Forklift training system and method based on dangerous operation prediction and adaptive intervention

Abstract

The application discloses a forklift training system and method based on dangerous operation prediction and adaptive intervention, and particularly relates to the technical field of forklift training, which comprises obtaining vehicle position, vehicle orientation, driving speed, fork state, mast state, site boundary relationship and steering input corresponding to the same training task, performing sequential arrangement through time sequence alignment, and outputting training records; according to the training records, corresponding action execution sequence of steering action, braking action, advancing and retreating action, lifting and lowering action and observation is identified through action perception, and out-of-sequence identification and remedial identification are performed according to the prescribed action sequence, and out-of-sequence chains are output; for the out-of-sequence chains, bottom item identification is performed through position rollback amount, repeated correction number and insertion deviation amount, and bottom item records are output; correlation identification is performed on action sequence, remedial action and bottom item in the training process, and a retraining task is reconstructed under the condition of keeping the training task target unchanged to make error operation explicit and implement corresponding intervention.

Description

Technical Field

[0001] This invention relates to the field of forklift training technology, and more specifically, to a forklift training system and method based on dangerous operation prediction and adaptive intervention. Background Technology

[0002] In forklift training practice, existing technologies mainly focus on detecting improper operation by trainees during driving and operation as early as possible, and providing training evaluation or safety tips accordingly. In engineering implementation, the speed, steering, braking, fork lifting, mast changes, vehicle position, and relative status with racks, pallets, and aisle boundaries are continuously acquired on the training vehicle or simulation device. Then, according to the established judgment criteria, prompts, deductions, replay analysis, or suspension are carried out for situations such as collisions, crossing lines, sharp turns, speeding, cargo shaking, stacking deviations, or exceeding the operation time limit. For example, when organizing continuous practical training in warehouse training around picking up goods in narrow aisles, turning around while carrying goods, and stacking goods at high levels, the training arrangement should maintain the continuous rhythm within the existing class time, and should not frequently stop to dissect and comment on each suspicious action. At the same time, it should not set up extreme conditions that are obviously inconsistent with the actual operation process. Instead, trainees should be required to complete the entire operation under conditions that are close to the actual warehouse size, aisle width, and task connection method. However, under these training conditions, a situation that can be directly verified will repeatedly occur on site: some trainees have obvious problems in the order of observation, deceleration timing, steering correction, and fork posture control, but they will still complete the task because there is extra turning space in the passage, the position allows a certain deviation, there is no obstruction for a short time, or the subsequent remedial space has not been exhausted. The existing handling method, which is mainly based on the result event and obvious abnormality, can only record such samples as completed or slightly non-compliant, and cannot further distinguish whether the completion is due to standard operation or due to accidental conditions covering up the wrong method. As a result, the wrong operation method is retained as an acceptable path in the training stage and quickly evolves into a real risk in subsequent tighter working conditions. The technical problem this application aims to solve is: how to identify the accidental conditions that undermine the correct operating methods during the completion of training tasks without interrupting the continuous process of forklift training or deviating from real operating conditions, and thereby reveal the erroneous operations that are masked by the completed results. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a forklift training system and method based on dangerous operation prediction and adaptive intervention. By performing correlation identification on the action sequence, remedial actions, and bottoming-out items during the training process, and reconstructing the retraining task while keeping the training task objective unchanged, the system can make erroneous operations explicit and implement corresponding interventions, thereby solving the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: Forklift training methods based on hazardous operation prediction and adaptive intervention include: S1. Obtain the vehicle position, vehicle orientation, driving speed, fork status, mast status, site boundary relationship and operation input corresponding to the same training task, align the execution order according to the time sequence, and output the training record. S2. Based on the training records, the sequence recognition of steering, braking, forward and backward, lifting and lowering actions and observation actions is performed through motion perception. The sequence recognition and remediation recognition are performed according to the prescribed action sequence, and the error chain is output. S3. For misaligned chains, identify the support item by the position back-off amount, the number of repeated corrections, and the insertion deviation, and output the support item record. S4. Based on the bottom-line item record, while keeping the task objective unchanged, reduce the return space, the number of repeated corrections and the alignment margin, and output the retraining task. S5. During the retraining task, the current action is compared in real time through action perception, and combined with the sequence prediction of the wrong chain execution, the danger manifestation segment is located, a prompt is issued before the danger manifestation segment is reached, and the corresponding remedial action is restricted when the wrong sequence action occurs again, and the retraining record is output. S6. By comparing the completion results of the retraining record with the training record and comparing the explicit results, identify the erroneous operation methods that are covered by the completion results of the training task, generate subsequent intervention content, and output the training intervention results.

[0005] In a preferred embodiment, S1 includes: S11. Based on the collection time, perform simultaneous classification of vehicle position, vehicle orientation, driving speed, fork status, mast status, site boundary relationship and operation input, extract the time of vehicle orientation change, driving speed change, fork status switch, mast status switch and site boundary relationship change, and determine the time when two or more types of changes occur at the same collection time as common time, and output the common time sequence. S12. Using the acquisition records between adjacent common moments in the common moment sequence as the same segment record, the operation inputs in each same segment record are ranked in the order of appearance of steering input, driving input, fork input and gantry input, and the acquisition records that do not fall into the common moment are attached to the previous or next common moment according to the time adjacency relationship, and the sequence segment record is output. S13. Perform head-to-tail concatenation on the sequence records according to the common time sequence, and write the segment number, input order within the segment, corresponding vehicle position, corresponding vehicle orientation, corresponding driving speed and corresponding site boundary relationship in each sequence record into a unified record format, and output the training record.

[0006] In a preferred embodiment, S2 includes: S21. Based on the changes in the control input, vehicle position, vehicle orientation, driving speed, fork status and mast status corresponding to the training records, the starting point and ending point of the action are identified at each collection time through motion perception. Adjacent collection times that change continuously in the same direction are merged into the same action segment, and the action segments are formed into steering action segments, braking action segments, forward and backward action segments and lifting action segments according to the action category. The action segment sequence is output. S22. Obtain the trainee's head turning record and gaze dwell record corresponding to the training record. Perform observation corresponding position recognition on the head turning record and gaze dwell record through motion perception. Form observation action segments from the continuous observation records facing the direction of travel, fork direction, target cargo location direction and site boundary direction respectively. Perform joint ranking according to the sequence of action segments and the order of observation action segments, and output the action order table. S23. Based on the prescribed action sequence corresponding to the same training task, perform a front-to-back sequence comparison on the action sequence table, determine the action segment that appears before the prescribed sequence as the misordered action segment, and determine the turning action segment, braking action segment, forward and backward action segment or lifting action segment that is located after the misordered action segment and is used to restore the prescribed operation relationship as the remedial action segment, and output the misordered chain according to the succession order between the misordered action segment and the remedial action segment.

[0007] In a preferred embodiment, S3 includes: S31. Around each misaligned action segment in the misaligned chain, take the starting time of the corresponding remedial action segment as the return point and the ending time of the corresponding misaligned action segment as the return point. Perform reverse return on the training record. Accumulate the additional travel distance corresponding to the return to the specified travel relationship after deviating from the return point and the return point to obtain the position return amount. Count the number of times the steering correction, stopping correction and forward / backward correction occur between the return point and the return point to restore the specified operation relationship to obtain the number of repeated corrections. Accumulate the lateral deviation distance of the fork centerline relative to the target cargo position centerline before and after insertion and the remaining lateral deviation distance after correction to obtain the insertion deviation amount. Output the bottoming candidate table.

[0008] In a preferred embodiment, S3 further includes: S32. Based on the candidate list for bottoming out, perform bottoming out by following the sequence of each out-of-order action segment and its corresponding remedial action segment. Records with non-zero position backoff are listed as moving bottoming out items, records with non-zero number of repeated corrections are listed as correction bottoming out items, and records with non-zero insertion deviation are listed as insertion bottoming out items. For records where the same remedial action segment is referenced by two or more out-of-order action segments, perform unique attribution processing. Specifically, the remedial action segment is assigned to an out-of-order action segment with one interval action segment between the retracement start point and the start point of the remedial action segment. When there are two or more out-of-order action segments with one interval action segment, the remedial action segment is assigned to an out-of-order action segment whose end time is later than the end time of other out-of-order action segments. Output the bottoming out list. S33. Perform deletion replay for each support item in the support sublist. Delete each support item from the corresponding training record one by one while keeping the remaining records unchanged, and then reconstruct the operation process. For support items that fail to restore the specified movement relationship, the specified insertion relationship, or the training task after deletion, record them as established support items. For two or more established support items corresponding to the same misordered action segment, write them in the order of movement support item, correction support item, and insertion support item. Write an empty support mark for misordered action segments without established support items, and output the support item record.

