A power inspection robot body intelligence system based on fast-slow system architecture and a control method thereof

Abstract

The application belongs to the technical field of robots and artificial intelligence, and particularly relates to a power inspection robot body intelligence system based on a fast-slow system architecture and a control method, the system comprising an input layer: inputting an inspection task; a slow system cognitive planning layer: used for semantic analysis and long-range logic planning of a complex inspection task; a fast system reflex control layer: used for high-frequency end-to-end mapping from visual observation to action Token; a cross-modal safety shielding adapter: commonly used for dynamically solving a power safety distance and combining a Lyapunov Barrier function to perform hard physical red line checking and safety correction on a candidate action; and a robot execution layer: used for converting a safe action into an actuator control vector, collecting real-time observation feedback to the double system, and synchronously updating a long-term memory module and a subsequent planning path according to a deviation by the slow system cognitive planning layer. The system effectively improves the reliability and generalization performance of the robot in autonomous operation under complex power working conditions.

Description

Technical Field

[0001] This application belongs to the field of robotics and artificial intelligence technology, specifically relating to an embodied intelligent system and control method for a power inspection robot based on a fast-slow system architecture. Background Technology

[0002] The power system is the core of a nation's infrastructure, and its stability and security directly affect the sound operation of the social economy. Transmission lines, substations, and distribution facilities are typically located in areas with complex geographical environments and harsh electromagnetic conditions, making inspection work a challenging task. Traditional power inspection methods mainly rely on manual handheld equipment or remote-controlled drones, which are not only labor-intensive and inefficient, but also pose significant safety risks to inspection personnel when facing high-voltage, live environments and severe weather.

[0003] With the development of robotics technology, various inspection drones, quadruped robots, and wheeled inspection robots have gradually been put into use. However, existing inspection systems still have the following technical problems in terms of intelligence:

[0004] (1) Most of the existing inspection robots are “tool-type” equipment, which are highly dependent on preset waypoints or remote manual commands. When faced with unstructured dynamic scenarios (such as sudden foreign object obstruction, unknown equipment deformation or extreme weather interference), the robots lack sufficient embodied perception and on-site self-decision-making ability, making it difficult to achieve closed-loop autonomous operation under complex physical constraints.

[0005] (2) Traditional computer vision algorithms and preset control scripts have extremely poor generalization ability; for different inspection tasks or different types of power equipment, it is usually necessary to develop dedicated software modules, resulting in high system development costs and difficulty in cross-platform reuse; although multimodal large model (MLLM) has shown strong logical reasoning ability in the field of task planning in recent years, large models have "logical illusions" and high reasoning delays, making it difficult to directly undertake motor-level feedback control with extremely high physical real-time requirements.

[0006] (3) In the power industry, a special field that places great emphasis on safety, existing embodied intelligence solutions often struggle to balance “flexibility of action” and “safety red line”. Fast-paced end-to-end neural network control often lacks mathematical interpretability, making it difficult to ensure that the robot always strictly adheres to the safety distance regulations for live work when performing complex actions. Once a large model logic deviation or a fast system response error occurs, it is very easy to trigger a serious electrical fire or equipment damage accident.

[0007] Therefore, there is an urgent need for a new system that possesses both high-level logical planning capabilities for large models and low-level millisecond-level instinctive reflexes, while also strictly adhering to the red line of power safety. Summary of the Invention

[0008] To address the aforementioned technical problems in the existing technology, the purpose of this invention is to design a universal embodied intelligent control system that possesses both high-level logic planning capabilities for large models and low-level millisecond-level instinctive reflexes, while strictly adhering to the red lines of power safety. The technical solution is as follows:

[0009] An embodied intelligent system for power line inspection robots based on a fast-slow system architecture includes:

[0010] Input layer: Input inspection tasks;

[0011] Slow system cognitive planning layer: Based on a multimodal large language model, the linear temporal logic (LTL) and power knowledge graph are used to perform semantic parsing and long-range logical planning for the inspection task;

[0012] Fast system reflex control layer: Based on the visual-language-action integrated VLA model and world model prediction engine with neural circuit structure, high-frequency end-to-end mapping from visual observation to action tokens is performed;

[0013] Cross-modal safety shielding adapter: By dynamically calculating the electrical safety distance and combining it with the Lyapunov Barrier function, candidate actions are subjected to hard physical red line verification and safety correction;

[0014] Robot execution layer: The verified and safety-corrected actions are transformed into the actuator control vectors of the target carrier at the bottom layer through a hierarchical mapping matrix. Real-time observations are collected and fed back to the slow system cognitive planning layer and the fast system reflection control layer. The slow system cognitive planning layer updates the long-term memory module and subsequent planning paths synchronously according to the deviation.

