A substation unmanned aerial vehicle inspection service flow dynamic derivation and optimization method

Abstract

The application discloses a substation unmanned aerial vehicle inspection service flow dynamic derivation and optimization method, which is driven by an inspection service flow dynamic derivation and optimization model. First, the input layer receives the multimodal input data of the substation and performs semantic unification and weighted fusion. Then, the processing layer performs trigger identification and selects multiple candidate variant service flows. The decision layer performs multi-objective quantitative evaluation and selects the optimal service flow. The execution layer converts the optimal service flow into unmanned aerial vehicle inspection actions and control instructions, collects multimodal inspection logs, and finally generates an inspection report and calculates a reward value. According to the reward value, the parameters of the inspection service flow dynamic derivation and optimization model are fine-tuned online. The application proposes a large model driven dynamic derivation and reinforcement learning closed loop optimization mechanism, which enables the inspection service flow to automatically derive a new version according to the multimodal trigger condition, and continuously evolves through actual flight logs, achieving higher coverage, lower missed detection rate and better energy consumption performance.

Description

Technical Field

[0001] This invention relates to the field of substation drone inspection technology, specifically a method for dynamic derivation and optimization of substation drone inspection workflow. Background Technology

[0002] As the lifeline of the national economy, the stable operation of the power system is directly related to social production and people's livelihood. Substations, as key nodes in power transmission and distribution, bear core functions such as voltage transformation, current distribution, and power dispatch. Their internal equipment, such as transformers, GIS switchgear, circuit breakers, disconnect switches, and instrument transformers, are exposed to the outdoor environment for extended periods, making them susceptible to factors such as wind and rain erosion, insulation aging, thermal stress accumulation, and foreign object adhesion. If equipment defects are not detected in time, they may escalate from localized faults to large-scale power outages, or even cause fires or explosions, resulting in huge economic losses and social impacts. Therefore, regular, accurate, and efficient substation equipment inspections have become a necessary measure to ensure the safe operation of the power grid.

[0003] In recent years, with the rapid development of drone technology, the application of drones in substation inspection has received widespread attention. Traditional substation inspection methods mainly include manual inspection from elevated positions and fixed camera monitoring. The former suffers from high workload, high safety risks, and low efficiency, while the latter is limited by blind spots, lighting conditions, and resolution, making it difficult to achieve comprehensive and detailed inspection. The introduction of drone inspection systems effectively alleviates these problems. Equipped with multiple sensors such as visible light, infrared, and ultraviolet, drones can achieve non-contact, multi-angle equipment status acquisition, significantly improving inspection efficiency and safety.

[0004] Existing methods for unmanned aerial vehicle (UAV) inspection of substations have the following drawbacks: 1. Manual Programming and Preset Path Control: Before a drone inspection mission, technicians manually draw the flight path, set the camera position and shooting parameters using ground station software, and the drone executes the mission along a fixed trajectory. While this method is stable, it requires manual replanning when there are changes in the substation equipment layout, new temporary obstacles, or seasonal defects, resulting in poor flexibility and high maintenance costs.

[0005] 2. Automatic generation based on traditional path planning algorithms: This method uses algorithms such as A*, RRT, Dijkstra's algorithm, or genetic algorithms combined with a digital twin model of the substation to generate flight routes. While this method achieves a certain degree of automation, it remains a static planning approach. If dynamic deviations occur during actual flight, such as wind disturbances, temporary bird nests, or equipment hotspot shifts, it cannot adaptively adjust, leading to increased missed detection rates or decreased flight efficiency.

[0006] 3. Visual Servo and Simple LLM-Assisted Control: Introducing machine vision or large-scale language models to parse natural language tasks can convert verbal instructions from maintenance personnel into initial business flows. However, existing solutions mostly remain at the one-time generation stage, lacking closed-loop feedback on execution logs and model iteration mechanisms. When the site environment or equipment status changes, the business flow quickly becomes invalid and needs to be regenerated, making it difficult to achieve continuous learning and continuous optimization.

[0007] In summary, existing drone inspection methods lack adaptability to the dynamic nature of actual substation operations, and their performance degrades significantly after long-term use. Furthermore, traditional methods are highly dependent on senior algorithm engineers, making them difficult for ordinary maintenance personnel to use directly, thus limiting the technology's widespread adoption. Summary of the Invention

[0008] The technical problem to be solved by this invention is to provide a dynamic derivation and optimization method for the unmanned aerial vehicle (UAV) inspection workflow in substations. It proposes a large model-driven dynamic derivation and reinforcement learning closed-loop optimization mechanism, which enables the UAV inspection workflow to automatically generate new versions based on multimodal triggering conditions and continuously evolve through actual flight logs. This significantly reduces the threshold for manual intervention and achieves higher coverage, lower missed detection rate and better energy consumption performance in complex substation environments.

