Intelligent isolation switch full life cycle defect early warning system

Abstract

The present application relates to the technical field of deep learning, in particular to a smart disconnecting switch life cycle defect early warning system. The system comprises a life event construction module, a life stage constraint module, a deep learning time sequence prediction module, a defect evolution constraint module and a hierarchical alarm module; the commissioning acceptance, opening and closing actions, operation inspection, maintenance defect elimination and historical alarm records are converted into life event sequences arranged according to time, the service stage, action load, maintenance recovery and defect correlation identifier are configured, the stage deviation features are generated based on deep learning, and the defect stage results are generated in combination with the evolution correlation between mechanism jamming, contact contact abnormality, conductive loop overheating, insulation support abnormality and opening and closing misplacement. The present application makes the alarm judgment subject to the life stage, time sequence deviation and defect evolution constraint at the same time, can distinguish normal stage fluctuations, maintenance recovery process and recurrence process after defect elimination, and improves the continuity of implicit defect precursor warning and the stability of alarm level.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, specifically to a full life-cycle defect early warning system for intelligent disconnect switches. Background Technology

[0002] Intelligent disconnect switch full lifecycle defect early warning falls under the field of power equipment status alarm technology. Conventional systems typically establish equipment status ledgers based on disconnect switch commissioning and acceptance, opening and closing actions, operation inspections, maintenance and defect elimination, and historical alarm records. Rules for temperature changes, action duration, location status, defect level, and alarm confirmation are set in the monitoring backend. The closest existing implementation involves writing records from different sources into a database by equipment number, location name, and occurrence time. Operating status quantities, inspection conclusions, maintenance content, and alarm results are mapped to discrete states such as normal, warning, and abnormal, respectively. The alarm program then generates alarms according to preset thresholds, manual rules, or abnormal probability ranges. Some systems introduce deep learning classification models, converting inspection descriptions, action process records, maintenance texts, and historical alarm records into feature vectors. Anomaly recognition models are trained using labeled defect samples, and the defect category or anomaly score is output upon the entry of a new record. Alarm processing is typically triggered by a single record exceeding a limit, the appearance of an abnormal label in the same location, or repeated alarms from the same equipment within a short period. The technical focus of this type of solution is on status aggregation, anomaly identification, and alarm triggering. It can connect scattered records to a unified platform, but its judgment objects are still mainly single records, single parts, or combinations of status within a short time window.

[0003] In existing technologies, full lifecycle data typically exists only as historical archives, lacking unified technical constraints across the initial commissioning phase, normal operation phase, operational load accumulation phase, maintenance and recovery phase, and recurrence observation phase. The conventional approach is to store commissioning time, cumulative number of actions, most recent maintenance time, defect number, and alarm closure status in a database, which is then manually reviewed by maintenance personnel for post-event analysis. Automatic alarm mechanisms often employ uniform thresholds, models, or fixed weights. Deep learning models, during training, frequently mix data from different lifecycle phases as samples. Even when the model input includes operating years, number of actions, or maintenance intervals, these are treated as ordinary fields in the calculation, without embedding lifecycle phase boundaries, operational load changes, maintenance and recovery status, and recurrence relationships as constraints into the defect judgment process. Even when time-series modeling is used, it often extracts nearest-neighbor records using fixed-length windows, with no correspondence between the window's start and end points and the equipment's lifecycle phase. For defects such as mechanism jamming, abnormal contact of contacts, overheating of conductive circuits, abnormal insulation support, and incomplete opening and closing of circuit breakers, existing alarm systems mostly set identification rules according to independent defect types, lacking a unified expression of sequential relationships, co-occurrence relationships, location correlations, and recurrence paths, resulting in insufficient connection between defect judgment results and the long-term evolution process of equipment.

[0004] Therefore, the main technical problem with existing intelligent disconnect switch defect early warning systems lies in the failure to integrate lifespan stage, timing deviation, and defect evolution logic into the same constraint chain for alarm judgment. This results in the same state description being assigned similar alarm meanings at different service stages, making it difficult for alarm signals to correspond to the actual defect stage of the disconnect switch. Slight fluctuations in the disconnect switch's operation during the stable commissioning period may be normal convergence after break-in; similar fluctuations during the operational load accumulation period may correspond to early signs of mechanical stagnation; and the temperature rise and fall during the maintenance recovery period may have different meanings than the temperature stability during normal operation. The core issue is not the lack of individual state records, but the failure of the existing alarm chain to jointly define the lifespan stage of the record, the direction of continuous change, and the sequence of defects. If the system only uses static thresholds or isolated deep learning classification results, it will be difficult to identify latent defect precursors that have not yet exceeded the alarm threshold but have already deviated from the normal evolution trajectory of the current lifespan stage, and it will also be difficult to distinguish between the post-maintenance recovery process and the defect recurrence process. The resulting defect stage judgment is prone to break between adjacent records, and the alarm level jumps with the change of a single record, making the whole life cycle early warning lack continuous stage basis and defect evolution basis. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent disconnect switch full life cycle defect early warning system, which can effectively solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] The intelligent disconnect switch full life cycle defect early warning system includes a life event construction module, a life stage constraint module, a deep learning time series prediction module, a defect evolution constraint module, and a hierarchical alarm module;

[0008] The life event construction module constructs the commissioning and acceptance of disconnecting switches, opening and closing actions, operation and inspection, maintenance and troubleshooting, and historical alarm records into a life event sequence arranged by time.

[0009] The life stage constraint module configures service stages, operational loads, maintenance and recovery, and defect association identifiers for life event sequences.

[0010] The deep learning time series prediction module generates stage deviation features based on the lifetime event sequence after configuration identification;

[0011] The defect evolution constraint module generates defect stage results based on stage deviation characteristics and defect evolution correlation;

[0012] The graded alarm module outputs the defect warning level according to the defect stage results.

[0013] Preferably, the life event construction module includes an event normalization processing link, which converts commissioning and acceptance records, opening and closing action records, operation and inspection records, maintenance and defect elimination records and historical alarm records into a unified event format with equipment identifier, location identifier, event time, event source, status description and defect object, and establishes continuous life segments according to event time;

[0014] For state descriptions and defect objects of the same location appearing in adjacent lifetime segments, establish cross-segment reference tags;

[0015] For any recurrence of abnormal records after maintenance and defect elimination, a recurrence marker is established, and cross-segment reference markers and recurrence markers are written into the lifetime event sequence.

[0016] Preferably, the life stage constraint module includes a stage boundary generation and processing link. The stage boundary generation and processing link generates life stage boundaries based on the disconnector switch commissioning time, cumulative number of opening and closing operations, continuous running time, recent maintenance time, and historical intervals of similar defects, and divides the life event sequence into a stable commissioning stage, a normal operation stage, an operational load accumulation stage, a maintenance recovery stage, and a recurrence observation stage.

[0017] The boundaries of different life stages, along with the action load identifier and maintenance recovery identifier, are written into the input sequence of the deep learning time series prediction module, so that the same state description has different sequence position constraints in different life stages.

[0018] Preferably, the defect evolution constraint module includes a defect evolution graph processing link, which takes mechanism jamming, abnormal contact of contacts, overheating of conductive circuit, abnormal insulation support and incomplete opening and closing as defect nodes, takes sequential occurrence relationship, co-occurrence relationship, recurrence relationship after defect elimination and location correlation relationship as node edges, and maps the stage deviation features generated by the deep learning time series prediction module to the defect nodes.

[0019] When multiple defective nodes receive stage deviation features from the same location within the same lifespan stage, the candidate defect stage sequence is determined according to the node edges and then called by the graded alarm module.