[0015] Furthermore, the slow system cognitive planning layer parses the inspection task into a linear temporal logic with spatiotemporal constraints through neural symbolic computation, and the expression of the linear temporal logic is as follows:

[0016]

[0017] in, As a global operator, it means that the robot must satisfy the electrical safety clearance constraint and avoid obstacles throughout the entire process; For power safety clearance constraints; For obstacles; This is the final operator, indicating that the target inspection point must be completed; Target inspection points; This is the operator for the next time step, indicating that a status report must be triggered after the task is completed; For status reporting.

[0018] Furthermore, the operator is optimized based on operational compliance through contrastive physics learning, and the decision quality is improved by minimizing the contrastive loss function, the expression of which is:

[0019]

[0020] in, The features of the currently generated inspection plan, These are standard compliance operation characteristics extracted from the power safety expert system. These are characteristics of violations and misoperations found in the historical accident case database. This is the temperature coefficient.

[0021] Furthermore, the fast system reflection control layer utilizes a causal Transformer structure to construct an action generation sequence. It generates control action tokens by maximizing the conditional log-likelihood estimate under the current observation sequence, and then linearly maps the generated tokens to motor torque control commands using a lightweight decoder. Its objective loss function is defined as:

[0022]

[0023] in, For the generated discrete action token, This is a sliding window observation sequence composed of visual depth maps, radar point clouds, and infrared features. This is a semantically guided target vector issued by a slow system. To record short-term characteristics of physical interaction states within the last three seconds.

[0024] Furthermore, in the fast system reflection control layer, the world model prediction engine is a variational latent space dynamics prediction engine based on the world model, and its state evolution equation is expressed as:

[0025]

[0026] in, For the encoded latent space environment features, For pre-execution actions, The latent random variable, which characterizes the uncertainty in the power scenario, follows a Gaussian distribution. .

[0027] Furthermore, the cross-modal safety shielding adapter integrates a multi-criteria safety assessment model for special power operating conditions; the dynamic calculation of the power safety distance is achieved by acquiring the voltage level of the energized equipment in real time. Environmental inductive coupling strength Wind speed disturbance and air humidity The dynamic safety distance threshold is calculated, and its expression is as follows:

[0028]

[0029] in, and For environmental adaptability coefficients, The electric field breakdown characteristic index is used. To predict emergency braking displacement based on the robot's current momentum, This is for sensor measurement noise compensation.

[0030] Furthermore, in the cross-modal safety shielding adapter, the process of performing safety correction is as follows:

[0031] ① Define the set of safe areas for inspection tasks as follows: ;

[0032] ② Ensure that the candidate actions output by the fast system reflection control layer satisfy the following inequality constraint:

[0033]

[0034] in, Let be the state transition function of the robot. It is a monotonically increasing function;

[0035] ③ Perform minimum intervention bias correction on candidate actions that do not satisfy the above inequality constraints.

[0036] Furthermore, the system provides generalized underlying support for heterogeneous power inspection robots through a configurable array of motion mapping operators, specifically:

[0037] (1) Project the high-dimensional action tokens onto the physical constraint space of the target robot using a hierarchical mapping matrix;

[0038] (2) For the inspection UAV carrier, the token is mapped to an attitude control vector based on the rotor speed increment;

[0039] (3) For the quadrupedal inspection robot, the token mapping is performed as a joint torque vector based on foot kinematics;

[0040] The mapping matrix automatically adjusts the scaling factor based on the robot's URDF model to compensate for the differences in dynamic response of different carriers.

[0041] Furthermore, the system also has a multi-dimensional, embodied inspection efficiency real-time evaluation function, and its quantitative evaluation index set includes:

[0042] Slow system cognitive consistency: measures the mutual information between the semantic goal generated by MLLM and the subsequent execution token stream;

[0043] Fast system control smoothness: Calculate the root mean square value of the second derivative of the actuator output sequence to assess the risk of mechanical wear of robot hardware;

[0044] Power safety violation redundancy: Quantifying the marginal distance distribution of the robot's real-time position deviation from the safety red line;

[0045] Task time-energy ratio: Calculates the number of power equipment inspections completed per unit of power consumption and the image coverage efficiency.

[0046] A control method for a power inspection robot based on a fast-slow system architecture and an embodied intelligent system, characterized by the following steps:

[0047] Step 1: Enter the original inspection task;

[0048] Step 2: Perform semantic logic parsing and temporal logic decomposition on the original inspection task, break down the original inspection task into an executable sub-task chain, and output semantic guidance features and security logic discriminant formulas, while dynamically maintaining long-term memory and short-term memory.

[0049] Step 3: Generate candidate action tokens based on the current sensor multimodal observation vectors, semantic guidance features, and short-term memory sequences, using the VLA model.

[0050] Step 4: Based on the candidate actions and the current latent space state, output the pre-simulated state of the environment and the physical consistency score for the next moment using the world model prediction engine;

[0051] Step 5: Real-time calculation of power safety distance constraints. Based on candidate actions, pre-simulation status and real-time safety distance constraints, determine whether candidate actions touch the power safety red line, and perform secondary programming solution through minimum deviation correction operator to output the corrected safe actions;

[0052] Step 6: The safety action is converted into the actuator control vector at the bottom layer of the target carrier through a hierarchical mapping matrix, and the drone rotor or robot joint is driven to complete the physical inspection action.

