High-quality training data set generation method based on multi-source heterogeneous data fusion

Abstract

The invention belongs to the technical field of machine learning training, discloses a high-quality training data set generation method based on multi-source heterogeneous data fusion, and aims to solve the problems of single data source, non-uniform space-time reference, quality management and control deficiency and the like in the prior art. According to the method, multi-source data of an electric power unmanned vehicle sensor, a high-precision map of an inspection area, electric power V2X communication and the like are acquired and analyzed, and a training data set containing a structured truth value label and context metadata is generated through space-time reference alignment, cross-source deep fusion and quality conflict detection and repair. The method gives full play to the complementary advantage of multi-source data, guarantees the quality and reliability of a data set, improves the model inspection target recognition capability and scene adaptability, and meets the demands of intelligent inspection in the power industry for high-quality training data.

Description

Technical Field

[0001] This invention belongs to the field of machine learning training technology, specifically relating to a method for generating high-quality training datasets based on the fusion of multi-source heterogeneous data. Background Technology

[0002] With the accelerated intelligent transformation of the power industry, unmanned power vehicle inspection, with its advantages of high efficiency, safety, and all-weather operation, is gradually replacing traditional manual inspection and becoming a core means to ensure the safe and stable operation of transmission lines, power distribution equipment, and power protection zones. The core of unmanned power vehicle inspection relies on onboard target detection and tracking models. The detection accuracy and robustness of these models directly determine the reliability of the inspection task, and high-quality training datasets are a key prerequisite for improving model performance.

[0003] Currently, there are still many problems to be solved in the technology for generating training datasets for unmanned electric vehicle inspection models: The data sources are singular and lack heterogeneity adaptation: Existing datasets mostly rely on data collected by single sensors, making it difficult to cope with the challenges of complex inspection environments. For example, in rainy, foggy, or backlit scenarios, cameras are easily affected by environmental interference and cannot accurately identify minor defects such as tower cracks and hanging objects on lines. Furthermore, data from sensors such as LiDAR and millimeter-wave radar are not effectively integrated, resulting in incomplete data information and failing to provide comprehensive environmental perception for the model. At the same time, static data specific to power inspection scenarios (such as high-precision maps of the inspection area and line voltage level information) and dynamic data (such as the real-time status of intelligent power monitoring equipment and the trajectory of surrounding construction machinery) are not effectively correlated, and the data value is not fully exploited.

[0004] Inconsistent spatiotemporal benchmarks make data fusion difficult: The timestamp benchmarks and spatial coordinate systems of multi-source data from electric autonomous vehicles, such as onboard sensors, electric V2X communication equipment, and static maps, vary, leading to inaccurate matching of information from different data sources. For example, the location of a suspected defect detected by a sensor may deviate from the location of the tower marked on a high-precision map, making it impossible to accurately associate the defect with the corresponding equipment and affecting the effectiveness of the dataset.

[0005] Lack of data quality control and poor annotation reliability: Existing technologies lack automated quality conflict detection and repair mechanisms for power line inspection scenarios. Inconsistencies easily arise between multi-source data (such as discrepancies between manually annotated tower defects and sensor fusion results, and conflicts between vehicle trajectories within transmission line protection zones and safety regulations). These conflicts cannot be identified and processed in a timely manner, resulting in a large amount of noisy data in the dataset. At the same time, manual annotation is costly, inefficient, and easily affected by subjective factors, making it difficult to meet the annotation needs of massive inspection data.

[0006] The dataset lacks semantic information, resulting in weak model generalization ability: Existing datasets only contain simple target locations and category labels, lacking semantic information specific to power inspection scenarios (such as line voltage levels, tower numbers, power equipment operating status, and environmental constraints of the inspection area). During model training, the model cannot learn scenario logic related to power safety regulations, leading to insufficient reasoning ability for complex scenarios in actual inspection situations and limited generalization performance.

