Unmanned aerial vehicle and robot linkage power grid intelligent patrol method based on AI large model

Abstract

The application provides an unmanned aerial vehicle and robot linkage power grid intelligent patrol method based on an AI large model, relates to the technical field of power grid intelligent inspection, and comprises the following steps: a double-layer game model of an aircraft and a robot is established, competition for task execution right is carried out, information is cooperatively shared, and a dynamic task allocation scheme is generated. Differentiation perception is executed and a complete device image is synthesized, model parameters are updated according to execution feedback to optimize the competition and cooperation balance, and the patrol coverage and identification accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent power grid inspection technology, and in particular to an intelligent power grid inspection method based on AI large-scale models that combines drones and robots. Background Technology

[0002] In the field of power system operation and maintenance, regular inspections of transmission and transformation equipment are a crucial link in ensuring the safe and stable operation of the power grid. Current conventional practices mainly rely on manual inspections or single types of automated equipment. Manual inspections require personnel to trek through mountains and rivers to conduct close-up checks on equipment such as transmission lines, towers, and substations, which is inefficient and poses safety risks. With technological advancements, using drones for aerial inspections or ground robots for point-to-point testing is gradually becoming mainstream. Drones can quickly cover vast areas and acquire macroscopic images of equipment; ground robots can approach equipment and perform more detailed local inspections, such as using infrared thermal imagers to detect heat points or high-definition cameras to capture surface defects.

[0003] However, the aforementioned conventional single-platform or simple combination inspection modes have significant drawbacks. On the one hand, the actions of drones and ground robots are often planned independently, lacking a systematic coordination mechanism. This can easily lead to unreasonable task allocation; for example, multiple platforms may compete to go to high-value equipment points while neglecting other areas, or repeatedly inspect the same equipment due to a lack of information sharing, resulting in resource waste and reduced inspection coverage. On the other hand, the inspection process is usually static or pre-programmed, making it difficult to adapt to dynamically changing field environments. Factors such as equipment status, platform energy consumption, and communication conditions constantly change during task execution. Fixed task allocation and path planning strategies cannot be optimized and adjusted in real time, leading to inspection task interruptions, missed critical defects, or overall inspection efficiency failing to reach its optimal level. Summary of the Invention

[0004] This invention provides a method for intelligent power grid inspection using drones and robots based on an AI large model, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for intelligent power grid inspection using a combination of drones and robots based on an AI large-scale model, comprising: Obtain the spatial distribution topology of the power grid, the importance level of equipment, and the initial position and energy state of aircraft and robots; The aircraft is defined as a global vision acquisition agent, and the robot is defined as a local depth exploration agent. A two-layer game model of multiple agents is established. In the first layer game, the aircraft and the robot construct utility functions through equipment movement costs, energy status and task benefits, and compete for the right to execute the target value task. In the second layer game, the equipment that wins the right to execute maximizes the overall inspection coverage by cooperating and sharing perception data and path information, and generates a dynamically balanced task allocation scheme and execution sequence. Based on the task allocation scheme, differentiated perception is performed. The aircraft collects wide-area images and extracts the equipment outline and layout information through target detection. The robot collects close-range data and extracts surface texture and heat distribution information through fine-grained analysis. The complete equipment image is synthesized through the feature cascade module. Record the energy consumption rate, communication latency, and anomaly detection rate during task execution. Calculate the difference between actual and expected revenue as the reward signal to update the utility function weight and collaboration strategy threshold. Adjust the task revenue calculation rules inversely based on the accuracy of anomaly identification in the device profile, driving continuous optimization of the competitive and collaborative balance between the two platforms.

[0006] In the first level of the game, aircraft and robots construct utility functions based on equipment movement costs, energy states, and task benefits, competing for the right to execute a target-value task, including: Based on the initial position and current energy state of the aircraft and robot, the spatial distribution topology is dynamically divided into reachable domains, generating reachable task domains for the aircraft and robots. The task nodes in the two domains are sorted according to the importance level of the equipment, and the respective reachable target value task sequences are output. Taking the achievable target value task sequence as input, the equipment movement cost of each task node in the arrival sequence of the aircraft and robot is extracted respectively. An energy feasibility coefficient with energy state as a parameter is introduced to impose an upper limit constraint on the equipment movement cost, and task nodes that exceed the energy feasibility boundary are filtered out to generate a candidate task set filtered by energy. Based on the gain mapping relationship between the required perception depth of each task node in the candidate task set and the task benefits, the task benefit gain value of the aircraft and the robot for each task node in the candidate task set is calculated respectively. The difference between the task benefit gain value and the equipment movement cost is substituted into the utility function to calculate the competitive utility value. An exclusive competitive allocation is performed based on the competitive utility value. The party with the highest utility value obtains the execution right of the corresponding task node. In the case of the same competitive utility value, the energy state surplus is used as the arbitration basis to complete the unique ownership determination. The execution right allocation result is written back to the candidate task set to complete the state closure.

[0007] In the second-level game, the device that gains execution rights maximizes overall inspection coverage by collaboratively sharing sensing data and path information, generating a dynamically balanced task allocation scheme and execution sequence, including: Each device that obtains execution rights will broadcast the collected sensing data and current path information. The spatial coverage distribution characteristics of the sensing data and the trajectory extension directionality of the path information will be used as the fusion basis, and the density gradient of the covered area and the gap coordinates of the uncovered area will be used as the status output to construct a global coverage density map. Using the global coverage density map as input, we extract the coverage gap area and low-density coverage area. Based on the spatial continuity range of each area, the perception importance weight and the current accessibility of the device that has obtained the execution right, we score the coverage urgency of each area and obtain a coverage completion demand sequence with priority tags. Using the complete demand sequence as the allocation source, the real-time location, path information and remaining energy status of each device that obtains the execution right as the three-element allocation constraint, and the difference between the perception adaptation gain of each device to each demand node and the movement cost to reach the node, an allocation benefit evaluation function is constructed to dynamically assign each demand node and generate a dynamically balanced task allocation scheme. Using the task allocation scheme as the timing orchestration input, the priority markers and movement costs of each device's assigned requirement node are extracted. The execution timing of each device is generated by using the priority markers to apply pre-excitation to the execution timing and the movement costs to apply delay suppression to the execution timing as bidirectional orchestration rules.

[0008] Using the task allocation scheme as the timing orchestration input, the priority markers and movement costs of each device's assigned request node are extracted. The execution timing of each device is generated using bidirectional orchestration rules: priority markers apply pre-incentives to the execution timing, and movement costs apply delays to the execution timing. This includes: Using the task allocation scheme as the parsing input, the priority mark quantization value and the inter-node movement cost quantization value of each device's assigned demand node are extracted. The priority mark quantization value is the pre-excitation amplitude and the movement cost quantization value is the post-suppression amplitude. The sign of the difference between the pre-excitation amplitude and the post-suppression amplitude of each node determines its pre-excitation attribute or post-suppression attribute, generating the pre-excitation attribute node sequence and the post-suppression attribute node sequence of each device. Using the pre-excitation attribute node sequence and the post-suppression attribute node sequence of each device as the timing filling input, and using the dual-attribute filling rule of filling the pre-excitation attribute node sequence in descending order of pre-excitation magnitude to the front end of the execution timing and filling the post-suppression attribute node sequence in ascending order of post-suppression magnitude to the back end of the execution timing, candidate execution timings for each device are generated. The candidate execution sequence of each device is used as the input for amplitude continuity verification. The cumulative value of the pre-excitation amplitude of each node in the entire candidate execution sequence is not lower than the cumulative value of the delayed suppression amplitude as the amplitude continuity judgment criterion. If the verification fails, the bidirectional amplitude is recalculated for the nodes that do not meet the criterion and a new round of dual attribute filling is triggered. The cycle continues until the verification is passed. The candidate execution sequence that passes the verification is confirmed as the execution sequence of each device.

