An intelligent park operation and maintenance management method based on AI autonomous learning

Abstract

The application discloses a kind of wisdom park operation and maintenance management methods based on AI autonomous learning, it is related to park operation and maintenance management technical field, for solving the problem that wisdom park is not accurate to abnormal event disposal;The application constructs park operation and maintenance snapshot flow, abnormal symptom fragment, composite abnormal evolution chain, candidate disposal script and the closed-loop processing mechanism of learning back-irrigation update, realize the unified of multi-source heterogeneous operation and maintenance data, the associated identification and disposal path dynamic optimization of cross-system abnormality, again in combination with permission state, non-stop operation requirement, resource occupation and task dependency generate target disposal path, and according to execution receipt continuous correction scene matching relationship, script ordering relationship and failure transfer condition, to improve the accuracy of research and judgment, disposal efficiency and recovery stability under the scene of park composite abnormality.

Description

Technical Field

[0001] This invention relates to the field of park operation and maintenance management technology, and more specifically, to a smart park operation and maintenance management method based on AI autonomous learning. Background Technology

[0002] In the field of smart park operation and maintenance management, park operation and maintenance methods based on equipment monitoring, alarm linkage, and work order flow have been widely used. These methods typically access business data such as building automation, security, access control, fire protection, energy consumption monitoring, environmental monitoring, and work order inspection through computer programs. They enable centralized monitoring and collaborative handling of equipment operating status, environmental changes, and repair incidents in office buildings, R&D buildings, underground parking garages, energy rooms, and public areas. Existing technical solutions focus on optimizing the processing efficiency of single equipment anomalies, single alarm events, or single work order processes. By using preset rules, threshold triggers, and human experience to dispatch maintenance personnel, inspection personnel, and related resources, they can achieve the execution of daily park operation and maintenance tasks and fault handling. These solutions generally adopt a subsystem access, submodule analysis, and sub-stage handling approach to improve operation and maintenance response speed. Their technical implementation involves the collaborative application of digital data processing, equipment status monitoring, and property operation and maintenance process management.

[0003] However, multi-source operation and maintenance data is scattered across different subsystems and different handling links. It is difficult to form a unified correlation between equipment status changes, environmental anomalies, repair reports, and repair results in terms of time and object dimensions. As a result, the system can only identify isolated alarms and it is difficult to reconstruct the continuous evolution process of anomalies from the equipment layer, environmental layer, to the management and handling layer. Especially in the real scenario of integrated office parks, when situations such as deviation of air conditioning equipment operation, abnormal environmental indicators, concentrated repair reports by personnel, and recurrence of anomalies after repair occur one after another, traditional solutions often rely on fixed rules and manual judgment for handling. This is prone to problems such as inaccurate root cause identification, unreasonable task sorting, delayed path switching, and difficulty in continuously reusing handling experience, thereby weakening the handling accuracy, response efficiency, and recovery stability of park operation and maintenance management. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the existing technology, the following solution is proposed to solve the problem of inaccurate handling of abnormal events in smart parks in the above-mentioned background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A smart park operation and maintenance management method based on AI autonomous learning includes the following steps:

[0007] Collect data from building automation, security, access control, fire protection, energy consumption, environmental monitoring, as well as work orders, inspections, and maintenance. Organize the data according to a unified time scale and by region, object, and event identifier to generate a snapshot stream of park operations and maintenance.

[0008] Extract abnormal symptom fragments corresponding to alarm mutations, operating condition deviations, repeated repair requests, and recovery failures from the park's operation and maintenance snapshot stream, and construct a composite abnormal evolution chain based on time connection, regional propagation, equipment dependence, and handling correlation.

[0009] Retrieve historical handling memory entries that are formed by work orders, inspection and maintenance data and are isomorphic to or contained in sub-chains of the composite anomaly evolution chain, and call the self-learning inference model to output a set of candidate handling scripts that include the verification order, collaborative roles, task connection and failure transfer conditions;

[0010] The permission status, non-stop operation requirements, resource consumption, and task dependencies are compiled into an execution constraint table. The candidate disposal script set is then trimmed and arranged to obtain the target disposal path.

[0011] Based on the arrival, verification, repair, retesting and recovery records corresponding to the target handling path, an execution receipt is generated. The root cause judgment bias, task sorting bias and path switching bias are inverted, and learning feedback items are generated and the self-learning inference model is updated.

[0012] Furthermore, by organizing the data according to a unified time scale and region, object, and event identifier, a snapshot stream of park operations and maintenance is generated, including:

[0013] Determine the target time slice set, and perform time mapping on building automation data, security data, access control data, fire protection data, energy consumption data, environmental monitoring data, work order data, inspection data, and maintenance data according to the target time slice set to generate data entries aligned with the time slice;

[0014] Based on the region identifier, object identifier, and event identifier, perform same-source merging and heterogeneous-source association on data entries of aligned time slices to generate park operation and maintenance snapshot items;

[0015] The park operation and maintenance snapshot items are written into the same snapshot sequence in chronological order, and status field, alarm field, work order field, inspection field and maintenance field are written into each park operation and maintenance snapshot item to form a park operation and maintenance snapshot stream.

[0016] Furthermore, performing same-source merging and heterogeneous-source association on data entries for aligned time slices based on region identifier, object identifier, and event identifier includes:

[0017] Perform field deduplication and status merging on data entries from the same data source that have the same region identifier, object identifier, and event identifier to generate merged entries with the same source.

[0018] Establish association edges for data entries from different data sources that have the same regional identifier, have a mapping relationship between object identifiers, or have overlapping event occurrence times, and generate heterogeneous association entries;

[0019] Based on the same-source merged entries and heterogeneous-source associated entries, the corresponding park operation and maintenance snapshot item within each time slice is determined, and the park operation and maintenance snapshot item is written into the park operation and maintenance snapshot stream.

[0020] Furthermore, the abnormal symptom segments corresponding to alarm mutations, operational deviations, duplicate repair requests, and recovery failures extracted from the park's operation and maintenance snapshot stream include:

[0021] Perform change detection on the status and alarm fields of the park operation and maintenance snapshot items within a continuous time slice, and generate alarm mutation fragments and operating condition deviation fragments based on the change direction, change duration and abnormal trigger boundary.

[0022] Perform event merging on the work order field, inspection field and maintenance field within a continuous time slice, and generate duplicate repair fragments and recovery failure fragments based on the duplicate repair relationship and recovery loop failure relationship of the same object within a preset window.

[0023] Alarm mutation fragments, operating condition deviation fragments, repeated repair fragments, and recovery failure fragments are written into the abnormal symptom fragment set according to time boundaries, area identifiers, and object identifiers.

[0024] Furthermore, based on temporal coherence, regional propagation, equipment dependence, and handling correlation, a complex anomaly evolution chain is constructed, including:

[0025] Establish time-connecting edges based on the chronological order and time intervals of abnormal symptom fragments;

[0026] Establish regional propagation edges based on regional connectivity, regional adjacency, or energy supply coverage;

[0027] Establish equipment dependency edges based on equipment control relationships, power supply and distribution relationships, or information transmission relationships;

[0028] Establish a processing association edge based on the same work order link, continuous processing records, or consistent execution entity identifier;

[0029] Based on time-connection edges, regional propagation edges, device-dependent edges, and handling-related edges, chain segment splicing is performed on abnormal symptom fragments to generate a composite abnormal evolution chain.

[0030] Furthermore, the historical handling memory entries retrieved, which are formed from work orders, inspection and maintenance data and are isomorphic to or contained within sub-chains of the composite anomaly evolution chain, include:

[0031] Based on work order data, inspection data, and maintenance data, extract anomaly chain identifiers, handling task sequences, execution role sequences, recovery results, and failure reasons to generate a historical handling memory entry set;

[0032] The composite anomaly evolution chain is mapped to a chain feature sequence. Isomorphic matching and sub-chain inclusion matching are performed on the anomaly chain identifiers in the historical disposal memory entry set to generate candidate historical disposal memory entries.

[0033] Candidate historical disposal memory entries are sorted according to the consistency of recovery results, the completeness of task sequence, and the timeliness of disposal to determine the set of historical disposal memory entries used for self-learning deduction.

[0034] Furthermore, the self-learning inference model outputs a set of candidate disposal scripts that include the verification order, collaborating roles, task connections, and failure transfer conditions.

[0035] Establish a scene coding subnet to map the complex anomaly evolution chain and the historical handling memory entry set into scene representation results;

[0036] Establish a script generation subnet, which includes a check order output branch, a collaborative role output branch, a task connection output branch, and a failure transfer output branch.

[0037] The output branches output the sequence of verification tasks based on the verification order, the set of collaborative roles based on the collaborative roles, the task sequence relationship based on the task connection, and the failure transfer conditions based on the failure transfer, thereby generating a set of candidate disposal scripts.