[0009] The technical solution of this invention is as follows: A method for dynamic derivation and optimization of unmanned aerial vehicle (UAV) inspection workflow in substations is specifically driven by a dynamic derivation and optimization model of the inspection workflow, which includes an input layer, a processing layer, a decision layer, an execution layer, and an output layer. The method for dynamically deriving and optimizing the workflow of unmanned aerial vehicle (UAV) inspection of substations includes the following steps: (1) The input layer is used to receive multimodal input data from the substation and map it to a unified semantic space. Then, the semantically unified multimodal input data and the current business flow master are input to the processing layer. (2) The processing layer performs trigger identification and determines whether to activate the trigger for dynamic derivation optimization. When the trigger for dynamic derivation optimization is activated, the current business flow master and the trigger reason description are used as conditions. Combined with the power inspection-specific standardized atomic module library, multiple candidate variant business flows are generated through the Chain-of-Thought prompting framework. (3) The decision-making layer performs multi-objective quantitative evaluation on multiple candidate variant business flows, that is, calculates and sorts the candidate optimization score of each candidate variant business flow, and takes the candidate variant business flow with the highest candidate optimization score as the optimal business flow; (4) The execution layer converts the optimal business flow into UAV inspection actions and control commands, while monitoring the UAV inspection flight process in real time and collecting multimodal inspection logs. (5) The output layer receives the multimodal inspection logs collected by the execution layer, generates an inspection report, defines a reward function, calculates the reward value based on the multimodal inspection logs, and fine-tunes the parameters of the dynamic derivation optimization model of the inspection business flow online based on the reward value to achieve continuous iterative optimization of the inspection business flow.

[0010] The multimodal input data of the substation includes historical inspection log sets, equipment operation status matrices, digital twin update record sequences, meteorological environmental variable vectors, and real-time flight deviation signal matrices.

[0011] The multimodal input data are all mapped to a unified semantic space, so that each modal input data forms a corresponding high-dimensional semantic feature vector; Each high-dimensional semantic feature vector is weighted and integrated using the following equation (1) to generate a unified embedding vector. : (1); In equation (1), This represents the i-th feature in the high-dimensional semantic feature vector; The attention weight represents the i-th feature; This represents a learnable attention weight vector; Represents all features in a high-dimensional semantic feature vector; Unified embedding vectors corresponding to multimodal input data Composition to form a unified embedded vector database , Represents the total number of modalities in the multimodal input data; The unified embedding vector database The current business flow master is input into the processing layer.

[0012] Each atomic module in the substation inspection-specific atomic library represents an inspection task. Multiple atomic modules selected from the substation inspection-specific atomic library are combined in sequence to form a complete UAV inspection task. The current business flow master and candidate variant business flow are both ordered atomic module sequences formed by combining multiple atomic modules from the substation inspection-specific atomic library. The ordered atomic module sequences are described in structured JSON format.

[0013] The trigger recognition specifically refers to: S211. First, the semantic similarity score is calculated using the following formula (2). : (2); In equation (2), This represents the semantic similarity score of the j-th modality input data; The unified embedding vector representing the j-th modal input data; The historical baseline vector representing the j-th modal input data; Represents the magnitude of the vector; S212, Calculate the dynamic weighted trigger index , see the following formula (3) for details: (3); In equation (3), The dynamic weighting coefficients represent the j-th modal input data; S213. Set the trigger threshold; when the calculated dynamic weighted trigger index... When the threshold is exceeded, dynamic derivation optimization is triggered for activation.

[0014] When the dynamic derivation optimization is activated, based on the current business flow master and the trigger reason description, and combined with the power inspection-specific standardized atomic module library, multiple candidate variant business flows are generated through the Chain-of-Thought hint framework, specifically: S221. Perform semantic parsing on the trigger reason description, extract key change entities, and generate a trigger embedding vector. Then, it is embedded with the description vector of each atomic module in the standardized atomic module library specifically for power inspection. Matching and calculating the relevance score of each atomic module in the standardized atomic module library specifically for power line inspection. , see the following formula (4) for details: (4); In equation (4), This represents the sigmoid activation function, which maps the correlation scores to [0,1]. Select The atomic modules are added to the list of targets that need to be adjusted; S222. Based on the target list that needs adjustment, generate diverse paths through temperature sampling and sampling probability distribution. The calculation formula is shown in the following formula (5): (5); In equation (5), This represents a temperature parameter used to control diversity, with a value range of 0.7 to 1.5. Represents the utility function. An index representing any candidate variant of the service flow; , , Representing candidate variant business flows respectively The estimated scores in coverage, security, and efficiency dimensions are obtained through digital twin simulation calculations; , , Represent , , Dimensional weights, ; An index representing all candidate variant business flows; Multiple schemes are generated from temperature sampling, and each scheme corresponds to an adjustment combination of a specified atomic module; S223. Output a complete structured JSON format business flow description: For each solution, execute the forced inheritance logic: First, copy the JSON structure of the current business flow master, then overwrite the parameters of the atomic modules that need to be adjusted; finally, verify the security constraints and performance benchmarks of the inheritance. If the verification fails, adjust the parameters until they meet the requirements, and finally output the JSON format as a candidate variant business flow. The set of multiple candidate variant business flows is , The total number of candidate variant business flows generated. .

