Standardized automatic driving operation method and system based on multi-source heterogeneous data cleaning

Abstract

The invention relates to the technical field of automatic driving, in particular to a standardized automatic driving operation method and system based on multi-source heterogeneous data cleaning, and the method comprises the following steps: S1, multi-source heterogeneous data collection and structuring, S2, layered differentiation cleaning, S3, standardized data fusion, 4, operation scene adaptation and application, and S5, dynamic feedback optimization. According to the standardized automatic driving operation method and system based on multi-source heterogeneous data cleaning, high-quality processing and efficient operation of automatic driving data are realized through hierarchical differentiation cleaning, standardized fusion of time-space unification, scene application and dynamic feedback optimization, the algorithm reliability and operation safety are improved, and the method and the system are suitable for popularization and application. The method is suitable for large-scale commercial operation of L4-level and above automatic driving vehicles (such as Robotaxi and automatic driving trucks), and data-driven safety monitoring, algorithm iteration and operation efficiency optimization are achieved.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and in particular to a standardized autonomous driving operation method and system based on multi-source heterogeneous data cleaning. Background Technology

[0002] The commercialization of autonomous driving technology relies on the efficient processing and application of massive amounts of multi-source data, but the following key issues exist in current operations:

[0003] 1. Data heterogeneity conflict: Significant differences in data format, spatiotemporal reference, and accuracy among vehicle sensors (cameras, LiDAR, millimeter-wave radar), positioning systems (GPS / IMU), vehicle bus (CAN), and roadside equipment (V2X) (e.g., the time synchronization error between LiDAR point clouds and camera images can reach hundreds of milliseconds) directly lead to difficulties in data fusion and affect the accuracy of environmental perception.

[0004] 2. Poor data quality: Sensors generate noisy data due to environmental interference such as rain, fog, and obstruction (e.g., outliers in point clouds, motion blur in images), occasional data packet loss or erroneous frames on the CAN bus, delays or false alarms in roadside V2X signals, and low-quality data that has not been cleaned can lead to misjudgments by autonomous driving algorithms (e.g., identifying shadows as obstacles).

[0005] 3. Lack of data standardization: The feature definitions of different operating scenarios (urban roads, highways, parking lots) are not uniform (such as the unit of obstacle speed and coordinate system benchmark), resulting in low model training sample reuse rate and low algorithm iteration efficiency.

[0006] 4. Operational loop disruption: The lack of a feedback mechanism linking data quality and operational effectiveness makes it impossible to reverse-engineer data processing procedures based on misjudgments in actual operations (such as delayed collision warnings), leading to the recurrence of similar problems.

[0007] To address the aforementioned issues, this invention proposes a standardized autonomous driving operation method and system based on multi-source heterogeneous data cleaning. Summary of the Invention

[0008] The main objective of this invention is to provide a standardized autonomous driving operation method and system based on multi-source heterogeneous data cleaning, which can effectively solve the problems in the background art.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A standardized autonomous driving operation method based on multi-source heterogeneous data cleaning includes the following steps:

[0011] S1. Acquisition and structuring of multi-source heterogeneous data:

[0012] Collect multi-source data from autonomous vehicles, including sensor data (camera images, LiDAR point clouds, millimeter-wave radar signals), positioning data (GPS / IMU fusion trajectory), vehicle bus data (vehicle speed and steering angle on the CAN bus), and roadside equipment data (V2X communication information).

[0013] The collected data is structured and the timestamps (accuracy ≤ 1ms) and spatial coordinate system (ENU coordinate system with the vehicle's center of mass as the origin) are unified to generate an original dataset containing data type, acquisition device ID, and spatiotemporal label.

[0014] S2, Layered Differentiated Cleaning:

[0015] Different cleaning strategies are used for different types of data:

[0016] Image data: Denoising is performed using adaptive thresholding (to remove rain and fog noise) and motion blur detection (discarding blurred frames with inter-frame displacement > 0.5m);

[0017] Point cloud data: Outlier points are removed based on radius filtering (points with less than 5 neighboring points are considered noise), and ground point segmentation is performed (the RANSAC algorithm is used to separate the road surface from obstacles).

