A method and system for robot training data governance based on data lineage constraints

By constructing a data lineage graph and indivisible lineage groups, the problems of identifying homologous samples and version state dispersion in robot training data management are solved, enabling effective state management and contamination detection of data usage interfaces, and improving the clarity and accuracy of data management.

CN123115118APending Publication Date: 2026-11-17HEFEI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611457627.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-18
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing robot training data management, homologous derived samples and related version states are recorded in a scattered manner, the boundaries of the purpose of the data set are unclear, and it is difficult to uniformly identify near-duplicate or homologous samples across sets. After the data license calibration tool scenario or segmentation rules change, the reference relationship between the old segmentation list and the old verification record and the downstream training release lacks clear constraints.

Method used

By establishing data lineage constraints and constructing a data lineage graph, the source job segment identifier, source relationship and version relationship between data lineage nodes are recorded to form indivisible lineage groups. Based on the lineage groups, the data is allocated to the training set, validation set, calibration set and independent real machine test set, generating pollution reports and data release tokens, and realizing effective state management of the data usage interface.

Benefits of technology

It implements constraint management of training data source relationships, version status, set ownership, and validation records, ensuring that the reading conditions of the data usage interface correspond to the valid status of the model validation records. When versions are inconsistent, it locates the affected objects and blocks the reading of old data or old validation evidence, thus improving the clarity and accuracy of data management.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115118A_ABST
    Figure CN123115118A_ABST
Patent Text Reader

Abstract

This invention relates to the field of artificial intelligence training data management technology, specifically to a robot training data governance method and system based on data lineage constraints. The method includes the following steps: receiving raw job segment data and version events triggered by the robot controller; establishing a data lineage graph recording source job segment identifiers, source relationships, and version relationships; forming indivisible lineage groups based on hard and soft association risks; allocating these indivisible lineage groups to the training set, validation set, calibration set, and independent real-machine test set; generating a contamination report, a frozen segmentation list, a data release token, and a model verification record; and revoking the data release token and blocking data reading when the current version binding information is inconsistent with the token binding information. This invention establishes a correspondence between same-source data segmentation, verification record status, and reading conditions through lineage grouping, contamination detection, token verification, and invalidation marking.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence training data management technology, and in particular to a robot training data governance method and system based on data lineage constraints. Background Technology

[0002] The field of artificial intelligence training data management technology mainly revolves around the collection, storage, annotation, verification, segmentation, and retrieval management of model training samples, label records, version information, and usage records, involving data interaction between data collection devices, annotation terminals, training servers, and databases.

[0003] Among them, robot training data governance methods and systems refer to the technical processing methods for managing training samples and their derived samples generated during robot operations. Typically, the robot controller and related sensing devices collect operation process data, write it into a database via an industrial network or data interface, and then the server collects, verifies, indexes, segments, and reads the data according to the data source, collection time, device identification, labeling status, version record, and task purpose, so as to serve as the data management basis for the robot model training, verification, and testing stages.

[0004] Existing robot training data management often involves collection, verification, and segmentation based on file samples, task batches, or data versions. After continuous operation segments undergo windowed resampling, enhanced annotation, revision, and fault replay, the status of derived samples and related versions is recorded in a scattered manner. The boundaries of the data set's purpose are unclear, and the usage status of calibration data and real machine test data is difficult to separate. Nearly duplicate or homogeneous samples across sets are difficult to identify uniformly. After changes in data license calibration tools, scenarios, or segmentation rules, the reference relationship between the old segmentation list and old verification records and downstream training releases lacks clear constraints. Summary of the Invention

[0005] The purpose of this invention is to provide a robot training data governance method and system based on data lineage constraints. It establishes a continuous governance relationship around the original job fragment data, derived data, version events, dataset segmentation, pollution detection, release reading and failure rollback in robot training data, so that the source relationship, version status, set ownership, verification record and downstream reading conditions of training data can be constrained and managed in the same technical chain.

[0006] To achieve the above objectives, this invention provides a robot training data governance method based on data lineage constraints, comprising the following steps: Receive raw job segment data and version events triggered by the robot controller, assign source job segment identifiers to the raw job segment data, and obtain derived data formed by the derivation process; Establish data lineage nodes for the original job fragment data and derived data, record the source job fragment identifier, the source relationship between data lineage nodes and the version relationship corresponding to version events, and construct a data lineage graph; Based on the hard and soft association risks in the data lineage graph, merge or mutual exclusion constraints are applied to data lineage nodes to form indivisible lineage groups. The data is assigned to the training set, validation set, calibration set and independent real machine test set using indivisible bloodline groups as the smallest unit. Bloodline group assignment results are generated, and cross-set contamination detection is performed on the data of different sets to generate contamination reports. A frozen segmentation list is generated based on the bloodline group allocation results, contamination reports, and version relationships. A data release token is issued when the data quality, data permission, and contamination threshold meet the preset conditions. The data release token is bound to the frozen segmentation list and version binding information. When the data release token is valid, the data is allowed to use the interface to read data or verify evidence, and model verification records are generated based on the frozen segmentation list, model summary, threshold summary and robot version tuple; When the current version binding information is inconsistent with the version binding information bound to the data release token, locate the affected data lineage nodes, indivisible lineage groups, frozen split lists and model verification records along the data lineage graph and reference relationships, revoke the data release token, mark the model verification records as invalid, and block the data usage interface from reading old data or old verification evidence.

[0007] Furthermore, the original job segment data includes at least two of the following: images, depth maps, point clouds, robot joint states, end-effector poses, force or torque sequences, workstation events, and task logs, all triggered by the robot controller during a continuous task, subtask, contact phase, or fault phase. Version events include at least two of the following: sensor calibration version, tool or fixture identifier, workpiece batch identifier, scene map version, safety rule version, and associated robot version. Derived data includes data formed from the original job segment data through windowing, resampling, enhancement, synthesis, annotation revision, or fault playback processing.

[0008] Furthermore, when receiving raw work segment data and version events triggered by the robot controller, the trigger timestamp generated by the robot controller at the start of the task, end of the task, contact establishment, contact release, or fault reset is obtained, and the trigger timestamp is matched with the collection time range of the raw work segment data. When the trigger timestamp falls within the collection time range or meets the synchronization error condition with the boundary of the collection time range, the corresponding version event is written to the source job segment identifier of the original job segment data. When the trigger timestamp and the collection time range do not meet the synchronization error condition, stop writing the corresponding version event to the source job segment identifier; The process of writing version events includes reading the sensor calibration version, tool or fixture identifier, workpiece batch identifier, scene map version, and safety rule version. The reading results are combined with the trigger timestamp, acquisition time range, and source operation segment identifier to form a version record. When the same source operation segment identifier corresponds to multiple version records, the valid version record is selected according to the time matching order of the trigger timestamp and acquisition time range, and the unselected version records are marked as pending review.

[0009] Furthermore, the data lineage nodes include original job fragment nodes and derived nodes. The original job fragment node is written with the source job fragment identifier, collection time range, robot identifier, sensor identifier, content summary and version relationship. The derived node is written with the parent node identifier, derived processing type, processing parameter summary and content summary. When the parent node identifier of a derived node does not match an existing node in the data lineage graph, stop writing the derived node into the data lineage graph and output a node relationship pending confirmation status.