[0020] Preferably, the event normalization processing link includes a lifetime event trusted labeling processing method, which configures trusted labels for lifetime events based on the event source, consistency of adjacent records in the same part, connection relationship between records before and after maintenance, and historical alarm closed-loop status.

[0021] When there are lifetime events with different sources and conflicting state descriptions within the same lifetime segment, the conflicting events are preserved and a conflict group identifier is generated, without overwriting other sources with a single source;

[0022] The deep learning temporal prediction module receives trusted tags and conflict group identifiers, uses the conflict group identifiers as sequence masks, and writes the trusted tags as event embedding weights into the stage deviation feature generation process.

[0023] Preferably, the stage boundary generation processing link includes a dynamic stage reassignment processing method. After the maintenance and defect elimination record, recurrence marker, and action load surge record enter the life event sequence, the dynamic stage reassignment processing method reallocates the start and end positions of the maintenance recovery segment and the recurrence observation segment, and saves the life stage boundaries before and after the reallocation in the stage boundary version sequence.

[0024] When the deep learning time series prediction module has multiple stage boundary versions of the same lifetime event, it generates corresponding stage deviation features respectively, and the defect evolution constraint module performs consistency screening on the defect stage results under different stage boundary versions.

[0025] Preferably, the defect evolution map processing link includes a stage-restricted path screening method, which writes the life stage identifier into the availability conditions of the defect node and node edge, and selects different candidate defect stage sequences according to the stable commissioning stage, normal operation stage, action load accumulation stage, maintenance and recovery stage, and recurrence observation stage.

[0026] For node edges that do not match the current lifespan stage, set them as non-participating paths;

[0027] For node edges that match the current lifespan stage and have a recurrence marker after defect elimination, set them as recurrence paths;

[0028] The hierarchical alarm module only receives candidate defect stage sequences after they have been filtered by stage-restricted paths.

[0029] Preferably, the deep learning time series prediction module includes a trusted conflict joint coding network, which encodes the unified event format, trusted tag, conflict group identifier, cross-segment reference tag and post-recurrence tag of lifetime events into an event vector, and introduces a reference direction identifier between adjacent lifetime segments;

[0030] When multiple lifetime events corresponding to a conflict group identifier belong to the same location and the event times fall into the same lifetime segment, the Trusted Conflict Joint Coding Network generates a conflict-preserving vector, does not delete event vectors with opposite abnormal directions, and incorporates the conflict-preserving vector into the stage deviation feature.

[0031] Preferably, the defect evolution constraint module includes a joint verification method for boundary versions and evolution paths. The joint verification method for boundary versions and evolution paths inputs the stage boundary version sequence into the stage restricted path filtering method to obtain the candidate defect stage sequence corresponding to each stage boundary version, and compares the first defect node, recurrence path and alarm level source in the candidate defect stage sequence under different stage boundary versions.

[0032] When different stage boundary versions correspond to the same first defect node and the recurrence path is consistent, a stable defect stage result is formed;

[0033] When the first defect node is inconsistent, the inconsistent node and its corresponding lifetime event are returned to the deep learning time series prediction module for stage deviation feature regeneration.

[0034] Preferably, the graded alarm module includes a stage deviation and defect stage linkage output mode. The stage deviation and defect stage linkage output mode reads the conflict preservation vector, the stable defect stage result, the recurrence path and the stage boundary version sequence, and generates four types of outputs: attention, early warning, alarm and emergency alarm according to the stage deviation feature change direction of the same part in a continuous life segment.

[0035] When the abnormal direction corresponding to the conflict holding vector has not yet formed a stable defect stage result, an observation warning with a conflict group identifier is generated.

[0036] When the stable defect stage result contains a recurrence path and the stage deviation feature continuously points to the same defect node, an alarm output with defect node and life stage identifier is generated.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] 1. This invention constructs a time-ordered lifespan event sequence from disconnector commissioning and acceptance, opening and closing operations, operational inspections, maintenance and troubleshooting, and historical alarm records. It assigns service stage, operational load, maintenance recovery, and defect association identifiers to this lifespan event sequence, transforming the input for alarm judgment from isolated records into a continuous event chain with lifespan stage implications. A deep learning-based time-series prediction module generates stage deviation features based on the configured identifiers in the lifespan event sequence, extracting the deviation direction corresponding to the current lifespan stage from continuous state changes in the same equipment and the same location. A defect evolution constraint module combines stage deviation features with defect evolution association, generating defect stage results based on the sequential occurrence, co-occurrence, recurrence after troubleshooting, and location association relationships among mechanism jamming, abnormal contact, conductive circuit overheating, abnormal insulation support, and incomplete opening and closing. A graded alarm module outputs corresponding warning levels according to the defect stage results, transforming the alarm basis from a single threshold or single classification result into a combined judgment chain of lifespan stage limitation, time-series deviation identification, and defect stage constraint. The lifespan stage limitation distinguishes the same state descriptions in different service stages, the time sequence deviation identification extracts the abnormal direction in continuous events, and the defect stage constraint restricts the alarm results to conform to the technical sequence of disconnector defect development. These three processing steps operate continuously on the same lifespan event sequence. The stage identifier formed in the previous step directly enters the subsequent feature generation, and the subsequent defect stage results are constrained by the defect evolution correlation, avoiding alarms being determined solely by a single anomaly. Therefore, for latent defect precursors that have not yet reached the static alarm conditions but have deviated from the normal trajectory of this lifespan stage, an alarm output consistent with the defect evolution stage can be generated. For state changes at the boundary between the maintenance recovery period and the recurrence observation period, continuous judgment can also be made according to the stage boundary and recurrence relationship, ensuring that adjacent alarm levels have a stable stage basis.

[0039] 2. This invention also limits inconsistencies in records, changes in stage boundaries, and post-maintenance recurrence relationships in the entire lifecycle data through an event normalization processing link, a stage boundary version sequence, a stage-restricted path filtering method, and a trusted conflict joint coding network. The event normalization processing link converts records from different sources into a unified event format with equipment identifier, location identifier, event time, event source, status description, and defect object. It also establishes cross-segment reference markers, post-remediation recurrence markers, trusted markers, and conflict group markers, ensuring that anomalies in the same location within adjacent lifecycle segments remain correlated, and records with conflicting sources are not overwritten by a single source. The dynamic stage reassignment processing method reallocates maintenance recovery segments and recurrence observation segments after maintenance and remediation records, recurrence markers, and records of sudden increases in action load enter the system, while retaining the stage boundary versions before and after reassignment. The stage-restricted path filtering method limits the defect node edges that can participate according to the current lifecycle stage, and compares candidate defect stage sequences under different stage boundary versions using a joint verification method of boundary versions and evolution paths. The trusted conflict joint coding network encodes a unified event format, trusted markers, conflict group identifiers, cross-segment reference markers, and post-recurrence markers into an event vector, and generates a conflict-preserving vector when there are records with opposite abnormal directions. The conflict-preserving vector, stable defect stage results, recurrence paths, and stage boundary version sequences are all fed into the alarm output, ensuring that observation warnings, alarms, and emergency alarms all carry corresponding lifetime stages, defect nodes, or conflict group identifiers. For cases where different stage boundary versions point to the same first defect node and the recurrence paths are consistent, the alarm output is generated along the stable defect stage results; for cases where the first defect node is inconsistent, the relevant lifetime events are returned to regenerate the stage deviation features. Subsequent new records can continue to be updated along the same defect evolution chain, reducing alarm breakpoints caused by record overwriting, stage switching, or loss of recurrence relationships. Attached Figure Description

[0040] Figure 1 This is a flowchart of the overall processing flow of the intelligent disconnector switch full life cycle defect early warning system of the present invention;

[0041] Figure 2 Flowchart for lifetime event normalization and continuous lifetime segment construction of the present invention;

[0042] Figure 3 This is a flowchart of the combined processing of lifetime stage constraints and defect evolution maps in this invention;

[0043] Figure 4 This is a flowchart illustrating the stage boundary version, conflict retention, and hierarchical alarm linkage output of the present invention. Detailed Implementation

[0044] 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, not all, of the embodiments of the present invention. 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.

