Robot feedback-driven practical training closed-loop optimization method

Abstract

The application relates to the technical field of industrial robot practical training control, and discloses a robot feedback driving practical training closed loop optimization method; the method first collects robot joint state data, end feedback data, practical training area vision data and teach pendant input events, establishes a data time sequence corresponding relationship under the same practical training task, divides student operation segments and determines corresponding practical training states, associates the operation segments to corresponding skill subitems, identifies operation deviation types according to the data time sequence, establishes a skill defect attribution relationship and updates student skill state records, then matches a training task template according to the updated student skill state records, and writes the screened practical training data into a practical training knowledge base, so that the practical training evaluation accuracy, the training task pertinence and the equipment operation safety are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial robot training and control technology, specifically a robot feedback-driven training closed-loop optimization method. Background Technology

[0002] With the upgrading of the intelligent manufacturing industry and the digital development of vocational education, industrial robot operation training has become an important part of cultivating skilled talents. The accuracy of the training system's evaluation of trainees' operational abilities and the adaptability of training tasks directly affect the efficiency of skill acquisition. For example, the published invention patent application CN113408621B discloses a rapid imitation learning method, system, and device for robot skill learning. By extracting keyframe teaching data and training a control strategy model, the robot can reproduce the taught actions. Another example is the published invention patent application CN114660934B, which discloses a robotic arm autonomous operation strategy learning method based on vision-tactile fusion. This method trains the robotic arm's autonomous operation strategy by integrating visual information, tactile information, and proprioceptive information. While the above solutions can enhance the robot's skill learning or autonomous operation capabilities, their optimization focuses primarily on the robot's control strategy itself, and they have not yet established a corresponding analytical relationship between robot feedback data and trainee operational capabilities for vocational training scenarios. In actual training, multi-source data such as joint torque, end effector force perception, vision, inertial measurement unit, and encoder can reflect the sources of deviations in the trainee's operation process. However, existing training systems typically evaluate based only on result indicators such as final pose error and task completion time, making it difficult to determine whether operational errors specifically stem from inaccurate tool coordinate system establishment, insufficient singular posture avoidance capability, insufficient joint limit prediction, or substandard force perception coordination. Due to the lack of an attributable diagnostic mechanism for the operational process, subsequent training tasks are prone to deviating from the trainee's actual skill deficiencies, leading to repeated trial and error on similar mistakes. Low-quality operational trajectories and parameter data entering the knowledge base can also affect the continuous optimization of training paths and increase the risks of collisions, joint overspeed, and equipment wear. Therefore, there is an urgent need to provide a closed-loop optimization method for practical training that can diagnose the trainee's operational process, attribute skill deficiencies, update the skill profile, and adaptively push training tasks using multi-source sensor data from the robot itself. This method aims to achieve the technical effects of accurate identification of trainee skill deficiencies, targeted improvement of training tasks, improved training data quality, and reduced wear and tear on practical training equipment. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a robot feedback-driven closed-loop optimization method for practical training. This method solves the problems of traditional methods, such as failing to fully utilize multi-source sensor data from the robot itself to analyze the sources of deviations during trainee operation, making it difficult to accurately attribute skill deficiencies, resulting in training tasks that are difficult to match with trainees' actual weaknesses, and thus affecting training efficiency, training data quality, and the safety of training equipment.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A robot feedback-driven training closed-loop optimization method includes: S1. After the industrial robot training task is started, robot joint status data, end effector feedback data, training area visual data and teach pendant input events are collected synchronously, and the data timing correspondence under the same training task is established. S2. Based on the teach pendant input events and the robot's motion state, the trainee's operation process is divided into multiple continuous operation segments, and the corresponding training state for each operation segment is determined. S3. Associate each operation segment with a preset skill sub-item, and determine the operation deviation type based on the data timing corresponding to each operation segment; S4. Based on the type of operational deviation, training status and corresponding operational segments, establish the attribution relationship between trainees' operational deviations and skill deficiencies, and update the trainees' skill status records according to the attribution relationship. S5. Match the training task template according to the updated student skill status record, generate the next round of training tasks, and write the training data of this round into the training knowledge base after quality screening.

[0005] Furthermore, after the industrial robot training task is started, robot joint status data, end effector feedback data, training area visual data, and teach pendant input events are collected synchronously, and a data time sequence correspondence relationship under the same training task is established, including: When establishing the data time sequence correspondence, the unified clock of the edge computing unit is used as the reference, and timestamps are added to the joint status data, end feedback data, training area visual data and teach pendant input events; The original timestamps are corrected based on the time deviation, and a unified time index is established; Set validity and quality tags for the data, retain the operation context through sliding cache, and fill in short-term missing data. When critical data sources are continuously missing, pause the attribution processing of the corresponding segments.

[0006] Furthermore, based on the teach pendant input events and the robot's motion state, the trainee's operation process is divided into multiple continuous operation segments, including: When dividing consecutive operation segments, a unified time index data stream is used as input, and candidate segment boundaries are extracted from teach pendant input events. The motion change boundary is extracted based on the fluctuations of joint velocity, end-effector velocity, end-effector force, and end-effector attitude within the sliding window; After performing temporal proximity matching between the candidate segment boundary and the motion change boundary, it is determined to be a composite boundary, event boundary, or implicit boundary. Then, continuous operation segments are formed in chronological order, and operation segments with a duration less than the preset duration and no abnormal motion events are merged with adjacent operation segments.

[0007] Furthermore, the training status corresponding to each operation segment is determined, including: When determining the training status corresponding to each operation segment, a training status table is established. Candidate training statuses are first determined based on the teaching pendant input events within the segment, and then the training status is determined by combining the joint velocity distribution, end-effector velocity direction, end-effector force change, and visual proximity relationship. The manual operation state in joint mode is determined based on the dominant motion characteristics of a single joint, while the contact operation state is determined based on the resultant force value of the end effector, the change in the end effector velocity, and the proximity relationship between the robot end effector and the work object. When multiple training states simultaneously meet the judgment criteria, state confidence is generated based on event consistency, motion feature consistency, and visual relationship consistency.

[0008] Furthermore, each operation segment is associated with a preset skill sub-item, including: When associating skill sub-items, configure skill sub-items according to coordinate establishment, point teaching, motion control, contact operation, safety prediction and rhythm control, and establish a training task flow table; Matching is performed based on the time and location of the operation segment, the training status, the teaching pendant input events, and the dominant motion characteristics. When multiple skill sub-items are matched simultaneously, the primary skill sub-item, auxiliary skill sub-item, or parallel relationship is determined.

[0009] Furthermore, the type of operational deviation is determined based on the data timing sequence corresponding to each operational segment, including: When determining the type of operational deviation, deviation identification features are extracted according to the data time series corresponding to the associated skill sub-items. The deviation identification features are then compared with the reference range established according to the training task type, robot model, end-effector type, and difficulty level to determine the deviation level. Process thresholds are used to constrain force control, joint limit and unusual posture related features. When the conclusions corresponding to the reference range are inconsistent with the conclusions corresponding to the process thresholds, the operation deviation type is determined according to the safety-related process thresholds. When the data quality is lower than the judgment requirements, candidate operation deviation types are generated.

[0010] Furthermore, based on the type of operational deviation, training status, and corresponding operational segments, an attribution relationship between trainees' operational deviations and skill deficiencies is established, including: When establishing attribution relationships, unusable data is removed, the skill defect rule table is invoked, and joint attribution is performed by combining the type of operational deviation, the applicable training state, and the context of the segment. First, identify non-student factor abnormal segments based on exclusion criteria. Then, generate attribution confidence based on core bias characteristics, supporting evidence conditions, and attribution weights. When multiple rules are hit simultaneously, determine the primary attribution, secondary attribution, or parallel attribution.

[0011] Furthermore, the learner skill status record is updated based on attribution relationships, including: When updating trainees' skill status records, a skill sub-item status table is used to record the skill sub-item identifier, mastery value, update time, confidence interval, recent deviation type, recent attribution frequency, and key training mark. The mastery value is adjusted down based on effective negative attribution and adjusted up based on continuous normal performance. Low-quality candidate attributions are not included in the mastery value update. When the cumulative change exceeds the limit, the automatic update of the corresponding skill sub-item is frozen, and a key training mark is set when there are consecutive negative attributions.

[0012] Furthermore, based on the updated trainee skill status records, a training task template is matched to generate the next round of training tasks, including: When matching training task templates, the training task template library is called. The training task template library records skill sub-item tags, task type, difficulty level, parameter range and applicable conditions. Based on the mastery value, attribution confidence, and key training markers in the trainees' skill status records, target training items are determined, and the task difficulty, task parameters, and prompt levels are adjusted in combination with skill dependencies to generate the next round of training tasks.

[0013] Furthermore, the training data from this round underwent quality screening and was then written into the training knowledge base, including: When screening the training data for this round, the operation segments are written into the training sample area, error case area, manual review area or anomaly exclusion area based on the deviation level, data validity label, state confidence, attribution results and exclusion conditions. When high-quality samples are automatically screened or manually reviewed and confirmed to meet the update conditions, the reference range boundary values ​​and training task template parameters in the training knowledge base are incrementally updated.

