An adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights
The adaptive closed-loop training system, which uses dynamic fusion of multiple evidence-based weights, solves the problems of indistinguishable causes of generalization failure and cancellation of training units in collaborative training between institutions and families, and achieves dynamic optimization and effective execution of training strategies.
Patent Information
- Application Number
- CN202611534472.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-30
- Publication Date
- 2026-11-17
AI Technical Summary
Existing rehabilitation training systems, when used in collaboration between institutions and families, cannot effectively isolate the causes of generalization failure, leading to problems such as limited detection opportunities being mismatched and multiple evidence-based training units canceling each other out in the same session.
An adaptive closed-loop training system employing dynamic fusion of multiple evidence-based weights is used to standardize and analyze training data from both institutional and home environments through a data collection and normalization module, a profile construction module, a weight orchestration module, and an execution control module. This allows for dynamic adjustment of training strategies, ensuring independent execution and effective generalization of each training unit.
It enables accurate identification and separation of the causes of failure at the home end, optimizes the arrangement of training strategies, avoids mutual cancellation of the same lesson, and improves the effectiveness of training and the success rate of execution at the home end.
Smart Images

Figure CN123120210A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer application technology, and more specifically, to an adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights. Background Technology
[0002] The rehabilitation of preschool children on the autism spectrum typically consists of one-on-one training at an institution and generalization tasks at home. A child often has deficits in multiple dimensions simultaneously, including obedience, joint attention, emotional regulation, and stereotyped behaviors. Rehabilitation therapists need to incorporate training units with different control models—applied behavior analysis, floor time, interpersonal relationship development intervention, structured teaching, and sensory integration—into the same weekly prescription; placing them adjacently can send conflicting signals to the child. Each lesson allows for very few trials for conditional comparison, and actual training does not allow for repeated trial runs on the child to determine the correct strategy. Most publicly available rehabilitation training systems rely on assessment scales or static rule bases to match the fixed content of a particular evidence-based system. Subsequent programs have begun to adjust the difficulty based on physiological signals or behavioral feedback during training and to perform deduplication and conflict detection on candidate strategies. However, they still treat multiple evidence-based systems as interchangeable items, failing to break them down into independently executable meta-operators according to control models, and failing to differentiate between observation criteria under institutional controlled conditions and those under real-world home conditions. When the success rate at the home level declines, the result is often recorded as a failure, making it impossible to determine whether the skill has not yet been established, has been lost after changing conditions, or was due to parents not following the procedure. The problems left by the above treatment are: under the constraint of not allowing exploratory scheduling, evidence-based training units with different control modes will cancel each other out when they are arranged into the same lesson; and the limited opportunities for conditional comparison in each lesson cannot be used on the variation dimension whose cause is not yet clear. Summary of the Invention
[0003] This invention provides an adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights, which solves the technical problems in related technologies such as the inability to separate the causes of generalization failure, the mismatch of limited detection opportunities, and the mutual cancellation of multiple evidence-based training units in the same lesson when institutions and families conduct collaborative training.
[0004] This invention provides an adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights, comprising: The data collection and standardization module receives audio and video data from the institution and the family, training terminal records, and parent selection records. After behavioral event recognition, trial segmentation, execution fidelity labeling, variation condition labeling, and state sequence calculation, it outputs a standardized trial record sequence and a child state sequence. The profile construction module calculates multidimensional progress indicators, writes generalization degree, estimates state tolerance band and updates cumulatively according to the upper limit of step size based on the standardized trial record sequence, child state sequence and generalization discrimination results, and outputs defect subtype profile, state tolerance band and profile displacement. The weighted orchestration module, based on defect subtype profiles, state tolerance bands, profile displacement, child state sequences, execution logs, meta-operator outcome effects, and next-period detection conditions, outputs scenario-specific meta-operator prescriptions through meta-operator attribute encoding, profile dimension weighted aggregation, relationship graph correction, coverage density shrinkage, and constrained time block orchestration. The execution control module receives scenario-specific meta-operator prescriptions, defect subtype profiles, state tolerance bands, and child state sequences, completes personality parameter initialization, fast layer state switching, and slow layer support quantity adjustment, and outputs execution logs. The reflux analysis module, based on the standardized trial record sequence, execution log, defect subtype profile, and scenario-specific meta-operator prescriptions, outputs generalization discrimination results, meta-operator outcome effects, next-period detection conditions, and parent execution deviation prompts after detection condition scheduling, generalization difference calculation by variation dimension, and meta-operator outcome effect statistics.
[0005] Furthermore, the trial segmentation uses the instruction-type events output by the behavior event recognition as anchor points to extract the response window; For the first training trial with the same training objective in the same lesson, the time period after the instruction-type event ends, which is the waiting time specified by the scenario-specific meta-operator prescription, is taken out as a no-prompt observation window. The results corresponding to the child's reaction-type events in the window are recorded in the first no-prompt result. When a prompt-type event is identified in the window, it is recorded as disturbed and is not included in the training no-support accuracy rate. The state sequence calculation outputs the arousal index and engagement index according to a fixed-length time slice.
[0006] Furthermore, the execution fidelity labeling checks the integrity of the delivery instructions, the waiting interval, and the consistency of the consequences in the family trial. Trials in which prompt-type events or corrective consequences occur are recorded as detection disturbances, and trials that pass all three checks and do not experience detection disturbances are recorded as fidelity valid. When the proportion of failed checks in the family trials this week exceeds the third threshold, the feedback analysis module generates a parent execution deviation prompt. The variation condition labels are marked according to three variation dimensions: the trainee, the training materials, and the context, indicating the similarities and differences between the trial and the standard training conditions at the institution.
[0007] Furthermore, the dimensional progress indicators are defined and normalized to the same interval according to the type of defect dimension. The skill-type dimension takes the low support accuracy rate, which is the ratio of the number of trials in the institutional training trials in which the prompt level is achieved as independent completion or gesture prompt to the total number of institutional training trials. The state-type dimension takes the negative normalized value of the intensity of negative state events. The generalization degree is normalized according to the generalization difference of the dimensional dimension and its complement value is written. The defect subtype profile is updated at the weekly boundary. When the absolute value of the difference between the newly calculated index and the current value of the profile is multiplied by the merging coefficient and exceeds the step size limit, the value is taken according to the step size limit and the sign is retained before being added to the current value of the profile. When the difference exceeds the first threshold and the direction is consistent for two consecutive weeks, the step size limit is released.
[0008] Furthermore, the meta-operator attribute encoding maps the meta-operator attribute vector into a meta-operator embedding and query vector through a two-layer feedforward network; The image dimension weighted aggregation concatenates the values of each defect dimension in the defect subtype image with its availability flag, and then maps them to key vectors and value vectors through two layers of feedforward network. The value vectors are weighted and summed according to the attention obtained by normalizing the dot product of the query vector and each key vector to obtain the image context vector. The image context vector is concatenated with the meta-operator embedding and scene identifier vector and mapped to the initial adaptation score. The relationship graph is corrected by making a round of corrections to the initial fit score along the relationship graph. The coefficients of mutually exclusive edges are negative, and the coefficients of successive affinities are positive. When the support accuracy of the preceding dependent edge in the target dimension of the preceding meta-operator has not reached the threshold, the subsequent meta-operator is suppressed with a negative coefficient to obtain the corrected fit score.