[0018] Positioning data: Smooth trajectory jumps through Kalman filtering (removing outliers with position deviations > 1m) and time synchronization calibration (time difference with vehicle bus data ≤ 50ms);

[0019] Cross-source data conflict detection: When the obstacle distance deviation between the lidar and the camera is >0.3m, confidence-weighted fusion is initiated (weights are dynamically allocated based on device accuracy);

[0020] S3, Standardized Data Fusion:

[0021] A spatiotemporally unified feature library is constructed based on the cleaned data, including:

[0022] Static features: coordinates and attributes of road markings and traffic lights (generated through image semantic segmentation and point cloud matching);

[0023] Dynamic characteristics: the speed and acceleration of obstacles (fusion of radar and visual tracking results);

[0024] Vehicle status characteristics: normalized vehicle speed (0-100km / h mapped to 0-1) and steering angle (-45°~45° mapped to -1~1);

[0025] S4. Operational Scenario Adaptation and Application:

[0026] Standardized data is categorized according to operational scenarios (such as urban roads, highways, and parking lots) to generate scenario feature templates;

[0027] Template-based implementation:

[0028] Model training: Select typical scenario data for autonomous driving algorithm iteration (generate 1000+ training samples every 100km of travel);

[0029] Real-time monitoring: The system compares standardized features with preset thresholds (e.g., triggering an alert when the distance to an obstacle is less than 5m and the relative speed is greater than 10m / s).

[0030] S5, Dynamic Feedback Optimization:

[0031] Record cases of misjudgment during operation (such as false obstacle warnings), trace back the original data and cleaning process, and locate the source of deviation;

[0032] Adjust cleaning parameters (such as relaxing / tightening the outlier threshold in the point cloud) and standardization rules (such as optimizing dynamic feature weights) to improve data availability by ≥5% with each iteration.

[0033] Preferably, the layered differential cleaning in S2 further includes:

[0034] Vehicle bus data cleaning: CAN frame errors are eliminated through checksum verification (backup communication link is activated when the error rate is >1%), and missing frames are filled based on historical data prediction (linear interpolation is used when there are ≤3 consecutive missing frames).

[0035] Roadside equipment data cleaning: V2X messages with signal strength < -80dBm are filtered, and the truth value is determined by majority voting for multi-source reported information of the same traffic event (such as traffic light status).

[0036] Preferably, the standardized data fusion in S3 achieves spatiotemporal alignment in the following manner:

[0037] Time alignment: Based on the system time of the vehicle bus, interpolate / thin the sensor data (unify 10Hz for LiDAR and 20Hz for camera to 15Hz);

[0038] Spatial alignment: The image pixel coordinates and point cloud 3D coordinates are transformed to the vehicle coordinate system by calibration parameters (intrinsic parameters + extrinsic parameters), with an error ≤0.1m.

[0039] Preferably, the scene data selection for model training in S4 includes:

[0040] The scene complexity (including the number of obstacles, lane curvature, and types of traffic participants) is calculated using the entropy method, and high-complexity scenes (entropy > 0.8) are retained first.

[0041] Data from similar scenarios is deduplicated (based on a dynamic time warping algorithm, only one copy of samples with a similarity greater than 90% is retained) to reduce redundancy.

[0042] Preferably, the real-time monitoring in S4 further includes:

[0043] Establish standardized health indicators: sensor data integrity rate (≥95%), positioning accuracy (<1m), and communication latency (V2X message latency <100ms);

[0044] If any indicator fails to meet the standard for three consecutive sampling periods, the vehicle will be downgraded (e.g., from Level 4 autonomous driving to Level 2).

[0045] Preferably, the dynamic feedback optimization in S5 further includes:

[0046] Establish a correlation model between cleaning parameters and data quality (such as the mapping relationship between point cloud filtering radius and obstacle recognition accuracy);

[0047] Automatic parameter adjustment based on reinforcement learning: The reward function is "operational accident rate + data storage volume", and the optimization is performed once per quarter to reduce the accident rate by ≥10%.

