Infrastructure unmanned aerial vehicle inspection multi-agent operation platform and method

Abstract

The application discloses an infrastructure unmanned aerial vehicle (UAV) inspection multi-agent operation platform and method, relates to the technical field of UAV inspection, and comprises the following: a platform core, which is used for providing multi-agent operation scheduling and long-term problem object management, the object being independent of a single inspection operation instance; a sensing boundary, which is used for converting multi-source original data into standardized evidence units, observation objects and evidence packages; a scene adaptation rule package, which is used for loading classification, risk, inspection points and other rules for different industry scenes; and a plurality of functional agents, which generate evaluation results by processing the evidence packages under the scheduling of the platform core, implement creating, adding, updating, reopening or closing operations on the long-term problem objects according to the matching relationship between the evaluation results and the historical long-term problem objects, and further trigger downstream disposal actions. Through long-term problem object evolution across operation instances and decoupling architecture, the application solves the problems of serious scene coupling, short-term results and difficulty in cross-industry reuse of the existing system.

Description

Technical Field

[0001] This invention relates to the fields of drone inspection, artificial intelligence, multi-agent systems, intelligent infrastructure operation and maintenance, and inspection platform technology. More specifically, it relates to a drone inspection multi-agent operation platform for infrastructure scenarios and its problem object-driven closed-loop operation method. Background Technology

[0002] With the continuous expansion of infrastructure such as power grids, energy, water utilities, highways and bridges, and rail transit, drone-based inspection methods are gradually replacing some manual inspection methods. Existing technologies can generally achieve functions such as flight path planning, automatic flight, image acquisition, video transmission, target recognition, defect detection, remote monitoring, and report generation. However, existing inspection platforms as a whole still mainly follow the technical path of traditional SaaS inspection systems, which are centered on task forms, manual workflows, static rules, and result reports.

[0003] The existing technology has at least the following problems: 1. Lack of a unified platform kernel across different industry scenarios. Existing systems are typically built around single scenarios such as bridges, power grids, highways, or industrial parks. Scenario rules, object models, and processing logic are tightly coupled, making it difficult to support cross-industry reuse and continuous expansion.

[0004] 2. Most existing systems directly identify and judge based on raw images, videos, point clouds, or logs, lacking a unified perception boundary and structured observation object definition, which makes it difficult to standardize subsequent reasoning, review, approval, work order generation, and auditing.

[0005] 3. Existing systems mostly treat inspection results as the result of a single task or a one-time alarm, lacking long-term problem objects independent of a single inspection instance. This makes it difficult to support additional evidence, status updates, problem reopening, problem closure, and historical tracing for the same physical problem in multiple inspections.

[0006] 4. The artificial intelligence capabilities in existing systems are mostly limited to single-model recognition, single-step reasoning, or result summarization. They lack a multi-agent operation mechanism built around the inspection business closed loop, and cannot form a unified dynamic closed loop from evidence processing, anomaly assessment, rule matching, human responsibility nodes to deliverable generation.

[0007] Therefore, a new platform technology is needed to establish a unified platform core, perception boundary, scenario adaptation mechanism, and long-term problem object evolution mechanism in infrastructure inspection scenarios, thereby replacing and upgrading the capabilities of traditional inspection SaaS platforms. Summary of the Invention