[0015] The decision layer performs multi-objective quantitative evaluation on multiple candidate variant service flows, that is, calculates and ranks the candidate optimization score of each candidate variant service flow, and selects the candidate variant service flow with the highest candidate optimization score as the optimal service flow, specifically: S31, The decision-making layer receives the trigger signal from the processing layer and the complete set of candidate service flows. The trigger signals include dynamically weighted trigger indicators. and activation mark, Represents the current business flow master; S32. Calculate the complete candidate service flow set using the following formula (6). Candidate optimization score for each candidate business flow : (6); In equation (6), The safety score is derived from a simulation of minimum obstacle avoidance distance and collision risk. The efficiency score is calculated based on the estimated flight time. The coverage score is calculated based on the coverage ratio of device locations. The energy consumption score is calculated based on the projected battery consumption of the drone. , , , These represent the weight coefficients for the four dimensions corresponding to the scores above. ; S33, Candidate optimization scores for all candidate business flows After normalizing to the [0,1] interval, candidate Pareto fronts are constructed, and then candidate optimization scores are selected from the candidate Pareto fronts. The highest-ranking candidate business flow is selected as the optimal business flow. .

[0016] After the optimal service flow is determined, the historical service flow versions are managed through a differential storage mechanism. The historical service flow versions are all the optimal service flows determined at the current time and at historical times. After the optimal service flow at the current time is determined, simulation verification is performed. When the performance of the UAV inspection drops beyond the threshold during the simulation verification, it is automatically rolled back to the previous service flow, i.e., the current service flow master.

[0017] The execution layer converts the optimal business flow into UAV inspection actions and control commands, while simultaneously monitoring the UAV inspection flight process in real time and collecting multimodal inspection logs, specifically: S41. Perform business flow parsing: Parse the optimal business flow in JSON format. The process is decomposed into an ordered sequence of atomic modules, and a list of atomic modules is extracted. Each atomic module contains an ID, configuration parameters, and dependencies. The integrity of the ordered sequence of atomic modules is verified by depth-first traversal to ensure that there are no circular dependencies, and an intermediate action chain is generated. S42. Initialize the UAV's state vector based on the action chain. , Represents the time step and the state vector. Used to represent the position, velocity, and attitude of the UAV; predicts the future optimized trajectory, as shown in the following formula (7): (7); In equation (7), Represents the state transition matrix. Represents the control command input matrix. Represents control commands; Optimize the trajectory to minimize the deviation cost function and generate a smooth path point sequence; S43, Control Signal Conversion: Maps the smooth path point sequence to low-level control signals of the UAV. Each path point is converted into a control command, which is either a PWM signal or a CAN bus command. Sensor scheduling is integrated. S44. During the execution of control commands, continuously compare the actual trajectory with the optimized trajectory, calculate the deviation index, and if an abnormal deviation index is detected, apply feedback control to adjust the control commands. And record the event; S45. Continuously collect multimodal inspection logs to quantify performance.

[0018] The output layer receives multimodal inspection logs collected by the execution layer, generates an inspection report, defines a reward function, calculates a reward value based on the multimodal inspection logs, and fine-tunes the parameters of the dynamic derivation optimization model of the inspection business flow online based on the reward value. Specifically: S51. The output layer receives the multimodal inspection logs collected by the execution layer and generates an inspection report. The inspection report includes defect detection results, quantitative performance indicators, and a business flow version evolution tree diagram. The business flow version evolution tree diagram visualizes the change path from the current business flow master to the best business flow. When a performance deviation is detected, an alarm is issued and a version rollback is triggered. S52. Define the reward function and calculate the reward value based on the multimodal inspection log. The specific reward function is shown in the following formula (8): (8); In equation (8), Represents the reward value; This represents the equipment coverage rate, which is the actual number of covered devices divided by the total number of devices that should be covered. It represents the reciprocal of the missed detection rate, which is one minus the proportion of defects missed as determined by manual review or cross-validation; It represents the reciprocal of the standardized flight time, i.e., the shortest historical flight time / the actual flight time; This represents the reciprocal of the collision risk, which is one minus the normalized value of the maximum collision risk for this flight. It represents the reciprocal of energy efficiency, which is the theoretical minimum energy consumption / actual energy consumption; , , , , These represent the weight coefficients for the corresponding dimensions of the five parameters mentioned above. ; S53. Fine-tune the parameters of the dynamic derivation optimization model of the inspection business flow online based on the reward value; S54. Feed back the structured feedback data to the input layer to support the triggering identification and derivation optimization process of the next UAV inspection. The structured feedback data includes the parameters of the fine-tuned inspection business flow dynamic derivation optimization model, multimodal inspection log summary, and business flow version record.

