Looped network box multi-source data fusion fault early warning method
Patent Information
- Application Number
- CN202611536782.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-30
- Publication Date
- 2026-11-03
AI Technical Summary
Existing ring main unit fault early warning schemes rely on single-point thresholds to identify anomalies, which cannot effectively distinguish between normal changes and anomalies under different operating conditions, leading to false triggering or missing early warnings. Furthermore, the sampling period and data gaps between multiple data sources result in inaccurate alarm results.
By acquiring multi-source state sequences and operating condition records of each monitoring point in the ring network box, operating condition codes are generated, condition baselines are constructed, state events are mapped and a fault evidence chain is formed, and graded early warning results are generated by fusing evidence based on the fault rule diagram. Data sources with different sampling frequencies are processed using cross-source association indexes and electrical connection relationships.
It effectively suppressed baseline jumps caused by load and environmental changes, reduced false triggers and early anomaly detection, ensured the consistency and accuracy of alarm levels, and avoided false mutations and duplicate entries caused by data gaps.
Smart Images
Figure CN123031192A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of power distribution equipment monitoring data processing, multi-source information fusion and fault early warning technology, specifically a method for fault early warning of ring network boxes using multi-source data fusion. Background Technology
[0002] Existing ring main unit fault early warning schemes typically deploy acquisition modules at cable joints, busbar connections, gas chambers, and the enclosure environment to collect data on temperature, partial discharge, sulfur hexafluoride gas status, current, voltage, and switch status. Local acquisition terminals obtain monitoring values according to their respective sampling cycles, which are then uploaded to the monitoring platform after protocol conversion, time stamping, filtering, missing value imputation, and engineering quantity conversion. The monitoring platform sets fixed upper and lower limits or historical statistical intervals for different status quantities and generates alarms for overheating, insulation abnormality, gas abnormality, or mechanical abnormality based on duration, hysteresis range, and logical combination relationships. Some schemes also calculate a comprehensive status score based on the number of alarm status quantities, the extent of exceeding limits, and the duration, classifying the equipment status into normal, warning, abnormal, and severe levels, and then presenting the alarm results through trend curves, event lists, and message pushes. This processing chain relies on single-point thresholds to identify abnormalities and then uses pre-compiled rules to summarize the judgment results of multiple monitoring points. The data sources are mainly combined at the alarm result layer, without corresponding modeling of the normal change trajectory of each status quantity under different operating conditions.
[0003] Changes in the load level and environmental conditions of the ring main unit can cause the normal distribution of temperature, current, and insulation-related quantities to shift with the operation phase. Fixed thresholds still compare real-time values with a single boundary, which can easily identify continuous deviations caused by the transition from normal operating conditions as faults, or keep anomalies in the early stages of evolution but not yet crossing the absolute boundary within the normal range. The sampling period, uploading interval, and data gaps of different data sources are different, and the abnormal segments formed by the same physical process are not consistent in terms of time boundaries and duration intervals. Directly filling, resampling, or summarizing alarm results will change the sequence of events. Pre-programmed logical rules can only determine whether several conditions are met simultaneously, and cannot distinguish between independent concurrent anomalies and continuous anomalies with a forward-backward transmission relationship based on electrical connection relationships. Repeatedly uploaded data may also be repeatedly counted in multiple rules, ultimately making the graded status reflect more the number of limits exceeded than the evolutionary position of the fault evidence chain, resulting in alarm level jumps, false triggers, or missing early warnings. Summary of the Invention
[0004] The purpose of this invention is to provide a method for fault early warning by multi-source data fusion in ring network boxes, which can solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0006] A method for fault early warning by multi-source data fusion of ring network boxes includes: acquiring multi-source state sequences of each monitoring point of the same ring network box and operating condition records generated synchronously with the multi-source state sequences, and generating operating condition codes for each collection period.
[0007] Select data segments from historical normal data that match the current operating condition code, construct conditional baselines for each data source, and encode continuous deviations of real-time state quantities relative to the conditional baselines as state events;
[0008] The state events are mapped to the electrical connection relationships of the ring network box, and the state events with electrical reachability are connected in chronological order to form a fault evidence chain.
[0009] Based on the extension position of the fault evidence chain in the preset fault rule diagram, evidence from various data sources is integrated to generate a graded fault early warning result;
[0010] The construction of the conditional baseline includes: forming the current operating condition trajectory by arranging the operating condition codes of the current acquisition period and its adjacent acquisition periods in the order of occurrence; and retrieving historical operating condition trajectories from historical normal data whose operating condition change direction is consistent with the current operating condition trajectory.
[0011] The historical state quantities are aligned according to the relative positions of the corresponding working condition codes in the trajectory to form the normal state distribution of each data source.
[0012] Establish a conditional baseline interval based on the center trajectory and the two side distribution boundaries of the normal state distribution, and register the unmatched segments between the historical working condition trajectory and the current working condition trajectory as baseline vacancy segments.
[0013] Preferably, the processing of the multi-source state sequence includes: retaining the original acquisition sequence number of each state quantity according to the data source identifier, and converting the abnormal change segment in the continuous quantity sequence and the state switching segment in the discrete quantity sequence into the in-source event interval respectively;
[0014] Configure component association identifiers for the event intervals within the source, and write the dimensionally normalized change direction and duration into the corresponding event intervals;
[0015] A cross-source association index is established based on the time overlap relationship between different event intervals, so that data sources with different sampling frequencies maintain their original collection sequences, and a unified status event record is formed through the cross-source association index.
[0016] Preferably, the fault rule diagram includes event nodes corresponding to the ring main unit components and directed rule edges connecting the event nodes, wherein the directed rule edges are jointly defined by the electrical reachability relationship between components and the fault state transition order;
[0017] Write the state events into the corresponding event nodes according to the component association identifiers, and activate the directed rule edges that satisfy the state transition order according to the order of the event intervals.
[0018] Candidate fault paths are formed by connecting consecutive event nodes along the activated directed rule edges, and state events that do not satisfy the electrical reachability relationship or whose event order is reversed are retained as isolated evidence records.
[0019] Preferably, a trusted data record is established for each state event, and the trusted data record consists of a continuous interval of the original collection sequence number and the corresponding event interval in the cross-source association index;
[0020] When the original acquisition sequence number is interrupted, the normal state range of the interruption location is determined based on the conditional baseline interval, and the candidate occurrence interval of the missing event is limited by the event interval of other data sources with electrical correlation.
[0021] The candidate occurrence interval is written as an interval-type state event into the cross-source association index, and the data trust record is revised according to the connection relationship between the state quantity collected in subsequent recovery and the candidate occurrence interval.