[0007] In summary, existing technologies struggle to generate training datasets that are comprehensive, reliable, and suitable for the complex inspection scenarios of unmanned power vehicles, thus hindering the improvement of inspection model performance and failing to meet the development needs of the power industry for high-precision and intelligent inspections. Therefore, developing a method for generating high-quality training datasets based on the fusion of multi-source heterogeneous data has become an urgent technical problem to be solved in the field of unmanned power vehicle inspections. Summary of the Invention

[0008] To overcome the above-mentioned technical problems, this invention provides a method for generating high-quality training datasets based on the fusion of multi-source heterogeneous data.

[0009] The present invention adopts the following technical solution: A method for generating high-quality training datasets based on multi-source heterogeneous data fusion includes the following steps: Acquire and parse multi-source heterogeneous data, which includes at least on-board sensor data of the unmanned electric vehicle, high-precision map data of the power inspection area, and electric vehicle network communication data. The multi-source heterogeneous data is aligned with a spatiotemporal reference to unify them into a common time coordinate system and world space coordinate system; Perform cross-source deep fusion on the aligned data to generate a fusion object containing multi-source attributes for the inspection targets in the scene, and fuse dynamic perception data with static inspection environment semantics; Perform data quality conflict detection and remediation, and identify and address inconsistencies and inconsistencies between different data sources; Based on the fused and repaired data, training samples containing structured truth labels and contextual metadata are generated, forming a high-quality training dataset.

[0010] Preferably, the spatiotemporal reference alignment step specifically includes: The timestamps of all data are calibrated using the Coordinated Universal Time (UTC) output by the onboard inertial measurement unit and the Global Positioning System of the electric unmanned vehicle as the global master clock. By using preset sensor extrinsic parameters, the data from each on-board sensor are unified into a coordinate system centered on the electric unmanned vehicle; By utilizing real-time high-precision positioning information, the coordinate system centered on the unmanned electric vehicle and the data therein are further transformed to the world spatial coordinate system defined by the high-precision map of the power inspection area.

[0011] Preferably, the step of generating a fusion object for the inspection target includes: The system correlates the target location data from lidar, the target category data from cameras, and the distance and relative motion data from millimeter-wave radar. The location and size of the inspection target are determined by lidar data, the category of the inspection target is determined by camera data, and the smooth moving speed of the electric unmanned vehicle relative to the inspection target is estimated by fusing millimeter-wave radar measurements and continuous position changes through a Kalman filter.

[0012] Preferably, the step of fusing dynamic sensing data with static inspection environment semantics includes: Match the location of the unmanned power vehicle or the inspection target with a high-precision map of the power inspection area to obtain static semantic information such as the line voltage level, tower type, and inspection path attributes of the inspection area. By combining information such as facility status broadcast by intelligent power monitoring equipment obtained from power vehicle network communication data, a complete inspection scenario including dynamic perception and static environment is constructed.

[0013] Preferably, the data quality conflict detection step is executed through a rule engine, which includes at least the following verification logic: Physical consistency verification is used to detect whether the state of the inspected target or the trajectory of the electric unmanned vehicle violates the laws of physics. Cross-source consistency verification is used to detect whether there are contradictions in the descriptions of the same inspection target from different data sources; Scenario rationality verification is used to detect whether the status of the inspection target or the inspection behavior of the unmanned vehicle is consistent with the power safety specifications or environmental constraints of the inspection area defined by the high-precision map of the power inspection area.

[0014] Preferably, the repair steps include: When a conflict is detected, the conflicting data is automatically corrected using a preset data source with higher confidence. For complex conflicts that cannot be automatically resolved, generate a visual report containing comparisons of data from multiple sources and submit it for manual review.

