Multi-agent autonomous inspection closed-loop control method and system based on open architecture

Abstract

The application provides a multi-agent autonomous inspection closed-loop control method and system based on an open architecture, relates to the technical field of intelligent operation and maintenance of power systems, and comprises the following steps: acquiring multi-source heterogeneous sensing data in an inspection area, generating global situation information through space-time unified mapping, generating a candidate inspection operation scheme, simulating and deducing the instruction sequence in the candidate scheme in a digital equivalent model that is kept synchronous with the operation state of the inspected entity equipment to verify safety, determining the verified scheme as a to-be-executed instruction sequence, generating a decision control instruction based on an open multi-agent architecture, enabling a physical inspection terminal to execute the decision control instruction and collect execution feedback data, returning the execution feedback data to the digital equivalent model, and generating a structured inspection report. Through space-time benchmark unification, digital equivalent model forward verification, open multi-agent collaborative decision and closed-loop feedback, the application solves the problems of rigid strategy, closed decision and low efficiency in existing inspection technology.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power systems, and in particular to a multi-agent autonomous inspection closed-loop control method and system based on an open architecture. Background Technology

[0002] In the operation and maintenance management of power systems and industrial facilities, inspection is a crucial link in ensuring the safe operation of equipment. Traditional inspections mainly rely on regular manual inspections or automated systems based on fixed logic. Manual inspections have inherent drawbacks such as low efficiency, high labor intensity, easy omissions, and inability to respond to sudden anomalies in real time. Early automated inspection systems introduced robots, fixed cameras, and various sensors, significantly improving data acquisition capabilities. However, their control logic is still mainly based on preset rules and hard-coded scripts, leading to the following prominent problems in complex dynamic environments: First, the inspection strategy is rigid, making it difficult to adjust the inspection path and frequency according to changes in equipment status, environmental disturbances, or task priorities, lacking adaptive capabilities; Second, the decision-making process is closed, the system can only execute preset instructions, and cannot achieve end-to-end autonomous closed loop from anomaly perception to handling and control, with anomaly response relying on manual intervention; Third, multi-terminal collaboration is inefficient, there is a lack of a unified task coordination mechanism between different types of inspection equipment, centralized scheduling is prone to task conflicts and has poor scalability; Fourth, the perception model is statically deployed, unable to continuously evolve using cross-site operational data, resulting in insufficient generalization capabilities. In summary, existing inspection systems have significant shortcomings in terms of strategy flexibility, decision-making autonomy, system scalability, and the ability to continuously evolve models. Summary of the Invention

[0003] To address the shortcomings of existing inspection technologies in terms of strategy flexibility, decision autonomy, system scalability, and continuous model evolution, this invention provides a multi-agent autonomous inspection closed-loop control method and system based on an open architecture. This method enables dynamic strategy generation, autonomous decision execution, efficient multi-terminal collaboration, and continuous self-evolution, thereby improving the safety, efficiency, and intelligence of inspection operations.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a multi-agent autonomous inspection closed-loop control method based on an open architecture in its first aspect, comprising: acquiring multi-source heterogeneous sensing data within an inspection area and mapping it to a unified spatiotemporal coordinate system to generate global situational information with a consistent spatiotemporal reference; the multi-source heterogeneous sensing data includes at least the status parameters of the inspected physical equipment, environmental physical quantities, and image data; generating candidate inspection operation schemes based on the global situational information; simulating and deducing the instruction sequence in the candidate inspection operation schemes in a digital equivalent model that is synchronized with the operating status of the inspected physical equipment to verify its safety, and determining the candidate inspection operation schemes that pass the safety verification as the instruction sequence to be executed; combining the instruction sequence to be executed and the global situational information, generating decision control instructions based on an open multi-agent architecture, and enabling physical inspection terminals to execute the decision control instructions and collect execution feedback data; and transmitting the execution feedback data back to the digital equivalent model and automatically generating a structured inspection report.

[0005] The present invention provides a preferred embodiment in its first aspect, wherein the mapping to a unified spatiotemporal coordinate system to generate global situational information with a consistent spatiotemporal reference includes: constructing a data alignment pipeline integrating stream and batch processing, which receives real-time acquired data streams and historically stored data batches, and sequentially performs the following alignment processing flow on each data unit in both types of data: using a unified time reference, interpolation or timestamp matching methods are used to synchronize the data of the data unit at the time of acquisition to a preset time segment to complete time axis alignment; for the data unit after time axis alignment, using a unified spatial reference, a coordinate transformation matrix is ​​used to map the coordinates of the data unit to a preset spatial coordinate system to complete spatial coordinate mapping; and all data units after time axis alignment and spatial coordinate mapping are merged into global situational information in a unified format.

[0006] This solution supports efficient processing of both real-time data streams and historical data batches. By interpolating or matching timestamps for each data unit for time synchronization and spatial alignment through coordinate transformation, it ensures low latency, consistency, and traceability of global situational information, providing high-fidelity input data for subsequent strategy generation and simulation verification, thereby improving reliability and response speed.

[0007] The present invention provides a preferred embodiment in a first aspect, wherein generating a candidate inspection operation plan based on the global situation information includes: receiving the global situation information, which includes the latest status parameters, environmental physical quantities, and spatial location distribution of each device within the inspection area; receiving inspection rules configured by maintenance personnel through an interactive strategy editing interface based on the global situation information, wherein the rules include inspection paths, triggering conditions, and handling actions; and the effective scope of the inspection rules is limited according to the spatial location distribution in the global situation information, specifically including: binding the inspection rules to preset spatial area boundaries or device coordinate positions, wherein the inspection rules only take effect when the current position of the physical inspection terminal falls within the spatial area boundary or the distance to the device coordinate position is less than a preset threshold; the interactive strategy editing interface adopts a low-code graphical orchestration engine to generate a plan description through visual drag-and-drop operations, wherein the plan description includes at least a sequence of inspection path points, detection action instructions for each inspection point, and condition triggering logic; and the plan description is standardized and represented in a domain-specific language as the candidate inspection operation plan.

[0008] This solution effectively lowers the technical threshold for configuring inspection strategies, enabling maintenance personnel to flexibly customize inspection paths, triggering conditions, and handling actions through visual drag-and-drop operations without writing complex code. At the same time, it limits the scope of rule effectiveness to specific spatial areas or equipment coordinate locations, achieving refined spatial control of inspection tasks, avoiding invalid or redundant inspections, and further improving the flexibility and execution efficiency of the strategy.

[0009] The present invention provides a preferred embodiment in a first aspect, wherein the step of simulating and deducing the instruction sequence in the candidate inspection operation scheme to verify its safety in a digital equivalent model that is synchronized with the operating status of the inspected physical equipment includes: loading the instruction sequence in the candidate inspection operation scheme into the digital equivalent model; performing a forward-looking simulation of the instruction sequence in the digital equivalent model to generate simulation result data; the simulation result data includes the movement trajectory of each physical inspection terminal in the spatiotemporal coordinate system after each physical inspection terminal executes the instruction sequence and the evolution curve of the operating parameters of the inspected physical equipment over time; performing a safety analysis based on the simulation result data, the safety analysis including the following detection items: detecting whether there is spatiotemporal overlap and conflict between the movement trajectories of each physical inspection terminal, and detecting whether the evolution curve of the equipment operating parameters exceeds a preset safety threshold range; if all detection items meet the corresponding safety requirements, the candidate inspection operation scheme is determined to have passed the safety verification; otherwise, the candidate inspection operation scheme is rejected, and conflict information or over-limit information corresponding to the failed detection items is output.

[0010] This solution breaks down security verification into two independent dimensions: multi-terminal collaborative security and the operational security of the inspected equipment. The former ensures that the inspection terminals do not interfere with or collide with each other, while the latter ensures that inspection commands do not cause the equipment to enter a dangerous state. Through this dual detection mechanism, this invention eliminates potential risks before the strategy is executed, effectively improving the overall safety of inspection operations.

[0011] The present invention provides a preferred embodiment in its first aspect, which generates decision control instructions based on an open multi-agent architecture by combining the sequence of instructions to be executed and global situational information. The embodiment includes: constructing an open multi-agent architecture that supports dynamic registration of agent types and elastic expansion of the number of agents, and includes several command agents and execution agents, employing a command-execution decoupled collaborative mode. At least one command agent is responsible for global task decomposition and scheduling, and at least one execution agent is responsible for executing specific inspection or diagnostic tasks. The command agent receives global situational information and the sequence of instructions to be executed, decomposes the overall task into multiple sub-tasks, each sub-task carrying a list of required resources and spatial constraints. Simultaneously, a global task allocation table is created to record the allocation status of each sub-task and the responsible execution agent. The system analyzes the overall cost of executing a subtask. After the command agent issues a subtask announcement, each execution agent assesses the comprehensive cost of executing the subtask based on its own capabilities. This comprehensive cost is calculated by weighting the resource consumption cost corresponding to currently available resources, the movement cost corresponding to the path distance to the target area, and the confidence cost corresponding to the historical task execution success rate. Each execution agent reports its calculated comprehensive cost to the command agent. The command agent collects all reported comprehensive costs from the execution agents, selects the execution agent with the lowest comprehensive cost as the executor of the subtask, assigns the subtask to that agent, and updates the status of the corresponding subtask in the global task allocation table. The selected execution agent generates decision control instructions for the specific physical inspection terminal based on the subtask requirements.

