A Smart Identification System for Latent Defects in Electricity Meters Based on Fusion Density Clustering
By integrating density clustering and implicit relational databases, latent faults in dismantled electricity meters are identified, solving the identification problem of low-proportion meter groups, improving the accuracy and reliability of the sorting process, and achieving the suppression of detection station drift and effective identification of latent faults.
Patent Information
- Application Number
- CN202611305766.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-26
- Publication Date
- 2026-11-17
AI Technical Summary
In the absence of sufficient similar latent fault labels, how can we utilize the full amount of test data and historical re-inspection results generated from the sorting of dismantled electricity meters to identify a low-proportion group of meters with a common failure mechanism and suppress pseudo-group anomalies caused by test station drift?
By fusing density clustering, a comparable sample domain and a latent relation database are constructed. Using explicit feature vectors and historical re-examination conclusions, latent relation edges are established, latent activation vectors are generated, and early warning trigger results are identified and output. Through targeted re-examination tasks, latent faults are confirmed or denied, forming a stable identification system.
It effectively identifies a low-proportion group of meters with common failure mechanisms, reduces the impact of detection station drift, improves the accuracy and reliability of the sorting process, and ensures the interpretability and verifiability of latent faults.
Smart Images

Figure CN123116007A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electricity meter condition detection and sorting data processing technology, and more specifically, to an intelligent identification system and method for latent defects in electricity meters based on density clustering. Background Technology
[0002] Sorting dismantled electricity meters is a crucial step in the quality assessment and disposal of metering assets after they are removed from field operation. The sorting line typically retrieves asset files based on meter identification, and generates inspection records by examining items such as visual inspection, metering performance, clock, communication, cost control, event logs, load switches, and power supply status. Based on this, a conclusion is given regarding whether the meter is pending verification, maintenance, review, or scrapping. Because automated workstations can continuously generate multiple values and status codes for the same meter, the sorting data already provides a basis for describing the meter's remaining service life and signs of failure.
[0003] Publication No. CN110348401A discloses a "System and Method for Classifying and Processing Retrieved Smart Meters." This solution associates the images, numbers, specifications, recycling inspection contents, and accuracy test results of the retrieved meters to form classification and disposal results. This document solves the problem of connecting the information collection and classification process of retrieved meters and can be regarded as a relatively close engineering processing path in this field; however, its judgment still mainly relies on the pre-set inspection contents and input results, and lacks an active discovery process for group anomalies that are not directly defined by existing projects.
[0004] In the field of data analysis, publication number CN105866725A discloses a "Smart Meter Fault Classification Method Based on Cluster Analysis and Cloud Model." This method utilizes historical smart meter fault samples for K-means clustering, then establishes a fault cloud model based on cluster centers and further subdivides fault types. This technology introduces clustering into electricity meter fault classification, providing an analytical supplement to the aforementioned sorting data processing path. However, its starting point is existing fault samples and a preset number of clusters, making it difficult to cover situations where the proportion of fault samples is very low, cluster shapes are irregular, and fault names are yet to be determined.
[0005] Publication No. CN111401785A discloses a "Power System Equipment Fault Early Warning Method Based on Fuzzy Association Rules." This scheme uses clustering to determine attribute intervals and establishes a rule base through association rules to predict equipment fault states. Compared with simple classification techniques, this document further utilizes the co-occurrence relationship between attributes; however, rule support is easily affected by common factors such as equipment model, service stage, and inspection station. If historical co-occurrence is directly used for sorting in the data recovery table, accidental associations may become fixed as the data volume increases.
[0006] The above technologies respectively provide a scrapped meter sorting process, a fault clustering method and an association rule idea, which have a scenario inheritance and analysis complementation relationship with each other. The existing sorting still faces a type of situation that is difficult to handle directly: a small number of meters from the same manufacturing batch or with similar software and hardware versions may not have a single out-of-limit item, but show consistent gradual change or correlation on multiple indicators such as battery voltage, daily timing error, communication success rate, and fee control authentication status; these meters are scattered near normal samples, fixed thresholds may judge them as qualified, and ordinary clustering tends to merge them into a large normal cluster or regard them as noise points. On the other hand, the drift of the detection station will also create similar groups. If this confounding effect is not controlled, it is difficult for the early warning result to distinguish equipment defects from test condition abnormalities.
[0007] Therefore, the technical problem to be solved is: in the absence of sufficient labels for similar hidden faults, how to use the full amount of detection data and historical re-inspection results already generated by scrapped meter sorting, convert the group distribution change of explicit indicators into interpretable hidden fault evidence, and stably identify meter groups with a common failure mechanism when sparse small clusters, missing measurement channels and station drift coexist. Contents of the Invention
[0008] In order to identify the low-proportion meter groups with a common failure mechanism from the sorting data of scrapped electric energy meters under the condition that hidden fault samples are rare and a single detection result does not necessarily exceed the limit, and suppress the pseudo-group abnormality caused by the drift of the detection station, the present invention proposes an intelligent identification system for hidden defects of electric energy meters integrating density clustering.
[0009] Comprising: a detection data access unit, configured to collect the full amount of sorting detection indicators, asset pedigree information and historical re-inspection conclusions according to the electric energy meter identifier, and form an explicit feature vector from the full amount of sorting detection indicators; A comparable domain construction unit, configured to construct a comparable sample domain according to the asset pedigree information, and establish a reference distribution of each explicit detection feature in the comparable sample domain according to historical stable samples; A hidden relationship library construction unit, configured to control the confounding effects caused by asset pedigree and detection stations, determine hidden relationship edges with cross-stratified codirectional conditional co-occurrence residuals between the historical explicit feature distribution offset and the re-inspection confirmed fault, and form a hidden relationship library; A distribution activation unit, configured to calculate the interval proportion change of the current sorting window in the risk direction according to the unified quantile boundary determined by the reference distribution, and generate a hidden activation vector for each electric energy meter to be identified along the hidden relationship edge; A dual-space density discovery unit, configured to establish density adjacency for the electric energy meters to be identified that simultaneously satisfy the explicit neighborhood condition and the hidden neighborhood condition, so as to obtain candidate meter clusters; The cluster triggering unit is used to generate early warning triggering results based on the proportion of node triggering samples of candidate meter clusters, the coverage of independent evidence channels, and the concentration of asset lineage. The early warning output unit is used to output candidate meter identifiers, evidence paths, and suggested re-inspection items based on the early warning triggering result. The re-inspection and backfeeding unit is used to execute the targeted re-inspection task generated by the mapping relationship between the implicit relationship edge and the re-inspection item, and to backfeed the confirmed data and negative data to the implicit relationship database respectively.
