An unmanned aerial vehicle-based substation equipment inspection path planning system

Abstract

The application relates to the field of intelligent inspection of substation equipment, and discloses a substation equipment inspection path planning system based on a unmanned aerial vehicle, which aims to solve problems such as difficulty in real-time return of high-definition images and single path planning under strong electromagnetic interference of a substation. The system contains five core modules, a three-dimensional digital reference and a feature library are constructed, new equipment is adapted by combining operator similarity migration, and an observation pose is accurately reproduced to guarantee image comparison accuracy; data is selectively returned after onboard edge analysis and is adapted to a low-frequency transmission link; and a globally optimal path is planned by combining path length and flight risk according to equipment similarity in batches. The application realizes efficient transmission of inspection data, accurate and reliable detection, balances inspection efficiency and flight safety, and meets the core requirements of safety, efficiency and accuracy of substation inspection.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent inspection of substation equipment, in particular to a substation equipment inspection path planning system based on a UAV. BACKGROUND

[0002] As the core hub of the power system, the operation state of the substation equipment is directly related to the safe and stable power supply of the power grid, and the frequency and accuracy of the inspection are extremely high. Traditional manual inspection is labor-intensive and inefficient, and faces safety hazards such as strong electromagnetic radiation and high-altitude operation in the substation, making it difficult to achieve full-coverage and accurate inspection.

[0003] The UAV inspection technology gradually replaces manual inspection due to its flexibility and wide coverage. However, the existing technology has significant pain points in the substation scenario: strong electromagnetic interference in the substation leads to unstable high-frequency data transmission, and the narrow bandwidth of low-frequency transmission cannot meet the real-time backhaul demand of high-definition images. The traditional full-quantity image backhaul mode either sacrifices detection accuracy by compressing image quality or affects inspection efficiency due to transmission delay. The path planning is too single, mostly targeting the shortest path, without considering the relevance of equipment types and environmental consistency, and without integrating electromagnetic interference risks and other safety constraints, which poses a flight safety hazard. In summary, the existing UAV inspection system has deficiencies in electromagnetic environment adaptability, operator management flexibility, path planning scientificity, and data transmission efficiency, and cannot simultaneously meet the core requirements of safe, efficient, and accurate substation inspection.

[0004] Therefore, in order to solve the problems existing in the prior art, the present application provides a substation equipment inspection path planning system based on a UAV. SUMMARY

[0005] In view of the deficiencies of the prior art, the purpose of the present application is to provide a substation equipment inspection path planning system based on a UAV.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] A substation equipment inspection path planning system based on a UAV, characterized in that it comprises:

[0008] A reference and feature management module for constructing and managing a three-dimensional digital reference and feature library of substation equipment, the feature library being associated with the geometric features of the equipment, the preset accurate observation pose, and the corresponding reference images collected thereby;

[0009] An intelligent operator management module for carrying equipment detection operators, generating new operators according to the similarity of equipment features; and for scoring the confidence of the corresponding equipment detection operators according to the current environmental conditions and the type of the equipment to be inspected, and prioritizing all operators to be used based on the confidence score.

[0010] A precise pose control module receives the precise observation pose and adjusts the unmanned aerial vehicle image acquisition point and the acquisition angle according to the precise observation pose;

[0011] An on-board edge processing module is configured to perform real-time analysis and processing on the real-time collected images, and perform preliminary comparison between the processed images and the reference images through a device detection algorithm, make a decision based on the comparison result, and perform selective data return;

[0012] A multi-target fusion path planning module selects a high-priority algorithm as a reference algorithm based on the priority sorting result, identifies similar devices covered by the reference algorithm through parameter migration, sets the devices covered by the reference algorithm and the similar devices as the same inspection batch, generates an optimal inspection sub-path for the inspection batch by minimizing the flight path and combining the flight risk comprehensive target, and iterates each priority algorithm and its corresponding batch to generate a global inspection path composed of the optimal inspection sub-paths.

[0013] As a further improvement of the application, the reference and feature management module further comprises:

[0014] A three-dimensional point cloud processing unit is configured to construct a high-precision three-dimensional point cloud model of the substation, and perform semantic segmentation on the three-dimensional point cloud model to identify and label each device instance therein;

[0015] An observation pose construction unit is configured to determine, for each identified device instance, a final observation pose for image acquisition and a reference image based on the final observation pose based on a preset observation pose construction strategy, the final observation pose comprising three-dimensional space coordinates and acquisition angle information.

[0016] As a further improvement of the application, the observation pose construction strategy comprises, for each identified device instance, identifying a key component region for inspection on the device based on its three-dimensional point cloud data, generating a plurality of candidate observation poses based on preset line-of-sight unobstructed conditions and safe operating distance conditions, performing multi-factor comprehensive evaluation on each candidate observation pose, including observation angle quality, coverage integrity and environmental adaptability, and selecting a final observation pose based on the observation angle quality, coverage integrity and environmental adaptability; the observation angle quality is the included angle between the observation axis of the pose pointing to the key component region and the average normal direction of the surface of the region, and the smaller the included angle, the higher the observation angle quality score; the coverage integrity is the proportion of the key component region that can be covered by the image frame taken at this pose calculated by simulating the camera field of view, and the higher the proportion, the higher the coverage integrity score; the environmental adaptability is the score value of the electromagnetic interference intensity and the background visual complexity at the location of the pose according to the pre-stored electromagnetic interference distribution map and the illumination model.

[0017] As a further improvement of the present invention, the intelligent operator management module includes:

[0018] The operator library unit is used to store and manage device detection operators, as well as the baseline feature vectors of the device types corresponding to the device detection operators;

[0019] The similarity calculation and transfer unit extracts the geometric and performance features of the new device instance and generates a feature vector. It performs similarity matching calculation between the feature vector and the benchmark feature vectors of various types of devices stored in the operator library unit. When the similarity is greater than a preset threshold, it uses the existing operator corresponding to the type of device as a pre-trained model, adjusts the parameters of the model through the sample data of the new device instance, and generates a similar device detection operator.

[0020] The operator adjustment unit acquires light intensity, weather conditions, and background complexity as real-time environmental data, and dynamically adjusts the image preprocessing parameters of the called operators based on the mapping relationship between the environmental data and the image preprocessing parameter adjustment amount preset by the Bayesian optimization framework.

[0021] As a further improvement of the present invention, the airborne edge processing module includes:

[0022] The real-time analysis unit is used to call the device detection operator output by the intelligent operator management module, process the acquired real-time image, compare the processed image with the corresponding benchmark image obtained by the benchmark and feature management module, and output the comparison result containing difference region information and confidence level.

[0023] The data feedback decision control unit is used to output a data feedback command based on the comparison result and a preset data feedback strategy.

[0024] As a further improvement of the present invention, the data backhaul strategy includes: if the comparison result indicates no significant difference and the confidence level is higher than a first threshold, a silent mode is triggered, image data backhaul is not started, and only inspection logs are recorded locally; if the comparison result indicates a difference and the confidence level is higher than a second threshold, a difference precision transmission mode is triggered, the pixel coordinates of the difference region of the real-time analysis unit are received, and the corresponding difference region sub-image is cropped from the acquired original high-definition image, and the difference region sub-image, device identifier, difference type, confidence level and timestamp are encapsulated into source data and backhauled to the data terminal; if the confidence level of the comparison result is lower than a third threshold, or an unrecognizable scene is detected, an abnormal degradation mode is triggered, and a high-resolution panoramic image and an abnormal alarm code are backhauled to the data terminal.