[0038] Furthermore, the permission status, non-stop operation requirements, resource consumption, and task dependencies are compiled into an execution constraint table, including:

[0039] Extract area access permissions, equipment shutdown permits, personnel availability status, resource occupancy status, and task sequence dependencies to generate a set of constraints.

[0040] The constraint set is merged according to region identifier, object identifier, role identifier, and task identifier to generate an execution constraint table;

[0041] Based on the execution constraint table, the verification tasks, disposal tasks, retesting tasks, and failure transfer tasks in the candidate disposal script set are marked with executable and blocking tags.

[0042] Furthermore, by trimming and arranging the candidate disposal script set, the target disposal paths are obtained, including:

[0043] Remove task segments marked with blocking flags, retain task segments marked with executable flags, and generate candidate execution chains;

[0044] The candidate execution chain is executed according to the priority of tasks, the availability of collaborative roles, and the failure transfer conditions to generate the target disposal path.

[0045] Generate on-site records, verification records, maintenance records, retest records, and recovery records according to the target handling path.

[0046] Furthermore, based on the on-site records, verification records, maintenance records, retest records, and recovery records corresponding to the target handling path, execution receipts are generated. Root cause determination bias, task sorting bias, and path switching bias are inverted. Learning feedback items are generated and the self-learning inference model is updated, including:

[0047] The execution sequence of on-site records, verification records, maintenance records, retest records, and recovery records is merged to generate an execution receipt;

[0048] The root cause determination bias, task sorting bias, and path switching bias are determined based on the differences between the execution receipt and the target handling path;

[0049] Write the deviation type, deviation location, anomaly chain identifier, target handling path identifier, and actual recovery result into the learning and backfeeding entry;

[0050] The scene matching relationship, script ranking relationship, and failure transition condition in the self-learning inference model are updated based on the learned feedback entries.

[0051] The technical effects and advantages of the intelligent park operation and maintenance management method based on AI autonomous learning in this invention are as follows:

[0052] This invention achieves unified organization of multi-source heterogeneous operation and maintenance data in smart parks, correlation identification of cross-system anomalies, and dynamic optimization of disposal paths by constructing a closed-loop processing mechanism that includes building a park operation and maintenance snapshot stream, extracting abnormal symptom fragments, constructing a composite anomaly evolution chain, generating candidate disposal scripts, and learning and feeding back updates. It can effectively overcome the problems of data fragmentation, isolated alarms, and disposal relying on manual experience and being difficult to continuously optimize in traditional park operation and maintenance.

[0053] Meanwhile, candidate disposal scripts are generated based on historical disposal memory entries and self-learning inference models. The target disposal path is arranged by combining permission status, non-stop operation requirements, resource consumption and task dependencies. This makes the disposal process more scenario-specific, adaptable to execution constraints and flexible in path switching. By reversing the root cause judgment bias, task sorting bias and path switching bias through execution receipts, the scenario matching relationship, script sorting relationship and failure transfer conditions are continuously corrected, thereby improving the accuracy of judgment, disposal efficiency and recovery stability in complex abnormal scenarios. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating a smart park operation and maintenance management method based on AI autonomous learning according to the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] In order to achieve the above objectives, Figure 1 A structural diagram of a smart park operation and maintenance management method based on AI autonomous learning according to the present invention is given, which specifically includes the following steps;

[0057] Collect data from building automation, security, access control, fire protection, energy consumption, environmental monitoring, work orders, inspections, and maintenance. Organize this data according to a unified time scale and by region, object, and event identifier to generate a park operation and maintenance snapshot stream. Specific implementation includes:

[0058] The application scenario is illustrated using a comprehensive office and technology park. This park includes Building A (office building), Building B (R&D building), an underground parking garage, and a public energy data room. The park is equipped with building automation data acquisition points, video security points, access control systems, fire detectors, sub-item energy consumption metering devices, and temperature, humidity, and air quality sensors, along with work order records, inspection records, and maintenance records. The park is divided into three levels of physical units: buildings, floors, and functional areas. For example, the western office area on the fifth floor of Building A can be designated as area A-05-W, and the eastern area of ​​the underground parking garage as area P-01-E. Each monitored object within the park is assigned a unique object identifier; for example, a fresh air handling unit is designated as AHU-A-05-W-01, an access control system as AC-A-05-W-03, a fire detector as FD-A-05-W-12, and a sub-item electricity meter as EM-A-05-W-01.

[0059] Meanwhile, a unified event identification table is established for events such as building automation alarms, access control anomalies, video alarms, fire triggers, sudden changes in energy consumption, environmental limit violations, work order creation, on-site inspections, and maintenance completion.

[0060] The data accessed includes equipment operating status, start / stop status, setpoints, feedback values, and alarm codes from building automation data; video analysis alarms, area loitering alarms, and intrusion alarms from security data; card swipe records, door opening records, forced door opening records, and door left open for extended periods records from access control data; smoke detector status, temperature detector status, manual alarm trigger status, and fire alarm linkage status from fire protection data; electricity consumption, instantaneous power, and energy consumption status of individual items from energy consumption data; temperature, humidity, carbon dioxide concentration, and particulate matter concentration from environmental monitoring data; repair content, repair time, repair area, and work order status from work order data; arrival time, inspection conclusion, and on-site remarks from inspection data; and repair actions, completion time, replaced parts, and recovery results from maintenance data.

[0061] After the above types of data are accessed, the timestamps are first standardized and converted into absolute time under the park's standard clock, while retaining the original source identifier for subsequent time mapping, object tracing, and result verification.

[0062] Meanwhile, the clock deviation and arrival delay of various types of data are further corrected. For the collection records whose timestamp deviation does not exceed the preset tolerance, the corrected absolute time is directly written to the corresponding target time slice. For the collection records whose timestamp deviation exceeds the preset tolerance but whose original occurrence order can be determined, time compensation is performed according to the original source identifier and a delay mark is written.

[0063] For data that arrives after the current target time slice is closed, if the occurrence time still falls within the preset backfill window, it will be written back to the corresponding target time slice and the park operation and maintenance snapshot item will be updated. If it exceeds the preset backfill window, the original record will be retained and a cross-slice delay flag will be written.

[0064] For missing fields or fields that failed to be collected, a missing test marker is written to ensure that the extraction of abnormal symptom fragments and the construction of composite abnormal evolution chains can distinguish between the real normal state and the data missing state.

[0065] When generating a snapshot stream of park operations and maintenance, first determine the target time slice set. Considering that the building automation data is updated frequently, while work orders, inspection and maintenance data are generated in the form of events, one minute can be set as a target time slice, and a natural day can be divided into a set of consecutive target time slices.

[0066] For continuously sampled data, such as temperature, humidity, energy consumption and equipment feedback values, records falling within the same target time slice are mapped to data entries in the same aligned time slice, and the last valid value, peak value, direction of change marker and limit violation marker are retained.

[0067] For event-triggered data, such as forced door opening, smoke alarm, work order creation, inspection arrival, and maintenance completion, the target time slice to which the event occurred is used as the mapping segment.

[0068] For data with start and end times, such as doors left open for extended periods, continuous out-of-limit events, and maintenance operations, the event is overlaid onto all corresponding target time slices, and a continuous status flag is written into each time slice. After the above processing, the original records from different sources are converted into aligned time slice data entries with a unified time granularity.

[0069] After aligning time-slice data entries, same-source merging and dissimilar-source merging are performed based on region identifier, object identifier, and event identifier. Same-source merging refers to performing field deduplication and status merging on data entries from the same data source that have the same region identifier, object identifier, and event identifier.

[0070] For example, during the time slot of 09:15 to 09:16 in the west office area on the fifth floor of Building A, the building automation system collected three records for the fresh air unit AHU-A-05-W-01, which corresponded to high supply air temperature, decreased fan frequency, and continuous operation alarm, respectively. At this time, duplicate fields are first deleted, and then the operation status, alarm status, feedback parameters, and change markers in this time slot are merged into a single entry with the same source. This avoids multiple separate records for the same object in the same time slot. If the same object has both a normal status and an alarm status in the same time slot, the alarm status is retained, and the status change marker is written synchronously.

[0071] Heterogeneous association refers to establishing association edges for data entries from different data sources that meet the requirements of regional consistency, valid object mapping relationship, or overlapping event occurrence time. An object mapping relationship table can be pre-established to correspond the fresh air unit AHU-A-05-W-01, electricity meter EM-A-05-W-01, environmental monitoring point ENV-A-05-W-07 with the region A-05-W and its service object, the air conditioning subdomain of the western office area.

[0072] During the time slice from 09:15 to 09:16, in addition to the abnormality of the fresh air unit, there were also instances of rising carbon dioxide levels at the environmental monitoring point, abnormally decreasing power at the electricity meter, and a newly created repair work order for the stuffy western office area in the work order platform. At this time, as long as the data entries in the time slice are aligned to meet the requirements of regional consistency and object mapping to the same service object, or different events overlap within the same target time slice, a heterogeneous association edge can be established.