[0009] Based on the task allocation scheme, differentiated perception is performed. The aircraft acquires wide-area imagery and extracts equipment contour and layout information through target detection. The robot acquires close-range data and extracts surface texture and heat distribution information through fine-grained analysis. A complete equipment profile is synthesized through a feature cascade module, including: The aircraft performs wide-area image acquisition according to the mission allocation plan, applies target detection processing to the wide-area image, extracts contour geometric features and layout topology features with the equipment response area as the boundary, decomposes the contour geometric features into global contour features and local boundary features according to the perception scale, and aggregates them together with the layout topology features into a macroscopic structural feature set with a hierarchical structure. The robot performs close-range data acquisition according to the task allocation plan, applies fine-grained analysis to the acquired data, and extracts surface texture anisotropic features and thermal distribution thermal field gradient features using material response boundary and thermal field temperature gradient boundary as dual segmentation constraints. The anisotropic features are decomposed into coarse texture features and fine texture features according to texture frequency, and together with the thermal field gradient features, they are aggregated into a set of microscopic physical features with hierarchical structure. The feature cascade module takes macroscopic structural feature set and microscopic physical feature set as input, uses global contour feature quantity and thermal field gradient feature quantity to form macroscopic corresponding layer, and uses local boundary feature quantity and fine texture feature quantity to form microscopic corresponding layer. It calculates cross-modal feature complementarity layer by layer, and uses complementarity as weighting basis to drive cross-layer feature fusion to generate a complete device profile.

[0010] The difference between actual and expected revenue is calculated as the reward signal to update the utility function weight and collaboration strategy threshold. Furthermore, the task revenue calculation rules are adjusted inversely based on the accuracy of anomaly identification in the device profile, driving continuous optimization of the competitive and collaborative balance between the two platforms. This includes: The difference between the actual and expected benefits of the aircraft and the robot is used to synthesize a joint benefit deviation vector. The deviation distance between the current task allocation scheme and the Nash equilibrium solution set is calculated using the joint benefit deviation vector. The equilibrium deviation driving signal is constructed using the magnitude of the deviation distance. The equilibrium deviation driving signal amplitude drives the gradient update of the moving cost weight component and energy state weight component in the utility function towards the Nash equilibrium solution set. The deviation direction drives the reverse compensation adjustment of the path sharing trigger threshold in the cooperation strategy threshold, and outputs the updated utility function weight and cooperation strategy threshold. The identification deviation is constructed by the difference between the accuracy of the device image anomaly identification and the benchmark value. The Pareto improvement feasible region is constructed by the identification deviation. It is determined whether the wide-area image benefit coefficient and the near-field data benefit coefficient are located at the boundary of the Pareto feasible region. The two types of benefit coefficients are adjusted in the opposite direction and magnitude of deviation from the Pareto boundary to converge towards the Pareto optimum. The updated task benefit calculation rules are output. The termination condition is determined by the combined criteria of the Nash equilibrium convergence radius and the Pareto boundary distance, driving the dual platforms to continuously optimize until the effectiveness of feature complementarity is stably improved.

[0011] A second aspect of this invention provides a smart power grid inspection system based on an AI large-scale model, integrating drones and robots, comprising: The information acquisition unit is used to acquire the spatial distribution topology of the power grid, the importance level of equipment, and the initial position and energy state of aircraft and robots; The game modeling unit defines the aircraft as a global vision acquisition agent and the robot as a local depth exploration agent. It establishes a two-layer game model for multiple agents. In the first layer of the game, the aircraft and the robot construct utility functions based on equipment movement costs, energy states, and task benefits, and compete for the right to execute the target value task. In the second layer of the game, the device that wins the right to execute maximizes the overall inspection coverage by collaborating and sharing perception data and path information, and generates a dynamically balanced task allocation scheme and execution sequence. The perception synthesis unit is used to perform differentiated perception based on the task allocation scheme. The aircraft collects wide-area images and extracts the equipment outline and layout information through target detection. The robot collects close-range data and extracts surface texture and heat distribution information through fine-grained analysis. The complete equipment image is synthesized through the feature cascade module. The strategy optimization unit records the energy consumption rate, communication latency, and anomaly detection rate during task execution. It calculates the difference between actual and expected revenue as the reward signal to update the utility function weight and collaboration strategy threshold. It also adjusts the task revenue calculation rules in reverse based on the accuracy of anomaly identification in the device profile, driving the dual platforms to continuously optimize the competitive and collaborative balance.

[0012] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0014] This method constructs a two-layer game model between an aircraft and a robot, achieving dynamic optimization of task allocation and efficient resource allocation. The aircraft, acting as a global vision agent, and the robot, acting as a local depth-sensing agent, compete for the right to execute high-value tasks through utility functions, effectively balancing equipment movement costs, energy status, and task benefits. The device that wins the execution right collaborates in the second-layer game, sharing perception data and path information, significantly improving the overall inspection coverage, avoiding task overlap or omissions, and forming a dynamically balanced task allocation and execution sequence.

[0015] This method performs differentiated perception, combining the advantages of wide-area and local inspection. Wide-area images acquired by the aircraft can be quickly extracted for macroscopic contours and layout information of equipment through target detection, while close-range data acquired by the robot can be accurately obtained for microscopic conditions such as surface texture and heat distribution through fine-grained analysis. By synthesizing a complete equipment profile through feature cascading modules, a comprehensive and multi-level state perception of power grid equipment from the overall picture to details is achieved, significantly improving the completeness and accuracy of state assessment.

[0016] The system has the ability to continuously self-optimize. It uses the accuracy of anomaly identification in the device profile to adjust the task reward calculation rules in reverse, so that the task allocation strategy can adapt to environmental changes and performance feedback, gradually converge to a better resource allocation scheme, and ensure long-term inspection efficiency and robustness. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the intelligent power grid inspection method based on an AI large model using drones and robots, as described in an embodiment of the present invention. Figure 2 A flowchart illustrating the process of generating a dynamically balanced task allocation scheme and execution timing method for embodiments of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1This is a flowchart illustrating the intelligent power grid inspection method based on an AI large-scale model using drones and robots, as described in this embodiment of the invention. Figure 1 As shown, the method includes: Obtain the spatial distribution topology of the power grid, the importance level of equipment, and the initial position and energy state of aircraft and robots; The aircraft is defined as a global vision acquisition agent, and the robot is defined as a local depth exploration agent. A two-layer game model of multiple agents is established. In the first layer game, the aircraft and the robot construct utility functions through equipment movement costs, energy status and task benefits, and compete for the right to execute the target value task. In the second layer game, the equipment that wins the right to execute maximizes the overall inspection coverage by cooperating and sharing perception data and path information, and generates a dynamically balanced task allocation scheme and execution sequence. Based on the task allocation scheme, differentiated perception is performed. The aircraft collects wide-area images and extracts the equipment outline and layout information through target detection. The robot collects close-range data and extracts surface texture and heat distribution information through fine-grained analysis. The complete equipment image is synthesized through the feature cascade module. Record the energy consumption rate, communication latency, and anomaly detection rate during task execution. Calculate the difference between actual and expected revenue as the reward signal to update the utility function weight and collaboration strategy threshold. Adjust the task revenue calculation rules inversely based on the accuracy of anomaly identification in the device profile, driving continuous optimization of the competitive and collaborative balance between the two platforms.