[0014] Compared with existing technologies, this invention provides a robot feedback-driven training closed-loop optimization method, which has the following beneficial effects: 1. This invention synchronously collects robot joint state data, end effector feedback data, training area visual data, and teach pendant input events, and establishes a data time sequence correspondence under the same training task. It divides the trainee's continuous operation process into analyzable operation segments, and then combines training status, skill sub-items, and corresponding data time sequences to identify operation deviation types. This expands the evaluation process from outcome indicators such as final pose error and completion time to operation process data. By combining operation deviation types, corresponding operation segments, and training status for skill defect attribution, and updating trainee skill status records, matching the next round of training tasks, and filtering training data for writing into the training knowledge base based on the attribution results, this invention reduces problems such as mismatch between training tasks and trainees' actual weaknesses, low-quality data interfering with knowledge base updates, collisions, joint overspeeding, and equipment wear. This improves the accuracy of trainee skill defect identification, the relevance of training tasks, the reliability of training data, and the safety of equipment operation.

[0015] 2. This invention, by retaining operation segment indexes, teach pendant event sequences, sensor data slices, deviation identification features, attribution confidence levels, and supporting evidence lists during practical training, and by classifying normal segments, error case segments, manually reviewed segments, and anomaly elimination segments according to their uses and writing them into the practical training knowledge base, enables practical training evaluation to correspond to specific operation segments and data evidence. Instructors can review the students' operation processes in coordinate system establishment, point confirmation, contact control, and limit avoidance based on the corresponding operation segments, reducing the reliance on on-site observation and subjective experience for evaluation, and facilitating the review, archiving, and teaching application of typical errors, thereby improving the traceability of practical training evaluation, the efficiency of teaching review, and the consistency of evaluation standards. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the training closed-loop optimization method of robot feedback driven by the present invention. Figure 2 This is a schematic diagram illustrating the multi-source data acquisition and unified time indexing of the present invention; Figure 3 This is a schematic diagram illustrating the operation segment division and training status determination of the present invention; Figure 4 This is a schematic diagram illustrating the association between skill sub-items and the identification of operational deviations in this invention; Figure 5 This is a schematic diagram of the skill defect attribution and training task closed-loop update of the present invention; Figure 6 This is a schematic diagram of the shaft and hole assembly training scenario of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: Figures 1-5 A feedback-driven closed-loop optimization method for robot training is presented, including: S1. After the industrial robot training task is started, robot joint status data, end effector feedback data, training area visual data and teach pendant input events are collected synchronously, and the data timing correspondence under the same training task is established. S2. Based on the teach pendant input events and the robot's motion state, the trainee's operation process is divided into multiple continuous operation segments, and the corresponding training state for each operation segment is determined. S3. Associate each operation segment with a preset skill sub-item, and determine the operation deviation type based on the data timing corresponding to each operation segment; S4. Based on the type of operational deviation, training status and corresponding operational segments, establish the attribution relationship between trainees' operational deviations and skill deficiencies, and update the trainees' skill status records according to the attribution relationship. S5. Match the training task template according to the updated student skill status record, generate the next round of training tasks, and write the training data of this round into the training knowledge base after quality screening.

[0019] This method is applicable to industrial robot teaching, assessment, and training scenarios, especially for practical training tasks of six-DOF industrial robots such as handling, palletizing, trajectory teaching, shaft and hole assembly, electrical terminal insertion, and welding path teaching. Participants include at least trainees performing robot operation training, the industrial robot executing the training actions, a teach pendant inputting operation commands, sensing devices collecting robot operating status and training area status, and local computing devices performing data processing. Input information includes at least robot joint state data, end effector feedback data, training area visual data, and teach pendant input events. Robot joint state data includes at least joint position, joint velocity, and joint torque; end effector feedback data includes at least end effector force perception, end effector posture, and end effector motion state; and training area visual data includes at least robot... Image information of the human end effector, workpiece, fixture, and obstacles; teaching pendant input events include at least task initiation, motion mode switching, coordinate system switching, parameter modification, point confirmation, pause, and emergency stop; by uniformly associating the above data with time sequence, dividing operation segments, determining training status, associating skill sub-items, identifying operation deviations, and attributing skill defects, the student's skill status record is updated, and the next round of training tasks is generated based on the updated student skill status record; the output information includes at least the operation segment sequence, training status identifier, skill sub-item association result, operation deviation type, skill defect attribution result, student skill status record, next round of training task, and filtered training data, thus forming a continuous closed loop of data acquisition, process division, deviation identification, defect attribution, task generation, and knowledge base update.

[0020] Specifically, such as Figure 2 As shown: When the teach pendant generates a task start event, or when the training terminal completes student identity verification and issues a training task start command to the robot controller, the edge computing unit uses a unified clock as the time reference and adds timestamps to the robot joint state data, end effector feedback data, training area visual data, and teach pendant input events respectively. The robot joint state data includes joint position, joint velocity, and joint torque. The joint position and joint velocity are obtained by the robot joint encoder, with a sampling frequency set to 500 Hz to 1000 Hz. The joint torque is obtained by feedback data from the joint torque sensor or servo driver, with a sampling frequency set to 1000 Hz. The reason for setting the above frequencies is that the process changes such as sudden changes in joint velocity, load changes, emergency stop response, and collision approach during industrial robot training are mainly concentrated below 200 Hz. Sampling at 500 Hz to 1000 Hz can retain sufficient sampling margin for changes in this frequency band and meets the high-frequency sampling requirements in industrial robot state monitoring. The end-effector feedback data includes end-effector force data, end-effector posture data, and end-effector inertial data. The end-effector force data is acquired by a six-dimensional force sensor installed between the robot's end flange and the actuator, with a sampling frequency set to 500 Hz. This setting is based on the fact that effective contact force changes during shaft-hole assembly, electrical terminal connection, and teach contact training typically concentrate in the 50 Hz to 150 Hz range, and 500 Hz sampling can meet the needs of contact process identification and force coordination analysis. The end-effector posture data is obtained from joint encoder data through robot forward kinematics conversion, or by auxiliary correction from an inertial measurement unit installed near the end flange. The sampling frequency for the inertial measurement data is set to 200 Hz. This setting is based on the fact that end-effector posture changes and low-frequency vibrations during practical training operations typically concentrate in the 20 Hz to 80 Hz range, and 200 Hz sampling can cover this frequency band and reduce the real-time processing load on the edge computing unit. The visual data for the training area is acquired by industrial cameras deployed around the training workstations, with a frame rate set to 30 to 60 frames per second and an image resolution of no less than 1280 x 720 pixels. The rationale for these frame rate and resolution settings is that in common training tasks such as handling, trajectory teaching, and assembly training, the average movement speed of the robot's end effector is usually no more than 300 millimeters per second. When acquiring data at 60 frames per second, the end effector displacement between two adjacent frames is approximately 5 millimeters. Within the common field of view of the training workstations, which ranges from 1.0 to 2.0 meters, an image resolution of no less than 1280 x 720 pixels can provide sufficient spatial resolution for the robot's end effector, workpiece edges, and obstacle contours to meet the requirements for position recognition and proximity judgment. The teach pendant input events are event-driven data. Each event generates an event identifier, event time, event type, and corresponding operation content. Event types include task start, motion mode switching, coordinate system switching, parameter modification, teaching point confirmation, pause, and emergency stop. When establishing the data timing correspondence, the edge computing unit periodically acquires the time deviation between the local time of each data source and the unified clock, corrects the original timestamp t of each data source to t plus the time deviation Δt, and sets the correction period to once every 1 second. The reason for setting this period is that the 1-second period can limit the cumulative clock drift of each data source and will not significantly occupy the internal communication bandwidth of the training workstation. When the internal network of the training workstation supports precise time synchronization, the alignment error of each data source is controlled within 100 microseconds. When using ordinary industrial Ethernet, fieldbus or serial port for acquisition, the alignment error of each data source is controlled in the millisecond range, preferably not greater than 1 millisecond. The above alignment accuracy can meet the time proximity matching requirements between the teach pendant input event and the robot motion state in the subsequent operation segment division process, and avoid the same operation process being incorrectly divided due to data timing misalignment. After completing the timestamp correction, a unified time index is established according to the corrected timestamps, and data from different sources are written into the data sequence of the same training task. Each record in the data sequence includes at least the following fields: task identifier, timestamp, data source identifier, data value, validity flag, and quality flag. The validity flag is used to distinguish between normal data, short-term missing data, and abnormal data. The quality flag is used to indicate whether the data falls within the sensor range, the robot's rated motion range, and the allowable range of the training task. When the collected value exceeds the corresponding sensor range, exceeds the joint soft limit range fed back by the robot controller, or the change amplitude of adjacent sampling points exceeds the maximum allowable change rate of the device, the corresponding data is marked as abnormal data. To preserve the context of the trainee's operation, a sliding cache is used to save data for a period of time prior to the current moment. The cache window is set to 20 to 40 seconds, preferably 30 seconds. The reason for this window length is that a single basic operation segment in industrial robot training typically lasts 5 to 20 seconds, and a 30-second window can cover the entire operation segment while preserving the context data before and after the operation switch. When the continuous missing time of a single data source is less than 100 milliseconds, the joint position, joint velocity, joint torque, end effector force, and inertial measurement data are temporarily supplemented by linear interpolation of adjacent valid data or by keeping the previous valid value, and the corresponding time period is marked as low-quality data. Teach pendant input events are not interpolated and are written only according to the actual event time. When visual data is missing for a short time, the target position in the previous valid frame is retained and marked as low-quality visual data. 100 milliseconds is less than the duration of most basic operation segments and can cover common anomalies such as short-term communication jitter or single-frame visual loss. The key data sources are joint status data, end effector feedback data, visual data of the training area, and teaching pendant input events, which are involved in the segmentation of operation segments, determination of training status, and attribution of operation deviations. If any key data source is missing for more than 500 milliseconds, operation deviation attribution within the missing time period and the buffer intervals before and after it is paused. Only the data acquisition records and safety prompts are retained until the data is recovered or the coach confirms that training can continue. 500 milliseconds covers multiple visual sampling cycles and multiple joint status and end effector force sampling cycles. Continuing to perform deviation attribution may result in the loss of temporal evidence. After the above processing is completed, a unified time-indexed data stream is formed and used as the input for subsequent segmentation of operation segments and determination of training status.