[0022] Preferably, when the current working condition trajectory enters an adjacent working condition area from a matched working condition area, the pre-migration condition baseline and candidate condition baseline are obtained respectively, and the data source that has not formed a state event within the migration interval is selected as a common reference source.
[0023] The candidate conditional baseline is corrected based on the difference in the center trajectory of the common reference source in the two conditional baselines, so that the corrected candidate conditional baseline takes over the conditional baseline before migration at the working condition migration boundary.
[0024] The candidate condition baselines that have been completed are written into the current acquisition period, and the working condition migration intervals where a common reference source cannot be obtained are kept as the baseline gap segments.
[0025] Preferably, the generation of the state event includes: calculating the directed deviation of the real-time state quantity according to the upper and lower distribution boundaries of the conditional baseline interval, and accumulating the directed deviation within the continuous deviation interval according to the original acquisition sequence number;
[0026] The candidate occurrence intervals in the reliable data records are converted into uncertain intervals of deviation, and then merged with the continuous deviation intervals.
[0027] The event start boundary is determined based on the cumulative deviation distribution formed by the historical normal data under the same working condition trajectory. The event termination boundary is determined by restoring the position to the condition baseline interval along the directional deviation amount, thus forming a state event with a direction identifier.
[0028] Preferably, a propagation interval is established for each directed rule edge in the fault rule graph, formed by the difference between historical matching event intervals;
[0029] When the same state event is associated with multiple candidate fault paths, the state event is temporarily stored as a shared predecessor of the candidate fault paths, and a successor event retrieval interval is generated according to the corresponding propagation interval.
[0030] The occurrence interval and direction identifier of subsequent state events are matched with the retrieval interval of each subsequent event, and the shared predecessor is included in the candidate fault path where the matching relationship is established.
[0031] A single evidence identifier is retained for shared predecessors that still match multiple paths, and the event evidence of the shared predecessors is prohibited from being repeatedly accumulated in different candidate failure paths.
[0032] Preferably, a counterfactual check is performed on the candidate fault paths that have completed the shared predecessor attribution, the earliest time position of the state event in the candidate fault path is selected, the corresponding real-time state quantity is replaced with the conditional baseline interval under the same working condition code, and the state event records of other data sources are kept unchanged.
[0033] According to the fault rule graph, the replaced event nodes and the directed rule edges are rematched, and the successor events that can still be connected by other predecessors are transferred to the corresponding candidate fault path. The successor events that have lost their continuous predecessors are disconnected from the path association, and the fault evidence chain is revised according to the newly formed path connection relationship.
[0034] Preferably, a hierarchical state sequence corresponding to the evolution stage of the fault evidence chain is established. Event nodes that form directed deviation state events but have not yet been connected to directed rule edges are registered as source-end abnormal states. Paths that pass the counterfactual verification and are connected to the subsequent event retrieval interval are registered as propagation warning states. Paths that allow subsequent state events to enter the subsequent event retrieval interval are registered as target alarm states.
[0035] Perform hierarchical state migration according to the connection direction of the fault evidence chain, and after the state event is restored, sequentially cancel the corresponding state according to the reverse path to form a fault warning record containing fault path identifier and current hierarchical state.
[0036] Preferably, the generation of the operating condition code includes: establishing ordered intervals for load status and environmental status based on historical stable operation records; mapping the operating condition records of the current acquisition period to the corresponding ordered intervals; combining the interval numbers into the current area code; reading the area codes of adjacent acquisition periods; comparing the relative positions of the current area code and the previous area code in each ordered interval; generating a migration direction code; combining the current area code and the migration direction code into the operating condition code; and retaining the corresponding direction codes when the load status and environmental status change in different directions within the same acquisition period.
[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0038] 1. By associating multi-source state sequences with synchronous operating condition records, historical normal segments are retrieved using continuous operating condition coded trajectories. A conditional baseline is constructed for each data source to migrate with the operating conditions, transforming the judgment object of real-time state quantities from absolute measured values to directed deviations relative to the normal distribution of the current operating conditions. The conditional baseline corrects the center trajectory difference between adjacent operating condition regions using data sources that have not formed state events, and is continuously connected at the migration boundary, which can suppress baseline jumps caused by load changes or environmental changes. Continuous deviations are accumulated according to the original acquisition sequence number to form state events with time intervals and direction identifiers. Data sources with different sampling frequencies participate in fusion through cross-source association indexes, without the need to force all monitoring values to be resampled to the same time point. State events are then written into the fault rule graph according to component association identifiers, electrical reachability relationships, and event sequence, so that multi-source evidence is aggregated along paths that conform to the state transition order. The hierarchical results correspond to the source formation, link propagation, and target arrival stages of the fault evidence chain, thereby reducing false triggers caused by operating condition changes and early anomaly omissions that have not reached the absolute threshold.
[0039] 2. Reliable data records retain continuous intervals of the original acquisition sequence numbers and the correspondence between cross-source related events. When acquisition is interrupted, definite values are not directly generated. Instead, the conditional baseline and electrical correlation data are used to limit the candidate occurrence intervals of missing events. After acquisition is resumed, the candidate occurrence intervals are revised so that the data gaps enter the subsequent judgment as uncertain intervals, avoiding false abrupt changes caused by interpolation. The start and end boundaries of state events are determined by the cumulative deviation distribution under the same operating conditions. A single instantaneous fluctuation will not immediately change the event state. Continuous deviations accumulate along the acquisition sequence number and retain directional information. When the same event supports multiple candidate fault paths at the same time, the event evidence is temporarily stored as a shared predecessor. The path assignment is completed after the subsequent event matches the propagation interval. The single evidence identifier restricts its repeated accumulation in multiple paths. Counterfactual verification replaces the earliest event of the candidate path with the conditional baseline and reconnects the fault rule graph. Paths formed by accidental concurrency can be split. The graded status is canceled in reverse order when the event is resumed, so that the alarm record maintains the consistency of event boundaries, path assignment, and grade migration under the conditions of data loss, repeated uploading, and short-term disturbances. Attached Figure Description
[0040] Figure 1 A flowchart of the overall process for fault early warning based on multi-source data fusion in ring network boxes;
[0041] Figure 2 Boundary discrimination diagram;
[0042] Figure 3 This is a continuous cumulative graph;
[0043] Figure 4 Flowchart for fault evidence chain verification and early warning level migration;
[0044] Figure 5 This is a diagram showing the comparison of performance metrics for candidate paths. Detailed Implementation
[0045] In one embodiment, please refer to Figure 1This paper provides a method for multi-source data fusion fault early warning for ring main units (RNBs), including: acquiring multi-source state sequences of each monitoring point in the same RNB and operating condition records synchronously formed with the multi-source state sequences, generating operating condition codes for each acquisition period; selecting data segments matching the current operating condition codes from historical normal data, constructing condition baselines for each data source, and encoding continuous deviations of real-time state quantities relative to the condition baselines as state events; mapping the state events to the electrical connection relationships of the RNB, connecting state events with electrical reachability in chronological order to form a fault evidence chain; fusing evidence from each data source based on the extension position of the fault evidence chain in a preset fault rule diagram to generate a graded fault early warning result; the entire process uses the acquisition period as the data organization boundary, the operating condition code as the association key between historical normal data and real-time state data, the state event as the unified expression object for different data sources participating in fusion, and the fault evidence chain as the direct basis for graded state migration, so that the original measured values, operating condition baselines, event intervals, rule paths, and early warning records are continuously transmitted according to the determined data reference relationships.