[0009] Furthermore, the coverage density shrinkage uses the number of samples in the training samples whose distance from the current image is less than the neighborhood radius as the coverage quantity. When the coverage quantity is lower than the coverage threshold, the modified adaptation score is pulled towards the shrinkage target score according to the gap ratio relative to the coverage threshold and converted into adaptation weight. The shrinkage target score is obtained by weighting the frequency of adoption of each meta-operator in the snapshots of historical child images that are closest to the current image and have similar training weeks, based on the meta-operator outcome effect at that time. When the number of retrieved snapshots is lower than the snapshot threshold, the value corresponding to the default priority in the supervision rule table is taken.
[0010] Furthermore, the constrained time block arrangement reserves a probe block at the end of the lesson according to the maximum number of probe blocks, and the remaining time blocks are training blocks; Candidate meta-operators are masked according to the scenario and those with intolerance markers are removed. The adjustment-type meta-operator with the largest adaptation weight is selected as the reserved meta-operator for this lesson. Among the candidates whose wake-up applicable band covers their expected wake-up state, whose prerequisites are satisfied, and whose meta-operators have no mutually exclusive edges with the selected meta-operators, the meta-operator with the largest adaptation weight is selected for each training block. The meta-operators selected in the previous week are used until the adaptation weight of the largest candidate in this week exceeds the sum of its and the switching margin. The switching margin increases as the image displacement decreases.
[0011] Furthermore, the personality parameters are initialized by setting initial values for each dimension according to the defect subtype profile and the state tolerance band. The smaller the width and the smaller the upper bound of the state tolerance band, the smaller the initial values of emotional expression intensity and instruction density, and the larger the initial values of follow-up ratio and waiting time. When the wake-up index exceeds the upper limit of the state tolerance band and continues to exceed the second duration threshold, the fast layer state switching stops the current meta-operator, switches to the reserved meta-operator of this lesson, reduces the instruction density, and extends the waiting time and the prompt progression interval. After the wake-up index falls back into the state tolerance band and continues to exceed the third duration threshold, the stopped meta-operator is restored.
[0012] Furthermore, the adjustment of the slow layer support is carried out by comparing the training unsupported accuracy rate of each meta-operator target dimension with the target interval after each lesson. The training unsupported accuracy rate is the proportion of independent completion under the condition of no support at the institution. If it is higher than the upper limit of the target interval, the starting point of the prompt level is adjusted to the independent completion side and the waiting time and prompt progression interval are shortened. If it is lower than the lower limit of the target interval, the adjustment is reversed. If the comparison results are inconsistent, the adjustment direction is determined by weighted voting based on the number of training trials of each dimension in this lesson. If the frequency of wake-up exceeding the limit increases relative to the previous period and exceeds the threshold within several lessons after the adjustment is made to independently complete one side, the adjustment is rolled back by one level, and the support amount before the rollback is recorded as the upper limit of the support removal tolerance. When the same element operator is terminated by fast layer state switching in several consecutive lessons and the reason for the termination is that the wake-up index exceeds the upper limit of the state tolerance band, an intolerance mark is generated.
[0013] Furthermore, the generalization difference calculation for the sub-variation dimension is based on the difference between the training unsupported accuracy under the baseline condition and the detection unsupported accuracy under the single-variation condition. The difference is weighted and summarized according to the effective number of trials for each training objective. The effective number of trials is the number of trials that are undisturbed and faithfully effective or the number of observations without prompts on the first trial. When the value under the baseline condition is lower than the fourth threshold, the generalization judgment result is taken as not yet learned. When the sub-variation dimension generalization difference is higher than the fifth threshold and the effective number of trials is not lower than the sixth threshold, the corresponding variation dimension is taken as not generalized. The detection condition scheduling takes the combination of undetermined defect dimension and variation dimension, the effective number of trials, the interval width estimated by the binomial ratio interval, the severity of the defect dimension to which the combination belongs, and the current adaptation weight of the corresponding generalization class operator, and divides it by the implementation cost coefficient that increases in order of changing training materials, changing training personnel, and changing the situation to obtain the detection priority, and takes the combination with the largest value as the detection condition for the next period. The outcome effect of the meta-operator is calculated by taking the slope of the linear regression of the achievement prompt level moving toward independent completion with each trial as the target level effect. The outcome effect of the meta-operator is the target level effect minus the mean of the target level effects of other meta-operators in the same lesson, normalized by the scaling coefficient and limited, and deducted by the penalty coefficient when there is an intolerance mark.
[0014] The beneficial effects of this invention are as follows: To address the complex causes of failure at the home level and the limited time available for probing, the data collection standardization module divides the trials into sessions based on the respective perspectives of the institutional and home levels, and separates the initial trial without prompts into a separate layer. The backflow analysis module calculates the generalization difference based on three variation dimensions: person, material, and context. It then ranks the conditions for the next probing session based on interval width, defect severity, corresponding generalization class operator adaptation weight, and implementation cost coefficient. The probing block only performs single-variable comparisons; failure to learn, lack of generalization, and execution deviation are respectively backflowed to the profile, prescription, and parent prompts, and the three reasons no longer share the same set of subsequent prescriptions. To address the issue of conflicting lesson arrangements within multiple evidence-based systems, and the prohibition of trial arrangements of learning strategies on children during actual training, the weighted arrangement module, after revising the relationship graph, pulls scores towards historically similar children or supervisory rules based on coverage density, and then arranges them block by block according to mutual exclusion, arousal applicability bands, and prescription stability. The execution control module initializes personality parameters based on defect subtype profiles and state tolerance bands, separating fast layer switching due to arousal exceeding limits from slow layer support adjusted according to training unsupported accuracy, and constrains the speed of support withdrawal with the support withdrawal tolerance upper limit. The profile construction module cumulatively updates defect subtype profiles according to the step size upper limit. Prescriptions are adjusted based on actual records at both ends, and each adjustment can be traced back to the trial, detection results, and judgment conclusion. Attached Figure Description
[0015] Figure 1 This is a structural diagram of an adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights according to the present invention. Figure 2 This is a block diagram of an adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights according to the present invention. Figure 3 This is a flowchart of an adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights according to the present invention. Detailed Implementation
[0016] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.