[0009] In a preferred embodiment, S4 includes: S41. Extract the positions of the misordered action segments, remedial action segments, and task segments corresponding to each support item from the support item record. Extract the starting position, target cargo location, ending position, cargo loading status, and task sequence corresponding to the same training task from the training record. Write the starting position, target cargo location, ending position, cargo loading status, and task sequence as target retention items. Then, write the corresponding task segments as correction reduction segments, correction reduction segments, and alignment reduction segments according to the support item type. Output the reconstruction base table. S42. Based on the reconstruction base table, for each straightening reduction segment, shift the segment's end boundary in the opposite direction of the corresponding position's backtracking amount while keeping the segment's starting boundary unchanged to rewrite the straightening space. For each correction reduction segment, delete the second and subsequent similar correction actions in the remedial action segment according to the corresponding number of repeated corrections, and retain the first similar correction action to rewrite the correction sequence. For each alignment reduction segment, simultaneously deduct the lateral entry range from both sides of the target location with the corresponding insertion deviation amount while keeping the fork entry depth range unchanged to rewrite the alignment range. When there are two or more rewriting results for the same task segment, write them in the order of straightening space, correction sequence, and alignment range, and use the segment's end state of the previous rewriting result as the input of the next rewriting result to output the candidate retraining table.

[0010] In a preferred embodiment, S4 further includes: S43. Based on the candidate retraining table, perform forward replay and reverse anaphora verification on each task segment. Forward replay is used to verify whether the beginning and end of each task segment are continuously connected in the task order and reach the target retention item. Reverse anaphora is used to verify whether each reduced segment still corresponds to the task segment where the original support item is located and has not been rewritten across segments. For candidate results that simultaneously satisfy forward replay and reverse anaphora, connect them in the order of task segments to form a retraining task. For candidate results that do not satisfy forward replay, cancel and rewrite them item by item in the reverse order of alignment range, correction sequence, and return space, and repeat forward replay and reverse anaphora after each cancellation. For candidate results that do not satisfy reverse anaphora, cancel the cross-segment rewriting part and repeat forward replay and reverse anaphora until each task segment forms a rewritten result that passes the verification, and output the retraining task.

[0011] In a preferred embodiment, S5 includes: S51. During the retraining task, obtain the current vehicle position, current vehicle orientation, current driving speed, current fork status, current mast status, current control input, and current observed corresponding action. Through action perception, classify and register the current control input and current observed corresponding action, and compare the current action with the out-of-order actions, remedial actions, and succession order in the chain of errors, and output the current order table. S52. Based on the current sequence table, perform forward continuation expansion on the misaligned chain, and match the registered action sequence in the current sequence table with the continuation sequence in the misaligned chain according to the registration order; when each registered action sequence is consistent with the continuous continuation sequence in the misaligned chain, determine the task segment between the starting time of the remedial action after the continuous continuation sequence and the corresponding completion result turning point time as the danger manifestation segment, and determine the action switching time immediately before the starting time of the remedial action as the prompt time, and output the manifestation positioning table; S53. For the explicit positioning table, issue a sequence prompt associated with the corresponding misordered action at the prompt time, and when the misordered action is identified again in the danger explicit segment, block the writing of the remedial action input corresponding to the misordered chain, delete the repeated correction instruction corresponding to the remedial action and retain the non-remedial action input, and write the prompt time, danger explicit segment, time when the misordered action reappears and the remedial action restriction result into the retraining record.

[0012] In a preferred embodiment, S6 includes: S61. Based on the training records, extract the original task completion status, the original remedial action sequence, and the original support item distribution. Based on the retraining records, extract the retraining completion status, the retraining out-of-order action sequence, the retraining remedial action sequence, and the retraining explicit segment distribution. Perform a correspondence comparison between the original task completion status and the retraining completion status. Perform a deletion comparison between the original remedial action sequence and the retraining remedial action sequence. Perform same-segment correspondence between the original support item distribution and the retraining explicit segment distribution. Output a result comparison table. S62. Based on the result comparison table, the disordered action chains that are completed in the training record but not completed in the retraining record are recorded as completion masking chains. The disordered action chains that have remedial actions in the training record and whose corresponding remedial actions are restricted and then continuously expanded in the retraining record are recorded as explicit masking chains. The completion masking chains and explicit masking chains are merged and classified according to the disordered action category, remedial action category and task segment position, and the error mode table is output. S63. Based on the error mode table, generate sequential prompts for each error operation mode according to the corresponding wrong sequence action category, generate restricted training content according to the corresponding remedial action category, and generate retraining insertion content according to the corresponding task segment position. Write the sequential prompts, restricted training content, and retraining insertion content into the intervention release table according to the task order of the same training task, and output the training intervention results.

[0013] A forklift training system based on hazardous operation prediction and adaptive intervention includes a record organization module, an error correction and identification module, a bottom-out identification module, a reconstruction module, an explicit intervention module, and a results module. The record organization module is used to obtain the vehicle position, vehicle orientation, driving speed, fork status, mast status, site boundary relationship and operation input corresponding to the same training task, and to organize the execution order through time alignment and output the training record. The error correction module is used to perform sequential recognition of steering, braking, forward and backward, lifting and lowering actions and observation corresponding actions based on training records and through action perception. It also performs error recognition and correction recognition according to the prescribed action sequence and outputs the error chain. The bottom-support identification module is used to identify bottom-support items for misaligned chains by measuring the position back-off amount, the number of repeated corrections, and the insertion deviation, and outputs bottom-support item records. The reconstruction module is used to reduce the back-to-center space, the number of repeated corrections, and the alignment margin based on the support item record, while keeping the task objective unchanged, and output the retraining task. The explicit intervention module is used to compare the current action execution in real time through action perception during the retraining task, and combined with the sequence prediction of the wrong chain execution to locate the dangerous explicit segment, issue a warning before the dangerous explicit segment is reached, restrict the corresponding remedial action when the wrong sequence action occurs again, and output the retraining record. The results module is used to compare the completion results of retraining records with those of training records and to compare the results of explicit training. It identifies erroneous operation methods that are masked by the completion results of training tasks, generates subsequent intervention content, and outputs training intervention results.

[0014] The technical effects and advantages of this invention are as follows: This solution performs association recognition on out-of-order actions, remedial actions and backup items in training records, and reconstructs the retraining task while keeping the task objective unchanged. It can distinguish whether the task completion is due to standard operation or backup support, thereby relatively improving the ability to identify errors that are masked by the completion result. The vehicle status, operation status, site relationship and operation input are organized on the same time axis and training records, action sequence table and staggered chain are formed. This can stably unfold the action succession relationship in the continuous training process, thus providing a consistent data foundation for subsequent bottom-out identification and explicit judgment. By calculating the position backoff amount, the number of repeated corrections, and the insertion deviation, and by combining the deletion replay to verify the supporting role of each backup item in the task completion, it is possible to separate the additional correction, repeated correction, and additional alignment from the completion result, thereby relatively reducing the single dependence on the result event and explicit anomaly. By reducing the space for re-alignment, the number of repeated corrections, and the alignment margin in retraining tasks, while keeping the starting position, target location, ending position, and task order unchanged, incorrect operation methods can be made explicit in advance without deviating from the original training scenario, thereby relatively improving the pertinence of the training process. Attached Figure Description

[0015] Figure 1 This is a flowchart outlining the method steps of the present invention; Figure 2 This is a schematic diagram of the system module structure of the present invention. Detailed Implementation

[0016] 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.