[0010] The detection data, comparable sample domains, relation edges, meter-level activation, candidate clusters, early warning outputs, and targeted re-inspections are connected according to a clear data connection relationship. The early warning conclusion includes candidate meters, as well as retaining the explicit evidence, implicit nodes, and re-inspection items that form the conclusion; after the confirmation and rejection results are returned to the relation database, subsequent windows can use the updated applicable stratification and credibility, avoiding the long-term solidification of a single correlation result.
[0011] Furthermore, the implicit relationship database construction unit stratifies historical data according to manufacturing source, specifications, hardware batch, software version, service duration range, and testing station. Within each stratum, the failure confirmation ratio in samples showing the same explicit feature shift is subtracted from the failure confirmation ratio in samples not showing the same explicit feature shift, yielding the conditional co-occurrence residual pointing to the corresponding implicit fault node. Only when the conditional co-occurrence residual is positive in multiple strata satisfying the minimum sample size, and the explicit feature shift ratio in unconfirmed samples does not reach the rejection threshold, is a corresponding implicit relationship edge established. The evidence channel, shift direction, credibility, applicable stratum, re-inspection item mapping, and version are recorded for this implicit relationship edge. The conditional co-occurrence residual is defined by the difference in failure confirmation ratios, giving the co-occurrence strength a statistically reproducible meaning. The minimum sample size, rejection threshold, and version recording limit the impact of sparse stratification and accidental co-occurrence on the relationship edge, while the re-inspection item mapping provides a definite entry point for subsequent verification.
[0012] Furthermore, the comparable domain construction unit also establishes a detection station comparison domain, using the same-direction distribution offset generated by the same detection station for different asset series energy meters within adjacent sorting windows as evidence of station drift. When the station drift evidence covers multiple comparable sample domains and the candidate meter clusters are not concentrated in the same asset series, the transmission of the corresponding implicit relationship edge is stopped, and the early warning output unit outputs a detection station verification identifier. The detection station comparison domain identifies the same-direction offset concentrated in the same station across asset series as abnormal test conditions, enabling the data path to switch between meter defect identification and station verification.
[0013] Furthermore, the distribution activation unit uses the same group quantile boundaries determined by the reference distribution to statistically analyze the proportions of reference samples and current sorting window samples in each quantile interval; the positive difference between the current proportion of adjacent quantile intervals in the risk direction and the reference proportion is weighted according to the quantile interval order and the number of continuous windows to form an explicit offset; the explicit offset is combined with the credibility of the implicit relationship edge, the applicable hierarchical matching degree, and the evidence channel integrity to form an edge propagation quantity, and the edge propagation quantity pointing to the same implicit fault node is accumulated with attenuation to form the node trigger value in the implicit activation vector. After the reference distribution uniformly determines the quantile boundaries, the changes in the proportions of each interval in the current window have a common scale; by only accumulating the positive difference in the risk direction, and combining the continuous window, the credibility of the relationship edge, and the channel integrity, the gradual signal can be preserved and short-term disturbances can be eliminated with attenuation.
[0014] Furthermore, the dual-space density detection unit determines the explicit distance based on the normalized difference of the explicit detection channels jointly effective by the two energy meters to be identified, and determines the implicit distance based on the difference in trigger values and the consistency of trigger directions of the implicit fault nodes jointly activated by the two energy meters to be identified. Density adjacency is established between the two energy meters to be identified only when the explicit distance is not greater than the radius of the explicit neighborhood and the implicit distance is not greater than the radius of the implicit neighborhood, and the candidate meter cluster is expanded through the density adjacency. The explicit distance limits the proximity of observable detection phenomena, while the implicit distance limits the consistency of the potential fault's direction and orientation. Density adjacency is established only after both types of neighborhoods are established simultaneously, which can reduce the merging of normal large clusters and small implicit fault clusters due to the proximity of a single indicator.
[0015] Furthermore, the dual-space density discovery unit determines the local density threshold for each energy meter to be identified within the basic density threshold range based on the proportion of effective evidence channels and the credibility of corresponding implicit relationship edges. The implicit fault node with the highest percentage of samples reaching the node trigger state within the candidate meter cluster is identified as the dominant implicit fault node. The evidence channels supporting the dominant implicit fault node and whose cumulative reach of the channel coverage threshold after being sorted by the number of effective samples within the cluster are identified as dominant evidence channels. When an energy meter to be identified lacks any dominant evidence channel, it is retained as a boundary point to be re-examined rather than as a cluster expansion core point. Local density thresholds respond to evidence availability within the basic threshold range, while dominant evidence channels limit the expansion of cluster boundaries for samples lacking key evidence, thus balancing the utilization of missing samples and the stability of candidate cluster shapes.
[0016] Furthermore, the dual-space density discovery unit identifies the latent fault node with the uniquely largest node trigger value in the latent activation vector as the dominant latent fault node. When the dominant latent fault nodes of two meters to be identified are different or the triggering directions for the same dominant latent fault node are opposite, the density connectivity between the two meters to be identified is severed. When there are two nodes with the largest trigger values, the corresponding meter to be identified is used as a boundary point rather than a cluster expansion core point. Severing the bridging relationship when node directions are inconsistent and prohibiting the corresponding sample from being used as a core point when values are the largest can prevent a transitional sample lacking a clear fault indication from connecting different failure mechanisms.