[0012] In this scheme, the command agent is only responsible for task decomposition and allocation decisions, while the execution agent autonomously calculates the overall cost and participates in the best-choice competition. This mechanism avoids the risk of overload of the central node, while allowing multiple execution agents to calculate the overall cost in parallel, thus improving the overall efficiency of task allocation.

[0013] In a first aspect, the present invention provides a preferred solution that, by combining the sequence of instructions to be executed and global situational information, generates decision control instructions based on an open multi-agent architecture. The solution further includes a dynamic agent scheduling mechanism based on task load awareness, specifically comprising: each execution agent periodically reporting current task queue information to the command agent; the task queue information including at least: the number of subtasks to be processed in the queue, the estimated processing delay of the task at the head of the queue, and the remaining power and available bandwidth of the physical inspection terminal controlled by the agent; when allocating newly generated subtasks, the command agent, based on the information reported by each execution agent... The task queue information is used to calculate the overall load of each execution agent and prioritize the allocation of subtasks to the execution agent with the lowest overall load. The overall load is obtained by weighting the number of subtasks to be processed, the estimated processing latency, and the level of remaining resources. If the number of subtasks to be processed by an execution agent exceeds the preset task backlog limit, or the estimated processing latency of the task at the head of the queue exceeds the preset response time limit, the agent is determined to be in an overloaded state. The command agent selects subtasks that have not yet started execution from the task queue of the overloaded agent, reassigns them to other agents with the lowest overall load, and updates the global task allocation table.

[0014] This solution achieves dynamic load balancing among multiple agents, effectively avoiding response latency or resource deadlock caused by task backlog, and further enhancing resilience and real-time performance. When an agent or terminal malfunctions, the task migration mechanism ensures that the overall inspection task can continue to execute, thereby improving the system's robustness and fault tolerance.

[0015] The present invention provides a preferred embodiment in its first aspect, wherein the decision control command is generated through a hybrid decision mechanism that integrates a self-learning decision model and an experiential knowledge reasoning model. Specifically, the self-learning decision model, based on a multi-agent reinforcement learning algorithm, takes the currently updated global situation information as input and outputs multiple candidate control actions and an alarm priority evaluation value corresponding to each candidate action. The candidate control actions include executable operations for equipment anomalies, and the alarm priority evaluation value characterizes the urgency of the anomaly addressed by the action. The experiential knowledge reasoning model, based on a case-based reasoning expert system, uses key state parameters in the current global situation information as query conditions, retrieves successful handling cases from a historical case database whose similarity to the key state parameters exceeds a preset threshold, extracts the control action sequences used in the successful handling cases as recommended actions, and applies safety boundary constraints to the candidate control actions output by the self-learning decision model. The safety boundary constraints include equipment parameter threshold limits, operation sequence compliance rules, and spatiotemporal safety rules. The candidate control actions output by the self-learning decision model, the recommended actions output by the experiential knowledge reasoning model, and the safety boundary constraints are fused using a priority arbitration mechanism to obtain the final decision control command.

[0016] This solution balances strategy exploration capabilities in dynamic environments with security and compliance. Reinforcement learning enables the system to continuously optimize, while the expert system ensures that decisions do not exceed preset security boundaries. The two work together to solve the problem of low exploration efficiency of a single decision model under boundary constraints, achieving intelligent and safe autonomous control.

[0017] The present invention provides a preferred embodiment in a first aspect, wherein the step of enabling physical inspection terminals to execute the decision control instructions and collect execution feedback data includes: distributing the decision control instructions to the corresponding physical inspection terminals through a unified terminal scheduling interface; each physical inspection terminal generating execution feedback data after executing the instructions, the execution feedback data including at least the terminal execution status, anomaly detection results, and resource occupancy information; the execution feedback data being reported to the command-type intelligent agent in real time, for the command-type intelligent agent to dynamically adjust the allocation scheme of unassigned sub-tasks or re-plan the execution order of the remaining tasks based on the real-time status and unfinished tasks of each physical inspection terminal; and being stored in a local cache for subsequent transmission back to the digital equivalent model.

[0018] In this scheme, the command-type intelligent agent can dynamically adjust the allocation scheme of unassigned sub-tasks or re-plan the execution order of remaining tasks based on the real-time status and remaining tasks of each terminal, so as to respond quickly when anomalies or disturbances occur and ensure the continuity and completion of inspection tasks.

[0019] The present invention provides a preferred embodiment in a first aspect, wherein the method further includes a continuous evolution step based on edge incremental learning and cloud federated aggregation, specifically including: using image data and state parameters in the execution feedback data as new training samples to incrementally update the parameters of the self-learning decision model; the incremental parameter update is performed on edge computing nodes deployed locally in the inspection area, the edge computing nodes and the physical inspection terminal being in the same network domain; multiple edge computing nodes deployed in different inspection areas respectively encrypt and upload their updated self-learning decision model parameters to a cloud server; the cloud server uses a federated averaging algorithm to aggregate the model parameters uploaded by each edge computing node, generates a globally updated self-learning decision model, and then distributes it to each edge computing node; each edge computing node replaces its local self-learning decision model with the received globally updated self-learning decision model and uses it for subsequent generation of decision control instructions.

[0020] This solution enables cross-site collaborative model training without leaking the original data, solving the problem of weak generalization ability caused by insufficient samples at a single site. At the same time, incremental updates reduce communication overhead and computational load, enabling the model to continuously improve and evolve with actual operating data, adapting to the long-term deployment needs of different inspection scenarios.

[0021] In a second aspect, this invention provides a multi-agent autonomous inspection closed-loop control system based on an open architecture, used to execute the method, comprising: a perception data acquisition and spatiotemporal alignment module, used to acquire multi-source heterogeneous perception data within the inspection area and map it to a unified spatiotemporal coordinate system to generate global situational information with a consistent spatiotemporal reference; the multi-source heterogeneous perception data includes at least equipment status parameters, environmental physical quantities, and image data; a strategy generation and verification module, used to generate candidate inspection operation schemes based on the global situational information; in a digital equivalent model that keeps synchronized with the operating status of the inspected physical equipment, the instruction sequence in the candidate inspection operation scheme is simulated and deduced to verify its safety, and the candidate inspection operation scheme that passes the safety verification is determined as the instruction sequence to be executed; a multi-agent decision and execution module, used to combine the instruction sequence to be executed and the global situational information to generate decision control instructions based on an open multi-agent architecture, and to enable the physical inspection terminal to execute the decision control instructions and collect execution feedback data; and a closed-loop feedback and report generation module, used to transmit the execution feedback data back to the digital equivalent model and automatically generate a structured inspection report.

[0022] Compared with the prior art, the present invention has the following advantages: This invention first maps multi-source heterogeneous sensing data to a unified spatiotemporal reference to generate global situational information, providing a real-time and consistent environmental perception foundation for dynamic decision-making. Next, it simulates and extrapolates candidate solutions in a digital equivalent model synchronized with the inspected physical equipment to verify safety, transforming the rigid mode of traditional fixed preset scripts into a dynamic and verifiable strategy generation mode. This achieves flexible strategy arrangement and proactive risk avoidance while ensuring safety. Then, based on an open multi-agent architecture, it generates decision instructions and schedules physical terminals for execution, breaking the limitations of closed hard-coded logic and transforming it into a negotiable and scalable multi-agent closed-loop autonomous system. This enables the system to have autonomous decision-making, multi-terminal collaboration, and dynamic agent expansion capabilities, effectively improving decision autonomy and system scalability, as well as enhancing anomaly response capabilities and multi-device collaboration efficiency. Finally, it feeds back execution feedback data to the digital equivalent model and automatically generates structured reports, forming a complete closed loop of perception, verification, decision-making, execution, and feedback. This allows the digital equivalent model to be continuously corrected based on actual execution results, providing a data foundation for model evolution. Therefore, this invention achieves a unified approach to flexible and configurable strategies, autonomous decision-making, open and scalable systems, and continuous self-evolution of models, effectively addressing the shortcomings of traditional inspection technologies in terms of strategy flexibility, decision-making autonomy, system scalability, and continuous model evolution capabilities. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0024] Figure 1 This is a flowchart of the multi-agent autonomous inspection closed-loop control method based on an open architecture according to Embodiment 1 of the present invention; Figure 2 This is an architecture diagram of the multi-agent autonomous inspection closed-loop control system based on an open architecture, as described in Embodiment 1 of the present invention. Figure 3 This is a flowchart of the dynamic agent scheduling mechanism based on task load perception in the multi-agent autonomous inspection closed-loop control method based on open architecture in Embodiment 2 of the present invention. Figure 4 This is a flowchart of the continuous evolution steps based on edge incremental learning and cloud federated aggregation in the multi-agent autonomous inspection closed-loop control method based on an open architecture in Embodiment 3 of the present invention. Detailed Implementation

[0025] 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.