[0025] As a further improvement of the present invention, the multi-objective fusion path planning module includes receiving a set of precise observation poses corresponding to all equipment in the inspection batch, and obtaining an electromagnetic interference risk distribution map of the substation area. Each three-dimensional coordinate point in the electromagnetic interference risk distribution map is associated with a risk coefficient based on electromagnetic interference settings. Decision variables are defined, including the order of UAV access pose points and the path segments connecting each point. An optimization problem is established with the dual objectives of minimizing the total flight path length and minimizing the total exposure risk. The total flight path length is the sum of the Euclidean distances of all access segments. The total exposure risk is calculated by integrating the risk coefficients along the path, calculating the average risk coefficient of several uniformly sampled points of the access segment, and combining it with the segment length to obtain the risk cost of the access segment. The risk costs of all segments are calculated as the total exposure risk value. The optimization problem is solved using a non-dominated sorting genetic algorithm with an elite strategy, and a set of Pareto optimal solutions are output. Each solution corresponds to an access sequence and the corresponding total flight path length and total exposure risk value. The priority value of each solution is calculated comprehensively, the final solution is selected according to the priority value, and the access sequence and flight path corresponding to the solution are output as the optimal inspection sub-path.

[0026] As a further improvement of the present invention, the multi-target fusion path planning module further includes, when generating the optimal inspection sub-path for an inspection batch, using the accurate observation poses of all devices in the batch as necessary waypoints for path optimization, so that when performing inspection according to the optimal inspection sub-path, the airborne edge processing module can complete the image analysis of all devices on the path by relying only on the one reference operator.

[0027] As a further improvement of the present invention, the multi-objective fusion path planning module further includes: obtaining the optimal inspection sub-path corresponding to each batch and sorting it according to the priority of the benchmark operator; for each pair of sequentially adjacent optimal inspection sub-paths, obtaining the planning endpoint of the previous sub-path and the planning starting point of the next sub-path, and planning a transfer segment connecting the two points through a path search algorithm; during planning, the generation of the transfer segment is determined by evaluating its flight length and the cumulative risk of crossing the electromagnetic interference zone; connecting the sorted optimal inspection sub-paths with the transfer segment to generate a global inspection path, and calculating the total flight distance and total risk integral of the global inspection path, verifying whether it meets the total endurance constraint and global risk budget constraint of the UAV, and if it meets the constraint, outputting the global inspection path.

[0028] As a further improvement of the present invention, the substation equipment inspection path planning system performs the following:

[0029] Receive inspection task instructions and determine the target equipment set; call the reference and feature management module to obtain the three-dimensional digital model, preset precise observation pose and corresponding reference image of each device in the target equipment set;

[0030] The intelligent operator management module is invoked. For the target device, its features are matched with the operator library for similarity. If the match is successful, it is directly associated with an existing operator. If the matching degree is higher than the migration threshold, a dedicated operator is generated based on similar operators through parameter migration. If the matching degree is low, it is marked as needing an independent operator. According to the current environmental parameters, the preprocessing parameters and decision thresholds of all associated operators are dynamically adjusted through a preset environment and parameter mapping relationship. Combining environmental conditions and device type, the expected confidence score is calculated for all associated operators and prioritized.

[0031] The multi-objective fusion path planning module is invoked, and operators are sorted according to their priority. High-priority operators are selected as the baseline operators, and the equipment covered by them is grouped into the same inspection batch. For each inspection batch, with the goal of minimizing the combined cost of flight path length and electromagnetic exposure risk, the optimal inspection sub-path that traverses the observation poses of all equipment in the batch is generated. According to the batch priority order, transfer segments that also consider risk and efficiency are planned for adjacent sub-paths. All sub-paths and transfer segments are combined into a global inspection path sequence, and its compliance with endurance and safety constraints is verified.

[0032] The UAV loads the global inspection path sequence and executes flight; when it arrives near the preset observation pose of a device, the precise pose control module starts the visual servo closed loop: it performs feature matching between the real-time image and the reference image, generates a pose error signal, and controls the UAV and gimbal to fine-tune until the error is lower than the threshold and triggers high-definition image acquisition.

[0033] The airborne edge processing module is invoked to call the equipment inspection operator preset for the corresponding device on the acquired image, and the real-time image is compared and analyzed with the reference image. Based on the comparison result and confidence level, a three-state decision is executed: if there is no difference, the system remains silent and logs are recorded; if a difference is detected, a high-definition sub-image of the difference area is extracted and metadata is encapsulated and sent back; if the analysis fails, a high-resolution image and alarm are sent back.

[0034] The beneficial effects of this invention are:

[0035] (1) To resolve the contradiction of data transmission under electromagnetic interference, high-definition defect information can be efficiently transmitted in low-frequency links through real-time edge analysis and selective backhaul, taking into account both transmission stability and detection accuracy.

[0036] (2) Improve the flexibility and environmental robustness of operator adaptation. New operators are generated through feature transfer and parameters are dynamically adjusted to reduce the cost of adapting to new equipment and reduce false detections and missed detections caused by factors such as lighting and weather.

[0037] (3) Optimize the quality of inspection path planning, balance path efficiency and flight safety with multi-objective fusion decision-making, ensure consistency of the inspection environment through batch inspection, and improve overall inspection efficiency.

[0038] (4) Ensure the accuracy of image acquisition and comparison. Reproduce the standard observation pose through precise pose control, provide a high-quality comparison basis for defect identification, and enhance the reliability of inspection results. Attached Figure Description

[0039] Fig. 1 This is a block diagram of a UAV-based substation equipment inspection path planning system according to an embodiment of the present invention.

[0040] Fig. 2 This is a system workflow diagram of an embodiment of the present invention. Detailed Implementation

[0041] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0042] This invention proposes a substation equipment inspection path planning system based on unmanned aerial vehicles (UAVs), such as... Figs. 1-2 As shown, it includes:

[0043] The benchmark and feature management module is used to construct and manage a 3D digital benchmark and feature library for substation equipment. This feature library associates the geometric features of the equipment, preset precise observation poses, and corresponding acquired benchmark images. This module plays a crucial role in providing a unified and accurate comparison benchmark and core feature support for the entire inspection process. The 3D digital benchmark and feature library it constructs is the foundation for equipment detection, image comparison, and path optimization. This feature library does not simply store equipment-related data; rather, it establishes a mapping of multi-dimensional information about the equipment. The geometric features of the equipment encompass core attributes that reflect the physical form of the equipment, such as its shape, size parameters, and surface contours. These attributes are important bases for distinguishing different equipment types and judging equipment similarity. The preset precise observation poses are ideal acquisition parameter combinations determined after multi-dimensional optimization by the system, ensuring that the UAV accurately captures key equipment information when acquiring images. The corresponding benchmark images are standard images of defect-free equipment acquired under these ideal acquisition conditions, providing a reliable reference for subsequent real-time image defect comparison.

[0044] Specifically, such as Figs. 1-2 As shown, the benchmark and feature management module further includes:

[0045] A 3D point cloud processing unit is used to construct a high-precision 3D point cloud model of a substation, and to perform semantic segmentation on the 3D point cloud model to identify and label each equipment instance therein.

[0046] The core task of the 3D point cloud processing unit is to construct a high-precision 3D model that can realistically reproduce the substation scene and accurately identify equipment instances. The construction process of the high-precision 3D point cloud model involves using professional measuring equipment equipped with LiDAR to perform a comprehensive, multi-angle scan of the entire substation area, collecting massive amounts of 3D coordinate data that reflects the substation's topography, building distribution, and equipment spatial locations. Before entering the model construction stage, the raw data undergoes a series of preprocessing operations. First, invalid and noisy data caused by measurement errors and environmental interference are removed. Then, data registration technology is used to integrate the point cloud data acquired from different scanning perspectives into a complete data set under a unified coordinate system, ultimately forming a 3D point cloud model that accurately replicates the real substation scene. Semantic segmentation is based on deep learning algorithms to perform fine classification processing on the constructed 3D point cloud model. By training a mature recognition model, it captures the unique features of point cloud data from different devices, aggregates and classifies point cloud data belonging to the same device, and clearly delineates the boundary range between devices and the background environment, as well as between different devices. This enables accurate identification and independent labeling of each device instance in the substation, laying the foundation for subsequent observation pose planning and inspection operations for individual devices.

[0047] The observation pose construction unit, for each identified device instance, determines the final observation pose for image acquisition and the reference image based on the final observation pose, based on a preset observation pose construction strategy. The final observation pose includes three-dimensional spatial coordinates and acquisition viewpoint information.