[0021] Specifically, such as Figure 3As shown: Using the aforementioned unified time index data stream as input, when the cumulative duration of the data in the sliding buffer window reaches the minimum segment analysis duration, the operation segment division is initiated; the minimum segment analysis duration is set to 3 seconds. The reason for this setting is that during the teaching process of industrial robots, analyzable basic operations such as approaching and confirming the point, adjusting the posture before and after coordinate switching, and short-distance movement usually last for more than 2 seconds. Using 3 seconds as the analysis starting point can cover the complete basic actions and avoid identifying instantaneous jitter or brief button mis-touches as independent operation segments; When segmenting, candidate segment boundaries are first extracted from the teach pendant input events. These candidate segment boundaries correspond to events such as task initiation, motion mode switching, coordinate system switching, speed scaling adjustment, parameter modification, teach point confirmation, pause, and emergency stop. Then, motion change boundaries are extracted based on the robot's motion state. The robot's motion state includes joint velocity changes, end-effector velocity changes, end-effector force changes, and end-effector posture changes. The end-effector velocity is calculated from joint position, joint velocity, and the robot's kinematic model, and can also be assisted by continuous changes in the robot's end-effector position in the visual data of the training area. When extracting the boundary of motion change, the square root of the sum of the squares of the absolute values ​​of each joint velocity is used to obtain the composite value of the joint velocity; the square root of the sum of the squares of the three translational velocities of the end effector is used to obtain the composite value of the end effector velocity; and the square root of the sum of the squares of the three forces of the end effector is used to obtain the composite value of the end effector force. The fluctuation of the above composite values ​​is calculated within a sliding window of 100 to 200 milliseconds, and the fluctuation is the standard deviation of the composite value within the corresponding sliding window. The reason for choosing this window length is that, at the aforementioned sampling frequency, a 100-millisecond window can cover multiple joint state sampling points, end effector force sampling points, and inertial measurement sampling points, while a 200-millisecond window can improve noise resistance while still preserving the time characteristics of the student's operation state switching. The sliding window step size is set to half of the window length to balance continuous monitoring and computational load control. The static noise level is obtained by collecting data for no less than 2 seconds in the waiting state before the start of the training task, and is judged based on 1 to 2 times the standard deviation of the corresponding composite value. When the fluctuation ratio of adjacent sliding windows is greater than 2.0 or less than 0.5, and this change occurs consecutively for at least three windows, a motion change boundary is determined to exist near the corresponding time point. The reason for using 2.0 and 0.5 as relative ratios is that different trainees' operating speeds, different training task loads, and different tool installation states will lead to differences in motion amplitude, and the relative ratio can adapt to different training intensities. Three consecutive windows can eliminate single sampling spikes, single-frame visual disturbances, and short-term communication jitter. Before making the ratio judgment, the fluctuation of the current window must be higher than the static noise level. When identifying segment boundaries, candidate segment boundaries are matched with motion change boundaries based on temporal proximity. If the time difference between the two is no more than 200 milliseconds, the location is identified as a composite boundary. If only candidate segment boundaries exist, the location is identified as an event boundary. If only motion change boundaries exist, and the difference in motion state before and after the boundary lasts for more than 300 milliseconds, the location is identified as a latent boundary identified by motion state changes. The reason for setting 200 milliseconds is that after a student triggers a command through the teach pendant, the robot controller response and mechanical motion response usually have a delay of tens to hundreds of milliseconds. 200 milliseconds can cover common control response delays and reduce incorrect matching. 300 milliseconds is longer than the ordinary control response delay and the duration of short-term jitter, which can avoid misjudging transient disturbances as real segment boundaries. After determining the segment boundaries, continuous operation segments are generated in chronological order. Each operation segment includes at least the following fields: segment start point, segment end point, boundary type, data slices within the segment, teach pendant event sequence within the segment, and data quality markers, and maintains a corresponding relationship with the aforementioned unified time index. For segments with a duration of less than 1 second, except for segments containing emergency stop events, collision approach events, or abnormal motion events, they are merged with adjacent operation segments. 1 second is shorter than the duration of most independently evaluable basic operations, and is usually difficult to support stable training state recognition and operation deviation judgment. Abnormal motion events include sudden changes in joint velocity, sudden increases in end-effector force, sudden changes in end-effector trajectory direction, emergency stop triggering, or situations where the distance between the end-effector and an obstacle is less than the safe distance. When determining the training state, a training state table is established. The training states can be selected as coordinate system establishment state, point teaching state, joint mode manual operation state, rectangular coordinate mode manual operation state, trajectory reproduction state, contact operation state, parameter adjustment state, waiting state, and abnormal pause state. The state determination adopts a hierarchical logic of event first and motion second. First, candidate training states are determined based on the teaching pendant input events within the segment. When coordinate system switching and coordinate point confirmation events occur, the coordinate system establishment state or point teaching state is added to the candidate training states. When the motion mode switches to joint mode, the joint mode manual operation state is added to the candidate training states. When end-effector force changes, the contact operation state is added to the candidate training states. When pause or emergency stop events occur, the waiting state or abnormal pause state is added to the candidate training states. Then, by combining the joint velocity distribution, end-effector velocity direction, end-effector force changes, and visual proximity relationship, the training state corresponding to the current operation segment is determined from the candidate training states. Visual proximity is determined by the minimum distance between the robot's end effector position and the edge of the workpiece, fixture, or target work area. When this minimum distance is less than 50 mm to 100 mm, the robot's end effector is considered to be approaching the work object. The rationale for this distance range is that 50 mm is suitable for proximity judgment in delicate operations such as shaft hole assembly and electrical terminal insertion, while 100 mm is suitable for proximity judgment in general operations such as handling, palletizing, and trajectory teaching. The reduction in end effector speed is determined by the composite end effector speed value decreasing by more than 30% relative to the average end effector speed in the 300 milliseconds before contact, or by being lower than 50% of the peak speed of the current segment. 300 milliseconds can cover the short-term motion trend before contact, and the speed change ranges of 30% and 50% can eliminate misjudgments caused by normal speed fine-tuning. During the state determination process, the manual operation state in joint mode is identified by the percentage of single-joint velocity. This percentage is the ratio of the absolute value of the maximum single-joint velocity to the sum of the absolute values ​​of all joint velocities. When this ratio is higher than 70%, the segment is determined to have single-joint dominant motion characteristics. The reason for setting 70% is that this ratio is significantly higher than the velocity distribution level during balanced multi-joint linkage, which can distinguish between manual operation in joint mode and multi-joint cooperative motion in Cartesian coordinate mode. The manual operation state in Cartesian coordinate mode is identified by the synchronous movement of multiple joints, the continuity of the Cartesian motion direction at the end effector, and the smoothness of the end effector posture change. The synchronous movement of multiple joints is identified by the consistency of the velocity change direction of at least two joints within a continuous window, and the continuous Cartesian motion direction at the end effector is identified by the consistency of the velocity change direction of at least two joints within a continuous window. The identification is based on the angle between the velocity directions of adjacent windows being less than 30 degrees, and the identification is based on the change in attitude angle between adjacent windows being less than 5 degrees. The reason for setting 30 degrees and 5 degrees is that manual teaching allows for small adjustments in direction and attitude, but there should be no large jumps during continuous movement. The contact operation state is identified by the combined value of the end force and the change in end velocity. When the combined value of the end force is higher than 5 Newtons for no less than 100 milliseconds, accompanied by a decrease in end velocity or the end position is close to the workpiece area, the contact operation state is taken as the training state of the corresponding segment. 5 Newtons is higher than the zero drift and slight contact noise level of common six-dimensional force sensors, but lower than the normal working force range of most light-load training contact operations. 100 milliseconds can eliminate misjudgment caused by a single impact spike. If multiple training states simultaneously meet the judgment criteria, a state confidence score is generated based on the consistency of the teaching pendant input events, motion features, and visual proximity relationships. The state confidence score is obtained by weighting the event consistency, motion feature consistency, and visual relationship consistency, where the event consistency weight can be set to 0.4, the motion feature consistency weight can be set to 0.4, and the visual relationship consistency weight can be set to 0.2. The values ​​of event consistency, motion feature consistency, and visual relationship consistency all range from 0 to 1, with 1 for fully conforming to the training state table rules, 0.5 for partially conforming, and 0 for not conforming. When visual data is missing, the visual relationship consistency weight is incorporated into the motion feature consistency weight. When the state confidence score is below 0.7, the corresponding operation segment is marked as a segment to be reviewed, and the weight of this segment is reduced in the attribution process. The reason for setting 0.7 is that this threshold requires at least two types of evidence to be consistent, which can reduce the impact of single data source errors on the training state judgment. After the above processing is completed, a sequence of continuous operation segments with training status identifiers is formed. Each operation segment is temporally associated with the corresponding data slice and teach pendant event sequence, and serves as input for skill sub-item association and operation deviation type determination.