[0017] Furthermore, the cluster triggering unit determines the median node trigger value of the same latent fault node within the candidate meter cluster as the cluster aggregation trigger value, determines the ratio of the number of valid samples reaching the node trigger threshold to the number of valid samples within the cluster as the node trigger sample proportion, determines the ratio of the number of valid evidence channels supporting different detection mechanisms of the latent fault node to the required number of evidence channels as the evidence channel coverage rate, and determines the ratio of the maximum number of samples of the same asset lineage to the number of valid samples within the cluster as the asset lineage concentration. When the cluster aggregation trigger value, node trigger sample proportion, evidence channel coverage rate, and asset lineage concentration all reach the corresponding thresholds, an early warning triggering result is generated. When no existing latent fault node meets the above thresholds and the candidate meter cluster maintains density connectivity across windows, a temporary latent fault node is established based on the dominant explicit offset combination that maintains the same direction across windows, and a re-inspection output is generated. The median reduces the impact of a single abnormal trigger value on cluster-level judgment. The sample proportion, independent detection mechanism coverage, and asset genealogy concentration respectively constrain group consistency, evidence independence, and defect source. Temporary nodes only generate outputs to be re-inspected, reserving an entry point for unknown defects without directly polluting the formal relational database.
[0018] Furthermore, the re-inspection and re-feedback unit determines the sum of edge transmission amounts generated by the same implicit relation edge within the candidate meter cluster as the cluster-level transmission contribution of the implicit relation edge. Implicit relation edges are selected from high to low according to the cluster-level transmission contribution until a contribution coverage threshold is reached. A targeted re-inspection task is generated based on the re-inspection item mapping recorded by the selected implicit relation edge. During re-inspection confirmation, the confirmation count of the corresponding implicit relation edge in the corresponding applicable layer is increased; during re-inspection non-confirmation, a negation count is increased and the re-inspection conditions are recorded. The credibility of the implicit relation edge is updated based on the confirmation and negation counts, and the transmission of the implicit relation edge in the corresponding applicable layer is stopped when the number of consecutive non-confirmations reaches a freezing threshold. The cluster-level transmission contribution allows re-inspection items to be sorted according to actual evidence contribution, and the mapping of re-inspection items in the relation edge ensures task execution. The layer update and continuous negation freezing mechanism simultaneously retain positive confirmation and negative evidence, reducing the risk of a self-evident loop forming in the relation database.
[0019] To enable the above identification process to be implemented as a data processing step, this invention also proposes an intelligent identification method for latent defects in electricity meters based on fusion density clustering.
[0020] This includes: collecting and sorting full-volume inspection indicators, asset lineage information, and historical re-inspection conclusions according to the electricity meter identification; forming an explicit feature vector from the sorting full-volume inspection indicators; and constructing a comparable sample domain and a reference distribution of explicit inspection features based on the asset lineage information. By controlling the confounding effects caused by asset lineage and inspection stations, we can identify implicit relationship edges between historical explicit feature distribution shifts and re-inspection confirmation faults that have cross-layered co-occurrence residuals, thus forming an implicit relationship library. According to the unified quantile boundary determined by the reference distribution, calculate the change in the interval proportion of the current sorting window in the risk direction, and generate the implicit activation vector of each energy meter to be identified along the implicit relationship edge. Density adjacency is established for the energy meters to be identified that simultaneously meet both explicit and implicit neighborhood conditions to obtain candidate meter clusters. The warning trigger result is generated based on the node trigger sample ratio of candidate meter clusters, the coverage of independent evidence channels, and the concentration of asset lineage. The candidate meter identifier, evidence path, and suggested re-inspection items are output. Execute a targeted re-inspection task generated by the mapping relationship between implicit relation edges and re-inspection items, and feed back the confirmed data and negative data into the implicit relation database respectively.
[0021] This method establishes a complete data chain with unified quantile boundaries, computable conditional co-occurrence residuals, dual spatial neighborhoods, and traceable re-examination tasks, enabling the discovery of candidate populations, explanation of fault mechanisms, and subsequent verification to use the same set of versioned evidence.
[0022] Overall, this invention integrates historical relationship construction, current window activation, dual spatial density connectivity, cluster-level evidence triggering, early warning output, and bidirectional re-injection into a continuous processing procedure. The workstation comparison domain is used to identify abnormal test conditions, while confirmation and rejection counts are used to update the state of relationship edges in the specific applicable layer. Temporary hidden fault nodes are isolated and saved from formal nodes before the re-inspection is completed. Attached Figure Description
[0023] Figure 1 This is a functional connection diagram of an intelligent identification system for latent defects in electricity meters.
[0024] Figure 2 A schematic diagram illustrating the process of constructing the implicit relation database and identifying the current window.
[0025] Figure 3 This is a schematic diagram illustrating the data relationships between explicit fault nodes, implicit relation edges, and implicit activation vectors.
[0026] Figure 4 This diagram illustrates the migration of explicit detection feature distribution and changes in node trigger values.
[0027] Figure 5 This is a schematic diagram illustrating density connectivity constrained by both explicit and implicit neighborhoods.
[0028] Figure 6 This is a schematic diagram of the process for determining the early warning of existing nodes in the candidate meter cluster and outputting the output of nodes awaiting re-inspection for temporary hidden faults.
[0029] Figure 7 A schematic diagram illustrating the deployment of the identification system in the sorting process of dismantled electricity meters.
[0030] Figure 8 This is a closed-loop diagram illustrating the process of feeding targeted re-examination data back into the implicit relation database.
[0031] Figure labeling: 110, Detection data access unit; 120, Comparable domain construction unit; 130, Implicit relation library construction unit; 140, Distributed activation unit; 150, Dual spatial density discovery unit; 160, Cluster triggering unit; 170, Re-inspection and reinjection unit; 180, Early warning output unit; 200, Sorting station; 210, Asset data management terminal; 220, Full-volume inspection equipment group; 230, Identification server; 240, Re-inspection equipment; 250, Verification terminal; 300, Explicit feature vector; 310, Implicit relation library; 320, Implicit relation edge; 330, Implicit activation vector; 340, Comparable sample domain; 350, Candidate meter cluster; 360, Implicit defect early warning; 370, Temporary implicit fault node; 380, Targeted re-inspection task; 390, Confirmed data; 400, Negative data. Detailed Implementation
[0032] The embodiments of the present invention will be further described below with reference to the accompanying drawings. Latent defects refer to defects in which no direct conclusion items are set for routine sorting items, or where a single test value has not yet reached the non-compliance limit, but multiple observable indicators exhibit repeatable linkage in a group of energy meters with a common asset spectrum, and a common failure link can be located through re-inspection. Latent defects belong to the engineering objects to be identified; latent fault nodes belong to data objects in the relational database, used to represent a common failure link that can be confirmed by re-inspection items; the two are not used synonymously.