[0017] At least one embodiment of the present invention discloses an adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights, such as... Figure 1 The diagram shows the overall structure of an adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights according to the present invention. In the diagram, ① represents the institutional end, ② the home end, ③ the data acquisition and standardization module, ④ the profile building module, ⑤ the weight arrangement module, ⑥ the execution control module, ⑦ the feedback analysis module, ⑧ the virtual tutor and training terminal, and ⑨ the parent application. The data acquisition end consists of ① the institutional end and ② the home end, with audio, video, and recordings from both ends being fed into the system server. Within the system server, ③ the data acquisition and standardization module, ④ the profile building module, ⑤ the weight arrangement module, ⑥ the execution control module, and ⑦ the feedback analysis module are sequentially connected. ⑦ The feedback analysis module connects back to ④ the profile building module via closed-loop feedback. ⑥ The execution control module sends execution control information to ⑧ the virtual tutor and training terminal, and ⑦ the feedback analysis module sends parent execution deviation prompts to ⑨ the parent application. Specifically... Figures 2 to 3 As shown, it includes: The data collection and standardization module receives audio and video data from the institution and the family, training terminal records, and parent selection records. After behavioral event recognition, trial segmentation, execution fidelity labeling, variation condition labeling, and state sequence calculation, it outputs a standardized trial record sequence and a child state sequence. Specifically, the inputs at the institution's end include video captured by a fixed camera in the training room, audio captured by a microphone, and records entered by the therapists' training terminals for each trial. Each record in the terminal includes the training goal number, trial type, achievement prompt level, child's response result, and type of reinforcer used. The recording time is aligned with the audio and video timeline. The audio and video at the institution's end are processed on the server side. The input for the family end includes video and audio recorded by the parents' mobile phones, as well as the daily task number, materials used, context, and child's reaction results selected by the parents in the application; the behavioral event recognition on the family end is completed locally on the mobile phone, and only the event-level features and trial records are uploaded to the server, while the original audio and video are not uploaded; this processing allows the subsequent analysis to be directly based on events, maintaining the same granularity as the institutional end; Behavioral event recognition employs conventional perception methods in the field; this embodiment only illustrates its function. Posture estimation provides a sequence of joint points of the child's torso and limbs, used to calculate the amplitude of movement and detect repetitive movement fragments as stereotyped behavioral events. Facial orientation estimation provides the orientation of the child's face relative to the trainer and the material, used to generate gaze partner and gaze material events. Speech activity detection and keyword recognition distinguish between the trainer's speech and the child's speech, and identify instruction statements, hierarchical prompts, and reinforcement statements from the trainer's speech, and identify vocalization and imitation events from the child's speech. In the identified behavioral event stream, each event includes event type, start and end time, issuing subject, and recognition confidence level; event types are divided into five groups: instruction type, prompt type, consequence type, child reaction type, and child state type, among which the child state type includes stereotyped behavior, gaze direction, movement amplitude grading, and vocal intensity grading; the event stream is sorted by time and then enters trial segmentation; Trial segmentation is anchored by instruction-type events; the system takes a response window starting from the end of an instruction event, and the child response-type events appearing in the window constitute the response of that trial. The first consequence-type event appearing after the window constitutes the consequence of that trial, and the next instruction event starts a new trial; the response window length of the training trial is the waiting time specified by the meta-operator, plus the product of the prompt progression interval and the number of progressive levels of the prompt hierarchy; the exploration trial does not provide prompts, and its response window length is only the waiting time. Within each training session at the institution level, it is also necessary to identify the first-trial no-prompt observation. For the first training session of the same training objective within the same lesson, the system extracts a separate time period equal to the waiting time after the instruction event ends as a no-prompt observation window. The results corresponding to the child's reaction events within this window are recorded in the first-trial no-prompt result field of this session, taking one of correct, incorrect, or no response. If a prompt event is identified within the window, this observation is recorded as disturbed and is not included in the training no-support accuracy rate and baseline observation level. This field is not set for subsequent training sessions of this lesson for this objective. If the child has not completed the training by the end of the window, the teaching continues at the prompt level starting point specified in the prescription for this session, and the achievement of the prompt level is still recorded as usual. There is no initial trial without prompts for observation at the home end; parents often fail to wait enough or give premature prompts when implementing the training. Requiring them to remain unprompted in the first trial will only increase deviations in implementation. Therefore, the accuracy of training without support is only given by the institution's trial, and the home's performance without support is determined by the exploration test. When the system is first run for a child, before the prescription is generated, the waiting time and the interval between prompts are taken as the midpoint of the range of the execution parameters of the meta-operator as the basis for calculating the reaction window and the no-prompt observation window. After the first prescription is issued, the prescription value will be used instead. The segmentation results from the institution are aligned with the training terminal records one by one. When the timestamp of the terminal record falls within a certain segmented trial interval, the training objective, trial type, achievement prompt level, and response result of that trial are based on the terminal record, with audio and video recognition results as a supplement. If the number of terminal records and segmentation results is inconsistent, the trial boundary is redefined based on the terminal record to ensure the accuracy of the trial records from the institution. The home terminal lacks trial-by-trial recording, relying entirely on event streams for segmentation, which presents challenges due to inconsistent parental execution. Parents often repeat instructions continuously when children don't respond. In such cases, the system merges consecutive similar instruction events with intervals shorter than a first duration threshold into a single instruction event, using the end time of the last merged instruction as the anchor point, and recording the number of repetitions in the trial's attributes for subsequent fidelity checks. The home terminal's prompt level is mapped by prompt-type events within that trial: no prompt and correct response are recorded as independent completion; the highest prompt, such as gestures, demonstrations, or physical assistance, is recorded as the corresponding level; all others are recorded as incomplete. In one embodiment of the present invention, the fidelity check is performed on each trial at the home end and includes three items; the first item is instruction delivery integrity, which checks whether the identified instruction statement belongs to the instruction statement family of the meta-operator corresponding to the task of the day, and whether the instruction is given once; if the instruction statement is not in the statement family, or the number of repetitions before merging exceeds the first number threshold, it is recorded as instruction delivery incomplete. The second item is the waiting interval requirement. Check whether the interval between the end of the instruction event and the next prompt event or the re-instruction is not shorter than the waiting time specified by the prescription for the meta-operator. If the parent gives a prompt before the child has responded, the trial will not reflect the child's true ability. Such trials are recorded as insufficient waiting. The third item is consistency of consequences, which examines the correspondence between the child's response and the consequences. Consistency is indicated by the occurrence of a reinforcing statement or the delivery of a reinforcing object after the child's correct response. Consistency is indicated by the occurrence of a corrective prompt after the child's incorrect or no response. Consistency is indicated by the occurrence of a reinforcing event after the child's incorrect or no response, or by the absence of any consequences after the child's correct response. Trials with inconsistent consequences cannot reflect ability and may reinforce incorrect behavior. There is a separate check for the probing test, which checks whether there are any prompting events or corrective consequences in the test. The purpose of the probing test is to observe the actual performance without support. Once a prompt or correction occurs, the observation becomes meaningless. Such tests are recorded as probe disturbances and are not included in the generalization difference calculation. It is common for parents to give prompts when performing the probing task. This check is necessary for the family. Family trials where all three general checks are passed without detection interference are marked as authentic and valid; otherwise, they are marked as authentic and invalid, and the deviation type is recorded. The deviation type refers to the failed check item, and multiple deviation types can be recorded in one trial. The authenticity mark and deviation type are output along with the trial record for use by the feedback analysis module. Institutional trials are performed by rehabilitation therapists in conjunction with virtual instructors. The instructors do not provide prompts or corrections within the detection block, so institutional trials are marked as authentic and valid by default. The variation condition label marks the similarities and differences from the standard training conditions at the institution in three dimensions for each trial; the trainer is identified by voiceprint and facial recognition as a rehabilitation therapist, parent A, or parent B, etc., and the trainer is different from the trainer in the standard training conditions, which is considered a difference in person; the training materials and the situation are entered by the terminal at the institution and selected by the parents at the home. If they are consistent with the teaching aids specified in the prescription, it is considered the same material; if a substitute is used, it is considered the different material; if they are on the training table, it is considered the same situation; if they are in the play area or daily life situation, it is considered the different situation. The child's state sequence is calculated in time slices of fixed length. For example, the default length of the time slice in this embodiment is two seconds. The arousal index of each time slice is obtained by normalizing and weighting the sum of five features: the amplitude of movement, the intensity of vocalization, the frequency of stereotyped behaviors, the response delay, and the proportion of gaze deviation within that slice. The coefficients are obtained by using the arousal level marked by the rehabilitation therapist at the institution as supervision, and then using ordered categorical regression to fit the feature coefficients and fix them. The higher the arousal index value, the closer the child is to an agitated state. The engagement index is obtained by fitting three features in the same way: the ratio of time spent on the fixation partner to time spent on the fixation material within the time frame, the rate of response after instructions, and the number of times the child initiates interaction. The higher the value, the more trainable the child is in the engagement state. The arousal index and engagement index are calculated in real time by the server at the institution and output to the execution control module. At the home, they are calculated locally by the mobile phone and recorded and uploaded with each trial. The output of the data collection and standardization module is a standardized trial record sequence and a child state sequence. Each trial record includes the trial number, scene identifier, trial type, training target number, associated meta-operator number, trainer, instruction statement, achievement prompt level, first trial no prompt result, response result, consequence type, fidelity mark, deviation type, and three variation labels: trainer, material, and context. The child state sequence includes the start and end times of each time slice, arousal index, and engagement index.