[0053] Step 7: Collect actual observations after execution and feed them back to Steps 2 and 3, and update the long-term memory and subsequent planned paths synchronously based on the deviation;

[0054] Step 8: Summarize all observation sequences and safety intervention logs, evaluate the system's performance throughout the entire task process, and output a PDF inspection report containing coordinates, timestamps, fault determination, and safety margin distribution.

[0055] Beneficial effects: (1) By using the multimodal large language model only for high-level semantic understanding, task decomposition and logical planning, the risk of large model logic illusion to the physical execution layer is reduced. (2) By outputting linear temporal logic constraints and semantic guidance features through the slow system, the robot has clear sequence and boundary conditions when performing inspection tasks such as "arrival first, then detection, maintaining a safe distance throughout the process, and reporting after completion", avoiding task omissions caused by relying solely on empirical rules. (3) By mapping multimodal observations such as vision, radar and infrared directly to action tokens through the fast system reflection control layer, the delay caused by the traditional "recognition-planning-control" link is shortened, and the robot's ability to avoid obstacles and compensate for posture in unstructured chemical conditions such as crosswinds, slippery ground and sudden obstacle appearance is improved. (4) By making short-term predictions before the actual execution of actions through the world model, the robot can first rehearse the consequences of candidate actions in the internal model; when collision, boundary crossing, sideslip or posture instability risks are predicted, the action can be suppressed or corrected in advance, thereby improving the safety redundancy in complex power scenarios. (5) By calculating the power safety distance in real time through the cross-modal safety shielding adapter and combining the Barrier function and quadratic programming to make the action with minimum deviation correction, the robot can ensure that it does not touch the safety red line of the live body while preserving the VLA action intention as much as possible. (6) By using a hierarchical mapping matrix to adapt the unified action token to the rotor speed of the UAV, the joint torque of the quadruped robot or the differential speed control of the wheeled robot, the same fast and slow system can be migrated to multiple power inspection carriers, reducing the cost of repeated development. (7) By continuously recording the task completion quality, control smoothness, safety margin and energy consumption efficiency through a multi-dimensional performance evaluation system, the execution deviation of a single inspection is fed back to the slow system, which can be used for subsequent path replanning, knowledge memory update and inspection strategy optimization. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the overall architecture of an embodied intelligent system for power inspection robots based on a fast-slow system architecture, according to the present invention.

[0057] Figure 2 This is a flowchart illustrating the logical planning process of the cognitive planning layer for a slow system according to the present invention.

[0058] Figure 3 This is a schematic diagram of the "prediction-feedback" closed loop inside the fast system reflection control layer of the present invention;

[0059] Figure 4 This is a schematic diagram illustrating the safety correction principle of the cross-modal safety shielding adapter of the present invention.

[0060] Figure 5 This is a schematic diagram of system interaction in a multi-machine collaborative emergency inspection scenario according to Embodiment 3 of the present invention. Detailed Implementation

[0061] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0062] like Figure 1 As shown, the embodied intelligent system for power inspection robots based on a fast-slow system architecture of the present invention includes:

[0063] Input layer: Input inspection tasks;

[0064] Slow system cognitive planning layer: Based on a multimodal large language model, the linear temporal logic (LTL) and power knowledge graph are used to perform semantic parsing and long-range logical planning for the inspection task;

[0065] Fast system reflex control layer: Based on the visual-language-action integrated VLA model and world model prediction engine with neural circuit structure, high-frequency end-to-end mapping from visual observation to action tokens is performed;

[0066] Cross-modal safety shielding adapter: By dynamically calculating the electrical safety distance and combining it with the Lyapunov Barrier function, candidate actions are subjected to hard physical red line verification and safety correction;

[0067] Robot execution layer: The verified and safety-corrected actions are transformed into the actuator control vectors of the target carrier at the bottom layer through a hierarchical mapping matrix. Real-time observations are collected and fed back to the slow system cognitive planning layer and the fast system reflection control layer. The slow system cognitive planning layer updates the long-term memory module and subsequent planning paths synchronously according to the deviation.

[0068] The slow system cognitive planning layer serves as the high-level logic hub, equipped with a multimodal large language model (MLLM), a symbolic reasoning engine, and a power expert knowledge graph. It is used to parse complex power inspection natural language tasks into long-term task plans with physical and logical constraints, and dynamically maintain a long short-term memory module containing historical fault characteristics and environmental priors.