[0008] To overcome the shortcomings of the existing technologies, this invention discloses a multi-agent operation platform and method for infrastructure drone inspection. By constructing a platform core, perception boundary, and scenario adaptation rule package, and introducing a long-term problem object evolution mechanism centered on Case and a human-machine shared responsibility closed-loop link for multi-agent collaborative orchestration, this invention addresses the technical problems of severe scenario coupling, short-term results, fragmented closed-loop chains, and difficulty in cross-industry reuse in existing inspection systems. This enables unified platform-based operation and maintenance, long-term problem tracking and governance, and standardized and auditable closed-loop processes from evidence processing to disposal and delivery across multiple industry scenarios.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a multi-agent operation platform for unmanned aerial vehicle (UAV) inspection of infrastructure, comprising: The core of the platform is used to provide multi-agent operation scheduling, lifecycle management, long-term problem object management, checkpoint control, approval control, deliverable assembly, audit records and platform interface capabilities. The long-term problem objects exist independently of a single inspection operation instance and are used to carry the continuous evolution closed loop of the same infrastructure problem in different times and different tasks. Perception boundary is used to convert multi-source raw data collected by UAV inspections into standardized evidence units, structured observation objects, and evidence packages; The scenario adaptation rule package is used to load object classification rules, anomaly classification rules, risk rules, checkpoint rules, supplementary evidence collection templates, report templates, capability binding configurations, and intelligent agent execution graph configurations for different infrastructure industry scenarios. And multiple functional intelligent agents, under the core scheduling of the platform, perform at least one of the following operations on the evidence package based on the scenario adaptation rule package: evidence processing, anomaly assessment, reasoning, rule matching, audit routing, task orchestration and delivery generation, generate assessment results, and perform creation, appending, updating, reopening or closing operations on the long-term problem objects according to the matching relationship between the assessment results and the long-term problem objects, thereby triggering downstream handling actions.

[0010] Preferably, the platform adopts a layered architecture, which includes: The infrastructure industry scenario layer is used to support at least one infrastructure industry scenario from power grids, energy, water utilities, highways and bridges, or rail transit. The Domain Pack layer, which is the carrier layer of the scenario adaptation rule package, is used to inject object classification rules, anomaly classification rules, risk rules, checkpoint rules, supplementary evidence collection templates, report templates, capability binding configurations, and agent execution graph configurations. The Perception Boundary layer, which is the implementation layer of the perception boundary, is used to convert multi-source raw inspection data into evidence units, observation objects, and evidence packages. The Platform Core layer, which is the implementation layer of the platform core, is used to carry out multi-agent operation scheduling, lifecycle management, long-term problem object management, checkpoint control, approval control, deliverable assembly, audit records and platform interface capabilities. The functional agent layer includes at least an evidence analysis agent, an anomaly detection agent, a reasoning agent, a route review agent, and a task orchestration agent. A closed-loop delivery layer is used to drive the generation and flow of checkpoints, approvals, work order candidates, and deliverables by the long-term problem objects.

[0011] Preferably, the platform internally defines the following core objects: AssetRecord is used to represent the infrastructure assets being inspected; ComponentNode is used to represent the component node within the asset; MissionRecord is used to represent inspection tasks; RunRecord is used to represent a specific instance of an inspection run. EvidenceUnit is used to represent a standard evidence unit after access and cleansing. Observation is used to represent structured observation objects; Assessment is used to represent the evaluation results generated after the operation of multiple agents; The Case, or long-term problem object, is used to represent a long-term evolution object for infrastructure problems. The Case exists independently of the RunRecord and is created, appended, updated, reopened, or closed by new EvidenceUnits, Observations, and Assessments generated by the RunRecord. Checkpoint and Signoff are used to represent human responsibility nodes; WorkOrderCandidate and DeliveryArtifact are used to represent work order candidates and deliverables derived from the convergence of the issue object.

[0012] Preferably, the method by which the perception boundary converts the multi-source raw data into standardized evidence units, structured observation objects, and evidence packages includes: first, accessing and cleaning the raw data to generate standardized evidence units; then, performing spatiotemporal alignment, object association, and feature extraction on the evidence units to generate structured observation objects; and finally, aggregating at least one observation object targeting the same asset or the same problem into an evidence package for use by the functional agent.

[0013] Preferably, the plurality of functional intelligent agents include at least: An intelligent agent for evidence analysis, used for evidence cleaning, evidence aggregation, and evidence correlation; An anomaly analysis agent is used to output the anomaly type, risk level, and confidence level. Based on the reasoning agent, it is used to generate rule bases, historical comparisons, and explanatory conclusions; The intelligent agent that reviews the routing is used to determine whether to proceed to manual review, supplementary evidence collection, or approval. A task orchestration agent is used to drive the generation of work order candidates, reports, alarms, and downstream deliverables; Under the constraints of the platform's core, each functional intelligent agent collaborates around the same evidence package and its matched or newly created long-term problem objects.