[0019] Advantages of this invention: (1) This invention continuously collects multimodal data, including historical inspection logs, equipment operating status, digital twin update records, meteorological environmental variables and real-time flight deviation signals. The above multimodal data is mapped to a unified semantic space through the input layer to form a high-dimensional semantic feature vector, and then fused to form a unified embedding vector. The unified embedding vector is compared and learned with the historical benchmark vector to calculate a dynamic weighted trigger index for quantifying the significance of environmental or equipment changes. When the dynamic weighted trigger index exceeds the set trigger threshold, the business flow derivative process is activated. This trigger identification mechanism ensures that the derivative is only triggered when there is a substantial change in actual operation and maintenance needs (such as equipment capacity expansion, the arrival of the icing season or the high incidence of historical hot spots), thus avoiding invalid calculation overhead.

[0020] (2) After activation, the processing layer of this invention uses the current business flow master as a condition and combines the dedicated atomic library for power inspection to generate multiple candidate variant business flows through a three-stage Chain-of-Thought prompting process. This ensures that each generated candidate variant business flow is superior to the master in terms of coverage, security and efficiency, and forcibly inherits the security constraints and performance benchmarks of the master.

[0021] (3) The decision-making layer of this invention balances multiple objectives such as security, efficiency, coverage and energy consumption according to the Pareto optimality principle, thereby selecting the best business flow, and manages historical versions through differential storage mechanism. Even if deviations occur in complex substation environments, this mechanism can roll back or adjust in time to ensure the continuity and reliability of the inspection process. The problem of computing resource limitation is solved through efficient version management.

[0022] (4) This invention constructs a multi-dimensional reward function (covering equipment coverage, defect miss rate, flight efficiency, collision risk, and energy consumption indicators) that takes the actual collected multimodal inspection logs as direct input, and uses online Low-Rank Adaptation (LoRA) technology to dynamically derive and optimize the relevant parameters of the inspection business flow model in real time. This closed-loop mechanism automatically completes reward evaluation and parameter updates after each real inspection task, realizing the continuous evolution and performance improvement of the model in zero-sample real scenarios, ensuring that the business flow strategy adapts to the dynamic changes of complex substation environments in the long term, and continuously optimizing coverage integrity and defect detection accuracy. Attached Figure Description

[0023] Figure 1 This is a flowchart of the present invention.