[0073] For maintenance records that specify insufficient air supply from the fresh air unit on the west side of the fifth floor of Building A, even if the maintenance record does not directly carry the sensor object identifier, it can be mapped to object AHU-A-05-W-01 based on the area information and equipment name, and associated with building automation, energy consumption and environmental data within the same time slice.

[0074] Park maintenance snapshots are generated based on time slices, area identifiers, object identifiers, or service object identifiers. Each park maintenance snapshot must contain at least a status field, an alarm field, a work order field, an inspection field, and a maintenance field. The status field records equipment operating status, environmental status, and energy consumption status; the alarm field records anomaly flags for building automation, access control, security, and fire protection; the work order field records newly created or ongoing work order information within the current time slice; the inspection field records whether an inspection was conducted, the inspection conclusion, and on-site remarks; and the maintenance field records… Record maintenance actions and recovery conclusions. Taking the data from the west office area on the fifth floor of Building A during the time slot from 09:15 to 09:16 as an example, the corresponding generated park operation and maintenance snapshot item can record the fresh air unit's operating status as "operating but with abnormal air supply", the environmental status as "carbon dioxide rising", the energy consumption status as "power lower than the adjacent time slot of the same day", the alarm field as "equipment alarm valid", the work order field as "stuffy repair work order created", the inspection field as "not on site", and the maintenance field as "no maintenance record". After writing all park operation and maintenance snapshot items into the same snapshot sequence in chronological order, a park operation and maintenance snapshot stream is formed.

[0075] To ensure that the processing can be implemented directly, the criteria for determining homogeneous merging and heterogeneous association are further clarified as follows:

[0076] Same-source merging is only performed when the region identifier, object identifier, and event identifier are consistent within the same data source; heterogeneous association can establish an association edge if at least one of the following conditions is met: the region identifier is consistent and the object mapping relationship is established, the event occurrence time segment is entered into the same target time slice, or the corresponding object identifier or region identifier is explicitly referenced in the management record;

[0077] Through the above processing, scattered equipment status data, environmental change data, energy consumption data, and management data such as work orders, inspections, and maintenance can be uniformly organized into a continuous snapshot stream of park operation and maintenance, providing a complete, clear, and executable data foundation for subsequent extraction of abnormal symptom fragments and construction of composite abnormal evolution chains.

[0078] Extracting anomalous symptom fragments corresponding to alarm mutations, operational deviations, duplicate repair requests, and recovery failures from the park's operation and maintenance snapshot stream, and constructing a composite anomaly evolution chain based on time sequence, regional propagation, equipment dependence, and handling correlation, the specific implementation includes:

[0079] Taking the western office area on the fifth floor of Building A in the aforementioned integrated office technology park as an example, the process of extracting abnormal symptom fragments and constructing composite abnormal evolution chains will be further explained. This area is characterized by the operation status of the fresh air unit AHU-A-05-W-01, the sub-meter EM-A-05-W-01, the environmental monitoring point ENV-A-05-W-07, the access controller AC-A-05-W-03, and the corresponding work orders, inspections, and maintenance records. The target time slice still adopts a one-minute granularity. After the park operation and maintenance snapshot stream is formed, the park operation and maintenance snapshot item in each time slice already includes the status field, alarm field, work order field, inspection field, and maintenance field. Therefore, continuous change detection and event merging can be directly performed around the same area, the same object, or the same service object to extract abnormal symptom fragments.

[0080] To extract alarm mutation segments and operating condition deviation segments, first read the park operation and maintenance snapshot items in the continuous time slice according to the object identifier, and then perform change detection on the status field and alarm field respectively. For the alarm field, focus on detecting status jumps such as normal to alarm, low-level alarm to high-level alarm, and alarm reappearing shortly after being cleared.

[0081] When an object experiences an alarm state transition between adjacent time slices, and this transition lasts for a set minimum duration, an alarm mutation fragment is generated. For the state field, corresponding abnormal trigger boundaries are pre-defined according to the object type. For example, for fresh air handling units, upper boundary for supply air temperature, lower boundary for fan frequency, and lower boundary for supply air volume can be set; for environmental monitoring points, upper boundary for carbon dioxide concentration can be set; and for energy consumption metering points, lower boundary for power can be set.

[0082] The abnormal trigger boundary is determined by calling the boundary parameter table according to the object type. The boundary parameter table shall at least record the object type, reference state source, upper boundary, lower boundary and minimum duration; the reference state source can be the current running set value, the baseline value of the historical stable running range or the range allowed by the device nameplate;

[0083] When an object is in a fixed-setting operating mode, the current operating setting value is used as the reference state. When an object is in a load-fluctuation operating mode, the baseline value of the historical stable operating range is used as the reference state. The minimum duration length is used to exclude short-term jitter and transient disturbances. Only when the same deviation direction continuously reaches the minimum duration length of the corresponding object category is the operating condition deviation segment written. Through the boundary parameter table calling method, different device objects, environmental objects, and energy consumption objects can use a unified data structure to perform anomaly judgment.

[0084] After reading the status field in the continuous time slice, determine its deviation direction relative to the previous stable state or relative to the current running set value. When the same deviation direction is maintained in multiple consecutive time slices and exceeds the abnormal trigger boundary, generate the operating condition deviation segment.

[0085] Specifically, when the state deviates by a certain amount The absolute value of the exception continuously exceeds the exception triggering boundary of the corresponding object class. And deviates from the duration Reaching minimum duration When the time interval is reached, the corresponding time interval will be written into the operating condition deviation segment;

[0086] The deviation from the state is determined by the following formula: The deviation duration is determined by the following formula: ;

[0087] When object When the following formula is satisfied within a continuous time slice, a condition deviation segment is generated: ,and ,in, Representation Object In the State deviation in each time slice; Representation Object In the The current status value of the status field corresponding to each time slice; Representation Object In the The reference status value corresponds to each time slice; the reference status value can be the current operating setting value or the benchmark value corresponding to the historical stable operating interval. Representation Object The duration of the current deviation segment; Representation Object The sequence number of the first time slice to meet the deviation judgment criteria; Representation Object The last consecutive time segment that met the deviation judgment criteria is the end time segment number; Representation Object The exception trigger boundary corresponding to the status field; Representation Object The minimum duration of the corresponding deviation segment.

[0088] Taking the scenario of the west office area on the fifth floor of Building A from 09:15 to 09:18 as an example, the supply air temperature of the fresh air unit is continuously higher than the set value and the fan frequency is continuously decreasing. At the same time, the corresponding power of the electricity meter is continuously lower than the previous stable stage, and the carbon dioxide concentration at the environmental monitoring point continues to rise. Then, the deviation segments of the fresh air unit operating conditions, the deviation segments of the electricity meter power, and the deviation segments of the environmental operating conditions are extracted respectively.

[0089] When extracting duplicate repair reports and recovery failure reports, event merging is performed on the work order field, inspection field, and maintenance field within a continuous time slice. During event merging, the work order records are first clustered according to the object identifier, area identifier, and event topic. Then, the inspection arrival record, inspection conclusion, and maintenance completion record are written into the same event cluster in chronological order. If the same object has two repair reports with similar content in a preset window, and the previous repair report has been closed or is in a pending state, it is identified as a duplicate repair relationship, and a duplicate repair report is generated.

[0090] To ensure consistent identification of similar repair requests, the repair subject, object, and region are first normalized. During normalization, the equipment name, region name, fault symptom, and handling keywords in the repair text are mapped to a unified subject code. When two repair requests have the same object identifier and subject code, or when the object identifier is mapped to the same service object and the subject code belongs to the same fault subject set, they are considered similar. A preset window is then used to determine the window based on the object category, calling the window parameter table. The window parameter table records at least the object category, the duplicate repair judgment window, and the recovery failure review window. Environmental complaint objects can use a shorter window, while equipment repair verification objects can use a longer window. By linking subject normalization and the window parameter table, the determination process for duplicate repair relationships and recovery failure relationships can be standardized.

[0091] If, after a maintenance is completed, the same or similar alarm mutation segment or operating condition deviation segment reappears for the same object in the preset review window, or if the original work order is reopened or reassigned, it is considered a failure to restore the closed-loop relationship, and a recovery failure segment is generated. Taking the above scenario as an example, if two maintenance work orders regarding stuffiness in the west office area are generated consecutively at 09:18 and 09:26, and the corresponding objects are both mapped to the fresh air unit service area, then a duplicate maintenance segment is generated for that object and area; if the maintenance record at 09:24 shows that the air supply valve has been reset, but at 09:31 the air supply temperature and carbon dioxide levels rise again, then a recovery failure segment is generated.