[0014] Preferably, the scenario adaptation rule package is configured to be instantiated for at least one industry scenario among power grid inspection, energy facility inspection, water facility inspection, highway bridge inspection, or rail transit inspection; when switching industry scenarios, only the corresponding scenario adaptation rule package is replaced or added, while the main structure of the platform core, perception boundary, and functional intelligent agent remains unchanged.

[0015] Secondly, the present invention also provides a multi-agent operation method for unmanned aerial vehicle (UAV) inspection of infrastructure, characterized by comprising the following steps: Acquire multi-source raw inspection data generated by drone inspections; The multi-source raw inspection data is transformed into standardized evidence units, structured observation objects, and evidence packages by sensing boundaries; Under the core scheduling of the platform, multiple functional intelligent agents are driven to process the evidence package according to the scenario-adaptive rule package, and an evaluation result containing anomaly type, risk level, confidence level and rule basis is generated; Based on the matching relationship between the evaluation results and historical long-term problem objects, perform creation, appending, updating, reopening, or closing operations on the long-term problem objects; Based on the status, confidence level, risk level, and rule constraints of long-term problem objects, trigger one or more downstream actions among supplementary sampling and re-flight, manual review, supplementary evidence collection and confirmation, signature confirmation, continued observation, planned disposal, immediate disposal, work order candidate generation, report generation, and alarm push. All objects throughout the process are stored in an associated manner to form a closed-loop operation record that is traceable, replayable, and auditable.

[0016] Preferably, the step of performing creation, appending, updating, reopening, or closing operations on long-term problem objects based on the matching relationship between the evaluation results and historical long-term problem objects specifically includes: When the evaluation results cannot match any historical long-term problem objects, create a new long-term problem object; When the evaluation result can match one or more historical long-term problem objects, the current evaluation result and the corresponding evidence reference are added to the matched long-term problem object, and the life cycle status of the long-term problem object is updated according to the preset rules. When a closed long-term issue object is matched again and evidence indicates that the issue has recurred, the long-term issue object is automatically reopened. When a long-term problem object meets the preset convergence criteria, it is closed and archived.

[0017] Preferably, the step of triggering downstream handling actions based on the status, confidence level, risk level, and rule constraints of long-term problem objects further includes: when the rules determine that the current problem requires manual intervention, the platform core automatically generates checkpoints and pushes them to the review node, and after the review is completed, it enters the signature and confirmation stage; if the problem meets the handling conditions, work order candidates and deliverables are further generated, and the deliverables include inspection reports, alarm information, or handling suggestions.

[0018] Preferably, the collaboration among the multiple functional agents is constrained by the platform's core agent registration, capability binding, execution graph, and lifecycle scheduling mechanism; wherein, the execution graph defines the calling order and data flow direction among the evidence analysis agent, anomaly judgment agent, basis reasoning agent, audit routing agent, and task orchestration agent, ensuring that the same evidence package and the same long-term problem object do not deviate from the unified operating context of the platform core when flowing between different agents.

[0019] The beneficial effects of this invention are: 1. By decoupling the platform core from the scenario rule package, the coupling degree between industry scenarios and the platform base is reduced, making it easier to reuse across industries; 2. By unifying the entry point for raw inspection data through the perception boundary, the standardization of evidence processing, evaluation, review, and approval has been improved; 3. By abstracting the inspection problem into a long-term problem object Case, the same physical problem can be continuously evolved and tracked across multiple Runs; 4. By using a multi-agent operation mechanism, evidence processing, anomaly detection, reasoning, audit routing, and task orchestration are integrated into a unified platform operation framework; 5. Through checkpoint and signoff mechanisms, the ability to share responsibility between humans and machines and the auditing capabilities in the handling of high-risk issues have been enhanced; 6. Through the downstream derivation mechanism of work order candidates and deliverables, the platform is equipped with dynamic closed-loop execution capabilities to replace traditional inspection SaaS platforms. Attached Figure Description

[0020] Figure 1 This is a layered framework diagram of the platform of the present invention, illustrating the relationship between the infrastructure industry scenario, Domain Pack, PerceptionBoundary, Platform Core, multi-agent layer, and closed-loop delivery layer. Figure 2 This is the core object relationship diagram of the present invention, illustrating the relationships between AssetRecord, ComponentNode, MissionRecord, RunRecord, EvidenceUnit, Observation, Assessment, Case, Checkpoint, and deliverable objects; Figure 3 This is an end-to-end technical flowchart of the present invention, illustrating the overall process of multi-source raw inspection data passing through the perception boundary, multi-agent operation platform, case evolution, and delivery closed loop. Detailed Implementation

[0021] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with the embodiments and accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention.