[0022] Specifically, such as Figure 4 As shown: Using the aforementioned continuous operation segment sequence, training status identifiers, and data slices corresponding to each operation segment as input, when the training status is confirmed in any operation segment and the key data used for skill judgment within that operation segment meets the validity requirements, the skill sub-item association and operation deviation type determination are performed. The key data meeting the validity requirements means that the joint state data, end-effector feedback data, training area visual data, and teach pendant input events used to judge the corresponding skill sub-item within that operation segment do not have data missing for more than 500 milliseconds consecutively, and the training status confidence of that operation segment is not lower than 0.7. 500 milliseconds is consistent with the aforementioned key data source missing handling threshold, used to avoid deviation judgments in the event of missing multiple consecutive visual sampling cycles, joint state sampling cycles, and end-effector force sampling cycles. 0.7 indicates that at least two types of evidence in the teach pendant input event, robot motion characteristics, and visual relationship are consistent. When the training status confidence is lower than 0.7 or low-quality data is marked, only candidate skill sub-items and candidate operation deviation types are generated. The skill sub-items are categorized according to the operational capabilities that can be observed and evaluated through sensor data during practical training. They can be selected into nine categories: tool coordinate system establishment, workpiece coordinate system calibration, manual operation in joint mode, manual operation in rectangular coordinate mode, teach point confirmation, singular posture avoidance, joint limit prediction, force-sensory coordination control, and cycle control. These sub-items cover the core training content in industrial robot practical training, such as coordinate establishment, point teaching, motion control, contact operation, safety prediction, and cycle optimization. All of the above skill sub-items can be represented through the aforementioned joint state data, end effector feedback data, visual data of the training area, and teach pendant input events, enabling trainees to identify their operational capabilities from continuous process data. When associating skill sub-items, a training task flow table is first established. This table includes at least the following fields: task stage number, stage start and end conditions, allowed training states, target operation content, corresponding skill sub-item, allowable error range, and safety constraints. The allowable error range is determined based on the training task template, robot repeatability accuracy, workpiece dimensional tolerance, and instructor configuration parameters. When allowable error ranges from multiple sources are inconsistent, the stricter range between the training task safety constraints and workpiece dimensional tolerance is used as the current allowable error range. Subsequently, the time position, training state, teach pendant input event, and dominant motion characteristics within the operation segment are matched with the training task flow table. When the time position, training state, and teach pendant input event all match the same task stage, the skill sub-item corresponding to that task stage is identified as the associated skill sub-item for that operation segment. When only some conditions match, candidate skill sub-items are generated and assigned matching weights. In coordinate system establishment mode, if multiple positioning point confirmation events occur within an operation segment, the operation segment is associated with tool coordinate system establishment or workpiece coordinate system calibration. In joint mode manual operation mode, if a single joint dominates motion features within an operation segment, the operation segment is associated with joint mode manual operation. In Cartesian coordinate mode manual operation mode, if multi-joint cooperative motion and continuous changes in the end effector's Cartesian direction occur within an operation segment, the operation segment is associated with Cartesian coordinate mode manual operation. In contact operation mode, if continuous changes in end effector force accompanied by changes in end effector velocity occur within an operation segment, the operation segment is associated with force-sensing coordination control. When a robot joint approaches a soft limit or the end effector's posture approaches a singular posture region within an operation segment, the operation segment is associated with joint limit prediction or singular posture avoidance. When an operation segment matches multiple skill sub-items simultaneously, the primary skill sub-item and auxiliary skill sub-item are determined based on the proportion of dominant events and dominant data features within the segment. The proportion of dominant events is the ratio of the number of events corresponding to the target skill sub-item to the total number of teach pendant input events within the segment. The proportion of dominant data features is the ratio of the degree of feature satisfaction corresponding to the target skill sub-item to the sum of the degree of feature satisfaction corresponding to all candidate skill sub-items. The degree of feature satisfaction is determined based on whether the feature meets the valid judgment conditions of the corresponding skill sub-item: 1 for complete satisfaction, 0.5 for partial satisfaction, and 0 for non-satisfaction. When the proportion of dominant events or dominant data features is higher than 60%, ... The corresponding skill sub-items are designated as primary skill sub-items, with a weight ranging from 0.7 to 0.9. The weight of auxiliary skill sub-items is 1 minus the weight of the primary skill sub-item. When multiple auxiliary skill sub-items exist, the remaining weights are allocated according to the matching degree of each auxiliary skill sub-item. The reason for setting 60% is that this ratio can indicate that the segment is mainly dominated by a certain type of operational ability, while retaining the influence of other auxiliary abilities on the segment's performance. When the difference in matching weights between two skill sub-items is less than 0.1, it indicates that the matching degree of the two skill sub-items is close. Forcibly distinguishing between primary and secondary sub-items can easily lead to misattribution. Therefore, parallel associations are retained, and the results of parallel associations are passed to subsequent attribution processing. After determining the skill sub-item, deviation identification features are extracted based on the corresponding data time series. The tool coordinate system establishes corresponding deviation identification features, which include at least the fields of multiple positioning point coordinate dispersion, calibration point residual, and end-effector stabilization time before confirmation. The multiple positioning point coordinate dispersion is determined by the standard deviation or maximum distance difference of multiple confirmation coordinates of the same positioning point. The calibration point residual is determined by the distance difference between the tool center point calculated by calibration and the actual positioning point. The end-effector stabilization time before confirmation is determined by the duration during which the end-effector velocity is continuously lower than the stabilization velocity threshold before the teaching point confirmation event. The stabilization velocity threshold can be set from 5 mm / s to 10 mm / s. 5 mm / s is suitable for fine training such as shaft hole assembly and electrical terminal insertion, while 10 mm / s is suitable for general training such as handling and palletizing. This range is lower than the common teaching approach speed and can characterize the end-effector entering a stationary or quasi-stationary state. The deviation identification features corresponding to the workpiece coordinate system calibration include at least the fields of three-point positioning residual, coordinate axis angle deviation, and coordinate system transformation matrix orthogonality deviation. The three-point positioning residual is determined by the fitting deviation of the three reference points during the workpiece coordinate system establishment process. The coordinate axis angle deviation is determined by the angle difference between the actual established coordinate axis and the target coordinate axis. The coordinate system transformation matrix orthogonality deviation is determined by the degree of deviation of the product of the rotation matrix and its transpose matrix relative to the identity matrix. The deviation identification features corresponding to manual operation in joint mode include at least the fields of target joint velocity stability, non-target joint disturbance, and the number of joint velocity mutations. The target joint velocity stability is determined by the standard deviation of the target joint velocity sequence or the rate of change of velocity between adjacent sampling points. The non-target joint disturbance is determined by the maximum value of the angle changes of other joints during the target joint's movement. The number of joint velocity mutations is determined by the number of times the rate of change of joint velocity exceeds the corresponding reference range. The deviation identification features corresponding to manual operation in Cartesian coordinate mode include at least the fields of end-effector trajectory smoothness, end-effector orientation retention, and the degree of multi-joint cooperative change. The end-effector trajectory smoothness is determined by the mean of the second-order difference modulus of the end-effector position sequence. The end-effector orientation retention is determined by the... The degree of fluctuation of the angle between the velocity directions at the ends of adjacent windows is determined, and the degree of multi-joint coordinated change is determined by the consistency of the velocity change directions and timing of multiple joints. The deviation recognition features corresponding to the teaching point confirmation include at least the following fields: residual velocity value before confirmation, attitude fluctuation value before confirmation, and attitude consistency between adjacent confirmation points. The residual velocity value before confirmation is determined by the average of the synthesized end velocity values ​​within 300 milliseconds before the teaching point confirmation, the attitude fluctuation value before confirmation is determined by the change in end attitude angle within 300 milliseconds before confirmation, and the attitude consistency between adjacent confirmation points is determined by the difference in end attitude angle between adjacent teaching points. 300 milliseconds can cover the short-term stabilization process before the student presses the confirmation button and connects with the short-term state change judgment in the aforementioned segment boundary recognition. The deviation identification features for singular posture avoidance include at least the following fields: Jacobian matrix condition number change, joint velocity mutation events, and end-effector trajectory detour. The Jacobian matrix condition number is calculated from the Jacobian matrix corresponding to the robot's current joint position. Joint velocity mutation events are determined by the number of times the joint velocity change rate exceeds the reference range. The end-effector trajectory detour is determined by the deviation distance between the actual end-effector trajectory and the target path. The deviation identification features for joint limit prediction include at least the following fields: minimum distance from the joint position to the soft limit, velocity direction when approaching the soft limit, and backtracking delay. The minimum distance from the soft limit is determined by the difference between the current joint position and the corresponding soft limit position. The velocity direction is determined by whether the current joint velocity sign is consistent with the direction of approaching the soft limit. The backtracking delay... The time difference between entering the soft limit buffer and the first occurrence of reverse motion is determined; the deviation identification features corresponding to force coordination control include at least the fields of end force composite value, force direction change, force change rate, and end velocity decay relationship, wherein the force change rate is obtained by dividing the difference of end force composite values ​​at adjacent sampling times by the time interval, and the end velocity decay relationship is determined by the ratio of end velocity change before and after contact; the deviation identification features corresponding to beat control include at least the fields of waiting time ratio, effective motion time ratio, and repeated retreat number, wherein the waiting time ratio is determined by the ratio of the cumulative waiting time to the total task time, the effective motion time ratio is determined by the ratio of the cumulative effective motion segment time to the total task time, and the repeated retreat number is determined by the number of times the end trajectory repeatedly moves away from and approaches the same target area; Operational deviation types are determined based on the relationship between deviation identification features and reference ranges. The reference ranges are derived from sample data generated by experienced operators completing similar training tasks. Experienced operators are defined as those with at least three years of experience in operating, debugging, or teaching industrial robots, or those holding corresponding vocational skill level certificates. Reference ranges are established separately for training task type, robot model, end effector type, and difficulty level to avoid mixing feature values ​​under different task conditions. The sample size for each type of operational segment is no less than 30 groups. The reason for using 30 groups as the lower limit for the sample size is that this number can meet the basic stability requirements for ranking and statistically analyzing similar operational segments, while also considering sample collection costs. For sample characteristics… After sorting the eigenvalues ​​from smallest to largest, the sample value range at the 25th to 75th percentile is defined as the normal range, the sample value ranges at the 10th to 25th percentile and the 75th to 90th percentile are defined as the slightly biased range, and the sample value ranges below the 10th percentile or above the 90th percentile are defined as the significantly biased range. The reason for using sorted position ranges is that different robot models, different task difficulties, and different end-effector loads will lead to differences in the absolute values ​​of the features, and sorted position ranges can stably represent the relative degree of deviation in the same type of task. The type of operational deviation is jointly determined by the deviated skill sub-item and the deviation identification feature, and the deviation level is used to distinguish between normal, slightly biased, and significantly biased deviations. For safety-related characteristics, process thresholds are used for constraint. When the end force composition value exceeds 50 Newtons and the force change rate exceeds 200 Newtons per second in light-load assembly training, the deviation type is marked as force control abnormality. 50 Newtons is higher than the target contact force range of 10 to 30 Newtons in most light-load assembly and insertion training. 200 Newtons per second is used to identify rapid impact loading processes and can be distinguished from normal slow loading processes. When the joint distance from the soft limit is less than 5% of the joint's range of motion and the current joint velocity direction still points to the corresponding soft limit, the deviation type is marked as insufficient limit prediction. 5% serves as a safety buffer range before the soft limit, which can reserve space for retreat or deceleration before the trainee continues to operate. Using the aforementioned sliding window setting, when the Jacobian matrix condition number exceeds 500 and shows an upward trend within at least 3 consecutive sliding windows, the deviation type is marked as singular posture approach risk. 500 is lower than the typical singular region condition number threshold of 1000, which is suitable as an early identification threshold in teaching and training. At least 3 consecutive sliding windows can eliminate instantaneous numerical fluctuations. When the reference range judgment result is inconsistent with the process threshold judgment result, the safety-related process threshold is used first to determine the operation deviation type, and the deviation degree of the corresponding feature relative to the reference range is recorded simultaneously. When there are low-quality data markers in the operation segment, the final operation deviation type is not directly generated, but candidate operation deviation types are generated and data quality weights are attached. If there is only short-term missing data or visual low quality, the data quality weight is 0.7 to 0.8. If there are multiple sources of slight missing data or low confidence in the training status, the data quality weight is 0.5 to 0.7. The reason for setting this weight range is that low-quality data can still reflect some trends, but it should not have the same judgment effect as complete high-quality data. After completing the above processing, the output information includes at least the skill sub-item identifier, corresponding data segment, deviation identification feature value, deviation level, and operation deviation type, and serves as the input for establishing the attribution relationship between trainees' operation deviations and skill defects.