[0048] The main task of the observation pose construction unit is to determine the optimal image acquisition parameters for each independently labeled device instance, ensuring that the acquired images meet the accuracy requirements of subsequent defect detection. Before determining the final observation pose, this unit first uses the device's 3D point cloud data to accurately locate the key component areas that need to be inspected. These areas are typically parts of the device prone to failure, aging, or damage during operation, and their detection results directly affect the judgment of the device's operating status. Based on the spatial location information of the key component areas, the system generates several candidate observation poses, strictly adhering to two core constraints during the generation process. The unobstructed line-of-sight condition requires that there are no obstacles on the acquisition path corresponding to the candidate observation pose that could obstruct the key component areas, including other equipment, buildings, vegetation, etc., ensuring that the acquired images can fully present the true state of the key components. The safe operating distance condition comprehensively considers the electrical characteristics of the equipment, the flight safety threshold of the UAV, and the electromagnetic environment distribution within the substation, ensuring that the UAV maintains a sufficient safe distance from the equipment when acquiring images in this pose, avoiding both electromagnetic interference affecting the normal operation of the UAV and preventing safety accidents caused by physical contact.

[0049] Specifically, such as Figs. 1-2 As shown, the observation pose construction strategy includes, for the identified equipment instance, identifying the key component area used for inspection on the equipment based on its 3D point cloud data, generating several candidate observation poses based on preset unobstructed line-of-sight conditions and safe operating distance conditions, comprehensively evaluating each candidate observation pose based on multiple factors, including observation view quality, coverage integrity, and environmental adaptability, and selecting the final observation pose by comprehensively considering the observation view quality, coverage integrity, and environmental adaptability; the observation view quality is the angle between the observation axis pointing to the key component area of ​​the pose and the average normal direction of the surface of the area, the smaller the angle, the higher the view quality score; the coverage integrity is the proportion of the key component area that can be covered by the image frame taken under the pose, calculated by simulating the camera's field of view, the higher the proportion, the higher the coverage integrity score; the environmental adaptability is the score value of judging the electromagnetic interference intensity and background visual complexity at the location of the pose based on the pre-stored electromagnetic interference distribution map and illumination model.

[0050] The core of evaluating the quality of the observation viewpoint is determining whether the acquisition viewpoint can most clearly present the surface features of the key component area. Essentially, it measures the degree to which the observation direction matches the natural orientation of the key component area's surface. The average normal direction of the key component area's surface reflects its natural orientation. The degree of deviation between the direction of the observation axis pointing towards the area and this normal direction directly affects the image's clarity and feature recognition. The smaller the deviation, the closer the observation viewpoint is to the optimal presentation angle of the key component, and the higher the corresponding viewpoint quality score.

[0051] The assessment of coverage integrity involves simulating the actual shooting range and imaging effect of the camera to analyze the size of the area of ​​the critical component that the acquired image can cover under a specific candidate observation pose. Based on information such as camera lens parameters and shooting distance, the system simulates and calculates the coverage range of the image frame, and then determines the overlap ratio between this range and the critical component area. The higher the overlap ratio, the greater the possibility of detecting the critical component defect under that pose, and the higher the coverage integrity score.

[0052] Environmental adaptability assessment requires combining pre-mapped data and environmental models of the substation to comprehensively determine whether the environment at the candidate observation pose location is suitable for image acquisition. Specifically, the system will refer to pre-stored electromagnetic interference distribution data to determine whether the electromagnetic interference intensity at the location is within the tolerance range of the UAV equipment, avoiding abnormal image acquisition or equipment failure due to strong electromagnetic interference. At the same time, it will combine illumination models to analyze whether the illumination conditions at the location are stable and the complexity of the background environment, avoiding the inability to distinguish key component areas from the background due to excessively strong or weak illumination or overly complex backgrounds, which would affect the subsequent image comparison results. The environmental adaptability score is formed by combining the assessment results of these two aspects.

[0053] During the comprehensive evaluation process, the system assigns appropriate weights to the three evaluation dimensions based on the core requirements of the inspection task. For example, in inspection tasks where detection accuracy is a primary concern, the weights of observation view quality and coverage integrity are relatively high; while in substation areas with severe electromagnetic interference, the weight of environmental adaptability is appropriately increased. Through weight allocation, the scores of the three dimensions are transformed into a unified comprehensive score, and the candidate observation pose with the highest comprehensive score is selected as the final observation pose. Simultaneously, the system acquires standard images of the equipment in this final observation pose, correlates them with the equipment's geometric features and observation pose parameters, and stores them in a 3D digital benchmark and feature library. This provides standardized basic data support for subsequent image acquisition and comparative analysis during UAV inspections.

[0054] The intelligent operator management module is used to carry device detection operators and generate new operators based on the similarity of device features; it is also used to perform confidence scoring on the corresponding device detection operators according to the current environmental conditions and the type of device to be inspected, and to prioritize all operators to be used based on the confidence scores.

[0055] Specifically, such as Figs. 1-2 As shown, the intelligent operator management module includes:

[0056] The operator library unit is used to store and manage device detection operators, as well as the baseline feature vectors of the device types corresponding to the device detection operators;

[0057] The equipment detection operator is not a single algorithm, but a set of algorithms designed for specific equipment types or defect types. It includes a series of coherent algorithmic logics such as image preprocessing, feature extraction, defect identification, and result judgment, and can independently complete the entire process from image input to defect judgment. The benchmark feature vector is a digital extraction of the core features of each equipment type, covering multi-dimensional key information such as the equipment's geometric shape features, surface texture features, and spectral response features. This information is the core basis for distinguishing different equipment types and judging equipment similarity, and it is also the comparison benchmark for subsequent similarity calculations.

[0058] In terms of storage management, the operator library unit manages device detection operators and their corresponding baseline feature vectors using a categorized archiving approach. The system establishes a classification system based on dimensions such as device function and structural type, associating and storing the operators and baseline feature vectors corresponding to each type of device, and constructing an efficient indexing mechanism. This indexing mechanism can quickly locate the corresponding operators and baseline feature vectors based on device type, feature keywords, etc., significantly improving retrieval efficiency. Simultaneously, the operator library unit supports dynamic updates and iterations of operators, enabling optimization and adjustments to existing operators or the addition of operators adapted to new devices based on actual inspection results and equipment upgrades, ensuring the timeliness and applicability of the operator library. Furthermore, the unit also has version management capabilities, recording the iteration process of operators. When a new operator encounters compatibility issues, it can be quickly rolled back to a stable version, ensuring system stability.

[0059] The similarity calculation and transfer unit extracts the geometric and performance features of the new device instance and generates a feature vector. It then performs similarity matching calculations between the feature vector and the benchmark feature vectors of various types of devices stored in the operator library unit. When the similarity is greater than a preset threshold, it uses the existing operator corresponding to that type of device as a pre-trained model. The model's parameters are adjusted using the sample data of the new device instance, and a similar device detection operator is generated.

[0060] The core value of the similarity calculation and migration unit lies in enabling operator reuse and rapid adaptation, avoiding the need to develop dedicated operators for each new device, significantly reducing operator development costs, and improving the system's adaptability to new devices. The geometric features of a new device instance encompass characteristics reflecting the device's physical form, such as its external dimensions, structural layout, relative positions of components, and surface contour curves; while the performance features include surface texture patterns, color distribution, material reflectivity, and thermal radiation characteristics under normal operating conditions, reflecting the device's appearance and operating status.

[0061] The feature vector generation process involves standardizing and digitizing multi-dimensional features. The system uses specialized feature extraction algorithms to transform geometric and representational features into a set of ordered values, forming a feature vector that quantifies the characteristics of the device. The core logic of similarity matching calculation is to measure the degree of fit between the new device's feature vector and the baseline feature vectors of various devices in the operator library. A specific distance metric is used to analyze the difference between the two sets of vectors; the smaller the difference, the higher the feature similarity between the new device and that type of device.