[0092] After the anomalous symptom fragments are formed, they are written into the anomalous symptom fragment set according to time boundaries, regional identifiers, and object identifiers, specifically including:

[0093] Each abnormal symptom segment must include at least the segment identifier, segment type, start time slice, end time slice, region identifier, object identifier, trigger field, pre-trigger status, post-trigger status, and associated work order identifier. For alarm mutation segments, the segment type is written as alarm mutation, and the direction and duration of alarm level change are recorded.

[0094] For segments that deviate from the operating condition, the segment type is written as "operating condition deviation", and the deviation direction, boundary category, and deviation duration are recorded.

[0095] For duplicate repair requests and failed recovery requests, the corresponding work order link identifier, repair action identifier, and follow-up result are recorded respectively.

[0096] In the construction phase of the complex anomaly evolution chain, multiple types of connection edges are first established around the anomalous symptom segments. Temporal connection edges are used to represent the temporal continuity of the anomalous symptom segments. The establishment method is as follows:

[0097] If the start time of the next segment is within the set time interval after the end of the previous segment, and the two segments have the same object, the same service object, or belong to the same business processing context, then a time connection edge is established between the two segments. The set time interval is determined by calling the edge interval parameter table according to the edge type. The edge interval parameter table records at least the time connection edge interval, the regional propagation edge interval, the device dependency edge interval, and the disposal association edge interval.

[0098] Time-seam edge is used to determine the continuation of anomalies for the same object or service object within consecutive time slices; area propagation edge is used to determine the time range of anomalies spreading from the current area to adjacent areas or power supply coverage areas.

[0099] Device dependency edges are used to determine the allowable time difference for anomalies from upstream objects to downstream objects.

[0100] The handling of related edges is used to determine the continuous handling relationship between work orders, inspections, and maintenance actions;

[0101] When splicing chain segments, if no subsequent segment satisfying the edge interval condition is detected in multiple consecutive time slices, or if a stable segment has appeared and reached the preset stable duration, the current chain segment is terminated from further expansion to prevent unrelated anomalous segments from being mistakenly spliced ​​into the same composite anomalous evolution chain.

[0102] Taking the above scenario as an example, the end time of the fresh air unit's operating condition deviation segment is 09:18, and the start time of the environmental operating condition deviation segment in the west office area is 09:17. Since the two overlap in time and serve the same object, a time connection edge can be established. The stuffy repair work order segment generated at 09:18 follows immediately after the environmental deviation segment, so a time connection edge can also be established.

[0103] Regional propagation edges are used to represent the spatial diffusion process of anomalies. A pre-established campus regional topology table is created, which records at least regional connectivity, regional adjacency, and power supply coverage relationships. Among them, regional connectivity indicates direct physical connection, regional adjacency indicates adjacent positions on the same floor or vertically adjacent, and power supply coverage indicates that the same equipment or the same power supply branch serves multiple regions;

[0104] When an abnormal symptom segment in a region reappears in an adjacent region or a region affected by the same power supply branch after a set time, a regional propagation edge is established between the two segments. For example, the fresh air unit AHU-A-05-W-01 serves both the office area on the west side of the fifth floor of Building A and the adjacent conference room. If the office area first experiences a segment of rising carbon dioxide concentration, and then the conference room experiences a segment of higher temperature, a regional propagation edge can be established based on the same air supply coverage relationship.

[0105] Equipment dependency edges are used to represent the dependencies between equipment in terms of control, power supply and distribution, or information transmission. For this purpose, an equipment dependency table is established in advance. The equipment dependency table records the upstream object, downstream object, dependency type, and dependency direction. Control relationships can correspond to the dependencies between valves and actuators, and start / stop commands and controlled objects. Power supply and distribution relationships can correspond to the dependencies between meters, distribution branches, and equipment loads. Information transmission relationships can correspond to the dependencies between sensors, controllers, and actuators. When an upstream object experiences an alarm mutation segment or a deviation from its operating condition segment, and a downstream object experiences a similar or derived abnormal segment in its subsequent time slice, an equipment dependency edge can be established between the two segments.

[0106] Taking the above scenario as an example, there is a power supply and load correspondence between the segment of power deflection of the sub-meter and the segment of frequency reduction of the fresh air unit fan, and there is a service object dependency between the segment of fresh air unit operating condition deviation and the segment of carbon dioxide increase at the environmental monitoring point. Equipment dependency edges can be established in both cases.

[0107] The handling association edge is used to represent the organizational relationship of abnormal symptom segments in the handling process. When establishing the handling association edge, the key is to determine whether the segments belong to the same work order link, whether they correspond to consecutive handling records, or whether the executing entity identifiers are consistent. If abnormal symptom segments of the same object are linked together by the same work order link, or if the same inspection personnel or maintenance personnel perform consecutive operations such as on-site inspection, verification, maintenance, and retesting on the relevant object within a continuous time window, then these segments can be considered to have a consistent context in handling, and a handling association edge can be established.

[0108] For example, if the work order for reporting sultry weather created at 09:18, the inspection arrival record at 09:21, the repair completion record at 09:24, and the work order reassignment record at 09:31 all point to the same object and the same area, and the executing entities have a continuous succession relationship, then a handling association edge should be established between the corresponding abnormal symptom segments.

[0109] After the multi-type connection edge is established, starting from each alarm mutation segment or operating condition deviation segment as a seed segment, chain segment splicing is performed along the time connection edge, regional propagation edge, equipment dependency edge and handling association edge. When splicing, the time unidirectional increment principle is followed, that is, the start time of the subsequent access segment must not be earlier than the start time of the current segment.

[0110] Simultaneously, object delooping is performed. If a segment has already entered the current chain segment, it will not be added again. For segments that satisfy multiple edge types, time connection relationships are retained first, followed by device dependency relationships and handling relationships, and finally, regional propagation relationships are added. This forms a composite anomaly evolution chain that can reflect both the sequential evolution of anomalies and the device dependency and handling process.

[0111] Taking the above scenario as an example, a complete chain can be formed: the segment of deviation of the fresh air unit's operating condition, the segment of deviation of the electricity meter's power, the segment of rising carbon dioxide at the environmental monitoring point, the segment of repeated repairs due to stuffiness, and the segment of failure to recover after repair. This chain can clearly reflect the continuous evolution process of the anomaly from the equipment layer, the environmental layer, the management and handling layer, and then to the failure to recover layer.

[0112] Through the above processing, the abnormal symptom fragment extraction stage not only clarified the change detection method, the determination method of change direction and change duration, and the usage method of abnormal trigger boundary for status and alarm fields, but also clarified the merging rules for repeated repair relationships and recovery loop failure relationships. The composite abnormal evolution chain construction stage further disclosed the establishment criteria for time connection edges, regional propagation edges, equipment dependency edges, and disposal association edges, as well as the order and termination method of chain segment splicing. Based on this, the extraction of abnormal symptom fragments and the construction of composite abnormal evolution chains can be directly completed, and they can be used for subsequent historical disposal memory entry retrieval and candidate disposal script set generation.

[0113] Retrieve historical handling memory entries compiled from work orders, inspections, and maintenance data that are isomorphic to or contain sub-chains of the composite anomaly evolution chain. Call the self-learning inference model to output a set of candidate handling scripts containing verification order, collaborating roles, task connections, and failure transfer conditions. Specific implementation includes:

[0114] Continuing with the example of the abnormal fresh air situation in the west office area on the fifth floor of Building A in the aforementioned integrated office technology park, a complex abnormal evolution chain has been formed in the preliminary processing. This complex abnormal evolution chain includes segments such as the deviation of the fresh air unit's operating conditions, the rise in ambient carbon dioxide, repeated repair requests due to stuffiness, and the failure to recover after repair. Based on this complex abnormal evolution chain, historical handling memory entries are further sorted out from the park's existing work order records, inspection records, and maintenance records for the generation of subsequent candidate handling script sets.

[0115] The formation of historical disposal memory entries is based on historical disposal events that have been completed in a closed loop or whose reasons for failure have been confirmed, specifically including:

[0116] First, extract the work order creation time, work order closing time, repair area, repair object, repair subject, handling status, and closed-loop result from the work order data. Then, extract the arrival time, verification actions, on-site judgment, temporary handling actions, and inspection conclusions from the inspection data. Finally, extract the repair steps, replacement parts, parameter adjustment actions, completion time, and recovery results from the maintenance data.

[0117] Subsequently, the data is merged according to the region identifier, object identifier, and work order link identifier to obtain a cluster of handling records corresponding to each historical abnormal event. For each handling record cluster, the abnormal chain identifier, handling task sequence, execution role sequence, recovery result, and failure reason are further extracted. The abnormal chain identifier is used to characterize the abnormal evolution structure corresponding to the historical event. The handling task sequence records tasks such as remote verification, on-site verification, parameter reset, component replacement, linkage notification, and retest confirmation in chronological order. The execution role sequence records the participating roles such as HVAC maintenance personnel, on-duty electricians, security patrol personnel, and floor customer service personnel. The recovery result records states such as normal recovery, partial recovery, and short-term recovery followed by recurrence of abnormality. The failure reason records reasons such as valve jamming, actuator failure, sensor drift, and fan inverter failure. After the above sorting, a set of historical handling memory entries can be formed.