[0018] The profile construction module calculates multidimensional progress indicators, writes generalization degree, estimates state tolerance band and updates cumulatively according to the upper limit of step size based on the standardized trial record sequence, child state sequence and generalization discrimination results, and outputs defect subtype profile, state tolerance band and profile displacement. Specifically, the defect dimension system is determined by the evaluation system adopted by the institution. In this embodiment, it includes dimensions such as instruction obedience, joint attention, turn-taking interaction, emotion regulation, stereotyped behavior, language expression, motor imitation, fine motor and sensory regulation, etc. The number of dimensions is not limited. Each training target is pre-assigned to a dimension, and each meta-operator is pre-associated with one or more target dimensions, so that the trial records can be collected by dimension. Severity reflects the current degree of deficiency in this dimension; the intensity of negative state events is defined separately for each dimension. For the skill-based dimension, it is the proportion of time slices in which the arousal index exceeds the upper limit of the tolerance band in the time slices corresponding to the training trials at the institution for this dimension. For the emotion regulation dimension, it is the proportion of time slices in which the arousal index exceeds the upper limit of the tolerance band in all time slices within the statistical window. For the stereotyped behavior dimension, it is the weighted value of the frequency of stereotyped events and the proportion of cumulative duration. For the sensory regulation dimension, it is the number of time slices required for the arousal index to fall back into the tolerance band after the regulation class operator is executed. The slower the fall, the higher this item is. The severity of skill-related dimensions is obtained by weighting and summing two parts: the failure rate of training trials for that dimension at the institution level in the past week and the intensity of negative state events for that dimension. The two parts are first normalized to the same interval according to the upper and lower quantiles of their historical distributions. The coefficients are set by the supervisor based on whether the performance of that dimension is dominated by failure or state events. Dimensions such as emotion regulation, stereotyped behavior, and sensory regulation, which do not have failure counts, are referred to as state-related dimensions. Their severity is determined only by the intensity of the corresponding negative state events and is also normalized according to the historical distribution. Volatility reflects the stability of this dimension's performance. It is calculated by taking the standard deviation of severity over a number of recent lessons and normalizing it according to the historical distribution. The severity of each lesson is calculated in the same way as described above, but the window is reduced from the past week to the number of trials and time slices of the lesson itself. High volatility means that the performance of this dimension is greatly affected by the state, and the weighting will suppress meta-operators with high instruction load. Low support accuracy is given by the proportion of trials in the institutional training sessions for this dimension where the prompt level is achieved independently or with gesture prompts. The denominator is the number of institutional training trials for this dimension, and trials that are not achieved are included in the denominator but not in the numerator. Training without support accuracy is only counted for institutional trials: trials with no prompt result on the first attempt are counted only once, and only if the result is correct are they included in the numerator; other trials are only counted in the numerator if the prompt level starts at independent completion and the prompt level is achieved independently. The denominator is the sum of the above two types of trials. Trials that are terminated by the fast layer are not included in either the numerator or the denominator. The three items are not complementary. Children may have a higher achievement rate with physical assistance but a lower rate of independent completion. The three items together characterize which stage of the learning process this dimension is in. Dimensional progress indicators are defined and normalized to the same interval according to dimension type, with the value direction being uniformly higher for better progress. Skill-type dimensions and low support accuracy have the same value, while state-type dimensions and severity have the opposite origin. The profile stores progress information in a separate column so that the network reads progress information in the same column with the same direction. Skill-type dimensions directly take low support accuracy. Emotion regulation dimension takes the proportion of time slices in which the arousal index exceeds the upper limit of the tolerance band, and takes a negative value and is normalized to one. Stereotyped behavior dimension takes the weighted value of the number of stereotyped events and the proportion of cumulative time within a unit of training time, and takes a negative value and is normalized to one. The weighting coefficient is consistent with the coefficient used in the severity calculation. The adjustment dimension is the number of time slices required for the wake-up index to fall back into the tolerance band after the adjustment class operator is executed. It is then negative and normalized. The faster the fall back, the better the progress of this dimension. After the four categories are normalized to the same interval, they are used in the same way for the portrait value, the normalization conversion of the generalization difference, and the calculation of the portrait displacement. Generalization is divided into three items: character generalization, material generalization, and situational generalization, which respectively reflect the degree to which the skill in this dimension is retained after changing the person, material, and situation. The value is obtained by converting the generalization difference of the returned dimension. The conversion method is to normalize the generalization difference according to the range of the progress index of this dimension and take its complement value. The larger the generalization difference, the lower the generalization. When the generalization difference is negative, that is, the performance under the detection condition is better than the baseline, it is counted as zero to avoid the generalization exceeding the upper bound. When the system runs for the first time, all three generalization values are missing. They are written after the detection layer accumulates enough effective trials. For dimensions such as emotion regulation and sensory regulation, which do not have training objectives, there is no way to obtain baseline observations and detection trials. Their three generalization values remain unavailable for a long time and are occupied by missing vectors. The tolerance band is given by the arousal index distribution of the time slices corresponding to the correct response tests in the most recent lessons. The lower and upper quantiles are taken as the lower and upper bounds of the tolerance band, and the lower quantile of the input index distribution of the same batch of time slices is taken as the lower bound of the input index. When the system is first run, there are no time slices available for children with this book. The tolerance band is temporarily taken as the median interval of the tolerance band distribution of historical samples of the same age group as the default value. After accumulating to the minimum number of time slices, the system will switch to its own distribution. The system simultaneously labels the child's state type according to the direction of change in the engagement index when arousal exceeds the limit. When arousal increases, engagement decreases accordingly, indicating an avoidant state. When arousal increases, engagement increases instead of decreasing, indicating a seeking state. Time slices where the arousal index is below the lower limit of the tolerance band and the engagement index is below the lower limit of the engagement index are counted separately. These time slices require increased emotional expression intensity rather than reduced requirements. This count, along with the state type, is output to the execution control module. In one embodiment of the present invention, the profile is updated once a week at the boundary; a newly calculated indicator within the week is subtracted from the present value of that indicator in the profile, and the difference is multiplied by a merging coefficient. If the absolute value of the difference exceeds the weekly step size limit of that indicator, the value is taken according to the step size limit and the sign is retained. The result is added to the present value as the profile value for the next week; for example, the default value of the weekly step size limit in this embodiment is one-tenth of the value range of that indicator. The step size limit ensures that weekly abnormalities can only change the profile to a limited extent, avoiding frequent prescription rewrites due to illness, changes in work and rest, etc. When the difference between a new indicator in a certain dimension and the profile value exceeds the first threshold and the direction is consistent for two consecutive weeks, the system determines that the dimension has experienced a continuous shift rather than an accidental fluctuation. The system temporarily relaxes the upper limit of the step size for that dimension, allowing the profile to follow the new level before restoring the upper limit. The generalization score is driven by the generalization discrimination results of the feedback; dimensions judged as insufficient evidence retain the previous period's generalization score and are marked with insufficient evidence. This mark is output with the profile. When the weighting module encodes, the generalization score item with this mark is treated as unusable to avoid driving prescription adjustments with outdated values; dimensions judged as generalized or not generalized by a certain variant dimension are written according to the aforementioned conversion; dimensions judged as not yet learned are not written with generalization score, because when the institution has not yet learned it, the performance under any condition does not reflect the generalization ability; Each indicator in the profile is accompanied by an availability flag; the three generalization parameters are marked as unavailable when the system runs for the first time, and gradually become available as detection and collection accumulate; indicators that do not exist in the training history data are also marked as unavailable; the state tolerance-related parameters take the default value and are marked as the default source when the system runs for the first time, and after accumulating to the minimum number of time slices, they are changed to their own distribution and marked as the source of this child, and are not included in the unavailable range; this flag is output along with the profile, so that the weight orchestration module can distinguish between the real value, the default value and the missing value during encoding; The image displacement is used to measure the overall change of the image this week relative to last week; all indicators have been normalized to the same range, and the displacement is the sum of the absolute values of the changes of each indicator divided by the number of available indicators; the image displacement is output to the weighted orchestration module to determine the switching margin value in the orchestration stage. When the displacement is low, the prescription remains stable, and when the displacement exceeds the second threshold, the prescription is allowed to change with the image. The output of the portrait construction module is a defect subtype portrait, a state tolerance band, and a portrait displacement. The portrait is organized in matrix form, with each row corresponding to a defect dimension and each column corresponding to one of the following: severity, volatility, low support accuracy, training without support accuracy, dimension progress index, person generalization, material generalization, and context generalization. Each value is accompanied by an availability flag. The state tolerance band includes the upper bound of arousal, the lower bound of arousal, the lower bound of the input index, the state type, and the count of low arousal time slices.