[0045] Please refer to Figure 1 This embodiment provides an intelligent disconnector switch full lifecycle defect early warning system, including a lifecycle event construction module, a lifecycle stage constraint module, a deep learning time series prediction module, a defect evolution constraint module, and a hierarchical alarm module. The lifecycle event construction module constructs a time-ordered lifecycle event sequence from the disconnector switch's commissioning and acceptance, opening and closing actions, operation inspections, maintenance and defect elimination, and historical alarm records. The lifecycle stage constraint module configures service stages, operational loads, maintenance recovery, and defect association identifiers for the lifecycle event sequence. The deep learning time series prediction module generates stage deviation features based on the configured lifecycle event sequence. The defect evolution constraint module generates defect stage results based on the stage deviation features and defect evolution associations. The graded alarm module outputs the defect warning level according to the defect stage results. In this embodiment, the disconnect switch is the equipment object, the continuous records from commissioning to decommissioning are the processing object, and the alarm signal generation is the output object. It belongs to the data processing and defect warning system corresponding to alarm classification. The life event construction module receives data from the existing operation record library, inspection record library, maintenance record library and alarm record library. It does not change the disconnect switch body structure and does not add hardware components unrelated to the product object. The system performs time-based, stage-based, characteristic and evolution constraint processing on the existing records so that the same state description no longer participates in alarm judgment separately from the service stage, and makes the graded alarm results have the corresponding life stage source and defect evolution source.

[0046] Specifically, the life event construction module performs field extraction, time correction, object attribution, and segment arrangement on the input records. Commissioning and acceptance records are used to form the initial state of the life event sequence; opening and closing action records are used to form the trajectory of action load changes; operation and inspection records are used to form a continuous description of the visible and operational states; maintenance and defect elimination records are used to form time anchors for recovery status and recurrence judgment; and historical alarm records are used to form the basis for the correlation between existing defects and alarm closed loops. When processing records, the life event construction module retains the correspondence between the record source and the original description, and merges records from different sources pointing to the same disconnector, the same location, and adjacent time ranges into event sets within the same life segment. The life event sequence is arranged according to the event occurrence time, record generation time, and defect closed loop time. For maintenance or inspection texts where the occurrence time cannot be directly determined, the record generation time is used as the sorting benchmark, and the time reliability is written into the event attributes. The processed life event sequence can support the subsequent deep learning time series prediction module to uniformly input long-cycle trends and short-cycle disturbances.

[0047] In this embodiment, the life stage constraint module does not use a fixed calendar period as the sole basis for division. Instead, it incorporates commissioning time, continuous operating time, cumulative number of opening and closing operations, recent maintenance time, historical intervals of similar defects, and defect elimination closed-loop status into the life stage determination, forming service stage identifiers, action load identifiers, maintenance recovery identifiers, and defect association identifiers. The service stage identifier is used to express whether the disconnector is in the stable commissioning stage, normal operation stage, action load accumulation stage, maintenance recovery stage, or recurrence observation stage. The action load identifier is used to express the cumulative impact of opening and closing operations on the changes in the mechanism's state. The maintenance recovery identifier is used to express the state recovery window after maintenance is completed. The defect association identifier is used to express the relationship between the current life event and historical defects, defects in the same location, defects of the same type, or abnormalities that reappear after defect elimination. The above identifiers are written into the input vector of the deep learning time series prediction module, so that the model can form different sequence constraints based on the life stage when dealing with the same temperature rise change, the same action time series offset, or the same inspection anomaly description.

[0048] The deep learning time series prediction module adopts an embedding encoding and time series prediction structure oriented towards lifetime event sequences. The input consists of a unified event vector generated by the lifetime event construction module and multi-class identifiers generated by the lifetime stage constraint module. The event vector includes equipment identifier, location identifier, event time, event source, status description, defect object, action load, maintenance and recovery status, and defect association status. Text-based status descriptions are transformed into semantic vectors after word segmentation, entity merging, and defect object mapping. Numerical status descriptions are transformed into numerical vectors after normalization and missing marker completion. Classification-based status descriptions are transformed into category vectors after embedding mapping. The deep learning time series prediction module establishes time interval encoding between adjacent lifetime segments and generates stage deviation features of the current lifetime event relative to the baseline of the current stage using the lifetime stage identifier as a conditional input. These stage deviation features are not directly equivalent to the anomaly score, but include the deviation direction, deviation persistence, deviation location, and deviation stage source.

[0049] The stage deviation feature can be generated according to the following formula:

[0050] ;

[0051] in, This represents the stage deviation feature vector corresponding to the i-th lifetime event. Represents a nonlinear mapping function. This indicates deviation from the feature mapping matrix. Represents a unified event vector. Represents a lifespan stage identifier vector. Represents the defect association identifier vector. This represents the time interval between the current lifetime event and the previous lifetime event. Represents the bias vector, when Take the corresponding code of the maintenance and recovery section and When the encoding corresponding to the recurrence after the deficiency is removed, the same temperature rise recovery description will be mapped to a deviation feature with recovery window and recurrence association. Take the code corresponding to the normal operation segment and When no historical defect is pointed to, the same description will be mapped to the deviation feature of normal operation. The two values ​​correspond to different sources of defect evolution input.

[0052] After receiving the stage deviation features, the defect evolution constraint module maps the stage deviation features to defect nodes such as mechanism jamming, abnormal contact, overheating of conductive circuit, abnormal insulation support, and incomplete opening and closing. Based on the defect evolution correlation, it generates candidate defect stage results. The defect evolution correlation includes sequential occurrence relationship, co-occurrence relationship, post-defect recurrence relationship, and location correlation relationship. The sequential occurrence relationship is used to limit the possible evolution order of a certain type of defect in the life event sequence. The co-occurrence relationship is used to limit the co-occurrence possibility of multiple defect nodes within the same time range. The post-defect recurrence relationship is used to limit the recurrence path corresponding to the abnormality after maintenance record. The location correlation relationship is used to limit the defect transmission relationship between the same location or adjacent functional parts. The defect evolution constraint module does not directly generate defect stages based on isolated abnormality scores. Instead, it puts the stage deviation features into the graph composed of defect nodes and node edges for path screening, so that the defect stage results are consistent with the full life cycle event chain of the disconnecting switch.

[0053] The graded alarm module reads the defect stage results and outputs defect warning levels such as attention, warning, alarm, and emergency alarm. The alarm level is jointly determined by the defect node, defect stage, life stage identifier, deviation persistence state, and recurrence path. When generating alarm output, the graded alarm module retains the equipment identifier, location identifier, life stage identifier, defect node, and event source index, so that the life event sequence position corresponding to the alarm signal can be traced during subsequent review. If the stage deviation feature only shows short-term fluctuations and fails to match a stable defect node, attention or observation warning is output. If the stage deviation feature points to the same defect node within a continuous life segment and is consistent with the defect evolution path, a higher-level alarm is output. If the stage deviation feature appears in the recurrence observation segment and corresponds to the recurrence relationship after defect elimination, an alarm result with recurrence path identifier is output. This embodiment uses a closed processing chain between life event sequence, life stage identifier, deep learning stage deviation feature, defect evolution constraint, and graded alarm output to ensure that latent defect precursors can still be included in the staged warning judgment when they do not exceed the static threshold, and to distinguish the maintenance and recovery process from the defect recurrence process in the alarm logic.