[0026] Example 1: Please refer to Figure 1 This embodiment provides a multi-agent autonomous inspection closed-loop control method based on an open architecture, mainly implemented through steps S1 to S4: S1. Acquire multi-source heterogeneous sensing data within the inspection area and map it to a unified spatiotemporal coordinate system to generate global situational information with a consistent spatiotemporal reference; the multi-source heterogeneous sensing data includes at least the status parameters of the inspected physical equipment, environmental physical quantities, and image data. S2. Based on the global situation information, generate candidate inspection operation plans; in a digital equivalent model that is synchronized with the operating status of the inspected physical equipment, simulate and deduce the instruction sequence in the candidate inspection operation plan to verify its security, and determine the candidate inspection operation plan that passes the security verification as the instruction sequence to be executed. S3. Combining the sequence of instructions to be executed and the global situation information, generate decision control instructions based on an open multi-agent architecture, and enable the physical inspection terminal to execute the decision control instructions and collect execution feedback data; S4. The execution feedback data is sent back to the digital equivalent model and a structured inspection report is automatically generated.

[0027] Please refer to Figure 2 Corresponding to the method described in this embodiment, a multi-agent autonomous inspection closed-loop control system based on an open architecture is provided for executing the above method, and mainly consists of the following parts: The sensing data acquisition and spatiotemporal alignment module 1 is used to acquire multi-source heterogeneous sensing data within the inspection area and map it to a unified spatiotemporal coordinate system to generate global situational information with a consistent spatiotemporal reference; the multi-source heterogeneous sensing data includes at least equipment status parameters, environmental physical quantities and image data. The strategy generation and verification module 2 is used to generate candidate inspection operation plans based on the global situation information; in a digital equivalent model that is synchronized with the operating status of the inspected physical equipment, the instruction sequence in the candidate inspection operation plan is simulated and deduced to verify its security, and the candidate inspection operation plan that passes the security verification is determined as the instruction sequence to be executed. The multi-agent decision-making and execution module 3 is used to combine the sequence of instructions to be executed and the global situation information to generate decision control instructions based on an open multi-agent architecture, and to enable the physical inspection terminal to execute the decision control instructions and collect execution feedback data. The closed-loop feedback and report generation module 4 is used to send the execution feedback data back to the digital equivalent model and automatically generate a structured inspection report.

[0028] The specific implementation process of steps S1 to S4 above will be described in detail below. The implementation process of each module of the corresponding system should be referred to in the corresponding step. This embodiment uses a 220kV substation as the inspection area as the application scenario to illustrate the complete implementation process of the present invention under a standard environment: S1. Acquire multi-source heterogeneous sensing data within the inspection area and map it to a unified spatiotemporal coordinate system to generate global situational information with a consistent spatiotemporal reference; the multi-source heterogeneous sensing data includes at least the status parameters of the inspected physical equipment, environmental physical quantities, and image data.

[0029] Furthermore, in step S1, acquiring multi-source heterogeneous sensing data within the inspection area is mainly achieved through the following sub-steps: S101. Deploy multiple sensing terminals, wherein the sensing terminals include at least a fixed camera device, a mobile inspection robot, and an environmental status sensor; S102. Collect the device status parameters, environmental physical quantities, and image data through the sensing terminal; S103. A protocol adaptation mechanism is adopted to complete the dynamic parsing and secure access of different industrial communication protocols. The protocol adaptation mechanism is based on the protocol feature code fingerprint library and dynamic syntax parsing tree, and supports zero-code dynamic loading of multiple industrial communication protocols.

[0030] Specifically, in this embodiment, multiple types of sensing terminals are deployed in the key equipment area within the switchyard, including: fixed infrared / visible light binocular cameras, wireless temperature, humidity, and vibration sensors, and autonomous mobile inspection robots (equipped with LiDAR and visible light cameras). Each sensing terminal communicates using multiple industrial communication protocols. Each sensing terminal continuously collects data according to its own sampling frequency and communication protocol, forming a multi-source heterogeneous sensing data stream.

[0031] As is understandable, industrial communication protocols refer to communication protocols used for industrial control and data acquisition. Common ones include Modbus TCP / RTU (serial and Ethernet), OPC UA (service-oriented cross-platform protocol), MQTT (lightweight publish / subscribe protocol), IEC 61850 (substation automation), and DNP3 (power system). These protocols differ significantly in data format, transmission method, and security mechanism, and traditional gateways require separate firmware flashing for each protocol. This embodiment employs a protocol adaptive mechanism to achieve dynamic parsing and secure access for different industrial communication protocols, supporting zero-code dynamic loading of multiple industrial communication protocols. It can automatically identify and load the parsing module for the protocol used by newly accessed devices without modifying, compiling, or restarting the edge gateway software, automatically identify the protocol type, and dynamically load the protocol parsing plugin at runtime, achieving plug-and-play communication access. The core of this protocol adaptive mechanism is a two-stage identification and parsing based on a protocol feature fingerprint database and a dynamic syntax parse tree.

[0032] The first stage involves protocol signature fingerprint matching, i.e., coarse identification. Each industrial protocol has a fixed sequence of characteristic bytes in its message header or at specific offset positions. A signature fingerprint database is pre-built to record the mapping relationship between protocol types and signatures. Then, signature extraction, fingerprint database construction, and real-time matching are performed sequentially to achieve protocol signature fingerprint matching. For Modbus TCP, the signature can be a fixed pattern of the transaction identifier and function code range in the first few bytes of the message; for OPC UA, the signature is the fixed string "OPC UA" in the HELO message; for MQTT, the signature is a fixed bit pattern in the first byte of the Connect message. These signatures, along with information such as protocol type, version, and default port, are encoded into a hash table or prefix tree to construct the fingerprint database for fast matching. When the edge gateway receives the first message from an unfamiliar device, it extracts several bytes from the message header and searches the fingerprint database for a match. If a match is successful, the protocol type is determined; otherwise, the process proceeds to the second stage. The second stage, Dynamic Syntax Tree (DSLT), or fine-grained parsing, is primarily applicable to certain proprietary or variant protocols lacking pre-defined signatures, or where signatures are encrypted / obfuscated. This stage gradually achieves DLT through message sampling, byte pattern clustering, and semantic verification. Specifically, message sampling involves collecting 10-20 sample messages sent by the device. Byte pattern clustering uses sequence alignment algorithms to identify constant and variable bytes at fixed positions within multiple messages. Constant byte fields may be protocol identifiers or command codes; variable bytes may be addresses or data payloads. Based on the clustering results, the hierarchical structure of the messages is derived, with each structure corresponding to a generative rule, forming a DLT. Semantic verification applies the generated DLT to new messages, checking the rationality of parsed fields, such as matching the length field with the data payload length and ensuring a correct checksum. If verification passes, the protocol's pattern is stored in a temporary fingerprint database for rapid matching in the future.

[0033] Furthermore, in step S1, the mapping to a unified spatiotemporal coordinate system to generate global situational information with a consistent spatiotemporal reference is achieved through the following sub-steps: S111. Construct a data alignment pipeline that integrates streaming and batch processing. This pipeline receives real-time acquired data streams and historically stored data batches, and sequentially performs the following alignment process on each data unit in both types of data: S112. Using a unified time base as a reference, interpolation or timestamp matching methods are used to synchronize the data of the data unit at the time of collection to the preset time section, thus completing the time axis alignment; S113. Using a unified spatial reference, the coordinates of the data unit after time axis alignment are mapped to the preset spatial coordinate system using a coordinate transformation matrix, thus completing the spatial coordinate mapping. S114. Merge all data units after time axis alignment and spatial coordinate mapping into a unified format of global situation information.

[0034] Specifically, for real-time data streams, using a unified time reference (NTP), an interpolation method is employed to synchronize data from different sampling frequencies to a preset time segment. .

[0035] Assume the sensor is at time... The sampled value Target time section satisfy The interpolation result is: ; in, , These are adjacent original sampling times; for The sampled value at the sampling time, for The sampled value at the sampling time; To unify the preset time sections on the timeline; This is the synchronized data obtained after interpolation.