[0021] In the first level of the game, aircraft and robots construct utility functions based on equipment movement costs, energy states, and task benefits, competing for the right to execute a target-value task, including: Based on the initial position and current energy state of the aircraft and robot, the spatial distribution topology is dynamically divided into reachable domains, generating reachable task domains for the aircraft and robots. The task nodes in the two domains are sorted according to the importance level of the equipment, and the respective reachable target value task sequences are output. Taking the achievable target value task sequence as input, the equipment movement cost of each task node in the arrival sequence of the aircraft and robot is extracted respectively. An energy feasibility coefficient with energy state as a parameter is introduced to impose an upper limit constraint on the equipment movement cost, and task nodes that exceed the energy feasibility boundary are filtered out to generate a candidate task set filtered by energy. Based on the gain mapping relationship between the required perception depth of each task node in the candidate task set and the task benefits, the task benefit gain value of the aircraft and the robot for each task node in the candidate task set is calculated respectively. The difference between the task benefit gain value and the equipment movement cost is substituted into the utility function to calculate the competitive utility value. An exclusive competitive allocation is performed based on the competitive utility value. The party with the highest utility value obtains the execution right of the corresponding task node. In the case of identical competitive utility values, the energy state margin is used as the arbitration basis to determine unique ownership. The execution right allocation result is written back to the candidate task set to complete the state closure. When constructing the first-layer game competition mechanism, it is necessary to dynamically divide the reachability domain of the power grid equipment and sort the task nodes. Based on the initial position and energy state of the UAV and robot, the power grid inspection space is dynamically divided into reachability domains. Starting from power tower A, the current power of the UAV is collected as 85%, and the power of the robot is 70%. Through spatial topology analysis algorithm, the inspection range that each device can cover under energy constraints is calculated. Drones, leveraging their flight capabilities, can cover high-altitude power grid facilities within a 300-meter radius, including suspended components such as high-voltage line insulators and surge arresters. Robots, on the other hand, cover ground facilities within a 120-meter radius, such as transformer bases and grounding devices. According to the State Grid equipment classification standards, equipment within the reachable area is categorized by importance: Level 1 equipment (core transmission towers, main transformers), Level 2 equipment (secondary lines, distribution equipment), and Level 3 equipment (auxiliary monitoring equipment). A value weight for each piece of equipment is calculated using an importance-level mapping function, resulting in a drone-reachable target value task sequence {T1, T3, T5, T8} and a robot-reachable target value task sequence {T2, T4, T6, T7}. Here, T1 represents the highest-value main line insulator inspection task, and T2 represents the substation perimeter grounding device inspection task.

[0022] Energy feasibility analysis and candidate task filtering were conducted. Based on the previously generated sequence of achievable target value tasks, the movement costs of the UAV and robot to each task node were extracted. The movement costs were calculated using a path planning algorithm, taking into account equipment speed, terrain complexity, and obstacle avoidance costs. The movement costs of the UAV from its current position to T1, T3, T5, and T8 were 10 units, 15 units, 25 units, and 30 units of energy, respectively; the movement costs of the robot to T2, T4, T6, and T7 were 8 units, 12 units, 20 units, and 22 units of energy, respectively. An energy feasibility coefficient η was introduced, which is a nonlinear function of the ratio of the equipment's current energy to its full-load energy, reflecting the constraint relationship between remaining energy and the tolerable movement cost. When the UAV's energy state was 85%, the value of η was 0.78, indicating that the maximum tolerable movement cost was 32 units of energy; when the robot's energy state was 70%, the value of η was 0.65, indicating that the maximum tolerable movement cost was 26 units of energy. After energy filtering, the candidate task set for UAVs retains {T1, T3, T5} and removes T8; the candidate task set for robots retains {T2, T4, T6} and removes T7.

[0023] This study calculates the task benefit gain and competitive utility value. Power grid equipment inspection requires different perception depths, affecting inspection quality and fault detection rate. Perception depth is divided into five levels: visual recognition, infrared thermal imaging, ultrasonic detection, electromagnetic field measurement, and multimodal fusion analysis. For each task node in the candidate task set, the mapping relationship between the required perception depth and expected benefit is analyzed. For example, in the T1 main line insulator inspection task, a drone equipped with high-definition cameras and infrared thermal imaging equipment can achieve a level 4 perception depth with an expected benefit of 45 units; a robot, only providing visual recognition, achieves a level 1 perception depth with an expected benefit of only 15 units. The study calculates the benefit gain value of the equipment for each task, i.e., the difference between the actual benefit brought by the equipment's perception capability and the basic benefit. The drone's gain value for T1 is 30 units, and the robot's is 0 units. The final competitive utility value is calculated by subtracting the movement cost from the gain value and then multiplying it by the task priority coefficient. The competitive utility value of the drone for T1 is 20 × 1.5 = 30, and the competitive utility value of the robot for T1 is (-10) × 1.5 = (-15). Similarly, the competitive utility value matrix of all task nodes is calculated. The process involves exclusive competitive allocation and task attribution confirmation. Based on the competitive utility value, a task allocation decision is made. For each task node, the device with the highest utility value obtains the execution right. For example, the execution right of T1 is allocated to the drone (utility value 30 is higher than -15), and the execution right of T2 is allocated to the robot (utility value 24 is higher than 12). When there are cases where the competitive utility values ​​are the same, such as T3 being calculated to have the same utility value of 22 by both the drone and the robot, the energy state margin is used as the arbitration basis. The drone has 70% remaining energy, and the robot has 58% remaining energy. Therefore, the execution right of T3 is allocated to the drone. After completing the attribution determination of all task nodes, the results are written back to the candidate task set, the task status is marked as "allocated", and the execution device identifier is recorded, forming a closed-loop task allocation scheme.

[0024] In a practical application scenario, a provincial power grid company deployed this method for intelligent inspection of transmission lines. Initially, drones and robots departed from substations, automatically defining reachable areas and filtering tasks. Drones were responsible for tasks such as high-altitude insulator infrared scanning and conductor wear detection, while robots performed tower base stability checks and grounding device corrosion assessments. Through a competitive mechanism, inspection tasks were dynamically adjusted; when dense vegetation was found around a tower, the task was automatically reassigned from the drone to a more suitable robot, improving inspection efficiency and fault detection rate.

[0025] This intelligent power grid inspection method, based on dynamic accessibility partitioning and a competitive utility model, which combines drones and robots, effectively solves the problems of energy waste and blind spots in traditional fixed allocation methods, improving the overall coverage and inspection quality of power grid facilities. The method is highly adaptable, capable of adjusting task allocation strategies in real time according to changes in equipment energy status and sensing capabilities, ensuring maximum value for power grid inspection under limited energy conditions.