[0118] In order to enable historical treatment memory entries to be effectively retrieved by the current composite anomaly evolution chain, the current composite anomaly evolution chain needs to be mapped into a chain feature sequence. During the mapping, each anomaly symptom fragment in the composite anomaly evolution chain is first transcribed into a fragment feature node. The node records at least the fragment type, region identifier, object category, trigger field category, and occurrence order.

[0119] Next, the time-connecting edges, regional propagation edges, equipment-dependent edges, and disposal-related edges are transcribed into relational features respectively; finally, the node features and relational features are concatenated according to the occurrence order of abnormal symptom fragments to form a chain feature sequence.

[0120] Taking the above scenario as an example, the current chain can be mapped to a continuous sequence of segments such as deviation of fresh air unit operating conditions, environmental deviation, repeated repair requests, and recovery failure. Relationship markers such as equipment dependence, time connection, and handling association are marked between adjacent segments, thereby obtaining a standardized chain feature sequence that can be used for matching.

[0121] When performing historical disposal memory entry retrieval, the same chain feature transformation is first performed on each abnormal chain identifier in the historical disposal memory entry set. Then, isomorphic matching and sub-chain inclusion matching are performed in sequence. Isomorphic matching is used to filter historical events that are consistent with the current composite abnormal evolution chain in terms of fragment type, relation structure and sequence. When matching, the historical chain and the current chain are required to correspond as a whole in terms of the number of nodes, node type and edge relationship.

[0122] For isomorphic matching, at least the fragment type sequence, edge relationship sequence, object category sequence, and region role sequence are compared. Isomorphic matching is considered successful only when the historical chain and the current chain correspond entirely in terms of node quantity, node type, node order, and edge relationship type. For sub-chain inclusion matching, at least the historical chain must contain a fragment sub-sequence that is continuously corresponding to the current chain, and the corresponding fragments must have the same type order, edge relationship type, object category, or be mapped to the same service object. If the same historical disposal memory entry satisfies both isomorphic matching and sub-chain inclusion matching, it is prioritized for inclusion as a candidate historical disposal memory entry based on isomorphic matching.

[0123] If multiple matching results exist, they are sorted hierarchically according to the consistency of recovery results, the integrity of the task sequence, and the timeliness of handling.

[0124] Sub-chain matching is used to filter historical events that, while longer than the current chain, contain the complete evolution path of the current chain. Matching requires that there be a segment sub-sequence in the historical chain that corresponds continuously to the current chain. Taking the above scenario as an example, if there were historical events such as deviations in the operating conditions of the fresh air unit, environmental deviations, or stuffy conditions requiring repair, which were ultimately resolved by valve reset, then this historical event can be identified as an isomorphic matching entry.

[0125] If another historical event, in addition to the above-mentioned segments, also includes segments about a new round of carbon dioxide increases caused by post-meeting crowd gatherings, but its first half completely covers the current chain, then this historical event can be identified as a sub-chain containing a matching entry. Candidate historical disposal memory entries are obtained after matching.

[0126] After obtaining candidate historical handling memory entries, they are further sorted according to the consistency of recovery results, the completeness of task sequence, and the timeliness of handling. Among them, the consistency of recovery results is used to determine whether the handling results of historical events are consistent with the current operation and maintenance goals. For example, if the goal of the current scenario is to restore the air supply capacity of the west office area and relieve the stuffy complaints, then memory entries that have achieved stable recovery in the past and have not recurred will be retained first.

[0127] Task sequence integrity is used to determine whether historical entries cover complete processing steps such as remote verification, on-site verification, maintenance actions, and retest confirmation. Entries with higher integrity are given priority.

[0128] The handling timeliness relationship is used to compare the time span from the first occurrence of an anomaly to the completion of recovery for different historical entries. When the recovery results are consistent and the completeness of the task sequence is the same, the entry with shorter handling time is given priority. Through this hierarchical sorting method, the set of historical handling memory entries most suitable for the current scenario can be selected without using weighted mixed expression.

[0129] After the candidate historical disposal memory entry set is determined, the self-learning deduction stage begins. First, a scene encoding subnet is established to map the current composite anomaly evolution chain with the historical disposal memory entry set into scene representation results. The scene encoding subnet includes a chain structure encoding path and a disposal memory encoding path. The chain structure encoding path reads the fragment type sequence, edge relationship sequence, region sequence, and object category sequence in the current composite anomaly evolution chain. The disposal memory encoding path reads the disposal task sequence, execution role sequence, recovery result, and failure reason sequence in the historical disposal memory entries.

[0130] The chain-structure encoding pathway first encodes the anomalous symptom segments in the current chain sequentially, and then writes the dependencies between adjacent segments into the encoding result according to edge relationships. The handling memory encoding pathway encodes the verification actions, maintenance actions, recovery results, and failure reasons in the order of historical handling, and retains the corresponding role information. The outputs of the two encoding pathways are concatenated and organized according to time order and object correspondence to form a scene representation result. The scene representation result obtained in this way not only retains the evolutionary structure of the current anomaly, but also retains the execution trajectory of historical handling experience.

[0131] After the scene representation results are generated, a script generation subnet is established to output a set of candidate disposal scripts, specifically including:

[0132] The script generation subnet includes a verification sequence output branch, a collaborative role output branch, a task connection output branch, and a failure transfer output branch. The verification sequence output branch outputs the verification task sequence in the current scenario based on the sequential relationship of each segment in the current chain and the corresponding valid verification path in the historical entries. For example, first verify the building automation trend record, then verify the fresh air unit valve position feedback, then verify the fan inverter status, and finally verify the environmental monitoring point drift.

[0133] The collaborative role output branch outputs the set of collaborative roles required for the current scenario based on the actual effective role participation order in the historical handling. For example, HVAC maintenance personnel will first confirm remotely, then the on-duty electrician will assist in checking the power supply and drive status, and if necessary, the floor customer service personnel will handle the situation.

[0134] The task connection output branch is used to determine the sequential relationship and prerequisite constraints between verification tasks, maintenance tasks, and retesting tasks. For example, the actuator replacement step cannot be entered without completing the valve position verification, and the work order cannot be closed without completing the retesting confirmation.

[0135] The failure transfer output branch is used to provide the transfer conditions when the current path is invalid. For example, if the air supply temperature does not drop within two time slices after the valve is reset, the path will switch to the actuator maintenance path. If the environmental indicators are still abnormal after the actuator maintenance, the path will switch to the sensor calibration path.

[0136] For the above-mentioned abnormal fresh air scenario in the west office area on the fifth floor of Building A, the script generation subnet can output multiple candidate handling scripts. The first candidate script is the valve position abnormality handling script. Its verification order is to check the valve position feedback, compare the set value with the feedback value, perform remote reset, and check the actuator on site. The collaborating roles are HVAC maintenance personnel and on-duty electricians. The task connection relationship is that on-site verification can only be carried out after the remote reset is completed. The failure transfer condition is that if the remote reset is ineffective, the process will be transferred to actuator maintenance.

[0137] The second candidate scenario is the fan drive anomaly handling scenario. Its verification order is to check the inverter alarm, compare the fan frequency and power changes, and test the power supply circuit on site. The collaborating roles are the on-duty electrician and HVAC maintenance personnel. The failure transfer condition is to transfer to drive replacement when the power supply is normal but the fan frequency cannot be restored.

[0138] The third candidate scenario is the sensor misalignment troubleshooting scenario. Its verification order is to compare the data of environmental monitoring points with the backup detection equipment, review the building automation feedback curve, and retest the environmental indicators on site. The collaborating roles are inspection personnel and HVAC maintenance personnel. The failure transfer condition is to switch to the sensor calibration path when the on-site retest is normal but the platform data is continuously abnormal.

[0139] The aforementioned candidate disposal scenarios together constitute a candidate disposal scenario set, which can be tailored and arranged in accordance with execution constraints in the future.

[0140] The permission status, non-operation requirements, resource consumption, and task dependencies are compiled into an execution constraint table. The candidate disposal script set is then trimmed and arranged to obtain the target disposal path. The specific implementation includes:

[0141] Continuing with the example of the abnormal fresh air situation in the west office area on the fifth floor of Building A in the integrated office technology park, in the current scenario, the previous steps have already output multiple candidate handling scripts, corresponding to the valve position abnormality handling path, the fan drive abnormality handling path, and the sensor misalignment troubleshooting path, respectively. Since the candidate handling scripts only provide an optional handling framework and have not yet combined with the park's regional access conditions, equipment operating boundaries, personnel attendance status, and resource occupancy status at that time, it is necessary to further compile the permission status, non-stop operation requirements, resource occupancy, and task dependencies into an execution constraint table. Then, based on the execution constraint table, the candidate handling script set is trimmed and arranged to obtain the target handling path that can be directly implemented and executed.