[0062] The preset threshold is a similarity threshold determined by the system based on a large amount of historical data and engineering practice experience. When the similarity exceeds this threshold, it indicates that the new device and the corresponding existing device type have little difference in core features, and the operators of the existing device have the basis for transfer and adaptation. At this time, the system uses the existing operator as a pre-trained model and fine-tunes the key parameters of the model using a small amount of sample data from the new device instance. This parameter adjustment is not a model reconstruction, but rather, while retaining the core algorithm framework of the original operator, it allows the model to adapt to the specific features of the new device, ultimately generating a similar device detection operator that can accurately detect the new device. This process does not require training the model from scratch, which saves computing resources, shortens the operator adaptation cycle, and achieves rapid reuse and efficient transfer of operators.

[0063] The operator adjustment unit acquires light intensity, weather conditions, and background complexity as real-time environmental data, and dynamically adjusts the image preprocessing parameters of the called operators based on the mapping relationship between the environmental data and the image preprocessing parameter adjustment amount preset by the Bayesian optimization framework.

[0064] The main function of the operator adjustment unit is to improve the operator's adaptability to the real-time environment, reduce the interference of environmental factors such as lighting and weather on the detection results, and ensure that the operator can maintain stable detection performance in complex and ever-changing inspection environments. Light intensity refers to the brightness of the light in the inspection area, and its changes directly affect the image's brightness, contrast, and other quality parameters. Weather conditions include different meteorological conditions such as sunny days, cloudy days, fog, haze, and light rain. Fog can cause image blurring, and light rain can cause image reflection and occlusion, both of which can interfere with the operator's feature extraction and defect identification. Background complexity refers to the number and distribution of interfering elements in the environment surrounding the equipment. Too many or messy interfering elements will increase the difficulty for the operator to distinguish the equipment from the background, easily leading to false detections.

[0065] The operator adjustment unit collects environmental data in real time, including light intensity, weather conditions, and background complexity at the inspection site. This data is input into a pre-defined mapping model. Based on learned correlation patterns, the model outputs the optimal adjustment direction and magnitude for each image preprocessing parameter under the current environment. Following this adjustment instruction, the operator adjustment unit dynamically updates the parameters of the operators to be invoked, enabling the image preprocessing stage to specifically compensate for image quality defects caused by environmental factors. For example, in low-light conditions, the system increases the contrast enhancement coefficient and brightness adjustment parameters to make image details clearer; in hazy weather, it enhances noise suppression to reduce image blur. Through this dynamic adjustment mechanism, the operators can always operate with optimal parameter configurations, significantly improving environmental robustness and ensuring the accuracy of the detection results.

[0066] In addition, the intelligent operator management module also features confidence scoring and priority ranking functions. Confidence scoring comprehensively evaluates the detection reliability of each available operator by considering current environmental conditions and the type of device under inspection. The evaluation criteria include multiple dimensions such as the operator's historical accuracy in similar environments and device types, the matching degree between the operator's computational complexity and the UAV's computing power, and the fit between the current environmental data and the operator's optimal operating environment. The system comprehensively analyzes information from these dimensions to derive a confidence score for each operator; a higher score indicates more reliable detection performance in the current scenario.

[0067] Priority sorting arranges all available operators sequentially based on their confidence scores, assigning the highest priority to the operator with the highest confidence score, which is then called by the system to execute the detection task first. This sorting mechanism ensures that the system prioritizes the operator most suitable for the current scenario during each detection, avoiding false positives and false negatives caused by low-confidence operators, and improving the overall inspection speed by prioritizing the use of computationally efficient operators, thus achieving a dual optimization of detection accuracy and computational efficiency.

[0068] The precise pose control module receives the precise observation pose and adjusts the image acquisition points and acquisition angle of the UAV according to the precise observation pose.

[0069] The precise pose control module is the core module that ensures a high degree of consistency between the inspection images and the reference images. Its core objective is to enable the UAV to accurately reproduce the preset precise observation pose in a dynamic flight environment, ensuring that the image acquisition points and viewing angles perfectly match the reference conditions. This provides high-quality image data without viewing angle deviation or positional offset for the subsequent feature comparison by the onboard edge processing module. This module does not simply execute position movement commands, but rather achieves high-precision closed-loop control of the UAV's spatial position and camera attitude through multi-dimensional perception, real-time calibration, and dynamic compensation mechanisms.

[0070] Precise observation of the drone's pose serves as the control target benchmark for the module, encompassing key parameters across two core dimensions. The three-dimensional spatial coordinates define the drone's precise position within the substation's global coordinate system. This coordinate system not only includes horizontal and vertical distance parameters but also requires maintaining a preset safe operating distance from the equipment under inspection, while avoiding areas with strong electromagnetic interference to ensure drone flight safety and prevent environmental interference during equipment inspection. The acquisition perspective defines the camera's attitude parameters, including pitch, roll, and yaw angles. These parameters collectively determine the camera's framing direction during image capture, ensuring that critical component areas of the equipment are captured from an angle perfectly consistent with the reference image, avoiding misjudgments in feature comparison due to perspective deviations.

[0071] In the data reception and parsing phase, the precise pose control module receives precise observation pose data from the reference and feature management module via the UAV's communication link. After standardized encoding, the module first decodes and verifies the data, eliminating any abnormal data that may have occurred during transmission to ensure the accuracy of the control target. Subsequently, the module converts the decoded 3D spatial coordinates and acquired viewpoint parameters into control command parameters that can be recognized by the UAV flight control system and gimbal control system, establishing a mapping relationship between the control target and the actuators. Specifically, the spatial coordinate parameters are converted into UAV flight position commands, and the acquired viewpoint parameters are converted into gimbal attitude adjustment commands, providing a clear execution basis for subsequent precise control.

[0072] In the position and attitude adjustment phase, the module integrates multiple high-precision control technologies to achieve coordinated control. For precise positioning in three-dimensional space, the module combines high-precision positioning technology with inertial navigation technology. High-precision positioning technology can acquire the UAV's current position information in real time, compare it with the target space coordinates, and calculate the position deviation; inertial navigation technology, by sensing the UAV's acceleration, angular velocity, and other motion parameters, predicts the trend of position changes and intervenes in advance. When a position deviation exists, the module sends fine-tuning commands to the UAV's propulsion system, controlling the difference in propeller speed to achieve precise movement of the UAV in the horizontal and vertical directions, gradually reducing the position deviation.

[0073] For precise adjustment of the acquisition perspective, the module achieves this by controlling the drone's gimbal system. As the core component carrying the camera, the gimbal has independent attitude adjustment capabilities. Based on the analyzed acquisition perspective parameters, the module sends control signals to the gimbal's pitch, roll, and yaw axes, driving the gimbal to rotate at the corresponding angles. During adjustment, the gimbal's built-in angle sensor provides real-time feedback on the current attitude information. The module compares this feedback information with the target perspective parameters, continuously fine-tuning until the perspective deviation meets the preset requirements. This gimbal control method, independent of the drone's fuselage, effectively isolates the impact of fuselage vibration or attitude changes on the camera's perspective, ensuring the stability of the shooting angle.

[0074] To address external interference and accumulated errors during flight, the module establishes a real-time calibration and dynamic compensation mechanism, achieving high-precision correction through a visual servo closed-loop system. When the UAV flies to a preset range near the precise observation pose, the module automatically initiates the visual servo closed-loop process. First, the camera acquires image frames of the current scene in real time. The module then calls a feature matching algorithm to extract key feature points from the image frame, which correspond one-to-one with preset feature points in the reference image. By calculating the positional offset of the two sets of feature points on the image plane, the module reverse-engineers the actual error signals between the UAV's current position and the target position, and between the current viewing angle and the target viewing angle.

[0075] This error signal is fed back to the control unit in real time. The control unit dynamically adjusts the intensity of position and view control commands based on the magnitude and direction of the error. For example, if feature point offset indicates the drone is positioned to the left, a command to increase power is sent to the right-side propulsion system; if feature point offset indicates insufficient camera pitch angle, a command to adjust downwards is sent to the gimbal's pitch axis. This real-time calibration mechanism based on image features can accurately compensate for position and view deviations caused by wind interference, inertial drift, and other factors, ensuring that the pose error of the final acquired image is below a preset threshold.