[0023] Specifically, such as Figure 5 As shown: Using the aforementioned skill sub-item association results, operation deviation type, deviation identification feature value, deviation level, training status, and operation segment boundary information as input, when the operation segment has completed skill sub-item association, generated operation deviation type, and is not marked as unusable data, attribution processing between student operation deviation and skill deficiency is performed; unusable data refers to data that has been continuously missing for more than 500 milliseconds, or data quality weight is less than 0.5; 500 milliseconds is consistent with the aforementioned key data source missing processing threshold, which can avoid forming attribution conclusions when multiple consecutive visual sampling cycles, joint state sampling cycles, and end force sampling cycles are missing; a data quality weight less than 0.5 indicates that the operation segment has a high degree of data missingness or state uncertainty, and no student skill deficiency attribution is generated; when the data quality weight is between 0.5 and 0.6, only candidate attribution records are generated, and the mastery value is not updated; when the data quality weight is not less than 0.6, the attribution result can be used to update the student skill status record; using 0.6 as the data quality update threshold can avoid operation segments with missing data, low state confidence, or inconsistent multi-source evidence directly affecting the student's skill status; To avoid directly attributing a single abnormal value to a skill deficiency, a joint attribution is performed using operational deviation type, training status, and segment context. Segment context includes at least the preceding and following operational segments, segment boundary type, teach pendant input event sequence, and the operational order of adjacent segments within the same training task. Attribution relationships are established using a skill deficiency rule table, which includes at least the following fields: skill sub-item, operational deviation type, applicable training status, core deviation feature, supporting evidence conditions, exclusion conditions, and attribution weight. Core deviation features characterize data anomalies directly corresponding to skill deficiencies; supporting evidence conditions verify whether the anomaly is caused by the trainee's operation in conjunction with the training status and context; and exclusion conditions exclude non-trainee factors such as equipment malfunction, sensor anomalies, task template errors, or external interference. When an exclusion condition is met, no trainee skill deficiency attribution is generated; the corresponding operational segment is marked as a non-trainee factor anomalous segment and written into the anomaly record. The skill defect rule table covers all skill sub-items, with each skill sub-item configured with at least one core rule and one auxiliary rule. The attribution criteria for tool coordinate system establishment defects can be: in coordinate system establishment state, the calibration point residual is higher than the value corresponding to the 90th position of the same type of reference sample sorting; the coordinate dispersion of multiple positioning points is higher than the value corresponding to the 90th position of the same type of reference sample sorting; and the confirmed end-to-end stabilization time is lower than the value corresponding to the 10th position of the same type of reference sample sorting. The attribution criteria for singular attitude avoidance defects can be: in trajectory reproduction state or manual operation state in Cartesian coordinate mode, using the aforementioned sliding window settings, the Jacobian matrix condition number exceeds 500 and shows an upward trend within at least three consecutive sliding windows, while the number of joint velocity mutations is higher than the value corresponding to the 90th position of the same type of reference sample sorting; 500 is lower than the typical singular region condition number threshold of 1000, suitable as an early identification threshold in teaching and training; at least three consecutive sliding windows can exclude instantaneous... Numerical fluctuations; the attribution criteria for joint limit prediction defects can be selected as follows: in the manual operation state of the joint mode, the distance between the joint position and the soft limit is less than 5% of the joint's range of motion, and the current joint velocity direction still points to the corresponding soft limit, and there is no reverse retreat trend within 200 milliseconds; 5% serves as a safe buffer range before the soft limit, which can reserve retreat or deceleration space for the trainee, and 200 milliseconds is consistent with the aforementioned time proximity matching threshold of the teach pendant event and robot motion response, which can cover the common delay from a teach pendant operation to the robot motion response; the attribution criteria for force coordination control defects can be selected as follows: in the contact operation state, the end force composite value exceeds the target contact force upper limit set in the training task template or exceeds the reference range upper limit of the same type of task, the force direction fluctuation is higher than the value corresponding to the 90th position of the same type of reference sample ranking, and the end speed does not decrease with the increase of contact force; the other skill sub-items are set with corresponding core deviation features, auxiliary evidence conditions and exclusion conditions in the same way; Attribution weights are used to represent the strength of support for the same skill deficiency from different pieces of evidence; the weights for core bias features range from 0.5 to 0.7, state consistency from 0.2 to 0.3, and fragment context from 0.1 to 0.2, with the sum of the three weights after normalization being 1; the rationale for these weights is that core bias features directly reflect skill deficiencies and have the highest weight; training state is used to determine the scenario in which the bias occurs and has a moderate weight; fragment context is used to assist in verifying the relationship before and after an operation and has a lower weight; the attribution confidence C satisfies the following formula:

[0024] in Let i be the hit value of the i-th piece of evidence. Let be the attribution weight for the i-th term, and