[0026] Figure 2 This is a flowchart illustrating the dynamic balance task allocation scheme and execution sequence method for embodiments of the present invention. In the second-level game, the device that obtains the execution right maximizes the overall inspection coverage by collaboratively sharing sensing data and path information, generating a dynamic balance task allocation scheme and execution sequence. This includes: each device that obtains the execution right broadcasts the collected sensing data and current path information; using the spatial coverage distribution characteristics of the sensing data and the trajectory extension directionality of the path information as the fusion basis; and using the density gradient of the covered area and the gap coordinates of the uncovered area as the state output to construct a global coverage density map. Using the global coverage density map as input, we extract the coverage gap area and low-density coverage area. Based on the spatial continuity range of each area, the perception importance weight and the current accessibility of the device that has obtained the execution right, we score the coverage urgency of each area and obtain a coverage completion demand sequence with priority tags. Using the complete demand sequence as the allocation source, the real-time location, path information and remaining energy status of each device that obtains the execution right as the three-element allocation constraint, and the difference between the perception adaptation gain of each device to each demand node and the movement cost to reach the node, an allocation benefit evaluation function is constructed to dynamically assign each demand node and generate a dynamically balanced task allocation scheme. Using the task allocation scheme as the timing orchestration input, the priority markers and movement costs of each device's assigned requirement node are extracted. The execution timing of each device is generated by using the priority markers to apply pre-excitation to the execution timing and the movement costs to apply delay suppression to the execution timing as bidirectional orchestration rules.

[0027] The aircraft and robot that obtain execution rights need to broadcast their collected perception data and real-time path information to the collaborative network via wireless communication modules. The perception data includes the device's geographic coordinates, coverage radius, and perception type identifier. The path information includes historical trajectory point sequences and movement direction vectors. After receiving the broadcast information, the central processing node normalizes and maps the perception data according to a spatial grid, dividing the power grid inspection area into square grid cells with a side length of 5 meters. The number of perception records in each grid cell is counted to form a coverage density numerical matrix. At the same time, the trajectory extension direction vector in the path information is extracted, and the azimuth angle change rate of adjacent timestamp trajectory points is calculated. This change rate is used as a dynamic weight and superimposed on the density value of the grid through which the trajectory passes to reflect the coverage trend of the device. For the covered area, the gradient difference between its density value and the density values ​​of the surrounding eight neighboring grids is calculated. Areas with gradient values ​​greater than the threshold of 2.5 are marked as high-density areas. For the uncovered area, connected grid clusters with a density value of zero are extracted, and their centroid coordinates and boundary coordinate sets are recorded to construct a global coverage density map that includes density distribution, gradient characteristics, and blank area annotations.

[0028] Based on the global coverage density map, blank areas with a density value of zero and low-density areas with a density value less than 30% of the average density are extracted. The spatial continuity range of each area is calculated. The spatial continuity range is obtained by multiplying the number of grid cells within the area by the area per unit grid cell. The importance weights of power grid equipment within each area are queried from the equipment importance level database and normalized to the range of 0 to 1. The Euclidean distance between the executing device and the centroid of each area is calculated. Combined with the device's current remaining energy state and the energy consumption coefficient per unit distance, it is determined whether the device can reach the area under the current energy constraints, yielding a Boolean value for reachability. The spatial continuity range, importance weights, and reachability Boolean value are input into the coverage urgency scoring formula, and the score is calculated. Where A is the area of ​​the region, W is the importance weight, R is the accessibility indicator variable, and D is the distance to the nearest device. To prevent small constants from being divided by zero, To adjust the coefficients, all regions are sorted in descending order of their score values, generating a cover completion requirement sequence with priority numbers.

[0029] For each demand node in the coverage completion demand sequence, the perception type demand label for the corresponding region is extracted. Based on the differences in perception characteristics between the aircraft and the robot, the perception adaptation gain of each device for that node is calculated. The adaptation gain for the aircraft is set to 0.8 for wide-area monitoring nodes and 0.3 for detail detection nodes; the adaptation gain for the robot is set to 0.4 for wide-area monitoring nodes and 0.9 for detail detection nodes. The centroid coordinates of the current position of each device and the demand node are extracted, and the Euclidean distance is calculated as the movement path length. This length is multiplied by the energy consumption coefficient per unit distance of the device to obtain the movement cost. An allocation benefit evaluation function is constructed, and the difference between the perception adaptation gain and the movement cost is calculated as the allocation benefit value. All combinations of devices and all demand nodes are traversed. Under the constraints that the remaining energy of the device is greater than the movement cost and there is no conflict in the display of device path information, a greedy strategy is used to establish a belonging relationship for the device-node pair with the highest benefit value in turn, until all demand nodes are allocated or all device energy is exhausted, generating a dynamically balanced task allocation scheme.

[0030] Based on the task allocation scheme, the set of required nodes assigned to each device, along with their priority markers and movement costs, are extracted. For multiple required nodes of a single device, a time-series orchestration queue is established. The reciprocal of the priority marker is calculated as a pre-activation factor; the higher the priority, the larger the activation factor. The ratio of movement cost to the device's maximum single-move energy consumption is calculated as a delay suppression factor; the higher the cost, the larger the suppression factor. A time-series score is calculated by combining these two factors. Required nodes with high scores are placed at the front of the execution queue, generating the execution time sequence for each device and guiding the aircraft and robot to perform inspection tasks sequentially.

[0031] Using the task allocation scheme as the timing orchestration input, the priority markers and movement costs of each device's assigned request node are extracted. The execution timing of each device is generated using bidirectional orchestration rules: priority markers apply pre-incentives to the execution timing, and movement costs apply delays to the execution timing. This includes: Using the task allocation scheme as the parsing input, the priority mark quantization value and the inter-node movement cost quantization value of each device's assigned demand node are extracted. The priority mark quantization value is the pre-excitation amplitude and the movement cost quantization value is the post-suppression amplitude. The sign of the difference between the pre-excitation amplitude and the post-suppression amplitude of each node determines its pre-excitation attribute or post-suppression attribute, generating the pre-excitation attribute node sequence and the post-suppression attribute node sequence of each device. Using the pre-excitation attribute node sequence and the post-suppression attribute node sequence of each device as the timing filling input, and using the dual-attribute filling rule of filling the pre-excitation attribute node sequence in descending order of pre-excitation magnitude to the front end of the execution timing and filling the post-suppression attribute node sequence in ascending order of post-suppression magnitude to the back end of the execution timing, candidate execution timings for each device are generated. The candidate execution sequence of each device is used as the input for amplitude continuity verification. The cumulative value of the pre-excitation amplitude of each node in the entire candidate execution sequence is not lower than the cumulative value of the delayed suppression amplitude as the amplitude continuity judgment criterion. If the verification fails, the bidirectional amplitude is recalculated for the nodes that do not meet the criterion and a new round of dual attribute filling is triggered. The cycle continues until the verification is passed. The candidate execution sequence that passes the verification is confirmed as the execution sequence of each device.