[0010] Furthermore, when performing merge or mutual exclusion constraints on data lineage nodes, hard association relationships are determined based on whether data lineage nodes have the same source job segment identifier, parent-child source relationship, same fault playback source, same synthesis processing seed, or same annotation revision chain. When two data lineage nodes satisfy a hard association relationship, the two data lineage nodes are merged into the same initial lineage group; Calculate the soft association risk between different initial lineage groups. When the soft association risk meets the merging condition, merge the corresponding initial lineage groups. When the soft association risk does not meet the merging condition but meets the mutual exclusion condition, establish mutual exclusion constraints for the corresponding initial lineage groups to form indivisible lineage groups. When calculating the soft association risk, the workpiece batch identifier, scene map version, tool or fixture identifier, collection time range, trajectory features, visual features and sensor calibration version are extracted from the data lineage nodes corresponding to the two initial lineage groups. These are used to form batch similarity, scene similarity, tool similarity, time proximity, trajectory similarity, visual similarity and calibration similarity, and then normalized and weighted to obtain the soft association risk.

[0011] Furthermore, when generating lineage group allocation results and contamination reports, set coverage constraints are established for indivisible lineage groups based on at least two of the following: task tags, success / failure status, fault type, object domain, tool domain, scenario domain, calibration domain, and risk level. Under the conditions of satisfying set coverage constraints and mutual exclusion constraints, the indivisible lineage groups are assigned to the training set, validation set, calibration set and independent real machine test set; Perform at least two of the following detection methods on data from different sets: content duplication detection, visual or point cloud near duplication detection, trajectory or force sequence similarity detection, object or scene overlap detection, and metadata homology detection. Write the associated nodes, source paths, detection scores, judgment thresholds, and disposal status into the pollution report.

[0012] Furthermore, the version binding information includes data content summary, sensor calibration version, tool or fixture identifier, scene map version, data license status, segmentation rule version, and associated robot version; Before issuing a data release token, data quality, data permission, and pollution threshold are checked sequentially. If any check result fails to meet the preset conditions, the issuance of the data release token is stopped. When the current version binding information is inconsistent with the version binding information, the affected scope is determined according to the source relationship in the data lineage diagram and the set affiliation in the frozen segmentation list, and the model verification record corresponding to the affected scope is marked as invalid.

[0013] This invention also provides a robot training data governance system based on data lineage constraints, which supports the above-mentioned robot training data governance method. The system includes a robot data access module, a lineage graph construction module, a lineage grouping module, a constraint segmentation module, a contamination detection module, a segmentation list module, a data release token module, an invalidation and rollback module, a data usage interface, and a database. The robot data access module is used to receive raw task segment data, derived data, and version events, and write them to the database; The lineage graph construction module generates a data lineage graph based on the original job fragment data, derived data, and version events; the lineage grouping module generates indivisible lineage groups based on the data lineage graph. The constraint segmentation module generates bloodline allocation results based on indivisible bloodline groups, and the contamination detection module generates a contamination report based on the bloodline allocation results. The segmentation list module generates a frozen segmentation list based on the bloodline group allocation results and contamination reports. The data release token module generates a data release token that binds the frozen segmentation list and version binding information when the data quality, data permission, and contamination threshold meet preset conditions. The data usage interface reads data or verifies evidence when the data release token is valid, and generates model verification records; When the current version binding information is inconsistent with the version binding information bound to the data release token, the failure and rollback module locates the affected object along the data lineage graph and reference relationships, revokes the data release token, marks the model verification record as invalid, and controls the data usage interface to stop reading old data or old verification evidence.

[0014] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this invention, by establishing data lineage nodes for the original job fragment data and derived data, and recording the source job fragment identifier, source relationship, and version relationship, a queryable link is formed between the training samples and the generation process. Indivisible lineage groups are formed through hard association and soft association risks, and set allocation and cross-set pollution detection are performed according to the lineage groups, so that the segmentation status of the same source data corresponds to the pollution report. By freezing the segmentation list, data release token, and version binding information, the reading conditions of the data use interface correspond to the valid status of the model verification record. When the versions are inconsistent, the affected objects are located and the reading of old data or old verification evidence is blocked. Attached Figure Description

[0015] Figure 1 This is a flowchart of the robot training data governance process of the present invention; Figure 2 This is a schematic diagram of the data lineage diagram structure of the present invention; Figure 3 This is a schematic diagram of bloodline grouping and set allocation in this invention; Figure 4 This is a schematic diagram illustrating the data publishing token and expiration handling of this invention; Figure 5 This is a schematic diagram illustrating the collaborative operation of the robot training data governance system modules of the present invention. Detailed Implementation

[0016] The technical solutions described in this invention will be further explained below with reference to the embodiments. The original work segment data in the embodiments refers to the data set collected by the robot controller for continuous tasks, sub-tasks, contact phases, or fault phases; derived data refers to data formed from the original work segment data through windowing, resampling, enhancement, synthesis, annotation revision, or fault playback processing; version events refer to events related to sensor calibration versions, tool or fixture identifiers, workpiece batch identifiers, scene map versions, safety rule versions, or associated robot versions.

[0017] This embodiment uses the training data management of a precision insertion robot as an application example. The robot system includes a six-axis industrial robot, an RGB-D camera, a six-dimensional force or torque sensor at the end effector, and a workstation controller. The robot controller has a control cycle of 4 milliseconds, the RGB-D camera has a data acquisition frequency of 30 frames per second, and the six-dimensional force or torque sensor has a data acquisition frequency of 1000 Hz. When an insertion operation segment lasts for 8 seconds, the RGB-D camera acquires approximately 240 frames of images and depth maps, the robot controller outputs approximately 2000 sets of robot joint states and end effector poses, the six-dimensional force or torque sensor outputs approximately 8000 sets of force or torque sequences, and the workstation controller outputs corresponding workstation events and task logs.

[0018] Images and depth maps are acquired by the color imaging unit and depth imaging unit of an RGB-D camera. Point clouds are generated using depth maps, camera intrinsics, and point cloud back-projection techniques. Robot joint states are acquired by the robot controller's servo state registers or real-time state interface. End-effector pose is output by the robot controller based on the robot joint states and according to forward kinematics. Force or torque sequences are acquired by six-dimensional force or torque sensors mounted on the robot's end effector or wrist. Workstation events and task logs are output by the workstation controller or programmable logic controller via industrial Ethernet, fieldbus, message queues, or file interfaces.

[0019] Please see Figures 1 to 4 This embodiment provides a robot training data governance method based on data lineage constraints. This method addresses the processes of robot training data acquisition, derivation, grouping, segmentation, detection, publication, and failure handling. The execution environment includes a training server, a robot edge control computer, or a data governance environment composed of a training server and a robot edge control computer. The robot edge control computer is responsible for accessing the original job segment data and aligning its timestamps. The training server stores the data lineage graph, the frozen segmentation list, and the data publication token. The data usage interface corresponds to the model training interface, testing interface, or robot deployment verification interface. The training server and the robot edge control computer transmit the source job segment identifier, data lineage node, frozen segmentation list, and data publication token status via industrial Ethernet or message queues. Before reading data through the data usage interface, the data publication token status is accessed first, and the decision to read the data corresponding to the frozen segmentation list is made based on the token status.