[0046] In this embodiment, the multi-source state sequence is formed by each monitoring point in the ring network box according to its own acquisition cycle. Continuous quantities save the measured value, acquisition time and data source identifier according to the original acquisition sequence number. Discrete quantities save the status code and its change boundary according to the state holding interval. A fixed mapping is established between the data source identifier and the component association identifier of the monitored component. The operating condition record is organized according to the same acquisition period as the state sequence. Each operating condition record is associated with the load status and environmental status within the corresponding period and retains the state change direction when entering the current acquisition period from the previous acquisition period. During the data reception process, a uniform frequency resampling is not performed on different data sources. Instead, an independent original sequence and continuous acquisition sequence number interval are maintained for each data source. Then, the acquisition period identifier is used to establish a period-level association between the multi-source state sequence and the operating condition record, so that subsequent processing can distinguish the data boundaries formed by the change of the measured value itself, the migration of the operating condition, and the interruption of the acquisition link.
[0047] Specifically, the operating condition code is formed by sequentially partitioning the operating condition records. The processing end establishes ordered intervals for load status and environmental status based on historical stable operating records, maps the operating condition records of the current collection period to the corresponding ordered intervals, and combines the interval numbers into the current area code. The processing end then reads the area codes of adjacent collection periods, compares the relative positions of the current area code and the previous area code in each ordered interval, and encodes the results of remaining unchanged, migrating to a higher interval, or migrating to a lower interval as migration direction codes. The current area code and migration direction codes are then combined into the operating condition code. When the load status and environmental status change in different directions within the same collection period, the corresponding direction codes are retained instead of being compressed into a single rise / fall indicator. This ensures that operating states with the same current area code but different entry paths form different operating condition codes, thereby avoiding the sharing of the same historical baseline when the load increases into the stable interval and the load decreases into the same interval.
[0048] Preferably, historical normal data is divided into historical data segments according to continuous collection periods. The inclusion conditions of historical data segments include no confirmed state events formed within the segment, no fault evidence chain connected to the segment appearing in the observation interval after the segment ends, and no undefinable data gaps in the collection sequence number within the segment. Historical data segments that meet the inclusion conditions are combined into historical operating condition trajectories according to the operating condition coding order, and each historical operating condition trajectory is indexed with the original state sequence of each data source in the same period. For the current operating condition trajectory, the processing end compares the relative positions of the current area code, migration direction code, and operating condition migration boundary in the trajectory in turn, retains only the historical operating condition trajectory corresponding to the area code and with the same migration direction, and then aligns the migration boundary of each retained trajectory to the migration boundary of the current operating condition trajectory, so that the state quantities in different historical segments are aligned according to the operating condition evolution position rather than absolute date or natural time.
[0049] In this embodiment, conditional baselines are constructed for each data source. After aligning historical operating condition trajectories, the processing end extracts the set of state variables at the same relative position as the current acquisition time in each trajectory, arranges the set according to the size of the state variables, and forms a central trajectory with continuous values in the middle. The conditional baseline interval is formed by the normal distribution boundaries on both sides of the central trajectory. For state variables with obvious directionality, conditional baselines are calculated separately for load increase and load decrease trajectories to avoid mixing the two operating paths due to hysteresis. For discrete state variables, the conditional baseline is composed of the stable state of the historical operating condition trajectory at the corresponding position and its allowable holding interval, without numerical averaging of discrete state codes. Each conditional baseline record saves the data source identifier, operating condition code, relative position of the trajectory, central trajectory, and upper and lower distribution boundaries. During real-time processing, the corresponding record is directly obtained according to the current operating condition code.
[0050] In a preferred embodiment, when the current working condition trajectory does not have a complete corresponding trajectory in the historical normal data, the processing end does not directly replace it with data segments of adjacent working condition codes. Instead, it registers the unmatched trajectory position as a baseline gap segment and saves the historical segments that can correspond to the front or back of the current trajectory as candidate segments. In the case where only a single intermediate working condition code is missing and the migration direction of adjacent codes is continuous, the processing end limits the candidate range of the intermediate baseline according to the change direction of the center trajectory of the candidate segments on both sides. However, the candidate range only participates in the interval expression of the reliable data record and does not directly trigger state events. When the subsequent normal data collection covers the baseline gap segment and meets the inclusion conditions of historical normal data, the new data is written to the corresponding working condition trajectory position, the center trajectory and distribution boundary are recalculated, and the candidate state events in the gap segment are revised by backtracking using the newly formed conditional baseline.
[0051] Furthermore, continuous and discrete quantities in the multi-source state sequence are segmented in different ways within the source. For continuous quantities, the conditional baseline interval is used as a reference. Continuous sampling points located within the conditional baseline interval and whose change direction remains stable are merged into stable intervals, while sampling points that continuously cross the upper or lower distribution boundary are merged into abnormal change segments. For discrete quantities, the state holding interval is used as a reference. Continuous records with the same state code are merged into state holding segments, while acquisition boundaries where the state code changes are registered as state switching segments. Abnormal change segments and state switching segments are uniformly converted into intra-source event intervals. Intra-source event intervals store data source identifiers, component association identifiers, original acquisition sequence number ranges, occurrence time intervals, direction identifiers, and duration. The direction identifier of continuous quantities is taken as positive or negative relative to the conditional baseline, while the direction identifier of discrete quantities is taken as the ordered relationship before and after the state transition.