[0017] Refer to the instruction manual appendix Figures 1-2 The forklift training method based on hazardous operation prediction and adaptive intervention of the present invention includes: S1. Obtain the vehicle position, vehicle orientation, driving speed, fork status, mast status, site boundary relationship and operation input corresponding to the same training task, align the execution order according to the time sequence, and output the training record. This implementation method is used to organize the same training task into training records that can be directly read for subsequent action recognition under the original data acquisition state. Its mechanism is as follows: First, vehicle position, vehicle orientation, speed, forklift status, mast status, site boundary relationships, and operational inputs are unified onto the same timeline. Then, common moments that characterize the switching of operational states are extracted from the unified data acquisition results. Subsequently, records are divided into segments around these common moments, and the inputs within each segment are ranked. Finally, the records of each ranked segment are concatenated end-to-end according to a unified field format to form a training record, ensuring that subsequent misorder recognition, remedial recognition, and bottom-line item recognition are all performed based on the same record caliber. This implementation process includes the following steps: In S11, the input quantities include vehicle position sequence, vehicle orientation sequence, driving speed sequence, fork state sequence, mast state sequence, site boundary relationship sequence, and operation input sequence corresponding to the same training task. First, the above sequences are mapped to the same time axis according to the collection timestamp, and all fields at the same collection time are simultaneously grouped to form a basic grouping record. Then, the change events are determined moment by moment in the basic grouping record. Among them, the vehicle orientation change moment is the collection moment when the vehicle orientation change direction changes from clockwise to counterclockwise, from counterclockwise to clockwise, or from change to hold. The driving speed change moment is the collection moment when the driving speed changes from increase to decrease, from decrease to increase, from zero to non-zero, or from non-zero to zero. The fork state switching moment is the collection moment when the fork lifting direction, forward and backward state, or load-bearing state changes. The mast state switching moment is the collection moment when the mast tilting forward, tilting backward, or holding state changes. The site boundary relationship change moment is the collection moment when the relationship between the vehicle center point and the channel boundary, cargo space boundary, and obstacle boundary changes. Records of two or more change events occurring at the same acquisition time are written as common time records, and the common time sequence is output in chronological order. The common time record must include at least the task number, acquisition time, change category set, and corresponding basic classification record number. When a single field is missing at a certain acquisition time, the field value with the same name from the previous acquisition time is written to that field and a carry-over mark is added. When the same field is missing in two or more consecutive acquisition times, a missing mark is written to the corresponding acquisition time and that acquisition time is not written to the common time sequence. In S12, the input quantities include common time sequence and basic classification records. First, all basic classification records between two adjacent common times in the common time sequence are taken as a group of same segment records, and written into the same segment record table according to the acquisition time of the previous common time, the acquisition time of the next common time, and the numbers of all basic classification records in the segment. Then, the operation inputs in each same segment record are read, and the steering input, driving input, fork input, and gantry input are ranked according to the first occurrence time in the segment. Among them, inputs of the same category that remain unchanged continuously in the same segment record are merged into one input item in the segment. If two or more types of operation inputs appear at the same acquisition time, the input order in the segment is registered according to the fixed order of steering input, driving input, fork input, and gantry input. Subsequently, basic sequence records that do not fall within a common time are joined by adjacency. Specifically: basic sequence records located between two common times are assigned to the common time with the smaller time difference, while those with the same time difference are assigned to the next common time; basic sequence records located before the first common time are assigned to the first common time, and basic sequence records located after the last common time are assigned to the last common time; after processing, sequential segment records are output. Sequential segment records must at least include the task number, segment number, segment start time, segment end time, set of input items within the segment, order of input items within each segment, and set of linked record numbers; when there is no manipulation input in a segment record, an empty input flag is written to that segment record and the segment number is retained to maintain time continuity during subsequent head-to-tail connections; In S13, the input includes sequence segment records and common time sequences. First, segment numbers are generated for each sequence segment record according to the time sequence of the common time sequence. The sequence segment record between the first common time and the second common time is written as the first segment, and the subsequent sequence segment records are numbered sequentially according to time. Then, the first and last segments are joined in order of segment number. For each sequence segment record, the vehicle position, vehicle orientation, driving speed and site boundary relationship in the corresponding basic classification record are read, and the segment number, input order within the segment, segment start time, segment end time, corresponding vehicle position, corresponding vehicle orientation, corresponding driving speed, corresponding site boundary relationship and corresponding basic classification record number are written into a unified record format. The unified record format includes at least the task number, segment number, record time, input order within the segment, vehicle position, vehicle orientation, driving speed, fork status, mast status, site boundary relationship, and operation input. After all sequential segment records are written, a connection check is performed based on the continuity of segment numbers, the connection relationship between the end time of a segment and the start time of the next segment, and the consistency of task numbers. Records that pass the check are output as training records and written to the training record table for S2 to read directly. When there are missing segment numbers, inverted time between segments, or inconsistent task numbers, the current first and last connection is stopped, an abnormal number is written for abnormal sequential segment records, and the process is rolled back to S12 to re-execute the adjacent connection and intra-segment ranking. Through the above implementation process, the original collected data from different sources, with different sampling rhythms and different field completeness under the same training task can be organized into training records with unified fields, continuous time, and clear order within the segment. This ensures that subsequent action perception, out-of-order recognition, and remedial recognition are all based on a unified timeline and a unified recording standard, avoiding the break between steps caused by misaligned collection times, mixed inputs within the segment, or inconsistent record connection. In practical applications: When trainees perform training tasks such as retrieving goods in narrow aisles and reversing out, the vehicle position, vehicle orientation, driving speed, fork status, mast status, site boundary relationship, and operation input are synchronously collected by on-board sensors and training terminals. The system first identifies events such as steering changes, speed changes from zero to non-zero, fork changes from holding to rising, and changes in the vehicle's relationship with the cargo location boundary on a unified time axis, and then forms a common time sequence, sequence segment record, and training record based on this. Subsequently, S2 directly reads the training record and can identify the trainee's action sequence in the process of approaching the cargo location, inserting the pallet, and reversing back to center according to the segment number and the input order within the segment.

[0018] S2. Based on the training records, the sequence recognition of steering, braking, forward and backward, lifting and lowering actions and observation actions is performed through motion perception. The sequence recognition and remediation recognition are performed according to the prescribed action sequence, and the error chain is output. This implementation method, based on existing training records, further organizes the manipulative and observational behaviors during training into action sequence results that can be directly used for misorder identification and remedial identification. Its mechanism is as follows: First, based on vehicle movement, changes in operating devices, and changes in manipulative input in the training records, the start and end boundaries of various operational actions are identified and action segments are formed. Then, the trainee's head turning records and gaze fixation records are converted into observation action segments with clearly defined observation objects. Subsequently, the operational action segments and observation action segments are jointly ranked along a unified timeline and compared item by item with the pre-defined action sequence corresponding to the same training task, thereby resolving misordered action segments, remedial action segments, and their succession relationships. The pre-defined action sequence is taken from the training plan table corresponding to the same training task. The training plan table at least includes the task number, task segment position, prescribed travel relationship, prescribed operational relationship, prescribed insertion relationship, and the action category sequence of each task segment. This implementation process includes the following steps: In S21, the continuous change process in the training record is divided into comparable operation action segments. Its working mechanism is to lock the continuous same-direction interval of the same type of operation input and corresponding state change into the same action segment through action perception. The input includes operation input, vehicle position, vehicle orientation, driving speed, fork status and mast status in the training record. First, the action start point recognition and action end point recognition are performed for each acquisition time. Among them, the steering action start point is the acquisition time when the steering input changes from zero to left turn input or right turn input, and the steering action end point is the acquisition time when the steering input returns to zero input or the steering input direction changes. The braking action start point is the acquisition time when the braking input changes from no braking to braking, and the braking action end point is the acquisition time when the braking input returns to no braking. The forward and backward action start point is the acquisition time when the vehicle's longitudinal displacement direction changes from stationary to forward or backward, and the forward and backward action end point is the acquisition time when the vehicle's longitudinal displacement direction returns to stationary or forward and backward changes. The lifting action start point is the acquisition time when the fork status or mast status changes from holding to rising, lowering, tilting forward or tilting backward, and the lifting action end point is the acquisition time when the corresponding state returns to holding or the direction changes. Then, adjacent acquisition moments that change continuously in the same direction are merged. Continuous acquisition moments with the same steering input direction and the same vehicle orientation change direction are formed into a steering action segment. Continuous acquisition moments with continuous braking input and continuously decreasing speed are formed into a braking action segment. Continuous acquisition moments with continuous changes in vehicle position along the same longitudinal direction are formed into a forward / backward action segment. Continuous acquisition moments with continuous changes in fork status or mast status along the same change direction are formed into a lifting action segment. Subsequently, the action segment is written into the action segment table according to the action category, action start time, action end time, task segment position, and corresponding training record number. The action segment sequence is output and read by S22. When the control input exists at a certain acquisition moment but the corresponding state field is missing, a check mark is written to that acquisition moment and the action start point is not formed temporarily. When the state field is restored after two consecutive acquisition moments and the change direction is consistent with the original control input, the aforementioned check mark acquisition moment is written as the action start point; otherwise, the check mark is deleted. In S22, the trainee's observation behavior is converted into observation action segments that can be uniformly compared with the operation action segments. Its mechanism is to map head turning records and gaze placement records to a specific observation object through motion perception and complete joint ranking on a unified time axis. The input includes the trainee's head turning records, gaze placement records corresponding to the training records, and the action segment sequence output from S21. First, a simultaneous mapping is performed on the head turning records and gaze placement records, mapping the head orientation and gaze direction at each acquisition moment to four types of observation objects: travel direction, forklift direction, target cargo location direction, and site boundary direction. Among these, the gaze... When the direction of the gaze is consistent with the direction of the head, the direction is directly written as the corresponding observation position. When the direction of the gaze is inconsistent with the direction of the head, the direction of the gaze is written as the corresponding observation position and the direction of the head is written as an auxiliary field. Then, adjacent collection times that keep the corresponding observation position unchanged are merged into the same observation action segment, forming observation action segments facing the direction of travel, observation action segments facing the forklift direction, observation action segments facing the target cargo location direction, and observation action segments facing the site boundary direction. Each observation action segment is written into the observation action segment table according to the start time, end time, observation object category, and the position of the task segment to which it belongs. Subsequently, the action segment sequence and the observation action segment table are read, and joint ranking is performed according to the starting time in the unified time axis. When the operation action segment and the observation action segment appear at the same acquisition time, they are written into the action order table in a fixed order of observation action segment, turning action segment, braking action segment, forward and backward action segment, and lifting action segment. The action order table must at least write the task number, task segment position, action segment number, action category, order number, starting time and ending time, and be available for S23 to read. When the head turning record is missing but the line of sight stay record exists, only the line of sight stay record is used to form the observation action segment. When the line of sight stay record is missing but the head turning record exists, an observation missing mark is written to the acquisition time and no observation action segment is formed. In S23, the misordered action segments and remedial action segments are extracted from the action sequence table. The mechanism is to match the actual action sequence with the prescribed action sequence in the training plan table item by item, and to use whether the prescribed operation relationship is restored as the basis for the establishment of the remedial action segment. The input includes the action sequence table and the training plan table corresponding to the same training task. First, the prescribed action sequence is read according to the task segment position, and the action category order in the action sequence table is compared with the order before and after the execution of the prescribed action sequence. When a certain action category in the action sequence table appears before its preceding action category in the prescribed action sequence, the action segment is written as a misordered action segment. Then, following the misordered action segment, continue searching for steering, braking, forward / backward, and lifting action segments according to their sequence numbers. For each candidate action segment, read its post-execution vehicle position, vehicle orientation, speed, fork status, and mast status, and perform corresponding judgments against the specified travel, operation, and insertion relationships in the training plan table. When a candidate action segment, after execution, causes the current task segment to satisfy the corresponding specified relationship again, write the candidate action segment as a remedial action segment. If two or more candidate action segments participate in the recovery process sequentially, write them consecutively as the remedial action segment sequence corresponding to the same misordered action segment according to the recovery occurrence order. Subsequently, write the misordered action segment number, remedial action segment number, succession order, position of the task segment to which it belongs, and the recovery result of the corresponding specified relationship into the misordered linked list, and output the misordered linked list for S3 to read. When no candidate action segment that restores the specified relationship appears after a certain misordered action segment until the end of the current task segment, write "no remedial" for the misordered action segment and keep it in the misordered linked list for direct reading by subsequent bottom-line identification. Through the above implementation process, under a unified timeline and a unified task segment caliber, the manipulation and observation behaviors in the training records can be organized into a chain of misaligned actions with clear order, clear recovery relationship, and clear succession relationship. This allows the subsequent identification of backup items to no longer rely on a single abnormal event, but to directly focus on the formation process between the misaligned action segments and the remedial action segments. At the same time, this implementation process provides fixed criteria for the start and end of actions, the observation object, the source of the prescribed action sequence, and the conditions for the establishment of remediation, which can avoid the disconnect between the action perception results and the task rules. In practical applications: When trainees perform narrow aisle picking training tasks, the system first identifies the reversing, turning, and forklifting action segments from the training records. Then, it identifies the observation action segments facing the target cargo location and the site boundary from the head turning and line-of-sight records. If the action sequence table shows that the trainee performs the forklifting action first and then the observation action facing the target cargo location, but the training plan table stipulates that the current task segment should be observed first and then raised, then the forklifting action segment is recorded as an out-of-order action segment. If the subsequent forward and backward action segments and turning action segments cause the vehicle to return to the prescribed working relationship, then the forward and backward action segments and turning action segments are recorded as remedial action segments and written into the out-of-order chain according to their order, so that the subsequent S3 can directly identify the position return amount, the number of repeated corrections, and the insertion deviation amount.