[0054] In a preferred embodiment, reference Figure 2The lifespan event construction module includes an event normalization processing link. This link converts commissioning and acceptance records, circuit breaker operation records, operation and inspection records, maintenance and troubleshooting records, and historical alarm records into a unified event format with equipment identifier, location identifier, event time, event source, status description, and defect object. It then establishes continuous lifespan segments based on event time. For the same location's status description and defect object appearing in adjacent lifespan segments, a cross-segment reference marker is established. For recurring anomaly records after maintenance and troubleshooting, a post-troubleshooting recurrence marker is established. Finally, the cross-segment reference marker and the post-troubleshooting recurrence marker are linked together. The event normalization process first establishes a mapping relationship between equipment name, interval number, phase, disconnector number and location description. It then merges the name differences of the same disconnector in different systems into a unified equipment identifier, merges the descriptions of contacts, transmission mechanisms, post insulators, operating mechanism boxes and conductive connection parts into a unified location identifier, merges infrared temperature measurement, inspection defects, operation ticket action records, maintenance records and alarm records into a unified event source field, and maps terms such as overheating, jamming, incomplete positioning, cracks, loosening and abnormal sounds in text records into a defect object field.

[0055] When establishing continuous lifetime segments, the event normalization processing link determines segment affiliation based on the event time difference between adjacent records, location consistency, defect object consistency, and the credibility of the event source. Records belonging to the same equipment and the same location that are temporally continuous and whose status descriptions point to the same or related defect objects are assigned to the same lifetime segment. If a maintenance and defect elimination record is located between two abnormal segments, the preceding segment is marked as the pre-maintenance abnormal segment, and the following segment is marked as the post-maintenance observation segment. Cross-segment reference markers are used to express the inheritance relationship of location, defect object, and status description between preceding and following segments. Post-defect recurrence markers are used to express the situation where the same or related defect objects reappear in the same location after maintenance. The event normalization processing link does not delete original records, nor does it simply merge multiple records into an average state. Instead, it retains their respective sources and participates in time series modeling with a unified event format to avoid lifetime event sequence breaks due to differences in record format. The unified event format and processing meaning of lifetime events are shown in Table 1 below.

[0056] Table 1. Standardized Event Format and Processing Meaning for Lifetime Events

[0057]

[0058] The fields in Table 1 are used to limit the way records from different sources enter the same processing chain. Device identifier and part identifier solve the object attribution problem, event time and event source solve the sorting and reliable source problem, state description and defect object solve the mapping problem between deep learning input and defect node, cross-segment reference marker and recurrence marker after defect elimination solve the continuity problem between adjacent lifetime segments. In specific implementation, text records are written into the defect object field after being processed by terminology dictionary and defect object mapping rules, numerical records are written into the state description field after unit unification and missing state marking, and classified records are written into the state description field after category encoding. If the same record contains both text and numerical values, both types of expression are retained and the source sub-identifier is set in the unified event vector. This embodiment converts the whole life cycle records into data objects that can be continuously processed by the model through a unified event format, and ensures that the defect evolution chain is not interrupted due to changes in the record source through cross-segment reference and recurrence marker.

[0059] In a preferred embodiment, reference Figure 3 The lifespan stage constraint module includes a stage boundary generation and processing link. This link generates lifespan stage boundaries based on the disconnector switch commissioning time, cumulative opening and closing times, continuous operating time, most recent maintenance time, and historical intervals of similar defects. It also divides the lifespan event sequence into a stable commissioning phase, a normal operating phase, an action load accumulation phase, a maintenance recovery phase, and a recurrence observation phase. Different lifespan stage boundaries, along with action load identifiers and maintenance recovery identifiers, are written into the input sequence of the deep learning time series prediction module. This ensures that the same state description has different sequence position limitations in different lifespan stages. In this embodiment, the stage boundaries are not isolated fields but are sequence constraint objects generated together with the lifespan event sequence. The stable commissioning phase is based on commissioning acceptance records and initial operation records. The normal operating phase is based on continuous operating events with empty or low-association defect identifiers. The action load accumulation phase is based on the dense appearance of opening and closing action records or a continuous increase in cumulative action load. The maintenance recovery phase is based on the continuous observation window formed after the maintenance and defect elimination records are closed. The recurrence observation phase is based on the shortening of historical intervals of similar defects or the appearance of recurrence markers after defect elimination.

[0060] The stage boundary generation processing link calculates the stage attribution value for each event in the lifetime event sequence. The stage attribution value is determined by time location, action load, maintenance status, and historical defect interval. If a lifetime event is close to the commissioning and acceptance record and no maintenance record or recurrence marker appears, the stage attribution tends to the stable commissioning stage. If a lifetime event is located in the long-term stable operation range and the same defect object does not appear continuously, the stage attribution tends to the normal operation stage. If there is a continuous accumulation of opening and closing action records before and after a lifetime event, the stage attribution tends to the action load accumulation stage. If a lifetime event is located within the observation window after maintenance and defect elimination closure, the stage attribution tends to the maintenance recovery stage. If a lifetime event has the same or related relationship with the defect object before defect elimination, the stage attribution tends to the recurrence observation stage. In the deep learning time series prediction input, the stage attribution value is encoded as a stage identifier vector and participates in the event vector concatenation together with the action load identifier and maintenance recovery identifier to avoid the model describing the same state as ordinary anomalies without stage differences.

[0061] The stage attribution value can be calculated using the following formula:

[0062] ;

[0063] in, This represents the attribution value of the i-th lifetime event to the k-th lifetime stage, where k ranges from 1 to 5, corresponding to the commissioning stability stage, normal operation stage, operational load accumulation stage, maintenance and recovery stage, and recurrence observation stage, respectively. This indicates the location of an event relative to its commissioning time. Indicates the workload. This indicates the status of maintenance and recovery. Indicates the historical interval of the same type of defect. , , and The denominator represents the weight coefficient corresponding to the k-th life stage, and the denominator represents the sum of the scores for the five life stages. After the maintenance loop is closed, observe the status variables in the observation window and When pointing to a recent, similar defect, the attribution value of the recurrence observation segment increases. Take the state of continuous accumulation of motion load and When maintenance recovery is not specified, the attribute value of the cumulative action load segment increases. After the stage attribute value is written into the input sequence, the same state description will form different model input combinations depending on the different life stages.

[0064] In a preferred embodiment, the defect evolution constraint module includes a defect evolution graph processing link. This link treats mechanism jamming, abnormal contact, overheating of conductive circuits, abnormal insulation support, and incomplete opening / closing as defect nodes, and uses sequential occurrence relationships, co-occurrence relationships, recurrence relationships after defect elimination, and location association relationships as node edges. It also maps the stage deviation features generated by the deep learning time-series prediction module to the defect nodes. When multiple defect nodes receive stage deviation features from the same location within the same lifespan stage, a candidate defect stage sequence is determined according to the node edges and provided to the graded alarm module. In this embodiment, defect nodes do not exist in isolation by manually enumerating alarm results, but are bound together with the lifespan stage, location identifier, and event source index. The mechanism jamming node can receive action timing offsets; the incomplete opening / closing node can receive position status abnormalities and operation record abnormalities; the abnormal contact node can receive contact location status descriptions and conductive connection abnormality descriptions; the overheating conductive circuit node can receive temperature rise records and related inspection descriptions; and the abnormal insulation support node can receive defect object descriptions of the support location and maintenance text mapping results.