[0069] like Figure 2 As shown, the slow system's cognitive planning layer uses neural symbolic computation to parse fuzzy inspection task instructions into linear temporal logic (LTL) formulas with strict spatiotemporal constraints. The expression for these LTL constraints is as follows:

[0070]

[0071] in, As a global operator, it means that the robot must satisfy the electrical safety clearance constraint and avoid obstacles throughout the entire process; For power safety clearance constraints; For obstacles; This is the final operator, indicating that the target inspection point must be completed; Target inspection points; This is the operator for the next time step, indicating that a status report must be triggered after the task is completed; For status reporting, the slow system's cognitive planning layer generates the optimal sub-task chain that satisfies the above formula through a heuristic search algorithm. This means that the goal must be completed and a report given while maintaining safety constraints throughout the process.

[0072] The operator is optimized based on operational compliance through contrastive physics learning, and the decision quality is improved by minimizing the contrastive loss function, the expression of which is:

[0073]

[0074] in, The features of the currently generated inspection plan, These are standard compliance operation characteristics extracted from the power safety expert system. These are characteristics of violations and misoperations found in the historical accident case database. This represents the temperature coefficient. By narrowing the gap between planning features and expert experience features in the embedding space, it ensures that the long-term plans generated by the large model strictly comply with power safety regulations.

[0075] The fast system reflection control layer, as a low-level execution reflection arc, is based on a visual-language-action integrated VLA model with a neural circuit structure. It realizes high-frequency end-to-end mapping from multi-sensor observation sequences to robot actuator tokens and has the ability to perform microsecond-level motion compensation and posture adjustment in unstructured power environments.

[0076] like Figure 3 As shown, the fast system's reflective control layer utilizes a causal Transformer structure to construct an action generation sequence. It generates control action tokens by maximizing the conditional log-likelihood estimate under the current observation sequence, and then linearly maps the generated tokens to motor torque control commands using a lightweight decoder. Its objective loss function is defined as:

[0077]

[0078] in, For the generated discrete action token, This is a sliding window observation sequence composed of visual depth maps, radar point clouds, and infrared features. This is a semantically guided target vector issued by a slow system. To record short-term characteristics of physical interaction states within the last three seconds, the fast system reflection control layer linearly maps the generated tokens to motor torque control commands using a lightweight decoder.

[0079] In the fast system reflection control layer, a world model prediction engine is introduced. This world model prediction engine is a variational latent space dynamics prediction engine based on a world model, and its state evolution equation is expressed as:

[0080]

[0081] in, For the encoded latent space environment features, For pre-execution actions, The latent random variable, which characterizes the uncertainty in the power scenario, follows a Gaussian distribution. The fast system's reflective control layer predicts multiple evolution trajectories in parallel and calculates the probability distribution of actions that cause the robot to enter an unstable state. If the expected value of the collapse probability is higher than the safety threshold, the system is triggered to switch to the logical defense mode controlled by the slow system's cognitive planning layer.

[0082] The cross-modal safety shielding adapter serves as a hard real-time monitoring operator between fast and slow systems. It has a built-in formal logic verification engine to intercept abnormal action tokens output by the fast system and to execute mandatory instruction rewriting and safety degradation strategies based on preset power safety red lines and robot dynamics boundaries.

[0083] like Figure 4 As shown, the cross-modal safety shielded adapter integrates a multi-criteria safety assessment model for special power operating conditions; it obtains the voltage level of the energized equipment in real time. Environmental inductive coupling strength Wind speed disturbance and air humidity Dynamic safety distance threshold The dynamic safety distance threshold calculation expression is as follows:

[0084]

[0085] in, and For environmental adaptability coefficients, The electric field breakdown characteristic index is used. To predict emergency braking displacement based on the robot's current momentum, This is for sensor measurement noise compensation; when the robot body coordinates With the surface of charged body distance At that time, the jammer immediately intervenes and forces an avoidance action.

[0086] The cross-modal safety shielding adapter possesses an embodied behavior safety guarantee mechanism based on the Lyapunov Barrier Function (CBF), and its safety correction process is as follows:

[0087] Define the set of safe areas for inspection tasks as follows ;

[0088] The candidate action output by the fast system reflection control layer The following inequality constraints must be satisfied:

[0089]

[0090] in, Let be the state transition function of the robot. It is a monotonically increasing function;

[0091] The safety shielding adapter solves the above quadratic programming problem and performs minimum intervention bias correction on candidate actions that do not satisfy the above inequality constraints, ensuring that the robot's inspection behavior is always within the safe reachable set.

[0092] The embodied intelligent system for power inspection robots based on a fast-slow system architecture of the present invention achieves generalized underlying support for heterogeneous power inspection robots through a configurable motion mapping operator array, specifically as follows:

[0093] Using hierarchical mapping matrix Project high-dimensional action tokens onto the physical constraint space of the target robot;

[0094] For inspection drones, the token is mapped to an attitude control vector based on the rotor speed increment. ;

[0095] For quadruped inspection robots, the execution token is mapped to a joint torque vector based on foot kinematics. ;

[0096] The mapping matrix The scaling factor is automatically adjusted based on the robot's URDF model to compensate for the differences in dynamic response between different carriers.