[0048] A standardized autonomous driving operation system based on multi-source heterogeneous data cleaning includes:

[0049] Multi-source data acquisition module: includes vehicle sensor interface (camera, lidar, millimeter-wave radar), positioning module (GPS / IMU), CAN bus reader and roadside V2X receiver, used to simultaneously acquire multiple types of data;

[0050] Heterogeneous data cleaning module: including image cleaning unit (denoising, deblurring), point cloud cleaning unit (outlier removal, ground segmentation), positioning calibration unit (filtering, time synchronization) and cross-source conflict processing unit (confidence fusion);

[0051] Data standardization and fusion module: includes a spatiotemporal alignment unit (time interpolation, spatial coordinate transformation) and a feature normalization unit (static / dynamic / vehicle state feature standardization), outputting a scene feature library in a unified format;

[0052] The autonomous driving operation module includes a scene classification unit (divided by road type), a model training unit (sample generation and algorithm iteration), and a real-time monitoring unit (threshold comparison and early warning).

[0053] Feedback optimization module: Locates data deviation sources through the misjudgment case analysis unit, and coordinates with the parameter adjustment unit to optimize the cleaning strategy and standardization rules.

[0054] Preferably, the heterogeneous data cleaning module further includes:

[0055] Bus data verification unit: performs error detection and missing data filling on CAN frames, and supports backup link switching for SOME / IP protocol;

[0056] Roadside data filtering unit: Filters valid V2X messages based on signal strength and fuses multi-source event information through majority voting.

[0057] Preferably, the feature normalization unit of the data standardization and fusion module includes:

[0058] Static feature encoding subunit: converts road element attributes (such as road marking type and traffic light status) into One-Hot vectors;

[0059] Dynamic feature normalization sub-unit: obstacle velocity and acceleration are normalized using min-max (mapped to the 0-1 interval);

[0060] Vehicle state standardization sub-unit: converts physical quantities such as vehicle speed and steering angle into dimensionless characteristic values.

[0061] Preferably, the autonomous driving operation module further includes:

[0062] Health assessment unit: Calculates indicators such as sensor integrity and positioning accuracy in real time, triggering vehicle degradation mechanism;

[0063] Fleet Management Unit: Based on standardized data statistics, analyze the operational efficiency of each vehicle (such as the number of breakdowns per 100 kilometers) and optimize dispatch strategies.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] 1. Significantly improved data quality: Through layered and differentiated cleaning, the noise rate of multi-source data was reduced from 30% to below 5%, and the cross-source conflict error was reduced from 0.5m to 0.15m, providing high-quality data support for environmental perception.

[0066] 2. High degree of standardization: The unified spatiotemporal benchmark and feature format increase the cross-scene data reuse rate from 30% to 80%, improve the efficiency of model training sample generation by 3 times, and shorten the algorithm iteration cycle by 50%.

[0067] 3. Enhanced operational safety: Real-time monitoring and health assessment reduce the false judgment rate of autonomous driving systems by 40%, shorten accident response time to less than 100ms, and increase safety redundancy by 2 times in extreme scenarios.

[0068] 4. Closed-loop optimization capability: The dynamic feedback mechanism enables adaptive adjustment of data processing parameters, improving data availability by 5% per quarter and reducing the recurrence rate of similar failures from 20% to below 5%.

[0069] 5. Operational efficiency optimization: Standardized data supports unified fleet management, reducing operating costs per 100 kilometers by 15% and increasing equipment resource utilization by 25%. Attached Figure Description

[0070] Figure 1 This is a flowchart illustrating the overall method of the standardized autonomous driving operation method and system based on multi-source heterogeneous data cleaning, as presented in this invention.

[0071] Figure 2 This is a system framework diagram of the standardized autonomous driving operation method and system based on multi-source heterogeneous data cleaning, as presented in this invention.

[0072] Figure 3 This is a dynamic feedback optimization closed-loop diagram of the standardized autonomous driving operation method and system based on multi-source heterogeneous data cleaning, as presented in this invention. Detailed Implementation

[0073] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0074] like Figure 1 As shown, this invention provides a standardized autonomous driving operation method based on multi-source heterogeneous data cleaning, including the following steps:

[0075] S1. Multi-source heterogeneous data acquisition and structuring

[0076] Data acquisition: Data is collected synchronously via vehicle-mounted sensors (128-line LiDAR, 8-megapixel camera, 77GHz millimeter-wave radar), a combined positioning module (GPS+IMU, positioning accuracy 10cm), a CAN bus interface, and a roadside V2X communication unit.