[0036] For historical batch data, a timestamp matching method is used. Spatial coordinate mapping: By establishing a global coordinate system, a pinhole camera model is adopted for the fixed camera. ; in, As a scale factor, For the camera intrinsic parameter matrix, Let be a rotation matrix. Let (X,Y,Z) be the translation vector, and (X,Y,Z) be the world coordinates of the point. , The coordinates are the image pixel coordinates. After obtaining the parameters through calibration, each pixel coordinate is mapped to the global coordinate system. For the mobile robot, the real-time pose is acquired using LiDAR SLAM for coordinate transformation. Finally, global situational information is generated, represented as a spatiotemporal tensor. ,in The number of time sections, , , The size of each dimension of the spatial grid. Features are defined as dimensions such as temperature, vibration amplitude, and image feature vectors.

[0037] S2. Based on the global situation information, generate candidate inspection operation plans; in a digital equivalent model that is synchronized with the operating status of the inspected physical equipment, simulate and deduce the instruction sequence in the candidate inspection operation plan to verify its security, and determine the candidate inspection operation plan that passes the security verification as the instruction sequence to be executed.

[0038] Furthermore, in step S2, the step of generating candidate inspection operation plans based on the global situation information is specifically implemented through the following sub-steps: S201. Receive the global situation information, which includes the latest status parameters, environmental physical quantities and spatial location distribution of each device in the inspection area; S202. Based on the global situation information, receive the inspection rules configured by the operation and maintenance personnel through the interactive strategy editing interface. The rules include inspection paths, triggering conditions, and handling actions. The effective scope of the inspection rules is limited according to the spatial location distribution in the global situation information. Specifically, the inspection rules are bound to preset spatial area boundaries or device coordinate positions. The inspection rules only take effect when the current position of the physical inspection terminal falls within the spatial area boundary or the distance from the device coordinate position is less than a preset threshold. S203. The interactive strategy editing interface adopts a low-code graphical orchestration engine to generate a scheme description through visual drag-and-drop operation. The scheme description includes at least the inspection path point sequence, the detection action instructions for each inspection point, and the condition triggering logic. S204. The scheme description is standardized in a domain-specific language and used as the candidate inspection operation scheme.

[0039] Furthermore, in step S2, the step of simulating and deducing the instruction sequence in the candidate inspection operation plan in a digital equivalent model that is synchronized with the operating status of the inspected physical equipment to verify its safety includes: S211. Load the instruction sequence from the candidate inspection operation plan into the digital equivalent model; S212. Perform a forward-looking simulation of the instruction sequence in the digital equivalent model to generate simulation result data; the simulation result data includes the movement trajectory of each physical inspection terminal in the spatiotemporal coordinate system after each physical inspection terminal executes the instruction sequence, as well as the evolution curve of the operating parameters of the inspected physical equipment over time. S213. Perform a security analysis based on the simulation results data. The security analysis includes the following detection items: detecting whether there is a spatiotemporal overlap or conflict between the movement trajectories of each physical inspection terminal, and detecting whether the evolution curve of the equipment operating parameters exceeds the preset safety threshold range. S214. If all inspection items meet the corresponding safety requirements, the candidate inspection operation plan is deemed to have passed the safety verification; otherwise, the candidate inspection operation plan is rejected, and conflict information or over-limit information corresponding to the failed inspection items is output.

[0040] Specifically, operations and maintenance personnel can configure inspection rules by dragging and dropping in the low-code policy editing interface, such as periodic comprehensive inspections and key inspections under specific environmental conditions. The system dynamically generates candidate solutions based on global situational information, described in a domain-specific language (DSL), including the coordinates of inspection path points, detection action types, triggering conditions, and handling logic.

[0041] Constructing a digital equivalent model: A fusion of Physical Information Neural Network (PINN) and Graph Attention Network (GAT) is employed. Device state vector. Its dynamic evolution satisfies preset physical laws, such as the heat conduction equation. The loss function of PINN is: ; in This refers to the device state vector, such as temperature, voltage, vibration, partial discharge, etc. Given a function based on known physical laws, These are boundary condition functions, such as the maximum allowable temperature of the equipment. To control the input, , , where is the weighting coefficient. GAT constructs a graph with devices as nodes and causal relationships obtained from Granger causality tests as edges, and outputs a feature representation that incorporates neighborhood information.

[0042] Load the instruction sequences from the candidate schemes into the digital equivalent model and perform a prospective simulation with a simulation step size of [value missing]. Calculate the movement trajectory of each physical inspection terminal: given the initial position. and path point sequence Cubic spline interpolation is used: ; in, For a moment The terminal position vector; , , , The coefficient vector is determined by the boundary conditions (starting position). Finish line Starting speed Final speed Determine. Predict the evolution of operating parameters of the inspected equipment: based on control input. Substitute the values ​​into the PINN model to iteratively solve for the state at the next time step: ; in, The device state vector at time t contains key operating parameters of the inspected physical device, such as temperature, voltage, vibration amplitude, and partial discharge. The simulation step size is the time interval between two consecutive state updates. For integration variables The state vector at time t, from Continuous change to ; For integration variables The control input at any given time is determined by the sequence of instructions in the inspection plan; It is a physical law function that describes the dynamic evolution of the state vector over time, such as the heat conduction equation or the electromechanical coupling equation; The results obtained from simulation and deduction The device state vector at any given moment. The physical meaning of this formula is: from the current state... Departure, control input Substitute physical law functions In the time interval Integrate internally and accumulate the state changes to obtain the state prediction for the next time step after simulation.

[0043] Security analysis: Detect spatiotemporal overlap conflicts between multiple terminals, and calculate the conflicts between any two terminals. , At any moment Spatial distance: ; in, , Terminals and exist Position vector at any given time; It is the Euclidean norm.

[0044] If it exists make ( If a preset safety distance is used, a conflict is recorded. The system checks whether the operating parameters of the detection equipment exceed preset safety thresholds, such as whether the temperature exceeds preset upper or lower temperature limits. If all detection items meet the requirements, the candidate solution is selected as the instruction sequence to be executed; otherwise, the solution is rejected and conflict or over-limit information is output.

[0045] S3. Combining the sequence of instructions to be executed and the global situation information, generate decision control instructions based on an open multi-agent architecture, and enable the physical inspection terminal to execute the decision control instructions and collect execution feedback data.

[0046] Furthermore, in step S3, combining the sequence of instructions to be executed and the global situational information, decision control instructions are generated based on an open multi-agent architecture, including: S301. Construct an open multi-agent architecture that supports dynamic registration of agent types and elastic expansion of the number of agents. It includes several command agents and execution agents and adopts a command-execution decoupled cooperation mode. At least one command agent is responsible for global task decomposition and scheduling, and at least one execution agent is responsible for executing specific inspection or diagnostic tasks. S302. The command-type intelligent agent receives global situational information and a sequence of instructions to be executed, decomposes the overall task into multiple sub-tasks, each sub-task carrying a list of required resources and spatial constraints; at the same time, it creates a global task allocation table to record the allocation status of each sub-task, the responsible execution-type intelligent agent, and the current execution progress. S303. After the command-type intelligent agent issues a sub-task announcement, each execution-type intelligent agent assesses the comprehensive cost of executing the sub-task based on its own capabilities. The comprehensive cost is calculated by weighting the resource consumption cost corresponding to the currently available resources, the movement cost corresponding to the path distance to the target area, and the confidence cost corresponding to the historical task execution success rate. Each execution-type intelligent agent reports the calculated comprehensive cost to the command-type intelligent agent. S304. After collecting the comprehensive costs reported by all executive agents, the command agent selects the executive agent with the lowest comprehensive cost as the executor of the subtask, assigns the subtask to the executive agent, and updates the status of the corresponding subtask in the global task allocation table. S305. The selected executive agent generates decision control instructions for specific physical inspection terminals based on the requirements of the subtask.

[0047] Furthermore, in step S3, the decision control command is generated through a hybrid decision-making mechanism that integrates a self-learning decision model and an experiential knowledge reasoning model, specifically implemented through the following sub-steps: S311. The self-learning decision model is based on a multi-agent reinforcement learning algorithm. It takes the current real-time updated global situation information as input and outputs multiple candidate control actions and an alarm priority evaluation value corresponding to each candidate action. The candidate control actions include executable operations for equipment abnormalities, and the alarm priority evaluation value is used to characterize the urgency of the abnormal event that the action responds to. S312. The experience-based knowledge reasoning model is based on a case-based reasoning expert system. Using key state parameters in the current global situation information as query conditions, it retrieves successful handling cases from the historical case database whose similarity to the key state parameters exceeds a preset threshold. It extracts the control action sequences used in the successful handling cases as recommended actions and applies safety boundary constraints to the candidate control actions output by the self-learning decision model. The safety boundary constraints include equipment parameter threshold limits, operation sequence compliance rules, and spatiotemporal safety rules. S313. The candidate control actions output by the self-learning decision model are fused with the recommended actions output by the empirical knowledge reasoning model and the safety boundary constraints using a priority arbitration mechanism to obtain the final decision control command. Specifically, the priority arbitration mechanism for fusion decision-making means: if a candidate control action satisfies all safety boundary constraints, then the candidate control action is output as the final decision control command; if a candidate control action violates any safety boundary constraint, then the candidate control action is rejected, and the recommended action output by the empirical knowledge reasoning model is output as the final decision control command.