[0052] Preferably, the cross-source association index is established according to the component association identifier and the collection period identifier. Each index entry points to the same component or components with direct electrical connection relationship within adjacent collection periods. The processing end compares the occurrence time range of different intra-source event intervals. When two occurrence time ranges have an intersection, a synchronous association is established. When there is no intersection, but the end boundary of the previous event interval is earlier than the start boundary of the next event interval, the sequential association is retained. Data sources with different sampling frequencies do not generate additional sampling points through interpolation. The lower frequency data source participates in the intersection judgment with its original collection interval, and the higher frequency data source retains the original change details inside the interval. The cross-source association index only represents the temporal candidate relationship between events and does not directly merge all associated events into the same fault evidence. Whether an event enters the same fault evidence chain is still determined by the electrical reachability relationship and the fault rule graph.
[0053] In this embodiment, the electrical connection relationship of the ring network box is represented by component nodes and connecting edges. Component nodes correspond to electrical components that can be located by monitored state quantities, and connecting edges correspond to the electrical transmission relationship between components in the current operating state. The processing end writes the state events into the corresponding component nodes according to the component association identifier of the event interval within the source, and reads the connecting edges to determine the electrical reachable path between event nodes. The fault rule graph adds state transition constraints to the electrical connection relationship. Each event node is associated with the allowed abnormal direction to enter, and each directed rule edge records the predecessor event type, the successor event type, and the allowed sequential relationship between the two. When the component connection state changes, the processing end updates the reachable path with the changed electrical connection relationship. The new reachable relationship is only used for state events formed after the update time, and the already formed fault evidence chain continues to retain the connection relationship version referenced when it was established.
[0054] Specifically, the fault evidence chain is established starting from the earliest state event with a reachable successor node. The processing end searches for successor event nodes that satisfy the electrical reachability relationship along the fault rule graph, and then compares the temporal and directional relationships between the predecessor event interval and the successor event interval. When the successor event occurs later than the predecessor event and the directional combination conforms to the directed rule edge record, the corresponding directed rule edge is activated and the successor event is connected to the current path. When a common predecessor occurs, multiple candidate paths are retained. When a common successor occurs, each predecessor path is recorded separately without immediately merging the event evidence. State events that cannot be connected to any directed rule edge are saved as isolated evidence records. Isolated evidence can re-enter path matching after a connectable event occurs in a subsequent collection period, but the propagation state of the fault evidence chain is not changed before a connection is formed.
[0055] In this embodiment, the graded fault warning result is formed based on the extension position of the fault evidence chain. When there is only a state event but no directed rule edge has been activated, the source end abnormal state is registered. The directed rule edge between the predecessor event and the successor event is registered as a propagation warning state after being confirmed by the time relationship and direction relationship. When the fault evidence chain reaches the event node corresponding to the target fault type in the rule graph, the target alarm state is registered. Each data source evidence uses the state event record as the only recorded object. Even if the same state event is referenced by multiple cross-source association indexes, only one evidence identifier is retained. The warning record saves the current graded state, associated fault path, source end event interval and current path end. When a subsequent state event arrives, the path extension position is updated in the same record to avoid splitting the same evolution process into multiple independent alarm records.
[0056] In a preferred embodiment, the determination of continuous state events simultaneously considers the conditional baseline boundary, data reliability, and cumulative continuous deviation. The processing end calculates the directional deviation contribution for each data source according to the original acquisition sequence number. Sampling points located within the conditional baseline interval are recorded as zero contribution, sampling points above the upper distribution boundary are recorded as positive contribution, and sampling points below the lower distribution boundary are recorded as negative contribution. The reliability coefficient given by the data reliability record is then used to correct the contribution of each sampling point to the continuous deviation interval. The cumulative deviation when the acquisition sequence number continues from u to v is calculated according to the following formula.
[0057] ;
[0058] Where p represents the data source identifier, n represents the original acquisition sequence number of the data source, and u and v represent the starting and ending acquisition sequences of the continuous deviation interval, respectively. This represents the real-time state quantity after dimension normalization. and These represent the lower and upper distribution boundaries of the conditional baseline interval under the current operating condition coding, respectively. This represents the confidence coefficient determined by reliable data records. When the original collection sequence numbers are continuous and there are no candidate missing intervals, the complete confidence value is taken. When the sampling point is located within the coverage area of an interval-type state event, it is determined according to the ratio of the length of the confirmed sub-interval to the length of the candidate occurrence interval. This represents the cumulative amount formed by continuous deviations in the same direction within the range of the acquisition sequence number. Positive values correspond to continuous deviations above the conditional baseline, while negative values correspond to continuous deviations below the conditional baseline.
[0059] Please refer to Figure 2 , Figure 2 The horizontal axis represents the sampling window number, ranging from 0 to 500, and the vertical axis represents the monitored state values, mainly used to present real-time state quantities. Baseline boundary with current operating conditions , The relative positional relationship between them. The black solid line is the dynamic state quantity curve, the gray dashed line is the lower boundary of the baseline, and the gray dotted line is the upper boundary of the baseline; the three curves together correspond to the boundary discrimination process before each sampling point enters the directional deviation accumulation in the cumulative deviation formula.
[0060] from Figure 2 It can be seen that the initial part of the sampling window Long-term and The fact that the real-time state variables at this stage are within the normal distribution range of the current operating conditions indicates that the variables are within this range. No effective deviation contribution is generated. Approximately 180 degrees after the sampling window, the state variable rapidly increases and crosses the upper boundary. Figure 2The black curve in the middle forms a clear gap with the upper boundary. This gap corresponds to the positive deviation above the conditional baseline, indicating that the anomaly is not triggered by a fixed threshold alone, but is jointly defined by the upper and lower boundaries under the current operating conditions.
[0061] Further observation of the later stages revealed another surge in the state quantity within a window of approximately 300 to 340, significantly exceeding [the previous range]. This interval corresponds to the key segment of continuous positive deviation; subsequently, the curve falls back to the vicinity of the baseline interval and slowly changes along the lower boundary, indicating that although the state quantity is still in a relatively high background, its degree of exceeding the boundary relative to the current operating condition baseline has weakened. Figure 2 Therefore, the three segmentation relationships in the cumulative deviation formula—above the upper boundary, below the lower boundary, and within the boundary—are transformed into visual boundary intervals.
[0062] Please refer to Figure 3 , Figure 3 The horizontal axis also represents the sampling window number, the left vertical axis represents the contribution of a single point, and the right vertical axis represents the continuous cumulative contribution. The gray bars represent the contribution of each sampling point based on... and , The difference and the confidence coefficient The corrected weighted contributions are represented by the black line, which indicates how these contributions are generated as they progress along the sampling window. trajectory of change; Figure 3 The key point is to demonstrate the transformation relationship of the cumulative deviation formula from point-by-point discrimination to interval accumulation.