[0065] When generating node edges in the defect evolution graph processing link, the sequential occurrence relationship is used to describe whether a certain defect node may be located before or after another defect node in time. The co-occurrence relationship is used to describe the situation where two defect nodes receive stage deviation features together in the same or adjacent life segments. The recurrence relationship after defect elimination is used to describe the situation where similar or related defect nodes receive stage deviation features again after maintenance and defect elimination records. The location association relationship is used to describe the defect association when the location identifier is the same, adjacent, or functionally related in the same equipment. When the stage deviation feature is mapped to the defect node, candidate nodes are first determined according to the location identifier and defect object. Then, the node reception strength is determined according to the degree of matching between the state description vector and the node feature vector. If multiple defect nodes receive stage deviation features of the same location in the same life segment, a candidate defect stage sequence is formed according to the node edges. The candidate defect stage sequence includes the starting defect node, intermediate associated nodes, recurrence path, and termination alarm node.

[0066] The received signal strength at a defective node can be calculated using the following formula:

[0067] ;

[0068] in, This represents the reception strength of the stage deviation feature of the i-th lifetime event for the j-th defect node. This represents the feature vector of the j-th defect node. Describing the vector norm, This indicates the correlation between the location and the defect node. A higher value is taken when the location of the life event coincides with or is associated with the defect node location. This represents the defect object matching quantity. A higher value is taken when the lifetime event defect object is of the same type as or associated with the defect node. and These represent the weighting coefficients of the location correlation quantity and the defect object matching quantity, respectively. When the stage deviation characteristics of the contact location are similar to the feature vector of the overheated node of the conductive circuit, and both the location correlation quantity and the defect object matching quantity are high, the receiving intensity of the overheated node of the conductive circuit increases. When the action record deviates but the location correlation quantity does not point to the contact or the conductive circuit, the receiving intensity of the relevant node is limited.

[0069] Furthermore, based on the event normalization processing link, the event normalization processing link includes a lifetime event trusted labeling processing method. The lifetime event trusted labeling processing method configures trusted labels for lifetime events based on the event source, the consistency of adjacent records in the same part, the connection relationship between records before and after maintenance, and the historical alarm closed-loop status. When there are lifetime events with different sources and conflicting status descriptions within the same lifetime segment, the conflicting events are retained and conflict group identifiers are generated, without covering other sources with a single source. The deep learning time series prediction module receives the trusted labels and conflict group identifiers, uses the conflict group identifiers as sequence masks, and writes the trusted labels as event embedding weights into the stage deviation feature generation process. In this embodiment, the trusted label does not simply discard records, but converts source reliability, adjacent consistency, maintenance connection, and alarm closed-loop status into event embedding weights, so that different records are retained with different weights in deep learning time series prediction. Records with conflicting sources enter the model through conflict group identifiers and maintain their respective abnormal directions.

[0070] The specific processing method for trustworthy labeling of life events is as follows: A source trustworthiness quantity is generated based on the event source; an adjacent consistency quantity is generated based on the similarity of the state descriptions of the same part in adjacent life segments; a maintenance connection quantity is generated based on the connection relationship of defect objects before and after maintenance and repair records; and an alarm closure quantity is generated based on whether historical alarms are closed-loop and whether they become abnormal again after closure. These quantities are then combined into a trustworthy label. If an inspection record within the same life segment shows a contact abnormality but the operating status quantity does not show a corresponding temperature rise, the system does not delete either record but instead groups the two records into the same conflict group identifier. The deep learning time series prediction module uses the conflict group identifier as a sequence mask during encoding, ensuring that the model maintains the parallel existence of conflict sources within the same life segment. The trustworthy label, as an event embedding weight, enters the stage deviation feature generation process, ensuring that event vectors with stable sources and consistency with adjacent records have a higher influence, while event vectors with conflicting sources but pointing to anomalies are still retained as observation criteria.

[0071] In a preferred embodiment, the trust tag can be calculated according to the following formula:

[0072] ;

[0073] in, This represents the value of the trust tag for the i-th lifetime event. Indicates the credibility of the source of the event. This indicates the consistency of adjacent records in the same location. This indicates the amount of data transfer between records before and after maintenance. This represents the closed-loop status of historical alarms. , , and This represents the corresponding weight coefficient, which can be learned during training. If a similar abnormality occurs again in the same part and with the same defect after a certain maintenance record, this indicates that the error is related to the maintenance record. and The value of will allow the trusted marker to retain recurrence associations. If a single record has a reliable source but is completely disconnected from adjacent lifetime segments, then... This will limit the impact of the event vector on the stage deviation feature. After the trusted label is included in the event embedding weight, the deep learning time series prediction module can distinguish the contribution of each record to the life stage deviation judgment while retaining the conflict record.

[0074] Furthermore, based on the stage boundary generation and processing link, the stage boundary generation and processing link includes a dynamic stage redistribution processing method. After maintenance and defect elimination records, recurrence markers, and action load surge records enter the life event sequence, the dynamic stage redistribution processing method reallocates the start and end positions of the maintenance recovery segment and the recurrence observation segment, and simultaneously saves the life stage boundaries before and after redistribution in the stage boundary version sequence. When the deep learning time series prediction module has multiple stage boundary versions for the same life event, it generates corresponding stage deviation features respectively, and the defect evolution constraint module performs consistency screening on the defect stage results under different stage boundary versions. In this embodiment, the dynamic stage redistribution processing method is used to handle the stage boundary changes caused by the continuous supplementation of full life cycle data. If maintenance and defect elimination records are subsequently added, some life events that were originally classified as normal operation segments may enter the maintenance recovery segment. If recurrence markers appear after defect elimination, some life events after the maintenance recovery segment may enter the recurrence observation segment. If action load surge records appear, the boundaries between the normal operation segment and the action load accumulation segment need to be redistributed.

[0075] The dynamic stage redistribution processing method does not overwrite existing stage boundaries when redistributing stage boundaries. Instead, it writes both the boundaries before and after redistribution into a stage boundary version sequence. The stage boundary version sequence includes a version number, start and end lifetime event indexes, stage type, lifetime event that triggered redistribution, and redistribution reason identifier. After reading multiple stage boundary versions, the deep learning time series prediction module generates stage deviation features for the same lifetime event. The defect evolution constraint module performs defect node mapping and candidate defect stage sequence generation for the stage deviation features corresponding to each stage boundary version. It also performs consistency screening on the first defect node, recurrence path, and alarm level source under different versions. If different versions all point to the same defect node, the defect stage result is retained as a stable result. If different versions point to different defect nodes, the relevant lifetime event is returned to regenerate the stage deviation features, avoiding the direct overwriting of historical alarm basis due to new records.

[0076] Furthermore, based on the defect evolution graph processing link, the defect evolution graph processing link includes a stage-restricted path filtering method. The stage-restricted path filtering method writes the life stage identifier into the availability conditions of defect nodes and node edges, and selects different candidate defect stage sequences according to the stable commissioning stage, normal operation stage, action load accumulation stage, maintenance recovery stage, and recurrence observation stage. For node edges that do not match the current life stage, they are set as non-participating paths. For node edges that match the current life stage and have a recurrence mark after defect elimination, they are set as recurrence paths. The hierarchical alarm module only receives candidate defect stage sequences after stage-restricted path filtering. In this embodiment, the stage-restricted path filtering method directly writes the life stage constraints into the defect evolution graph, instead of performing simple filtering at the alarm output end. The stable commissioning stage allows abnormal paths related to initial assembly, debugging, and break-in to participate in the filtering. The normal operation stage allows long-term operation deviation paths to participate in the filtering. The action load accumulation stage allows paths related to action load, such as mechanism jamming and incomplete opening and closing, to participate in the filtering. The maintenance recovery stage allows restorative fluctuation paths to participate in the filtering. The recurrence observation stage allows recurrence paths after defect elimination to participate in the filtering.