[0020] S1 receives raw task segment data and version events triggered by the robot controller. The robot controller triggers the acquisition process during continuous tasks, subtasks, contact phases, or fault phases, forming raw task segment data corresponding to the robot's operation process. Raw task segment data includes at least two of the following: images, depth maps, point clouds, robot joint states, end-effector poses, force or torque sequences, workstation events, and task logs. Version events include at least two of the following: sensor calibration version, tool or fixture identifier, workpiece batch identifier, scene map version, safety rule version, and associated robot version. After receiving the raw task segment data, a source task segment identifier is assigned to it. The source task segment identifier corresponds to a physical operation time range and serves as the basis for establishing the source relationship between the raw task segment data and derived data.

[0021] In a batch of data access for a production line, a single source job segment identifier corresponds to a continuous job segment. The source job segment identifier is associated with robot number ROBOT-01, workstation number STATION-02, workpiece batch BATCH-B, fixture identifier FIXTURE-03, camera calibration version CAL-XXXX-07-18, and scene map version MAP-07. When the same robot completes 120 insertion actions in one shift, 120 source job segment identifiers are formed. Each source job segment identifier corresponds to a data acquisition time range and a set of version events. When each job segment generates an average of 1 original job segment node and 29 derived nodes, approximately 3600 data lineage nodes are formed in one shift.

[0022] Workpiece batch identifiers are obtained by the workstation controller reading the workstation recipe or batch records issued by the manufacturing execution system. Tool or fixture identifiers are obtained from the robot controller register, fixture identification sensors, or equipment identification information. Sensor calibration versions are obtained from the camera controller or sensor management configuration files. Scene map versions and safety rule versions are obtained from the workstation controller, robot controller, or signed configuration messages.

[0023] S101: Obtain the trigger timestamps generated by the robot controller at task start, task end, contact establishment, contact disengagement, or fault reset. The trigger timestamps are used to define the time correspondence between the version event and the original job segment data. Match the trigger timestamps with the acquisition time range of the original job segment data. When the trigger timestamp falls within the acquisition time range, or the difference between the trigger timestamp and the boundary of the acquisition time range meets the synchronization error condition, write the corresponding version event to the source job segment identifier of the original job segment data. At this time, the version event and the source job segment identifier form a version binding relationship under the same job segment. If the trigger timestamp and the acquisition time range do not meet the synchronization error condition, stop writing the corresponding version event to the source job segment identifier.

[0024] In this embodiment, the synchronization error condition is set to the difference between the trigger timestamp and the boundary of the acquisition time range not exceeding 20 milliseconds. When the task start timestamp is 10:15:30.120 and the acquisition time range is from 10:15:30.112 to 10:15:38.096, the trigger timestamp falls within the acquisition time range, and the version event is written to the source job segment identifier. When the tool replacement event timestamp is 10:15:38.180, the difference between this timestamp and the end boundary of the acquisition time range is 84 milliseconds, exceeding the 20-millisecond synchronization error condition, and the tool replacement event is not written to the source job segment identifier. The 20-millisecond synchronization error condition is used to match the robot controller trigger time, camera frame timestamp, and force or torque sensor timestamp; all comparison objects are time quantities with consistent dimensions.

[0025] S102, the version event writing process includes reading the sensor calibration version, tool or fixture identifier, workpiece batch identifier, scene map version, and safety rule version from the version event. The reading result, along with the trigger timestamp, acquisition time range, and source job segment identifier, forms a version record. If the same source job segment identifier corresponds to multiple version records, the valid version record is selected according to the time matching order of the trigger timestamp and acquisition time range, and the unselected version records are marked as pending review. The pending review status does not enter subsequent version binding information.

[0026] Version records include the source job segment identifier, trigger timestamp, acquisition time range, sensor calibration version, tool or fixture identifier, workpiece batch identifier, scene map version, and safety rule version. For example, a version record might include the source job segment identifier E-ROBOT01-000418, sensor calibration version CAL-XXXX-07-18, tool or fixture identifier FIXTURE-03, workpiece batch identifier BATCH-B, scene map version MAP-07, and safety rule version SAFE-02. When the same source job segment identifier corresponds to two tool or fixture identifier records, the record with its trigger timestamp within the acquisition time range is prioritized as the valid version record, and the other record is marked as pending review. If both version records have trigger timestamps within the acquisition time range, the version record closest to the acquisition start time and originating from a valid configuration message from the robot controller or workstation controller is selected as the valid version record, and the remaining version records enter the pending review state. If neither version record meets the time matching condition, the source job segment identifier is not written to this type of version event.

[0027] S103, acquire the derived data formed by the derivation process. Derivation processes include windowing, resampling, enhancement, compositing, annotation revision, or fault playback. Derived data does not replace the original job segment data; it retains the source relationships generated from the original job segment data. For derived data generated from multiple original job segment data, the derived data saves the source of the corresponding parent node; for data formed by further processing of other derived data, the derived data saves the source of the previous derived data.

[0028] When an 8-second original task segment is windowed with a 2-second window length and a 1-second sliding step, it forms 7 window samples, covering 0-2 seconds, 1-3 seconds, 2-4 seconds, 3-5 seconds, 4-6 seconds, 5-7 seconds, and 6-8 seconds respectively. Applying three types of enhancements—brightness perturbation, cropping, and time scaling—to each window sample results in 21 enhanced samples based on the 7 window samples. If the original task segment corresponds to a fault playback process, the fault playback data maintains the same source task segment identifier as the original task segment data. Resampling processing uses timestamp interpolation or nearest neighbor alignment to ensure that image frames, robot states, and force or torque sequences at different sampling frequencies fall within the same window sample. Annotation revisions are generated by the annotation terminal or annotation management system, and fault playback data is generated by the robot controller or training server based on the source fault task segment.

[0029] S2 establishes data lineage nodes for the original job segment data and derived data. Data lineage nodes include original job segment nodes and derived nodes. Original job segment nodes contain the source job segment identifier, acquisition time range, robot identifier, sensor identifier, content summary, and version relationship. Derived nodes contain the parent node identifier, derived processing type, processing parameter summary, and content summary. The source relationship between data lineage nodes represents the parent-child source relationship between the original job segment data and derived data, and between derived data and subsequent derived data; the version relationship corresponding to version events represents the correspondence between data nodes and sensor calibration versions, tool or fixture identifiers, workpiece batch identifiers, scene map versions, safety rule versions, or associated robot versions.

[0030] The original job segment node includes node identifier N-RAW-000418, node type raw_episode, source job segment identifier E-ROBOT01-000418, acquisition time range 10:15:30.112 to 10:15:38.096, robot identifier ROBOT-01, sensor identifiers CAM-01 and FT-01, and a content summary. Derived nodes include node identifier N-DER-000418-03, parent node identifier N-RAW-000418, derived processing type sliding_window, processing parameter summary, and content summary. The content summary is generated using hash digest technology and is used for subsequent duplicate content detection. The content summary field is used to identify whether the data content is consistent and is not used to indicate data acquisition time, version status, or sample category.