[0063] Figure 3 The initial gray bars are close to the zero line, indicating that although there are slight fluctuations in the sampling points, they do not constitute a valid out-of-bounds contribution. Within a window of approximately 180 to 210, the gray bars concentrate above the zero line, and the black cumulative curve rises rapidly in sync, indicating that continuous positive deviations are read in point by point. This forms a stable foundation of internal state events. This stage is related to... Figure 2 The intervals where the black curve in the middle exceeds the upper boundary correspond to each other, verifying the consistency between the boundary discrimination results and the cumulative results.
[0064] Around the 285 to 310 window, the gray bars turn below the zero line, and the black cumulative curve drops accordingly, indicating that the real-time state quantity deviates in the opposite direction from the conditional baseline. The system will not directly regard this part as a continuation of the positive anomaly. A positive contribution reappears in the 320 to 340 window, causing the cumulative curve to rise briefly, reflecting that there may be multiple deviation segments in different sampling segments of the same data source.
[0065] For example, if the normalized conditional baseline range of a data source under the current operating condition coding is 68 to 72, and the current sample value is 75 with a confidence coefficient of 0.8, the corresponding sample point generates a positive deviation contribution of 2.4. When the next sample value is 74 with a confidence coefficient of 1, it generates a positive deviation contribution of 2. The two consecutive sample points form a cumulative deviation of 4.4. If the corresponding conditional baseline range changes to 72 to 76 after the operating condition is migrated, the sample point with the same value of 75 is located within the new conditional baseline range, and its deviation contribution becomes zero. Based on this, the processing end separates the state variable rise caused by the operating condition migration from the anomaly that continues to grow relative to the current operating condition. The event start boundary is taken as the earliest collection sequence number corresponding to the position where the cumulative deviation leaves the historical normal cumulative distribution of the same operating condition, and the event termination boundary is taken as the position where the deviation contribution returns to zero and a stable range is formed again.
[0066] In a preferred embodiment, please refer to Figure 4 The reliable data record is based on the continuous interval of the original acquisition sequence number of each data source, and is associated with other intra-source event intervals corresponding to it in the cross-source association index. When the processing end receives two adjacent status records, it compares the acquisition sequence number. If the acquisition sequence number is continuous, the current continuous interval is expanded. If the acquisition sequence number is not continuous, a data gap record is established in the missing sequence number range. For cases where the data gap is located within the condition baseline interval before and after and no associated status event occurs, the missing interval is limited to a normal candidate interval. For cases where a directional deviation has already formed before the gap and other electrical associated data sources have directional compatible events during the gap period, the candidate occurrence interval is limited according to the occurrence time range of the associated event and the acquisition cycle of the current data source, and the candidate occurrence interval is written into the interval-type status event without generating a definite sample value.
[0067] Furthermore, the post-recovery processing unit reads the state and direction information of the recovery record, traces the recovery record back to the data gap boundary, and compares it with the candidate occurrence interval of the interval-type state event. When the recovery record continues the directed deviation before the gap, the candidate occurrence interval is merged with the continuous deviation interval after recovery, the event starting boundary is retained inside the candidate occurrence interval and shrunk according to the time intersection of the associated data sources. When the recovery record returns to the conditional baseline interval, the interval-type state event is terminated and the confidence coefficient of the sampling contribution in the missing interval is reduced. When the recovery record is opposite to the candidate direction, the original candidate state event is revoked, the missing interval is registered as a direction-uncertain record, and its qualification to connect to the fault evidence chain is canceled until other data sources form an associated event that can determine the direction or a new continuous sampling interval covers the missing position.
[0068] In this embodiment, when the current working condition trajectory moves from a matched working condition area to an adjacent working condition area, the processing end reads the pre-migration condition baseline and the candidate condition baseline corresponding to the target working condition code, and establishes a transition acquisition interval on both sides of the working condition migration boundary. The processing end retrieves data sources that have not formed state events within the transition acquisition interval, uses these data sources as common reference sources, calculates the difference in the center trajectory between the end of the pre-migration condition baseline and the starting position of the candidate condition baseline for each common reference source, and then synthesizes a common migration amount according to the historical stability of each common reference source. The common migration amount is simultaneously deducted from the center trajectory and the distribution boundaries on both sides of the candidate condition baseline, so that the corrected candidate condition baseline connects with the pre-migration condition baseline at the migration boundary, and the connected candidate condition baseline is written into the current acquisition time period.
[0069] Preferably, the selection of common reference sources is based on the common conditions of empty state events and continuous original acquisition sequence numbers. When any reference source shows cumulative deviation within the transition acquisition interval, it is removed from the calculation of common migration amount. The processing end uses the remaining reference sources to redetermine the correction amount of the candidate condition baseline. When all associated data sources form state events within the transition acquisition interval, the candidate condition baseline is not corrected with abnormal data. Instead, the corresponding interval is kept as the baseline void segment, and the end boundary of the condition baseline before migration and the start boundary of the candidate condition baseline are jointly saved as the uncertain baseline range. The uncertain baseline range only participates in the calculation of the confidence coefficient of state events. The baseline is transferred after subsequent stable data forms a new common reference source, and the candidate events already formed in the void segment are backtracked and verified.
[0070] In this embodiment, the start and end boundaries of state events are determined by the continuity of directional deviation and the historical normal cumulative distribution. The processing end accumulates positive and negative deviation contributions respectively, and sampling contributions in opposite directions are not allowed to directly cancel each other within the same event. When the continuous cumulative deviation amount leaves the historical normal cumulative distribution under the current operating condition code, the process traces back from the current acquisition sequence number to find the position where the continuous same-direction contribution begins as the event start boundary. After the deviation contribution returns to zero, the subsequent stable interval is checked. When the stable interval is consistent with the direction of the center trajectory of the conditional baseline, the first zero-contribution acquisition sequence number is used as the event termination boundary. When the same-direction deviation reappears after a short period of decline, the decline interval is retained within the original event and marked as a transient recovery segment, thereby separating the continuous evolution process from mutually independent repetitive events.