[0024] Figure 2 This is a structural diagram of the dynamic derivation optimization model for the inspection workflow of this invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] A method for dynamic derivation and optimization of unmanned aerial vehicle (UAV) inspection workflow in substations is specifically driven by a dynamic derivation and optimization model of the inspection workflow, which includes an input layer 1, a processing layer 2, a decision layer 3, an execution layer 4, and an output layer 5. The method for dynamically deriving and optimizing the workflow of unmanned aerial vehicle (UAV) inspection of substations includes the following steps: (1) Input layer 1 is used to receive multimodal input data from the substation and perform semantic unification and weighted fusion, specifically: S11. The multimodal input data of the substation includes historical inspection log sets, equipment operation status matrix, digital twin update record sequence, meteorological environmental variable vector and real-time flight deviation signal matrix; The historical inspection log set is used to reflect the defect distribution and execution performance of historical inspection tasks; the equipment operation status matrix includes the number of equipment and its status dimensions, such as normalized indicators like load rate and temperature; the digital twin update record sequence is a JSON-formatted description of layout change events; the meteorological environmental variable vector is the collected meteorological environmental variables used to normalize external influencing factors; the real-time flight deviation signal matrix is ​​constructed from sensor readings to quantify trajectory deviation and risk; S12. All multimodal input data are mapped to a unified semantic space, so that each type of input data forms a corresponding high-dimensional semantic feature vector; S13. Each high-dimensional semantic feature vector is weighted and integrated using the following formula (1) to generate a unified embedding vector. : (1); In equation (1), This represents the i-th feature in the high-dimensional semantic feature vector; The attention weight represents the i-th feature; This represents a learnable attention weight vector; Represents all features in a high-dimensional semantic feature vector; S14. Unified embedding vectors corresponding to multimodal input data Composition to form a unified embedded vector database , Represents the total number of modalities in the multimodal input data; Unified Embedded Vector Database The current business flow master is input into the processing layer; Each atomic module in the substation inspection-specific atomic library represents an inspection task. Multiple atomic modules selected from the substation inspection-specific atomic library are combined in sequence to form a complete UAV inspection task. The current business flow master and the selected candidate variant business flow are both ordered atomic module sequences formed by combining multiple atomic modules from the substation inspection-specific atomic library. The ordered atomic module sequence is described in structured JSON format, including take-off and landing sequences, standard equipment surround shooting positions, multi-modal sensor acquisition strategies, and defect post-processing logic. (2) The processing layer 2 first performs trigger identification and generates multiple candidate variant service flows, specifically: S21, Trigger Recognition: S211. First, the semantic similarity score is calculated using the following formula (2). : (2); In equation (2), This represents the semantic similarity score of the j-th modality input data; The unified embedding vector representing the j-th modal input data; The historical baseline vector representing the j-th modality input data is derived from a unified embedding vector database. The normal status input is retrieved and pre-stored in the database; Represents the magnitude of the vector; S212, Calculate the dynamic weighted trigger index , see the following formula (3) for details: (3); In equation (3), The dynamic weighting coefficients representing the j-th modal input data are adjusted according to the station type through domain fine-tuning or online learning, prioritizing high-risk modalities such as real-time deviations. S213, Set trigger threshold When the dynamically weighted trigger indicator is calculated Greater than the trigger threshold At that time, dynamic derivation optimization is triggered and activated; S22. After activation is triggered, candidate variant business flows are generated: S221. Perform semantic parsing on the trigger cause description, extract key change entities (such as "new equipment coordinates" or "seasonal icing risk"), and generate a trigger embedding vector. (Obtained from the descriptive text in the input layer), and then compared with the description embedding vector of each atomic module in the standardized atomic module library specifically for power inspection. Matching and calculating the relevance score of each atomic module in the standardized atomic module library specifically for power line inspection. , see the following formula (4) for details: (4); In equation (4), This represents the sigmoid activation function, which maps the correlation scores to [0,1]. subscript This represents the index of atomic modules in the standardized atomic module library specifically for power line inspection. Select The atomic modules are added to the list of targets that need to be adjusted; S222. Based on the target list that needs adjustment, generate diverse paths through temperature sampling and sampling probability distribution. The calculation formula is shown in the following formula (5): (5); In equation (5), This represents a temperature parameter used to control diversity, with a value range of 0.7 to 1.5. Represents the utility function. An index representing any candidate variant of the service flow; , , Representing candidate variant business flows respectively The estimated scores in coverage, security, and efficiency dimensions are obtained through digital twin simulation calculations; , , Represent , , Dimensional weights, ; An index representing all candidate variant business flows; Multiple schemes are generated from temperature sampling, and each scheme corresponds to an adjustment combination of a specified atomic module; S223. Output a complete structured JSON format business flow description: For each solution, execute the forced inheritance logic: First, copy the JSON structure of the current business flow master, then overwrite the parameters of the atomic modules that need to be adjusted; finally, verify the security constraints and performance benchmarks of the inheritance. If the verification fails, adjust the parameters until they meet the requirements, and finally output the JSON format as a candidate variant business flow. The final output JSON format is: {"sequence": [list of atomic modules, including ID and configuration], "constraints": [inherited security parameters, such as {"avoidance_distance": 5.0}, {"energy_limit":80%}], "benchmarks": [performance thresholds, such as {"coverage_min": 95%}, {"efficiency_min":90%}]}; the above JSON format is stored after syntax validation. The set of multiple candidate variant business flows is , The total number of candidate variant business flows generated. ; (3) The decision-making level 3 conducts multi-objective quantitative evaluation and selects the optimal business flow, specifically: S31, The decision-making layer receives the trigger signal from the processing layer and the complete set of candidate service flows. The trigger signals include dynamically weighted trigger indicators. and activation mark, Represents the current business flow master; S32. Calculate the complete candidate service flow set using the following formula (6). Candidate optimization score for each candidate business flow : (6); In equation (6), The safety score is derived from a simulation of minimum obstacle avoidance distance and collision risk. The efficiency score is calculated based on the estimated flight time. The coverage score is calculated based on the coverage ratio of device locations. The energy consumption score is calculated based on the projected battery consumption of the drone. , , , These represent the weight coefficients for the four dimensions corresponding to the scores above. ; S33, Candidate optimization scores for all candidate business flows After normalizing to the [0,1] interval, candidate Pareto fronts are constructed, and then candidate optimization scores are selected from the candidate Pareto fronts. The highest-ranking candidate business flow is selected as the optimal business flow. ; S34. After the optimal service flow is determined, the historical service flow version is managed through the differential storage mechanism. The historical service flow version is all the optimal service flows determined at the current time and the historical time. After the optimal service flow at the current time is determined, simulation verification is performed. When the performance of the drone inspection in the simulation verification decreases beyond the threshold (e.g., the coverage rate decreases by more than 5%), it automatically rolls back to the previous service flow, i.e. the current service flow master version. (4) Execution layer 4 converts the optimal business flow into UAV inspection actions and control commands, while simultaneously monitoring the UAV inspection flight process in real time and collecting multimodal inspection logs, specifically: S41. Perform business flow parsing: Parse the optimal business flow in JSON format. The sequence is decomposed into an ordered atomic module sequence. The list of atomic modules in the "sequence" field is extracted. Each atomic module contains an ID, configuration parameters (such as surround angle and sensor mode) and dependencies. The integrity of the ordered atomic module sequence is verified by depth-first traversal to ensure no circular dependencies, and an intermediate action chain is generated (such as take-off → navigation to device location → hovering data acquisition → return). S42. Initialize the UAV's state vector based on the action chain. , Represents the time step and the state vector. Used to represent the position, velocity, and attitude of the UAV; predicts the future optimized trajectory, as shown in the following formula (7): (7); In equation (7), Represents the state transition matrix. Represents the control command input matrix. Represents control commands; The trajectory is optimized to minimize the deviation cost function (in conjunction with safety constraints, such as obstacle avoidance distance), generating a smooth path point sequence; S43, Control Signal Conversion: Maps the smooth path point sequence to low-level control signals for the UAV. Each path point is converted into a control command, which is either a PWM signal or a CAN bus command (such as motor speed or servo angle). Sensor scheduling is integrated (such as activating an infrared camera at a specified point). The signal format conforms to standard protocols (such as MAVLink) to ensure compatibility with commercial UAV hardware. S44. During the execution of control commands, continuously compare the actual trajectory with the optimized trajectory and calculate deviation indicators (such as triggering correction when the position error exceeds 0.5 meters). If abnormal deviation indicators are detected (such as wind disturbance causing deviation), apply feedback control to adjust the control commands. And record the event; S45. Continuously collect multimodal inspection logs to quantify performance. The multimodal inspection logs are a structured data set, including deviations between actual flight trajectory and planning, distribution of defect identification confidence, battery energy consumption curves, collision risk peaks, and a list of missed equipment. (5) Output layer 5 receives the multimodal inspection logs collected by the execution layer, generates an inspection report, defines a reward function, calculates the reward value based on the multimodal inspection logs, and fine-tunes the parameters of the dynamic derivation optimization model of the inspection business flow online based on the reward value to achieve continuous iterative optimization of the inspection business flow, specifically: S51. The output layer receives multimodal inspection logs collected by the execution layer and generates an inspection report. The inspection report includes defect detection results, quantitative performance indicators, and a business flow version evolution tree diagram. The business flow version evolution tree diagram visualizes the change path from the current business flow master to the best business flow. When a performance deviation is detected (such as a missed detection rate exceeding 5% or energy consumption exceeding the threshold), an alarm is issued and a version rollback is triggered. S52. Define the reward function and calculate the reward value based on the multimodal inspection log (used to quantify the performance of business flow execution). The specific reward function is shown in the following formula (8): (8); In equation (8), Represents the reward value; This represents the equipment coverage rate, which is the actual number of covered devices divided by the total number of devices that should be covered. It represents the reciprocal of the missed detection rate, which is one minus the proportion of defects missed as determined by manual review or cross-validation; It represents the reciprocal of the standardized flight time, i.e., the shortest historical flight time / the actual flight time; This represents the reciprocal of the collision risk, which is one minus the normalized value of the maximum collision risk for this flight. It represents the reciprocal of energy efficiency, which is the theoretical minimum energy consumption / actual energy consumption; , , , , These represent the weight coefficients for the corresponding dimensions of the five parameters above, which are dynamically adjusted through AHP or online learning. ; S53. Fine-tune the parameters of the dynamic derivation optimization model of the inspection business flow online based on the reward value; S54. Feed back the structured feedback data to the input layer to support the triggering identification and derivation optimization process of the next UAV inspection. The structured feedback data includes the parameters of the fine-tuned inspection business flow dynamic derivation optimization model, multimodal inspection log summary, and business flow version record.