[0077] The stage-restricted path filtering method configures an available stage set, location association conditions, recurrence marker conditions, and direction conditions for each node edge. If the current life event belongs to the maintenance and recovery stage and the stage deviation feature points to the overheated node of the conductive circuit, but the available stage set configured for the node edge does not include the maintenance and recovery stage, then the node edge is set as an unparticipating path. If the current life event belongs to the recurrence observation stage and there is a recurrence marker after defect elimination, and the recurrence marker conditions configured for the node edge are met, then the node edge is set as a recurrence path. The graded alarm module can only receive candidate defect stage sequences after stage-restricted path filtering. Node edges that have not been matched with life stages cannot enter the alarm level generation process. This embodiment limits the connection relationship between defect nodes to the service stage, avoiding the direct interpretation of short-term fluctuations in the normal operation stage as recurrence alarms, and also avoiding the misinterpretation of continuous anomalies in the recurrence observation stage as maintenance and recovery processes.

[0078] In a preferred embodiment, reference Figure 4 The deep learning time series prediction module includes a trusted conflict joint coding network. The trusted conflict joint coding network encodes the unified event format, trusted marker, conflict group identifier, cross-segment reference marker, and post-recurrence marker of lifetime events into an event vector, and introduces a reference direction identifier between adjacent lifetime segments. When multiple lifetime events corresponding to the conflict group identifier belong to the same location and the event time falls into the same lifetime segment, the trusted conflict joint coding network generates a conflict-preserving vector, does not delete event vectors with opposite abnormal directions, and incorporates the conflict-preserving vector into the stage deviation feature. In this embodiment, the trusted conflict joint coding network consists of event field encoding, marker encoding, segment reference encoding, and conflict preservation encoding. Event field encoding processes device identifier, location identifier, event time, event source, status description, and defect object. Marker encoding processes trusted marker, conflict group identifier, cross-segment reference marker, and post-recurrence marker. Segment reference encoding processes the reference direction between adjacent lifetime segments. Conflict preservation encoding processes the set of events with opposite abnormal directions or inconsistent sources within the same lifetime segment.

[0079] When generating event vectors, the Trusted Conflict Joint Encoding Network multiplies the trusted label as a weight coefficient with the event field encoding, converts the conflict group identifier into a mask vector and applies it to the attention calculation within the same lifetime segment, converts cross-segment reference labels into forward reference, backward reference, or bidirectional reference direction encoding, and converts post-defect recurrence labels into recurrence association encoding. If there are conflict events within the same lifetime segment where the running record is normal but the inspection record is abnormal, the network does not overwrite the inspection record with the running record, nor does it overwrite the running record with the inspection record. Instead, it retains the two event vectors separately and calculates the conflict-preserving vector. The conflict-preserving vector expresses the source difference, abnormal direction difference, part consistency, and defect object correlation between conflict events. When generating stage deviation features, both ordinary event vectors and conflict-preserving vectors are received simultaneously, enabling the subsequent defect evolution constraint module to identify whether the deviation originates from a stable and consistent record or from a conflict record that needs to be observed.

[0080] The conflict-preserving vector can be generated as follows:

[0081] ;

[0082] in, This represents the conflict preservation vector corresponding to the c-th conflict group identifier. This represents the set of lifetime events belonging to the c-th conflict group. This represents the number of lifetime events within a conflict group. The first term represents the event representation weighted according to the trust label, and the second term represents the mean of the events within the conflict group. This is true when multiple events within a conflict group are in the same direction. The tendency is towards smaller conflict expressions when events with reliable sources but anomalously opposite directions coexist. Preserve the direction of difference and enter the stage of deviation characteristics.

[0083] In one embodiment, the defect evolution constraint module includes a joint verification method for boundary versions and evolution paths. This method inputs the phase boundary version sequence into the phase restricted path filtering method to obtain the candidate defect phase sequence corresponding to each phase boundary version. It then compares the first defect node, recurrence path, and alarm level source in the candidate defect phase sequences under different phase boundary versions. When different phase boundary versions correspond to the same first defect node and have the same recurrence path, a stable defect phase result is formed. When the first defect node is inconsistent, the inconsistent node and the corresponding lifetime event are returned to the deep learning time series prediction module for phase deviation feature regeneration. This embodiment is used to handle the multi-version judgment problem caused by dynamic phase redistribution. If the addition of maintenance records causes some lifetime events to be transferred from the normal operation phase to the maintenance recovery phase, the system retains both the original phase boundary version and the new phase boundary version, and filters the defect phase sequences under the two versions respectively, so that the alarm output is not directly changed by a single phase redistribution.

[0084] The joint verification method of boundary version and evolution path generates a set of candidate defect stage sequences for each stage boundary version. It extracts the first defect node, the terminating defect node, whether it contains a recurrence path, the source of the alarm level, and the corresponding lifetime event index from the set. If multiple stage boundary versions all use the same defect node as the first defect node and all pass through the same recurrence path or do not contain a recurrence path, a stable defect stage result is formed. If the first defect nodes of different stage boundary versions are inconsistent, it indicates that the stage boundary division has a significant impact on the defect identification result. The system returns the inconsistent nodes and corresponding lifetime events to the deep learning time series prediction module. The deep learning time series prediction module rereads the trust marker, conflict group identifier, stage boundary version sequence, and reference direction identifier, and regenerates the stage deviation features. The regenerated features then enter the defect evolution constraint module for path screening. This embodiment enables historical lifetime events to be re-evaluated without losing the original judgment source after new maintenance records, sudden increase records of action load, or recurrence markers enter the system.

[0085] The consistency of the candidate defect stage sequence can be expressed by the following formula:

[0086] ;

[0087] Where K represents the consistency of the candidate defect stage sequence among stage boundary versions, and M represents the number of stage boundary versions. This represents the first defect node in the candidate defect stage sequence corresponding to the m-th stage boundary version. This indicates the first defect node that appears most frequently in the boundary versions of each stage. This represents the recurrence path identifier of the candidate defect stage sequence corresponding to the m-th stage boundary version. This indicates recurrence path identifiers that appear frequently in the boundary versions of each stage. K represents the condition indicator function, which takes the value 1 when the condition is met and 0 when the condition is not met. When all stage boundary versions point to the same first defect node and have the same recurrence path identifier, K takes a higher value and forms a stable defect stage result. When the first defect node is scattered between stage boundary versions, K decreases and triggers the regeneration of stage deviation features.

[0088] In a preferred embodiment, the graded alarm module includes a stage deviation and defect stage linkage output mode. This mode reads the conflict holding vector, stable defect stage results, recurrence paths, and stage boundary version sequences, and generates four types of outputs—attention, warning, alarm, and emergency alarm—based on the direction of change of stage deviation characteristics of the same location within a continuous lifespan segment. When the abnormal direction corresponding to the conflict holding vector has not yet formed a stable defect stage result, an observation warning with a conflict group identifier is generated. When the stable defect stage result contains a recurrence path and the stage deviation characteristics continuously point to the same defect node, an alarm output with a defect node and lifespan stage identifier is generated. In this embodiment, the graded alarm module binds the alarm level to the lifespan stage, defect node, and deviation continuity, rather than simply outputting based on the abnormal score. Attention-type outputs correspond to situations where short-term deviations or defect node reception strength is insufficient; warning-type outputs correspond to situations where stage deviation characteristics continuously appear but have not yet formed a stable defect stage result; alarm-type outputs correspond to situations where a stable defect stage result has formed and the deviation characteristics point to the same defect node within a continuous lifespan segment; and emergency alarm-type outputs correspond to situations where the stable defect stage result contains a recurrence path and the current lifespan stage is a recurrence observation segment or an action load accumulation segment.