[0071] Specifically, interval-type state events and deterministic state events use the same data structure. The difference is that the starting or ending boundary of interval-type state events is stored as a candidate range. The processing end writes the confidence coefficient within the candidate range into the corresponding collection sequence number position and retains the source of the associated evidence in the cross-source association index. When an interval-type state event enters the fault rule graph, it is only allowed to connect the successor nodes where its candidate occurrence interval and the directed rule edge propagation interval intersect. Connections are not allowed to be established solely based on the same direction. When the candidate range is shrunk after recovery collection, the processing end recalculates the time intersection of all relevant directed rule edges, deletes the edges that have lost intersection, and revises the fault evidence chain so that the uncertainty of the data gap can be propagated along the event interval and rule edge, instead of being converted into a deterministic fault state.
[0072] In a preferred embodiment, the event nodes in the fault rule graph are indexed by both component association identifiers and abnormal directions. Positive deviation events, negative deviation events, and discrete state switching events of the same component correspond to different event semantics. Directed rule edges are established by combining historical fault event sequences and electrical reachability relationships. The processing end maps historical fault events to component nodes according to the original acquisition order. Only when there is an electrical reachable path between the predecessor component and the successor component and the event direction combination is consistent in multiple historical sequences, is the corresponding directed rule edge retained. Each directed rule edge stores the semantics of the predecessor event, the semantics of the successor event, the allowed propagation direction, and the historical propagation interval. After the real-time state event is written to the node, it is checked item by item according to the records. When all record relationships are satisfied, the edge connection is activated.
[0073] In this embodiment, candidate fault paths are generated according to the principle of no time backoff. The processing end traverses backward from the source state event along the activated directed rule edge. For each subsequent event, its starting boundary is compared with the starting boundary of the current path end event. If the time position of the subsequent event is earlier than the path end, it is not connected to the current path. When multiple subsequent events simultaneously meet the directed rule edge, a branch path is formed. Each branch shares the event record before the fork but saves an independent path identifier. When multiple predecessor events connect to the same successor event, a convergence relationship is formed. The successor event only retains one event evidence identifier and records all reachable predecessor paths. State events that do not meet the path connection conditions are retained as isolated evidence. The occurrence interval of isolated evidence still participates in the cross-source retrieval of subsequent collection periods. After exceeding its connectable propagation interval, it is transferred to the historical event record.
[0074] Furthermore, the propagation interval of each directed rule edge is calculated by historical matching event pairs. The processing end reads historical fault paths with the same predecessor event semantics, successor event semantics, and electrical endpoint relationships, calculates the difference between the predecessor starting boundary and the successor starting boundary of each event pair, and converts event pairs with uncertain boundaries due to data gaps into time difference intervals. The time differences and time difference intervals are arranged in order of size, and isolated records that do not intersect with the main continuous interval are removed. The propagation interval is formed by the coverage of the remaining records. After the current end event of the real-time fault evidence chain is formed, its occurrence time interval is added to the propagation interval of the corresponding directed rule edge to obtain the successor event retrieval interval, and the state event with compatible direction and reachable component is searched within the retrieval interval.
[0075] In a preferred embodiment, when the same state event falls into multiple subsequent event retrieval intervals simultaneously, the processing end does not immediately copy the state event, but saves it as a shared predecessor or shared successor of the relevant candidate fault paths, and configures a single evidence identifier for the state event; after a new state event occurs in a subsequent collection period, a new subsequent event retrieval interval is calculated along the directed rule edges of each candidate path, and the occurrence interval, direction identifier, and component association identifier of the new state event are compared with each retrieval interval; when only one candidate path forms a continuous match, the shared event is assigned to the path and other path references are removed; when multiple candidate paths still form a match, the sharing relationship is maintained, but the single evidence identifier is read only once when calculating path evidence, and paths that do not form a subsequent match are retained as empty branches.
[0076] In this embodiment, counterfactual verification is performed on candidate fault paths that have completed shared event attribution. The processing end selects the earliest state event in the path with subsequent connections, copies the current path record, and replaces the real-time state quantity corresponding to the earliest state event in the copied record with the conditional baseline interval under the same operating condition code. The state events, trusted data records, and connection relationships of other data sources remain unchanged. The processing end re-executes state event generation, rule edge activation, and path connection. After replacement, subsequent events that can still be connected by other predecessor events are transferred to the corresponding candidate path. Subsequent events that lose continuous predecessors are decoupled from the original path association. By comparing the edge connection retention before and after replacement, the dependence of the current candidate path on the source event is determined, and the attribution relationship between the main path and the candidate path is revised accordingly.
[0077] Preferably, the temporal matching degree of the candidate fault path and the counterfactual verification result are uniformly expressed by the path consistency quantity. The path consistency quantity retains the overlap degree of the event intervals of each directed rule edge, and deducts the proportion of paths that still exist after the source event is replaced by the conditional baseline. The calculation relationship is shown in the following formula.
[0078] ;
[0079] Where 'c' represents the candidate fault path identifier. Let m represent the set of directed rule edges that have been activated in the candidate fault path, and m represent the directed rule edges in the set. The weight of the side evidence is determined by the credible records of the preceding and subsequent events. This indicates the overlap ratio between the actual occurrence interval of the subsequent event and the corresponding subsequent event retrieval interval. This represents the proportion of directed rule edges that remain after the source event is replaced by the conditional baseline, relative to the total number of directed rule edges before the replacement. This represents the path consistency value of the candidate faulty path; when all edge connections disappear after the replacement, Taking zero, the path consistency value is determined by the degree of time matching; when all edge connections still exist after the replacement... If the value is 1, the candidate fault path will no longer be used as the main path corresponding to the source event.
[0080] Please refer to Figure 5 , Figure 5 The horizontal axis represents the candidate fault path numbers, from CP-001 to CP-009, and the vertical axis represents the normalization result. The black bars represent the path consistency of each candidate path. The diagonal-striped hollow column represents the counterfactual retention ratio. The two sets of bars correspond to the relationship between the path time matching degree and the source dependency deduction in the path consistency formula, which is used to determine whether the candidate fault path has a sufficiently stable chain of evidence.
[0081] from Figure 5 It can be seen that CP-002, CP-003, CP-004, and CP-006 to CP-009 The values are all close to 0.64, significantly higher than 0.56 for CP-001 and 0.60 for CP-005, indicating that these candidate paths contain edge evidence weights for paths leading to regular edges. Overlap ratio with event interval The combination resulted in a higher weighted matching result. The height difference between the black bars reflects the differences in the sequence of events, the degree of hit within the propagation range, and the allocation of shared evidence among the different paths.