[0048] Furthermore, the step S3, which involves the physical inspection terminal executing the decision control command and collecting execution feedback data, is specifically implemented through the following sub-steps: S321. Distribute decision control commands to the corresponding physical inspection terminals through a unified terminal scheduling interface; S322. After each physical inspection terminal executes the instruction, it generates execution feedback data, which includes at least the terminal execution status, anomaly detection results, and resource usage information. S323. The execution feedback data is reported to the command-type intelligent agent in real time, so that the command-type intelligent agent can dynamically adjust the allocation scheme of the unassigned sub-tasks or re-plan the execution order of the remaining tasks according to the real-time status of each physical inspection terminal and the status of unfinished tasks; on the other hand, it is stored in the local cache for subsequent transmission back to the digital equivalent model.

[0049] Specifically, this embodiment first constructs a multi-agent architecture: one command agent and multiple execution agents, which are respectively bound to each physical inspection terminal.

[0050] The commanding agent receives a sequence of instructions to be executed, breaks it down into multiple subtasks, each carrying a list of required resources and spatial constraints. A global task allocation table is created to record the allocation status of each subtask, the responsible executing agent, and its current execution progress.

[0051] Each executive agent evaluates the overall cost of performing the sub-task: ; in, , , This is a comprehensive cost weighting coefficient, which can be dynamically adjusted according to the task type; The historical task success rate of this executive agent (number of successful tasks / total number of assignments); The cost of resource consumption is defined as: ; in, The remaining battery percentage (0~1); The proportion of available bandwidth (0~1); Sensor integrity rate (0~1); , , This refers to the resource cost weighting coefficient. The cost of movement is defined as: ; in, This represents the path distance from the agent's current position to the subtask target region. This is the maximum moving speed of the terminal.

[0052] Each executive intelligent agent will Report to the command agent. Command agent selection. The smallest executive agent acts as the executor of subtasks and updates the global task allocation table.

[0053] The selected executive agent generates decision control commands using a hybrid decision-making mechanism. The self-learning decision model is based on a multi-agent reinforcement learning algorithm (such as MADDPG), taking the current global situation information as input and outputting multiple candidate control actions. and the corresponding alarm priority evaluation value The experiential knowledge reasoning model is based on a case-based reasoning expert system, using the current working condition feature vector. Based on the query criteria, retrieve similarity from the historical case database: ; in, This is the current operating condition feature vector (equipment type, load rate, ambient temperature range, etc.). This represents the feature vector of the working conditions in historical cases; Let be the Euclidean norm (modulus) of the vector. When When a preset similarity threshold is met, extract the corresponding recommended action. and to Apply safety boundary constraints. Employ a priority arbitration mechanism: ; Output As decision-making and control commands, instructions are sent to physical inspection terminals via a unified terminal scheduling interface (based on a publish-subscribe mechanism). After execution, the terminals return execution feedback data, including terminal execution status, anomaly detection results, and resource usage information. The feedback data is reported in real time to the command-line intelligent agent for dynamic task adjustment and is also stored in the local cache.

[0054] S4. The execution feedback data is sent back to the digital equivalent model and a structured inspection report is automatically generated.

[0055] Specifically, the execution feedback data is fed back to the digital equivalent model, and an extended Kalman filter is used to update the equipment state estimate, correcting physical parameters in the model, such as the heat dissipation coefficient. Simultaneously, a structured inspection report is automatically generated based on retrieval-enhanced generation technology: using anomaly detection results as query conditions, matching related knowledge, such as handling experience for similar faults, is retrieved from the equipment knowledge graph and historical case database. This information is then fused to generate a report containing an inspection overview, anomaly details, handling suggestions, and attachments. Successful cases from the execution feedback data are stored in a case database and used to update the reward function of the self-learning decision model, forming a continuous evolutionary closed loop.

[0056] This embodiment systematically solves the problems of rigid strategies, closed decision-making, inefficient collaboration, and insufficient model evolution capabilities in existing inspection technologies through the four sequentially executed technical steps and the corresponding four modules. The specific beneficial effects are as follows: First, it enables the unification of spatiotemporal references for multi-source heterogeneous sensing data. It acquires multi-source heterogeneous sensing data within the inspection area, including at least equipment status parameters, environmental physical quantities, and image data, and maps this data to a unified time and space coordinate system, generating comprehensive situational information with a consistent spatiotemporal reference. This step eliminates the information silos caused by scattered data sources, varying formats, and spatiotemporal asynchrony in traditional systems. It allows subsequent strategy generation and decision execution to be based on a complete, consistent, and low-latency environmental understanding, fundamentally improving the reliability and real-time performance of the sensing layer.

[0057] Secondly, it enables strategy generation and proactive security verification based on digital equivalent models. Candidate inspection operation plans are dynamically generated based on comprehensive situational information. The instruction sequences in these candidate plans are then simulated and analyzed within a digital equivalent model synchronized with the operating status of the inspected equipment to verify their security. Only the verified plans are selected as the execution sequence. This step breaks through the rigid mode of traditional inspections relying on pre-set scripts, allowing inspection strategies to be dynamically adjusted based on the current health status of the equipment, environmental changes, and task priorities. Simultaneously, simulation and analysis can proactively detect multi-terminal spatiotemporal conflicts and predict safety risks such as exceeding equipment parameter limits before physical execution, transforming traditional post-event correction into pre-event prevention, effectively improving the security and strategy adaptability of inspection operations.

[0058] Third, it enables open multi-agent autonomous decision-making and collaborative execution. Combining the sequence of instructions to be executed and global situational information, decision-making and control instructions are generated based on an open multi-agent architecture, and the physical inspection terminals execute these instructions and collect execution feedback data. This step adopts a command-execution decoupled collaborative mode, supporting dynamic registration of agent types and elastic expansion of their numbers, enabling the system to have autonomous decision-making, distributed negotiation, and resource optimization allocation capabilities. Compared to traditional centralized hard-coded control, this embodiment effectively improves the timeliness of anomaly response, the efficiency of multi-terminal collaboration, and the system's scalability, adapting to inspection scenarios of varying scales and complexities.

[0059] Fourth, it enables closed-loop feedback and automatic generation of structured reports. Execution feedback data is fed back to the digital equivalent model, and structured inspection reports are automatically generated. The execution feedback data is used to correct the simulation parameters of the digital equivalent model, allowing the model to gradually approximate real physical behavior. Simultaneously, the structured report integrates perceived data, decision-making basis, anomaly detection results, and handling plans, reducing the burden of manual analysis and providing traceable decision-making basis. Overall, it forms a complete closed loop of perception, verification, decision-making, execution, and feedback, enabling continuous correction and evolution based on actual operational data.

[0060] Example 2: In the actual inspection and maintenance of power substations, in addition to routine periodic inspection tasks, the system faces two typical and challenging scenarios: rapid response to sudden abnormal events and difficulties in model performance degradation and cross-site knowledge transfer during long-term deployment. Specifically, in actual operation, equipment may experience sudden emergency anomalies such as overheating or increased partial discharge. Traditional inspection systems lack a dynamic task generation mechanism and usually still perform inspections according to a predetermined cycle, resulting in delayed anomaly detection. Even if an anomaly is detected, manual intervention or fixed scripts make it difficult to coordinate multiple robots, fixed cameras, and other terminals to work collaboratively in a short time, easily leading to path conflicts, resource contention, and blocking of critical handling actions. Therefore, based on Example 1, a better implementation method is given. This example introduces anomaly-triggered emergency inspection, which addresses how to dynamically trigger emergency inspection and efficiently schedule multi-agent responses when the system detects equipment anomalies.

[0061] S1. Acquire multi-source heterogeneous sensing data within the inspection area and map it to a unified spatiotemporal coordinate system to generate global situational information with a consistent spatiotemporal reference; the multi-source heterogeneous sensing data includes at least the status parameters of the inspected physical equipment, environmental physical quantities, and image data.