[0142] The compilation of the constraint table begins with the extraction of real-time constraint terms, and specifically includes:

[0143] Area access permissions are obtained from access control authorization records, maintenance access records, and key area access approval records, and are used to determine whether the on-site verification tasks and on-site maintenance tasks in the candidate disposal scripts meet the entry conditions;

[0144] Equipment shutdown permits are obtained from equipment operation plans, site usage plans, business continuity requirements, and critical equipment operation levels. They are used to determine whether shutdown maintenance tasks, power outage inspection tasks, and operation mode switching tasks in candidate disposal scenarios can be implemented. Personnel availability status is obtained from duty schedules, current work order execution status, and real-time location records. It is used to determine whether HVAC maintenance personnel, on-duty electricians, security patrol personnel, and floor customer service personnel are able to arrive on-site within the target time. Resource occupancy status is obtained from spare parts warehouse entry and exit records, special tool requisition records, spare sensor usage records, and on-site maintenance workstation occupancy records. It is used to determine whether resources such as actuators, drives, calibration instruments, and spare environmental probes are available.

[0145] The task sequence dependency is directly extracted from the task connection relationship already generated in the candidate disposal script, and is used to represent the sequence of remote verification, on-site verification, maintenance action, retest action and recovery confirmation.

[0146] In the scenario of the west office area on the fifth floor of Building A, the current time is a weekday morning, the west office area is in normal office status, and the adjacent meeting room is in use. Therefore, the fresh air unit is not allowed to be shut down for a long time. Only short-term parameter reset, partial valve position check and drive test without shutting down are allowed. It is not allowed to directly shut down and disassemble the whole unit for inspection.

[0147] Meanwhile, entering the fifth-floor west machine room requires HVAC maintenance authorization, and entering the power distribution room requires the accompaniment of the on-duty electrician and meeting the on-duty access requirements.

[0148] On-site HVAC maintenance personnel are available after 09:25, and the on-duty electrician is available after 09:35. The spare valve position actuator inventory is empty, and the portable environmental calibrator was already occupied by another area before 09:40. After extracting the above information, constraints are generated for area access permission, equipment shutdown permission, personnel availability, resource occupancy, and task dependency.

[0149] After extracting the constraints, they are merged according to region identifier, object identifier, role identifier, and task identifier to form an execution constraint table. Each record in the execution constraint table includes at least the region identifier, object identifier, task identifier, execution role identifier, constraint type, constraint status, and constraint effective time. For example, for a remote reset task in a valve position anomaly handling script, it can be recorded as: Region A-05-W, Object AHU-A-05-W-01, Task identifier: Remote Reset, Role identifier: HVAC Maintenance Personnel, Constraint type: Equipment Not Running Constraint, Constraint status: Allowed.

[0150] For actuator replacement tasks, the task can be recorded as area A-05-W, object AHU-A-05-W-01, task identifier as actuator replacement, role identifier as HVAC maintenance personnel, constraint type as spare parts occupancy constraint, and constraint status as blocked.

[0151] For the drive power distribution inspection task, it can be recorded as area A-05-W, object AHU-A-05-W-01, task identifier as power distribution inspection, role identifier as on-duty electrician, constraint type as personnel available constraint, and constraint status as delayed allowable.

[0152] In this way, each task in the candidate disposal script can be directly mapped to the corresponding real-time constraints.

[0153] After the execution constraint table is formed, the verification tasks, disposal tasks, retesting tasks, and failure transfer tasks in the candidate disposal script set are further marked according to the execution constraint table. If a task meets the area entry conditions, equipment operation conditions, personnel availability conditions, and resource availability conditions simultaneously within the current time slice, it is marked as executable; if a task fails to meet any hard constraint, such as no area access permission, equipment shutdown prohibition, or missing critical spare parts, it is marked as blocked.

[0154] If a task is not currently satisfied but can be satisfied in a subsequent time slice, its delayed executable status can be retained in the execution constraint table, and its inclusion will be determined based on the target time period during actual pruning. Taking the above scenario as an example, the valve position feedback reading task, remote parameter reset task, and non-stop environment retest task are marked as executable; the whole machine shutdown maintenance task and actuator replacement task are marked as blocked due to non-stop operation requirements and lack of spare parts; and the drive power distribution inspection task is marked as delayed executable because the on-duty electrician has not yet arrived.

[0155] During the candidate disposal script set trimming stage, task fragments marked with blocking tags are first removed, while task fragments marked with executable tags are retained to generate candidate execution chains. When generating candidate execution chains, the source script identifiers and sequential connections of the task fragments must still be retained to avoid task chain breaks after trimming. If a candidate disposal script no longer meets the basic closed-loop requirements after removing blocking task fragments, such as lacking necessary retesting tasks or essential verification steps, then the entire candidate disposal script is discarded; if a candidate disposal script can still form a complete closed loop after trimming, it is retained as a candidate execution chain.

[0156] When the number of remaining candidate execution chains is greater than one, further filtering is performed based on the number of blocking tasks, waiting time of key roles, historical stable recovery result level, and expected recovery time. First, the number of blocking tasks in each candidate execution chain is compared, with the chain having fewer blocking tasks taking priority. If the number of blocking tasks is the same, the waiting time before the key roles arrive is compared, with the chain having a shorter waiting time taking priority. If the waiting times are the same, the stable recovery result level in the corresponding historical handling memory entries is compared, with the chain having a higher recovery result level taking priority. If the recovery result levels are still the same, the expected recovery time is compared, with the chain having a shorter expected recovery time taking priority. After this sequential filtering, the candidate execution chain with the highest priority is retained as the target handling path.

[0157] For the current scenario in the west office area on the fifth floor of Building A, the valve position anomaly handling script, after being trimmed, can still retain tasks such as remote verification, valve position feedback comparison, remote reset, on-site non-stop verification, and environmental retesting, thus forming the first candidate execution chain. In the fan drive anomaly handling script, since the power distribution inspection requires waiting for the on-duty electrician to arrive, only the remote alarm verification and frequency curve comparison tasks are retained between 09:25 and 09:35. After 09:35, the on-site inspection task is connected, forming the second candidate execution chain. The sensor misalignment troubleshooting script can only retain the tasks of comparing historical curves and manual verification of backup measuring points because the calibration instrument is temporarily occupied, and it does not form a priority execution chain for the time being.

[0158] After the candidate execution chain is formed, it is then unfolded according to the order of task sequence, the availability of coordinating roles, and the failure transfer conditions to generate the target disposal path. The task sequence is directly derived from the task connection structure obtained in the script generation stage. For example, the remote reset step cannot be entered without completing the valve position feedback reading, the on-site verification step cannot be entered without completing the remote reset, and the recovery confirmation step cannot be entered without completing the on-site verification.

[0159] The availability of collaborative roles is used to determine whether a task requiring multiple people to collaborate has the conditions to be started. For example, a task involving electrical safety inspection can only be started after the on-duty electrician is in place.

[0160] The failure transfer condition is used to determine the subsequent succession direction when the current path is invalid. For example, if the air supply temperature is still higher than the trigger boundary within two time slices after the remote reset is executed, the path will switch from valve position abnormality handling to fan drive abnormality handling. If the drive path still cannot resolve the environmental deviation, the path will switch to sensor misalignment troubleshooting.

[0161] Through the above derivation, a target handling path with a clear execution order, clear execution roles, and clear switching rules can be obtained. Taking the above scenario as an example, the target handling path can be specifically derived as follows: at 09:25, HVAC maintenance personnel perform remote trend verification and valve position feedback reading; at 09:27, remote reset is performed; at 09:29, the air supply status is retested for the first time; if the air supply temperature still does not drop, at 09:35, the on-duty electrician will cooperate to carry out drive checks; after 09:40, a decision will be made on whether to switch to the backup measurement point verification path based on changes in environmental indicators.

[0162] When implementing the target disposal path, on-site records, verification records, maintenance records, retest records, and recovery records are generated step by step;

[0163] The attendance record should include at least the arrival time, the area of ​​arrival, the identifier of the executing entity, and the corresponding task identifier;

[0164] The verification record should include at least the object being verified, the verification action, the status value obtained, the verification conclusion, and the verification time.

[0165] Maintenance records should include at least the actions performed, the objects affected, the parts replaced, the parameters adjusted, the personnel involved, and the completion time.

[0166] The retest record should include at least the retest time, retest object, retest field, retest value, and retest conclusion; the recovery record should include at least the recovery confirmation time, recovery range, number of sustained stable time slices, and recovery status.