[0025] Evidence hit values ​​range from 0 to 1, with 1 for fully meeting the corresponding conditions, 0.3 to 0.5 for slightly meeting or slightly deviating from the conditions, 0.6 to 0.8 for close to significant deviation, and 0 for not meeting the conditions. An attribution confidence score of 0.6 generates a valid attribution, and 0.8 generates a highly credible attribution. 0.6 indicates that the core bias feature is hit and is supported by at least one piece of state evidence or contextual evidence, which can avoid misjudgment based on a single feature. 0.8 indicates that multiple pieces of evidence are consistent and are suitable for labeling as a stable skill deficiency. When an operation segment triggers multiple skill defect rules simultaneously, the attribution results are sorted according to their attribution confidence. The attribution with the highest confidence is the primary attribution, and the attribution with the second highest confidence is the secondary attribution. When the difference in confidence between the primary and secondary attributions is less than 0.1, parallel attributions are retained. 0.1 indicates that the credibility of the two attribution results is close, and forcibly assigning a single attribution can easily mask compound defects. For example, inaccurate workpiece coordinate system calibration may simultaneously cause trajectory deviation in Cartesian coordinate mode. In this case, workpiece coordinate system calibration defects and manual operation defects in Cartesian coordinate mode are retained in parallel to distinguish root causes and performance biases during training task matching. Trainee skill status records are represented by a skill sub-item status table. This table includes at least the following fields: skill sub-item identifier, mastery value, update time, confidence interval, most recent deviation type, most recent attribution frequency, and key training marker. The mastery value ranges from 0 to 1, where 0 indicates not yet mastered and 1 indicates stable mastery. The initial value can be set to 0.3 to represent the trainee's starting level after completing basic theory or introductory training but not yet having developed stable practical skills. For trainees with no prior knowledge, the initial value can also be set to 0.2. Negative attribution refers to... The attribution is based on the result pointing to a skill deficiency or significant operational deviation. When a valid negative attribution occurs for the corresponding skill sub-item, the new mastery value is equal to the original mastery value minus the corresponding downward adjustment step size. The downward adjustment step size can be determined by multiplying 0.02 by 0.15 and the difference between the attribution confidence level and 0.6, and is limited to the range of 0.02 to 0.08. Thus, when the attribution confidence level is 0.6, the downward adjustment step size is 0.02; when the attribution confidence level is 1.0, the downward adjustment step size is 0.08, resulting in a small correction for low-confidence attributions and a larger correction for high-confidence attributions. When the performance of a corresponding skill sub-item falls within the normal range in a continuous operation segment, the new mastery value equals the original mastery value plus the corresponding upward adjustment step size, with a single upward adjustment step size ranging from 0.01 to 0.05. The upward adjustment step size is lower than the downward adjustment step size to prevent short-term normal performance from quickly offsetting identified stable skill deficiencies. When there are two consecutive normal performances, the upward adjustment step size can be 0.01; when there are more than five consecutive normal performances, the upward adjustment step size can be 0.05. After the mastery value is updated, it is limited to the range of 0 to 1. If it exceeds the upper limit, it is recorded as 1; if it is below the lower limit, it is recorded as 0. To avoid a single anomaly or short-term sensor error causing a large fluctuation in the skill status record, the cumulative change of any skill sub-item within 10 minutes shall not exceed 0.2. If it exceeds this range, it will be frozen and automatically updated and enter the coach's review queue. 10 minutes can usually cover a complete training task or multiple continuous operation segments. 0.2 is equivalent to 20% of the full range of the mastery value, which is suitable for limiting short-term drastic jumps. The confidence interval is used to represent the maximum fluctuation range of the mastery value, and is determined based on the number of effective updates and the direction of the last five updates. When the number of effective updates is less than 5, the confidence interval width can be set to 0.3, indicating that the number of sample observations is small and the reliability of skill state estimation is low. When the number of effective updates exceeds 20 and the direction of the last five updates is consistent, the confidence interval width can be narrowed to 0.05, indicating that the performance trend of the skill sub-item is relatively stable. The upper and lower limits of the confidence interval should not exceed the range of 0 to 1. If the upper limit is exceeded, it is recorded as 1, and if the lower limit is exceeded, it is recorded as 0. If the same skill sub-item shows effective negative attribution in three consecutive training tasks, the skill sub-item is marked as a key training item. If three consecutive tasks can rule out single occasional errors, it indicates that the skill sub-item has a stable weakness. After completing the above processing, the output information should include at least the following fields: skill defect attribution results, attribution confidence, supporting evidence list, corresponding operation segment index, and updated student skill status record. These fields will be used as input for matching training task templates and filtering training data.

[0026] Specifically, such as Figure 5As shown: Using the aforementioned student skill status records, skill defect attribution results, and corresponding operation segment indexes as input, and calling the training task template library and practical training knowledge base, training task matching and practical training data filtering are performed when a single practical training task ends, the student actively requests the next training task, or a key training mark appears in the student's skill status record. Each training task template in the training task template library includes at least the following fields: skill sub-item label, task type, task difficulty level, applicable target, parameter range, target completion criteria, allowable deviation range, and applicable conditions. The applicable target is determined based on the student's mastery value range, key training mark, and training task type. The criteria for achieving the target and the allowable deviation range are determined based on the training task type, workpiece dimensional tolerance, robot repeatability, and instructor configuration parameters. The task difficulty level can be set from level 1 to level 5, where level 1 corresponds to a mastery value of approximately 0.3, level 2 to approximately 0.45, level 3 to approximately 0.6, level 4 to approximately 0.75, and level 5 to approximately 0.9. The rationale for this grading method is that the five difficulty levels can cover the beginner, basic, qualified, proficient, and advanced training levels, and the mastery value interval between adjacent levels is approximately 0.15, which facilitates the correspondence with mastery values ​​in the range of 0 to 1. When matching training task templates, skill sub-items with an attribution confidence score of not less than 0.6 are first selected from the trainee's skill status records as candidate target training items. Among them, skill sub-items with an attribution confidence score of not less than 0.8 or marked with a key training mark are given priority. An attribution confidence score of 0.6 is consistent with the aforementioned effective attribution threshold and is used to exclude occasional biases due to insufficient evidence. An attribution confidence score of 0.8 indicates that multiple pieces of evidence are consistent and is suitable as a basis for priority training. When multiple skill sub-items meet the candidate conditions, the skill sub-item with the lowest mastery value is selected as the target training item. If the difference in mastery values ​​between two skill sub-items is less than 0.15, both are selected as target training items. 0.15 is consistent with the mastery value interval corresponding to the difficulty level of adjacent tasks and is used to determine whether two weaknesses are at similar training levels to avoid omitting skill weaknesses of equal severity. When there are dependencies between candidate training items, the priority training object is determined by a skill dependency table. The skill dependency table includes at least the fields of basic skill sub-item, affected skill sub-item, and dependency strength. The dependency strength ranges from 0 to 1. When the dependency strength is not less than 0.6, the training task corresponding to the basic skill sub-item is selected first. 0.6 is used to indicate that the basic skill has an effective influence on the performance of the sub-skill, which can avoid training the affected skill directly when the basic skill has not been stably mastered. For example, when workpiece coordinate system calibration defects and Cartesian coordinate mode operation deviations occur at the same time, the workpiece coordinate system calibration training task is matched first to reduce the impact of coordinate reference errors on Cartesian coordinate mode operation training. After determining the target training item, the task difficulty level is matched according to the mastery value corresponding to the target training item, and training task templates that are 0.5 to 1 level higher than the current mastery value are preferentially selected; a difficulty increase of 0.5 to 1 level is a training range slightly higher than the current ability level, which can provide a moderate training challenge; the training improvement effect is weak when it is below 0.5 levels, and it is easy to lead to a decrease in task completion rate and an increase in the risk of incorrect operation when it exceeds 1 level; if the trainee's mastery value of the corresponding skill sub-item is below 0.4, the template parameters are adjusted to simplify; a mastery value below 0.4 indicates insufficient stability of basic operations, and reducing task complexity can reduce the number of invalid failures; if the mastery value is between 0.4 and 0.7, the standard parameters of the training task template are used; The interval indicates that the trainee is in the basic training to qualified training stage, suitable for consolidating skills through standard tasks; if the mastery value is higher than 0.7 but there is still a slight deviation, the template parameters should be adjusted to increase the difficulty; a mastery value higher than 0.7 indicates that the trainee has a certain degree of stability and can be strengthened through increased requirements; the adjustment range of template parameters can be controlled within 10% to 20% of the standard parameters. This range is a relatively mild difficulty adjustment range in teaching and training, which can reflect the difference in task difficulty without destroying the basic structure of the original training task; simplification directions include shortening the trajectory length, reducing the load mass, relaxing the beat constraints, or reducing the contact force target; increasing the difficulty direction includes increasing the number of target points, shortening the beat time, or increasing the force control stability requirements; For contact-based training, the end-target contact force can be set in the range of 10 to 30 Newtons. This range is suitable for light-load shaft-hole assembly, electrical terminal plugging, and contact teaching training, providing perceptible contact feedback while remaining below abnormal force levels that could easily cause workpiece damage or safety risks. The aforementioned 50 Newtons is used for abnormal risk assessment and is not used as the target training force. For trajectory teaching training, the path length can be set in the range of 200 to 800 millimeters. 200 millimeters is suitable for beginners to conduct short-distance point teaching and basic trajectory control, while 800 millimeters can cover medium-length trajectories within common training workbenches. Exceeding this range can easily introduce too many spatial obstacle avoidance and rhythm factors, which is not conducive to single-skill training. When generating the next round of training tasks, the generated training tasks are based on the matched training task templates and include personalized task parameters and prompt levels. The training task content includes at least the following fields: target skill sub-item, task type, task parameters, target completion criteria, allowable deviation range, and prompt level. The prompt level is determined based on the trainee's skill status record: when the mastery value of the target skill sub-item is below 0.4, item-by-item operation prompts are provided; when the mastery value is between 0.4 and 0.7, key node prompts are provided; when the mastery value is above 0.7, only task objectives and safety prompts are provided. This prompt grading method is consistent with the training stage corresponding to the mastery value, which can gradually reduce the reliance on prompts as the trainee's ability improves. If there is no template in the training task template library that meets the target training item and difficulty level, the difficulty range is first widened to the adjacent level; if a match still cannot be found, a coach manual configuration request is generated, and the target skill sub-item, expected difficulty level, and reason for absence are written into the template expansion record. When screening the training data for this round, the aforementioned deviation level, data validity label, state confidence, and attribution results were used for classification. Normal segments and slightly deviated segments could be written into the training sample area for reference range and task template optimization. Significantly deviated segments were not written into the training sample area, but their corresponding diagnostic records were retained and written into the error case area for teaching review and error pattern analysis. Data segments with continuous missing values ​​exceeding 500 milliseconds or segments marked as unusable data were not included in the training sample update. Segments with a state confidence of less than 0.7 were written into the manual review area for coach review. Segments caused by equipment failure, sensor inaccuracy, task template errors, or external interference were written into the anomaly exclusion area and were not included in the reference range and training template update. The 500 milliseconds and 0.7 values ​​are consistent with the aforementioned data validity and state confidence judgments, which can avoid the impact of insufficient temporal evidence or uncertain state data on the quality of the knowledge base. The data written into the training knowledge base must include at least the following fields: task identifier, student anonymity identifier, skill sub-item identifier, time segment index, deviation type, attribution result, quality level, data summary, and data usage label. To avoid low-quality data repeatedly entering the knowledge base and causing a shift in the reference range, the training knowledge base adopts a small incremental update mechanism. The reference range boundary values ​​and training task template parameters are only updated after 20 new high-quality samples are added for the same skill sub-item. The reason for setting 20 sets as the incremental update threshold is that this number provides the basic stability required for a small batch update, while preventing the knowledge base from remaining in an old sample state for an extended period. The aforementioned minimum of 30 sets... The initial sample is used to establish the reference range, and the 20 sets are used for incremental updates; their functions are different. Each time the reference range boundary value is adjusted using newly added high-quality samples, the adjustment range shall not exceed 1% of the original boundary value. 1% is considered a small update range, which can prevent the reference range from drifting rapidly due to a small number of newly added samples, thereby maintaining the stability of the training knowledge base. High-quality samples refer to data segments that are not marked as unusable data, have a state confidence level of not less than 0.7, a deviation level of normal or slight deviation, do not meet the exclusion criteria, and have been confirmed by automatic screening or manual review. Among them, only data that has passed manual review and been transferred to the training sample area can be included in the 20 sets of newly added high-quality samples. If the data in the manual review area is confirmed by the coach to be representative of the training, it will be transferred to the training sample area or error case area according to its deviation level; if it is confirmed that it is caused by equipment failure, sensor inaccuracy, external interference or task template error, it will be transferred to the anomaly elimination area; after the training task is generated and the data quality is screened, the generated next round of training task will be used as the starting condition for subsequent practical training tasks, and the screened data will be written into the practical training knowledge base, thus forming a closed loop of training task matching, practical training data screening and knowledge base updating.