[0077] Environmental perception data: LiDAR point cloud (10Hz), camera image (20Hz), millimeter-wave radar target list (50Hz);

[0078] Vehicle status data: vehicle speed, steering angle, and braking pressure (100Hz) via CAN bus;

[0079] Location and spatiotemporal data: GPS / IMU fused trajectory (50Hz), UTC timestamp;

[0080] Roadside cooperative data: V2X traffic light status, forward congestion warning (10Hz).

[0081] Structured processing: Add unique identifiers (device ID, vehicle ID) to all data, unify timestamps (based on the vehicle's master clock, with a synchronization error ≤1ms), transform the spatial coordinate system (unify to the ENU coordinate system with the vehicle's centroid as the origin), and generate a structured dataset containing "data type - acquisition device - spatiotemporal label - original value".

[0082] S2, Layered Differentiated Cleaning

[0083] Image data cleaning:

[0084] Adaptive threshold denoising: Identify rain and fog pixels based on dark channel prior algorithm, and remove noise through bilateral filtering (signal-to-noise ratio improvement ≥20dB);

[0085] Motion blur detection: Calculate the optical flow vectors of adjacent frames and discard blurry frames with inter-frame displacement > 0.5m (blurry frame rejection rate ≥ 95%).

[0086] Point cloud data cleaning:

[0087] Outlier removal: Radius filtering (radius 0.3m, less than 5 neighboring points are considered noise) is used to remove erroneous points from the lidar (noise point removal rate ≥90%).

[0088] Ground point segmentation: The ground plane is fitted using the RANSAC algorithm (plane error ≤ 5cm) to separate the point clouds of the road surface and obstacles.

[0089] Location and vehicle data cleaning:

[0090] Track smoothing: Extended Kalman filter (EKF) is used to process GPS jumps and remove outliers with position deviations >1m (track smoothness improved by ≥30%).

[0091] CAN data verification: Error frames are detected by CRC check (error frame recognition rate 100%). When ≤3 consecutive frames are missing, linear interpolation is used to fill in the gaps (filling error ≤5%).

[0092] Cross-source data conflict handling:

[0093] When the obstacle distance deviation between the lidar and the camera is greater than 0.3m, weights are dynamically allocated based on the device accuracy (lidar weight 0.7, camera weight 0.3) to perform weighted fusion, reducing collision error (deviation after fusion ≤ 0.15m).

[0094] S3, Standardized Data Fusion

[0095] Spatiotemporal alignment:

[0096] Time alignment: Interpolate / thin the data of different frequencies (unify to 15Hz) to ensure timestamp consistency (time difference ≤10ms);

[0097] Spatial alignment: The image pixel coordinates and point cloud 3D coordinates are transformed to the vehicle coordinate system through sensor calibration parameters (intrinsic parameter matrix, extrinsic parameter matrix) (conversion error ≤ 0.1m).

[0098] Feature standardization:

[0099] Static features: Road markings (type, coordinates) and traffic lights (location, status) are generated through image semantic segmentation and point cloud matching, and encoded as structured vectors;

[0100] Dynamic features: The speed and acceleration of obstacles (vehicles, pedestrians) are generated by fusing radar and visual tracking results and are standardized using z-score (mean 0, standard deviation 1);

[0101] Vehicle status characteristics: vehicle speed (0-100km / h mapped to 0-1) and steering angle (-45°~45° mapped to -1~1) are converted into dimensionless characteristic values.

[0102] Generate a scene feature library: Standardized features are stored in categories such as "urban roads - highways - parking lots", and retrieval is supported by scene complexity (obstacle density, lane curvature).

[0103] S4, Operational Scenario Adaptation and Application

[0104] Model training support:

[0105] Scene selection: The entropy method is used to calculate the scene complexity (entropy > 0.8 is a high complexity scene), and extreme scenes (such as rainstorm weather, construction sections) are retained first;

[0106] Sample deduplication: Based on the Dynamic Time Warping (DTW) algorithm, duplicate data of the same scene are deduplicated (only 1 copy of the sample with similarity > 90%), generating 1000+ effective training samples for every 100km journey.