[0097] Furthermore, the embodied intelligent system for power inspection robots based on a fast-slow system architecture of the present invention also has a multi-dimensional real-time evaluation function for embodied inspection efficiency, and its quantitative evaluation index set includes:

[0098] Slow System Cognitive Consistency: Measuring the mutual information between the semantic goal generated by MLLM and the subsequent execution token stream. ;

[0099] Fast system control smoothness: Calculate the root mean square value of the second derivative of the actuator output sequence to assess the risk of mechanical wear of robot hardware;

[0100] Power safety violation redundancy: Quantifying the marginal distance distribution of the robot's real-time position deviation from the safety red line;

[0101] Task time-energy ratio: Calculates the number of power equipment inspections completed per unit of power consumption and the image coverage efficiency.

[0102] The embodied intelligent system for power inspection robots based on a fast-slow system architecture of the present invention includes the following stages:

[0103] <1> Slow system reasoning and planning phase: Receiving the original inspection task instruction set Power Expert Knowledge Base and global topology map As input, the cognitive planning layer outputs task description semantic guidance features containing temporal logical constraints. With formal security discriminant Satisfying the relation: , used to define the logical boundaries of inspection tasks;

[0104] <2> Fast system candidate action generation stage: using the current sensor multimodal observation vector Semantic guidance features issued by slow systems and short-term memory sequences As input, candidate action tokens are output through the VLA integrated model. Its generation process is represented as follows: , This refers to the parameters of the model; the token represents the initial instinctive movement decision.

[0105] <3> Brain evolutionary rehearsal stage: candidate actions With the current latent space state Input is fed into the world model prediction engine, which outputs a preview of the environment's state at the next moment. and physical consistency score The calculation formula is:

[0106]

[0107] in, As the random disturbance component of the environment, this stage enables advanced evaluation of the action results;

[0108] <4> Shielding device safety correction phase: The safety shielding adapter receives various candidate actions output by the fast system. The pre-simulation state output by the world model and real-time calculated power safety distance constraints As input, ,in, This indicates the predicted physical position of the robot at a future time after performing the candidate action. Represents the set of spatial coordinates of hazardous objects or charged bodies; outputs the corrected optimal safety action by solving a quadratic programming problem. :

[0109]

[0110] This phase ensures that the output action, while fulfilling the VLA intent, absolutely does not violate the physical safety red line of electrical circuits;

[0111] <5> Hardware layer mapping execution phase: through hierarchical mapping matrix Receive correction action This is transformed into the actuator control vector at the underlying level of the target carrier. ,satisfy: And drive the drone rotors or robot joints to complete physical inspection actions;

[0112] <6> Feedback and memory update phase: actual observations after system collection and execution. And it is fed back to the dual system, the slow system according to the deviation The long-term memory module and subsequent planned paths are updated synchronously to complete a complete sensing-driven calculation control closed loop.

[0113] This invention also provides a control method for a power inspection robot based on a fast-slow system architecture and an embodied intelligent system, comprising the following steps:

[0114] Step 1: Enter the original inspection task;

[0115] Step 2: Perform semantic logic parsing and temporal logic decomposition on the original inspection task, break down the original inspection task into an executable sub-task chain, and output semantic guidance features and security logic discriminant formulas, while dynamically maintaining long-term memory and short-term memory.

[0116] Step 3: Generate candidate action tokens based on the current sensor multimodal observation vectors, semantic guidance features, and short-term memory sequences, using the VLA model.

[0117] Step 4: Based on the candidate actions and the current latent space state, output the pre-simulated state of the environment and the physical consistency score for the next moment using the world model prediction engine;

[0118] Step 5: Real-time calculation of power safety distance constraints. Based on candidate actions, pre-simulation status and real-time safety distance constraints, determine whether candidate actions touch the power safety red line, and perform secondary programming solution through minimum deviation correction operator to output the corrected safe actions;

[0119] Step 6: The safety action is converted into the actuator control vector at the bottom layer of the target carrier through a hierarchical mapping matrix, and the drone rotor or robot joint is driven to complete the physical inspection action.

[0120] Step 7: Collect actual observations after execution and feed them back to Steps 2 and 3, and update the long-term memory and subsequent planned paths synchronously based on the deviation;

[0121] Step 8: Summarize all observation sequences and safety intervention logs, evaluate the system's performance throughout the entire task process, and output a PDF inspection report containing coordinates, timestamps, fault determination, and safety margin distribution.

[0122] Example 1:

[0123] The following example illustrates the system's operational logic using a UAV autonomous detailed inspection mission on a 110kV high-voltage transmission line. The mission is highly dangerous, testing the system's real-time control, obstacle avoidance capabilities, and basic mission planning abilities.

[0124] Step 1: The slow system receives the user instruction "Inspect the suspension insulators of towers 101 to 105 for cracks, maintaining a safe distance." The slow system retrieves the power expert knowledge graph, identifies the conductor voltage level, safe operating procedures, and tower latitude and longitude coordinates for this section, decomposes the task into five sub-waypoint tasks, and generates LTL task logic. This ensures that the logic operator always forces the "safe distance monitoring" mode to be enabled during the time the device stays on each tower.