[0019] The weighted orchestration module, based on defect subtype profiles, state tolerance bands, profile displacement, child state sequences, execution logs, meta-operator outcome effects, and next-period detection conditions, outputs scenario-specific meta-operator prescriptions through meta-operator attribute encoding, profile dimension weighted aggregation, relationship graph correction, coverage density shrinkage, and constrained time block orchestration. Specifically, the meta-operator library is pre-built by the institutional supervisor based on the operation manuals of each evidence-based system; the decomposition principle is that each meta-operator can be executed independently within a time block, with a clear target dimension, a clear operation process and adjustable execution parameters; decomposition to this granularity allows training units of different systems to be placed adjacent to each other in the same lesson, and their interrelationships can be determined. Taking the system used in this embodiment as an example, applied behavior analysis is broken down into meta-operators such as step-by-step instruction training, differential reinforcement of alternative behaviors, behavior extinction, and task analysis chains; floor time is broken down into meta-operators such as emotional acceptance, interactive following, joint attention expansion, and emotional peak extension; interpersonal relationship development intervention is broken down into meta-operators such as turn-taking interaction, emotional reference, and social reference probing; structured teaching is broken down into meta-operators such as visual flowchart execution and work system completion; and sensory integration is broken down into meta-operators such as vestibular regulation activities, tactile regulation activities, and proprioceptive regulation activities. Each meta-operator is described by an attribute vector; the control mode attribute takes one of three options: adult-led, child-led, or co-regulated. Step-by-step instruction training is adult-led, interactive following is child-led, and alternating interaction is co-regulated; the instruction load attribute grading describes the density of demands made to the child per unit time; the arousal applicability band attribute gives the arousal index range that the meta-operator is suitable for execution. The upper bound of the applicability band for behavior extinction is lower than that for step-by-step instruction training, and the applicability band for emotion acceptance and sensory regulation meta-operators covers a higher arousal range. The attribute vector also includes a wake-up increment, which is the typical change in the wake-up index after the meta-operator executes one time block. It takes a negative value for regulating meta-operators and a positive value for high instruction load meta-operators. The value is obtained by statistically analyzing the difference between the average wake-up index before and after the execution of the meta-operator in the historical execution log. This item is used by the orchestration stage to estimate the expected wake-up status of subsequent time blocks, so that orchestration does not have to rely on actual measured data that has not yet occurred. When a new meta-operator is added to the library or the historical log is insufficient for the minimum number of statistical counts, this item takes the default value set by the supervisor according to the instruction load level and regulating class flag, and then uses the statistical value after accumulation. The attribute vector also includes family executability, prerequisite skills, execution parameter range, instruction statement family, and adjustment category markers; family executability indicates whether the meta-operator can be safely executed by parents at home, and behavior extinction is marked as not executable at home because it requires handling of extinction outbreaks; prerequisite skills list other dimensions of low support accuracy thresholds required to execute the meta-operator; execution parameter range gives the allowed intervals for six items: waiting time, cue progression interval, cue level starting point, emotional expression intensity, instruction density, and follow-through ratio; instruction statement family lists the instruction statements allowed to be used by the meta-operator; adjustment category markers indicate whether the meta-operator is used to bring the child back from a hyperarousal state to the tolerance zone; For generalization meta-operators, the attribute vector includes a generalization type attribute and a probe availability flag. The generalization type can be one of character generalization, material generalization, or contextual generalization. When a second trainer is introduced to perform the same training objective at the institutional end, it is a character generalization class; when the teaching aid is changed, it is a material generalization class; and when the training objective is embedded in the game area or daily process, it is a contextual generalization class. The probe availability flag indicates that the meta-operator can carry probe tests, that is, the objective is presented once without prompts to observe the actual performance. The relationship graph uses meta-operators as nodes and contains three types of edges: mutually exclusive edges connect meta-operators that cannot coexist in the same time block, typically for behavior extinction and interactive following for the same stereotyped behavior, where the former requires adults to ignore it while the latter requires adults to join, and their proximity will give children conflicting signals; sequential affinity edges connect meta-operators where the execution of the former is beneficial to the execution of the latter, typically for step-by-step instruction training after emotional acceptance; and pre-dependency edges connect meta-operators where the latter requires the former to reach the threshold of the target skill; the three types of edges are labeled by the supervisor, and sequential affinity edges can be supplemented by the outcome effect of backflow; The meta-operator attribute encoding employs a two-layer feedforward network. Categorical attributes in the attribute vector are encoded using one-hot or multi-hot methods, while numerical attributes are directly concatenated after interval normalization to form the input vector. The first layer is a linear transformation followed by a modified linear unit activation, and the second layer is a linear transformation that outputs the meta-operator embedding. This embedding is then subjected to a linear mapping to obtain the query vector. The encoding network shares parameters with all meta-operators, so when adding a new meta-operator to the library, there is no need to change the network structure; only its attribute vector needs to be provided for it to participate in scoring. In one embodiment of the present invention, the image side encoding also adopts a two-layer feedforward network; the input vector of each defect dimension is composed of the severity, volatility, low support accuracy, training without support accuracy, dimension progress index, three generalization terms and the center and width of the state tolerance band of that dimension, and each term is followed by its availability flag. When the flag is unavailable, the value of that term is replaced by a learnable missing vector component. The input vector is then concatenated with a learnable dimension identifier vector to distinguish different dimensions. After the first layer of linear transformation and the activation of the modified linear unit, and the second layer of linear transformation, the dimension representation is obtained. The dimension representation is then processed by two linear mappings to obtain the key vector and value vector of that dimension. The missing vector and the dimension identifier vector are trained together with the network, so the difference in feature availability between historical data and online data will not force the network to be redesigned. Weighted aggregation is performed as follows: The query vector of each meta-operator is multiplied by the key vector of each defect dimension, the result is divided by the square root of the key vector dimension, and then a normalized exponential transformation is performed on all dimensions to obtain the attention of the meta-operator to each dimension; then the value vectors of each dimension are weighted and summed with the attention to obtain the portrait context vector of the meta-operator. The role of attention is to allow each meta-operator to determine which dimensions of the profile to focus on and which values within each dimension, according to its own attributes. An adult-dominant meta-operator targeting stereotyped behaviors will focus on the severity of the stereotyped behaviors, as well as the volatility and