[0089] When generating alarm output, the stage deviation and defect stage linkage output method reads the direction of change of stage deviation features within consecutive life segments of the same location and matches the direction of change with the defect node in the stable defect stage result. If the conflict holding vector shows that the abnormal direction comes from the source conflict record, and different stage boundary versions have not yet formed the same first defect node, then an observation warning with a conflict group identifier is output. The observation warning includes the device identifier, location identifier, conflict group identifier, life segment index, and the defect object that needs to be observed. If the stable defect stage result contains a recurrence path, and the stage deviation features all point to the same defect node within consecutive life segments, then an alarm result with a defect node and life stage identifier is output. The alarm result includes the device identifier, location identifier, life stage, defect node, recurrence path, and source life event index. The hierarchical alarm module continues to update the alarm level along the same life event sequence after subsequent life events enter, without overwriting existing defect stage results due to a single new record.

[0090] In this embodiment, the alarm level mapping is not a hardware parameter configuration, but a logical mapping based on event sequences, defect stage results, and recurrence paths. The hierarchical alarm module maintains an alarm state chain for each part. The alarm state chain includes the most recent stable defect stage result, the most recent conflict group identifier, the most recent recurrence path, the most recent stage boundary version, and the most recent alarm level. If a new lifetime event enters and only changes the conflict retention vector without changing the stable defect stage result, the alarm state chain retains the original defect stage and adds an observation warning. If a new lifetime event causes the stage deviation feature to continuously point to the same defect node, the alarm state chain is updated to the corresponding alarm level. If a new lifetime event causes the stage boundary version and the evolution path joint verification to be inconsistent, the alarm state chain pauses the upgrade and returns to stage deviation feature regeneration. This embodiment can keep the alarm output, lifetime event construction, lifetime stage constraints, deep learning time series prediction, and defect evolution constraints in the same data closed loop.

[0091] In a specific application implementation, after the disconnecting switch is commissioned and accepted, an initial life event is generated. The life event construction module writes the acceptance conclusion, initial action record, and initial inspection record into the life event sequence. The life stage constraint module classifies the corresponding event into the stable commissioning stage. Subsequently, during long-term operation, inspection records, opening and closing action records, and historical alarm records are generated. The event normalization processing link converts records from different sources into a unified event format and establishes continuous life segments according to time. When a certain part experiences continuous changes in state description within an adjacent life segment, a cross-segment reference mark is written. When the same part reappears with a related defect object after maintenance and defect elimination record, a recurrence mark after defect elimination is written. The deep learning time series prediction module generates stage deviation features based on the unified event vector, life stage identifier, trusted mark, and conflict group identifier. The defect evolution constraint module maps the stage deviation features to defect nodes and forms a candidate defect stage sequence through stage-restricted path filtering. The hierarchical alarm module generates corresponding alarm outputs based on the stable defect stage results.

[0092] In this specific application implementation, if the inspection log shows an abnormal description at the same contact location, but the temperature record has not yet formed a persistent abnormality, the event normalization processing link will group the inspection log and the temperature record into the same lifetime segment and form a conflict group identifier. The trusted conflict joint coding network will encode the events within the conflict group and generate conflict holding vectors. When the defect evolution constraint module identifies that the conflict holding vector has not yet formed a stable defect stage result, the hierarchical alarm module will output an observation warning with a conflict group identifier. Subsequently, if the stage deviation features within the continuous lifetime segment all point to contact abnormality or overheating nodes in the conductive circuit, and the stage restricted path filtering method confirms that the current lifetime stage allows relevant node edges to participate in the path, then the boundary version and evolution path joint verification method will form a stable defect stage result. The hierarchical alarm module will output an alarm result with defect node and lifetime stage identifiers. This processing process ensures that the source conflict record is not simply overwritten and that the alarm output is supported by continuous lifetime events.

[0093] In another specific application implementation, after a disconnecting switch enters the maintenance recovery phase following a closed-loop maintenance and defect elimination process, a subsequent abnormality record of the same defective object in the same location reappears. The event normalization processing link establishes a recurrence marker after defect elimination. The dynamic phase redistribution processing method reclassifies some life events that originally belonged to the normal operation phase into the recurrence observation phase, and saves the life phase boundaries before and after redistribution as a phase boundary version sequence. The deep learning time series prediction module generates phase deviation features under different phase boundary versions. The defect evolution constraint module inputs the phase deviation features of different versions into the phase restricted path filtering method. If all versions point to the same first defect node and contain a consistent recurrence path, a stable defect phase result is formed. The graded alarm module outputs the corresponding alarm result based on the recurrence path and the direction of continuous deviation. If each version points to a different first defect node, the alarm is not directly upgraded, but the inconsistent node and the corresponding life event are returned to the deep learning time series prediction module for regeneration.

[0094] In this embodiment, data processing can be implemented in the software environment of an existing monitoring platform, operation and maintenance data platform, or alarm processing platform. The life event construction module, life stage constraint module, deep learning time series prediction module, defect evolution constraint module, and hierarchical alarm module can all be deployed as software functional units within the same data processing service, or they can be deployed as multiple service processes with data interfaces. The data interface transmits life event sequences, stage boundary version sequences, stage deviation features, candidate defect stage sequences, and alarm output records. It does not require changes to the mechanical or electrical structure of the disconnector switch. During the training phase, historical life event sequences and historical alarm closed-loop records can be used to form samples. During the operation phase, new life events can be used to continuously update stage deviation features and defect stage results. If the defect labels in the historical samples are incomplete, maintenance and defect elimination records and alarm closed-loop status can be used to weakly label the defect objects, and the impact of weakly labeled samples on the model output can be limited by the defect evolution map.

[0095] The training process in this embodiment can use historical lifetime event sequences as input, and historical defect objects, alarm closed-loop states, and post-repair recurrence markers as supervisory information. The deep learning time series prediction module learns the mapping relationship between lifetime event vectors and stage deviation features, and the defect evolution constraint module learns the mapping relationship between stage deviation features and defect node reception strength. During training, constraints are applied to candidate defect stage sequences that do not conform to the stage-limited path, so that the model output does not deviate from the lifetime stage and defect evolution map. When a new lifetime event is added during operation, the system first performs a unified event format conversion, then performs lifetime stage attribution calculation, then generates stage deviation features and candidate defect stage sequences, and finally the hierarchical alarm module outputs alarm records. This training and operation processing logic makes the model output subject to the lifetime stage boundary, event credibility marker, conflict group identifier, recurrence path, and defect evolution map.

[0096] In each implementation, the lifetime event construction module addresses the problem of continuous input of records from different sources; the lifetime stage constraint module addresses the problem of different technical meanings of the same state description at different service stages; the deep learning time series prediction module addresses the problem of stage deviation identification when long-term slow degradation and short-term disturbances coexist; the defect evolution constraint module addresses the problem of lack of evolution order constraints between defect nodes; and the hierarchical alarm module addresses the problem of lack of binding between alarm levels and defect stages, lifetime stages, and recurrence paths. The event normalization processing link, trusted label processing method, dynamic stage redistribution processing method, stage-restricted path filtering method, trusted conflict joint coding network, boundary version and evolution path joint verification method, and stage deviation and defect stage linkage output method respectively undertake the processing tasks of data continuity, source conflict retention, stage boundary update, defect path limitation, conflict event encoding, multi-version consistency verification, and alarm status chain output in the above main processing chain.