[0062] The specific implementation of this step is the same as step S1 in Embodiment 1, with real-time updates of the overall situational information. Furthermore, based on this, this embodiment requires continuous monitoring of the key operating parameters of each device. When the instantaneous rate of change or absolute value of a certain parameter (such as transformer winding temperature) exceeds a preset warning threshold, an emergency mode is triggered.

[0063] S2. Based on the global situation information, generate candidate inspection operation plans; in a digital equivalent model that is synchronized with the operating status of the inspected physical equipment, simulate and deduce the instruction sequence in the candidate inspection operation plan to verify its security, and determine the candidate inspection operation plan that passes the security verification as the instruction sequence to be executed.

[0064] In step S2, based on steps S201 to S204 and S211 to S214 of Embodiment 1, this embodiment automatically triggers an emergency mode when an equipment malfunction is detected, thereby generating an emergency strategy. Specifically, based on the current global situational information (including the location of the abnormal equipment, load rate, ambient temperature, fault propagation map, etc.), an emergency inspection plan is dynamically generated through a low-code engine. The instruction sequence of this plan includes: designating the optimal inspection terminal to go to the location of the abnormal equipment, performing a verification test, and performing a handling action. This emergency plan is loaded into a digital equivalent model for simulation, simulating movement path conflicts and predicting the evolution of equipment parameters after the handling action. After the simulation passes, it is determined as the instruction sequence to be executed. In this embodiment, performing a verification test includes infrared temperature measurement and partial discharge detection; performing a handling action is to start the cooling system.

[0065] S3. Combining the sequence of instructions to be executed and the global situation information, generate decision control instructions based on an open multi-agent architecture, and enable the physical inspection terminal to execute the decision control instructions and collect execution feedback data.

[0066] This step, based on steps S301 to S305 of Example 1, provides an additional implementation scheme: Please refer to Figure 3 That is, in step S3, combining the sequence of instructions to be executed and the global situational information, decision control instructions are generated based on an open multi-agent architecture. This also includes a dynamic agent scheduling mechanism based on task load awareness, specifically implemented through the following sub-steps: S331. Each execution agent periodically reports the current task queue information to the command agent. The task queue information includes at least: the number of subtasks to be processed in the queue, the estimated processing delay of the task at the head of the queue, and the remaining power and available bandwidth of the physical inspection terminal controlled by the agent. S332. When assigning newly generated subtasks, the command agent calculates the overall load of each execution agent based on the task queue information reported by each execution agent, and prioritizes assigning subtasks to the execution agent with the lowest overall load; the overall load is obtained by weighting the number of subtasks to be processed, the estimated processing delay, and the level of remaining resources. S333. If the number of pending subtasks of an execution agent exceeds the preset task backlog limit, or the estimated processing delay of the task at the head of the queue exceeds the preset response time limit, the agent is determined to be in an overloaded state. The command agent selects subtasks that have not yet started execution from the task queue of the overloaded agent, reassigns them to other agents with the lowest overall load, and updates the global task allocation table.

[0067] Specifically, in this embodiment, the command-type intelligent agent receives an emergency command sequence and decomposes it into multiple sub-tasks. Each execution-type intelligent agent periodically reports task queue information: the number of sub-tasks to be processed. Estimated latency at the head of the queue Remaining battery power Available bandwidth Overall computational load of the command-type intelligent agent: ; in, To assess overall workload, a higher value indicates a busier agent. The number of subtasks to be processed in the queue; The estimated processing latency for the task at the head of the queue; This represents the remaining battery percentage. This represents the percentage of available bandwidth. , , , This is the load weighting coefficient, configured according to system preferences (more emphasis on response time or energy consumption).

[0068] Prioritize assigning subtasks to The smallest executive intelligent agent.

[0069] An agent is considered overloaded if it meets any of the following conditions: ;or ; in, The preset task backlog limit, The preset response time limit is used. The command agent selects unexecuted subtasks from its queue and recalculates the current response time for each agent. Select the agent with the lowest load to migrate the task and update the global task allocation table.

[0070] The selected executive agent generates decision control instructions using the hybrid decision-making mechanism of Example 1. The self-learning model outputs candidate actions. If cooling is initiated immediately, the expert system will search for similar operating condition cases and recommend appropriate actions. After the security constraint verification is passed, the arbitration outputs the final instruction. The instruction is sent through the terminal scheduling interface, and execution feedback data is reported in real time, allowing the command-type intelligent agent to dynamically adjust the allocation of remaining tasks.

[0071] S4. The execution feedback data is sent back to the digital equivalent model and a structured inspection report is automatically generated.

[0072] The complete process of this abnormal event (including the operating condition feature vector) The sequence of actions taken and the effectiveness evaluation indicators are coded as new cases and stored in the historical case database to enhance the coverage of the experience-based reasoning model. Emergency inspection reports are automatically generated, highlighting the alarm level, response time, handling process, and effectiveness evaluation.

[0073] Example 2 addresses scenarios of sudden equipment malfunctions. Building upon Example 1, it adds an event-driven emergency response and load-aware scheduling mechanism. Its beneficial effects are reflected in the following aspects: First, it effectively shortens the response time to anomalies. By monitoring device parameters in the overall situational information in real time, emergency strategies are automatically triggered when parameters exceed the warning threshold, eliminating the need to wait for fixed inspection cycles. This allows the system to respond to sudden anomalies instantly, avoiding the risk of escalating faults due to delayed responses.

[0074] Second, dynamic load balancing among multiple agents is achieved. Each execution agent periodically reports task queue information, and the command agent calculates the overall load based on this, prioritizing the allocation of emergency subtasks to the agent with the lowest load. When an agent is overloaded, the system automatically migrates unexecuted tasks to other idle agents, effectively avoiding task backlog and resource deadlock, and ensuring the efficient execution of emergency tasks.

[0075] Third, it strengthens the safety decision-making boundaries in emergency scenarios. The hybrid decision-making mechanism introduces priority arbitration. When the aggressive actions recommended by the self-learning model violate safety constraints, it automatically degrades to the conservative actions of the expert system, ensuring that the safety boundaries of equipment operation are not exceeded even in emergency situations, thus balancing the timeliness and safety of emergency response.

[0076] Example 3: In actual multi-site deployments, each switching station has different equipment models, load characteristics, and environmental conditions, and the fault samples at each station are usually sparse. If each station trains its model independently, it is prone to overfitting and insufficient generalization ability; if all the original data is centralized in the cloud for training, it faces the dual pressures of data privacy (sensitive data of power grid operation) and communication bandwidth. Therefore, based on Example 1, another better implementation method is given. This example introduces cross-site collaborative evolution and long-term deployment. After multiple switching stations are deployed, how can the model be continuously evolved through cloud collaboration, and how can the inspection strategy be dynamically adjusted based on equipment health prediction?

[0077] S1. Acquire multi-source heterogeneous sensing data within the inspection area and map it to a unified spatiotemporal coordinate system to generate global situational information with a consistent spatiotemporal reference; the multi-source heterogeneous sensing data includes at least the status parameters of the inspected physical equipment, environmental physical quantities, and image data.

[0078] The specific implementation of this step can be referred to step S1 of Embodiment 1. Furthermore, based on this, this embodiment sets up multiple edge computing nodes deployed in different inspection areas, each independently completing multi-source data acquisition and spatiotemporal alignment to generate global situational information. The nodes do not directly share raw data with each other.

[0079] S2. Based on the global situation information, generate candidate inspection operation plans; in a digital equivalent model that is synchronized with the operating status of the inspected physical equipment, simulate and deduce the instruction sequence in the candidate inspection operation plan to verify its security, and determine the candidate inspection operation plan that passes the security verification as the instruction sequence to be executed.

[0080] The specific implementation of this step is basically the same as step S2 in Embodiment 1. However, when generating the strategy, this embodiment will additionally input the device health prediction information (remaining useful life RUL) from step S40, so that the generation of candidate solutions can take into account the health status of the device.

[0081] S3. Combining the sequence of instructions to be executed and the global situation information, generate decision control instructions based on an open multi-agent architecture, and enable the physical inspection terminal to execute the decision control instructions and collect execution feedback data.

[0082] The specific implementation of this step is basically the same as step S3 in Example 1, with each edge node independently running the multi-agent decision-making and execution process.

[0083] S4. The execution feedback data is sent back to the digital equivalent model and a structured inspection report is automatically generated.

[0084] Please refer to Figure 4The specific implementation process of this step, based on Example 1, adds a continuous evolution step based on edge incremental learning and cloud federated aggregation, which is implemented through the following sub-steps: S401. The image data and state parameters in the execution feedback data are used as new training samples to perform incremental parameter updates on the self-learning decision model; the incremental parameter updates are performed on the edge computing node deployed locally in the inspection area, and the edge computing node and the physical inspection terminal are in the same network domain.