[0033] The comparable sample domain is a set of samples to be compared that share the same manufacturing origin, specifications, hardware batch, and software version, and are within the same service life range. The testing station control domain organizes samples across asset lineages centered on the same testing station to identify common offsets caused by test conditions. The applicable stratification is a set of asset lineages and testing station conditions whose implicit relationship edges have been verified through historical re-inspection. The applicable stratification matching degree represents the proportion of current meter conditions that match fields in this set. The implicit activation vector is formed by arranging the node trigger values of each implicit fault node according to the node's unique identifier.
[0034] Figure 1 The illustrated implementation includes a detection data access unit 110, a comparable domain construction unit 120, a latent relation database construction unit 130, a distribution activation unit 140, a dual spatial density discovery unit 150, a cluster triggering unit 160, a re-inspection and re-feedback unit 170, and an early warning output unit 180. During initial database construction, the comparable domain construction unit 120 provides historical sample hierarchies to the latent relation database construction unit 130. After the relation database is formed, the latent relation database construction unit 130 feeds back the applicable hierarchies of the relation edge records to the comparable domain construction unit 120 for hierarchical matching in the current window. Figure 1 Connections 130 to 120 in the table represent subsequent feedback relationships. Units can be deployed on the same server, or the data can be shared by the sorting data service and the independent identification service. All data is associated with a unique meter identifier and a sorting task identifier.
[0035] The sorting full-volume inspection indicators received by the inspection data access unit 110 include not only pass and fail status. Continuous value items retain the original measured value, test load point, sampling time, workstation identifier, and validity flag; discrete items such as communication, authentication, and power on / off operations record the number of requests, the number of successes, the response delay interval, the failure code sequence, and the interruption location. Asset genealogy information is used to distinguish manufacturing source, specifications, hardware batch, software version, and service life interval. The evidence channel is the standardized data source for the above inspection fields and their validity status, while the re-inspection items are the specific inspection actions used to verify hidden fault nodes. The two are linked through the re-inspection item mapping in the implicit relationship edge.
[0036] The explicit feature vector 300 consists of values or states that can be directly measured at existing workstations. Continuous detection values within the same comparable sample domain 340 are centered using the median of the reference sample and scaled by the difference between the upper and lower quartiles of the reference sample; when the quartile difference is zero, the channel is not included in distance calculations and is recorded as an invalid channel. Discrete states are converted into occurrence frequency, consecutive failure length, or state transition counts. Explicit detection features represent a single measurable field, and explicit fault nodes represent stable offset patterns formed by that field or multiple fields within the same detection item relative to the reference distribution.
[0037] The comparable domain construction unit 120 first forms a basic stratification based on manufacturing origin, specifications, hardware batch, and software version, which are known to alter the detection distribution, and then further subdivides it according to service duration intervals. The minimum sample size for each basic stratification is no less than five times the basic density threshold required for subsequent density determination; when the sample size is insufficient, adjacent service duration intervals are merged only if the hardware batch and software version are the same, and the median difference of each detection feature in the two merged intervals is required to be no more than the difference of the corresponding reference sample quartiles. The resulting comparable sample domain 340 is used to establish a reference distribution and also to calculate the asset phylogenetic concentration.
[0038] Latent fault nodes in historical data are defined by common failure links that can be located through a set of re-inspection items. Each node record includes the node name, re-inspection confirmation conditions, required re-inspection items, allowed explicit fault nodes pointing to that node, and differences from adjacent nodes. For example, clock backup power supply chain degradation is jointly confirmed by battery load voltage, power-down retention test, and clock retest; digital control chain stability defects are jointly confirmed by multi-round communication, cost control authentication, reset records, and power supply ripple verification. Task scheduling anomalies are only classified as digital control chain stability defects if there is a identical time sequence between reset records and communication or authentication failures; otherwise, they are saved as independent phenomena to be re-inspected.
[0039] Figure 2 The construction process corresponds to the implicit relation database 310. Historical samples are first stratified according to the comparable sample domain 340 and the detection station. For a certain explicit feature shift, the failure confirmation ratio of samples with the shift and the failure confirmation ratio of samples without the shift are calculated in each stratum. The conditional co-occurrence residual is obtained by subtracting the latter from the former. A positive residual indicates that the shift is in the same direction as the failure confirmation, while a zero or negative residual indicates that the stratum does not support the establishment of a positive relation edge. The corresponding implicit relation edge 320 is established only when at least two strata that meet the minimum sample size requirement both obtain positive residuals, and the shift ratio in the unconfirmed samples of each stratum is lower than the rejection threshold.
[0040] Each implicit relationship edge 320 records the starting explicit fault node, ending implicit fault node, offset direction, evidence channel, credibility, applicable layer, re-inspection item mapping, confirmation count, rejection count, final verification window, and version. The re-inspection item mapping includes at least the re-inspection item identifier, applicable testing conditions, and the failure link that can be confirmed or excluded. The starting point can be a single explicit feature or an explicit node jointly defined by the value, response delay, and state transition of the same testing item; when multiple explicit nodes must appear together, the joint coverage condition is recorded in the relationship edge, without creating a comprehensive score name that cannot determine the boundary.