[0022] Example 1 A multi-agent operation platform and method for unmanned aerial vehicle (UAV) inspection of infrastructure is shown below: I. Overall Plan This invention constructs a multi-agent operating platform consisting of a platform core, a perception boundary, and a domain pack for scene adaptation rules.

[0023] in: 1. The core of the platform is used to provide unified operational lifecycle management, problem object management, checkpoint control, approval control, deliverable assembly, audit records, and platform interface capabilities.

[0024] 2. The perception boundary is used to convert the multi-source raw data collected during the UAV inspection process into standardized evidence units (EvidenceUnit), structured observation objects (Observation), and evidence packs (EvidencePack), so that subsequent judgments no longer directly depend on the original evidence input.

[0025] 3. The scenario adaptation rule package is used to load object classification rules, anomaly classification rules, risk rules, checkpoint rules, supplementary evidence collection templates, report templates, capability binding configurations, and agent execution graph configurations for different infrastructure industry scenarios.

[0026] Under the core scheduling of the platform, multiple functional intelligent agents perform evidence processing, anomaly assessment, reasoning, rule matching, audit routing, task orchestration, and delivery generation around the evidence package to form an assessment result, and drive the creation, addition, update, reopening, or closure of long-term problem object cases based on the matching relationship between the assessment and historical cases.

[0027] II. Platform Layered Structure like Figure 1 As shown, the platform structure of the present invention includes: Infrastructure industry scenario layer: used to support scenarios such as power grids, energy, water affairs, highways and bridges, and rail transit; Domain Pack layer: used to inject object classification rules, anomaly classification rules, risk rules, checkpoint rules, supplementary evidence templates, report templates, capability binding configurations, and agent execution graph configurations; The Perception Boundary layer is used to convert raw inspection data from multiple sources into EvidenceUnits, Observations, and EvidencePacks. Platform Core layer: Used to handle multi-agent operation scheduling, lifecycle management, long-term problem object management, checkpoint control, approval control, deliverable assembly, audit logging, and platform interface capabilities; Functional agent layer: includes at least evidence analysis agent, anomaly detection agent, reasoning agent, route review agent, and task orchestration agent; Closed-loop delivery layer: The generation and flow of Checkpoint, Signoff, WorkOrderCandidate, and DeliveryArtifact are driven by Case.

[0028] III. Core Object System like Figure 2 As shown, the following key objects are defined internally within the platform of this invention: 1. AssetRecord: Represents the infrastructure asset being inspected; 2. ComponentNode: Represents a component node within the asset; 3. MissionRecord: Represents an inspection task; 4. RunRecord: Represents a specific instance of an actual inspection run; 5. Evidence Unit: Represents a standard unit of evidence after accession and cleansing; 6. Observation: Represents a structured observation object; 7. Assessment: Represents the evaluation result after the multi-agent operation; 8. Case: Indicates a long-term problem object; 9. Checkpoint / Signoff: Indicates a human responsibility point; 10. WorkOrderCandidate / DeliveryArtifact: Represents the work order candidates and deliverables derived from the convergence of the problem object.

[0029] Here, Case is not a sub-object of a single Run, but a long-term evolution object for infrastructure problems; Run is responsible for generating new EvidenceUnits, Observations, and Assessments, while Case is responsible for carrying the continuous evolution loop of the problem itself.