[0027] Example 2: Based on Example 1, the specific application process of a robot feedback-driven training closed-loop optimization method is further explained: Specifically, such as Figure 6 As shown: Taking the six-degree-of-freedom industrial robot shaft and hole assembly training task as an example, the operation logic of the robot feedback-driven training closed-loop optimization method is further explained. This training task requires trainees to complete tool coordinate system confirmation, workpiece hole proximity, insertion path teaching, contact force control, and exit path teaching through a teach pendant. The training objects include at least a six-degree-of-freedom industrial robot, an assembly tool installed at the robot's end effector, a workpiece with holes set on the training platform, a teach pendant, an end effector six-dimensional force sensor, a joint encoder, an inertial measurement unit, an industrial camera in the training area, and an edge computing unit. After the trainee logs in and selects the shaft and hole assembly training task, the edge computing unit generates a task identifier for this training and calls the task flow table and training task template corresponding to the shaft and hole assembly training. After training begins, the edge computing unit receives robot joint state data, end effector feedback data, training area visual data, and teach pendant input events based on a unified clock, and writes the above data into a time index sequence under the same task identifier. When the trainee confirms the tool coordinate system through the teach pendant, the teach pendant generates events such as coordinate system switching, positioning point confirmation, and parameter confirmation. The robot end effector moves at low speed between different positioning points above the workpiece. Based on the teach pendant event time and the changes in robot motion state, the edge computing unit divides this stage into tool coordinate system establishment operation segments and records the end effector speed, end effector posture, and positioning point coordinates before each positioning point confirmation. If the coordinate dispersion obtained from multiple confirmations of the same positioning point is large, and the end effector speed before confirmation is still higher than the stable speed threshold, then the operation segment is identified as a tool coordinate system establishment related segment, and a tool coordinate system establishment deviation candidate is generated. When the trainee controls the robot's end effector to approach the workpiece hole, the distance between the robot's end effector and the edge of the workpiece hole gradually decreases, and the joint speed gradually decreases from a high value. The teach pendant generates point confirmation and motion mode switching events. The edge computing unit divides this stage into hole approach operation segments based on the relative position of the end effector and the workpiece hole, the joint speed changes, and the end effector speed changes in the visual data. It then associates these operation segments with the manual operation in Cartesian coordinate mode and the teach point confirmation skill sub-item. If the end effector approach path shows obvious backtracking, or if the angle between the end effector speed directions in adjacent windows changes significantly, an end effector trajectory smoothness deviation is generated. If there is still a large residual speed at the end effector before point confirmation, a teach point confirmation deviation is generated. During the insertion path teaching process, the trainer controls the robot's end effector to slowly descend along the hole axis and make contact with the workpiece. The end effector's six-dimensional force sensor detects an increase in the end effector force composite value, and the industrial camera identifies that the assembly tool and the hole are in close proximity, with the joint speed and end effector speed showing a decreasing trend. The edge computing unit identifies this stage as a contact operation segment and associates it with the force coordination control skill sub-item. If the end effector force composite value increases rapidly after contact and the end effector speed does not decrease accordingly, or if the force direction fluctuates frequently in a short period of time, the operation segment is identified as a force coordination control deviation. If the robot approaches the joint soft limit during insertion and the trainer does not adjust the joint movement direction in a timely manner, a candidate deviation of insufficient joint limit prediction is generated. During the exit path teaching and trajectory reproduction process, the robot performs the exit action according to the points set by the trainer. If the Jacobian matrix condition number continuously increases and the joint velocity changes abruptly during this stage, the edge computing unit associates the operation segment with the singular posture avoidance skill sub-item and generates a singular posture approach risk deviation. If the deviation distance between the actual trajectory at the end and the target path continuously increases during trajectory reproduction, the manual operation deviation in Cartesian coordinate mode is recorded. For the same operation segment involving multiple skill sub-items, the edge computing unit determines the main skill sub-item and auxiliary skill sub-item based on the degree of satisfaction of the teach pendant event, dominant motion feature, and deviation feature. When the matching degree of multiple skill sub-items is close, the parallel association results are retained for subsequent attribution processing. After this round of training, the edge computing unit establishes attribution relationships based on the type of operational deviation, training status, and context of each operation segment. If there are abnormalities in the dispersion of positioning point coordinates or the residual of calibration points during the tool coordinate system establishment stage, and it is confirmed that the end-effector stabilization time is insufficient, then this type of deviation is attributed to a defect in the tool coordinate system establishment. If there are abnormal increases in end-effector force or large fluctuations in force direction during the insertion stage, and the end-effector velocity does not decrease with the increase in contact force, then this type of deviation is attributed to a defect in force perception coordination control. If there are continuous increases in the Jacobian matrix condition number and sudden changes in joint velocity during the exit stage, then this type of deviation is attributed to a defect in singular posture avoidance. For abnormal segments caused by sensor frame loss, loose workpiece fixtures, or external collisions, if the exclusion conditions are met, no attribution of trainee skill defects is generated; instead, the corresponding operation segment is written into the abnormal record. After attribution is completed, the edge computing unit updates the trainee skill status record. If both the tool coordinate system establishment defect and the force coordination control defect in this round of training meet the valid attribution conditions, the mastery value of the corresponding skill sub-item is reduced respectively. If the demonstration point confirmation and trajectory exit segment performance are within the normal range, the mastery value of the corresponding skill sub-item is slightly increased. If the same skill sub-item shows valid negative attribution in multiple consecutive training sessions, the skill sub-item is marked as a key training item. The updated trainee skill status record includes at least the following fields: skill sub-item identifier, mastery value, recent deviation type, attribution confidence, update time, and key training mark. When generating the next training task, the edge computing unit prioritizes skill sub-items with lower mastery values ​​and higher attribution confidence as target training items. For example, when a trainee's tool coordinate system establishment mastery value is lower than their force coordination control mastery value, and tool coordinate system establishment defects will affect the accuracy of subsequent insertion paths, a tool coordinate system establishment reinforcement training task is generated first. This training task can reduce the number of workpiece holes, shorten the movement path, and provide step-by-step operation prompts, allowing the trainee to first complete the positioning point stability confirmation and coordinate system establishment training. When the tool coordinate system establishment mastery value improves, a force coordination control training task is then pushed, for example, using a lower target contact force, a longer insertion time, and a wider cycle constraint, so that the trainee can gradually establish the coordination relationship between contact force changes and end-effector speed control. Before the training data is written into the training knowledge base, the edge computing unit categorizes the data for different uses. Normal segments and segments with slight deviations are written into the training sample area as a data source for the reference range and training task template optimization. Segments with significant deviations are written into the error case area for teaching review and typical error analysis. Segments with low state confidence or missing data are written into the manual review area. Segments caused by sensor anomalies, loose fixtures, or external interference are written into the anomaly exclusion area and do not participate in the reference range update. After the above processing is completed, the next training task can be generated based on the operational deviations and skill defects attribution results of this round, and the training knowledge base can be continuously updated based on the selected high-quality data, thus forming a closed loop of operation diagnosis, skill attribution, task push, and data optimization driven by robot feedback data.