[0085] Specifically, in this sub-step, each edge node collects image and state parameters from the execution feedback data as new training samples. When the accumulated number of samples reaches a preset update threshold, incremental updates are performed on the self-learning decision model, using an experience replay buffer. Actor network parameters... and Critic network parameters The update rules are as follows: ; ;

[0086] in, For Actor network parameters (policy network parameters); These are the Critic network parameters (value network parameters). This indicates the parameters of the Actor network. Find the gradient; Describe the policy objective function Regarding Actor network parameters The gradient indicates how to adjust. To maximize expected cumulative returns for strategy improvement; This indicates calculating the gradient with respect to the parameters of the Critic network; Value loss function Regarding Critic network parameters The gradient indicates how to adjust. To minimize the value estimation error, it is used for value function approximation. The learning rate of the Actor network controls the step size for updating the policy parameters; The learning rate of the Critic network controls the step size for updating the value parameters; The objective function of the strategy measures the expected cumulative return of the strategy. This is the value loss function, which measures the error in value estimation.

[0087] S402. Multiple edge computing nodes deployed in different inspection areas encrypt and upload their updated self-learning decision model parameters to the cloud server; the cloud server uses a federated averaging algorithm to aggregate the model parameters uploaded by each edge computing node, generates a globally updated self-learning decision model, and then distributes it to each edge computing node.

[0088] Specifically, this sub-step involves each edge node updating the model parameters. (Excluding the original data) The encrypted data is uploaded to the cloud server. The cloud uses a federated average algorithm for aggregation. ; in, The number of edge nodes participating in the aggregation; For the first The number of local samples used by each node; For the first The updated model parameters for each node; These are the model parameters after global aggregation.

[0089] S403. Each edge computing node replaces its local self-learning decision model with the received globally updated self-learning decision model, and uses it for subsequent generation of decision control instructions.

[0090] After generating the global update model, it is distributed to each edge node, and the nodes are configured to... After updating the model, replace the local model for subsequent decision-making.

[0091] Furthermore, based on the aforementioned continuous evolution steps of edge incremental learning and cloud federated aggregation, a step of equipment health prediction and dynamic adjustment of inspection strategies based on degradation trajectory modeling is added. Specifically, this includes: constructing an equipment health degradation trajectory model based on the evolution sequence of the state parameters of the inspected equipment over time in historically accumulated execution feedback data to output a predicted value of the equipment's remaining service life; the degradation trajectory model uses a Wiener process or exponential decay function to fit the degradation trend of key equipment indicators; comparing the predicted value of the remaining service life with a preset threshold, and if it is lower than the threshold, marking the equipment as a key focus equipment, and dynamically adjusting the generation strategy of the candidate inspection operation plan, including increasing the inspection frequency of the equipment or adding detection items. The added detection items include at least one of the following types: adding infrared thermal imaging temperature measurement, adding partial discharge detection, adding vibration spectrum analysis, and adding online oil chromatography monitoring.

[0092] Specifically, this is based on the evolution sequence of the inspected equipment's status parameters (such as insulation resistance and temperature rise rate) in historical execution feedback data. A model of equipment health degradation trajectory is constructed. The Wiener process is employed. ;

[0093] in, For a moment The amount of equipment degradation is a normalized health indicator, such as insulation resistance value; This is the drift parameter, i.e., the degradation rate, which is the average amount of degradation per unit time. This is the diffusion parameter, i.e., the degree of random fluctuation; The increment of standard Brownian motion follows a normal distribution with a mean of 0 and a variance of dt.

[0094] The drift parameter estimates are obtained through maximum likelihood estimation. and diffusion parameter estimates The drift parameter estimate reflects the average trend of the equipment degradation process (average degradation per unit time); the diffusion parameter estimate characterizes the degree of random fluctuation in the equipment degradation process. Let the failure threshold be... This refers to the critical value at which the equipment is deemed to have failed, and the remaining service life is: ; in, Remaining Useful Life refers to the length of time that the equipment is expected to continue operating normally from the current moment. Indicates time The equipment degradation is a normalized health indicator (insulation resistance value, temperature rise rate, etc.), which monotonically increases or decreases with the operation time. This is the equipment failure threshold, which is the threshold at which the equipment is judged to be failed or needs to be taken out of operation when the degradation reaches this critical value. This is the infimum operator, which means taking the minimum value among all time points that satisfy the condition; This refers to the current moment, i.e., the moment when the prediction is made; This represents a combination of multiple moments that satisfy the condition that "the amount of degradation reaches the failure threshold".

[0095] Then the inspection strategy is dynamically adjusted, if If the health threshold is below a preset threshold, the device is marked as a key focus device, and the generation strategy for candidate inspection plans is dynamically adjusted: the inspection frequency is increased, and at least one of the following in-depth detection methods is added to the detection items: infrared thermal imaging temperature measurement, partial discharge detection, vibration spectrum analysis, and online oil chromatography monitoring. The adjusted strategy is then used to regenerate candidate plans through a low-code engine, proceeding to step S2 for verification.

[0096] Automatically generate periodic maintenance reports, including model evolution indicators for each node (such as the trend of average prediction error), equipment health trend curves, and recommended maintenance plans.

[0097] Example 3 addresses long-term cross-site deployment scenarios and adds edge incremental learning, cloud federated aggregation, and device degradation prediction mechanisms to Example 1. Its beneficial effects are reflected in the following aspects: First, it achieves cross-site collaborative evolution of the model and data privacy protection. Each edge node incrementally updates its self-learning decision model using only local execution feedback data and uploads the model parameters (rather than the original data) to the cloud. The cloud aggregates the parameter updates from each node using a federated averaging algorithm, generates a global model, and then distributes it to replace the local model. This mechanism fully leverages the statistical commonalities across sites to improve the model's generalization ability without leaking the original perception data.

[0098] Secondly, it addresses the issue of model performance degradation during long-term deployment. Through edge incremental learning, the model can continuously adapt to dynamic factors such as equipment aging and environmental changes based on actual operational data, avoiding the decline in decision accuracy caused by concept drift in static models. The cloud continuously aggregates optimization directions from various sites, enabling the global model to absorb diverse experiences and further enhancing the system's adaptability.

[0099] Third, it achieves a leap from passive response to predictive maintenance. Based on the Wiener process, a health degradation trajectory model for equipment is constructed. Drift and diffusion parameters are obtained through maximum likelihood estimation, dynamically predicting the remaining service life of the equipment. When the predicted service life falls below the health threshold, the system automatically increases the inspection frequency and adds in-depth inspection items, precisely focusing limited maintenance resources on high-risk equipment, thereby reducing the probability of sudden failures and extending the equipment's entire lifespan.

[0100] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0101] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. Furthermore, the above embodiments only illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A multi-agent autonomous inspection closed-loop control method based on an open architecture, characterized in that, include: Acquire multi-source heterogeneous sensing data within the inspection area and map it to a unified spatiotemporal coordinate system to generate global situational information with a consistent spatiotemporal reference. The multi-source heterogeneous sensing data includes at least the status parameters of the inspected physical equipment, environmental physical quantities, and image data. Based on the overall situational information, candidate inspection operation plans are generated; In a digital equivalent model that is synchronized with the operating status of the inspected physical equipment, the instruction sequence in the candidate inspection operation plan is simulated and deduced to verify its safety, and the candidate inspection operation plan that passes the safety verification is determined as the instruction sequence to be executed. Combining the sequence of instructions to be executed and the global situation information, decision control instructions are generated based on an open multi-agent architecture, and the physical inspection terminal executes the decision control instructions and collects execution feedback data. The execution feedback data is sent back to the digital equivalent model and a structured inspection report is automatically generated. Combining the sequence of instructions to be executed and the overall situational information, decision-making and control instructions are generated based on an open multi-agent architecture, including: Construct an open multi-agent architecture that supports dynamic registration of agent types and elastic expansion of the number of agents. It includes several command agents and execution agents and adopts a command-execution decoupled cooperation mode. At least one command agent is responsible for global task decomposition and scheduling, and at least one execution agent is responsible for executing specific inspection or diagnostic tasks. The command-type intelligent agent receives global situational information and a sequence of instructions to be executed, decomposes the overall task into multiple sub-tasks, each sub-task carrying a list of required resources and spatial constraints; at the same time, it creates a global task allocation table to record the allocation status of each sub-task, the responsible execution-type intelligent agent, and the current execution progress. After the command-type intelligent agent issues a sub-task announcement, each execution-type intelligent agent assesses the comprehensive cost of executing the sub-task based on its own capabilities. The comprehensive cost is calculated by weighting the resource consumption cost corresponding to the currently available resources, the movement cost corresponding to the path distance to the target area, and the confidence cost corresponding to the historical task execution success rate. Each execution-type intelligent agent reports the calculated comprehensive cost to the command-type intelligent agent. After collecting the comprehensive costs reported by all executive agents, the command agent selects the executive agent with the lowest comprehensive cost as the executor of the subtask, assigns the subtask to the executive agent, and updates the status of the corresponding subtask in the global task allocation table. The selected execution agent generates decision control instructions for specific physical inspection terminals based on the requirements of the subtasks; Combining the sequence of instructions to be executed and global situational information, decision-making and control instructions are generated based on an open multi-agent architecture. This also includes a dynamic agent scheduling mechanism based on task load awareness, specifically comprising: Each execution agent periodically reports the current task queue information to the command agent. The task queue information includes at least: the number of subtasks to be processed in the queue, the estimated processing delay of the task at the head of the queue, and the remaining power and available bandwidth of the physical inspection terminal controlled by the agent. When assigning newly generated subtasks, the command agent calculates the overall load of each execution agent based on the task queue information reported by each execution agent, and prioritizes assigning subtasks to the execution agent with the lowest overall load. The overall load is obtained by weighting the number of subtasks to be processed, the estimated processing latency, and the level of remaining resources. If the number of pending subtasks of an execution agent exceeds the preset task backlog limit, or the estimated processing latency of the task at the head of the queue exceeds the preset response time limit, the agent is determined to be in an overload state. The command agent selects subtasks that have not yet started execution from the task queue of the overloaded agent, reassigns them to other agents with the lowest overall load, and updates the global task allocation table. The command-type intelligent agent receives emergency command sequences and breaks them down into multiple sub-tasks; each execution-type intelligent agent periodically reports task queue information: the number of sub-tasks to be processed. Estimated latency at the head of the queue Remaining battery power Available bandwidth Overall computational load of the command-type intelligent agent: ； in, To assess overall workload, a higher value indicates that the agent is busier. The number of subtasks to be processed in the queue; The estimated processing latency for the task at the head of the queue; This represents the remaining battery percentage. This represents the percentage of available bandwidth. , , , This is the load factor, which can be configured according to system preferences.