[0030] IV. End-to-end operation process like Figure 3 As shown, the operation flow of the present invention includes: 1. Acquire multi-source raw inspection data generated by drone inspections; 2. Standardize the multi-source raw inspection data by sensing the boundary to generate EvidenceUnit, Observation, and EvidencePack; 3. Under the platform's core scheduling, multiple functional agents are driven to process the EvidencePack according to the scenario adaptation rule package and agent execution graph to generate an Assessment; 4. Based on the matching relationship between Assessment and historical Cases, perform creation, append, update, reopen, and close operations on Cases; 5. Based on the case's status, confidence level, risk level, and rule constraints, trigger one or more of the following actions: re-sample collection and re-testing, manual review, supplementary evidence collection and confirmation, approval confirmation, continued observation, planned handling, immediate handling, work order candidate generation, report generation, and alarm push. 6. Store all process objects in an associated manner to form a closed-loop operation record that is traceable, replayable, and auditable.

[0031] V. Cross-Run Problem Object Evolution Mechanism The key to this invention lies in abstracting inspection problems into long-term problem objects (Case), rather than treating each inspection result as an isolated conclusion. For problems that recur in different times, tasks, and operational instances of the same infrastructure asset or component node, the platform can perform the following operations based on the matching relationship between historical problem objects and current assessment results: Create a new case; Add new assessments and evidence citations to existing cases; Update the lifecycle status of an existing case; Reopen a closed case; Close the case that meets the convergence criteria.

[0032] As a result, the platform can form a continuous closed loop for long-term problem governance, rather than remaining in the single-task processing mode of traditional SaaS inspection platforms.

[0033] VI. Multi-agent operation mechanism The multi-agent operation mechanism in this invention is not a free-negotiation autonomous system, but a controllable orchestration mechanism under the constraints of the platform core. The platform core enables different functional agents to work collaboratively around the same EvidencePack and its matched or newly created Cases through agent registration, capability binding, execution graphs, and lifecycle scheduling.

[0034] The multi-agent system includes at least: Evidence analysis intelligent agents: used for evidence cleaning, evidence aggregation, and evidence correlation; Anomaly analysis agent: used to output anomaly type, risk level, and confidence level; Based on the reasoning agent: used to generate rule bases, historical comparisons, and interpretive conclusions; Review routing agent: Used to determine whether to proceed to manual review, supplementary evidence collection, or approval path; Task orchestration agent: used to drive the generation of work order candidates, reports, alarms, and downstream deliverables.

[0035] VII. Scene Adaptation Rule Package To enable the platform to scale across industries, this invention introduces the Domain Pack mechanism. Different infrastructure industry scenarios are adapted simply by loading different rule packages, without changing the platform's core structure.

[0036] The rule package can be applied to: Power grid inspection scenario; Energy facility inspection scenario; Scene of water facility inspection; Scenes of highway and bridge inspection; Inspection scene of rail transit.

[0037] Furthermore, the platform can be extended to adjacent scenarios such as forest and grassland patrol, but the core of the main platform is still built around infrastructure asset objects and component node objects.

[0038] Example 2 In this embodiment, taking the transmission line inspection scenario as an example, the platform first establishes an AssetRecord for the target line asset and a ComponentNode for objects such as insulators, conductors, and towers. The system receives images, videos, location, time, and environmental information collected by the UAV during an inspection run, converts them into EvidenceUnits through the perception boundary, and further generates Observations and EvidencePacks.

[0039] Subsequently, the platform's core system invokes the evidence analysis agent, anomaly detection agent, and reasoning agent to process the EvidencePack, generating an Assessment that includes the anomaly type, risk level, confidence level, and rule basis. If the Assessment matches a historical Case, the existing Case is updated; otherwise, a new Case is created.

[0040] When the platform determines that a problem requires manual intervention based on its rules, it automatically generates a Checkpoint and pushes it to the review node; after review, it enters the Signoff stage. If the problem meets the handling conditions, it further generates a WorkOrderCandidate and DeliveryArtifact, such as an inspection report, alarm information, or handling suggestions. All objects and status transitions in the process are recorded, thus forming a complete audit chain.

[0041] Taking bridge inspection, water facility inspection, or rail transit inspection scenarios as examples, only the object classification rules, anomaly classification rules, risk rules, and delivery templates in the Domain Pack need to be replaced, without changing the platform's core object chain and operating mechanism.