[0028] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0029] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented in software, the above embodiments can be implemented in whole or in part by a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions of the embodiments of this application are implemented in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted wirelessly or wiredly from one website, computer, server, or data center to another website, computer, server, or data center. Wired methods include optical fiber, twisted pair, coaxial cable, etc. Wireless methods include infrared, microwave, etc. Available media include any available media that can be accessed by a computer or data storage devices such as servers and data centers that contain one or more sets of available media. Available media can be magnetic media (floppy disks, hard disks, magnetic tapes), optical media (DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0030] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0031] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A robot feedback-driven training closed-loop optimization method, characterized in that, include: S1. After the industrial robot training task is started, robot joint status data, end effector feedback data, training area visual data and teach pendant input events are collected synchronously, and the data timing correspondence under the same training task is established. S2. Based on the teach pendant input events and the robot's motion state, the trainee's operation process is divided into multiple continuous operation segments, and the corresponding training state for each operation segment is determined. When dividing the continuous operation segments, a unified time index data stream is used as input to extract candidate segment boundaries from the teach pendant input events. The motion change boundaries are extracted based on the fluctuations of joint velocity, end-effector velocity, end-effector force, and end-effector posture within the sliding window. When determining the corresponding training state for each operation segment, a training state table is established. Candidate training states are first determined based on the teach pendant input events within the segment, and then the training state is determined by combining the joint velocity distribution, end-effector velocity direction, end-effector force change, and visual proximity relationship. S3. Associate each operation segment with a preset skill sub-item, and determine the operation deviation type based on the data time sequence corresponding to each operation segment; when determining the operation deviation type, extract deviation identification features according to the data time sequence corresponding to the associated skill sub-item, and compare the deviation identification features with the reference range established according to the training task type, robot model, end tool type and difficulty level to determine the deviation level; S4. Based on the type of operational deviation, training status, and corresponding operational segment, establish the attribution relationship between the student's operational deviation and skill deficiency, and update the student's skill status record according to the attribution relationship; when establishing the attribution relationship, remove unusable data, call the skill deficiency rule table, and combine the operational deviation type, applicable training status, and segment context to perform joint attribution; Attribution relationships are established through a skill defect rule table, which includes at least the following fields: skill sub-item, operational deviation type, applicable training state, core deviation feature, supporting evidence conditions, exclusion conditions, and attribution weight. The core deviation feature characterizes data anomalies directly corresponding to skill defects. Supporting evidence conditions verify whether the anomaly is caused by the trainee's operation in conjunction with the training state and context. Exclusion conditions exclude non-trainee factors caused by equipment malfunction, sensor anomalies, task template errors, or external interference. When an exclusion condition is met, no trainee skill defect attribution is generated; the corresponding operational segment is marked as a non-trainee factor anomaly segment and written into the anomaly record. Low-quality candidate attributions do not participate in mastery value updates. When the cumulative change exceeds the limit, the automatic update of the corresponding skill sub-item is frozen, and a key training marker is set when there are consecutive negative attributions. S5. Match the training task template according to the updated student skill status record, generate the next round of training tasks, and write the current round of training data into the training knowledge base after quality screening; determine the target training items based on the mastery value, attribution confidence and key training mark in the student skill status record, and determine the task difficulty, adjust the task parameters and prompt level in combination with the skill dependency relationship to generate the next round of training tasks; when screening the current round of training data for quality, write the operation segment into the training sample area, error case area, manual review area or anomaly exclusion area according to the deviation level, data validity mark, status confidence, attribution result and exclusion conditions.

2. The robot feedback-driven training closed-loop optimization method according to claim 1, characterized in that, After the industrial robot training task is started, robot joint status data, end effector feedback data, training area visual data, and teach pendant input events are collected synchronously, and a data time sequence correspondence relationship under the same training task is established, including: When establishing the data time sequence correspondence, the unified clock of the edge computing unit is used as the reference, and timestamps are added to the joint status data, end feedback data, training area visual data and teach pendant input events; The original timestamps are corrected based on the time deviation, and a unified time index is established; Set validity and quality tags for the data, retain the operation context through sliding cache, and fill in short-term missing data. When critical data sources are continuously missing, pause the attribution processing of the corresponding segments.

3. The robot feedback-driven training closed-loop optimization method according to claim 1, characterized in that, Based on the teach pendant input events and the robot's motion state, the trainee's operation process is divided into multiple continuous operation segments, including: After performing temporal proximity matching between the candidate segment boundary and the motion change boundary, it is determined to be a composite boundary, event boundary, or implicit boundary. Then, continuous operation segments are formed in chronological order, and operation segments with a duration less than the preset duration and no abnormal motion events are merged with adjacent operation segments.

4. The robot feedback-driven training closed-loop optimization method according to claim 1, characterized in that, Determine the training status corresponding to each operation segment, including: The manual operation state in joint mode is determined based on the dominant motion characteristics of a single joint, while the contact operation state is determined based on the resultant force value of the end effector, the change in the end effector velocity, and the proximity relationship between the robot end effector and the work object. When multiple training states simultaneously meet the judgment criteria, state confidence is generated based on event consistency, motion feature consistency, and visual relationship consistency.

5. The robot feedback-driven training closed-loop optimization method according to claim 1, characterized in that, Associate each operation segment with a preset skill sub-item, including: When associating skill sub-items, configure skill sub-items according to coordinate establishment, point teaching, motion control, contact operation, safety prediction and rhythm control, and establish a training task flow table; Matching is performed based on the time and location of the operation segment, the training status, the teaching pendant input events, and the dominant motion characteristics. When multiple skill sub-items are matched simultaneously, the primary skill sub-item, auxiliary skill sub-item, or parallel relationship is determined.

6. The robot feedback-driven training closed-loop optimization method according to claim 1, characterized in that, The type of operational deviation is determined based on the data timing sequence corresponding to each operational segment, including: Process thresholds are used to constrain force control, joint limit and unusual posture related features. When the conclusions corresponding to the reference range are inconsistent with the conclusions corresponding to the process thresholds, the operation deviation type is determined according to the safety-related process thresholds. When the data quality is lower than the judgment requirements, candidate operation deviation types are generated.

7. The robot feedback-driven training closed-loop optimization method according to claim 1, characterized in that, Based on the type of operational deviation, training status, and corresponding operational segments, establish the attribution relationship between trainees' operational deviations and skill deficiencies, including: First, identify non-student factor abnormal segments based on exclusion criteria. Then, generate attribution confidence based on core bias characteristics, supporting evidence conditions, and attribution weights. When multiple rules are hit simultaneously, determine the primary attribution, secondary attribution, or parallel attribution.

8. The robot feedback-driven training closed-loop optimization method according to claim 1, characterized in that, Update trainee skill status records based on attribution relationships, including: When updating trainees' skill status records, a skill sub-item status table is used to record the skill sub-item identifier, mastery value, update time, confidence interval, recent deviation type, recent attribution frequency, and key training mark. The mastery value is adjusted down based on effective negative attribution and adjusted up based on continuous normal performance.

9. The robot feedback-driven training closed-loop optimization method according to claim 1, characterized in that, Based on the updated trainee skill status records, a training task template is matched to generate the next round of training tasks, including: When matching training task templates, the training task template library is called. The training task template library records skill sub-item tags, task type, difficulty level, parameter range, and applicable conditions.

10. The robot feedback-driven training closed-loop optimization method according to claim 1, characterized in that, After quality screening, the training data from this round was written into the training knowledge base, including: When high-quality samples are automatically screened or manually reviewed and confirmed to meet the update conditions, the reference range boundary values ​​and training task template parameters in the training knowledge base are incrementally updated.

Images