[0019] S3. For misaligned chains, identify the support item by the position back-off amount, the number of repeated corrections, and the insertion deviation, and output the support item record. This implementation method is used to identify support items that contribute to the completion of training tasks based on misaligned action segments. The basic process is as follows: First, focusing on the transition between misaligned action segments and remedial action segments, the position backtracking amount, number of repetitions for correction, and insertion deviation are extracted from the training records. Then, a support item sublist is formed according to the support type and attribution rules. Finally, the supporting role of each support item in task completion is verified through deletion replay, and the support item record is output. This implementation process includes the following steps: In S31, the bottoming candidate quantity is extracted from the training records corresponding to the misaligned chain. The mechanism is to define the return interval by the end time of the misaligned action segment and the start time of the remedial action segment, and to calculate the additional travel, additional correction, and additional alignment within this interval. The input quantities include the misaligned chain, the training records, and the training plan table. First, the remedial action segment, the position of the task segment to which it belongs, and the succession order corresponding to each misaligned action segment are read, and the end time of the misaligned action segment and the start time of the remedial action segment are located in the training records. Then, the reverse return is executed with the end time of the misaligned action segment as the return start time and the start time of the remedial action segment as the return end time. The vehicle centerline and vehicle orientation are judged one by one in the training records within the interval to see if they deviate from the allowable channel range and allowable travel direction of the current task segment. The actual travel trajectory length between the first deviation from the specified travel relationship and the return to the specified travel relationship is summed to obtain the position backtracking amount. The number of repeated corrections is obtained by counting and summing the steering correction, parking correction, and forward / backward corrections used to restore the specified operational relationship within the same interval. Parking corrections are recorded when the vehicle changes from non-zero speed to zero speed and stops before continuing the action to restore the specified operational relationship. The insertion deviation is obtained by summing the lateral deviation distance of the fork centerline relative to the target cargo position centerline before insertion and the remaining lateral deviation distance after correction. The fork centerline is the longitudinal extension of the line connecting the midpoints of the left and right forks, and the target cargo position centerline is the longitudinal extension of the line connecting the midpoints of the left and right boundaries of the target cargo position. After processing, the mis-order action segment number, remedial action segment number, position retraction amount, number of repeated corrections, insertion deviation, position of the task segment, and start and end times of the return pointer are written into the bottoming candidate table for S32 to read. When there are consecutive missing records in the return pointer interval and the specified relationship cannot be determined, a missing mark is written to the bottoming candidate and the calculation stops. In S32, the candidate quantities for support are organized into verifiable support items. The mechanism is to first classify them according to the support type, and then complete the unique assignment of the same remedial action segment according to fixed rules. The input quantities include the candidate support table and the fault chain. First, the candidate support table is expanded according to the succession order between the faulty action segment and the remedial action segment. Records with non-zero position backoff are listed as moving support items, records with non-zero repeated corrections are listed as correction support items, and records with non-zero insertion deviation are listed as insertion support items. Then, the unique assignment process is performed on the records that are referenced by two or more faulty action segments for the same remedial action segment. Specifically, the number of interval action segments between the starting point of each candidate faulty action segment and the starting point of the remedial action segment is counted, and the faulty action segment with one interval action segment is listed as the assignment candidate. When there is only one candidate for attribution, the remedial action segment is assigned to that out-of-order action segment; when there are two or more candidates for attribution, the remedial action segment is assigned to the out-of-order action segment whose end time is later than the end time of the other out-of-order action segments; when there is no candidate for attribution, the remedial action segment is assigned to the out-of-order action segment that is registered earlier in the out-of-order chain; after processing, the out-of-order action segment number, the type of the support item, the remedial action segment number, the position of the task segment to which it belongs, and the attribution rule number are written into the support sublist for S33 to read; when all three additional quantities of the same support candidate are zero, the support candidate is deleted and not written into the support sublist. In S33, the mechanism for determining whether a support item actually supports the completion of the training task involves deleting each support item one by one and reconstructing the operation process, then using the recovery result of the specified relationship and the task completion result as the basis for determination. Inputs include a support item sublist, training records, and a training plan table. First, a replay copy is created for each support item. When deleting a traveling support item, only the additional traveling trajectory corresponding to the position retraction amount is removed; when deleting a corrective support item, only the corresponding repeated corrective action record is removed; when deleting an insertion support item, only the corresponding lateral alignment correction record is removed while retaining the fork entry depth record. Then, in the remaining training... The operation process is reconstructed while the training record remains unchanged. The prescribed travel relationship, prescribed insertion relationship and training task completion status are judged at each moment. The prescribed travel relationship is not restored, which means that the vehicle centerline has not fallen back into the allowed passage range of the current task segment or the vehicle orientation has not re-corresponded to the allowed travel direction of the current task segment. The prescribed insertion relationship is not restored, which means that the fork centerline has not re-corresponded to the target cargo position centerline or the fork entry depth has not fallen into the entry depth range required by the task segment. The training task is not completed, which means that the termination position has not been reached in the order of the task, the target cargo position insertion has not been completed, or the loading status requirements have not been met. Any support item that results in any of the above outcomes after deletion is recorded as a successful support item. Subsequently, successful support items are summarized according to the same out-of-order action segment. For two or more successful support items, they are written in the order of progress support item, correction support item, and insertion support item. For out-of-order action segments without successful support items, an empty support mark is written. The empty support mark includes at least the out-of-order action segment number, the empty support mark, and the reason code for failure. Finally, the out-of-order action segment number, successful support item type, corresponding remedial action segment number, the position of the task segment, the position backoff amount, the number of repeated corrections, the insertion deviation amount, or the empty support mark is written into the support item record for S4 to read. When the specified relationship cannot be determined due to missing records during the replay process, a replay abort mark is written to the corresponding support item, but it is not recorded as a successful support item. Through the above implementation process, the additional corrective, additional adjustment and additional alignment components that support the erroneous operation methods during the completion of the training task can be solved respectively, and the bottom-line judgment can be transformed into a verifiable result by deleting the replay, thereby providing a stable input for the reconstruction of the subsequent retraining task; In practical applications: When a trainee performs a staggered lifting action in a pickup task, and then completes the pickup by reversing, stopping to correct, and moving forward and backward again, the system first calculates the position back-off amount, the number of repeated corrections, and the insertion deviation between the end time of the staggered action segment and the start time of the remedial action segment. Then, it forms the travel support item, correction support item, and insertion support item. Subsequently, it deletes the corresponding additional trajectory, correction record, and alignment correction record one by one and reconstructs the operation process. For support items that result in the failure to restore the specified travel relationship, the failure to restore the specified insertion relationship, or the failure to complete the training task after deletion, they are written into the support item record for subsequent S4 to reduce the return space, the number of repeated corrections, and the alignment margin.