[0042] The embodiments of the present invention have been described in detail above, but the present invention is not limited to the described embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalents or substitutions are all included within the scope defined by the claims of the present invention.

Claims

1. A multi-agent operation platform for unmanned aerial vehicle (UAV) inspection of infrastructure, characterized in that, include: The core of the platform is used to provide multi-agent operation scheduling, lifecycle management, long-term problem object management, checkpoint control, approval control, deliverable assembly, audit records and platform interface capabilities. The long-term problem objects exist independently of a single inspection operation instance and are used to carry the continuous evolution closed loop of the same infrastructure problem in different times and different tasks. Perception boundary is used to convert multi-source raw data collected by UAV inspections into standardized evidence units, structured observation objects, and evidence packages; The scenario adaptation rule package is used to load object classification rules, anomaly classification rules, risk rules, checkpoint rules, supplementary evidence collection templates, report templates, capability binding configurations, and intelligent agent execution graph configurations for different infrastructure industry scenarios. And multiple functional intelligent agents, under the core scheduling of the platform, perform at least one of the following operations on the evidence package based on the scenario adaptation rule package: evidence processing, anomaly assessment, reasoning, rule matching, audit routing, task orchestration and delivery generation, generate assessment results, and perform creation, appending, updating, reopening or closing operations on the long-term problem objects according to the matching relationship between the assessment results and the long-term problem objects, thereby triggering downstream handling actions.

2. The multi-agent operation platform for infrastructure drone inspection as described in claim 1, characterized in that, The platform adopts a layered architecture, which includes: The infrastructure industry scenario layer is used to support at least one infrastructure industry scenario from power grids, energy, water utilities, highways and bridges, or rail transit. The Domain Pack layer, which is the carrier layer of the scenario adaptation rule package, is used to inject object classification rules, anomaly classification rules, risk rules, checkpoint rules, supplementary evidence collection templates, report templates, capability binding configurations, and agent execution graph configurations. The Perception Boundary layer, which is the implementation layer of the perception boundary, is used to convert multi-source raw inspection data into evidence units, observation objects, and evidence packages. The Platform Core layer, which is the implementation layer of the platform core, is used to carry out multi-agent operation scheduling, lifecycle management, long-term problem object management, checkpoint control, approval control, deliverable assembly, audit records and platform interface capabilities. The functional agent layer includes at least an evidence analysis agent, an anomaly detection agent, a reasoning agent, a route review agent, and a task orchestration agent. A closed-loop delivery layer is used to drive the generation and flow of checkpoints, approvals, work order candidates, and deliverables by the long-term problem objects.

3. The multi-agent operation platform for infrastructure drone inspection as described in claim 1, characterized in that, The platform internally defines the following core objects: AssetRecord is used to represent the infrastructure assets being inspected; ComponentNode is used to represent the component node within the asset; MissionRecord is used to represent inspection tasks; RunRecord is used to represent a specific instance of an inspection run. EvidenceUnit is used to represent a standard evidence unit after access and cleansing. Observation is used to represent structured observation objects; Assessment is used to represent the evaluation results generated after the operation of multiple agents; The Case, or long-term problem object, is used to represent a long-term evolution object for infrastructure problems. The Case exists independently of the RunRecord and is created, appended, updated, reopened, or closed by new EvidenceUnits, Observations, and Assessments generated by the RunRecord. Checkpoint and Signoff are used to represent human responsibility nodes; WorkOrderCandidate and DeliveryArtifact are used to represent work order candidates and deliverables derived from the convergence of the issue object.

4. The multi-agent operation platform for infrastructure drone inspection as described in claim 1, characterized in that, The perceptual boundary transforms the multi-source raw data into standardized evidence units, structured observation objects, and evidence packages in the following ways: first, the raw data is accessed and cleaned to generate standardized evidence units; then, the evidence units are spatiotemporally aligned, associated with objects, and feature extracted to generate structured observation objects; finally, at least one observation object targeting the same asset or the same problem is aggregated into an evidence package for use by the functional agent.