[0032] During the generation of the execution sequence, the attribute data of each node in the task allocation scheme needs to be quantitatively analyzed. For each assigned node, the aircraft and robot read the node's equipment importance level from the power grid topology as a priority marker quantification value. This value is typically represented as a normalized score between 0 and 1, with higher values ​​indicating greater urgency in the inspection task. Simultaneously, the product of the Euclidean distance from the current location to the node to be inspected and the expected energy consumption is calculated as the movement cost quantification value, reflecting the physical cost required for the equipment to reach the node. Based on the priority marker quantification value and the movement cost quantification value, the excitation characteristics of each node are calculated. The priority marker quantification value serves as the pre-excitation amplitude, reflecting the attractiveness of the task; the movement cost quantification value serves as the post-excitation suppression amplitude, reflecting the repulsive force of the task. By calculating the difference between the pre-excitation amplitude and the post-excitation suppression amplitude, the attribute characteristics of the task node are determined. When the difference is positive, the node exhibits pre-excitation attributes and should be executed first; when the difference is negative, the node exhibits post-excitation attributes and should be executed later.

[0033] For the UAV's task node T1, its priority marker quantization value (pre-excitation amplitude) is 85, and its movement cost quantization value (delayed suppression amplitude) from the initial position to T1 is 20, with a difference of 65, exhibiting a clear pre-excitation attribute. For the task node T5, its priority marker quantization value is 40, its movement cost quantization value is 50, and its difference is -10, exhibiting a delayed suppression attribute. In this way, the UAV's task nodes are divided into a pre-excitation attribute node sequence {T1, T3} and a delayed suppression attribute node sequence {T5}. Similarly, the robot's task nodes are divided into a pre-excitation attribute node sequence {T2} and a delayed suppression attribute node sequence {T4, T6}.

[0034] A dual-attribute filling rule is used to generate candidate execution time sequences. The sequence of preceding excitation attribute nodes is arranged in descending order of the preceding excitation amplitude, that is, the node with the larger excitation amplitude appears earlier. The sequence of delayed suppression attribute nodes is arranged in ascending order of the delayed suppression amplitude, that is, the node with the smaller suppression amplitude appears earlier. For UAVs, in the preceding excitation attribute node sequence {T1, T3}, the preceding excitation amplitude of T1 is 65, which is greater than the preceding excitation amplitude of T3, 45. Therefore, they are arranged in descending order as {T1, T3}. The delayed suppression attribute node sequence only contains T5, with a suppression amplitude of 10. According to the dual-attribute filling rule, the preceding excitation attribute node sequence is filled towards the front of the execution time sequence, and the delayed suppression attribute node sequence is filled towards the back of the execution time sequence, resulting in the candidate execution time sequence of UAVs as {T1, T3, T5}.

[0035] For the robot, its pre-excitation attribute node sequence contains only T2, with a pre-excitation amplitude of 32. In the delayed inhibition attribute node sequence {T4, T6}, the delayed inhibition amplitude of T4 is 8, which is less than the delayed inhibition amplitude of T6 (15), and they are arranged in ascending order as {T4, T6}. Applying the dual-attribute filling rule, the candidate execution sequence of the robot is obtained as {T2, T4, T6}. To ensure the energy feasibility and efficiency of the execution sequence, the amplitude continuity of the candidate execution sequence is verified. The verification criterion requires that the cumulative value of the pre-excitation amplitude of each node throughout the execution sequence is not lower than the cumulative value of the delayed inhibition amplitude, ensuring that the device has sufficient power to continue executing the task at any time without interruption due to insufficient energy or low efficiency.

[0036] The candidate execution sequence {T1, T3, T5} for the UAV was verified: When executing T1, the cumulative value of the pre-excitation amplitude was 65, and the cumulative value of the delayed suppression amplitude was 0, satisfying the verification criterion; when executing T3 after T1, the cumulative value of the pre-excitation amplitude was 65+45=110, and the cumulative value of the delayed suppression amplitude was 25 (the movement cost from T1 to T3), satisfying the verification criterion; when executing T5 after T3, the cumulative value of the pre-excitation amplitude was 110+0=110 (T5 does not contribute to the pre-excitation), and the cumulative value of the delayed suppression amplitude was 25+35=60 (the movement cost from T3 to T5 was increased), still satisfying the verification criterion. The candidate execution sequence for the UAV passed the verification and was confirmed as {T1, T3, T5}.

[0037] A similar verification was performed on the robot's candidate execution sequence {T2, T4, T6}: After executing T2, the cumulative value of the pre-excitation amplitude was 32, and the cumulative value of the delayed suppression amplitude was 0, which met the verification criteria; when executing T4 after T2, the cumulative value of the pre-excitation amplitude was still 32 (T4 does not contribute to the pre-excitation), and the cumulative value of the delayed suppression amplitude was 18 (the movement cost from T2 to T4), which met the verification criteria; however, when executing T6 after T4, the cumulative value of the pre-excitation amplitude was still 32, and the cumulative value of the delayed suppression amplitude increased to 18+30=48 (the movement cost from T4 to T6 increased), causing the cumulative value of the pre-excitation amplitude of 32 to be less than the cumulative value of the delayed suppression amplitude of 48, which did not meet the verification criteria.

[0038] If the verification fails, the bidirectional amplitudes of T4 and T6 are recalculated. To improve execution efficiency, the priority marker quantization value of T4 is adjusted (increased to 55), changing it from a delayed suppression attribute to a pre-excitation attribute. The new pre-excitation amplitude is 55-18=37. The attribute sequence is re-divided, and the robot's pre-excitation attribute node sequence becomes {T2, T4}, and the delayed suppression attribute node sequence becomes {T6}. The dual-attribute filling rule is applied to generate a new candidate execution sequence {T2, T4, T6}. The amplitude continuity of the new candidate execution sequence is verified again: the cumulative pre-excitation amplitude after executing T2 is 32, which meets the verification criterion; when executing T4 after T2, the cumulative pre-excitation amplitude is 32+37=69, and the cumulative delayed suppression amplitude is 18, which meets the verification criterion; when executing T6 after T4, the cumulative pre-excitation amplitude is 69+0=69, and the cumulative delayed suppression amplitude is 18+30=48, which still meets the verification criterion. The new candidate execution sequence was verified, confirming that the robot's final execution sequence is {T2, T4, T6}.

[0039] In practical power grid inspection applications, this method significantly improves inspection efficiency. In the inspection of a 110kV transmission line, the UAV completed the inspection according to the optimized execution sequence {T1, T3, T5}. This not only prioritized high-priority insulator inspection tasks but also rationally arranged subsequent paths, reducing round-trip flight distance and saving 25% of energy compared to traditional fixed-path inspection. Simultaneously, the robot performed ground inspection tasks according to the optimized path {T2, T4, T6}, maintaining a state of excitation greater than inhibition throughout the process. This avoided the risk of failing to complete the task due to energy depletion, thus improving the reliability and coverage of power grid inspection.