[0107] Real-time operation monitoring:

[0108] Safety threshold monitoring: Set standardized feature thresholds (e.g., trigger a collision warning when the distance to an obstacle is <5m and the relative speed is >10m / s);

[0109] System health assessment: Real-time calculation of sensor data integrity rate (≥95%), positioning accuracy (<1m), and V2X communication latency (<100ms). An alarm is triggered if any indicator fails to meet the standard.

[0110] S5, Dynamic Feedback Optimization

[0111] Misjudgment Case Retrospective: Record misjudgment events during operation (such as false obstacle warnings, missed pedestrian detection), extract the original data and cleaning logs for the corresponding time period, and locate the source of deviation (such as unreasonable point cloud filtering parameters).

[0112] Parameter adaptive adjustment:

[0113] Establish a correlation model between cleaning parameters and data quality (such as the mapping relationship between point cloud filtering radius and obstacle recognition accuracy);

[0114] Parameter optimization based on reinforcement learning: The reward function is "reduction in operational accident rate + reduction in data storage". The parameters are automatically adjusted once per quarter to improve data availability by ≥5% and reduce the accident rate by ≥10%.

[0115] like Figures 2-3 As shown, this invention provides a standardized autonomous driving operation system based on multi-source heterogeneous data cleaning, used to implement the above method, including:

[0116] Multi-source data acquisition module

[0117] It includes a lidar interface (supporting ROS / Protobuf protocol), a camera data acquisition card (USB3.0 / HDMI), a millimeter-wave radar signal processor, a GPS / IMU combined positioning unit (supporting RTK), a CAN bus reader (compatible with SAE J1939 protocol), and a V2X communication module (supporting 802.11p).

[0118] Function: Synchronously collects multiple types of data and outputs raw data streams with spatiotemporal labels (data transmission bandwidth ≥ 1Gbps).

[0119] Heterogeneous data cleaning module

[0120] It includes an image cleaning unit (adaptive denoising, blur detection), a point cloud cleaning unit (outlier filtering, ground segmentation), a positioning calibration unit (EKF filtering, time synchronization), a bus verification unit (CRC error detection, missing data filling), and a cross-source conflict processing unit (confidence-weighted fusion).

[0121] Function: Executes dedicated cleaning strategies for different types of data, with the noise rate of the processed data ≤5% and the cross-source conflict rate ≤3%.

[0122] Data standardization and fusion module

[0123] It includes a spatiotemporal alignment unit (temporal interpolation / thinning, spatial coordinate transformation), a static feature encoding subunit (road element One-Hot encoding), a dynamic feature normalization subunit (obstacle parameter z-score standardization), and a vehicle state standardization subunit (physical quantity dimensionless transformation).

[0124] Function: Outputs a scene feature library with unified spatiotemporal structure and standardized format (feature dimension ≤ 512 dimensions, storage efficiency improved by ≥ 40%).

[0125] Autonomous driving operation module

[0126] It includes a scene classification unit (scene type division based on decision tree), a model training unit (sample generation, algorithm iteration interface), a real-time monitoring unit (threshold comparison, multi-level early warning), a health assessment unit (sensor / positioning / communication index calculation), and a fleet management unit (vehicle operation efficiency statistics and scheduling).

[0127] Functions: Supports algorithm training data supply, real-time safety monitoring (early warning response time ≤100ms) and fleet operation optimization.

[0128] Feedback optimization module

[0129] It includes a misjudgment case analysis unit (data backtracking, deviation source location), a parameter adjustment unit (parameter cleaning / standardized rule optimization), and a reinforcement learning engine (reward function calculation, policy update);

[0130] Function: Continuously optimize the data processing flow based on operational feedback, improving the overall system performance by ≥8% with each iteration.

[0131] Example: Robotaxi urban road operation

[0132] System Configuration:

[0133] Vehicle platform: Equipped with a 128-line lidar (detection range 200m), 6-channel 8-megapixel cameras (FOV 120°), 5-channel millimeter-wave radar, and GPS / IMU combined positioning (RTK mode);

[0134] Data collection: Each vehicle generates approximately 50GB of raw data per hour (30GB point cloud, 15GB image, and 5GB other data);

[0135] Operating area: 100km of the city's core area 2 It covers complex scenarios such as intersections, business districts, and schools.