[0027] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for dynamic derivation and optimization of unmanned aerial vehicle (UAV) inspection workflow in substations, characterized in that: Specifically, it is driven by a dynamic derivation optimization model of the inspection business flow, which includes an input layer, a processing layer, a decision layer, an execution layer, and an output layer. The method for dynamically deriving and optimizing the workflow of unmanned aerial vehicle (UAV) inspection of substations includes the following steps: (1) The input layer is used to receive multimodal input data from the substation and map it to a unified semantic space. Then, the semantically unified multimodal input data and the current business flow master are input to the processing layer. (2) The processing layer performs trigger identification and determines whether to activate the trigger for dynamic derivation optimization. When the trigger for dynamic derivation optimization is activated, the current business flow master and the trigger reason description are used as conditions. Combined with the power inspection-specific standardized atomic module library, multiple candidate variant business flows are generated through the Chain-of-Thought prompting framework. (3) The decision-making layer performs multi-objective quantitative evaluation on multiple candidate variant business flows, that is, calculates and sorts the candidate optimization score of each candidate variant business flow, and takes the candidate variant business flow with the highest candidate optimization score as the optimal business flow; (4) The execution layer converts the optimal business flow into UAV inspection actions and control commands, while monitoring the UAV inspection flight process in real time and collecting multimodal inspection logs. (5) The output layer receives the multimodal inspection logs collected by the execution layer, generates an inspection report, defines a reward function, calculates the reward value based on the multimodal inspection logs, and fine-tunes the parameters of the dynamic derivation optimization model of the inspection business flow online based on the reward value to achieve continuous iterative optimization of the inspection business flow.