[0167] For example, in the current scenario, at 09:25, an arrival record of HVAC maintenance personnel arriving at the remote monitoring terminal is generated; at 09:26, a valve position feedback reading verification record is generated; at 09:27, a remote reset maintenance record is generated; and at 09:29, a record of the first supply air temperature and carbon dioxide retest is generated. If the supply air temperature returns to normal and the carbon dioxide concentration drops to the allowable range within multiple consecutive time slots from 09:33 to 09:36, a recovery record is generated. If the retest does not meet the recovery conditions, no recovery record is generated, but instead, the failure transfer condition is triggered to enter the next path.

[0168] Based on the arrival, verification, repair, retesting, and recovery records corresponding to the target handling path, execution receipts are generated. Root cause determination bias, task sequencing bias, and path switching bias are inverted. Learning feedback entries are generated and the self-learning inference model is updated. Specific implementation includes:

[0169] Continuing with the example of the abnormal fresh air situation in the west office area on the fifth floor of Building A in the integrated office technology park, after the prior constraint trimming and path orchestration, a target handling path has been generated. This target handling path initially sets the valve position abnormality handling path as the priority path, first performing remote trend verification, valve position feedback reading, remote reset, on-site non-stop verification, and environmental retesting, and then deciding whether to switch to the drive check path based on the retest results.

[0170] To enable the handling results to correct subsequent simulation processes, it is necessary to further consolidate the on-site records, verification records, maintenance records, retest records, and recovery records corresponding to the target handling path into execution receipts. Based on the differences between the execution receipts and the target handling path, the deviation type is inverted, learning feedback items are generated, and then the self-learning simulation model is updated using the learning feedback items.

[0171] The generation of execution receipts is based on time-series merging. First, the corresponding on-site records, verification records, maintenance records, retest records, and recovery records are read according to the task identifier in the target disposal path. Then, each record is sorted uniformly according to execution time, execution object, execution subject, and task category.

[0172] During the merging process, the task segments in the target handling path are used as the primary key. The on-site actions, verification actions, maintenance actions, and retesting actions belonging to the same task segment are aggregated into the same execution segment, and then a complete execution receipt is formed according to the order of the execution segments.

[0173] Each execution receipt should include at least the anomaly chain identifier, the target handling path identifier, the task execution order, the arrival time of each task, the verification conclusion, the maintenance action, the retest result, and the final recovery status.

[0174] For example, in the current scenario, at 09:25, a record of HVAC maintenance personnel arriving on-site is generated; at 09:26, a valve position feedback reading verification record is generated; at 09:27, a remote reset maintenance record is generated; at 09:29, a first retest record is generated; at 09:35, a record of the on-duty electrician arriving on-site is generated; at 09:37, a drive inspection verification record is generated; at 09:40, a drive parameter correction maintenance record is generated; at 09:43, a second retest record is generated; and at 09:47, a recovery record is generated. By merging these records according to the task order, a complete execution receipt corresponding to the current abnormal event can be formed.

[0175] After the execution receipt is generated, it is compared item by item with the target handling path to invert the root cause judgment bias, task sorting bias, and path switching bias, specifically including:

[0176] Root cause determination bias is used to determine whether the root cause initially selected in the target treatment path is consistent with the actual effective root cause. When making the determination, the assumed root cause corresponding to the priority path in the target treatment path can be compared with the maintenance object that actually produces an effective repair effect during the final recovery.

[0177] If the target handling path takes valve position jamming as the primary root cause, but the execution receipt shows that valve position reset did not bring effective improvement, and the indicators returned to normal after the drive parameters were corrected, then it is determined that there is a root cause judgment deviation. The deviation direction is that the valve position abnormality was misjudged as the drive abnormality and was not identified first.

[0178] Task sequencing deviation is used to determine whether the order of tasks in the current path is reasonable. This is done by comparing the critical verification and maintenance actions in the execution feedback that truly determine the subsequent recovery outcome with the original order in the target handling path. If the execution feedback indicates that the drive check should be performed first, followed by the valve position reset, for faster convergence, then there is a task sequencing deviation in the original target handling path.

[0179] The path switching deviation is used to determine whether the failure transfer condition is triggered too late, too early, or the switching direction is inappropriate. When making the determination, the preset failure transfer node of the target disposal path is compared with the actual optimal switching node.

[0180] If the target handling path continues to maintain the valve position path after the first retest failure, and only switches to the drive path after the second retest failure, while the execution receipt shows that switching immediately after the first retest failure would result in faster recovery, then it is determined that there is a path switching deviation.

[0181] To ensure that the deviation inversion process can be implemented directly, the criteria for determining the three types of deviations can be further clarified. The determination of root cause determination deviation is based on the causal consistency between the maintenance object corresponding to the actual recovery action, the maintenance action, and the recovery result. That is, when a maintenance action is implemented and a stable recovery occurs within a preset continuous time slice without recurrence, the object corresponding to the maintenance action is regarded as the actual effective root cause. If the actual effective root cause is inconsistent with the initial root cause assumption of the target treatment path, the root cause determination deviation is recorded.

[0182] The determination of task sequencing deviation is based on the pre-emptiveness of key actions. That is, if the verification or maintenance action that brings the first positive improvement in the execution feedback is moved to the earlier part of the target handling path, the number of invalid tasks can be significantly reduced or the recovery time can be shortened, then the task sequencing deviation is recorded.

[0183] The path switching deviation is determined based on the timing and direction of the switching. That is, if the current path continues to execute without producing a recovery effect, while another path has the conditions for switching at an earlier node, the path switching deviation is recorded. Through the above criteria, the deviation inversion can have a clear judgment standard, without relying on fuzzy experience judgment.

[0184] After the deviation type is determined, a learning and backfeed entry is generated. The learning and backfeed entry should at least include the deviation type, deviation location, anomaly chain identifier, target treatment path identifier, and actual recovery result. It can also further include the initial root cause pointer, the actual effective root cause pointer, the invalid task segment sequence, the effective task segment sequence, the actual switching node, and the recovery time interval.

[0185] Taking the above scenario as an example, the following can be recorded in the learning backfeed entry: the abnormal chain is identified as the A-05-W fresh air abnormal chain, the target treatment path is identified as the valve position priority path, the deviation type includes root cause judgment deviation, task sorting deviation and path switching deviation, the deviation positions are the initial root cause judgment node, the sorting node after the first retest and the switching node after the first retest, respectively, and the actual recovery result is that the driving parameters are corrected and the system returns to normal. The learning backfeed entry formed in this way can not only reflect the key dividing point between the failure and success of this treatment, but also serve as a direct sample for subsequent model updates.

[0186] During the model update phase, the scene matching relationship is first corrected by learning and refeeding back the entries, specifically including:

[0187] The scene matching relationship is updated based on the correspondence between the current anomaly chain and historical handling memory entries. If the learned and fed-back entries indicate that the current scene is closer to a historical driver anomaly recovery event than a valve position anomaly recovery event, then in subsequent matching, historical handling memory entries related to driver anomalies should be moved to a higher matching level, and memory entries related to valve position anomalies but with poor recovery effects should be moved to the back. Then, the script ordering relationship is corrected using the learned and fed-back entries.

[0188] The update object of the script sorting relationship is the arrangement order of the candidate disposal script set. If the current learning and backfeeding entry indicates that the driving anomaly handling script should be given priority in this type of scenario, then when encountering scenarios with the same anomaly chain structure, the same environmental deviation characteristics and the same meter power deviation characteristics in the future, the driving anomaly handling script should be output first, and then the valve position anomaly handling script should be output.

[0189] Finally, the failure transition conditions are corrected using the learned feedback entries. The failure transition conditions are updated based on the time window, the number of retest failures, and key metric boundaries required to switch from the current path to another path. For example, the condition that originally required switching to the drive path after two consecutive retest failures can be corrected to immediately switching to the drive path if the first retest fails and the power deviation segment persists, thereby reducing unnecessary waiting time.

[0190] To ensure the continuous operation of the update process, the learning backfeed entries can be written into the incremental sample pool according to the scene category. After reaching the preset number of samples or the preset time period, the self-learning inference model is updated incrementally. During the incremental update, the learning backfeed entries accumulated in the same scene are read first, and the scene matching relationship, script ranking relationship and failure transfer condition are grouped and corrected. Then the correction results are written back to the scene encoding layer and script generation layer of the self-learning inference model.

[0191] It should be noted that the triggering conditions for incremental updates include at least the following: the cumulative number of samples in the scene category reaches a preset number, the root cause judgment deviation occurs a preset number of times consecutively, the task ranking deviation occurs a preset number of times consecutively, and the path switching deviation occurs a preset number of times consecutively. After the incremental update is triggered, the learning backfeed entries under the same scene category are first read from the incremental sample pool, and then the scene matching relationship, script ranking relationship and failure transfer condition are respectively grouped and corrected, and the correction results are written back to the scene matching relationship table, script ranking relationship table and failure transfer condition table.