[0041] The minimum sample size and rejection threshold are determined by the historical stability window during database construction. In one implementation, the historical window without corresponding re-inspection faults is used as the control window. The distribution of the proportion of unconfirmed sample shifts is obtained according to the same stratification and the same shift identification process, and the upper quartile of this distribution is taken as the rejection threshold. The total sample size of each stratum is not less than five times the corresponding basic density threshold, and both the shifted group and the non-shifted group have no less than ten samples. If these conditions are not met, the co-occurrence residual is not calculated. Stratified comparison is used to handle mixed inspection stations; the association is only established when a single station is established. No relational edge is established for meter defects; instead, it is retained as a station drift candidate.
[0042] Risk direction is recorded item by item according to the detection characteristics. For characteristics such as low battery voltage that indicate increased risk, the risk direction is a decrease in value; for characteristics such as high communication response delay that indicate increased risk, the risk direction is an increase in value; discrete states are determined according to the state transition order in the re-inspection confirmation sample. Detection characteristics without failure mechanism basis and whose risk direction cannot be determined by historical re-inspection records do not establish a unidirectional implicit relationship edge and can only serve as undirected auxiliary evidence.
[0043] Figure 3 Two representative relationships are shown. The first group is characterized by the migration of battery voltage distribution to lower values and the expansion of daily timing errors in the same direction as explicit nodes. If historical re-examinations show that these two phenomena frequently precede power failures, then an implicit relationship edge 320 pointing to the degradation of the clock backup power supply chain can be formed. The second group is characterized by the increase in the length of consecutive 485 communication failures and the concentrated occurrence of fee control authentication failure codes as explicit nodes. If re-examinations pinpoint instability in the common power supply, reset, or task scheduling link, the relevant edges can point to stability defects in the digital control chain. These edges only express reusable evidence transmission relationships after condition stripping; they do not directly identify 485 communication failures as the physical cause of fee control authentication failures.
[0044] When a new sorting window arrives, the distribution activation unit 140 uses only the reference distribution to determine the quantile boundaries, and then uses the same group boundaries to calculate the sample proportion of the current window, avoiding the problem of taking quantile points according to two different distributions and making the proportion of each interval fixed and equal. Each detection feature pre-determines the risk direction based on the failure mechanism or historical re-inspection records. When the proportion of the current window in the adjacent intervals of the risk direction is higher than the reference proportion, the positive proportion difference is taken; incremental weights are given according to the order of interval distance from the middle of the reference distribution, and then multiplied by the number of windows that show consecutive changes in the same direction to obtain the explicit offset. Here, quantile migration refers to the change in the proportion of the overall probability quality under a unified boundary, and does not indicate that the same energy meter migrates between different windows.
[0045] Meter-level activation determines membership degree based on the quantile interval where each meter is located, its consistency with the risk direction, and the effectiveness of the evidence channel. Membership degree is zero in the middle interval of the reference distribution, increasing by one level for each quantile interval further away from the risk direction, reaching its maximum at the outermost interval. Membership degree is multiplied by explicit offset, relation edge credibility, applicable hierarchical matching degree, and evidence channel integrity to obtain the edge propagation amount of that relation edge to that meter; the edge propagation amounts pointing to the same latent fault node are summed and capped to form the node trigger value. When no further evidence in the same direction appears, the node trigger value decays according to the rule of decreasing the propagation amount corresponding to one membership degree level for each window.
[0046] Figure 4 The upper section uses the risk direction located to the right of the horizontal axis as an example. The solid line represents the previous reference distribution, and the dashed line represents the current distribution after workstation verification or evidence decay, showing the current distribution's decline from the risk side. The lower section shows, in chronological order, how the node entered the observation and activation state during the previous risk direction migration, and how it was deactivated after crossing the exit boundary after the distribution declined. The entry boundary is determined by the lower quartile of the node trigger value in the historical fault confirmation window, and the exit boundary is determined by the upper quartile of the node trigger value in the historical stability window; the entry boundary must be higher than the exit boundary to form a hysteresis interval.
[0047] The dual-space density detection unit 150 calculates the explicit and implicit distances separately. The explicit distance is the average of the normalized absolute differences between the two meters on their shared effective explicit detection channels; the implicit distance is the average of the normalized absolute differences of the node trigger values that jointly activate the implicit fault node. If the dominant nodes are different or the trigger directions are opposite, it is directly judged that the implicit neighborhood condition is not met. The explicit and implicit neighborhood radii are respectively taken as the upper quartiles of the pairwise distance distribution of samples in the same comparable sample domain within the historical stable window, and remain unchanged within a relational database version. Density adjacency is established only when both distances are not greater than the corresponding radii.
[0048] The base density threshold is set within a historical stability window, ensuring that no more than 5% of stable samples can become the core points of independent small clusters. For the current meter, the ratio of the number of valid evidence channels to the number of standard channels in the comparable sample domain is first calculated; as this ratio decreases, the local threshold is reduced by the same proportion, but it must not be lower than two, nor higher than the base density threshold. If the average confidence of the corresponding implicit relationship edge is lower than the median confidence of the relationship edge in the comparable sample domain, no reduction is performed. After the candidate cluster is formed, the dominant implicit fault node is first determined by the proportion of node-triggered samples, and then channels are selected from high to low based on the number of valid samples in the evidence channel until the selected channel covers 80% of the valid samples in the cluster. The selected channel constitutes the dominant evidence channel.
[0049] Bridging points in the explicit space need to be handled separately. The explicit distance between two meters may satisfy the neighborhood condition, but their implicit activation vectors 330 might point to the clock backup power chain and the digital control chain respectively, or have opposite triggering directions for the same node. In this case, the density connectivity between them must be severed. When a meter has two or more parallel largest node trigger values, the unique dominant implicit fault node cannot be determined. This meter can only serve as a boundary point connecting to adjacent core points and cannot continue to absorb other meters. Figure 5 The above connection constraint is represented by taking the intersection of the two types of neighborhood members.