[0076] Furthermore, the module possesses environmental adaptive compensation capabilities. By receiving environmental sensor data from the UAV, the module can perceive environmental interference factors such as wind speed, wind direction, and air turbulence in real time. For different types of interference, the module presets corresponding compensation strategies: when high wind speed is detected, the response intensity of position control commands is appropriately increased to improve the UAV's wind resistance stability; when encountering air turbulence causing slight vibrations in the fuselage, the damping control of the gimbal is strengthened to reduce the impact of vibration on the camera's field of view. This combination of environmental adaptive compensation and visual servo closed-loop calibration enables the precision pose control module to maintain stable high-precision control performance in complex and ever-changing inspection environments, providing reliable assurance for subsequent image comparison and defect detection.

[0077] The airborne edge processing module is used to perform real-time analysis and processing on the acquired images, and to perform a preliminary comparison between the processed images and the reference images through the device detection operator. Based on the comparison results, it makes decisions and performs selective data backhaul.

[0078] Specifically, such as Figs. 1-2 As shown, the airborne edge processing module includes:

[0079] The real-time analysis unit is used to call the device detection operator output by the intelligent operator management module to process the acquired real-time image, and compare the processed image with the corresponding benchmark image obtained by the benchmark and feature management module, and output the comparison result containing information on the difference region and confidence level.

[0080] The invocation of device detection operators follows the priority ranking results output by the intelligent operator management module. The system prioritizes the operator with the highest confidence level and the best fit for the current environment and device type, ensuring the efficiency and accuracy of the analysis process. The real-time image processing involves multiple refined operations: First, the acquired raw image is preprocessed, eliminating image quality loss caused by lighting changes and environmental interference through noise reduction, contrast enhancement, and distortion correction. Then, the operator extracts key features of the device in the image, covering the device's geometric contours, surface texture, color distribution, and other core criteria for comparison with the reference image. Finally, a feature matching algorithm precisely compares the real-time image features with the reference image features point-by-point and region-by-region to identify the differences between them.

[0081] The information on the difference region is the core output of the comparison results. It includes not only the specific location and range of the difference, but also detailed information such as the difference's morphological features, size, color changes, and texture differences. For example, rust on the surface of equipment will manifest as specific color and texture changes, and loose bolts will show a relative positional shift; these differences will be accurately captured and recorded. Confidence score is a quantitative assessment of the reliability of the difference recognition results. Its calculation is based on multiple dimensions, including the overlap of feature matching, the salience of the difference region, the historical accuracy of the operator in the current environment, and the impact of image quality on the recognition results. The higher the overlap, the more significant the difference, and the stronger the operator's adaptability, the higher the confidence score, meaning the more reliable the recognition result; conversely, the lower the confidence score, the more necessary further verification is required.

[0082] When outputting results, the real-time analysis unit calibrates the coordinates of the discrepancy regions, converting their positions on the image plane into precise pixel coordinates, providing accurate basis for subsequent discrepancy region cropping. Simultaneously, it converts confidence scores into intuitive numerical ratings, storing them in association with the discrepancy region information, jointly providing data support for decision-making. The entire analysis process is completed locally and in real-time on the drone, without relying on cloud computing power, avoiding efficiency losses due to data transmission latency, and reducing the risk of data loss or interference during transmission.

[0083] The data feedback decision control unit is used to output a data feedback command based on the comparison result and a preset data feedback strategy.

[0084] This unit first receives the complete comparison results output by the real-time analysis unit, including detailed information on the differences and corresponding confidence scores, and then analyzes and judges the information. The preset backhaul strategy is based on the actual needs of substation inspection, transmission link characteristics, and data value priority, covering three core backhaul modes. Each mode corresponds to clear triggering conditions and execution logic, and supports flexible adaptation according to the actual inspection scenario.

[0085] The three thresholds involved in the feedback strategy are quantitative standards determined by the system based on a large amount of historical inspection data, defect identification accuracy requirements, and engineering practice experience. They are used to divide different decision intervals. The first threshold is a high reliability standard for judging no difference. Only when the comparison results show no significant difference and the confidence level reaches this standard can the equipment be judged to be in normal condition. The second threshold is a critical standard for judging the reliability of the difference. It ensures that the identified difference is a real equipment defect, rather than caused by environmental interference or image error. The third threshold is a critical standard for judging the failure of the analysis. When the confidence level is lower than this standard, it means that the reliability of the analysis results cannot be guaranteed, and other methods are needed to supplement the verification.

[0086] Specifically, such as Figs. 1-2As shown, the data feedback strategy includes the following: if the comparison result indicates no significant difference and the confidence level is higher than a first threshold, a silent mode is triggered, image data feedback is not initiated, and only inspection logs are recorded locally. When the comparison result shows no significant difference between the real-time image and the reference image, and the confidence level is higher than the first threshold, the system triggers a silent mode. At this time, the device is determined to be in normal operating condition, and there is no need to feedback image data; only inspection logs are stored locally on the drone. The log content includes key information such as inspection time, device identification, inspection pose, comparison confidence level, and no-difference determination result, ensuring the traceability of the inspection process and avoiding invalid data occupying transmission bandwidth. The core value of this mode is to minimize transmission pressure, allowing the bandwidth of the low-frequency link to be concentrated on transmitting valuable abnormal data, thereby improving overall inspection efficiency.

[0087] If the comparison result indicates a difference and the confidence level is higher than the second threshold, the difference precision transmission mode is triggered. This mode receives the pixel coordinates of the difference region from the real-time analysis unit and crops the corresponding difference region sub-image from the acquired original high-definition image. The difference region sub-image, device identifier, difference type, confidence level, and timestamp are encapsulated as source data and transmitted back to the data terminal. When the comparison result shows a clear difference and the confidence level is higher than the second threshold, the system triggers the difference precision transmission mode. This mode is a precise transmission solution for device defects, aiming to transmit the most critical defect information with minimal data volume. First, the feedback decision control unit receives the pixel coordinates of the difference region output by the real-time analysis unit. Based on these coordinates, the original high-definition image is precisely cropped, retaining only the local sub-image containing the difference and removing worthless background areas, significantly compressing the data volume. Subsequently, the system standardizes and encapsulates the transmitted data, including not only the sub-images of the difference region, but also metadata such as equipment identifier (for quickly locating defective equipment), difference type (the defect category initially judged based on operator analysis, such as rust, loosening, and damage), confidence level (for terminal personnel to refer to for reliability identification), and timestamp (recording the time of defect discovery). The encapsulated source data has a unified format and complete information, which is convenient for low-frequency link transmission and allows ground terminals to quickly parse key defect information without processing complete high-definition images, thus improving the efficiency of defect handling.

[0088] If the confidence level of the comparison result is lower than the third threshold, or if an unrecognizable scene is detected, an abnormal degradation mode is triggered. A high-resolution panoramic image and anomaly alarm code are then sent back to the data terminal. This mode is a fallback solution to ensure the integrity of the inspection, aiming to prevent missed inspections due to analysis failure. In this case, the system abandons local data transmission and instead sends back a high-resolution panoramic image of the inspection point, ensuring that the ground terminal can obtain complete visual information of the device for subsequent manual verification. An anomaly alarm code is also included, containing a classification identifier for the cause of failure (such as lighting interference, image blurring, operator failure, etc.), helping ground personnel quickly locate the root cause of the problem and take targeted re-inspection or system optimization measures. Although this mode consumes more bandwidth, it is only triggered in scenarios of analysis failure and ensures the integrity of the inspection data, avoiding inspection loopholes caused by technical limitations.

[0089] The airborne edge processing module, through the accurate identification of the real-time analysis unit and the intelligent decision-making of the feedback decision control unit, realizes a differentiated transmission strategy. This not only solves the core contradiction that low-frequency bandwidth cannot support the real-time feedback of high-definition images, but also ensures the integrity and effectiveness of inspection data, providing key technical support for the large-scale application of UAV inspection in substations.