2. The method for dynamic derivation and optimization of unmanned aerial vehicle (UAV) inspection workflow in substations according to claim 1, characterized in that: The multimodal input data of the substation includes historical inspection log sets, equipment operation status matrices, digital twin update record sequences, meteorological environmental variable vectors, and real-time flight deviation signal matrices.

3. The method for dynamic derivation and optimization of substation drone inspection workflow according to claim 2, characterized in that: The multimodal input data are all mapped to a unified semantic space, so that each modal input data forms a corresponding high-dimensional semantic feature vector; Each high-dimensional semantic feature vector is weighted and integrated using the following equation (1) to generate a unified embedding vector. : （1）； In equation (1), This represents the i-th feature in the high-dimensional semantic feature vector; The attention weight represents the i-th feature; This represents a learnable attention weight vector; Represents all features in a high-dimensional semantic feature vector; Unified embedding vectors corresponding to multimodal input data Composition to form a unified embedded vector database , Represents the total number of modalities in the multimodal input data; The unified embedding vector database The current business flow master is input into the processing layer.

4. The method for dynamic derivation and optimization of unmanned aerial vehicle (UAV) inspection workflow in substations according to claim 1, characterized in that: Each atomic module in the dedicated atomic library for substation inspection represents an inspection task. Multiple atomic modules selected from the dedicated atomic library for substation inspection are combined in sequence to form a complete UAV inspection task. The current business flow master and candidate variant business flow are both ordered atomic module sequences formed by combining multiple atomic modules from the substation inspection dedicated atomic library. The ordered atomic module sequences are described in structured JSON format.

5. The method for dynamic derivation and optimization of unmanned aerial vehicle (UAV) inspection workflow in substations according to claim 1, characterized in that: The trigger recognition specifically refers to: S211. First, the semantic similarity score is calculated using the following formula (2). : （2）； In equation (2), This represents the semantic similarity score of the j-th modality input data; The unified embedding vector representing the j-th modal input data; The historical baseline vector representing the j-th modal input data; Represents the magnitude of the vector; S212, Calculate the dynamic weighted trigger index , see the following formula (3) for details: （3）； In equation (3), The dynamic weighting coefficients represent the j-th modal input data; S213. Set the trigger threshold; when the calculated dynamic weighted trigger index... When the threshold is exceeded, dynamic derivation optimization is triggered for activation.

6. The method for dynamic derivation and optimization of substation drone inspection workflow according to claim 5, characterized in that: When the dynamic derivation optimization is activated, based on the current business flow master and the trigger reason description, and combined with the power inspection-specific standardized atomic module library, multiple candidate variant business flows are generated through the Chain-of-Thought hint framework, specifically: S221. Perform semantic parsing on the trigger reason description, extract key change entities, and generate a trigger embedding vector. Then, it is embedded with the description vector of each atomic module in the standardized atomic module library specifically for power inspection. Matching and calculating the relevance score of each atomic module in the standardized atomic module library specifically for power line inspection. , see the following formula (4) for details: （4）； In equation (4), This represents the sigmoid activation function, which maps the correlation scores to [0,1]. Select The atomic modules are added to the list of targets that need to be adjusted; S222. Based on the target list that needs adjustment, generate diverse paths through temperature sampling and sampling probability distribution. The calculation formula is shown in the following formula (5): （5）； In equation (5), This represents a temperature parameter used to control diversity, with a value range of 0.7 to 1.

5. Represents the utility function. An index representing any candidate variant of the service flow; , , Representing candidate variant business flows respectively The estimated scores in coverage, security, and efficiency dimensions are obtained through digital twin simulation calculations; , , Represent , , Dimensional weights, ; An index representing all candidate variant business flows; Multiple schemes are generated from temperature sampling, and each scheme corresponds to an adjustment combination of a specified atomic module; S223. Output a complete structured JSON format business flow description: For each solution, execute the forced inheritance logic: First, copy the JSON structure of the current business flow master, then overwrite the parameters of the atomic modules that need to be adjusted; finally, verify the security constraints and performance benchmarks of the inheritance. If the verification fails, adjust the parameters until they meet the requirements, and finally output the JSON format as a candidate variant business flow. The set of multiple candidate variant business flows is , The total number of candidate variant business flows generated. .