[0040] Based on the task allocation scheme, differentiated perception is performed. The aircraft acquires wide-area imagery and extracts equipment contour and layout information through target detection. The robot acquires close-range data and extracts surface texture and heat distribution information through fine-grained analysis. A complete equipment profile is synthesized through a feature cascade module, including: The aircraft performs wide-area image acquisition according to the mission allocation plan, applies target detection processing to the wide-area image, extracts contour geometric features and layout topology features with the equipment response area as the boundary, decomposes the contour geometric features into global contour features and local boundary features according to the perception scale, and aggregates them together with the layout topology features into a macroscopic structural feature set with a hierarchical structure. The robot performs close-range data acquisition according to the task allocation plan, applies fine-grained analysis to the acquired data, and extracts surface texture anisotropic features and thermal distribution thermal field gradient features using material response boundary and thermal field temperature gradient boundary as dual segmentation constraints. The anisotropic features are decomposed into coarse texture features and fine texture features according to texture frequency, and together with the thermal field gradient features, they are aggregated into a set of microscopic physical features with hierarchical structure. The feature cascade module takes macroscopic structural feature set and microscopic physical feature set as input, uses global contour feature quantity and thermal field gradient feature quantity to form macroscopic corresponding layer, and uses local boundary feature quantity and fine texture feature quantity to form microscopic corresponding layer. It calculates cross-modal feature complementarity layer by layer, and uses complementarity as weighting basis to drive cross-layer feature fusion to generate a complete device profile.

[0041] The aircraft is equipped with a binocular vision sensor to perform wide-area image acquisition in the airspace 15 to 30 meters away from the device. The acquisition resolution is set to 4096×2160 pixels, and the frame rate is maintained at 30fps to ensure complete recording of dynamic scenes. The acquired wide-area images are input into an improved YOLOv8 target detection network. The network backbone adopts the CSPDarknet53 structure to extract multi-scale features. Non-maximum suppression is used to select detection boxes with a confidence score of more than 0.85 as the device response region. Within the response region, the Canny operator combined with Hough transform is used to extract the device contour. The perimeter, area, aspect ratio, and roundness are calculated as contour geometric features. The contour features are decomposed into 5 scales using a Gaussian pyramid. The top layer corresponds to global contour features, including the overall shape descriptor. The bottom 3 layers correspond to local boundary features, including edge curvature and corner distribution. A topological relationship graph between devices is constructed simultaneously. The centroid of the device is used as the node, and the line connecting the centroids is used as the edge. The relative distance and azimuth are recorded as layout topological relationship features. The above features are organized according to a hierarchical structure of "device type - global shape - local details - spatial relationship" to form a macroscopic structural feature set with a dimension of 512.

[0042] The robot is equipped with an infrared thermal imager and a high-resolution industrial camera to perform close-range data acquisition within a range of 0.5 to 2 meters from the equipment. The thermal imager has a resolution of 640×512 pixels and a temperature sensitivity better than 0.05 degrees Celsius. The industrial camera has a resolution of 8192×5460 pixels. Gabor filter banks are applied to the visible light images for texture decomposition. The filter banks contain 40 kernels across 8 directions and 5 scales, extracting texture energy, contrast, and correlation to form anisotropic features. Fourier transform is used to map the texture features to the frequency domain. Components above a cutoff frequency of 0.2 cycles / pixel are classified as fine texture features, while those below are classified as coarse texture features. Pixel-level temperature gradients are calculated from the infrared images. The Sobel operator is used to extract gradient magnitudes in the horizontal and vertical directions, and regions with gradients exceeding 3 degrees Celsius / cm are marked as thermal field temperature gradient boundaries. Different material regions are segmented using color space clustering at the material response boundaries, and intersection operations with the thermal field boundaries are performed to determine double-constrained regions. The histogram of thermal gradient direction, peak temperature and temperature variance within the region are extracted as thermal gradient feature quantities. The texture and thermal features are organized into a hierarchical structure of "material type-coarse texture-fine texture-heat distribution" to form a microscopic physical feature set with a dimension of 256.

[0043] After receiving the macroscopic structural feature set and the microscopic physical feature set, the feature cascade module first normalizes the shape descriptor of the global contour feature quantity and the temperature distribution peak value of the thermal gradient feature quantity to a unified numerical range. The cosine of the angle between the two in the principal component space is calculated as the complementarity; a value greater than 0.6 indicates strong complementarity. In the macroscopic correspondence layer, the global contour feature and the thermal gradient feature are fused using a weighted complementarity calculation, with the fusion weights dynamically allocated using the softmax function. In the microscopic correspondence layer, the edge curvature sequence of the local boundary feature quantity and the high-frequency components of the fine texture feature quantity are aligned through convolution to establish a pixel-level correspondence. The feature correlation coefficient is calculated as the complementarity; a value greater than 0.55 triggers cross-layer feature fusion. Cross-layer fusion employs an attention mechanism, using the microscopic layer complementarity as the query vector and the macroscopic layer features as the key vector, calculating a weighted sum to obtain a 768-dimensional fused feature. The fused features are compressed to 384 dimensions through a three-layer fully connected network, outputting a complete device profile, including device category labels, contour parameters, material distribution, thermal anomaly locations, and confidence scores. The profile data is stored in JSON format and associated with a unique device code.

[0044] The difference between actual and expected revenue is calculated as the reward signal to update the utility function weight and collaboration strategy threshold. Furthermore, the task revenue calculation rules are adjusted inversely based on the accuracy of anomaly identification in the device profile, driving continuous optimization of the competitive and collaborative balance between the two platforms. This includes: The difference between the actual and expected benefits of the aircraft and the robot is used to synthesize a joint benefit deviation vector. The deviation distance between the current task allocation scheme and the Nash equilibrium solution set is calculated using the joint benefit deviation vector. The equilibrium deviation driving signal is constructed using the magnitude of the deviation distance. The equilibrium deviation driving signal amplitude drives the gradient update of the moving cost weight component and energy state weight component in the utility function towards the Nash equilibrium solution set. The deviation direction drives the reverse compensation adjustment of the path sharing trigger threshold in the cooperation strategy threshold, and outputs the updated utility function weight and cooperation strategy threshold. The identification deviation is constructed by the difference between the accuracy of the device image anomaly identification and the benchmark value. The Pareto improvement feasible region is constructed by the identification deviation. It is determined whether the wide-area image benefit coefficient and the near-field data benefit coefficient are located at the boundary of the Pareto feasible region. The two types of benefit coefficients are adjusted in the opposite direction and magnitude of deviation from the Pareto boundary to converge towards the Pareto optimum. The updated task benefit calculation rules are output. The termination condition is determined by the combined criteria of the Nash equilibrium convergence radius and the Pareto boundary distance, driving the dual platforms to continuously optimize until the effectiveness of feature complementarity is stably improved.

[0045] During task execution, the system continuously monitors the operating parameters of the aircraft and robots. For the i-th aircraft, its actual benefit is the value gained from performing the task minus the movement and communication costs. The expected benefit is determined based on the utility function value calculated during task allocation, and the difference between the two is denoted as the aircraft benefit deviation. Similarly, the benefit deviation of the j-th robot is calculated. The benefit deviations of all aircraft and robots are assembled into a joint benefit deviation vector, which reflects the degree of deviation between the actual execution effect and the game theory prediction under the current task allocation scheme.