[0136] Implementation of standardized autonomous driving operation methods based on multi-source heterogeneous data cleaning:

[0137] S1. Data Acquisition and Structuring: The multi-source data acquisition module synchronously acquires sensor and vehicle data, adds the vehicle ID "RT-001" and the timestamp "2024-08-20T14:30:00.000", and converts it to the ENU coordinate system.

[0138] S2, Layered Cleaning:

[0139] Image data: Remove rain and fog noise from heavy rain (signal-to-noise ratio increased from 15dB to 35dB), and remove blurred frames caused by rapid acceleration (the proportion of blurred frames decreased from 12% to 1%).

[0140] Point cloud data: outlier points were filtered (noise points were reduced from 20% to 3%), and obstacle point clouds were preserved after segmenting ground points;

[0141] Cross-source conflict: obstacle distances from LiDAR and camera are fused (original deviation 0.4m, fused deviation 0.1m).

[0142] S3, Standardization Integration:

[0143] Spatiotemporal alignment: unifying 10Hz point cloud and 20Hz image to 15Hz, with spatial transformation error ≤0.08m;

[0144] Feature standardization: The phrase "pedestrian 50m ahead, speed 3.6m / s" is converted into a standardized vector ([0.25, 0.32], corresponding to the normalized values ​​of distance and speed).

[0145] S4, Operational Applications:

[0146] Model training: Select the morning rush hour business district scene (entropy value 0.85) as high-value samples, and retain 2000 duplicates for the pedestrian detection algorithm iteration;

[0147] Real-time monitoring: When an obstacle is detected at a distance of 3m and a relative speed of 12m / s, a level one warning is triggered, and the vehicle automatically decelerates to 20km / h.

[0148] S5, Feedback Optimization:

[0149] By reviewing the case of "missed pedestrians crossing the road", it was found that the point cloud filtering radius was too large (0.5m), causing pedestrian points in the distance to be mistakenly deleted.

[0150] Adjusting the radius to 0.3m and verifying through reinforcement learning, the pedestrian detection recall rate improved by 12%, and it was included in the next round of system updates.

[0151] Implementation results:

[0152] Data processing: The average daily effective data volume per vehicle was compressed from 800GB to 300GB (storage efficiency improved by 62.5%), and data availability increased from 65% to 96%.

[0153] Operational safety: Collision warning accuracy improved from 70% to 95%, and the false alarm rate in extreme scenarios decreased by 60%;

[0154] Algorithm iteration: The training cycle for new scenario models has been shortened from 14 days to 5 days, and the cross-regional operation adaptation time has been reduced by 70%.

[0155] In summary, the standardized autonomous driving operation method and system based on multi-source heterogeneous data cleaning of the present invention collects multi-source data from vehicle sensors, positioning, bus, and roadside V2X, and performs structured processing; it employs a hierarchical and differentiated strategy to clean image, point cloud, and positioning data, resolving noise and cross-source conflicts; it generates a unified scene feature library through spatiotemporal alignment and feature standardization to support model training and real-time monitoring; and it combines dynamic feedback to optimize data processing parameters. This invention solves the problems of heterogeneity, low quality, and inconsistent standards of multi-source data, improving data availability to over 95% and reducing the accident rate by 40%, making it suitable for large-scale commercial operation of autonomous driving.