[0082] Figure 5 The twill hollow column remains around 0.20, indicating that after replacing the source event with the corresponding conditional baseline, the proportion of edge connections that can still be retained is approximately 1 / 5. Because... It acts as a deduction in the path consistency formula; when the counterfactual retention ratio increases, even... and The weighted average is higher, and the final It will also be suppressed; therefore, Figure 5 It's not simply about comparing the number of anomalies along the path, but about observing both temporal consistency and the irreplaceability of the source event.
[0083] The lowest black bar in CP-001 indicates relatively weak overlapping of regular edges or weight combinations, or that the path contains a large number of components that can be explained by other precursors. While CP-005 is higher than CP-001, it is still lower than most paths, indicating that its chain of evidence stability is at an intermediate level. In contrast, if a path has a high black bar and the striped bars do not rise synchronously, it indicates that the path satisfies both propagation time matching and has a strong dependence on the earliest source event, and can be prioritized as the main path or the basis for higher-level warnings.
[0084] For example, a candidate fault path contains three directed rule edges with edge evidence weights of 0.9, 0.8, and 0.7, respectively. The corresponding event interval overlap ratios are 0.8, 0.7, and 0.6, respectively, and the weighted overlap result is 1.7 ÷ 2.4. After replacing the source event with the conditional baseline, the remaining edge connections account for 0.2 of the original edge connections. Therefore, the path consistency quantity is the weighted overlap result multiplied by 0.8, and the calculated result is approximately 0.567. If another candidate path has a high event interval overlap but most edge connections still exist after replacing the source event, its path consistency quantity will decrease due to the increased survival ratio. Based on this, the processing end retains concurrent paths that can be explained by other predecessors as candidate paths and registers paths where time matching and source dependency are both true as primary paths.
[0085] In this embodiment, the hierarchical state sequence consists of source-end abnormal state, propagation warning state, and target alarm state. The source-end abnormal state is formed after the event interval is formed and verified by the data trust record. The directed rule edge between the source-end state event and the subsequent event is activated, and the corresponding candidate fault path enters the propagation warning state after passing the counterfactual verification. The target alarm state is entered when the end of the path reaches the event node of the pre-associated target fault type in the fault rule graph and the actual subsequent event falls into the corresponding propagation interval. The state transition record is bound to the fault path identifier. Subsequent events on the same main path only update the current hierarchical state and the end of the path, and do not re-establish independent warning records. The alternative path recalculates the path consistency quantity after obtaining new subsequent event support and then decides whether to convert it into the main path.
[0086] Furthermore, state recovery is processed according to the reverse connection order of the fault evidence chain. When the deviation contribution of the target event returns to the conditional baseline interval and forms a stable interval, the processing end first cancels the target alarm state and checks its predecessor directed rule edge. When the subsequent event recovers but the predecessor event still deviates, the path reverts to the propagation warning state. When both the predecessor and subsequent events recover and the directed rule edge no longer satisfies the propagation interval, the path reverts to the source end abnormal state. When the source end event recovers and the downstream event is confirmed by counterfactual verification to be explainable by other paths, the original path record is canceled and the shared event attribution is transferred. The warning record continuously saves the entry boundary, exit boundary, corresponding state event and conditional baseline version of each hierarchical state, so that the state establishment, path expansion and reverse recovery of the same fault evolution process maintain a continuous reference relationship.
[0087] In a preferred embodiment, the processing method is applied to the thermal status early warning of the ring network box connection. The processing end combines the load-related status and environmental status into a working condition code, retrieves the temperature change trajectory under the same load increase / decrease direction from historical normal data, and constructs the conditional baseline of the temperature data source of the connection. When the real-time temperature rises with the load but remains within the corresponding conditional baseline range, no directional deviation event is formed. When the temperature deviates continuously from the conditional baseline, the positive deviation is accumulated according to the original acquisition sequence number and a source end status event is generated. When adjacent state quantities with electrical reachability to the connection form a directional compatible event within the corresponding propagation range, the fault rule graph connects the directed rule edges and forms a propagation early warning state. If no subsequent component event occurs and the source end temperature recovers to the conditional baseline range after the working condition stabilizes, the fault evidence chain is canceled according to the reverse path, and the transient temperature rise caused by load switching is not retained as a target alarm.
[0088] In another preferred embodiment, the processing method is applied to insulation state evolution early warning. The processing end converts insulation-related continuous quantities and environment-related continuous quantities into in-source event intervals and writes them into the corresponding event nodes with component association identifiers. Environmental state changes are entered into the condition baseline construction through operating condition coding, so that the normal fluctuation range of the wet operation stage and the dry operation stage are stored separately. After the insulation-related state quantity forms a continuous deviation outside the current condition baseline, the processing end retrieves environmental events and other insulation events of the same component and electrically related components based on the cross-source association index, and then establishes candidate paths according to the direction combination and propagation interval in the fault rule diagram. When a data source experiences an acquisition interruption, the candidate occurrence interval participates in path matching. After the acquisition is resumed, the event boundary is shrunk and the directed rule edges are re-verified. Only when the path time relationship, component reachability relationship and counterfactual dependency relationship are all true at the same time will the hierarchical state be advanced.
[0089] In another preferred embodiment, when the multi-source data of the ring network box simultaneously suffers from interrupted acquisition sequence numbers, repeated uploads, and short-term disturbances, the processing end uses the original acquisition sequence number to identify data gaps and duplicate records. For duplicate records with the same content and acquisition sequence number, only one evidence identifier is retained. For data gaps, interval-type state events are established. For short-term disturbances, whether an event is formed is determined based on whether the cumulative deviation deviates from the historical normal distribution under the same working conditions. When multiple events enter the same cross-source association index due to their close time, they still need to go through electrical reachability, propagation interval, and counterfactual verification before they can be connected into a fault evidence chain. When duplicate records are referenced by multiple candidate paths, the shared predecessor mechanism restricts the repeated accumulation of evidence. When an interval-type state event is revoked after acquisition is resumed, the relevant rule edges and path states are synchronously rolled back, thereby completing the closed-loop processing from original heterogeneous data, conditional baseline, state events, candidate paths to hierarchical fault warning records.