2. The method according to claim 1, characterized in that, The mapping to a unified spatiotemporal coordinate system generates global situational information with a consistent spatiotemporal reference, including: Construct a data alignment pipeline that integrates streaming and batch processing. This pipeline receives real-time acquired data streams and historically stored data batches, and sequentially performs the following alignment process on each data unit in both types of data: Using a unified time base as a reference, interpolation or timestamp matching methods are employed to synchronize the data collected at the time of the data unit to the preset time section, thereby completing the time axis alignment. After the data unit is aligned along the time axis, a coordinate transformation matrix is ​​used to map the coordinates of the data unit to a preset spatial coordinate system, thus completing the spatial coordinate mapping. All data units, after being aligned along the time axis and mapped to spatial coordinates, are merged into a unified format of global situational information.

3. The method according to claim 1, characterized in that, The step of generating candidate inspection operation plans based on the overall situation information includes: Receive the global situation information, which includes the latest status parameters, environmental physical quantities and spatial location distribution of each device in the inspection area; Based on the overall situation information, the system receives inspection rules configured by maintenance personnel through an interactive strategy editing interface. The rules include inspection paths, triggering conditions, and handling actions. The effective scope of the inspection rules is limited according to the spatial location distribution in the overall situation information. Specifically, the inspection rules are bound to preset spatial area boundaries or device coordinate positions. The inspection rules only take effect when the current position of the physical inspection terminal falls within the spatial area boundary or the distance from the device coordinate position is less than a preset threshold. The interactive strategy editing interface uses a low-code graphical orchestration engine to generate scheme descriptions through visual drag-and-drop operations. The scheme descriptions include at least the inspection path point sequence, the detection action instructions for each inspection point, and the condition triggering logic. The proposed solution is then represented in a domain-specific language and used as a candidate inspection operation solution.

4. The method according to claim 1, characterized in that, The simulation and deduction of the instruction sequence in the candidate inspection operation plan to verify its safety in a digital equivalent model that is synchronized with the operating status of the inspected physical equipment includes: Load the instruction sequences from the candidate inspection operation plan into the digital equivalent model; A forward-looking simulation of the instruction sequence is performed in a digital equivalent model to generate simulation result data. The simulation result data includes the movement trajectory of each physical inspection terminal in the spatiotemporal coordinate system after each physical inspection terminal executes the instruction sequence, as well as the evolution curve of the operating parameters of the inspected physical equipment over time. Based on the simulation results data, a security analysis is performed. The security analysis includes the following detection items: detecting whether there is a spatiotemporal overlap or conflict between the movement trajectories of each physical inspection terminal, and detecting whether the evolution curve of the equipment operating parameters exceeds the preset safety threshold range. If all inspection items meet the corresponding safety requirements, the candidate inspection operation plan is deemed to have passed the safety verification; otherwise, the candidate inspection operation plan is rejected, and conflict information or over-limit information corresponding to the failed inspection items is output.

5. The method according to claim 4, characterized in that, The decision control instructions are generated through a hybrid decision-making mechanism that integrates a self-learning decision model and an experiential knowledge reasoning model, specifically including: The self-learning decision model is based on a multi-agent reinforcement learning algorithm. It takes the real-time updated global situation information as input and outputs multiple candidate control actions and an alarm priority evaluation value corresponding to each candidate action. The candidate control actions include executable operations for equipment abnormalities, and the alarm priority evaluation value is used to characterize the urgency of the abnormal event that the action responds to. The experience-based knowledge reasoning model is based on a case-based reasoning expert system. It uses key state parameters in the current global situation information as query conditions, retrieves successful handling cases from the historical case database whose similarity to the key state parameters exceeds a preset threshold, extracts the control action sequences used in the successful handling cases as recommended actions, and applies safety boundary constraints to the candidate control actions output by the self-learning decision model. The safety boundary constraints include equipment parameter threshold restrictions, operation sequence compliance rules, and spatiotemporal safety rules. The candidate control actions output by the self-learning decision model are fused with the recommended actions output by the experience knowledge reasoning model and the safety boundary constraints using a priority arbitration mechanism to obtain the final decision control command.

6. The method according to claim 5, characterized in that, The process of enabling the physical inspection terminal to execute the decision control command and collect execution feedback data includes: Decision control commands are distributed to the corresponding physical inspection terminals through a unified terminal scheduling interface. After executing the command, each physical inspection terminal generates execution feedback data, which includes at least the terminal execution status, anomaly detection results, and resource usage information. The execution feedback data is reported to the command-type intelligent agent in real time, so that the command-type intelligent agent can dynamically adjust the allocation scheme of unassigned sub-tasks or re-plan the execution order of the remaining tasks based on the real-time status of each physical inspection terminal and the status of unfinished tasks; on the other hand, it is stored in the local cache for subsequent transmission back to the digital equivalent model.

7. The method according to claim 5, characterized in that, It also includes continuous evolutionary steps based on edge incremental learning and cloud federated aggregation, specifically including: The image data and state parameters in the execution feedback data are used as new training samples to incrementally update the parameters of the self-learning decision model; the incremental parameter update is performed on the edge computing node deployed locally in the inspection area, and the edge computing node and the physical inspection terminal are in the same network domain. Multiple edge computing nodes deployed in different inspection areas encrypt and upload their updated self-learning decision model parameters to the cloud server. The cloud server uses a federated averaging algorithm to aggregate the model parameters uploaded by each edge computing node, generates a globally updated self-learning decision model, and then distributes it to each edge computing node. Each edge computing node replaces its local self-learning decision model with the received globally updated self-learning decision model, and uses it for subsequent generation of decision control instructions.

8. A multi-agent autonomous inspection closed-loop control system based on an open architecture, used to execute the method described in any one of claims 1 to 7, characterized in that, include: The sensing data acquisition and spatiotemporal alignment module is used to acquire multi-source heterogeneous sensing data within the inspection area and map it to a unified spatiotemporal coordinate system to generate global situational information with a consistent spatiotemporal reference. The multi-source heterogeneous sensing data includes at least device status parameters, environmental physical quantities, and image data; The strategy generation and verification module is used to generate candidate inspection operation plans based on the global situation information. In a digital equivalent model that is synchronized with the operating status of the inspected physical equipment, the instruction sequence in the candidate inspection operation plan is simulated and deduced to verify its safety, and the candidate inspection operation plan that passes the safety verification is determined as the instruction sequence to be executed. The multi-agent decision-making and execution module is used to combine the sequence of instructions to be executed and global situation information to generate decision control instructions based on an open multi-agent architecture, and to enable the physical inspection terminal to execute the decision control instructions and collect execution feedback data. The closed-loop feedback and report generation module is used to send the execution feedback data back to the digital equivalent model and automatically generate a structured inspection report.

Images