[0031] S201, when constructing the data lineage graph, each data lineage node is treated as an addressable node, and source and version relationships are used as connections between nodes and version records. Original job fragment nodes are connected to their corresponding derived nodes via the source job fragment identifier. Derived nodes are connected to their upstream nodes via the parent node identifier. Version relationships connect the corresponding valid version records. The data lineage graph simultaneously represents data source, derivation processing, and version status. Subsequent lineage grouping, set splitting, pollution detection, token issuance, and failure location all use the data lineage graph as their indexing basis.

[0032] When one original job segment node corresponds to 7 window samples, 21 enhanced samples, and 1 fault playback sample, the data lineage graph forms 1 original job segment node, 29 derived nodes, and source relationships connected by parent node identifiers. Version relationships associate the above nodes together with the version records corresponding to CAL-XXXX-07-18, FIXTURE-03, BATCH-B, and MAP-07. When all 120 source job segment identifiers are processed using the same derivation method, the data lineage graph forms approximately 3600 data lineage nodes and connections corresponding to source and version relationships. These 3600 data lineage nodes consist of 120 original job segment nodes and 3480 derived nodes, consistent with the setting that each source job segment identifier forms an average of 30 data lineage nodes.

[0033] S202, when the parent node identifier of a derived node does not match an existing node in the data lineage graph, stop writing the derived node to the data lineage graph and output a node relationship pending confirmation status. The node relationship pending confirmation status indicates that the source relationship of the derived data has not yet met the writing conditions. If the parent node identifier of a derived node matches an existing node in the data lineage graph, then write the derived node to the data lineage graph and maintain the source relationship between the derived node and its parent node.

[0034] If the parent node identifier of the derived node N-DER-000418-03 is N-RAW-000418, and N-RAW-000418 already exists in the data lineage graph, the derived node is written into the data lineage graph. If the parent node identifier is N-RAW-999999 but there is no corresponding node in the data lineage graph, the derived node enters a node relationship pending confirmation state and does not participate in the generation of this round of indivisible lineage groups. The node relationship pending confirmation state indicates that the source node of the derived data cannot be determined. Data in this state will not enter the set splitting, cross-set pollution detection, and data release token issuance process until the parent node identifier is corrected or the corresponding derived data is excluded.

[0035] S3, based on the hard and soft association risks in the data lineage graph, applies merging or mutual exclusion constraints to data lineage nodes, forming indivisible lineage groups. Hard associations are determined by whether data lineage nodes share the same source job segment identifier, parent-child source relationship, same fault playback source, same synthesis processing seed, or same annotation revision chain. When two data lineage nodes satisfy a hard association, they are merged into the same initial lineage group. The initial lineage group accepts data nodes from the same job source or the same derived link; this source link is not split during subsequent set partitioning.

[0036] Node identifier N-RAW-000418, 7 window sample nodes derived from N-RAW-000418, 21 enhanced sample nodes, and 1 fault playback sample node satisfying the same source job segment identifier or parent-child source relationship, are merged to form the same initial lineage group G-000418. When all 120 source job segment identifiers have completed hard association processing, 120 initial lineage groups are first formed. The hard association relationship is determined by the source job segment identifier, parent node identifier, fault playback source identifier, synthesis processing seed identifier, and annotation revision chain identifier. These fields are written by the data access, derivation processing, or annotation revision stages.

[0037] S301, calculate the soft association risk between different initial lineage groups. When calculating the soft association risk, extract the workpiece batch identifier, scene map version, tool or fixture identifier, acquisition time range, trajectory features, visual features, and sensor calibration version from the data lineage nodes corresponding to the two initial lineage groups. Batch similarity is calculated based on the workpiece batch identifier, scene similarity based on the scene map version, tool similarity based on the tool or fixture identifier, time proximity based on the acquisition time range, trajectory similarity based on the trajectory features, visual similarity based on the visual features, and calibration similarity based on the sensor calibration version. These similarities are then normalized and weighted to obtain the soft association risk.

[0038] Workpiece batch identification is generated from the workstation recipe, workpiece batch record, or output from the workstation controller. Scene map version is obtained from the workstation controller or map management configuration. Tool or fixture identification is obtained from the robot controller register, fixture identification sensor, or device identity information. Acquisition time range is determined by the robot controller trigger timestamp and sensor sampling timestamp. Trajectory features are obtained from the robot joint states and end-effector poses through resampling, time alignment, and normalization. Visual features are obtained from images, depth maps, or point clouds acquired by the RGB-D camera through perceptual summarization, local feature matching, or embedded feature extraction techniques. Sensor calibration version is read from the camera controller, sensor controller, or calibration configuration file.

[0039] In soft association risk, all similarity scores are normalized to a dimensionless interval of 0 to 1, and the weights are dimensionless numbers. Therefore, soft association risk is a dimensionless result. Soft association risk is expressed using a weighted sum formula: ; in, This indicates the risk of soft association. Indicates batch similarity. Indicates scene similarity. Indicates the degree of similarity of tools. Indicates temporal proximity. Indicates the similarity of trajectory features or visual features. Indicates the degree of similarity in calibration. , , , , , These represent the weights corresponding to batch similarity, scene similarity, tool similarity, temporal proximity, trajectory feature or visual feature similarity, and calibration similarity, respectively. .

[0040] In this embodiment, , , , , , .

[0041] When two initial bloodlines , , , , , hour; ; The calculation result satisfies the merging condition. When two initial bloodline groups... , , , , , hour, ; The calculation results satisfy the mutual exclusion condition. Trajectory feature similarity and visual feature similarity are both merged. The candidate group is weighted once; when the same candidate group has both trajectory feature similarity and visual feature similarity, the higher value between the two is selected within the range of 0 to 1, or the two values ​​are combined according to the preset internal weights into one. The value is then entered into the above formula to avoid the risk of the same related factor being repeatedly included in the soft association.

[0042] S302: When the soft association risk meets the merging condition, the corresponding initial lineage groups are merged; when the soft association risk does not meet the merging condition but meets the mutual exclusion condition, mutual exclusion constraints are established for the corresponding initial lineage groups. The merging condition indicates that there is a degree of association between different initial lineage groups sufficient to allow them to jointly enter the same dataset. The mutual exclusion condition indicates that different initial lineage groups are not directly merged, but should not be separately entered into the training set and independent real-machine test set. After processing for hard associations and soft association risks, indivisible lineage groups are formed. Indivisible lineage groups serve as the smallest unit for subsequent set allocation, and the same indivisible lineage group is no longer split into different sets.

[0043] In this embodiment, the merging condition is a soft association risk of no less than 0.85, and the mutual exclusion condition is a soft association risk of no less than 0.65 and less than 0.85. Of the 120 initial lineage groups, 6 are merged into 3 lineage groups because their soft association risk is no less than 0.85. Simultaneously, 10 pairs of initial lineage groups establish mutual exclusion constraints because their soft association risk is between 0.65 and 0.85. This results in 117 indivisible lineage groups and 10 mutually exclusive constraint records. When the soft association risk is less than 0.65, the corresponding initial lineage group is not merged and no mutual exclusion constraint is established; subsequently, it only participates in set allocation based on set coverage constraints and existing mutual exclusion constraints.