[0125] Step 2: After the UAV arrives at the work site, the fast system's reflection control layer processes the visual flow data from the payload camera in real time. The VLA model is based on the current visual observations. With the target guidance issued by the slow system Combined with curing parameters The token sequence used to adjust the pose is calculated. At this point, the world model simulates the perturbation of the fuselage by high-altitude crosswinds in latent space, using state prediction equations. Predict the trajectory deviation within the next second.

[0126] Step 3: The safety shielding adapter monitors the drone's coordinates in real time. With live wires The spatial distance. If a gust of wind causes the fuselage to shift towards the guide line, the distance calculated by the system... Less than the dynamic threshold At this point, the adapter intervenes and, using the minimum deviation correction formula, corrects the original forward token action to a "minor backward movement followed by a hovering" correction. Furthermore, this correction process is transparent to the higher-level logic of the slow system, ensuring millisecond-level real-time control response.

[0127] Step 4: The corrected token is mapped through the matrix. Converted into rotor differential speed signal The system is then deployed. The evaluation system scores the data in real time. If an abnormally high Physical Red Line Trigger Rate (PRR) is detected, feedback is sent to the slower system for long-range path replanning, achieving closed-loop optimization between the two systems.

[0128] Example 2:

[0129] The following example illustrates the system's heterogeneous carrier adaptability using a quadruped robot autonomously reading meters and verifying circuit breaker status within a substation. The task is of moderate risk, testing the system's basic control, route planning, and topology logic capabilities.

[0130] Step 1: The slow system receives the task "Inspect all oil-immersed instrument transformer level gauges in the substation and verify the closing position." The slow system uses a power topology map to logically layer the substation equipment and plans an inspection route that avoids areas with strong electromagnetic interference.

[0131] Step 2: During the quadruped robot's walking process, the fast system's reflection control layer transmits the generated action tokens through a mapping matrix. Projected onto the joint dynamics space of the target vehicle, the foot coordinate trajectory and joint torque vector are output. The VLA model automatically adjusts the fuselage pitch angle to align with the visual axis by recognizing features on the meter panel.

[0132] Step 3: When traversing uneven gravel ground or cable trench covers, the world model engine predicts multiple physical evolution trajectories in parallel. If the prediction indicates that the current foot landing point will cause instability of the center of gravity, i.e., the state prediction entropy value is too high, the fast system will suppress the original action and automatically switch to the "high-frequency small step" steady-state mode based on world model feedback, ensuring that the robot always stays within the safe reach set.

[0133] Step 4: The safety shielding adapter verifies the distance between the robot and live components such as transformer bushings in real time to ensure that the robot does not cross the safety warning line during meter reading. After the task is completed, the system compares the characteristics of the generated inspection report. Compliance features in expert knowledge graphs The PLC consistency index is calculated. If the PLC score meets the standard, the result is synchronized to the cloud storage. The quadruped robot then executes the long-range return plan of the slow system.

[0134] Example 3:

[0135] like Figure 5 As shown, taking the task of "multi-machine collaborative emergency inspection after a suspected partial discharge anomaly alarm of a 500kV main transformer" in a substation as an example, this paper details the specific operation process and collaborative mechanism of this system when performing long-distance and highly complex tasks.

[0136] Step 1: Slow System Long-Term Task Analysis and Collaborative Logic Construction. Receive emergency command: "Main transformer No. 1 cabinet has an abnormal discharge alarm. Collaborate with drones and wheeled robots for a comprehensive inspection; the drone will perform infrared scanning of the upper bushings and oil conservator, and the wheeled robot will inspect the bottom oil valves and cooling system. Maintain a 500kV safety clearance throughout the process. If oil leakage is found, immediately mark it and request a decision from the slow system." The slow system retrieves the power expert knowledge graph. The system identifies the three-dimensional topological nodes, energized parts distribution, and a 5-meter hard safety distance of the 500kV transformer. The slow system utilizes a symbolic reasoning engine to generate multi-machine collaborative LTL task logic: first, it triggers the UAV to take off and perform high-level obstacle avoidance detection. Then guide the wheeled robot into the work area below. The slow system transforms the task chain into semantic guidance features for each robot. and And inject the LTL formula into the safety shield of each machine. .

[0137] Step 2: High-frequency induction drive closed loop of the fast system reflective control layer. The UAV and wheeled robot respectively activate their VLA models. While the UAV performs high-altitude patrol maneuvers, the fast system reflective layer processes 4K visible light and infrared dual-channel observations in real time. The VLA model is based on an empirical parameter set. The "shooting sleeve" command is tokenized into tilt compensation and attitude maintenance actions. At this point, the world model engine evolves and pre-demonstrates the equations in real time. The system anticipates the impact of the strong local airflow generated by the high-speed rotation of the transformer cooling fan on the drone's flight path. If the prediction indicates a 0.3-meter flight path drift in the next second, the fast system's reflective layer immediately performs minor compensation at the action token level to ensure the image remains aligned with the insulating sleeve node.