[0050] The cluster triggering unit 160 calculates four indicators for each candidate table cluster 350. The median node trigger value of the same latent fault node in the effective samples within the cluster is the cluster aggregation trigger value; the number of samples reaching the node trigger threshold divided by the number of effective samples within the cluster is the node trigger sample ratio; the number of effective evidence channels supporting the node from different detection mechanisms divided by the number of evidence channels required by the node is the evidence channel coverage rate; and the maximum number of samples of the same asset lineage divided by the number of effective samples within the cluster is the asset lineage concentration. Each threshold is calculated as a candidate value by the historical stability window and the re-inspection confirmation window, and then the boundary value that prevents the stability window from triggering while maintaining the confirmation window is selected is selected and recorded with the relational database version. When all four indicators reach the threshold, the cluster triggering unit 160 generates an early warning trigger result, and the early warning output unit 180 generates a latent defect early warning 360.
[0051] The node trigger threshold is the boundary between the upper quartile of the node trigger value in the historical stable window and the lower quartile of the node trigger value in the re-inspection confirmation window for the same latent fault node. When the two distributions overlap and there is no boundary value, the node only outputs the result to be re-inspected and does not output a formal warning. Evidence channels are grouped according to detection mechanisms such as metering, power supply, time base, communication, authentication, switching, and event logging. Repeated sampling within the same mechanism does not increase the number of independent evidence channels.
[0052] To eliminate group offsets caused by testing equipment, the comparable domain construction unit 120 maintains the testing station comparison domain. If the same station causes multiple meters from multiple manufacturing sources and with multiple specifications to shift in the same direction within adjacent windows, and these anomalies are not concentrated in the common asset lineage, the system identifies this evidence as station drift, stops the transmission of the corresponding relationship edge, and sends a verification flag to the station maintenance end. If the offset exists across multiple stations and is concentrated in the same hardware batch, it will continue to be processed as a latent defect in the meter.
[0053] Figure 6The determination process also retains unknown defect entry points. When candidate meter clusters 350 maintain density connectivity across windows, but no existing latent fault nodes meet the cluster triggering conditions, only samples with at least one continuous explicit offset that do not belong to station drift are retained for visible space verification; when the dominant explicit offset combination maintains the same direction and channel composition in two or more consecutive windows, a temporary latent fault node 370 is established and a re-inspection output is formed. Figure 6 The temporary node box in the image represents the endpoint of the output to be re-inspected. Subsequent targeted re-inspections and recharge paths are determined by... Figure 8 This indicates that temporary nodes do not participate in the formal relationship edge transmission and are only converted into formal nodes after the common failure link is identified through re-inspection; if the failure link is repeatedly identified in multiple rounds, the temporary node remains in a temporary state.
[0054] Figure 7 A deployable engineering workflow is provided. After the removed electricity meter completes its identification reading at sorting station 200, the asset data management terminal 210 returns asset lineage information. The full-scale testing equipment group 220 executes metering, clock, communication, fee control, power supply, and switch items according to the task, and sends the raw results to the identification server 230. The identification server 230 runs each identification unit, generating candidate clusters and evidence paths. When verification is required, the targeted re-inspection task 380 is sent to the re-inspection equipment 240, and the confirmed conclusion is then signed for and returned by the verification terminal 250. This workflow can be embedded into existing sorting lines without requiring changes to the internal structure of the electricity meter.
[0055] The targeted re-inspection task 380 is generated based on the re-inspection item mapping in the implicit relation edges. The re-inspection feedback unit 170 first sums the edge transmission amount of each relation edge in the candidate cluster to obtain the cluster-level transmission contribution amount, and then selects relation edges from high to low contribution amount until the cumulative contribution reaches 80% of all positive contributions; then it reads the re-inspection item mapping of the selected relation edges and deletes duplicate items with the same applicable conditions and the same confirmation object. For the candidate clusters of the clock backup power supply chain, battery load voltage, power failure retention and daily timing re-testing can be arranged; for the candidate clusters of the digital control chain, multi-round communication, fee control authentication, reset record and power supply stability verification can be arranged.
[0056] Figure 8In the process, when a shared failure link is confirmed during a re-inspection, confirmation data 390 increases the confirmation count of the corresponding relationship edge in the current applicable layer by one. If the re-inspection fails to confirm, rejection data 400 increases the rejection count by one, and the re-inspection equipment, conditions, and time are saved simultaneously. The reliability of the relationship edge is the ratio of the confirmation count increasing by one to the sum of the confirmation and rejection counts increasing by two, to avoid a small number of samples causing the reliability to directly reach zero or one. If three consecutive re-inspection tasks fail to confirm the same applicable layer, the relationship edge enters a frozen state in that layer. After new confirmation data arrives, it is first manually reviewed before a decision is made on whether to unfreeze it. Different applicable layers are updated separately to avoid the rejection result of one layer erroneously weakening the already confirmed relationships in other layers.
[0057] To illustrate the calculation process, a set of calculation examples, which do not represent actual identification accuracy, is given. A sorting window contains 480 meters, and one comparable sample domain contains 126 meters. After unified boundary statistics according to the reference distribution, 18 meters are located in the battery voltage risk direction interval, of which 15 also show an increase in the proportion of daily timing error risk direction, and the proportion of dual-channel support is 15 divided by 18. After relational edge propagation, 14 meters simultaneously satisfy the explicit neighborhood, implicit neighborhood, and local boundary conditions, forming a candidate cluster; the other meter lacks a dominant evidence channel and is only used as a boundary point to be re-inspected. This example only illustrates the interval proportion, channel coverage, and boundary point connection relationship. Before project submission, the applicant's actual historical records should replace or supplement the sample source, testing time, equipment calibration status, and re-inspection results.
[0058] Another calculation example illustrates workstation drift detection. When a testing workstation continuously processes samples from three different manufacturing sources, the communication response delay shifts towards the high-risk range in all cases, but the cost control certification results and communication data from other workstations do not change synchronously. Since the shift spans multiple asset families and is concentrated in the same workstation, the system writes it into the testing workstation's comparison domain, suspending the transmission of relational edges pointing to the digital control chain. After completing wiring and time base verification, the new window distribution returns below the exit boundary, the observation status is released with decay, and no meter defect warning is generated. This example also belongs to rule-based calculation and is not considered an actual workstation verification result.