[0020] S4. Based on the bottom-line item record, while keeping the task objective unchanged, reduce the return space, the number of repeated corrections and the alignment margin, and output the retraining task. This implementation method is used to perform controlled reconstruction of the original training task based on the baseline item records. While maintaining the same training task objective, it directionally compresses the correction space, repetition correction count, and alignment margin supporting erroneous operation methods, thereby generating a task structure suitable for retraining. The basic process is as follows: First, extract the target-preserving items and reduced segments required for reconstruction from the baseline item records and training records. Then, rewrite the correction space, correction sequence, and alignment range according to the baseline item type. Finally, through forward replay and backward indexing verification, candidate results that satisfy the task continuity and the correspondence with the original baseline items are screened out, and the retraining task is output. This implementation process includes the following steps: In S41, a unified input basis for retraining and reconstruction is established. The mechanism is to write the target information that must remain unchanged in the training task and the task segment information that needs to be reduced into the same structure. The input includes the bottom-line item record and the training record. First, the bottom-line item record reads the misordered action segment number, remedial action segment number, bottom-line item type, the position of the task segment to which it belongs, the position back-off amount, the number of repeated corrections, and the insertion deviation amount corresponding to each bottom-line item. Then, the training record reads the starting position, target cargo location, ending position, loading status, and task sequence corresponding to the same training task. Among them, the starting position is the vehicle position corresponding to the first record of the training task, the target cargo location is the target cargo location identifier and centerline parameter corresponding to the training task in the training plan table, the ending position is the ending position identifier corresponding to the training task in the training plan table, the loading status is the empty or loaded status required when the training task is completed, and the task sequence is the registration order of each task segment of the training task in the training plan table. Subsequently, the starting position, target cargo location, ending position, cargo loading status, and task sequence are written as target holding items. The corresponding task segment positions are then written as centering reduction segments, correction reduction segments, and alignment reduction segments according to the type of support item. Specifically, moving support items correspond to centering reduction segments, correction support items correspond to correction reduction segments, and inserted support items correspond to alignment reduction segments. After processing, the task number, target holding item, support item type, out-of-order action segment number, remedial action segment number, task segment position, and corresponding reduction segment identifier are written to the reconstruction base table for S42 to read. When a support item record contains an empty support mark, only the corresponding task segment position and target holding item are retained; the reduction segment identifier is not written. In S42, the original training task is rewritten according to the type of the support item. The mechanism is to reduce the alignment space, correction sequence and alignment range respectively, and to concatenate the rewriting results of multiple types in a fixed order. The input includes the reconstruction base table and training records. First, the alignment space is rewritten for each alignment reduction segment. Specifically, the corresponding position back-off amount and the original starting boundary and the original ending boundary of the task segment are read. The alignment space is rewritten in such a way that the original starting boundary remains unchanged and the original ending boundary is translated in the opposite direction of the position back-off amount. The translation direction corresponding to the position back-off amount is the reverse of the back-off direction after the vehicle deviates from the specified driving relationship. Then, the correction sequence is rewritten for each correction reduction segment. Specifically, the corresponding number of repeated corrections and the sequence of remedial action segments are read. The same type of correction action is sorted according to the order of appearance. The second and subsequent same type of correction actions are deleted, and the first same type of correction action is retained. The same type of correction action is judged independently according to steering correction, parking correction and forward and backward correction, and is not merged across categories. Next, the alignment range is rewritten for each alignment reduction segment. Specifically, the corresponding insertion deviation, the target cargo location centerline, and the original lateral entry range are read. Using the target cargo location centerline as a reference, the lateral distance corresponding to the insertion deviation is subtracted equidistantly from the left and right sides of the original lateral entry range towards the centerline, while keeping the fork entry depth range unchanged, to form the rewritten alignment range. When there are two or more rewriting results for the same task segment, they are written in the order of alignment space, correction sequence, and alignment range. The end position, end orientation, end fork status, and end mast status of the previous rewriting result are used as the input for the next rewriting result. After processing, the rewritten alignment space, rewritten correction sequence, rewritten alignment range, and end status of each task segment are written into the candidate retraining table for S43 to read. When a certain type of rewriting causes the task segment boundary to cross in the opposite direction or the lateral entry range to disappear, a rewriting failure mark is written for this type of rewriting and the original task segment record is retained. In S43, rewrite results that satisfy the continuity and correspondence with the original support item are selected from the candidate retraining table. The mechanism is to check the continuity of the task through forward replay and check the consistency of the reduced segment attribution through reverse back-pointing verification. Rewrites that fail the verification are cancelled in a fixed order. The input includes the candidate retraining table, training records, and support item records. First, forward replay is performed on each task segment in the candidate retraining table in the order of the tasks. Specifically, starting from the starting position, the starting state and ending state of the rewritten segment are read one by one. It is determined whether the ending state of the previous task segment is continuously connected with the starting state of the next task segment. It is also determined whether the last task segment has reached the target cargo location, the termination position, the cargo loading status, and the target maintenance item corresponding to the task order. At the same time, reverse back-pointing verification is performed on each reduced segment. Specifically, the rewritten record of each reduced segment is back-pointed to the original support item record. It is determined whether the reduced segment still corresponds to the task segment position of the original support item and whether the rewritten content only acts on the task segment position and does not cross to other task segment positions. For candidate results that simultaneously satisfy both forward replay and backward anaphora, connect them end-to-end in the task order to form a retraining task; for candidate results that do not satisfy forward replay, rewrite them item by item in reverse order of alignment range, correction sequence, and return space, and re-execute forward replay and backward anaphora after each item is rewritten; for candidate results that do not satisfy backward anaphora, rewrite the cross-segment part and re-execute forward replay and backward anaphora; until each task segment has formed a rewritten result that passes the verification, write the task number, the position of each task segment, the rewritten result of each task segment, and the overall connection result into the retraining task table, output the retraining task, and let S5 read it; when all rewrites are rewritten and still fail forward replay, write the original task segment mark to the task segment and continue the verification of subsequent task segments; Through the above implementation process, without changing the starting position, target cargo location, ending position, cargo loading status and task sequence, the supporting components of the original training task that support the incorrect operation mode can be compressed in a targeted manner. Through forward replay and reverse feedback, it can be ensured that the retraining task maintains the original training objective and does not deviate from the original supporting item's position in the task segment, thereby providing a stable task carrier for subsequent hazard manifestation and limitation remedial actions. In practical applications: When trainees rely on the reversing space, multiple steering corrections, and lateral alignment margin to complete the insertion in the cargo retrieval training task, the system first extracts the corresponding task segment position, position back-off amount, number of repeated corrections, and insertion deviation from the bottom-up item record, and writes the starting position, target cargo position, ending position, cargo loading status, and task sequence as target retention items; then, the reversing space is reduced in the reversing reduction segment, the second and subsequent steering correction actions are deleted in the correction reduction segment, and the lateral entry range is simultaneously deducted from both sides of the target cargo position in the alignment reduction segment; then, the rewrite results are re-performed by forward replay and reverse back-pointing verification, and the rewrites that fail are reversed in the order of alignment range, correction sequence, and reversing space, until a retraining task that passes the verification is formed, and the retraining task is sent to the subsequent S5 to perform the hazard manifestation segment positioning and remedial action restrictions.