[0015] This invention also includes a high-quality training dataset generation system based on multi-source heterogeneous data fusion, comprising: A data acquisition module is used to acquire and parse onboard sensor data of the unmanned electric vehicle, high-precision map data of the power inspection area, and electric vehicle network communication data; A spatiotemporal alignment module is used to unify the multi-source heterogeneous data into a common time coordinate system and world space coordinate system; A data fusion module is used to generate fusion objects for inspection targets and to fuse dynamic perception data with static inspection environment semantics; A quality control module is used to perform conflict detection and repair for data quality issues; A dataset generation module is used to generate the final training dataset based on the processed data.

[0016] Preferably, the data fusion module further includes: A target association submodule is used to spatially associate data about the same inspection target from different sensors; A state estimation submodule with a built-in Kalman filter is used to fuse multi-source information to output a smooth and stable motion state of the electric unmanned vehicle relative to the inspection target.

[0017] Preferably, the quality control module further includes: A rules engine configured with physical consistency, cross-source consistency, and scenario rationality verification rules; An automatic repair submodule is used to correct data conflicts according to a preset confidence level strategy; A report generation unit for generating visual reports on conflicts that require human intervention.

[0018] Preferably, the training samples generated by the dataset generation module include a structured ground truth label, which records the following information about the inspection target in detail: High-precision three-dimensional position and dimensions guaranteed by lidar data; Precise categories identified by the camera; The smoothed moving speed and acceleration of the electric unmanned vehicle relative to the inspection target, calculated through multi-source fusion. And contextual metadata related to the target, including the line voltage level, tower number, distribution of surrounding power facilities, and real-time status of power intelligent monitoring equipment in the current inspection area.

[0019] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention integrates heterogeneous data from multiple sources, including onboard sensors of unmanned power vehicles, high-precision maps of inspection areas, power V2X communication data, and manually labeled data. Through deep cross-source fusion, it fully leverages the complementary advantages of each data source. For example, it utilizes the high-precision positioning capabilities of lidar to compensate for the recognition deficiencies of cameras in adverse weather conditions, and combines power V2X data to supplement the real-time operating status of power equipment. This results in a dataset that not only includes the location and category information of targets but also integrates multi-dimensional attribute data, providing a more comprehensive and accurate perceptual basis for model training and improving the model's accuracy in identifying inspection targets such as tower defects, foreign objects on lines, and illegal construction.

[0020] 2. This invention uses the UTC time output by the IMU / GPS of the electric unmanned vehicle as the global master clock and achieves timestamp calibration of all data through technologies such as PTP. Simultaneously, through sensor extrinsic parameter calibration and mapping to the world coordinate system, it achieves precise spatial alignment of onboard sensor data, static map data, and V2X data. This solves the problem of inconsistent spatiotemporal references for multi-source data in existing technologies, ensuring accurate correlation of information from different data sources and providing a reliable data foundation for subsequent deep fusion and quality inspection.

[0021] 3. This invention implements three types of verification—physical consistency, cross-source consistency, and scenario rationality—through a rule engine. It can automatically identify contradictions and unreasonable information in the data (abnormal movement trajectories of unmanned vehicles, conflicts between defect annotations and sensor data, and illegal trajectories within power transmission line protection zones, etc.). It also handles conflicts by combining confidence-priority automatic repair with manual review, effectively eliminating noisy data and correcting erroneous information. At the same time, it uncovers difficult samples in complex inspection scenarios, improving the quality and reliability of the dataset and providing solid data support for model training.

[0022] 4. This invention integrates dynamically perceived data with the semantics of the static inspection environment, associating target information with semantic information such as line voltage levels, tower types, and power safety regulations in the inspection area. This results in a dataset containing complete inspection scenario logic. During model training, the model learns the unique constraints and inference rules of the power inspection scenario, improving its understanding and decision-making ability in complex inspection situations and enhancing its generalization performance in different inspection scenarios such as mountain lines, urban distribution networks, and substation perimeters.