[0090] The multi-objective fusion path planning module, based on the priority ranking result, selects the high-priority operator as the benchmark operator, identifies similar devices covered by the benchmark operator through parameter migration, and sets the devices covered by the benchmark operator and similar devices as the same inspection batch. Within the same inspection batch, it generates the optimal inspection sub-path for the inspection batch by minimizing the flight path and combining the comprehensive flight risk objective. It then traverses each priority operator and its corresponding batch to generate a global inspection path composed of each of the optimal inspection sub-paths.

[0091] Specifically, such as Figs. 1-2As shown, the multi-objective fusion path planning module includes receiving the precise observation pose set corresponding to all equipment in the inspection batch and obtaining the electromagnetic interference risk distribution map of the substation area. Each three-dimensional coordinate point in the electromagnetic interference risk distribution map is associated with a risk coefficient based on electromagnetic interference settings. Decision variables are defined, including the order of UAV access pose points and the path segments connecting each point. An optimization problem is established with the dual objectives of minimizing the total flight path length and minimizing the total exposure risk. The total flight path length is the sum of the Euclidean distances of all access segments. The total exposure risk is calculated by integrating the risk coefficient along the path, calculating the average risk coefficient of several uniformly sampled points of the access segment, and combining it with the segment length to obtain the risk cost of the access segment. The risk costs of all segments are calculated as the total exposure risk value. The optimization problem is solved using a non-dominated sorting genetic algorithm with an elite strategy, and a set of Pareto optimal solutions are output. Each solution corresponds to an access sequence and the corresponding total flight path length and total exposure risk value. The priority value of each solution is calculated comprehensively, the final solution is selected according to the priority value, and the access sequence and flight path corresponding to the solution are output as the optimal inspection sub-path.

[0092] Specifically, such as Figs. 1-2 As shown, the multi-target fusion path planning module further includes, when generating the optimal inspection sub-path for an inspection batch, using the accurate observation poses of all devices in the batch as necessary waypoints for path optimization, so that when performing inspections based on the optimal inspection sub-path, the airborne edge processing module can complete the image analysis of all devices on the path by relying only on the one reference operator.

[0093] Inspection batch division is a prerequisite for path planning. Its core logic is to group devices with strong adaptability and consistent environmental sensitivity together, reducing the frequency of operator switching and the interference of environmental factors on the detection results during the inspection process. The module first receives the operator priority ranking results output by the intelligent operator management module. The higher the priority of the operator, the better its detection confidence, adaptability and computational efficiency in the current environment, and therefore it is selected as the benchmark operator.

[0094] The coverage of the benchmark operator is determined by its parameter transfer capability. The module identifies all devices that the benchmark operator can adapt to, including the original corresponding devices and similar devices that can be covered through parameter transfer, and integrates the devices into the same inspection batch. The determination of similar devices is based on the degree of fit of the device feature vectors, that is, the feature similarity calculated by the intelligent operator management module, ensuring that the benchmark operator can accurately detect all devices in the batch without significant parameter adjustments.

[0095] Furthermore, batch segmentation fully considers the principle of environmental consistency. The module references environmental data such as the light distribution and electromagnetic interference distribution in the substation area to ensure that equipment in the same batch is placed under similar environmental conditions. This aims to reduce the number of operator parameter adjustments caused by drastic environmental changes, lower computational costs, and ensure consistent image acquisition conditions within the same batch, thereby improving the accuracy of feature comparison. For example, similar equipment located in the same area of ​​the substation with similar lighting conditions and comparable electromagnetic interference intensity is grouped together to ensure that the baseline operator remains in optimal working condition throughout the batch's inspection process.

[0096] The generation of the optimal inspection sub-path is a refined path planning for a single inspection batch. The core is to achieve the dual minimization of flight path length and flight risk based on the accurate observation of the pose of all equipment in the batch, while ensuring the adaptability of the detection operator.

[0097] The module first receives the set of precise observation poses corresponding to all devices in the batch. Each precise observation pose is an ideal acquisition parameter optimized by the reference and feature management module, which includes three-dimensional spatial coordinates and acquisition viewpoint information. It is a waypoint that the UAV must reach. Only by accurately reproducing the pose can the comparison accuracy between the acquired image and the reference image be guaranteed. Therefore, the pose cannot be omitted or replaced in the path planning.

[0098] Simultaneously, the module acquires an electromagnetic interference risk distribution map of the substation area. This map is a three-dimensional environmental model generated by mapping the electromagnetic environment of the entire substation area and combining data such as equipment operating parameters and transmission line distribution. Each three-dimensional coordinate point in the distribution map is associated with a risk coefficient. The magnitude of the risk coefficient is determined by the electromagnetic interference intensity at that location; the stronger the interference, the higher the risk coefficient, and the greater the probability of signal interference and equipment failure when the drone flies at that location.

[0099] Defining decision variables is a prerequisite for path optimization. The module clearly defines two core decision variables: first, the order in which the UAV visits the pose points, i.e., the traversal order of all necessary waypoints within the inspection batch; and second, the path segments connecting the waypoints, i.e., the specific flight trajectory of the UAV from one waypoint to the next. These two variables directly determine the total flight path length and the magnitude of the total exposure risk, and are the core control objects of the optimization problem.

[0100] The optimization problem of module construction has two core objectives: minimizing the total flight path length and minimizing the total exposure risk. These two objectives are complementary and mutually restrictive. Simply pursuing a short path may lead to crossing areas with high electromagnetic interference, increasing flight risk; simply avoiding risks may lead to a large detour of the path, reducing inspection efficiency.

[0101] The calculation logic for the total flight path length is to sum the straight-line distances of all path segments. The distance of each segment is calculated based on the three-dimensional coordinates between waypoints. The total length directly reflects the inspection efficiency. The shorter the length, the less inspection time is required and the lower the drone's endurance consumption.

[0102] The calculation of total exposure risk revolves around electromagnetic interference risk, employing a "segment risk cost accumulation" logic. For each flight segment, the module performs uniform sampling to obtain risk coefficients at several sampling points. By calculating the average risk coefficient of these sampling points, the average risk intensity of the segment is obtained. This average risk intensity is then multiplied by the segment length to obtain the risk cost of that segment, representing the total risk the UAV experiences while flying in that segment. The risk costs of all segments are accumulated to form the total exposure risk value for the entire sub-path, directly reflecting the flight safety level.

[0103] The module employs a non-dominated sorting genetic algorithm with an elite strategy to solve bi-objective optimization problems. This algorithm is an intelligent optimization method that can find balanced solutions among multiple mutually constrained optimization objectives. Its core advantage lies in avoiding the limitations of single-objective optimization while retaining high-quality solutions to improve solution efficiency.

[0104] The role of the elite strategy is to retain the best-performing solution in each generation of optimization during algorithm iteration, preventing the loss of high-quality solutions due to iteration, and ensuring the stability and convergence speed of the solution process. Non-dominated sorting, on the other hand, performs hierarchical screening of all candidate solutions, selecting those that perform well in one objective without significantly lagging behind in the other. These solutions are called Pareto optimal solutions, and together they constitute the optimal solution set for the bi-objective optimization problem.

[0105] The module outputs a set of Pareto optimal solutions using this algorithm. Each solution corresponds to a specific waypoint visit sequence, along with the total flight path length and total exposure risk value for that sequence. For example, one solution may have a shorter path length but slightly higher risk, while another solution may have lower risk but a slightly longer path. There is no absolute superiority or inferiority among the solutions; rather, they correspond to different operational strategy preferences.

[0106] To determine the final solution, the module calculates the priority of each Pareto optimal solution. The priority is calculated based on preset operational strategy weights: if efficiency is emphasized in the inspection task, path length is assigned a higher weight; if safety is emphasized, risk value is assigned a higher weight; and if a balance is sought, both are assigned equal weights. After weighted calculation, the solution with the highest priority is selected as the final solution, and its corresponding waypoint access sequence and flight trajectory constitute the optimal inspection sub-path for that inspection batch.

