in that: The edge terminal subtracts the corresponding upper bound of the time error from the start time of the observation interval and adds the upper bound to the end time to form a conservative observation interval; the sum of the operation end time, the upper bound of the operation time error and the stability margin after the operation is determined as the end boundary of the operation influence interval; Wherein, the conservative observation interval satisfies: In the formula, The index for observation records is a positive integer and dimensionless. and These are the actual start and end times of the observation recorded, respectively, satisfying... ; This is the upper bound of the non-negative time error of this record; This represents a conservative observation range that the actual sampling might fall into; The operation's influence range satisfies: In the formula, A positive integer index for the operation event; and These are the start and end times of the operation, respectively, satisfying... ; This is the upper bound for non-negative operation time error; and These are the non-negative pre-operation impact margin and the post-operation stability margin, respectively. The area affected by the operation; The newly added observation data will only be included in the verification task if the start time of the conservative observation interval is later than the boundary, and the valid records of the operating status and energized status cover the interval and indicate that the target chamber is continuously energized and the operating phase has not changed. Otherwise, the data will not be included and the alarm will be retained.
3. The method according to claim 2, characterized in that: The review task is a cross-sampling source review task; the edge terminal establishes the sampling source set for each of the two candidate pieces of evidence based on the physical sampling source; If the sampling source sets are all known and non-empty, mutually disjoint, and the corresponding physical quantities satisfy the preset pairing relationship, and the sum of the absolute value of the time difference at the center of the corresponding window and the upper bound of their respective time errors is not greater than the preset time tolerance at the center of the window, then the candidate evidence will be used as paired evidence for continuous review according to the diagnostic configuration set; otherwise, the judgment results of the candidate evidence will be retained and will not be included in the cross-sampling source review. Wherein, the set of sampling sources satisfies: In the formula, , These are the sets of physical sampling sources cited by candidate evidence A and B, respectively; To represent the intersection of sets, Represents the empty set; The window center time condition is satisfied as follows: In the formula, candidate evidence A and B correspond to the two observation windows to be paired in this study; and These are the start and end times of observation at window A. and These are the start and end times of observation in window B; and This refers to the center time of the corresponding window; and This serves as the upper bound for the non-negative time error of each window; This is the pre-registered positive window center time tolerance.
4. The method according to claim 3, characterized in that: The edge terminal establishes a training sample isolation index that records isolated samples and the reasons for isolation, and saves the index and the version reference of the diagnostic configuration set in the same task record; the relevant model update is postponed until the review task is completed, and the candidate model is trained with the training window that has passed the isolation check after the task is completed. Before release, the training sample index of the candidate model is reviewed. If newly labeled isolated samples are found, the release is canceled and retraining is performed. Otherwise, the candidate model is released if the check based on the fixed health validation set and the confirmed abnormal replay set passes. If the check fails, the diagnostic configuration set is maintained.
5. The method according to claim 4, characterized in that: The edge terminal sends a sampling request associated with the verification task to the acquisition front end that supports request-triggered sampling, and saves the request acceptance receipt and measurement completion record respectively; it performs a validity check and a verification task inclusion condition check on the measurement completion record. If the check is passed, the corresponding new observation data is included in the task; otherwise, it is not included. The inclusion conditions include: In the formula, For the first The start time of the newly added observation data. This is the upper bound of the time error of the observation. For the reviewed operation event At the end of the day, Let its time error be the upper bound. To provide a stable margin after operation; The receipt for acceptance of the request does not replace the measurement completion record. The observation end time corresponding to the measurement completion record shall not be later than the time when the record is received and used.
6. The method according to claim 5, characterized in that: When establishing a review task, the edge terminal saves the most recent original sampling identifier of the physical sampling source and waits for the collection front end to update the record naturally in a periodic manner, without sending a request to trigger a sampling command; If the original sampling identifier in the periodic natural update record is different from the saved original sampling identifier and meets the conditions for inclusion in the review task, it will be counted; otherwise, it will be rejected and the alarm will be retained. The original sampling identifier is associated with the device identifier, physical measurement channel, sampling session, and device sampling sequence number updated with actual conversion. New sampling is not identified by incrementing the message sequence number.
7. The method according to claim 6, characterized in that: The edge terminal acquires the cumulative peak value from the acquisition front end. When the acquisition front end has a real-time monitoring window and a verification cumulative window that are stored separately, the cumulative peak value before clearing is stored and a clearing command is only sent to the verification cumulative window. Otherwise, the value is not cleared and the output capacity is insufficient. After obtaining the clearing completion feedback, the start and end times of the new cumulative value are verified. If the conditions for the verification task to be included are met, the value is counted. Otherwise, the value is not counted. If the clearing completion feedback is not obtained, the return value is saved as the historical cumulative value. Each counting window corresponds to a different zeroing completion feedback and observation interval, and overlapping accumulation windows are not counted repeatedly.