[0192] For frequently occurring similar abnormal scenarios, the version identifiers before and after the update can be retained to verify whether the model adjustment has truly shortened the recovery time, reduced invalid tasks, and improved the recovery success rate. Taking the above scenario as an example, in multiple consecutive fresh air abnormality events in the west office area, if the learned backflow items repeatedly indicate that the driving abnormal path is better than the valve position abnormal path, the model will automatically improve the ranking position of the driving check task in subsequent similar scenarios and trigger the path switching condition in advance.

[0193] Through the above process, the methods for forming execution receipts, judging the three types of deviations, writing learning feedback items, and updating the scenario matching relationship, script ordering relationship, and failure transfer conditions in the self-learning inference model are clearly given. Based on this, the reverse transmission of the handling results to the model's cognition can be directly completed, so that each real operation and maintenance process can become an effective update sample for subsequent autonomous learning inference, thereby gradually improving the handling accuracy and path arrangement rationality under complex abnormal scenarios in the park.

[0194] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0195] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0196] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0197] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0198] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart park operation and maintenance management method based on AI autonomous learning, characterized in that: The specific steps include: Collect data from building automation, security, access control, fire protection, energy consumption, environmental monitoring, as well as work orders, inspections, and maintenance. Organize the data according to a unified time scale and by region, object, and event identifier to generate a snapshot stream of park operations and maintenance. Extract abnormal symptom fragments corresponding to alarm mutations, operating condition deviations, repeated repair requests, and recovery failures from the park's operation and maintenance snapshot stream, and construct a composite abnormal evolution chain based on time connection, regional propagation, equipment dependence, and handling correlation. Retrieve historical handling memory entries that are formed by work orders, inspection and maintenance data and are isomorphic to or contained in sub-chains of the composite anomaly evolution chain, and call the self-learning inference model to output a set of candidate handling scripts that include the verification order, collaborative roles, task connection and failure transfer conditions; The permission status, non-stop operation requirements, resource consumption, and task dependencies are compiled into an execution constraint table. The candidate disposal script set is then trimmed and arranged to obtain the target disposal path. Based on the arrival, verification, repair, retesting and recovery records corresponding to the target handling path, an execution receipt is generated. The root cause judgment bias, task sorting bias and path switching bias are inverted, and learning feedback items are generated and the self-learning inference model is updated.

2. The smart park operation and maintenance management method based on AI autonomous learning according to claim 1, characterized in that: Organized according to a unified time scale and region, object, and event identifier, the generated park operation and maintenance snapshot stream includes: Determine the target time slice set, and perform time mapping on building automation data, security data, access control data, fire protection data, energy consumption data, environmental monitoring data, work order data, inspection data, and maintenance data according to the target time slice set to generate data entries aligned with the time slice; Based on the region identifier, object identifier, and event identifier, perform same-source merging and heterogeneous-source association on data entries of aligned time slices to generate park operation and maintenance snapshot items; The park operation and maintenance snapshot items are written into the same snapshot sequence in chronological order, and status field, alarm field, work order field, inspection field and maintenance field are written into each park operation and maintenance snapshot item to form a park operation and maintenance snapshot stream.

3. The smart park operation and maintenance management method based on AI autonomous learning according to claim 2, characterized in that: Performing same-source merging and heterogeneous-source association on data entries for aligned time slices based on region identifier, object identifier, and event identifier includes: Perform field deduplication and status merging on data entries from the same data source that have the same region identifier, object identifier, and event identifier to generate merged entries with the same source. Establish association edges for data entries from different data sources that have the same regional identifier, have a mapping relationship between object identifiers, or have overlapping event occurrence times, and generate heterogeneous association entries; Based on the same-source merged entries and heterogeneous-source associated entries, the corresponding park operation and maintenance snapshot item within each time slice is determined, and the park operation and maintenance snapshot item is written into the park operation and maintenance snapshot stream.

4. The smart park operation and maintenance management method based on AI autonomous learning according to claim 1, characterized in that: The abnormal symptom segments corresponding to alarm mutations, operational deviations, duplicate repair requests, and recovery failures extracted from the park's operation and maintenance snapshot stream include: Perform change detection on the status and alarm fields of the park operation and maintenance snapshot items within a continuous time slice, and generate alarm mutation fragments and operating condition deviation fragments based on the change direction, change duration and abnormal trigger boundary. Perform event merging on the work order field, inspection field and maintenance field within a continuous time slice, and generate duplicate repair fragments and recovery failure fragments based on the duplicate repair relationship and recovery loop failure relationship of the same object within a preset window. Alarm mutation fragments, operating condition deviation fragments, repeated repair fragments, and recovery failure fragments are written into the abnormal symptom fragment set according to time boundaries, area identifiers, and object identifiers.

5. The smart park operation and maintenance management method based on AI autonomous learning according to claim 4, characterized in that: A complex anomaly evolution chain is constructed based on temporal coherence, regional propagation, equipment dependence, and handling correlation, including: Establish time-connecting edges based on the chronological order and time intervals of abnormal symptom fragments; Establish regional propagation edges based on regional connectivity, regional adjacency, or energy supply coverage; Establish equipment dependency edges based on equipment control relationships, power supply and distribution relationships, or information transmission relationships; Establish a processing association edge based on the same work order link, continuous processing records, or consistent execution entity identifier; Based on time-connection edges, regional propagation edges, device-dependent edges, and handling-related edges, chain segment splicing is performed on abnormal symptom fragments to generate a composite abnormal evolution chain.

6. The smart park operation and maintenance management method based on AI autonomous learning according to claim 1, characterized in that: The historical handling memory entries retrieved from work orders, inspections, and maintenance data that are isomorphic to or contained within sub-chains of the complex anomaly evolution chain include: Based on work order data, inspection data, and maintenance data, extract anomaly chain identifiers, handling task sequences, execution role sequences, recovery results, and failure reasons to generate a historical handling memory entry set; The composite anomaly evolution chain is mapped to a chain feature sequence. Isomorphic matching and sub-chain inclusion matching are performed on the anomaly chain identifiers in the historical disposal memory entry set to generate candidate historical disposal memory entries. Candidate historical disposal memory entries are sorted according to the consistency of recovery results, the completeness of task sequence, and the timeliness of disposal to determine the set of historical disposal memory entries used for self-learning deduction.

7. A smart park operation and maintenance management method based on AI autonomous learning according to claim 6, characterized in that: The set of candidate disposal scenarios output by the self-learning inference model, which includes the verification order, collaborating roles, task connections, and failure transfer conditions, includes: Establish a scene coding subnet to map the complex anomaly evolution chain and the historical handling memory entry set into scene representation results; Establish a script generation subnet, which includes a check order output branch, a collaborative role output branch, a task connection output branch, and a failure transfer output branch. The output branches output the sequence of verification tasks based on the verification order, the set of collaborative roles based on the collaborative roles, the task sequence relationship based on the task connection, and the failure transfer conditions based on the failure transfer, thereby generating a set of candidate disposal scripts.

8. The smart park operation and maintenance management method based on AI autonomous learning according to claim 7, characterized in that: The permission status, non-stop operation requirements, resource consumption, and task dependencies are compiled into an execution constraint table, including: Extract area access permissions, equipment shutdown permits, personnel availability status, resource occupancy status, and task sequence dependencies to generate a set of constraints. The constraint set is merged according to region identifier, object identifier, role identifier, and task identifier to generate an execution constraint table; Based on the execution constraint table, the verification tasks, disposal tasks, retesting tasks, and failure transfer tasks in the candidate disposal script set are marked with executable and blocking tags.

9. A smart park operation and maintenance management method based on AI autonomous learning according to claim 8, characterized in that: The candidate disposal script set is trimmed and arranged to obtain the target disposal path, including: Remove task segments marked with blocking flags, retain task segments marked with executable flags, and generate candidate execution chains; The candidate execution chain is executed according to the priority of tasks, the availability of collaborative roles, and the failure transfer conditions to generate the target disposal path. Generate on-site records, verification records, maintenance records, retest records, and recovery records according to the target handling path.

10. A smart park operation and maintenance management method based on AI autonomous learning according to claim 9, characterized in that: Based on the arrival records, verification records, maintenance records, retest records, and recovery records corresponding to the target handling path, an execution receipt is generated. Root cause determination bias, task sorting bias, and path switching bias are inverted. Learning feedback items are generated and the self-learning inference model is updated, including: The execution sequence of on-site records, verification records, maintenance records, retest records, and recovery records is merged to generate an execution receipt; The root cause determination bias, task sorting bias, and path switching bias are determined based on the differences between the execution receipt and the target handling path; Write the deviation type, deviation location, anomaly chain identifier, target handling path identifier, and actual recovery result into the learning and backfeeding entry; The scene matching relationship, script ranking relationship, and failure transition condition in the self-learning inference model are updated based on the learned feedback entries.

Images