[0023] 5. The multi-source data parsing, spatiotemporal alignment, fusion, quality detection and repair, and dataset partitioning of this invention are all automated, reducing the cost of manual annotation and data processing and improving the efficiency of dataset generation. Simultaneously, the output dataset conforms to industry standard formats and comes with detailed documentation, possessing good interpretability and ease of use. It can be directly used for training different types of unmanned power vehicle inspection models, adapting to the massive and high-frequency inspection data processing needs of the power industry, and promoting the overall improvement of the intelligence level of power inspection. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0025] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. Unless otherwise specified, the raw materials and equipment used can be purchased from the market or are commonly used in the art. The methods in the embodiments, unless otherwise specified, are conventional methods in the art. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0026] This embodiment will elaborate on the specific implementation process of a high-quality training dataset generation method based on multi-source heterogeneous data fusion. It aims to construct a high-quality training dataset for a target detection and tracking model in power grid inspection scenarios for an unmanned power vehicle inspection system.

[0027] The implementation process of this method is interconnected, starting with the extensive acquisition and parsing of data, followed by strict spatiotemporal benchmark alignment, then entering the core cross-source data deep fusion and feature enhancement stage, and then ensuring data purity through intelligent quality conflict detection and repair, finally generating a finished dataset that can be directly used for model training.

[0028] Step 1: Acquisition and preliminary analysis of multi-source heterogeneous data The first step of the method is to collect raw data from diverse sources and perform preliminary structured processing. In this embodiment, we integrate four types of key data.

[0029] First, the core data originates from a comprehensive sensor suite deployed on the unmanned power line inspection vehicle. This is a high-frequency, time-series data set, including LiDAR, high-definition cameras, and millimeter-wave radar for instantaneous environmental sensing, as well as inertial measurement units and IMU / GPS for recording the vehicle's own motion. These sensors record data precisely with nanosecond-level synchronization timestamps. We developed a dedicated parsing program to decode the raw data in different formats, such as the 3D point cloud from LiDAR and video streams from cameras, into a unified intermediate structure, preparing it for subsequent processing.

[0030] Secondly, we introduce static geospatial data, namely a high-precision map of the power inspection area. The map, in vector format, details relevant geographical elements for power inspection with centimeter-level precision, such as the location of power poles, transmission line routes, inspection paths, insulator distribution areas, and even terrain slope and curvature. By constructing a spatial index, we can quickly query the static inspection environment information surrounding the unmanned vehicle based on its real-time location.

[0031] Secondly, there is V2X communication data. This is event-driven data, such as power facility status information broadcast by intelligent monitoring equipment within the power inspection area (e.g., tower tilt warnings, line temperature data), or safety messages such as location and work progress shared by surrounding collaborative inspection equipment through vehicle-to-vehicle communication. The onboard communication unit of the unmanned vehicle attaches a precise timestamp when receiving this information.

[0032] Finally, we also utilized some manually annotated ground truth data. This data was created by a professional annotation team on some of the collected images and point clouds. They meticulously drew 3D bounding boxes for targets such as defects in power poles, foreign objects on lines, and illegal construction sites, and stored them as annotation files that could be accurately associated with the original data frames.

[0033] Step 2: Spacetime Reference Alignment and Standardization All data must be described under a unified spatiotemporal reference, which is a prerequisite for effective integration.

[0034] In terms of time, we use Coordinated Universal Time (UTC) output from the vehicle's IMU / GPS as the global master clock. Timestamps from all other sensors and V2X data are calibrated using technologies such as Precision Time Protocol (PTP) to ensure that all events have a unified and millisecond-accurate time coordinate.

[0035] In the spatial dimension, we established a dual alignment mechanism. First, through a pre-calibrated extrinsic parameter matrix, we transformed all sensor data into a coordinate system centered on the autonomous vehicle itself, achieving unification of all components within the vehicle. This transformation process satisfies formula (1): in, The raw 3D coordinates collected by sensors such as lidar and cameras; The rotation matrix of the sensor relative to the center of the electric autonomous vehicle. These are translation vectors (all obtained through offline calibration). These are the target coordinates in the transformed autonomous vehicle coordinate system.