[0156] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A standardized automatic driving operation method based on multi-source heterogeneous data cleaning, characterized in that, Comprising the following steps: S1, multi-source heterogeneous data collection and structuring: Collecting multi-source data of autonomous vehicles, including sensor data, positioning data, vehicle bus data, and roadside device data; Structuring the collected data, unifying the time stamp and spatial coordinate system, and generating an original data set containing data types, collection device ID, and space-time labels; S2, hierarchical differential cleaning: Adopting exclusive cleaning strategies for different types of data: Image data: adaptive threshold denoising, motion blur detection; Point cloud data: removing outliers based on radius filtering, ground point segmentation; Positioning data: smoothing trajectory jumps through Kalman filtering, time synchronization calibration; Cross-source data conflict detection: when the obstacle distance deviation between laser radar and camera is >0.3m, start confidence weighted fusion; S3, standardized data fusion: Based on the cleaned data, a unified spatio-temporal feature library is constructed, including: Static features: coordinates and attributes of road markings and traffic signal lights; Dynamic features: speed and acceleration of obstacles; Vehicle state features: normalized speed and steering angle; S4, operation scenario adaptation and application: Classify standardized data by operation scenario to generate scenario feature templates; Based on the templates, realize: Model training: select typical scenario data for automatic driving algorithm iteration; Real-time monitoring: compare standardized features with preset thresholds; S5, dynamic feedback optimization: Record misjudgment cases during operation, backtrack original data and cleaning process, locate bias sources; Adjust cleaning parameters and standardization rules, increase data availability by ≥5% after each iteration. 2.The standardized autonomous driving operation method based on multi-source heterogeneous data cleaning according to claim 1, wherein: The hierarchical differential cleaning in S2 also includes: Vehicle bus data cleaning: remove CAN frame errors through checksum verification, and fill in missing frames based on historical data prediction; Roadside device data cleaning: filter V2X messages with signal strength < -80dBm, and use majority voting to determine the true value for multi-source reporting information of the same traffic event. 3.The standardized autonomous driving operation method based on multi-source heterogeneous data cleaning according to claim 1, wherein: The standardized data fusion in S3 realizes spatio-temporal alignment through the following methods: Time alignment: use the system time of the vehicle bus as the reference to interpolate / decimate sensor data; Space alignment: convert image pixel coordinates and point cloud three-dimensional coordinates to vehicle coordinate system through calibration parameters, with an error ≤0.1m. 4.The standardized autonomous driving operation method based on multi-source heterogeneous data cleaning according to claim 1, wherein: The scene data selection for model training in S4 includes: Use entropy method to calculate scene complexity, and preferentially retain high complexity scenes; Remove duplicate data for similar scenarios to reduce redundancy. 5.The standardized autonomous driving operation method based on multi-source heterogeneous data cleaning according to claim 1, wherein: The real-time monitoring in S4 also includes: Build standardized health indicators: sensor data completeness, positioning accuracy, and communication delay; When any indicator fails for 3 consecutive sampling periods, trigger vehicle degradation. 6.The standardized autonomous driving operation method based on multi-source heterogeneous data cleaning according to claim 1, wherein: The dynamic feedback optimization in S5 also includes: Establish a cleaning parameter-data quality association model; Based on reinforcement learning, automatically adjust parameters: use "operation accident rate + data storage volume" as the reward function, optimize once every quarter, and reduce the accident rate by ≥10%.

7. The standardized autonomous driving operation system based on multi-source heterogeneous data cleaning according to any one of claims 1-6, characterized in that: Comprise: Multi-source data collection module: contains vehicle-mounted sensor interface, positioning module, CAN bus reader, and roadside V2X receiver, used for synchronous collection of multiple types of data; Isomeric data cleaning module: including image cleaning unit, point cloud cleaning unit, positioning calibration unit and cross-source conflict processing unit; Data standardization fusion module: containing space-time alignment unit and feature normalization unit, output uniform format scene feature library; Automatic driving operation module: including scene classification unit, model training unit, real-time monitoring unit; Feedback optimization module: through misjudgment case analysis unit to locate data bias source, linkage parameter adjustment unit to optimize cleaning strategy and standardization rule. 8.The multi-source heterogeneous data cleaning based standardized automatic driving operation system of claim 7, wherein: The isomeric data cleaning module further comprises: Bus data verification unit: error detection and missing data filling for CAN frame, supporting SOME / IP protocol backup link switching; Roadside data filtering unit: based on signal strength to filter effective V2X message, through majority voting method to fuse multi-source event information. 9.The multi-source heterogeneous data cleaning based standardized automatic driving operation system of claim 7, wherein: The feature normalization unit of the data standardization fusion module Comprises: Static feature coding subunit: convert road element attribute into One-Hot vector; Dynamic feature normalization subunit: min-max standardization for obstacle speed and acceleration; Vehicle state standardization subunit: convert physical quantities such as vehicle speed and steering angle into dimensionless characteristic value. 10.The multi-source heterogeneous data cleaning based standardized autonomous driving operation system of claim 1, wherein: The automatic driving operation module further comprises: Health degree evaluation unit: real-time calculation of sensor integrity, positioning accuracy and other indicators, triggering vehicle degradation mechanism; Vehicle fleet management unit: based on standardized data to calculate the operation efficiency of each vehicle, optimize scheduling strategy.

Images