[0044] S4 assigns indivisible lineage groups to the training set, validation set, calibration set, and independent real-device test set, generating lineage group assignment results. Before assignment, set coverage constraints are established for indivisible lineage groups based on at least two of the following: task label, success / failure status, fault type, object domain, tool domain, scenario domain, calibration domain, and risk level. Set coverage constraints limit the coverage of each set in the task label, job status, object category, or version domain. During assignment, under the conditions of satisfying set coverage constraints and mutual exclusion constraints, indivisible lineage groups are assigned to the training set, validation set, calibration set, and independent real-device test set. The training set is used for training data, the validation set is used for model selection-related reads, the calibration set is used for calibration-related reads, and the independent real-device test set is used for frozen test reads.

[0045] Task labels are obtained from the annotation terminal, workstation controller task number, or robot task log. Success / failure status is obtained from task completion events, workstation detection results, or manual annotation records. Fault type is obtained from fault reset events, protection shutdown events, or fault playback sources. Object domain, tool domain, scene domain, and calibration domain are obtained from workpiece batch identifier, tool or fixture identifier, scene map version, and sensor calibration version, respectively. Risk level is obtained from fault type, contamination detection candidate status, and soft correlation risk classification.

[0046] In dataset version DS-2.1, out of 117 inseparable lineage groups, 82 inseparable lineage groups are assigned to the training set, 12 to the validation set, 12 to the calibration set, and 11 to the independent real-machine test set. The sum of these sets is 117, consistent with the total number of inseparable lineage groups. Window samples, augmented samples, and fault replay samples within the same inseparable lineage group remain in the same set.

[0047] Set allocation is constrained by a weighted bias objective, expressed by the following formula: ; in, This represents the target value of the set allocation deviation. Indicates the set size deviation. This indicates a deviation in the distribution of task labels or success / failure status. This indicates the distribution deviation of the object domain, tool domain, scene domain, or calibration domain. This indicates a deviation in the distribution of fault types or risk levels. , , , These represent the weights corresponding to set size bias, label distribution bias, domain distribution bias, and risk distribution bias, respectively. . , , and All are normalized deviation values. , , and Since the weights are dimensionless, This is the dimensionless deviation target value.

[0048] In this embodiment, , , , .when , , , hour, ; This formula is used to allow the size of the training set, validation set, calibration set, and independent real-machine test set to participate in the segmentation in conjunction with the label domain coverage relationship, under the condition that the indivisible lineage group cannot be segmented.

[0049] S401 performs cross-set contamination detection on data from different sets and generates a contamination report. Cross-set contamination detection includes at least two of the following: content duplication detection, visual or point cloud near-duplicate detection, trajectory or force sequence similarity detection, object or scene overlap detection, and metadata homology detection. Content duplication detection uses content summaries as the comparison object. Visual or point cloud near-duplicate detection uses image, depth map, or point cloud-related features as the comparison object. Trajectory or force sequence similarity detection uses robot motion or contact process data as the comparison object. Object or scene overlap detection uses object domain, scene domain, or related version information as the comparison object. Metadata homology detection uses metadata such as acquisition time range, workpiece batch identifier, tool or fixture identifier, and sensor calibration version as the comparison object.

[0050] Content duplication detection is achieved by comparing the consistency of content summaries, which are generated by hash digest technology based on the byte content of the original task fragment data or derived data. Visual near-duplication detection uses perceptual hashing, local feature matching, or image embedding feature comparison techniques to obtain visual features. Point cloud near-duplication detection uses voxelized point cloud summarization, point cloud registration overlap rate, or 3D asset identification comparison to obtain point cloud similarity states. Trajectory or force sequence similarity detection first performs temporal alignment of robot joint states, end-effector poses, and force or torque sequences based on control cycles, contact establishment events, or key events, and then normalizes the dimensions of different units. The normalized values ​​are used for similarity comparison to avoid inconsistencies in units caused by directly mixing position, pose, force, and torque.

[0051] Visual near-repetition detection uses 64-bit perceptual summaries for initial screening. The Hamming distance formula between two perceptual summaries is: ; in, Representing two perceptual summaries and Hamming distance between them Indicates the number of bits in the perceived summary. Representation of perceptual summary The first in Bit, Representation of perceptual summary The first in Bit, Indicates when the first An indicator function that takes 1 when bits are different and 0 when they are the same. In a 64-bit perceptual summarization scenario, , The value ranges from 0 to 64. The Hamming distance represents the positional difference and is a dimensionless counting result. When the Hamming distance is not greater than 6, the corresponding image sample is marked as a visual near-duplicate candidate, and then written into the pollution report based on the source path and set affiliation. When the Hamming distance is greater than 6, the corresponding sample pair does not enter the visual near-duplicate candidate, but it is still subject to trajectory or force sequence similarity detection, object or scene overlap detection, or metadata homology detection. When the Hamming distance between sample N-DER-000418-03 in the training set and sample N-DER-000579-01 in the independent real device test set is 5, the sample pair enters the visual near-duplicate candidate; when the Hamming distance is 12, the sample pair does not enter the visual near-duplicate candidate.

[0052] Trajectory or force sequence similarity detection uses Euclidean distance as the basic measure of local sequence differences after alignment, with the formula: ; in, Represents sequence point With sequence points The Euclidean distance between them Indicates the number of dimensions involved in the comparison. Represents sequence point In the Normalized values ​​in each dimension Represents sequence point In the Normalized values ​​across each dimension. For end-effector pose comparison, The dimensions corresponding to the normalized position and orientation; for comparisons of force or torque sequences. The normalized dimensions for stress and moment. Because and All values ​​are normalized. The distance is dimensionless.

[0053] In this embodiment, the preset distance boundary is 0.20. Continuously below the preset distance boundary means that at least two key event alignment points in the contact establishment, contact maintenance, and contact disengagement processes are met. All values ​​are no greater than 0.20. If the distance between the normalized sequence points of two contact phases is continuously lower than the preset distance boundary, the corresponding trajectory or force sequence enters the similarity candidate; if only one key event alignment point meets this condition, no trajectory or force sequence similarity candidate is formed. The preset distance boundary is the normalized distance threshold and does not have dimensions of length, angle, force, or torque.

[0054] S402, the contamination report includes the associated nodes, source path, detection score, judgment threshold, and disposal status. Associated nodes represent the data lineage nodes involved in cross-set contamination detection. The source path represents the tracing path of the associated nodes in the data lineage graph. The detection score represents the degree of similarity or identicalness. The judgment threshold represents the boundary for determining whether a detection score enters the contamination state. The disposal status indicates whether the detection results have been merged, mutually excluded, removed, reviewed, or retained. The contamination report enters the frozen segmentation list generation step, serving as one of the inputs for determining whether the data segmentation status is available for publication and reading.