[0036] Next, using the high-precision real-time positioning information provided by IMU / GPS, we further transform this coordinate system centered on the unmanned vehicle, along with all the data within it, to the world coordinate system used by the high-precision map. The transformation formula is shown in (2): In the formula, This is the real-time rotation matrix of the autonomous vehicle relative to the world coordinate system. This is the real-time translation vector (calculated from the latitude, longitude, and attitude angles output by the IMU / GPS). These are the coordinates in the final world coordinate system.

[0037] Thus, both dynamic perception data from autonomous vehicles and static environmental data from maps have been aligned within the same grand spatial framework.

[0038] Step 3: Deep Fusion and Feature Enhancement of Cross-Source Data Next, we move on to the core aspect of this invention: deep fusion and feature enhancement of cross-source data. This step aims to leverage the unique advantages of different data sources to create features with richer information dimensions and more accurate content.

[0039] Imagine a scenario: on a rainy day, a camera might struggle to detect defects near distant power poles, while millimeter-wave radar can reliably detect them. This method leverages this complementarity. First, within a unified global coordinate system, we establish a preliminary association between the pole / line point clusters from the lidar, the camera's detection frame, and the target point from the millimeter-wave radar based on spatial proximity.

[0040] Once a target is simultaneously captured by multiple sensors, we create a fused dynamic / static inspection object. The object's position and size are primarily determined using the most accurate LiDAR data; its dynamic parameters (such as the inspection vehicle's speed relative to the tower) are smoothly combined with direct measurements from millimeter-wave radar and positional changes from continuous point cloud data using a Kalman filter. This is achieved through the following formula: 1. Equation of State: In the formula, The state vector contains the planar position of the unmanned vehicle relative to the inspection target (such as tower defects or foreign objects on the line). ) and speed ( ); The state transition matrix ( , (data sampling interval); For the control matrix (here) (No additional control input) The noise is a process noise that follows a Gaussian distribution. ( (This is the process noise covariance matrix).

[0041] 2. Observation equation: In the formula, The observation vector contains the position measured by the lidar ( ) and the speed measured by millimeter-wave radar ( ); For the observation matrix ( ); To observe the noise, it follows a Gaussian distribution. ( The noise covariance matrix is ​​determined by sensor accuracy calibration to observe it.

[0042] 3. Filter update: in, This is the predicted state value from the previous moment. To predict the covariance matrix, It is the identity matrix. Observation matrix transpose, This represents finding the inverse of a matrix.

[0043] The above formula yields a more stable and accurate velocity estimate than that from a single source; its precise category, such as tower cracks, line hanging objects, and construction machinery, is determined by image recognition results. If the target also broadcasts its own information via V2X (such as tower temperature data uploaded by smart monitoring devices), we can supplement this information as additional ground truth features to the object.

[0044] Simultaneously, we integrate dynamic perception data with static inspection environment semantics. By matching the location of the unmanned vehicle to a high-precision power inspection map, we can instantly obtain information such as the voltage level of the lines, tower types, and key inspection areas for the current inspection section. Combined with real-time status data of power facilities transmitted via V2X (such as line load data for a certain section), the system can construct a complete inspection scenario map: at what speed the unmanned vehicle is traveling on which inspection section, and whether the monitoring data of the towers ahead is normal. This fused environmental semantics provides crucial background for understanding the execution of inspection tasks and target identification.

[0045] Step 4: Data Quality Conflict Detection and Repair To ensure the "high quality" of the final dataset, we designed an automated data quality conflict detection and repair mechanism.

[0046] This mechanism includes a rules engine that performs multi-layered checks. The first layer is physical consistency verification; for example, the system detects whether the autonomous vehicle's trajectory exhibits any unreasonable, sudden changes in speed, and if so, marks it as suspicious. In the formula, The speed of the unmanned vehicle relative to the inspection target at the current moment. The velocity at the previous moment, Preset physical thresholds (set according to safety specifications for power grid inspection scenarios, such as urban distribution network inspection scenarios). ).