[0097] Compared to alarm processing methods that rely solely on static thresholds or isolated deep learning classification results, this embodiment organizes commissioning and acceptance, circuit breaker actions, operational inspections, maintenance and troubleshooting, and historical alarm records into a time-ordered sequence of life events. Life stage constraints are written into the model input through service phases, operational loads, maintenance recovery, and defect association identifiers. Furthermore, the system incorporates the correlation between mechanism jamming, abnormal contact, conductive circuit overheating, insulation support abnormalities, and incomplete circuit breaker operation into the defect stage generation process using a defect evolution graph. This ensures that alarm outputs correspond to stage deviation characteristics and defect stage results. In cases of record conflicts, post-maintenance recurrence, and stage boundary redistribution, the system retains the conflict group identifier, stage boundary version sequence, and recurrence path, and carries the corresponding identifier in the alarm output. This allows latent defect precursors, maintenance recovery processes, and defect recurrence processes to be distinguished within the same full life cycle early warning chain.

Claims

1. A full life-cycle defect early warning system for intelligent disconnect switches, characterized in that, It includes a lifespan event construction module, a lifespan stage constraint module, a deep learning time series prediction module, a defect evolution constraint module, and a hierarchical alarm module; The life event construction module constructs the commissioning and acceptance of disconnecting switches, opening and closing actions, operation and inspection, maintenance and troubleshooting, and historical alarm records into a life event sequence arranged by time. The life stage constraint module configures service stages, operational loads, maintenance and recovery, and defect association identifiers for life event sequences. The deep learning time series prediction module generates stage deviation features based on the lifetime event sequence after configuration identification; The defect evolution constraint module generates defect stage results based on stage deviation characteristics and defect evolution correlation; The graded alarm module outputs the defect warning level according to the defect stage results.

2. The intelligent disconnector switch full life cycle defect early warning system according to claim 1, characterized in that, The life event construction module includes an event normalization processing link. The event normalization processing link converts commissioning and acceptance records, opening and closing action records, operation and inspection records, maintenance and defect elimination records and historical alarm records into a unified event format with equipment identifier, location identifier, event time, event source, status description and defect object, and establishes continuous life segments according to event time. For state descriptions and defect objects of the same location appearing in adjacent lifetime segments, establish cross-segment reference tags; For any recurrence of abnormal records after maintenance and defect elimination, a recurrence marker is established, and cross-segment reference markers and recurrence markers are written into the lifetime event sequence.

3. The intelligent disconnector switch full life cycle defect early warning system according to claim 1, characterized in that, The life stage constraint module includes a stage boundary generation and processing link. The stage boundary generation and processing link generates life stage boundaries based on the disconnector switch commissioning time, cumulative number of opening and closing operations, continuous running time, recent maintenance time, and historical intervals of similar defects. It also divides the life event sequence into a stable commissioning stage, a normal operation stage, an action load accumulation stage, a maintenance recovery stage, and a recurrence observation stage. The boundaries of different life stages, along with the action load identifier and maintenance recovery identifier, are written into the input sequence of the deep learning time series prediction module, so that the same state description has different sequence position constraints in different life stages.

4. The intelligent disconnector switch full life cycle defect early warning system according to claim 1, characterized in that, The defect evolution constraint module includes a defect evolution graph processing link. The defect evolution graph processing link takes mechanism jamming, abnormal contact of contacts, overheating of conductive circuit, abnormal insulation support and incomplete opening and closing of the switch as defect nodes, takes sequential occurrence relationship, co-occurrence relationship, recurrence relationship after defect elimination and location correlation relationship as node edges, and maps the stage deviation features generated by the deep learning time series prediction module to the defect nodes. When multiple defective nodes receive stage deviation features from the same location within the same lifespan stage, a candidate defect stage sequence is determined according to the node edges and called by the graded alarm module.

5. The intelligent disconnector switch full life cycle defect early warning system according to claim 2, characterized in that, The event normalization processing link includes a lifetime event trusted labeling processing method, which configures trusted labels for lifetime events based on the event source, consistency of adjacent records in the same part, connection relationship between records before and after maintenance, and historical alarm closed-loop status. When there are lifetime events with different sources and conflicting state descriptions within the same lifetime segment, the conflicting events are preserved and a conflict group identifier is generated, without overwriting other sources with a single source; The deep learning temporal prediction module receives the trusted marker and the conflict group identifier, uses the conflict group identifier as a sequence mask, and writes the trusted marker as an event embedding weight into the stage deviation feature generation process.

6. The intelligent disconnector switch full life cycle defect early warning system according to claim 3, characterized in that, The stage boundary generation and processing link includes a dynamic stage reassignment processing method. After the maintenance and defect elimination record, recurrence marker, and action load surge record enter the life event sequence, the dynamic stage reassignment processing method reassigns the start and end positions of the maintenance recovery segment and the recurrence observation segment, and saves the life stage boundary before and after the reassignment in the stage boundary version sequence. When the deep learning time series prediction module has multiple stage boundary versions of the same lifetime event, it generates corresponding stage deviation features respectively, and the defect evolution constraint module performs consistency screening on the defect stage results under different stage boundary versions.

7. The intelligent disconnector switch full life cycle defect early warning system according to claim 4, characterized in that, The defect evolution map processing link includes a stage-restricted path screening method. The stage-restricted path screening method writes the life stage identifier into the availability conditions of the defect node and node edge, and selects different candidate defect stage sequences according to the stable operation stage, normal operation stage, action load accumulation stage, maintenance and recovery stage, and recurrence observation stage. For node edges that do not match the current lifespan stage, set them as non-participating paths; For node edges that match the current lifespan stage and have a recurrence marker after defect elimination, set them as recurrence paths; The hierarchical alarm module only receives candidate defect stage sequences after they have been filtered by stage-restricted paths.

8. The intelligent disconnector switch full life cycle defect early warning system according to claim 5, characterized in that, The deep learning time series prediction module includes a trusted conflict joint coding network, which encodes the unified event format, trusted tag, conflict group identifier, cross-segment reference tag and post-recurrence tag of lifetime events into an event vector, and introduces a reference direction identifier between adjacent lifetime segments; When multiple lifetime events corresponding to the conflict group identifier belong to the same location and the event time falls into the same lifetime segment, the trusted conflict joint coding network generates a conflict hold vector, does not delete event vectors with opposite abnormal directions, and incorporates the conflict hold vector into the stage deviation feature.

9. The intelligent disconnector switch full life cycle defect early warning system according to claim 7, characterized in that, The defect evolution constraint module includes a joint verification method for boundary versions and evolution paths. The joint verification method for boundary versions and evolution paths inputs the stage boundary version sequence into the stage restricted path filtering method to obtain the candidate defect stage sequence corresponding to each stage boundary version, and compares the first defect node, recurrence path and alarm level source in the candidate defect stage sequence under different stage boundary versions. When different stage boundary versions correspond to the same first defect node and the recurrence path is consistent, a stable defect stage result is formed; When the first defect node is inconsistent, the inconsistent node and its corresponding lifetime event are returned to the deep learning time series prediction module for stage deviation feature regeneration.

10. The intelligent disconnector switch full life cycle defect early warning system according to claim 9, characterized in that, The graded alarm module includes a stage deviation and defect stage linkage output mode. The stage deviation and defect stage linkage output mode reads the conflict preservation vector, stable defect stage result, recurrence path and stage boundary version sequence, and generates four types of outputs: attention, early warning, alarm and emergency alarm according to the stage deviation feature change direction of the same part in continuous life segment. When the abnormal direction corresponding to the conflict holding vector has not yet formed a stable defect stage result, an observation warning with a conflict group identifier is generated. When the stable defect stage result contains a recurrence path and the stage deviation feature continuously points to the same defect node, an alarm output with the defect node and life stage identifier is generated.

Images