[0021] S5. During the retraining task, the current action is compared in real time through action perception, and combined with the sequence prediction of the wrong chain execution, the danger manifestation segment is located, a prompt is issued before the danger manifestation segment is reached, and the corresponding remedial action is restricted when the wrong sequence action occurs again, and the retraining record is output. This implementation method is used to, during the execution of a retraining task, locate the task interval in advance where the erroneous operation will be manifested again based on the real-time correspondence between the current action and the existing error chain, issue a prompt before the interval is reached, and restrict the corresponding remedial action within the interval, thereby forming a retraining record that can be directly compared and read for subsequent results. Its basic process is as follows: First, the current manipulation input and the corresponding currently observed action during the execution of the retraining task are categorized and registered in sequence. Then, the current sequence is matched item by item with the succession order in the error chain to locate the danger manifestation segment and the prompt time. Finally, a sequence prompt is issued at the prompt time, and when the erroneous action is identified again, the corresponding remedial action is restricted and written into the retraining record. This implementation process includes the following steps: In S51, real-time actions during the retraining task are converted into current sequence results that can be directly correlated with the fault chain. The mechanism involves using motion perception to form a unified action category from the current operational input and the corresponding current observation action, then performing sequence registration according to the order of occurrence and comparing it item by item with the fault chain. Inputs include current vehicle position, current vehicle orientation, current travel speed, current fork status, current mast status, current operational input, current head turning record, and current gaze dwell record, as well as the fault chain. First, the current head turning record and current gaze dwell record form the corresponding action for the current observation. When the current gaze dwell direction is consistent with the current head orientation, the action is directly correlated with the current head orientation. Next, write the direction as the current observation corresponding action. When the current line of sight is inconsistent with the current head orientation, write the current line of sight as the current observation corresponding action and write the current head orientation as an auxiliary field. Then, classify the current operation input and the current observation corresponding action. Write the current steering input as the current steering action, the current braking input as the current braking action, the current vehicle forward or backward state as the current forward or backward action, the current fork state or the current mast state of lifting, tilting forward or backward change as the current lifting action, and write the current observation corresponding action according to the direction of travel, fork direction, target cargo position direction and site boundary direction as the current observation action. Subsequently, the above actions are sequentially registered according to the current time sequence in the unified timeline. When two or more types of actions occur at the same time, they are written into the current sequence table in a fixed order of observation action, turning action, braking action, forward and backward action, and lifting action. Then, the action category, sequence number, and registration time in the current sequence table are compared item by item with the out-of-order actions, remedial actions, and succession order in the staggered chain. The current sequence table is output and read by S52. The current sequence table includes at least the task number, task segment position, current action category, current sequence number, registration time, and corresponding staggered chain number. When the current gaze dwell record is missing, only the current head turning record is used to form an auxiliary observation field and no current observation action is generated. When the current manipulation input is missing, an action missing mark is written to the current time and the sequence registration for that time is stopped. In S52, the mechanism for locating the danger manifestation segment and prompting time based on the current sequence table and the fault chain is to match the currently registered actions with the consecutive sequence in the fault chain item by item. When the consecutive correspondence is established, it directly points back to the manifestation interval after the starting point of the remedial action in the fault chain. The input includes the current sequence table, the fault chain, and the training record. First, the registered action sequence is read according to the registration order in the current sequence table. Then, the faulty action, subsequent action, and remedial action are read in sequence according to the succession order in the fault chain, and the two are matched item by item. When the registered action category in the current sequence table is consistent with the consecutive sequence in the fault chain item by item, the task interval between the starting time of the remedial action after the consecutive sequence and the corresponding completion result inflection point time is determined as the danger manifestation segment. The completion result inflection point time is the first collection time in the training record where the training task returns to the completed state after the remedial action is executed. Next, read back the action switching time immediately preceding the start time of the remedial action from the current sequence table and write that action switching time as the prompt time; then write the fault chain number, the start time of the danger manifestation segment, the end time of the danger manifestation segment, the prompt time, the corresponding remedial action number, and the corresponding completion result inflection point time into the manifestation positioning table for S53 to read; when the consecutive succession order in the current sequence table and the fault chain cannot be consistent item by item, no danger manifestation segment is generated, and an unmanifested mark is written to the corresponding fault chain; when the completion result inflection point time is missing in the training record, a missing inflection mark is written to the corresponding manifestation positioning result and it is not included in the subsequent prompt release; In S53, a warning is issued before the danger manifest segment is reached, and the corresponding remedial action is restricted within the danger manifest segment. The mechanism is to interrupt the continuation of the incorrect operation mode at the warning time, and then cut off the remedial path corresponding to the wrong sequence when the wrong sequence action occurs again. The input includes the manifest positioning table, the current sequence table, the current operation input, and the retraining task execution control link. First, at the warning time, the corresponding wrong sequence number and wrong sequence action category in the manifest positioning table are read, and the sequence warning content corresponding to the wrong sequence action category is generated and published to the retraining task execution terminal. The sequence warning content includes at least the current task segment position, the prescribed action category that should be executed first, and the currently occurring wrong sequence action category. The current sequence table is continuously read within the hazard manifestation segment. When a current action of the same category and position as the misordered action in the corresponding misaligned chain is identified again, remedial action restriction processing is performed. Specifically, the remedial action input corresponding to the misaligned chain is blocked from being written into the execution queue, the duplicate correction instruction corresponding to the remedial action is deleted, and other manipulation inputs that do not belong to the remedial action are retained. Here, re-identifying a misordered action means that within the same hazard manifestation segment, after a corresponding misordered action has been registered once, a misordered action of the same category and position is registered again. After processing, the prompt time, the start and end times of the hazard manifestation segment, the time when the misordered action reappears, the number of the blocked remedial action input, the number of the deleted duplicate correction instruction, and the number of the retained non-remedial action input are written into the retraining record for S6 to read. When no misordered action is identified again within the hazard manifestation segment, only the prompt release result is written into the retraining record, and remedial action restriction is not performed. Through the above implementation process, a real-time connection and correspondence can be established between the current action and the historical errors during the retraining task. Before the erroneous operation mode is re-expanded but not yet remedied, the danger manifestation segment can be located, prompts can be issued and remedial actions can be restricted. This allows the erroneous operation mode that was originally covered by the training task completion result to be stably manifested during the retraining process. At the same time, it provides a unified standard for retraining records for subsequent comparison of completion results and comparison of manifestation results. In practical applications: When a trainee performs a lifting action before observing the target cargo location during a retraining task, and this sequence matches the continuous sequence in the existing misaligned chain, the system writes the task interval between the start time of the remedial action corresponding to the misaligned chain and the turning point of the completion result as a danger manifestation segment, and writes the action switching time immediately preceding the start time of the remedial action as a prompt time; at the prompt time, the system issues a "observe first, then lift" sequence prompt to the training terminal; if the trainee performs the same misaligned lifting action at the same location again within the danger manifestation segment, the system blocks the remedial turning input corresponding to the misaligned chain from being written into the execution queue, deletes the corresponding duplicate correction instruction, retains the remaining non-remedial action inputs, and writes the prompt time, danger manifestation segment, the time when the misaligned action reappears, and the remedial action restriction result into the retraining record for subsequent S6 to identify erroneous operation methods masked by the completion result.