[0047] The second layer is cross-source consistency verification. For example, if manual annotations show that a certain tower has a crack defect, but the fused image recognition results do not detect it, the system will mark this annotation as pending review. The third layer is scenario reasonableness verification. For example, within the transmission line protection zone indicated by the high-precision map, the fused data shows the presence of moving vehicle tracks, which may indicate illegal construction activities or potential problems with the positioning or map data.

[0048] When a conflict is detected, the system attempts to automatically correct it using data sources with higher confidence. For example, if a camera misses a foreign object in a circuit due to backlighting, but the LiDAR clearly outlines its contour, the system can generate an inferred bounding box on the image based on the point cloud. For complex conflicts that cannot be automatically resolved, the system generates a visual report containing comparisons of data from multiple sources and submits it to a human reviewer for final decision. This process not only corrects errors but also uncovers difficult samples that pose a significant challenge to model training.

[0049] Step 5: Generation and labeling of high-quality training datasets After the above processing steps, we finally generate a high-quality dataset that can be directly used for model training.

[0050] We divide the continuous data stream into a series of training samples at a fixed frequency. Each sample contains input data at a specific time, such as LiDAR point clouds and image sequences, as well as a structured ground truth label.

[0051] This label provides a detailed description of the status of all targets in the inspection scenario. This status information is the culmination of the deep integration and repair processes described above. It includes high-precision three-dimensional position and size guaranteed by LiDAR, accurate categories identified by cameras (such as tower defect types and line foreign object types), and relevant parameters that integrate multi-source information (such as unmanned vehicle inspection speed and the distance between the target and the line).

[0052] In addition, each sample is accompanied by rich contextual metadata, recording the weather and lighting conditions at the time, as well as the fused inspection scene semantics, such as the line grade, pole number, and distribution of surrounding power facilities in the current inspection area. We also score the overall quality of each sample based on the collision detection results.

[0053] Finally, we will take into account all the generated samples according to the inspection scenario type (such as mountain lines, urban distribution networks, substation surroundings), weather conditions, and target density (such as the number of defective targets and the density of foreign objects), and divide them into training set, validation set, and test set according to a preset ratio.

[0054] Compared to traditional methods that rely on a single data source or simple data patching, the dataset generated by the present invention through the above embodiments exhibits significant advantages in multiple dimensions. First, the accuracy and richness of the labels are greatly improved, providing in-depth information on power line inspections that transcends single-sensory perception. Second, the high consistency and reliability of the data are guaranteed, providing a solid foundation for training the unmanned power line inspection model. More importantly, the dataset contains complete semantics of power line inspection scenarios, enabling the model to learn to reason and make decisions under real inspection environments and power safety regulations. Finally, the entire generation process is highly automated and easily scalable, making it possible to continuously produce massive amounts of high-quality data.

[0055] Although embodiments of the present invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to the above embodiments without departing from the principles and spirit of the present invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for generating high-quality training datasets based on multi-source heterogeneous data fusion, characterized in that, Includes the following steps: Acquire and parse multi-source heterogeneous data, including at least on-board sensor data of the unmanned electric vehicle, high-precision map data of the power inspection area, and electric vehicle network communication data; The multi-source heterogeneous data is aligned with a spatiotemporal reference to unify them into a common time coordinate system and world space coordinate system; Perform cross-source deep fusion on the aligned data to generate a fusion object containing multi-source attributes for the inspection targets in the scene, and fuse dynamic perception data with static inspection environment semantics; Perform data quality conflict detection and remediation, and identify and address inconsistencies and inconsistencies between different data sources; Based on the fused and repaired data, training samples containing structured truth labels and contextual metadata are generated, forming a high-quality training dataset.