[0059] The current sorting window can be formed by task batch, fixed sample size, or continuous time period. The minimum sample size of the window must not be less than five times the basic density threshold of the corresponding comparable sample domain; the window is immediately truncated when the asset spectrum boundary or the calibration status of the inspection station changes. The reference distribution, neighborhood radius, basic density threshold, entry boundary, exit boundary, rejection threshold, and attenuation rules are all saved with the relational database version. Parameter updates only use historical windows that have completed station drift checks and do not directly incorporate the current window to be identified into the reference distribution.
[0060] The implicit relation database 310 can be implemented using a graph data structure or a relational data table. Regardless of the storage format, it retains the unique identifier of each node, node definition, edge direction, evidence channel, applicable layer, review item mapping, confirmation count, rejection count, credibility, freeze status, last verification time, and version. The warning record references the relation database version, reference distribution version, and parameter version at the time of generation, enabling reviewers to reconstruct the explicit distance, implicit distance, node trigger value, and selected review items at that time. Temporary implicit fault nodes 370 are stored separately from formal nodes.
[0061] When multiple implicit relationship edges point to the same implicit fault node, the transit quantities of each edge are calculated separately and then summed. The summation stops once the node trigger value reaches a preset upper limit, preventing a single high-frequency channel from amplifying indefinitely. Evidence from different detection mechanisms can collectively increase the node trigger value; duplicate fields from the same detection mechanism are counted only once as evidence channel coverage. Forward and reverse transit quantities are stored separately; during cluster-level judgment, directional consistency is checked first, and unfavorable evidence is not deleted during the data access phase.
[0062] The density clustering in this embodiment is implemented using a density connectivity algorithm that supports noisy points and irregular cluster shapes; the specific algorithm name is not limited. The algorithm performs at least the following processing relationships: calculating explicit distances using common effective channels; calculating implicit distances using common activated nodes; establishing density adjacency only when both types of distances simultaneously satisfy the neighborhood radius; core points satisfying local constraints; and performing bridging constraints when the dominant nodes are different, have opposite directions, or have the same maximum trigger value. Ordinary single-space density clustering lacking any of the above processing relationships does not belong to the dual-space density discovery process referred to in this embodiment.
[0063] The early warning output unit 180 should output at least the following: candidate meter identifier set, centralized asset spectrum, dominant latent fault node or temporary node, cluster aggregation trigger value, node trigger sample ratio, evidence channel coverage, explicit detection features supporting the node, trigger window, relational database version, parameter version, and suggested re-inspection items. For workstation drift, the output object is changed to the detection workstation and the affected window; for temporary nodes, the output status is "pending re-inspection" instead of using the formal latent defect early warning identifier.
[0064] The above implementation method illustrates an feasible path: forming versioned relation edges from historical re-examination data, establishing explicit offsets for the current window from unified quantile boundaries, discovering candidate clusters from dual-space neighborhoods, and then triggering cluster-level indicators and targeted re-examination feedback. Those skilled in the art can adjust the number of quantile intervals or the database format according to specific detection projects, but should retain the calculation caliber of conditional co-occurrence residuals, the co-establishment of two types of neighborhoods, restrictions on dominant evidence channels, hierarchical confirmation and rejection updates, and temporary node isolation mechanisms.
Claims
1. A smart identification system for latent defects in electricity meters based on density clustering, characterized in that, include: The detection data access unit is used to collect and sort full-quantity detection indicators, asset lineage information and historical re-inspection conclusions according to the electricity meter identification, and form an explicit feature vector from the sort full-quantity detection indicators; the comparable domain construction unit is used to construct a comparable sample domain based on the asset lineage information, and establish a reference distribution of each explicit detection feature within the comparable sample domain based on historical stable samples. The implicit relationship library construction unit is used to control the confounding effects caused by asset lineage and inspection station, determine the implicit relationship edge with cross-layer co-occurrence residual between historical explicit feature distribution offset and re-inspection confirmation fault, and form an implicit relationship library; the distribution activation unit is used to calculate the interval proportion change of the current sorting window in the risk direction according to the unified quantile boundary determined by the reference distribution, and generate the implicit activation vector of each energy meter to be identified along the implicit relationship edge; The dual-space density discovery unit is used to establish density adjacency between energy meters to be identified that simultaneously satisfy both explicit and implicit neighborhood conditions, thereby obtaining candidate meter clusters. The cluster triggering unit is used to generate early warning triggering results based on the proportion of node triggering samples of candidate meter clusters, the coverage of independent evidence channels, and the concentration of asset lineage. The early warning output unit is used to output candidate meter identifiers, evidence paths, and suggested re-inspection items based on the early warning triggering result; the re-inspection backfeeding unit is used to execute the targeted re-inspection task generated by the mapping relationship between the implicit relationship edge and the re-inspection item, and backfeed the confirmed data and negative data to the implicit relationship database respectively.
2. The intelligent identification system for latent defects in electricity meters based on fusion density clustering according to claim 1, characterized in that, The implicit relationship database construction unit stratifies historical data according to manufacturing source, specifications, hardware batch, software version, service duration range, and testing station. Within each stratum, the failure confirmation ratio in samples showing the same explicit feature shift is subtracted from the failure confirmation ratio in samples that do not show the same explicit feature shift, resulting in the conditional co-occurrence residual pointing to the corresponding implicit fault node. Only when the conditional co-occurrence residual is positive in multiple strata that meet the minimum sample size, and the explicit feature shift ratio in unconfirmed samples does not reach the negation threshold, is a corresponding implicit relationship edge established, and the evidence channel, shift direction, credibility, applicable stratum, re-inspection item mapping, and version are recorded for the implicit relationship edge.
3. The intelligent identification system for latent defects in electricity meters based on fusion density clustering according to claim 2, characterized in that, The comparable domain construction unit also establishes a detection station comparison domain, and uses the same distribution offset of the same detection station to different asset series of electricity meters in adjacent sorting windows as evidence of station drift; when the evidence of station drift covers multiple comparable sample domains and the candidate meter clusters are not concentrated in the same asset series, the transmission of the corresponding implicit relationship edge is stopped, and the detection station verification identifier is output by the early warning output unit.