7. The method for dynamic derivation and optimization of unmanned aerial vehicle (UAV) inspection workflow in substations according to claim 6, characterized in that: The decision layer performs multi-objective quantitative evaluation on multiple candidate variant service flows, that is, calculates and ranks the candidate optimization score of each candidate variant service flow, and selects the candidate variant service flow with the highest candidate optimization score as the optimal service flow, specifically: S31, The decision-making layer receives the trigger signal from the processing layer and the complete set of candidate service flows. The trigger signals include dynamically weighted trigger indicators. and activation mark, Represents the current business flow master; S32. Calculate the complete candidate service flow set using the following formula (6). Candidate optimization score for each candidate business flow : （6）； In equation (6), The safety score is derived from a simulation of minimum obstacle avoidance distance and collision risk. The efficiency score is calculated based on the estimated flight time. The coverage score is calculated based on the coverage ratio of device locations. The energy consumption score is calculated based on the projected battery consumption of the drone. , , , These represent the weight coefficients for the four dimensions corresponding to the scores above. ; S33, Candidate optimization scores for all candidate business flows After normalizing to the [0,1] interval, candidate Pareto fronts are constructed, and then candidate optimization scores are selected from the candidate Pareto fronts. The highest-ranking candidate business flow is selected as the optimal business flow. .

8. The method for dynamic derivation and optimization of unmanned aerial vehicle (UAV) inspection workflow in substations according to claim 1, characterized in that: After the optimal service flow is determined, the historical service flow versions are managed through a differential storage mechanism. The historical service flow versions are all the optimal service flows determined at the current time and at historical times. After the optimal service flow at the current time is determined, simulation verification is performed. When the performance of the UAV inspection drops beyond the threshold during the simulation verification, it is automatically rolled back to the previous service flow, i.e., the current service flow master.

9. The method for dynamic derivation and optimization of unmanned aerial vehicle (UAV) inspection workflow in substations according to claim 7, characterized in that: The execution layer converts the optimal business flow into UAV inspection actions and control commands, while simultaneously monitoring the UAV inspection flight process in real time and collecting multimodal inspection logs, specifically: S41. Perform business flow parsing: Parse the optimal business flow in JSON format. The process is decomposed into an ordered sequence of atomic modules, and a list of atomic modules is extracted. Each atomic module contains an ID, configuration parameters, and dependencies. The integrity of the ordered sequence of atomic modules is verified by depth-first traversal to ensure that there are no circular dependencies, and an intermediate action chain is generated. S42. Initialize the UAV's state vector based on the action chain. , Represents the time step and the state vector. Used to represent the position, velocity, and attitude of the UAV; predicts the future optimized trajectory, as shown in the following formula (7): （7）； In equation (7), Represents the state transition matrix. Represents the control command input matrix. Represents control commands; Optimize the trajectory to minimize the deviation cost function and generate a smooth path point sequence; S43, Control Signal Conversion: Maps the smooth path point sequence to low-level control signals of the UAV. Each path point is converted into a control command, which is either a PWM signal or a CAN bus command. Sensor scheduling is integrated. S44. During the execution of control commands, continuously compare the actual trajectory with the optimized trajectory, calculate the deviation index, and if an abnormal deviation index is detected, apply feedback control to adjust the control commands. And record the event; S45. Continuously collect multimodal inspection logs to quantify performance.

10. The method for dynamic derivation and optimization of unmanned aerial vehicle (UAV) inspection workflow in substations according to claim 1, characterized in that: The output layer receives multimodal inspection logs collected by the execution layer, generates an inspection report, defines a reward function, calculates a reward value based on the multimodal inspection logs, and fine-tunes the parameters of the dynamic derivation optimization model of the inspection business flow online based on the reward value. Specifically: S51. The output layer receives the multimodal inspection logs collected by the execution layer and generates an inspection report. The inspection report includes defect detection results, quantitative performance indicators, and a business flow version evolution tree diagram. The business flow version evolution tree diagram visualizes the change path from the current business flow master to the best business flow. When a performance deviation is detected, an alarm is issued and a version rollback is triggered. S52. Define the reward function and calculate the reward value based on the multimodal inspection log. The specific reward function is shown in the following formula (8): （8）； In equation (8), Represents the reward value; This represents the equipment coverage rate, which is the actual number of covered devices divided by the total number of devices that should be covered. It represents the reciprocal of the missed detection rate, which is one minus the proportion of defects missed as determined by manual review or cross-validation; It represents the reciprocal of the standardized flight time, i.e., the shortest historical flight time / the actual flight time; This represents the reciprocal of the collision risk, which is one minus the normalized value of the maximum collision risk for this flight. It represents the reciprocal of energy efficiency, which is the theoretical minimum energy consumption / actual energy consumption; , , , , These represent the weight coefficients for the corresponding dimensions of the five parameters mentioned above. ; S53. Fine-tune the parameters of the dynamic derivation optimization model of the inspection business flow online based on the reward value; S54. Feed back the structured feedback data to the input layer to support the triggering identification and derivation optimization process of the next UAV inspection. The structured feedback data includes the parameters of the fine-tuned inspection business flow dynamic derivation optimization model, multimodal inspection log summary, and business flow version record.

Images