[0107] Specifically, such as Figs. 1-2As shown, the multi-objective fusion path planning module further includes: obtaining the optimal inspection sub-path corresponding to each batch and sorting it according to the priority of the benchmark operator; for each pair of sequentially adjacent optimal inspection sub-paths, obtaining the planning endpoint of the previous sub-path and the planning starting point of the next sub-path, and planning a transfer segment connecting the two points through a path search algorithm; during planning, the generation of the transfer segment is determined by evaluating its flight length and the cumulative risk of crossing the electromagnetic interference zone; connecting the sorted optimal inspection sub-paths with the transfer segment to generate a global inspection path, and calculating the total flight distance and total risk integral of the global inspection path to verify whether it meets the total endurance constraint and global risk budget constraint of the UAV; if it meets the constraint, the global inspection path is output.

[0108] The global inspection path integrates the optimal inspection sub-paths of all inspection batches into a coherent and feasible complete path. The core is to solve the connection problem between sub-paths and verify the overall feasibility of the path.

[0109] The module first sorts all optimal inspection sub-paths according to the priority of the baseline operators for each batch. The higher the priority of the sub-path, the stronger the detection reliability and adaptability of the corresponding baseline operator, and the corresponding equipment usually also has a higher inspection priority (such as critical equipment or equipment with a high failure rate). This sorting logic ensures that high-priority equipment can be inspected first, enhancing the core value of the inspection task, while avoiding omissions or delays in the inspection of high-priority equipment due to low-priority sub-paths consuming too much battery life and time.

[0110] The transfer segment is a crucial link connecting two adjacent optimal inspection sub-paths, and its planning objective is to find a balance between the shortest flight distance and the lowest cumulative risk. The module first obtains the planned endpoint of the previous sub-path, i.e., the precise observation pose coordinates of the last device in the batch, and the planned starting point of the next sub-path, i.e., the precise observation pose coordinates of the first device in the next batch. Then, it uses a path search algorithm to find the optimal trajectory connecting the two points.

[0111] The core evaluation metrics of the path search algorithm include two aspects: first, the flight length of the transfer segment, which is more conducive to saving endurance and time; second, the cumulative risk of the transfer segment, namely the total electromagnetic interference risk coefficient of the area traversed by the segment, which is more conducive to ensuring flight safety. The algorithm traverses multiple possible trajectory schemes and selects the scheme with the best overall performance in terms of both length and risk as the final transfer segment. For example, if there is a direct path between two points but it crosses a high electromagnetic interference area, while a detour path only increases the length slightly but avoids the high-risk area, then the detour path is selected as the transfer segment.

[0112] After the connection is completed, the module needs to perform two key constraint verifications on the global inspection path to ensure that it meets the actual execution requirements of the project:

[0113] Total range constraint verification: Calculate the total flight distance of the global inspection path and, in conjunction with the UAV's range parameters (such as energy consumption per unit distance and maximum range), determine whether the UAV can complete the entire inspection task without recharging. If the total flight distance exceeds the UAV's maximum range, it is necessary to adjust the sub-path ordering or transfer segment planning, such as splitting longer sub-paths or optimizing transfer segments to shorten the total length.

[0114] Global risk budget constraint verification: Calculate the total risk score of the global inspection path, which is the sum of the risk costs of all sub-paths and transfer segments. The global risk budget is the maximum tolerable risk value preset based on the UAV equipment's tolerance and operational safety standards. If the total risk score exceeds this budget, high-risk segments need to be re-optimized, for example, by replacing some paths to avoid high electromagnetic interference areas, or by adjusting the order of sub-paths to reduce the number of crossings of high-risk areas.

[0115] Only when the global inspection path simultaneously meets both the total range constraint and the global risk budget constraint will it be output as an executable path; otherwise, it will return to the sub-path optimization or transfer segment planning stage for adjustment until it meets the constraint requirements.

[0116] The multi-objective fusion path planning module achieves intelligent decision-making throughout the entire process, from equipment grouping to path generation, through batch partitioning, sub-path dual-objective optimization, and global path integration verification. The generated inspection paths ensure both the accuracy of equipment inspection and flight safety while maximizing inspection efficiency, perfectly meeting the needs of UAV inspections in substations with strong electromagnetic interference.

[0117] Specifically, such as Figs. 1-2 Figs. 1-2 As shown, the substation equipment inspection route planning system performs the following steps:

[0118] Receive inspection task instructions and determine the target equipment set; call the reference and feature management module to obtain the three-dimensional digital model, preset precise observation pose and corresponding reference image of each device in the target equipment set;

[0119] The intelligent operator management module is invoked. For the target device, its features are matched with the operator library for similarity. If the match is successful, it is directly associated with an existing operator. If the matching degree is higher than the migration threshold, a dedicated operator is generated based on similar operators through parameter migration. If the matching degree is low, it is marked as needing an independent operator. According to the current environmental parameters, the preprocessing parameters and decision thresholds of all associated operators are dynamically adjusted through a preset environment and parameter mapping relationship. Combining environmental conditions and device type, the expected confidence score is calculated for all associated operators and prioritized.

[0120] The multi-objective fusion path planning module is invoked, and operators are sorted according to their priority. High-priority operators are selected as the baseline operators, and the equipment covered by them is grouped into the same inspection batch. For each inspection batch, with the goal of minimizing the combined cost of flight path length and electromagnetic exposure risk, the optimal inspection sub-path that traverses the observation poses of all equipment in the batch is generated. According to the batch priority order, transfer segments that also consider risk and efficiency are planned for adjacent sub-paths. All sub-paths and transfer segments are combined into a global inspection path sequence, and its compliance with endurance and safety constraints is verified.

[0121] The UAV loads the global inspection path sequence and executes flight; when it arrives near the preset observation pose of a device, the precise pose control module starts the visual servo closed loop: it performs feature matching between the real-time image and the reference image, generates a pose error signal, and controls the UAV and gimbal to fine-tune until the error is lower than the threshold and triggers high-definition image acquisition.

[0122] The airborne edge processing module is invoked to call the equipment inspection operator preset for the corresponding device on the acquired image, and the real-time image is compared and analyzed with the reference image. Based on the comparison result and confidence level, a three-state decision is executed: if there is no difference, the system remains silent and logs are recorded; if a difference is detected, a high-definition sub-image of the difference area is extracted and metadata is encapsulated and sent back; if the analysis fails, a high-resolution image and alarm are sent back.

[0123] The foregoing has illustrated and described the basic features, principles, and advantages of the present invention. It should be noted that the present invention is not limited to the above embodiments, but only to some embodiments. Any improvements and additions made without departing from the spirit and scope of the present invention are considered to be within the scope of protection of the present invention.

Claims

1. A substation equipment inspection path planning system based on unmanned aerial vehicles (UAVs), characterized in that, include: The benchmark and feature management module is used to build and manage a three-dimensional digital benchmark and feature library for substation equipment. The feature library is associated with the geometric features of the equipment, the preset precise observation pose, and the corresponding acquired benchmark images. The intelligent operator management module is used to carry device detection operators and generate new operators based on the similarity of device features; it is also used to score the confidence of the corresponding device detection operators according to the current environmental conditions and the type of device to be inspected, and to prioritize all operators to be used based on the confidence scores. The precise pose control module receives the precise observed pose and adjusts the image acquisition point and acquisition angle of the UAV according to the precise observed pose. The airborne edge processing module is used to perform real-time analysis and processing on the real-time acquired images, and to perform a preliminary comparison between the processed images and the reference images through the device detection operator, and to make decisions and perform selective data back transmission based on the comparison results. The multi-objective fusion path planning module, based on the priority ranking result, selects the high-priority operator as the benchmark operator, identifies similar devices covered by the benchmark operator through parameter migration, and sets the devices covered by the benchmark operator and similar devices as the same inspection batch. Within the same inspection batch, it generates the optimal inspection sub-path for the inspection batch by minimizing the flight path and combining the comprehensive flight risk objective. It then traverses each priority operator and its corresponding batch to generate a global inspection path composed of each of the optimal inspection sub-paths.