[0055] A record in the contamination report includes the left node N-DER-000418-03, the right node N-DER-000579-01, the left set training set, the right set independent real machine test set, the source path N-RAW-000418 to N-DER-000418-03, the detection type visual near duplication, the detection score 5, the judgment threshold 6, and the disposal status mutual exclusion verification. When the disposal status is undisposed and the detection score meets the contamination judgment condition, the record enters the contamination threshold verification. For the dataset version containing 117 indivisible lineage groups, contamination detection generates 8 visual near duplication candidates, 3 trajectory similarity candidates, and 2 metadata homology candidates. After merging, mutual exclusion, or verification, the number of undisposed high-risk cross-set contamination records is used as the input for contamination threshold verification. If the status of a contaminated record is "merged", the bloodline group to which the corresponding data lineage node belongs will be regrouped; if the status of a contaminated record is "mutually exclusive", the corresponding indivisible bloodline group cannot be entered into the training set and the independent real machine test set respectively; if the status of a contaminated record is "excluded", the corresponding sample will not be entered into the frozen segmentation list; if the status of a contaminated record is "review", the frozen segmentation list will not enter the data release token issuance process before the review is completed.

[0056] S5 generates a frozen segmentation list based on the lineage group allocation results, contamination reports, and version relationships. The frozen segmentation list records the set affiliation of indivisible lineage groups, the version status related to version relationships, and the detection results in the contamination reports. After the frozen segmentation list is formed, the allocation results are no longer used as temporary results for subsequent readings, but rather as the list basis for binding data release tokens. Version binding information is generated based on version relationships, data license status, and segmentation rule versions. Version binding information includes a data content summary, sensor calibration version, tool or fixture identifier, scene map version, data license status, segmentation rule version, and associated robot version.

[0057] The frozen split list includes the list identifier SM-XXXX-07-22-01, dataset version DS-2.1, indivisible lineage group identifier G-000418, set ownership independent_robot_test, source job fragment identifier E-ROBOT01-000418, version binding information CAL-XXXX-07-18, FIXTURE-03, MAP-07, data permission status approved, splitting rule version SPLIT-RULE-02, and pollution digest. When all 117 indivisible lineage groups have set ownership and version binding information, the frozen split list enters the threshold verification process. If any indivisible lineage group lacks set ownership, version binding information, or a pollution digest, the frozen split list remains unfrozen until the missing fields are filled in or the corresponding indivisible lineage group is excluded.

[0058] S501, before issuing a data release token, sequential checks are performed on data quality, data permissions, and contamination thresholds. Data quality check confirms whether the data corresponding to the frozen segmentation list meets preset data integrity or readable conditions. Data permission check confirms whether the data permission status meets the usage scope. Contamination threshold check confirms whether the detection results in the contamination report meet the release conditions. If any check result fails to meet the preset conditions, the issuance of the data release token is stopped; if data quality, data permissions, and contamination thresholds all meet the preset conditions, the data release token is issued. The data release token is bound to the frozen segmentation list and version binding information, enabling downstream reading to be performed based on the frozen segmentation list, version binding information, and the token's validity status.

[0059] Data quality verification uses data integrity rate as the criterion, and the formula is: ; in, Indicates data integrity rate. This indicates the number of data records that meet the criteria of complete timestamps, correct field formats, existing content summaries, and valid version relationships. This indicates the total number of data records corresponding to the frozen segmentation list. and Both refer to the number of data records, and they have the same unit of measurement. This is a dimensionless ratio. When the frozen segment list corresponds to 24,000 data records, and 23,880 of them meet the integrity condition... ; When the data integrity rate is not less than 0.995, the data permission status is approved, and the number of unprocessed high-risk cross-collection contamination records in the contamination report is 0, the data release token enters the issuance process. If the data integrity rate is 0.991, or the data permission status is expired, or the number of unprocessed high-risk cross-collection contamination records is greater than 0, the issuance of data release tokens is stopped. The 24,000 data records correspond to the total number of synchronized image frame records, depth map records, robot status records, force or torque sampling aggregation records, version records, and derived sample records in a list format; the number of data records is used for quality verification and is not equivalent to the total number of original sensor point-by-point samples. The execution result of sequential verification is as follows: if the data quality verification is not met, subsequent data permission verification and contamination threshold verification are stopped; if the data quality verification is met but the data permission verification is not met, contamination threshold verification is stopped; if both of the first two are met, contamination threshold verification is then performed.

[0060] S6, when the data release token is valid, allows the data usage interface to read data or verify evidence, and generates a model verification record based on the frozen segmentation list, model summary, threshold summary, and robot version tuple. The data usage interface includes at least one of a model training interface, a testing interface, or a robot deployment verification interface. The data usage interface reads the validity status of the data release token before reading data or verifying evidence. When the data release token is valid, the data usage interface reads the data or verification evidence corresponding to the frozen segmentation list; when the data release token is invalid, the data usage interface stops reading the corresponding data or verification evidence. The model verification record is bound to the frozen segmentation list, model summary, threshold summary, and robot version tuple, representing the data list, model status, threshold status, and robot version status used in the corresponding model verification process.

[0061] The model validation records include the model summary MODEL-DIGEST-03, the threshold summary THRESHOLD-02, the robot version tuple ROBOT-01, CAL-XXXX-07-18, FIXTURE-03, MAP-07, and SAFE-02. When the test interface reads an independent real-device test set, the model validation records simultaneously write the data release token identifier, the frozen segmentation list identifier, and the reading time. If the data release token status field is valid, the data usage interface allows reading; if the data release token status field is revoked, the data usage interface returns a read blocking status. The data read by the data usage interface comes from the set ownership and data node references in the frozen segmentation list. Validation evidence comes from the binding records between the model summary, threshold summary, test records, and robot version tuples. The model summary is generated from the model file content summary and model structure version; the threshold summary is generated from the threshold configuration file content summary; and the robot version tuple is formed by the robot controller, sensor controller, tool or fixture configuration, scene map configuration, and safety rule configuration.

[0062] S7. Retrieve the current version binding information, which shares the same fields as the version binding information. The current version binding information is derived from the currently readable data content summary, sensor calibration version, tool or fixture identifier, scene map version, data license status, segmentation rule version, and associated robot version. Compare the current version binding information with the version binding information bound to the data release token. If the current version binding information matches the version binding information bound to the data release token, maintain the validity of the data release token; if the current version binding information does not match the version binding information bound to the data release token, locate the affected data lineage nodes, indivisible lineage groups, frozen segmentation lists, and model validation records along the data lineage graph and reference relationships.

[0063] For example, if the tool or fixture identifier bound to the data release token is FIXTURE-03, but the tool or fixture identifier in the current version binding information is updated to FIXTURE-04, the current version binding information is inconsistent with the version binding information bound to the data release token. When the data permission status changes from approved to expired, the same failure location process is also initiated. If the current version binding information is consistent with the version binding information bound to the data release token in the fields of data content summary, sensor calibration version, tool or fixture identifier, scene map version, data permission status, segmentation rule version, and associated robot version, then the data release token remains valid, and the data usage interface continues to read the corresponding data or verify evidence according to S6. The source of the current version binding information is consistent with the version binding information itself. Specifically, the data content summary is reread from the current data object or read from the database index; the sensor calibration version is read from the sensor controller or calibration configuration file; the tool or fixture identifier is read from the robot controller register or fixture identification sensor; the scene map version and safety rule version are read from the workstation controller or configuration message; and the data permission status and segmentation rule version are read from the governance database.