Claims
1. A method for fault early warning by multi-source data fusion in ring network boxes, characterized in that, include: Acquire the multi-source status sequence of each monitoring point in the same ring network box and the operating condition record generated synchronously with the multi-source status sequence, and generate the operating condition code for each collection period. Select data segments from historical normal data that match the current operating condition code, construct conditional baselines for each data source, and encode continuous deviations of real-time state quantities relative to the conditional baselines as state events; The state events are mapped to the electrical connection relationships of the ring network box, and the state events with electrical reachability are connected in chronological order to form a fault evidence chain. Based on the extension position of the fault evidence chain in the preset fault rule diagram, evidence from various data sources is integrated to generate a graded fault early warning result; The construction of the conditional baseline includes: forming the current operating condition trajectory by arranging the operating condition codes of the current acquisition period and its adjacent acquisition periods in the order of occurrence; and retrieving historical operating condition trajectories from historical normal data whose operating condition change direction is consistent with the current operating condition trajectory. The historical state quantities are aligned according to the relative positions of the corresponding working condition codes in the trajectory to form the normal state distribution of each data source. Establish a conditional baseline interval based on the center trajectory and the two side distribution boundaries of the normal state distribution, and register the unmatched segments between the historical working condition trajectory and the current working condition trajectory as baseline vacancy segments.
2. The method for multi-source data fusion fault early warning of ring network boxes according to claim 1, characterized in that, The processing of the multi-source state sequence includes: retaining the original acquisition sequence number of each state quantity according to the data source identifier, and converting the abnormal change segment in the continuous quantity sequence and the state switching segment in the discrete quantity sequence into the in-source event interval respectively; Configure component association identifiers for the event intervals within the source, and write the dimensionally normalized change direction and duration into the corresponding event intervals; A cross-source association index is established based on the time overlap relationship between different event intervals, so that data sources with different sampling frequencies maintain their original collection sequences, and a unified status event record is formed through the cross-source association index.
3. The method for multi-source data fusion fault early warning of ring network boxes according to claim 2, characterized in that, The fault rule diagram includes event nodes corresponding to ring network box components and directed rule edges connecting the event nodes. The directed rule edges are jointly defined by the electrical reachability between components and the fault state transition order. Write the state events into the corresponding event nodes according to the component association identifiers, and activate the directed rule edges that satisfy the state transition order according to the order of the event intervals. Candidate fault paths are formed by connecting consecutive event nodes along the activated directed rule edges, and state events that do not satisfy the electrical reachability relationship or whose event order is reversed are retained as isolated evidence records.
4. The method for multi-source data fusion fault early warning of ring network boxes according to claim 3, characterized in that, A trusted data record is established for each state event. The trusted data record consists of a continuous interval of the original collection sequence number and the corresponding event interval in the cross-source association index. When the original acquisition sequence number is interrupted, the normal state range of the interruption location is determined based on the conditional baseline interval, and the candidate occurrence interval of the missing event is limited by the event interval of other data sources with electrical correlation. The candidate occurrence interval is written as an interval-type state event into the cross-source association index, and the data trust record is revised according to the connection relationship between the state quantity collected in subsequent recovery and the candidate occurrence interval.
5. The method for multi-source data fusion fault early warning of ring network boxes according to claim 4, characterized in that, When the current working condition trajectory moves from the matched working condition area to the adjacent working condition area, the pre-migration condition baseline and candidate condition baseline are obtained respectively, and the data source that has not formed a state event within the migration interval is selected as the common reference source. The candidate conditional baseline is corrected based on the difference in the center trajectory of the common reference source in the two conditional baselines, so that the corrected candidate conditional baseline takes over the conditional baseline before migration at the working condition migration boundary. The candidate condition baselines that have been completed are written into the current acquisition period, and the working condition migration intervals where a common reference source cannot be obtained are kept as the baseline gap segments.
6. The method for multi-source data fusion fault early warning of ring network boxes according to claim 5, characterized in that, The generation of the state event includes: calculating the directed deviation of the real-time state quantity according to the upper and lower distribution boundaries of the conditional baseline interval, and accumulating the directed deviation within the continuous deviation interval according to the original acquisition sequence number. The candidate occurrence intervals in the reliable data records are converted into uncertain intervals of deviation, and then merged with the continuous deviation intervals. The event start boundary is determined based on the cumulative deviation distribution formed by the historical normal data under the same working condition trajectory. The event termination boundary is determined by restoring the position to the condition baseline interval along the directional deviation amount, thus forming a state event with a direction identifier.
7. The method for multi-source data fusion fault early warning of ring network boxes according to claim 6, characterized in that, For each directed rule edge in the fault rule graph, establish a propagation interval formed by the difference between historical matching event intervals; When the same state event is associated with multiple candidate fault paths, the state event is temporarily stored as a shared predecessor of the candidate fault paths, and a successor event retrieval interval is generated according to the corresponding propagation interval. The occurrence interval and direction identifier of subsequent state events are matched with the retrieval interval of each subsequent event, and the shared predecessor is included in the candidate fault path where the matching relationship is established. A single evidence identifier is retained for shared predecessors that still match multiple paths, and the event evidence of the shared predecessors is prohibited from being repeatedly accumulated in different candidate failure paths.
8. The method for multi-source data fusion fault early warning of ring network boxes according to claim 7, characterized in that, Perform counterfactual verification on the candidate fault paths that have completed the shared predecessor attribution, select the state event with the earliest time position in the candidate fault path, replace the corresponding real-time state quantity with the condition baseline interval under the same working condition code, and keep the state event records of other data sources unchanged. According to the fault rule graph, the replaced event nodes and the directed rule edges are rematched, and the successor events that can still be connected by other predecessors are transferred to the corresponding candidate fault path. The successor events that have lost their continuous predecessors are disconnected from the path association, and the fault evidence chain is revised according to the newly formed path connection relationship.
9. The method for multi-source data fusion fault early warning of ring network boxes according to claim 8, characterized in that, Establish a hierarchical state sequence corresponding to the evolution stage of the fault evidence chain. Event nodes that form directed deviation events but have not yet been connected to directed rule edges are registered as source-end abnormal states. Paths that pass the counterfactual verification and are connected to the subsequent event retrieval interval are registered as propagation warning states. Paths that allow subsequent state events to enter the subsequent event retrieval interval are registered as target alarm states. Perform hierarchical state migration according to the connection direction of the fault evidence chain, and after the state event is restored, sequentially cancel the corresponding state according to the reverse path to form a fault warning record containing fault path identifier and current hierarchical state.
10. The method for multi-source data fusion fault early warning of ring network boxes according to claim 9, characterized in that, The generation of the operating condition code includes: establishing ordered intervals for load status and environmental status based on historical stable operation records; mapping the operating condition records of the current collection period to the corresponding ordered intervals; combining the interval numbers into the current area code; reading the area codes of adjacent collection periods; comparing the relative positions of the current area code and the previous area code in each ordered interval to generate a migration direction code; combining the current area code and the migration direction code into the operating condition code; and retaining the corresponding direction codes when the load status and environmental status change in different directions within the same collection period.