[0046] The Nash equilibrium solution set is pre-stored. This solution set is obtained by iteratively solving the two-level game model and represents the strategy combination space when competition and cooperation reach a steady state. The Euclidean distance between the joint payoff deviation vector and the center point of the Nash equilibrium solution set is calculated. This distance value constitutes the equilibrium deviation driving signal. When the deviation distance exceeds the preset threshold, it indicates that the current strategy has become unbalanced and the parameter update mechanism needs to be triggered.

[0047] The amplitude of the driving signal directly affects the weight adjustment stage of the utility function. The utility function comprises three components: movement cost weight, energy state weight, and task benefit weight. Gradient descent is used to approximate the utility value of the Nash equilibrium solution. Specifically, the movement cost weight is updated by multiplying the driving signal amplitude by the learning rate, reducing the device's sensitivity to movement distance. The energy state weight is updated incrementally with the same amplitude, increasing the priority of energy constraints. The cooperation strategy threshold includes a path sharing trigger threshold and a data transmission delay tolerance. The deviation direction determines whether cooperation is excessive. If the aircraft's benefit deviation is negative and the robot's benefit deviation is positive, it indicates that the aircraft is relinquishing too many tasks. In this case, the path sharing trigger threshold is increased to reduce the frequency of ineffective cooperation; conversely, the threshold is decreased to promote information exchange.

[0048] Anomaly detection accuracy is calculated by comparing the device profile recognition results with manually labeled actual defects. The accuracy is subtracted from a preset benchmark value; a positive value indicates performance exceeding expectations, while a negative value indicates insufficient performance. When constructing the Pareto improvement feasible region, the wide-area image gain coefficient and the near-range data gain coefficient are used as two-dimensional coordinate axes, with the current coefficient combination as the coordinate point. The Pareto feasible region boundary is defined as the set of coefficient combinations where one gain coefficient cannot be further improved while maintaining the previous one. The vector from the current coordinate point to the nearest point on the boundary is calculated; this vector points in the Pareto optimal direction. The two types of gain coefficients are adjusted according to the sign of the recognition deviation: when the recognition accuracy is lower than the benchmark, if the error rate of the device contour detection handled by the aircraft is high, the wide-area image gain coefficient is reduced, and the robot's near-range data gain coefficient is increased, guiding the system to allocate more tasks to the robot for fine-tuning. The step size is adjusted proportionally to the absolute value of the recognition deviation, ensuring that the gain coefficient vector moves tangentially along the Pareto boundary.

[0049] The termination decision employs a two-condition logic: continuously monitoring the deviation distance between the joint benefit deviation vector and the Nash equilibrium solution set. When this distance remains within the convergence radius for several consecutive iterations, and the deviation of the identification accuracy from the benchmark value converges to a tolerable range, the two-layer optimization is considered to have reached a stable equilibrium point. At this point, the utility function weights, cooperation strategy thresholds, and task benefit calculation rules form a self-consistent parameter combination. The aircraft and robot maintain a reasonable allocation of resources in competition and achieve feature complementarity in cooperation, simultaneously improving overall inspection efficiency and anomaly detection quality to an acceptable level.

[0050] This invention also provides an intelligent power grid inspection system based on an AI-driven large-scale model, integrating drones and robots, including: The information acquisition unit is used to acquire the spatial distribution topology of the power grid, the importance level of equipment, and the initial position and energy state of aircraft and robots; The game modeling unit defines the aircraft as a global vision acquisition agent and the robot as a local depth exploration agent. It establishes a two-layer game model for multiple agents. In the first layer of the game, the aircraft and the robot construct utility functions based on equipment movement costs, energy states, and task benefits, and compete for the right to execute the target value task. In the second layer of the game, the device that wins the right to execute maximizes the overall inspection coverage by collaborating and sharing perception data and path information, and generates a dynamically balanced task allocation scheme and execution sequence. The perception synthesis unit is used to perform differentiated perception based on the task allocation scheme. The aircraft collects wide-area images and extracts the equipment outline and layout information through target detection. The robot collects close-range data and extracts surface texture and heat distribution information through fine-grained analysis. The complete equipment image is synthesized through the feature cascade module. The strategy optimization unit records the energy consumption rate, communication latency, and anomaly detection rate during task execution. It calculates the difference between actual and expected revenue as the reward signal to update the utility function weight and collaboration strategy threshold. It also adjusts the task revenue calculation rules in reverse based on the accuracy of anomaly identification in the device profile, driving the dual platforms to continuously optimize the competitive and collaborative balance.

[0051] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0052] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0053] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent power grid inspection using drones and robots linked by an AI large-scale model, characterized in that: include: Obtain the spatial distribution topology of the power grid, the importance level of equipment, and the initial position and energy state of aircraft and robots; The aircraft is defined as a global vision acquisition agent, and the robot is defined as a local depth exploration agent. A two-layer game model of multiple agents is established. In the first layer game, the aircraft and the robot construct utility functions through equipment movement costs, energy status and task benefits, and compete for the right to execute the target value task. In the second layer game, the equipment that wins the right to execute maximizes the overall inspection coverage by cooperating and sharing perception data and path information, and generates a dynamically balanced task allocation scheme and execution sequence. Based on the task allocation scheme, differentiated perception is performed. The aircraft collects wide-area images and extracts the equipment outline and layout information through target detection. The robot collects close-range data and extracts surface texture and heat distribution information through fine-grained analysis. The complete equipment image is synthesized through the feature cascade module. Record the energy consumption rate, communication latency, and anomaly detection rate during task execution. Calculate the difference between actual and expected revenue as the reward signal to update the utility function weight and collaboration strategy threshold. Adjust the task revenue calculation rules inversely based on the accuracy of anomaly identification in the device profile, driving continuous optimization of the competitive and collaborative balance between the two platforms.

2. The method according to claim 1, characterized in that, In the first level of the game, aircraft and robots construct utility functions based on equipment movement costs, energy states, and task benefits, competing for the right to execute a target-value task, including: Based on the initial position and current energy state of the aircraft and robot, the spatial distribution topology is dynamically divided into reachable domains, generating reachable task domains for the aircraft and robots. The task nodes in the two domains are sorted according to the importance level of the equipment, and the respective reachable target value task sequences are output. Taking the achievable target value task sequence as input, the equipment movement cost of each task node in the arrival sequence of the aircraft and robot is extracted respectively. An energy feasibility coefficient with energy state as a parameter is introduced to impose an upper limit constraint on the equipment movement cost, and task nodes that exceed the energy feasibility boundary are filtered out to generate a candidate task set filtered by energy. Based on the gain mapping relationship between the required perception depth of each task node in the candidate task set and the task benefits, the task benefit gain value of the aircraft and the robot for each task node in the candidate task set is calculated respectively. The difference between the task benefit gain value and the equipment movement cost is substituted into the utility function to calculate the competitive utility value. An exclusive competitive allocation is performed based on the competitive utility value. The party with the highest utility value obtains the execution right of the corresponding task node. In the case of the same competitive utility value, the energy state surplus is used as the arbitration basis to complete the unique ownership determination. The execution right allocation result is written back to the candidate task set to complete the state closure.