[0064] S701, when locating the affected area, the affected data lineage nodes are determined based on the source relationships in the data lineage graph. The affected indivisible lineage groups and set allocation positions are determined based on the set affiliation in the frozen segmentation list. The affected verification records are determined based on the frozen segmentation list, model summary, threshold summary, and robot version tuple in the model verification records. After location is complete, the data release token is revoked, the model verification records are marked as invalid, and the data usage interface is blocked from reading old data or old verification evidence. The invalid state indicates that the consistency condition between the model verification record and the current version binding information is no longer met. The read blocking is used to prevent downstream training, testing, or deployment verification from continuing to use old data or old verification evidence.

[0065] If the number of data lineage nodes related to FIXTURE-03 in the data lineage graph is 360, corresponding to 18 indivisible lineage groups, and the frozen split list involves four sets: training set, validation set, calibration set, and independent real-device test set, then the failure and rollback processing will locate the aforementioned 18 indivisible lineage groups and their corresponding model validation records level by level according to the source relationship and set affiliation, and update the relevant data release token status from valid to revoked. If only two indivisible lineage groups are affected, then only the corresponding frozen split list and model validation records will be marked as invalid, without changing the reading status of the remaining unaffected lineage groups. There is no fixed ratio requirement between the 360 ​​data lineage nodes and the 18 indivisible lineage groups; they represent the node-level impact range and the group-level impact range, respectively. The node-level impact range is determined by the source relationship in the data lineage graph, and the group-level impact range is determined by the indivisible lineage group to which the data lineage node belongs. If the corresponding model verification record cannot be located, the failure and rollback process only revokes the corresponding data release token and blocks the data usage interface from reading old data; if the model verification record is located, the model verification record is marked as failed.

[0066] Please see Figure 2 and Figure 5 This embodiment also provides a robot training data governance system based on data lineage constraints. This system supports the aforementioned robot training data governance method and includes a robot data access module, a lineage graph construction module, a lineage grouping module, a constraint segmentation module, a contamination detection module, a segmentation list module, a data release token module, an invalidation and rollback module, a data usage interface, and a database. The database stores original job segment data, derived data, version events, data lineage graphs, indivisible lineage groups, lineage group allocation results, contamination reports, frozen segmentation lists, data release tokens, and model validation records.

[0067] In the engineering deployment, the database stores the following tables: original job segment table, derived node table, version record table, lineage edge table, lineage group table, segmentation list table, contamination report table, data release token table, and model validation record table. These tables are linked through source job segment identifiers, data lineage node identifiers, indivisible lineage group identifiers, and frozen segmentation list identifiers. The original job segment table stores the data indexes collected by the robot controller, RGB-D camera, six-dimensional force or torque sensor, and workstation controller. The derived node table stores the derived data indexes after windowing, resampling, enhancement, synthesis, annotation revision, or fault playback processing. The version record table stores the sensor calibration version, tool or fixture identifier, workpiece batch identifier, scene map version, and safety rule version.

[0068] The robot data access module receives raw task data, derived data, and version events, and writes them to the database. The input to the robot data access module comes from the data acquisition process triggered by the robot controller, and the output is data objects and version events that can be read by the lineage graph construction module. The lineage graph construction module generates a data lineage graph based on the raw task data, derived data, and version events, and writes the data lineage graph to the database. The output of the lineage graph construction module enters the lineage grouping module, which generates indivisible lineage groups based on the data lineage graph. When generating indivisible lineage groups, the lineage grouping module uses hard associations and soft association risks to ensure that the lineage groups cover definite source relationships and bear the merging or mutual exclusion constraints arising from soft association risks.

[0069] The robot data access module writes data in batches according to the source job segment identifier, with each batch corresponding to a continuous job segment. The lineage map construction module reads this batch of data and generates original job segment nodes and derived nodes. The lineage grouping module reads the nodes and their source relationships and outputs indivisible lineage group identifiers for use by the constraint segmentation module. The data received by the robot data access module is not manually entered or generated. Images, depth maps, and point clouds come from RGB-D cameras and point cloud generation technology; robot joint states and end-effector poses come from the robot controller's real-time status interface; force or torque sequences come from six-dimensional force or torque sensors; and workstation events and task logs come from the workstation controller or programmable logic controller.

[0070] The constraint segmentation module generates lineage group allocation results based on indivisible lineage groups. It receives these indivisible lineage groups and, in conjunction with set coverage constraints and mutual exclusion constraints, allocates them to the training set, validation set, calibration set, and independent real-device test set. The contamination detection module generates a contamination report based on the lineage group allocation results. It reads data from different sets and detects at least two of the following: content duplication, visual or point cloud near-duplication, trajectory or force sequence similarity, object or scene overlap, and metadata homogeneity. The detection results are then written to the contamination report. The contamination report is output to the segmentation list module as input for freezing the segmentation list generation.

[0071] When the bloodline group allocation results output by the constraint segmentation module contain mutually exclusive constraint conflicts, the constraint segmentation module does not write the conflict results into the frozen segmentation list. Instead, it returns the identifier of the conflicting indivisible bloodline group to the bloodline grouping module or the contamination detection module for verification. When the number of untreated high-risk records in the contamination report generated by the contamination detection module is 0, the segmentation list module enters the frozen segmentation list generation process. If there are untreated high-risk records in the contamination report output by the contamination detection module, the segmentation list module remains unfrozen, and the data release token module does not enter the issuance process.

[0072] The segmentation list module generates a frozen segmentation list based on lineage group allocation results and contamination reports. The data release token module generates a data release token that binds the frozen segmentation list and version binding information when data quality, data permissions, and contamination thresholds meet preset conditions. The data usage interface reads data or verification evidence and generates model verification records when the data release token is valid. The failure and rollback module locates the affected object along the data lineage graph and reference relationships when the current version binding information is inconsistent with the version binding information bound to the data release token, revokes the data release token, marks the model verification record as invalid, and controls the data usage interface to stop reading old data or old verification evidence. Each module in the system receives data status through the database, receives downstream read status through the data usage interface, and handles state transitions after version changes through the failure and rollback module.

[0073] When the data release token status is valid, the data usage interface allows reading the data corresponding to the frozen segment list; when the data release token status is revoked, the data usage interface returns a read blocking status, and will no longer provide old data or old verification evidence to the model training interface, testing interface, or robot deployment verification interface. The inputs to the invalidation and rollback module include the current version binding information, the version binding information bound to the data release token, the data lineage graph, and the model verification record. The outputs include the revoked data release token status, the invalidated model verification record, and the read blocking status.