[0022] S6. Compare the completion results of the retraining record with the training record and the explicit results to identify the erroneous operation methods that are covered by the completion results of the training task, generate subsequent intervention content, and output the training intervention results; This implementation method is used to compare training records and retraining records to identify erroneous operation methods masked by the completion results of training tasks, and generate subsequent intervention content accordingly. The basic process is as follows: First, extract result fields such as completion status, remedial actions, fallback items, and explicit segments from the training and retraining records to form a directly comparable result comparison table; then, distinguish between completion masking chains and explicit masking chains based on the result comparison table, and merge them according to the type of out-of-order action, remedial action type, and task segment position to form an error method table; finally, generate sequential prompts, restricted training content, and retraining insertion content based on the error method table, and write them into the intervention release table according to the task order of the same training task, outputting the training intervention results; this implementation process includes the following steps: In S61, a unified comparison basis is established between training records and retraining records. The mechanism is to map the completion results, remedial results, and explicit results in the two types of records to the same control structure according to the same task caliber. The input includes training records, retraining records, and support item records. First, the original task completion status, original remedial action sequence, and original support item distribution are extracted based on the training records. The original task completion status includes at least whether the terminal position was reached in the task order, whether the target cargo location insertion or stacking was completed, and whether the cargo loading status requirements were met. The original remedial action sequence includes at least the remedial action category, the order of occurrence, and the position of the task segment to which it belongs. The original support item distribution includes at least the support item type and the position of the task segment to which it belongs. Next, based on the retraining records, the retraining completion status, the retraining out-of-order action sequence, the retraining remedial action sequence, and the retraining explicit segment distribution are extracted. The explicit segment distribution includes at least the start time, end time, and position of the corresponding task segment. Then, a correspondence comparison is performed between the original task completion status and the retraining completion status. A deletion comparison is performed between the original remedial action sequence and the retraining remedial action sequence. The deletion comparison determines whether to retain, delete, or restrict actions based on their category, order of appearance, and position within the corresponding task segment. A same-segment correspondence is performed between the original support item distribution and the retraining explicit segment distribution, with the condition that the position of the support item matches the position of the explicit segment. After processing, the task number, original task completion status, retraining completion status, remedial action deletion result, support item correspondence result, and explicit segment correspondence result are written into a result comparison table for S62 to read. When a training record or retraining record lacks a task segment position field, a position missing marker is written to the corresponding record, and the correspondence comparison for that record is stopped. In S62, the disordered action chains masked by the completed results are determined from the result comparison table. The mechanism is to extract the two types of masking chains and merge them to form an error mode table, based on the difference in the completed results and the expansion of the explicit results respectively. The input includes the result comparison table, training records and retraining records. First, each disordered action chain in the result comparison table is traversed. Disordered action chains that are completed in the training records but not completed in the retraining records are written as completion masking chains. Here, completion in the training records means that the original task completion state is established, and incompleteness in the retraining records means that the retraining completion state is not established. Then, for disordered action chains that have remedial actions in the training records and whose corresponding remedial actions are restricted in the retraining records, an explicit judgment is performed. When the explicit segment corresponding to the same disordered action chain appears continuously in the retraining records along the task order and no completion state recovery record is written in the middle, the disordered action chain is written as an explicit masking chain. Subsequently, the completion masking chain and the explicit masking chain are merged and categorized according to the out-of-order action category, the remedial action category, and the task segment position. Records with all three fields are merged into the same error operation mode, while records with any inconsistency in any field are listed as different error operation modes. After processing, the error mode number, out-of-order action category, remedial action category, task segment position, masking type, and corresponding out-of-order action chain number are written into the error mode table for S63 to read. When the same out-of-order action chain simultaneously meets the conditions of completion masking chain and explicit masking chain, both masking types are written into the error mode table. In S63, subsequent intervention content is generated based on the error mode table. The mechanism is to generate three types of content—sequential prompts, restricted training, and retraining insertions—based on the error sequence action category, remedial action category, and task segment position, respectively, and form an executable release result according to the task order. The input includes the error mode table and the training plan table. First, the error sequence action category is read for each error operation mode, and sequential prompt content corresponding to the error sequence action category is generated. The sequential prompt content includes at least the task segment position, the prescribed action category that should be executed first, and the current error action category. Then, the remedial action category is read for each error operation mode, and restricted training content corresponding to the remedial action category is generated. The restricted training content includes at least the restricted remedial action category, the task segment position where the restriction takes effect, and the restriction method. Subsequently, the task segment position is read for each erroneous operation method, and corresponding retraining insertion content is generated. The retraining insertion content is written before the task segment corresponding to the task segment position in subsequent rounds of the same training task. Finally, the sequence prompt content, restricted training content, and retraining insertion content are written into the intervention release table in the task order of the same training task, and the task number, error method number, release order, and three types of intervention content are written as the training intervention result output. When there is an empty task segment position record in the error method table, only the sequence prompt content and restricted training content are generated for the corresponding erroneous operation method, and no retraining insertion content is generated. Through the above implementation process, the completion results in the training record and the explicit results in the retraining record can be uniformly converted into verifiable error operation methods, and further generate subsequent intervention content directly corresponding to the error operation methods, so that subsequent training will no longer focus solely on the surface failure results, but rather on the error operation methods that are masked by the completion results. In practical applications: When the training record shows that the trainee has completed the training task in the target location insertion task, and there is a sequence of turning remedial actions and a distribution of supporting items, while the retraining record shows that the training task was not completed after the remedial actions were restricted in the same task segment position, and the explicit segment continues to expand along the task segment, the system first writes the difference in completion status, the result of the reduction of remedial actions, and the corresponding result of the supporting items and the same segment of the explicit segment into the result comparison table. Then, the system records the disordered action chain as both the completion masking chain and the explicit masking chain. Subsequently, the system generates sequential prompts according to the corresponding disordered action category, generates restricted training content according to the corresponding remedial action category, and generates retraining insertion content before the task segment position. Finally, the system writes the results into the intervention release table according to the task order to form the training intervention results.

[0023] Furthermore, the present invention also includes a forklift training system based on dangerous operation prediction and adaptive intervention, the system comprising a record organization module, an error correction and identification module, a bottoming-out identification module, a reconstruction module, an explicit intervention module, and a result module: The record organization module is used to obtain the vehicle position, vehicle orientation, driving speed, fork status, mast status, site boundary relationship and operation input corresponding to the same training task, and to organize the execution order through time alignment and output the training record. The error correction module is used to perform sequential recognition of steering, braking, forward and backward, lifting and lowering actions and observation corresponding actions based on training records and through action perception. It also performs error recognition and correction recognition according to the prescribed action sequence and outputs the error chain. The bottom-support identification module is used to identify bottom-support items for misaligned chains by measuring the position back-off amount, the number of repeated corrections, and the insertion deviation, and outputs bottom-support item records. The reconstruction module is used to reduce the back-to-center space, the number of repeated corrections, and the alignment margin based on the support item record, while keeping the task objective unchanged, and output the retraining task. The explicit intervention module is used to compare the current action execution in real time through action perception during the retraining task, and combined with the sequence prediction of the wrong chain execution to locate the dangerous explicit segment, issue a warning before the dangerous explicit segment is reached, restrict the corresponding remedial action when the wrong sequence action occurs again, and output the retraining record. The results module is used to compare the completion results of retraining records with those of training records and to compare the results of explicit training. It identifies erroneous operation methods that are masked by the completion results of training tasks, generates subsequent intervention content, and outputs training intervention results.

[0024] Working principle: This solution first compiles the vehicle status, fork and mast status, site relationships, and trainee operations from a single forklift training session into a training record. Then, it identifies the trainee's steering, braking, forward / backward, lifting, and observation sequence, pinpointing which actions were performed incorrectly and which remedial actions were used to complete the task. Next, the system further determines which of these remedial actions constitute "bottleneck" components—extra clearance, repeated corrections, and alignment allowances. Without these backups, the task would not have been completed. Based on this result, the system maintains the original training objective and generates a more rigorous retraining task, tightening the previously limited space and margin for the trainee to complete the task. During retraining, if the trainee is found to be operating in the same incorrect sequence, an early warning is issued, and corresponding remedial actions are restricted. Finally, the original training and retraining results are compared to confirm the truly incorrect operating methods masked by the completed results. For example, when a trainee is retrieving goods in a narrow aisle, they may appear to have successfully lifted the pallet, but in reality, they made a mistake in the initial steps and only managed to correct it by repeatedly reversing, turning the steering wheel, and realigning the pallet multiple times. The system will identify these corrective actions during the first training session. In the second retraining session, the trainee will still perform the same retrieving task, but with reduced space for correction, fewer repetitive corrections, and smaller alignment margins. In this way, if the trainee continues to operate using the original incorrect method, the problem will be exposed earlier, and the system can provide timely warnings and limit the trainee's reliance on corrective actions. In other words, this solution does not only consider whether the task was completed successfully in the end, but specifically identifies the most easily overlooked problems where the method was incorrect despite the apparent success.

[0025] The above description is merely 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 forklift training method based on hazardous operation prediction and adaptive intervention, characterized in that, include: S1. Obtain the vehicle position, vehicle orientation, driving speed, fork status, mast status, site boundary relationship and operation input corresponding to the same training task, align the execution order according to the time sequence, and output the training record. S2. Based on the training records, the sequence recognition of steering, braking, forward and backward, lifting and lowering actions and observation actions is performed through motion perception. The sequence recognition and remediation recognition are performed according to the prescribed action sequence, and the error chain is output. S3. For misaligned chains, identify the support item by the position back-off amount, the number of repeated corrections, and the insertion deviation, and output the support item record. S4. Based on the bottom-line item record, while keeping the task objective unchanged, reduce the return space, the number of repeated corrections and the alignment margin, and output the retraining task. S5. During the retraining task, the current action is compared in real time through action perception, and combined with the sequence prediction of the wrong chain execution, the danger manifestation segment is located, a prompt is issued before the danger manifestation segment is reached, and the corresponding remedial action is restricted when the wrong sequence action occurs again, and the retraining record is output. S6. Compare the completion results of the retraining record with the training record and the explicit results to identify the erroneous operation methods that are covered by the completion results of the training task, generate subsequent intervention content, and output the training intervention results; S3 includes: S31. Around each mis-ordered action segment in the mis-ordered chain, take the starting time of the corresponding remedial action segment as the return point and the ending time of the corresponding mis-ordered action segment as the return point. Perform reverse return on the training record. Accumulate the additional travel distance corresponding to the return to the specified travel relationship after the return point and the return point deviate from the specified travel relationship to obtain the position return amount. Count the number of times the steering correction, stopping correction and forward / backward correction occur between the return point and the return point to restore the specified operation relationship to obtain the number of repeated corrections. Accumulate the lateral deviation distance of the fork centerline relative to the target cargo position centerline before and after insertion and the remaining lateral deviation distance after correction to obtain the insertion deviation amount. Output the bottoming candidate table. S32. Based on the candidate list for bottoming out, perform bottoming out by following the sequence of each out-of-order action segment and its corresponding remedial action segment. Records with non-zero position backoff are listed as moving bottoming out items, records with non-zero number of repeated corrections are listed as correction bottoming out items, and records with non-zero insertion deviation are listed as insertion bottoming out items. For records where the same remedial action segment is referenced by two or more out-of-order action segments, perform unique attribution processing. Specifically, the remedial action segment is assigned to an out-of-order action segment with one interval action segment between the retracement start point and the start point of the remedial action segment. When there are two or more out-of-order action segments with one interval action segment, the remedial action segment is assigned to an out-of-order action segment whose end time is later than the end time of other out-of-order action segments. Output the bottoming out list. S33. Perform deletion replay for each support item in the support sublist. Delete each support item from the corresponding training record one by one while keeping the remaining records unchanged, and then reconstruct the operation process. For support items that fail to restore the specified movement relationship, the specified insertion relationship, or the training task after deletion, record them as established support items. For two or more established support items corresponding to the same misordered action segment, write them in the order of movement support item, correction support item, and insertion support item. Write an empty support mark for misordered action segments without established support items, and output the support item record.