3. The method according to claim 1, characterized in that, In the second-level game, the device that gains execution rights maximizes overall inspection coverage by collaboratively sharing sensing data and path information, generating a dynamically balanced task allocation scheme and execution sequence, including: Each device that obtains execution rights will broadcast the collected sensing data and current path information. The spatial coverage distribution characteristics of the sensing data and the trajectory extension directionality of the path information will be used as the fusion basis, and the density gradient of the covered area and the gap coordinates of the uncovered area will be used as the status output to construct a global coverage density map. Using the global coverage density map as input, we extract the coverage gap area and low-density coverage area. Based on the spatial continuity range of each area, the perception importance weight and the current accessibility of the device that has obtained the execution right, we score the coverage urgency of each area and obtain a coverage completion demand sequence with priority tags. Using the complete demand sequence as the allocation source, the real-time location, path information and remaining energy status of each device that obtains the execution right as the three-element allocation constraint, and the difference between the perception adaptation gain of each device to each demand node and the movement cost to reach the node, an allocation benefit evaluation function is constructed to dynamically assign each demand node and generate a dynamically balanced task allocation scheme. Using the task allocation scheme as the timing orchestration input, the priority markers and movement costs of each device's assigned requirement node are extracted. The execution timing of each device is generated by using the priority markers to apply pre-excitation to the execution timing and the movement costs to apply delay suppression to the execution timing as bidirectional orchestration rules.

4. The method according to claim 3, characterized in that, Using the task allocation scheme as the timing orchestration input, the priority markers and movement costs of each device's assigned request node are extracted. The execution timing of each device is generated using bidirectional orchestration rules: priority markers apply pre-incentives to the execution timing, and movement costs apply delays to the execution timing. This includes: Using the task allocation scheme as the parsing input, the priority mark quantization value and the inter-node movement cost quantization value of each device's assigned demand node are extracted. The priority mark quantization value is the pre-excitation amplitude and the movement cost quantization value is the post-suppression amplitude. The sign of the difference between the pre-excitation amplitude and the post-suppression amplitude of each node determines its pre-excitation attribute or post-suppression attribute, generating the pre-excitation attribute node sequence and the post-suppression attribute node sequence of each device. Using the pre-excitation attribute node sequence and the post-suppression attribute node sequence of each device as the timing filling input, and using the dual-attribute filling rule of filling the pre-excitation attribute node sequence in descending order of pre-excitation magnitude to the front end of the execution timing and filling the post-suppression attribute node sequence in ascending order of post-suppression magnitude to the back end of the execution timing, candidate execution timings for each device are generated. The candidate execution sequence of each device is used as the input for amplitude continuity verification. The cumulative value of the pre-excitation amplitude of each node in the entire candidate execution sequence is not lower than the cumulative value of the delayed suppression amplitude as the amplitude continuity judgment criterion. If the verification fails, the bidirectional amplitude is recalculated for the nodes that do not meet the criterion and a new round of dual attribute filling is triggered. The cycle continues until the verification is passed. The candidate execution sequence that passes the verification is confirmed as the execution sequence of each device.

5. The method according to claim 1, characterized in that, Based on the task allocation scheme, differentiated perception is performed. The aircraft acquires wide-area imagery and extracts equipment contour and layout information through target detection. The robot acquires close-range data and extracts surface texture and heat distribution information through fine-grained analysis. A complete equipment profile is synthesized through a feature cascade module, including: The aircraft performs wide-area image acquisition according to the mission allocation plan, applies target detection processing to the wide-area image, extracts contour geometric features and layout topology features with the equipment response area as the boundary, decomposes the contour geometric features into global contour features and local boundary features according to the perception scale, and aggregates them together with the layout topology features into a macroscopic structural feature set with a hierarchical structure. The robot performs close-range data acquisition according to the task allocation plan, applies fine-grained analysis to the acquired data, and extracts surface texture anisotropic features and thermal distribution thermal field gradient features using material response boundary and thermal field temperature gradient boundary as dual segmentation constraints. The anisotropic features are decomposed into coarse texture features and fine texture features according to texture frequency, and together with the thermal field gradient features, they are aggregated into a set of microscopic physical features with hierarchical structure. The feature cascade module takes macroscopic structural feature set and microscopic physical feature set as input, uses global contour feature quantity and thermal field gradient feature quantity to form macroscopic corresponding layer, and uses local boundary feature quantity and fine texture feature quantity to form microscopic corresponding layer. It calculates cross-modal feature complementarity layer by layer, and uses complementarity as weighting basis to drive cross-layer feature fusion to generate a complete device profile.

6. The method according to claim 1, characterized in that, The difference between actual and expected revenue is calculated as the reward signal to update the utility function weight and collaboration strategy threshold. Furthermore, the task revenue calculation rules are adjusted inversely based on the accuracy of anomaly identification in the device profile, driving continuous optimization of the competitive and collaborative balance between the two platforms. This includes: The difference between the actual and expected benefits of the aircraft and the robot is used to synthesize a joint benefit deviation vector. The deviation distance between the current task allocation scheme and the Nash equilibrium solution set is calculated using the joint benefit deviation vector. The equilibrium deviation driving signal is constructed using the magnitude of the deviation distance. The equilibrium deviation driving signal amplitude drives the gradient update of the moving cost weight component and energy state weight component in the utility function towards the Nash equilibrium solution set. The deviation direction drives the reverse compensation adjustment of the path sharing trigger threshold in the cooperation strategy threshold, and outputs the updated utility function weight and cooperation strategy threshold. The identification deviation is constructed by the difference between the accuracy of the device image anomaly identification and the benchmark value. The Pareto improvement feasible region is constructed by the identification deviation. It is determined whether the wide-area image benefit coefficient and the near-field data benefit coefficient are located at the boundary of the Pareto feasible region. The two types of benefit coefficients are adjusted in the opposite direction and magnitude of deviation from the Pareto boundary to converge towards the Pareto optimum. The updated task benefit calculation rules are output. The termination condition is determined by the combined criteria of the Nash equilibrium convergence radius and the Pareto boundary distance, driving the dual platforms to continuously optimize until the effectiveness of feature complementarity is stably improved.

7. A smart power grid inspection system based on an AI large-scale model, integrating drones and robots, for implementing the method as described in any one of claims 1-6, characterized in that, include: The information acquisition unit is used to acquire the spatial distribution topology of the power grid, the importance level of equipment, and the initial position and energy state of aircraft and robots; The game modeling unit defines the aircraft as a global vision acquisition agent and the robot as a local depth exploration agent. It establishes a two-layer game model for multiple agents. In the first layer of the game, the aircraft and the robot construct utility functions based on equipment movement costs, energy states, and task benefits, and compete for the right to execute the target value task. In the second layer of the game, the device that wins the right to execute maximizes the overall inspection coverage by collaborating and sharing perception data and path information, and generates a dynamically balanced task allocation scheme and execution sequence. The perception synthesis unit is used to perform differentiated perception based on the task allocation scheme. The aircraft collects wide-area images and extracts the equipment outline and layout information through target detection. The robot collects close-range data and extracts surface texture and heat distribution information through fine-grained analysis. The complete equipment image is synthesized through the feature cascade module. The strategy optimization unit records the energy consumption rate, communication latency, and anomaly detection rate during task execution. It calculates the difference between actual and expected revenue as the reward signal to update the utility function weight and collaboration strategy threshold. It also adjusts the task revenue calculation rules in reverse based on the accuracy of anomaly identification in the device profile, driving the dual platforms to continuously optimize the competitive and collaborative balance.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images