tolerance band of the emotion regulation dimension, because the latter two determine the risk of the meta-operator triggering agitation. After the portrait context vector is concatenated with the embedded meta-operator and the scene identifier vector, it is transformed by the first layer of linear transformation and the activation of the modified linear unit, and the output of the second layer of linear transformation is a scalar, which serves as the initial adaptation score of the meta-operator in the scene. The scene identifier vector is a learnable vector for both the institutional end and the family end, so that the same portrait can obtain different adaptation scores in the two scenes. The adaptation score of the family end is biased towards meta-operators that can be executed by parents and do not have high requirements for execution accuracy. The initial adaptation score is corrected along the relation graph in one round. The correction is performed in the following steps: First, the initial adaptation score of each meta-operator is converted into an initial adaptation weight between zero and one using a logical function. Second, all meta-operators directly connected to the meta-operator to be corrected are extracted from the relation graph, and the contribution of each edge is calculated. The contribution is the initial adaptation weight of the opposite meta-operator multiplied by the coefficient corresponding to the edge type, and then multiplied by the gating value of the edge. Third, the sum of all contributions is added to the initial adaptation score of the meta-operator to obtain the corrected adaptation score. The coefficients corresponding to the edge types are negative for mutually exclusive edges and positive for sequential affinity edges. Their values are obtained through training and are constrained to one side of the corresponding sign during training, thereby inhibiting each other among mutually exclusive meta-operators and reinforcing each other among affinity meta-operators. The gating value is always one for mutually exclusive edges and sequential affinity edges. For preceding dependent edges, the value is zero when the low support accuracy of the target dimension of the preceding meta-operator reaches the threshold, so that the edge no longer takes effect. If the threshold is not reached, the value is one and the subsequent meta-operator is suppressed with a negative coefficient. The correction only performs one round because the edges of the relationship graph represent the direct relationships between meta-operators. Multiple rounds of propagation would mix in indirect relationships, deviating from the semantics of the supervised annotation. The correction first applies the coverage density shrinkage described later, and then converts it into the adaptation weight of the meta-operator through a logical function, which serves as the basis for the arrangement. In real training, it is not allowed to test the arrangement on children in order to learn the strategy. The values given by the scoring network in the areas not covered by the training samples are insufficient as the basis for arrangement. Therefore, after correcting the adaptation score, the coverage density shrinkage is applied. The system counts the number of samples in the training samples where the distance between the image and the current image is less than the neighborhood radius as the coverage number at the current image. The distance algorithm is the same as the image displacement, that is, the sum of the absolute values of the component differences of the indicators that are available in both images and then divided by the number of such indicators. When the coverage quantity is not lower than the coverage threshold, the adjustment score remains unchanged. When it is lower than the coverage threshold, the adjustment score is linearly pulled toward a shrinkage target score according to the gap ratio between the quantity and the coverage threshold. The larger the gap, the greater the degree to which it is pulled toward the target. The shrinkage target score is determined using a three-level concession system, with each level corresponding to the supervisor's actual judgment criteria under different data conditions. The first level is based on the child's own data; when the coverage is sufficient, no shrinkage occurs, and network scoring is used directly. The second level is based on the historical experience of children adjacent to the current image. This involves retrieving several snapshots from historical children's image snapshots that are closest to the current image and have a similar training week. The frequency of each meta-operator being actually adopted in the corresponding week of these snapshots is taken, and the weighted average is calculated based on the meta-operator's outcome effect at that time, which is then converted into a score as the shrinkage target. The third level is based on the supervisor's rule table. When the number of adjacent snapshots retrieved is lower than the snapshot threshold, the shrinkage target reverts to the score corresponding to the default priority of that meta-operator in the rule table. The second level was set up because the rule table was too coarse-grained, and when supervisors faced new children, they actually started with the previous arrangements of children with similar situations. The supervision rule table lists the default meta-operators and their priorities for each dimension in descending order of severity, which is the original basis for course scheduling in the institution; the three-level concession allows the system to revert to manual basis in areas with sparse data and to act according to network scoring in areas with sufficient data. The transition at each level is continuously given by the coverage quantity and the number of adjacent snapshots, without the need for manual switching. In one embodiment of the present invention, the orchestration is based on time blocks; each lesson on the institutional end is divided into several time blocks according to the typical execution time of the meta-operator, and each day on the home end is divided into one or two time blocks. All time blocks executed on the home end on a given day are included in subsequent statistics as a lesson. For example, the default value of the time block length on the institutional end in this embodiment is eight minutes. The time blocks of each lesson are divided into training blocks and probe blocks. The upper limit of the number of probe blocks is inversely linked to the recent wake-up out-of-bounds frequency. The wake-up out-of-bounds frequency refers to the number of events in the execution log that trigger fast layer switching because the wake-up index exceeds the upper limit of the tolerance band. When the frequency is too high, the upper limit is lowered. The actual number of probe blocks in this lesson is the number of backflow conditions that does not exceed the upper limit. The corresponding time blocks are reserved from the end of this lesson, and the rest are training blocks. The block-by-block arrangement process is as follows: First, the candidate set is masked according to the scenario. Meta-operators that cannot be executed by the home end are removed, and meta-operators with intolerance markers are removed from both ends. From the remaining candidates, the adjustment class meta-operator with the highest adaptation weight is selected as the reserve for this lesson. If there is no adjustment class candidate, the available meta-operator with high coverage range for wake-up is selected. This reserve does not occupy the training block. Then, the training block is arranged block by block. For each training block, candidates with mutually exclusive edges to the selected or reserved meta-operators in this block are first removed. Then, the expected wake-up state of this block is calculated. For the first time block, the average wake-up index of the opening time slice of the last few lessons is taken. If there are no historical lessons, the center of the state tolerance band is taken. For subsequent time blocks, the expected value of the previous block is taken plus the wake-up effect increment of the selected meta-operators in the previous block. Only candidates whose wake-up applicable band covers the expected state and whose prerequisites are satisfied are retained. Among the remaining candidates, the one with the highest adaptation weight is placed into this block. The orchestration imposes prescription stability constraints on training blocks, while probe blocks are not subject to these constraints and are directly arranged according to the probe conditions of the backflow. If the meta-operator selected in the previous week still satisfies all constraints in the current week, the candidate is replaced only if the fit weight of the highest candidate in the current week exceeds the sum of the fit weight of the meta-operator and