2. The forklift training method based on hazardous operation prediction and adaptive intervention according to claim 1, characterized in that: S1 includes: S11. Based on the collection time, perform simultaneous classification of vehicle position, vehicle orientation, driving speed, fork status, mast status, site boundary relationship and operation input, extract the time of vehicle orientation change, driving speed change, fork status switch, mast status switch and site boundary relationship change, and determine the time when two or more types of changes occur at the same collection time as common time, and output the common time sequence. S12. Using the acquisition records between adjacent common moments in the common moment sequence as the same segment record, the operation inputs in each same segment record are ranked in the order of appearance of steering input, driving input, fork input and gantry input, and the acquisition records that do not fall into the common moment are attached to the previous or next common moment according to the time adjacency relationship, and the sequence segment record is output. S13. Perform head-to-tail concatenation on the sequence records according to the common time sequence, and write the segment number, input order within the segment, corresponding vehicle position, corresponding vehicle orientation, corresponding driving speed and corresponding site boundary relationship in each sequence record into a unified record format, and output the training record.

3. The forklift training method based on hazardous operation prediction and adaptive intervention according to claim 2, characterized in that: S2 includes: S21. Based on the changes in the control input, vehicle position, vehicle orientation, driving speed, fork status and mast status corresponding to the training records, the starting point and ending point of the action are identified at each collection time through motion perception. Adjacent collection times that change continuously in the same direction are merged into the same action segment, and the action segments are formed into steering action segments, braking action segments, forward and backward action segments and lifting action segments according to the action category. The action segment sequence is output. S22. Obtain the trainee's head turning record and gaze dwell record corresponding to the training record. Perform observation corresponding position recognition on the head turning record and gaze dwell record through motion perception. Form observation action segments from the continuous observation records facing the direction of travel, fork direction, target cargo location direction and site boundary direction respectively. Perform joint ranking according to the sequence of action segments and the order of observation action segments, and output the action order table. S23. Based on the prescribed action sequence corresponding to the same training task, perform a front-to-back sequence comparison on the action sequence table, determine the action segment that appears before the prescribed sequence as the misordered action segment, and determine the turning action segment, braking action segment, forward and backward action segment or lifting action segment that is located after the misordered action segment and is used to restore the prescribed operation relationship as the remedial action segment, and output the misordered chain according to the succession order between the misordered action segment and the remedial action segment.

4. The forklift training method based on hazardous operation prediction and adaptive intervention according to claim 3, characterized in that: S4 includes: S41. Extract the positions of the misordered action segments, remedial action segments, and task segments corresponding to each support item from the support item record. Extract the starting position, target cargo location, ending position, cargo loading status, and task sequence corresponding to the same training task from the training record. Write the starting position, target cargo location, ending position, cargo loading status, and task sequence as target retention items. Then, write the corresponding task segments as correction reduction segments, correction reduction segments, and alignment reduction segments according to the support item type. Output the reconstruction base table. S42. Based on the reconstruction base table, for each straightening reduction segment, shift the segment's end boundary in the opposite direction of the corresponding position's backtracking amount while keeping the segment's starting boundary unchanged to rewrite the straightening space. For each correction reduction segment, delete the second and subsequent similar correction actions in the remedial action segment according to the corresponding number of repeated corrections, and retain the first similar correction action to rewrite the correction sequence. For each alignment reduction segment, simultaneously deduct the lateral entry range from both sides of the target location with the corresponding insertion deviation amount while keeping the fork entry depth range unchanged to rewrite the alignment range. When there are two or more rewriting results for the same task segment, write them in the order of straightening space, correction sequence, and alignment range, and use the segment's end state of the previous rewriting result as the input of the next rewriting result to output the candidate retraining table.

5. The forklift training method based on hazardous operation prediction and adaptive intervention according to claim 4, characterized in that: S4 also includes: S43. Based on the candidate retraining table, perform forward replay and reverse anaphora verification on each task segment. Forward replay is used to verify whether the beginning and end of each task segment are continuously connected in the task order and reach the target retention item. Reverse anaphora is used to verify whether each reduced segment still corresponds to the task segment where the original support item is located and has not been rewritten across segments. For candidate results that simultaneously satisfy forward replay and reverse anaphora, connect them in the order of task segments to form a retraining task. For candidate results that do not satisfy forward replay, cancel and rewrite them item by item in the reverse order of alignment range, correction sequence, and return space, and repeat forward replay and reverse anaphora after each cancellation. For candidate results that do not satisfy reverse anaphora, cancel the cross-segment rewriting part and repeat forward replay and reverse anaphora until each task segment forms a rewritten result that passes the verification, and output the retraining task.

6. The forklift training method based on hazardous operation prediction and adaptive intervention according to claim 5, characterized in that: S5 includes: S51. During the retraining task, obtain the current vehicle position, current vehicle orientation, current driving speed, current fork status, current mast status, current control input, and current observed corresponding action. Through action perception, classify and register the current control input and current observed corresponding action, and compare the current action with the out-of-order actions, remedial actions, and succession order in the chain of errors, and output the current order table. S52. Based on the current sequence table, perform forward continuation expansion on the misaligned chain, and match the registered action sequence in the current sequence table with the continuation sequence in the misaligned chain according to the registration order; when each registered action sequence is consistent with the continuous continuation sequence in the misaligned chain, determine the task segment between the starting time of the remedial action after the continuous continuation sequence and the corresponding completion result turning point time as the danger manifestation segment, and determine the action switching time immediately before the starting time of the remedial action as the prompt time, and output the manifestation positioning table; S53. For the explicit positioning table, issue a sequence prompt associated with the corresponding misordered action at the prompt time, and when the misordered action is identified again in the danger explicit segment, block the writing of the remedial action input corresponding to the misordered chain, delete the repeated correction instruction corresponding to the remedial action and retain the non-remedial action input, and write the prompt time, danger explicit segment, time when the misordered action reappears and the remedial action restriction result into the retraining record.

7. The forklift training method based on hazardous operation prediction and adaptive intervention according to claim 6, characterized in that: S6 includes: S61. Based on the training records, extract the original task completion status, the original remedial action sequence, and the original support item distribution. Based on the retraining records, extract the retraining completion status, the retraining out-of-order action sequence, the retraining remedial action sequence, and the retraining explicit segment distribution. Perform a correspondence comparison between the original task completion status and the retraining completion status. Perform a deletion comparison between the original remedial action sequence and the retraining remedial action sequence. Perform same-segment correspondence between the original support item distribution and the retraining explicit segment distribution. Output a result comparison table. S62. Based on the result comparison table, the disordered action chains that are completed in the training record but not completed in the retraining record are recorded as completion masking chains. The disordered action chains that have remedial actions in the training record and whose corresponding remedial actions are restricted and then continuously expanded in the retraining record are recorded as explicit masking chains. The completion masking chains and explicit masking chains are merged and classified according to the disordered action category, remedial action category and task segment position, and the error mode table is output. S63. Based on the error mode table, generate sequential prompts for each error operation mode according to the corresponding wrong sequence action category, generate restricted training content according to the corresponding remedial action category, and generate retraining insertion content according to the corresponding task segment position. Write the sequential prompts, restricted training content, and retraining insertion content into the intervention release table according to the task order of the same training task, and output the training intervention results.

8. A forklift training system based on hazardous operation prediction and adaptive intervention, used to implement the forklift training method based on hazardous operation prediction and adaptive intervention as described in any one of claims 1-7, the system comprising a record processing module, an error correction and identification module, a bottoming-out identification module, a reconstruction module, an explicit intervention module, and a result module, characterized in that: The record organization module is used to obtain the vehicle position, vehicle orientation, driving speed, fork status, mast status, site boundary relationship and operation input corresponding to the same training task, and to organize the execution order through time alignment and output the training record. The error correction module is used to perform sequential recognition of steering, braking, forward and backward, lifting and lowering actions and observation corresponding actions based on training records and through action perception. It also performs error recognition and correction recognition according to the prescribed action sequence and outputs the error chain. The bottom-support identification module is used to identify bottom-support items for misaligned chains by measuring the position back-off amount, the number of repeated corrections, and the insertion deviation, and outputs bottom-support item records. The reconstruction module is used to reduce the back-to-center space, the number of repeated corrections, and the alignment margin based on the support item record, while keeping the task objective unchanged, and output the retraining task. The explicit intervention module is used to compare the current action execution in real time through action perception during the retraining task, and combined with the sequence prediction of the wrong chain execution to locate the dangerous explicit segment, issue a warning before the dangerous explicit segment is reached, restrict the corresponding remedial action when the wrong sequence action occurs again, and output the retraining record. The results module is used to compare the completion results of retraining records with those of training records and to compare the results of explicit training. It identifies erroneous operation methods that are masked by the completion results of training tasks, generates subsequent intervention content, and outputs training intervention results.

Images