8. The method according to claim 7, characterized in that, in: The sampling request includes the task identity of the verification task, the target channel, the planned start and end times, the window length, and the request number. The measurement completion record includes the device sampling number, the observation interval, and the characteristic value. The edge terminal associates and saves the sampling request, request acceptance receipt and measurement completion record based on the task identity and request sequence number, and associates the diagnostic configuration version identifier, original sampling identifier and review conclusion of the diagnostic configuration set with the same task record; the request sequence number and the device sampling sequence number are saved separately.
9. The method according to claim 8, characterized in that: The edge terminal stores the basic alarm level, review conclusion, data availability and confirmation status respectively, and displays the trigger value, threshold and review conclusion on the local display. When a confirmation input is received from the alarm acknowledgment device, only the confirmation status is updated; otherwise, the confirmation status is maintained. The diagnostic configuration set also includes recovery rules. When the criteria for triggering an alarm meet the recovery rules, the corresponding current alarm status is lifted; otherwise, the status is retained. Unavailable data is not recorded as normal, and historical events are retained at each stage of occurrence, review, and recovery.
10. The method according to claim 9, characterized in that: The edge terminal stores the event identity, record version, and review conclusion corresponding to the alarm in non-volatile memory; When uplink communication is interrupted, data acquisition, diagnosis, verification, and local display continue, and unsent records are marked as pending. When uplink communication is normal, send pending records to the remote platform; When a receipt confirmation corresponding to the event identity and record version is received, the corresponding record is marked as sent; otherwise, it is kept pending. Retransmission does not change the event identity and record version. The review conclusion is updated by appending the higher record version, and the original record is retained.
11. A multi-source status analysis and early warning device for high-voltage switchgear, installed at an edge terminal, characterized in that, It should include at least: a model and training management module, a data access module, a branch diagnosis module, an operation phase management module, a review and scheduling module, and a continuous review module; among which: The model and training management module is used to store the diagnostic configuration set containing thresholds and model parameters, and to provide the diagnostic configuration set to the branch diagnostic module and the review scheduling module; The data access module is used to send monitoring data containing physical sampling sources, original sampling identifiers, and observation intervals to the quality and source management module for validity verification and deduplication based on the original sampling identifiers; The branch diagnostic module is used to diagnose the monitoring data verified by the quality and source management module according to the diagnostic configuration set, and send the valid limit violation results to the alarm and storage module and the review and scheduling module respectively. The alarm and storage module directly outputs alarms based on the valid limit violation results. The operation phase management module is used to determine the operation impact range and operation phase based on the received operation status and energized status, and provides this information to the review and scheduling module. The review scheduling module is used to establish a review task bound to the diagnostic configuration set based on the valid over-limit results, and send the status information of the task to the model and training management module. The model and training management module keeps the bound diagnostic configuration set unchanged during the task, and marks the samples in the backtracking interval before the anomaly and during the review period as isolated samples. The continuous review module is used to receive the review task, operation impact range and operation stage provided by the review scheduling module, and to receive the new observation data provided by the quality and source management module. Only the valid new observation data whose observation range is after the operation impact range and belongs to the same live operation stage corresponding to the task are included in the task, and the rest are not included. The continuous review module is used to count the number of abnormal windows according to the diagnostic configuration set and send the review conclusion to the alarm and storage module when a specified number of complete and valid windows are obtained within the task period. Otherwise, it sends the result of insufficient evidence when the deadline is reached. The alarm and storage module saves and outputs the received result. The model and training management module is used to remove training windows containing isolated samples in the target and all historical inputs when updating model parameters. The isolated samples are recorded by the training sample isolation index, which is stored in the same task record as the version reference of the diagnostic configuration set. When any historical input sample hits the isolation index, the entire candidate training window is not allowed to enter.
12. An electronic device, characterized in that, It includes a processor and a memory; the memory stores a computer program, which, when executed by the processor, enables the electronic device to act as an edge terminal to implement the multi-source status analysis and early warning method for high-voltage switchgear as described in any one of claims 1 to 10.
13. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program; when the computer program is executed by the processor of an edge terminal, it implements the high-voltage switchgear multi-source status analysis and early warning method as described in any one of claims 1 to 10.