4. The intelligent identification system for latent defects in electricity meters based on fusion density clustering according to claim 1, characterized in that, The distribution activation unit uses the same group gradation boundaries determined by the reference distribution to statistically analyze the proportion of reference samples and current sorting window samples in each gradation interval; it then calculates the positive difference between the current proportion of adjacent gradation intervals in the risk direction and the reference proportion, weighted according to the gradation interval order and the number of continuous windows to form an explicit offset; the explicit offset is combined with the credibility, applicable hierarchical matching degree, and evidence channel integrity of the implicit relationship edge to form an edge transmission quantity, and the edge transmission quantities pointing to the same implicit fault node are accumulated with attenuation to form the node trigger value in the implicit activation vector.
5. The intelligent identification system for latent defects in electricity meters based on fusion density clustering according to claim 4, characterized in that, The dual-space density discovery unit determines the explicit distance based on the normalized difference of the explicit detection channels that are commonly effective for the two energy meters to be identified, and determines the implicit distance based on the difference in trigger values and the consistency of trigger directions of the implicit fault nodes that are commonly activated by the two energy meters to be identified; only when the explicit distance is not greater than the radius of the explicit neighborhood and the implicit distance is not greater than the radius of the implicit neighborhood, the two energy meters to be identified are established in density adjacency, and the candidate meter cluster is expanded through the density adjacency.
6. The intelligent identification system for latent defects in electricity meters based on fusion density clustering according to claim 5, characterized in that, The dual spatial density discovery unit determines the local density threshold for each energy meter to be identified within the range of the basic density threshold based on the proportion of effective evidence channels and the credibility of the corresponding implicit relation edges. The hidden fault node with the highest percentage of samples reaching the node trigger state within the candidate meter cluster is identified as the dominant hidden fault node, and the evidence channel that supports the dominant hidden fault node and reaches the channel coverage threshold after being sorted by the number of valid samples within the cluster is identified as the dominant evidence channel. When the energy meter to be identified lacks any dominant evidence channel, it is retained as a boundary point to be re-examined rather than as a cluster expansion core point.
7. The intelligent identification system for latent defects in electricity meters based on fusion density clustering according to claim 5, characterized in that, The dual-space density discovery unit identifies the hidden fault node with the largest unique node trigger value in the hidden activation vector as the dominant hidden fault node; when the dominant hidden fault nodes of two energy meters to be identified are different or the triggering directions for the same dominant hidden fault node are opposite, the density connectivity between the two energy meters to be identified is cut off; when there are two nodes with the largest node trigger values, the corresponding energy meter to be identified is used as a boundary point instead of a cluster expansion core point.
8. The intelligent identification system for latent defects in electricity meters based on fusion density clustering according to claim 1, characterized in that, The cluster triggering unit determines the median node trigger value of the same latent fault node within the candidate meter cluster as the cluster aggregation trigger value, determines the ratio of the number of valid samples reaching the node trigger threshold to the number of valid samples within the cluster as the node trigger sample ratio, determines the ratio of the number of valid evidence channels supporting different detection mechanisms of the latent fault node to the required number of evidence channels as the evidence channel coverage rate, and determines the ratio of the maximum number of samples of the same asset lineage to the number of valid samples within the cluster as the asset lineage concentration. When the cluster aggregation trigger value, the proportion of node trigger samples, the evidence channel coverage, and the asset genealogy concentration all reach the corresponding thresholds, an early warning trigger result is generated. When no existing latent fault node meets the above threshold and the candidate meter cluster maintains density connectivity across the window, a temporary latent fault node is established based on the dominant explicit offset combination that maintains the same direction across the window, and a re-inspection output is generated.
9. The intelligent identification system for latent defects in electricity meters based on fusion density clustering according to claim 8, characterized in that, The re-inspection and re-feedback unit determines the sum of edge transmission amounts generated by the same implicit relation edge within the candidate meter cluster as the cluster-level transmission contribution of the implicit relation edge. It selects implicit relation edges from high to low according to the cluster-level transmission contribution until a contribution coverage threshold is reached. It then generates a targeted re-inspection task based on the re-inspection items recorded by the selected implicit relation edges. When the re-inspection is confirmed, the confirmation count of the corresponding implicit relation edge in the corresponding applicable layer is increased. When the re-inspection is not confirmed, a rejection count is increased and the re-inspection conditions are recorded. The credibility of the implicit relation edge is updated based on the confirmation count and rejection count. The transmission of the implicit relation edge in the corresponding applicable layer is stopped when the number of consecutive unconfirmed attempts reaches a freezing threshold.
10. A method for intelligent identification of latent defects in electricity meters based on density clustering, characterized in that, include: The system collects full-scale sorting inspection indicators, asset lineage information, and historical re-inspection conclusions based on the electricity meter identification. An explicit feature vector is formed from the full-scale sorting inspection indicators. A comparable sample domain and a reference distribution of explicit inspection features are constructed based on the asset lineage information. The system controls for confounding effects caused by the asset lineage and inspection stations, and determines implicit relationship edges between historical explicit feature distribution offsets and re-inspection confirmed faults that have cross-layered, co-occurring residuals, forming an implicit relationship library. According to the unified quantile boundary determined by the reference distribution, the current sorting window's interval occupancy in the risk direction is calculated. The changes are compared, and a latent activation vector for each energy meter to be identified is generated along the latent relation edges. Density adjacency is established for the energy meters to be identified that simultaneously satisfy both explicit and implicit neighborhood conditions to obtain candidate meter clusters. An early warning trigger result is formed based on the node trigger sample ratio, independent evidence channel coverage, and asset spectrum concentration of the candidate meter clusters, and the candidate meter identifier, evidence path, and suggested re-inspection items are output. A targeted re-inspection task generated by the mapping relationship between the latent relation edges and the re-inspection items is executed, and the confirmed data and negative data are respectively fed back into the latent relation database.
Citation Information
Patent Citations
Method for fault classification of smart electric meter based on cluster analysis and cloud model
CN105866725A
Classification processing system and method for dismounted intelligent meters
CN110348401A
Electric power system equipment fault early warning method based on fuzzy association rules
CN111401785A