[0138] Step 3: Command rewriting and safety intervention of the shielding device in a dynamic environment. When the wheeled robot travels in the narrow area at the bottom of the transformer, it must not only avoid various ground oil pipes, but also maintain a constant distance from the 500kV live busbar suspended above. Real-time calculation of the safety shielding adapter. Its parameters vary with the ambient humidity. With electric field strength Dynamic fluctuations. When the robot slips due to wet ground, its coordinates... When approaching the safety threshold, the shield immediately calculates the quadratic programming objective function. Although the VLA model was originally scheduled to continue moving forward to perform infrared imaging, the jammer, based on safety constraints, forcibly corrected the action token to "lock in place, brake, and reverse 0.5 meters." This correction command was transmitted through the mapping matrix. The differential motor driver is directly deployed to the wheeled robot to ensure that physical safety takes precedence over business logic.

[0139] Step 4: Dynamic Replanning and Knowledge Synchronization of the Slow System. During the inspection, the UAV detects an abnormal temperature rise below the oil tank using infrared visual flow. The VLA model semanticizes this visual feature and feeds it back to the slow system's cognitive planning layer. The slow system compares this with the power expert knowledge graph, assesses that the temperature rise is accompanied by an oil leakage risk, and immediately performs task replanning. The slow system suspends the wheeled robot's routine inspection task and issues new logical guidance features. They were instructed to immediately move to the area below the suspected oil leak point for high-definition video evidence collection. During this process, the slow system dynamically updates its long short-term memory module. Record the correlation between this discharge alarm and oil leakage.

[0140] Step 5: Task Closure Assessment and Compliance Archiving. Upon completion of the task, the slow system aggregates all observation sequences uploaded by the two robots. Along with safety intervention logs, the evaluation system performs a performance audit of the entire task process: PLC consistency assessment shows that the corrected action flow perfectly achieved the initial intent of "detecting oil leakage"; the PRR index records the intervention frequency of the shield under strong electric field conditions. The slow system ultimately generates a PDF inspection report containing coordinates, timestamps, fault determination, and safety margin distribution. The robot then executes the evacuation plan generated by the slow system and autonomously returns to the hangar using the obstacle avoidance reflection algorithm of the fast system.

[0141] The above long-range embodiments demonstrate how the present invention achieves autonomous and safe multi-machine collaborative operation in emergency inspection tasks with extremely high voltage and multiple interferences through advanced planning of slow systems, physical instincts of fast systems and rigid constraints of shielding devices.

[0142] The above embodiments demonstrate the technical advantages of this invention in ensuring the real-time performance, versatility, and safety of power line inspection through the organic combination of fast and slow systems, world models, and safety shields.

[0143] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power inspection robot body intelligence system based on fast-slow system architecture, characterized in that, include: Input layer: Input inspection tasks; Slow system cognitive planning layer: Based on a multimodal large language model, the linear temporal logic (LTL) and power knowledge graph are used to perform semantic parsing and long-range logical planning for the inspection task; Fast system reflex control layer: Based on the visual-language-action integrated VLA model and world model prediction engine with neural circuit structure, high-frequency end-to-end mapping from visual observation to action tokens is performed; Cross-modal safety shielding adapter: By dynamically calculating the electrical safety distance and combining it with the Lyapunov Barrier function, candidate actions are subjected to hard physical red line verification and safety correction; Robot execution layer: The verified and safety-corrected actions are transformed into the actuator control vectors of the target carrier at the bottom layer through a hierarchical mapping matrix. Real-time observations are collected and fed back to the slow system cognitive planning layer and the fast system reflection control layer. The slow system cognitive planning layer updates the long-term memory module and subsequent planning paths synchronously according to the deviation.

2. The embodied intelligent system for power inspection robots based on a fast-slow system architecture as described in claim 1, characterized in that, The slow system cognitive planning layer parses the inspection task into a linear temporal logic with spatiotemporal constraints through neural symbolic computation. The expression of the linear temporal logic is as follows: in, As a global operator, it means that the robot must satisfy the electrical safety clearance constraint and avoid obstacles throughout the entire process; For power safety clearance constraints; For obstacles; This is the final operator, indicating that the target inspection point must be completed; Target inspection points; This is the operator for the next time step, indicating that a status report must be triggered after the task is completed; For status reporting.

3. The embodied intelligent system for power inspection robots based on a fast-slow system architecture according to claim 2, characterized in that, The operator is optimized based on operational compliance through contrastive physics learning, and the decision quality is improved by minimizing the contrastive loss function, the expression of which is: in, The features of the currently generated inspection plan, These are standard compliance operation characteristics extracted from the power safety expert system. These are characteristics of violations and misoperations found in the historical accident case database. This is the temperature coefficient.