5. The multi-agent operation platform for infrastructure drone inspection as described in claim 1, characterized in that, The plurality of functional agents includes at least: An intelligent agent for evidence analysis, used for evidence cleaning, evidence aggregation, and evidence correlation; An anomaly analysis agent is used to output the anomaly type, risk level, and confidence level. Based on the reasoning agent, it is used to generate rule bases, historical comparisons, and explanatory conclusions; The intelligent agent that reviews the routing is used to determine whether to proceed to manual review, supplementary evidence collection, or approval. A task orchestration agent is used to drive the generation of work order candidates, reports, alarms, and downstream deliverables; Under the constraints of the platform's core, each functional intelligent agent collaborates around the same evidence package and its matched or newly created long-term problem objects.

6. The multi-agent operation platform for infrastructure drone inspection as described in claim 1, characterized in that, The scenario adaptation rule package is configured to be instantiated for at least one industry scenario, such as power grid inspection, energy facility inspection, water facility inspection, highway bridge inspection, or rail transit inspection. When switching industry scenarios, only the corresponding scenario adaptation rule package is replaced or added, while the main structure of the platform core, perception boundary, and functional intelligent agent remains unchanged.

7. A multi-agent operation method for unmanned aerial vehicle (UAV) inspection of infrastructure, characterized in that, Includes the following steps: Acquire multi-source raw inspection data generated by drone inspections; The multi-source raw inspection data is transformed into standardized evidence units, structured observation objects, and evidence packages by sensing boundaries; Under the core scheduling of the platform, multiple functional intelligent agents are driven to process the evidence package according to the scenario-adaptive rule package, and an evaluation result containing anomaly type, risk level, confidence level and rule basis is generated; Based on the matching relationship between the evaluation results and historical long-term problem objects, perform creation, appending, updating, reopening, or closing operations on the long-term problem objects; Based on the status, confidence level, risk level, and rule constraints of long-term problem objects, trigger one or more downstream actions among supplementary sampling and re-flight, manual review, supplementary evidence collection and confirmation, signature confirmation, continued observation, planned disposal, immediate disposal, work order candidate generation, report generation, and alarm push. All objects throughout the process are stored in an associated manner to form a closed-loop operation record that is traceable, replayable, and auditable.

8. The multi-agent operation method for infrastructure unmanned aerial vehicle (UAV) inspection as described in claim 7, characterized in that, The specific steps of performing creation, appending, updating, reopening, or closing operations on long-term problem objects based on the matching relationship between the evaluation results and historical long-term problem objects include: When the evaluation results cannot match any historical long-term problem objects, create a new long-term problem object; When the evaluation result can match one or more historical long-term problem objects, the current evaluation result and the corresponding evidence reference are added to the matched long-term problem object, and the life cycle status of the long-term problem object is updated according to the preset rules. When a closed long-term issue object is matched again and evidence indicates that the issue has recurred, the long-term issue object is automatically reopened. When a long-term problem object meets the preset convergence criteria, it is closed and archived.

9. A multi-agent operation method for infrastructure unmanned aerial vehicle (UAV) inspection as described in claim 7, characterized in that, The process of triggering downstream actions based on the status, confidence level, risk level, and rule constraints of long-term problem objects also includes: when the rules determine that the current problem requires manual intervention, the platform core automatically generates checkpoints and pushes them to the review node. After review, the checkpoints enter the approval and confirmation stage. If the problem meets the handling conditions, work order candidates and deliverables are further generated. The deliverables include inspection reports, alarm information, or handling suggestions.

10. A multi-agent operation method for infrastructure unmanned aerial vehicle (UAV) inspection as described in claim 7, characterized in that, The collaboration among the multiple functional agents is constrained by the platform's core agent registration, capability binding, execution graph, and lifecycle scheduling mechanism. The execution graph defines the calling order and data flow direction among the evidence analysis agent, anomaly judgment agent, reasoning agent, audit routing agent, and task orchestration agent, ensuring that the same evidence package and the same long-term problem object do not deviate from the unified operating context of the platform's core when flowing between different agents.

Images