[0074] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A robot training data governance method based on data bloodline constraints, characterized in that, Includes the following steps: Receive raw job segment data and version events triggered by the robot controller, assign a source job segment identifier to the raw job segment data, and obtain derived data formed by the derivation process; Establish data lineage nodes for the original job fragment data and the derived data, record the source job fragment identifier, the source relationship between data lineage nodes and the version relationship corresponding to the version event, and construct a data lineage graph; Based on the hard associations and soft association risks in the data lineage graph, merge or mutual exclusion constraints are applied to the data lineage nodes to form indivisible lineage groups; The indivisible bloodline group is used as the smallest unit to allocate data to the training set, validation set, calibration set, and independent real machine test set, generating bloodline group allocation results. Cross-set contamination detection is performed on the data of different sets to generate contamination reports. A frozen segmentation list is generated based on the bloodline group allocation results, contamination reports, and version relationships. A data release token is issued when data quality, data permission, and contamination threshold meet preset conditions. The data release token is bound to the frozen segmentation list and version binding information. When the data release token is valid, the data is allowed to use the interface to read data or verify evidence, and a model verification record is generated based on the frozen segmentation list, model summary, threshold summary and robot version tuple; When the current version binding information is inconsistent with the version binding information bound to the data release token, locate the affected data lineage node, indivisible lineage group, frozen split list and model verification record along the data lineage graph and reference relationship, revoke the data release token, mark the model verification record as invalid, and block the data usage interface from reading old data or old verification evidence.

2. The method of claim 1, wherein: The original job segment data includes at least two of the following: images, depth maps, point clouds, robot joint states, end-effector poses, force or torque sequences, workstation events, and task logs, all triggered by the robot controller during a continuous task, subtask, contact phase, or fault phase. The version events include at least two of the following: sensor calibration version, tool or fixture identifier, workpiece batch identifier, scene map version, safety rule version, and associated robot version. The derived data includes data formed by windowing, resampling, enhancement, synthesis, annotation revision, or fault playback processing of the original job segment data.

3. The method of claim 2, wherein: The steps for receiving raw job segment data and version events triggered by the robot controller include: Obtain the trigger timestamps generated by the robot controller at the start of a task, end of a task, contact establishment, contact disengagement, or fault reset; Match the trigger timestamp with the collection time range of the original job segment data; When the trigger timestamp falls within the collection time range or meets the synchronization error condition with the boundary of the collection time range, the corresponding version event is written into the source job segment identifier of the original job segment data. When the trigger timestamp and the collection time range do not meet the synchronization error condition, stop writing the corresponding version event to the source job segment identifier.

4. The method of claim 3, wherein: The process of writing the version event includes: reading the sensor calibration version, tool or fixture identifier, workpiece batch identifier, scene map version, and safety rule version from the version event; The reading result, along with the trigger timestamp, the collection time range, and the source job segment identifier, forms a version record; When the same source job segment identifier corresponds to multiple version records, the valid version records are selected according to the time matching order of the trigger timestamp and the collection time range, and the version records that are not selected are marked as pending review.

5. The method of claim 1, wherein: The data lineage nodes include original job segment nodes and derived nodes. The original job segment nodes are written with source job segment identifier, collection time range, robot identifier, sensor identifier, content summary and version relationship. The derived nodes are written with parent node identifier, derived processing type, processing parameter summary and content summary. When the parent node identifier of the derived node does not match an existing node in the data lineage graph, stop writing the derived node into the data lineage graph and output a node relationship pending confirmation status.

6. The method of claim 1, wherein: The process of performing merge or mutual exclusion constraints on the data lineage nodes includes: Hard association relationships are determined based on whether the data lineage nodes have the same source job segment identifier, parent-child source relationship, same fault playback source, same synthesis processing seed, or same annotation revision chain. When two data lineage nodes satisfy the hard association relationship, the two data lineage nodes are merged into the same initial lineage group; Calculate the soft association risk between different initial lineage groups. When the soft association risk meets the merging condition, merge the corresponding initial lineage groups. When the soft association risk does not meet the merging condition but meets the mutual exclusion condition, establish mutual exclusion constraints for the corresponding initial lineage groups to form indivisible lineage groups.

7. The method of claim 6, wherein: The process of calculating the soft association risk includes: Extract the workpiece batch identifier, scene map version, tool or fixture identifier, acquisition time range, trajectory features, visual features, and sensor calibration version from the data lineage nodes corresponding to the two initial lineage groups; Based on the workpiece batch identifier, the scene map version, the tool or fixture identifier, the acquisition time range, the trajectory features, the visual features, and the sensor calibration version, batch similarity, scene similarity, tool similarity, time proximity, trajectory similarity, visual similarity, and calibration similarity are formed respectively. The soft association risk is obtained by normalizing and weighting the batch similarity, scene similarity, tool similarity, time proximity, trajectory similarity, visual similarity, and calibration similarity.

8. The method of claim 1, wherein: The process of generating the bloodline allocation results and the contamination report includes: Establish set coverage constraints for indivisible lineage groups based on at least two of the following: task label, success / failure status, fault type, object domain, tool domain, scenario domain, calibration domain, and risk level. Under the conditions of satisfying the set coverage constraint and mutual exclusion constraint, the indivisible lineage group is assigned to the training set, validation set, calibration set and independent real machine test set; Perform at least two of the following detection methods on data from different sets: content duplication detection, visual or point cloud near duplication detection, trajectory or force sequence similarity detection, object or scene overlap detection, and metadata homology detection. Write the associated nodes, source paths, detection scores, judgment thresholds, and disposal status into the pollution report.

9. The method of claim 1, wherein: The version binding information includes data content summary, sensor calibration version, tool or fixture identifier, scene map version, data permission status, segmentation rule version, and associated robot version; Before issuing a data release token, data quality, data permission, and pollution threshold are checked sequentially. If any check result fails to meet the preset conditions, the issuance of the data release token is stopped. When the current version binding information is inconsistent with the version binding information, the affected scope is determined according to the source relationship in the data lineage diagram and the set affiliation in the frozen segmentation list, and the model verification record corresponding to the affected scope is marked as invalid.

10. A robot training data governance system based on data lineage constraints, configured to support the robot training data governance method of any one of claims 1-9, characterized in that, include: The system includes a robot data access module, a lineage graph construction module, a lineage grouping module, a constraint segmentation module, a contamination detection module, a segmentation list module, a data release token module, an invalidation and rollback module, a data usage interface, and a database. The robot data access module is used to receive raw task segment data, derived data, and version events, and write them into the database; The lineage graph construction module generates a data lineage graph based on the original job segment data, the derived data, and the version event; the lineage grouping module generates indivisible lineage groups based on the data lineage graph. The constraint segmentation module generates a bloodline group allocation result based on the indivisible bloodline group, and the contamination detection module generates a contamination report based on the bloodline group allocation result; The segmentation list module generates a frozen segmentation list based on the bloodline group allocation results and the contamination report. The data release token module generates a data release token that is bound to the frozen segmentation list and version binding information when the data quality, data permission, and contamination threshold meet preset conditions. The data usage interface reads data or verifies evidence when the data release token is valid, and generates model verification records; When the current version binding information is inconsistent with the version binding information bound to the data release token, the failure and rollback module locates the affected object along the data lineage graph and reference relationship, revokes the data release token, marks the model verification record as invalid, and controls the data usage interface to stop reading old data or old verification evidence.