4. The embodied intelligent system for power inspection robots based on a fast-slow system architecture as described in claim 1, characterized in that, The fast system reflection control layer utilizes a causal Transformer structure to construct an action generation sequence. It generates control action tokens by maximizing the conditional log-likelihood estimate under the current observation sequence, and then linearly maps the generated tokens to motor torque control commands using a lightweight decoder. Its objective loss function is defined as: in, For the generated discrete action token, This is a sliding window observation sequence composed of visual depth maps, radar point clouds, and infrared features. This is a semantically guided target vector issued by a slow system. To record short-term characteristics of physical interaction states within the last three seconds.

5. The embodied intelligent system for power inspection robots based on a fast-slow system architecture according to claim 1, characterized in that, In the fast system reflection control layer, the world model prediction engine is a variational latent space dynamics prediction engine based on the world model, and its state evolution equation is expressed as: in, For the encoded latent space environment features, For pre-execution actions, The latent random variable, which characterizes the uncertainty in the power scenario, follows a Gaussian distribution. .

6. The embodied intelligent system for power inspection robots based on a fast-slow system architecture according to claim 1, characterized in that, The cross-modal safety shielding adapter integrates a multi-criteria safety assessment model for special power operating conditions; the dynamic calculation of the power safety distance is achieved by acquiring the voltage level of the energized equipment in real time. Environmental inductive coupling strength Wind speed disturbance and air humidity The dynamic safety distance threshold is calculated, and its expression is as follows: in, and For environmental adaptability coefficients, The electric field breakdown characteristic index is used. To predict emergency braking displacement based on the robot's current momentum, This is for sensor measurement noise compensation.

7. The embodied intelligent system for power inspection robots based on a fast-slow system architecture according to claim 1, characterized in that, In the cross-modal safety shielding adapter, the process of performing safety correction is as follows: ① Define the set of safe areas for inspection tasks as follows: ; ② Ensure that the candidate actions output by the fast system reflection control layer satisfy the following inequality constraint: in, Let be the state transition function of the robot. It is a monotonically increasing function; ③ Perform minimum intervention bias correction on candidate actions that do not satisfy the above inequality constraints.

8. The embodied intelligent system for power inspection robots based on a fast-slow system architecture according to claim 1, characterized in that, The system provides generalized underlying support for heterogeneous power inspection robots through a configurable array of motion mapping operators, specifically: (1) Project the high-dimensional action tokens onto the physical constraint space of the target robot using a hierarchical mapping matrix; (2) For the inspection UAV carrier, the token is mapped to an attitude control vector based on the rotor speed increment; (3) For the quadrupedal inspection robot, the token mapping is performed as a joint torque vector based on foot kinematics; The mapping matrix automatically adjusts the scaling factor based on the robot's URDF model to compensate for the differences in dynamic response of different carriers.

9. The embodied intelligent system for power inspection robots based on a fast-slow system architecture according to claim 1, characterized in that, The system also has a multi-dimensional, embodied inspection efficiency real-time evaluation function, and its quantitative evaluation index set includes: Slow system cognitive consistency: measures the mutual information between the semantic goal generated by MLLM and the subsequent execution token stream; Fast system control smoothness: Calculate the root mean square value of the second derivative of the actuator output sequence to assess the risk of mechanical wear of robot hardware; Power safety violation redundancy: Quantifying the marginal distance distribution of the robot's real-time position deviation from the safety red line; Task time-energy ratio: Calculates the number of power equipment inspections completed per unit of power consumption and the image coverage efficiency.

10. A control method for a power inspection robot based on the embodied intelligent system of a power inspection robot according to the fast / slow system architecture of claim 1, characterized in that, Includes the following steps: Step 1: Enter the original inspection task; Step 2: Perform semantic logic parsing and temporal logic decomposition on the original inspection task, break down the original inspection task into an executable sub-task chain, and output semantic guidance features and security logic discriminant formulas, while dynamically maintaining long-term memory and short-term memory. Step 3: Generate candidate action tokens based on the current sensor multimodal observation vectors, semantic guidance features, and short-term memory sequences, using the VLA model. Step 4: Based on the candidate actions and the current latent space state, output the pre-simulated state of the environment and the physical consistency score for the next moment using the world model prediction engine; Step 5: Real-time calculation of power safety distance constraints. Based on candidate actions, pre-simulation status and real-time safety distance constraints, determine whether candidate actions touch the power safety red line, and perform secondary programming solution through minimum deviation correction operator to output the corrected safe actions; Step 6: The safety action is converted into the actuator control vector at the bottom layer of the target carrier through a hierarchical mapping matrix, and the drone rotor or robot joint is driven to complete the physical inspection action. Step 7: Collect actual observations after execution and feed them back to Steps 2 and 3, and update the long-term memory and subsequent planned paths synchronously based on the deviation; Step 8: Summarize all observation sequences and safety intervention logs, evaluate the system's performance throughout the entire task process, and output a PDF inspection report containing coordinates, timestamps, fault determination, and safety margin distribution.

Images