the switching margin; otherwise, the arrangement of the previous week is used. The switching margin increases as the image displacement decreases, and takes its maximum value when the displacement is below the second threshold, so that the prescriptions of weeks with basically unchanged images remain unchanged. There are two exceptions to the stability constraint, otherwise the conclusion of the generalization judgment cannot be put into the prescription in time. First, when the candidate belongs to the generalization class and its generalization type matches the generalization degree of the newly written low value this week, the switching margin takes half of its value. Second, when the meta-operator selected last week has an intolerance mark this week, the meta-operator has been excluded in the masking stage, and the stability constraint is no longer applied to this time block. It is directly rearranged according to the highest candidate in the current period. The scheduling also imposes global constraints: the number of time blocks in which each evidence-based system is selected within a week is not less than the diversity lower limit, to avoid the adaptation weight being concentrated on a single system and causing the prescription to degenerate into a single system plan; The diversity lower bound includes a concession: if all meta-operators in an evidence-based system have intolerance labels in the current period, then there are no available candidates for that system. The diversity lower bound is adjusted accordingly based on the number of systems with available meta-operators in the current period, and this situation, along with the systems and meta-operators involved, is recorded in the prescription for supervisor review. Without this concession, the masked candidate set will not be able to simultaneously satisfy the diversity lower bound and mutual exclusion constraints, resulting in no solution for the arrangement. The detection blocks are arranged according to the backflow detection conditions within the reserved time blocks, and only meta-operators with detection available markers, target dimensions and generalization types that match the conditions are selected; single variation detection is arranged at the institutional end, that is, the target is presented under the condition that only one dimension of the three dimensions of person, material and situation is different and the other two dimensions are the same; comprehensive layer detection is arranged at the family end, that is, the person is different and one or two dimensions of material or situation are different, reflecting the child's performance under real life conditions; The reference observation layer with identical three dimensions no longer occupies a separate probe block, but is undertaken by the first trial observation without prompts at the mechanism end; the first trial observation without prompts occurs before any prompts on the target on that day, its three variation dimensions are identical, and no prompts or corrections are given in the observation window, which is consistent with the observation conditions of the probe test, and therefore can be directly used as the comparison benchmark for each single variation layer. The output of each time block includes the meta-operator number, block type, target dimension, detection condition, initial value of execution parameters, and arousal precondition; the arousal precondition of the training block is taken as the entire tolerance band by default, and the detection block is taken as the middle of the tolerance band; the initial value of the execution parameter is given after being truncated by the range of the execution parameters of the meta-operator, with the corresponding value of the child's current personality parameter vector as the center. That is, if the center value falls within the range, the center value is taken, and if it falls outside the range, the endpoint closest to the center value in the range is taken. The online training is divided into two stages: pre-training and online fine-tuning. Pre-training uses the institution's anonymized historical course records. The supervisor compares the historical course item names to the meta-operator numbers in the meta-operator library, so there is no need to annotate each item individually. The profiling side only uses the severity, volatility, and low support accuracy that can be calculated from the historical trial records. The training unsupport accuracy cannot be calculated because the historical courses did not have a first trial or prompts for observation. It is marked as unusable along with generalization and tolerance band related items and is occupied by missing vectors. The imitation items calculate the binary cross-entropy based on the adaptation weight of each meta-operator and whether the therapist selected the label of that meta-operator, and sum it on the meta-operator. Online fine-tuning is conducted weekly. The sample consists of the final prescriptions that were actually issued and reviewed by the supervisor during the week. The labels are taken from the adjusted selection results rather than the system's own output to avoid the mimicry items degenerating into repeated confirmations of the system's existing preferences. Fine-tuning also introduces an outcome term, using the return meta-operator outcome effect to weight and correct the adaptation weights of the meta-operators actually executed within this evaluation unit. The outcome term is given by the following formula:
[0020] in, For the ending option, This is the set of meta-operators actually executed within this evaluation unit. For example operators The outcome effect, Its value is taken when the effect is positive, otherwise it is taken as zero. The absolute value is taken when the effect is negative; otherwise, it is taken as zero. For example operators Corrected adaptation score, For logical functions, Meta-operator The adaptation weights before shrinking; At most one of the positive and negative parts is non-zero. Therefore, meta-operators with positive effects are pushed to higher fitness weights, and meta-operators with negative effects are pushed to lower fitness weights. Meta-operators with effects close to zero do not participate in this term. The weighting coefficient is the absolute value of the effect. The more certain the effect, the stronger its correction to the score. This allows different results of the same meta-operator on different children to participate in training according to their reliability, without the need to introduce exploratory scheduling. The total loss is composed of the imitation term, the ending term, the mutual exclusion term, and the smoothing term, weighted and summed according to their respective coefficients. The mutual exclusion term is the product of the adaptation weights of the meta-operators at both ends of each mutual exclusion edge in the relation graph and summed on the mutual exclusion edge, so that the mutual exclusion meta-operators do not simultaneously obtain excessively high adaptation weights, thereby reducing the exclusion pressure in the orchestration stage. The smoothing term takes effect when the image displacement is lower than the second threshold. It is the sum of the absolute values of the differences between the adaptation weights of each meta-operator this week and the adaptation weights of last week, as an aid to the switching margin in the orchestration stage. The adaptation weights in the above terms are all taken from the network output before the coverage density shrinkage, that is, the values obtained by converting the corrected adaptation score through a logical function, because shrinkage is a treatment for the sparse data in the orchestration stage and should not participate in parameter updates. The optimization employs an adaptive moment estimation optimizer; during the pre-training phase, the coefficients of the outcome terms are set to zero, and iteration continues until the imitation terms on the validation set no longer decrease; during the fine-tuning phase, the learning rate is set to a value lower than that used in the pre-training phase, and mutually exclusive and smoothing terms are always retained; throughout the entire training process, supervision comes entirely from the rehabilitation therapist arrangements that have already occurred and the execution outcomes that have been reverted, without needing to change any children's training arrangements in order to obtain labels. The output of the weight orchestration module is a scenario-specific meta-operator prescription. The prescription is organized by training day and time block on a weekly basis, and includes the adaptation weights and coverage of each meta-operator for the week for supervisors to view and manually adjust. The supervisor's adjustment results are both included in the issued prescription and used as a simulation sample for the next fine-tuning.