2. The method according to claim 1, characterized in that, The spatiotemporal reference alignment step specifically includes: The timestamps of all data are calibrated using the Coordinated Universal Time (UTC) output by the onboard inertial measurement unit and the Global Positioning System of the electric unmanned vehicle as the global master clock. By using preset sensor extrinsic parameters, the data from each on-board sensor are unified into a coordinate system centered on the electric unmanned vehicle; By utilizing real-time high-precision positioning information, the coordinate system centered on the unmanned electric vehicle and the data therein are further transformed to the world spatial coordinate system defined by the high-precision map of the power inspection area.

3. The method according to claim 1, characterized in that, The steps for generating a fusion object for the inspection target include: The system correlates the target location data from lidar, the target category data from cameras, and the distance and relative motion data from millimeter-wave radar. The location and size of the inspection target are determined by lidar data, the category of the inspection target is determined by camera data, and the smooth moving speed of the electric unmanned vehicle relative to the inspection target is estimated by fusing millimeter-wave radar measurements and continuous position changes through a Kalman filter.

4. The method according to claim 1 or 3, characterized in that, The steps for fusing dynamic sensing data and static inspection environment semantics include: Match the location of the unmanned power vehicle or the inspection target with a high-precision map of the power inspection area to obtain static semantic information such as the line voltage level, tower type, and inspection path attributes of the inspection area. By combining information such as facility status broadcast by intelligent power monitoring equipment obtained from power vehicle network communication data, a complete inspection scenario including dynamic perception and static environment is constructed.

5. The method according to claim 1, characterized in that, The data quality conflict detection step is executed through a rule engine, which includes at least the following verification logic: Physical consistency verification is used to detect whether the state of the inspected target or the trajectory of the electric unmanned vehicle violates the laws of physics. Cross-source consistency verification is used to detect whether there are contradictions in the descriptions of the same inspection target from different data sources; Scenario rationality verification is used to detect whether the status of the inspection target or the inspection behavior of the unmanned vehicle is consistent with the power safety specifications or environmental constraints of the inspection area defined by the high-precision map of the power inspection area.

6. The method according to claim 5, characterized in that, The repair steps include: When a conflict is detected, the conflicting data is automatically corrected using a preset data source with higher confidence. For complex conflicts that cannot be automatically resolved, generate a visual report containing comparisons of data from multiple sources and submit it for manual review.

7. A high-quality training dataset generation system based on multi-source heterogeneous data fusion, characterized in that, include: A data acquisition module is used to acquire and parse onboard sensor data of the unmanned electric vehicle, high-precision map data of the power inspection area, and electric vehicle network communication data; A spatiotemporal alignment module is used to unify the multi-source heterogeneous data into a common time coordinate system and world space coordinate system; A data fusion module is used to generate fusion objects for inspection targets and to fuse dynamic perception data with static inspection environment semantics; A quality control module is used to perform conflict detection and repair for data quality issues; A dataset generation module is used to generate the final training dataset based on the processed data.

8. The system according to claim 7, characterized in that, The data fusion module further includes: A target association submodule is used to spatially associate data about the same inspection target from different sensors; A state estimation submodule with a built-in Kalman filter is used to fuse multi-source information to output a smooth and stable motion state of the electric unmanned vehicle relative to the inspection target.

9. The system according to claim 7, characterized in that, The quality control module further includes: A rules engine configured with physical consistency, cross-source consistency, and scenario rationality verification rules; An automatic repair submodule is used to correct data conflicts according to a preset confidence level strategy; A report generation unit for generating visual reports on conflicts that require human intervention.

10. The system according to claim 7, characterized in that, The training samples generated by the dataset generation module include a structured ground truth label, which records the following information about the inspection target in detail: High-precision three-dimensional position and dimensions guaranteed by lidar data; Precise categories identified by the camera; The smoothed moving speed and acceleration of the electric unmanned vehicle relative to the inspection target, calculated through multi-source fusion. And contextual metadata related to the target, including the line voltage level, tower number, distribution of surrounding power facilities, and real-time status of power intelligent monitoring equipment in the current inspection area.

Images