2. The substation equipment inspection path planning system based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The benchmark and feature management module also includes: A 3D point cloud processing unit is used to construct a high-precision 3D point cloud model of a substation, and to perform semantic segmentation on the 3D point cloud model to identify and label each equipment instance therein. The observation pose construction unit, for each identified device instance, determines the final observation pose for image acquisition and the reference image based on the final observation pose, based on a preset observation pose construction strategy. The final observation pose includes three-dimensional spatial coordinates and acquisition viewpoint information.

3. The substation equipment inspection path planning system based on unmanned aerial vehicles (UAVs) according to claim 2, characterized in that, The observation pose construction strategy includes, for the identified equipment instance, identifying the key component area of ​​the equipment used for inspection based on its three-dimensional point cloud data, generating several candidate observation poses based on preset unobstructed line of sight conditions and safe working distance conditions, comprehensively evaluating each candidate observation pose based on multiple factors, including observation view quality, coverage integrity and environmental adaptability, and selecting the final observation pose by comprehensively considering the observation view quality, coverage integrity and environmental adaptability. The observation view quality is the angle between the observation axis pointing to the critical component area from the pose and the average normal direction of the surface of that area. The smaller the angle, the higher the view quality score. The coverage integrity is the proportion of the critical component area that the image frame taken in this pose can cover, calculated by simulating the camera's field of view. The higher the proportion, the higher the coverage integrity score. The environmental adaptability is a score that judges the electromagnetic interference intensity and background visual complexity at the location of this pose based on a pre-stored electromagnetic interference distribution map and illumination model.

4. The substation equipment inspection path planning system based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The intelligent operator management module includes: The operator library unit is used to store and manage device detection operators, as well as the baseline feature vectors of the device types corresponding to the device detection operators; The similarity calculation and transfer unit extracts the geometric and performance features of the new device instance and generates a feature vector. It performs similarity matching calculation between the feature vector and the benchmark feature vectors of various types of devices stored in the operator library unit. When the similarity is greater than a preset threshold, it uses the existing operator corresponding to the type of device as a pre-trained model, adjusts the parameters of the model through the sample data of the new device instance, and generates a similar device detection operator. The operator adjustment unit acquires light intensity, weather conditions, and background complexity as real-time environmental data, and dynamically adjusts the image preprocessing parameters of the called operators based on the mapping relationship between the environmental data and the image preprocessing parameter adjustment amount preset by the Bayesian optimization framework.

5. A substation equipment inspection path planning system based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The airborne edge processing module includes: The real-time analysis unit is used to call the device detection operator output by the intelligent operator management module, process the acquired real-time image, compare the processed image with the corresponding benchmark image obtained by the benchmark and feature management module, and output the comparison result containing information on the difference region and confidence level. The data feedback decision control unit is used to output a data feedback command based on the comparison result and a preset data feedback strategy.

6. A substation equipment inspection path planning system based on unmanned aerial vehicles (UAVs) according to claim 5, characterized in that, The data feedback strategy includes the following steps: if the comparison result indicates no significant difference and the confidence level is higher than the first threshold, a silent mode is triggered, image data feedback is not initiated, and only inspection logs are recorded locally; if the comparison result indicates a difference and the confidence level is higher than the second threshold, a difference precision transmission mode is triggered, the pixel coordinates of the difference region of the real-time analysis unit are received, and the corresponding difference region sub-image is cropped from the acquired original high-definition image. The difference region sub-image, device identifier, difference type, confidence level, and timestamp are encapsulated into source data and fed back to the data terminal; if the confidence level of the comparison result is lower than the third threshold, or an unrecognizable scene is detected, an abnormal degradation mode is triggered, and a high-resolution panoramic image and an abnormal alarm code are fed back to the data terminal.

7. A substation equipment inspection path planning system based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The multi-objective fusion path planning module includes receiving the precise observation pose set corresponding to all equipment within the inspection batch and obtaining the electromagnetic interference risk distribution map of the substation area. Each three-dimensional coordinate point in the electromagnetic interference risk distribution map is associated with a risk coefficient based on electromagnetic interference settings. Decision variables are defined, including the order of UAV access pose points and the path segments connecting each point. An optimization problem is established with the dual objectives of minimizing the total flight path length and minimizing the total exposure risk. The total flight path length is the sum of the Euclidean distances of all access segments. The total exposure risk is calculated by integrating the risk coefficient along the path, calculating the average risk coefficient of several uniformly sampled points of the access segment, and combining it with the segment length to obtain the risk cost of the access segment. The risk costs of all segments are calculated as the total exposure risk value. The optimization problem is solved using a non-dominated sorting genetic algorithm with an elite strategy, outputting a set of Pareto optimal solutions. Each solution corresponds to an access sequence and the corresponding total flight path length and total exposure risk value. The priority value of each solution is calculated comprehensively, the final solution is selected based on the priority value, and the access sequence and flight path corresponding to the solution are output as the optimal inspection sub-path.

8. A substation equipment inspection path planning system based on unmanned aerial vehicles (UAVs) according to claim 7, characterized in that, The multi-target fusion path planning module further includes, when generating the optimal inspection sub-path for an inspection batch, using the precise observation poses of all devices in the batch as necessary waypoints for path optimization, so that when performing inspections based on the optimal inspection sub-path, the airborne edge processing module can complete the image analysis of all devices on the path by relying only on the single reference operator.

9. A substation equipment inspection path planning system based on unmanned aerial vehicles (UAVs) according to claim 8, characterized in that, The multi-objective fusion path planning module further includes: obtaining the optimal inspection sub-path for each batch and sorting them according to the priority of the benchmark operator; for each pair of sequentially adjacent optimal inspection sub-paths, obtaining the planning endpoint of the previous sub-path and the planning starting point of the next sub-path, and planning a transfer segment connecting the two points through a path search algorithm; during planning, the generation of the transfer segment is determined by evaluating its flight length and the cumulative risk of crossing the electromagnetic interference zone; connecting the sorted optimal inspection sub-paths with the transfer segment to generate a global inspection path, and calculating the total flight distance and total risk integral of the global inspection path to verify whether it meets the total endurance constraint and global risk budget constraint of the UAV; if it meets the constraint, the global inspection path is output.

10. A substation equipment inspection path planning system based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The substation equipment inspection route planning system performs the following: Receive inspection task instructions and determine the target equipment set; call the reference and feature management module to obtain the three-dimensional digital model, preset precise observation pose and corresponding reference image of each device in the target equipment set; The intelligent operator management module is invoked. For the target device, its features are matched with the operator library for similarity. If the match is successful, it is directly associated with an existing operator. If the matching degree is higher than the migration threshold, a dedicated operator is generated based on similar operators through parameter migration. If the matching degree is low, it is marked as needing an independent operator. According to the current environmental parameters, the preprocessing parameters and decision thresholds of all associated operators are dynamically adjusted through a preset environment and parameter mapping relationship. Combining environmental conditions and device type, the expected confidence score is calculated for all associated operators and prioritized. The multi-objective fusion path planning module is invoked, and operators are sorted according to their priority. High-priority operators are selected as the baseline operators, and the equipment covered by them is grouped into the same inspection batch. For each inspection batch, with the goal of minimizing the combined cost of flight path length and electromagnetic exposure risk, the optimal inspection sub-path that traverses the observation poses of all equipment in the batch is generated. According to the batch priority order, transfer segments that also consider risk and efficiency are planned for adjacent sub-paths. All sub-paths and transfer segments are combined into a global inspection path sequence, and its compliance with endurance and safety constraints is verified. The UAV loads the global inspection path sequence and executes flight; when it arrives near the preset observation pose of a device, the precise pose control module starts the visual servo closed loop: it performs feature matching between the real-time image and the reference image, generates a pose error signal, and controls the UAV and gimbal to fine-tune until the error is lower than the threshold and triggers high-definition image acquisition. The airborne edge processing module is invoked to call the equipment inspection operator preset for the corresponding device on the acquired image, and the real-time image is compared and analyzed with the reference image. Based on the comparison result and confidence level, a three-state decision is executed: if there is no difference, the system remains silent and logs are recorded; if a difference is detected, a high-definition sub-image of the difference area is extracted and metadata is encapsulated and sent back; if the analysis fails, a high-resolution image and alarm are sent back.

Images