[0021] The execution control module receives scenario-specific meta-operator prescriptions, defect subtype profiles, state tolerance bands, and child state sequences, completes personality parameter initialization, fast layer state switching, and slow layer support quantity adjustment, and outputs execution logs. Specifically, the virtual tutor's personality parameter vector includes six dimensions; waiting time is the time to wait for the child to react independently after an instruction is given; prompt progression interval is the waiting time between adjacent prompt levels; prompt level starting point is the prompt level given first when the child does not react; emotional expression intensity is the amplitude level of the tutor's facial expressions, tone of voice fluctuations, and body movements; instruction density is the level of the number of times requests are made to the child per unit of time; and the following ratio is the proportion of time the tutor follows the child's current interests and the time the tutor actively initiates activities. The six dimensions are divided into two groups according to their functions. The waiting time, the interval between prompts, and the starting point of the prompt level constitute the support quantity, while the intensity of emotional expression and the following ratio constitute the state quantity. The instruction density is constrained by both groups of signals because it affects both the difficulty and the arousal. These six dimensions are both the personality traits presented to the child by the tutor and the parameters of each meta-operator when it is actually executed. Personality parameters are initialized according to the profile; for children with a narrow state tolerance band and a low upper limit, the initial emotional expression intensity is set to a moderate level, the instruction density is set to a low level, the follow-up ratio is set to a high level, and the waiting time and the interval between prompts are set to a long level; for children with a low overall support accuracy rate, the starting point of the prompt level is set to a high level to reduce failure experiences; for children with a seeking state type or a high count of low arousal time slices, the initial value of emotional expression intensity is not set to a moderate level, because they need more vivid emotional signals to maintain engagement when arousal is increased or insufficient; The virtual tutor executes meta-operators sequentially according to the time blocks of the prescription. Before each time block begins, the tutor checks the wake-up preconditions of the block using the wake-up index of the current time slice. If the conditions are met, the prescription is executed. If not, the reserved meta-operators for this lesson are executed first. The block begins only after the wake-up index enters the allowed range. These reserved meta-operators can be repeatedly called within the lesson without being exhausted by a single use. Training blocks that are still not met before the end of the lesson are postponed to the next lesson. When a probe block is postponed, its probe conditions are returned to the backflow analysis module for rescheduling. Within the training block, the instructor selects a sentence from the instruction statement family of the meta-operator and issues an instruction. During the waiting period, the instructor monitors the child's reaction events. If the child responds correctly, reinforcement is given. If there is no response, the instructor will gradually advance the prompt level according to the prompt interval until the child completes or achieves physical assistance. If the child responds incorrectly, corrective prompts are given and the trial ends. Achieving the prompt level is recorded as incomplete. The prompt level achieved and the result of each trial are recorded by the training terminal. The first training session for each training target within the same lesson shall be conducted in a separate, unguided manner. That is, regardless of the starting level of the cueing level in the personality parameters, the instructor shall first issue an instruction and then not provide any cueing during the waiting period. The result of this session shall be recorded as the unguided result of the first session for that target on that day. Subsequently, the session shall return to the cueing level starting point specified in the prescription and continue teaching. The execution method of the first session and the cueing level starting point are independent of each other and shall not overlap. This rule is made because when the starting level of the prompt is set to a high level, no unsupported observation is generated in that trial, while the adjustment of the slow level requires this observation; the help is postponed to be given after the waiting period, so the teaching process is not reduced. The execution method of the detection block is different from that of the training block; the instructor presents the target according to the detection conditions specified in the prescription, issues only one instruction, does not give any prompts during the waiting time, does not correct the child if he / she does not complete it, and records and marks the result as correct or incorrect as the detection test. Failure to probe can lead to emotional risks. Therefore, each probe attempt is followed by a training attempt after the target has been mastered, so that children can have a successful experience. If two consecutive probe attempts in the same probe block are not correct, the remaining probes in this block are canceled and transferred to the reserved meta-operator of this lesson. The unfinished probe conditions are postponed to the next period and re-participated in the scheduling by the backflow analysis module. In each time slice, the fast-layer control reads and collects the arousal index and engagement index from the normalized module in real time and compares them with the state tolerance band. When the arousal index exceeds the upper limit of the tolerance band and continues to exceed the second duration threshold, the tutor stops the current meta-operator, switches to the reserved meta-operator for this lesson, and at the same time lowers the instruction density by one level and raises the waiting time and prompt progression interval by one level. The emotional expression intensity of avoidant children is lowered by one level, while the emotional expression intensity of seeker children remains unchanged. The trial that is in progress at the time of termination is recorded as terminated and is not included in the success or failure of the meta-operator, nor is it included in the outcome effect regression. When the engagement index is below the lower bound of the engagement index and continues to exceed the second duration threshold, while the arousal index is still within the tolerance zone, it indicates that the child is not agitated but has left the training context. At this time, the instructor increases the follow ratio by one level and inserts a child-dominated meta-operator that has no mutually exclusive edge with the current meta-operator in the current time block. The interaction is re-established starting from the item or activity that the child is currently focusing on. The interaction is returned to the original meta-operator when the engagement index returns to above the lower bound. When the arousal index is below the lower limit of the tolerance zone and the engagement index is also below the lower limit of the engagement index, the child is in a passive and low-arousal state. At this time, lowering the requirements will only cause the interaction to stagnate further. Instead, the instructor should increase the intensity of the emotional expression by one level and shorten the waiting time to bring the child back to the tolerance zone with a clearer signal and a more compact rhythm. This situation is the opposite of the handling direction when arousal exceeds the upper limit. Therefore, it is recorded with different switching reasons in the switching event. After the wake-up index falls back into the tolerance band and continues to exceed the third duration threshold, the tutor resumes the suspended meta-operator. During the resumption, the execution parameters remain at the adjusted level without rebounding until the end of the current time block. Each switch records a switch event, including the switch time, switch reason, suspended meta-operator number, incoming meta-operator number, and execution parameters before and after the switch. In one embodiment of the present invention, slow-layer control is performed once after each lesson. The system takes the training unsupported accuracy rate of each target dimension involved in the current lesson over the past few lessons and compares it with the target interval of that dimension. If the training unsupported accuracy rate is higher than the upper bound of the interval, it indicates that the current assistance provided is more than the child's actual needs, and the support amount is adjusted one level in the weakening direction, that is, the starting point of the prompt level is lowered, the waiting time and the prompt progression interval are shortened; if it is lower than the lower bound of the interval, it is adjusted one level in the strengthening direction; if it falls within the interval, the support amount remains unchanged; the target interval is set by the supervisor according to the child's learning pace. When multiple target dimensions are involved in this lesson and the comparison results are inconsistent, the support quantity, as a global parameter, cannot be adjusted in two directions at the same time. In this case, the weighted voting is based on the number of training trials for each dimension in this lesson, and the dominant direction determines the adjustment. If the number of trials in the positive and negative directions is close, no adjustment is made. This rule also applies to the adjustment of instruction density described later. The adjustment of the state quantity is based on the frequency of arousal overstepping and the frequency of insufficient engagement in this lesson. The frequency of insufficient engagement refers to the number of events that trigger fast-layer handling in this lesson because the engagement index is lower than the lower bound of the engagement index. Neither of the frequencies is related to the accuracy rate because these two dimensions address relationships and states rather than difficulty. When the frequency of arousal overstepping is higher than the threshold, the intensity of emotional expression is adjusted one level towards a more moderate direction and the follow-up ratio is increased one level. When the frequency of insufficient engagement is higher than the threshold, the follow-up ratio is increased one level, and the intensity of emotional expression is determined by the state type: avoidant types are adjusted towards a more moderate direction, and seeking types are adjusted towards a more pronounced direction. When both frequencies are lower than the threshold, the state quantity remains unchanged. The instruction density is increased by one level only when the training unsupported accuracy is higher than the upper bound of the target interval and no wake-up out-of-bounds occurs in the current lesson; it is decreased by one level when the training unsupported accuracy is lower than the lower bound or a wake-up out-of-bounds occurs in the current lesson; otherwise, it remains unchanged; this dual condition ensures that the instruction density will not be increased when the state is unstable or the accuracy is insufficient. If, within a few lessons after the support level is adjusted in a decreasing direction, the frequency of arousal exceeding the threshold rises relative to before the adjustment, then the adjustment is reversed by one level, and the support level before the reversal is recorded as the child's current support withdrawal tolerance limit. Thereafter, the support level in the slow layer must not exceed this limit; this limit is only lifted by one level when the frequency of arousal exceeding the threshold is below the threshold for several consecutive lessons. Removing support too quickly may cause behavioral regression, and the reversal and limit recording are used to limit this situation within one level. As the unsupported training accuracy rate continues to exceed the upper limit of the target range, the amount of support is gradually reduced while the instruction density is gradually increased. The virtual tutor gradually shifts from primarily providing companionship and emotional coordination to primarily guiding and making requests, keeping pace with the child's ability development. This advancement is a natural result of the above adjustment rules and no additional advancement conditions are set. When the same meta-operator is terminated by the fast layer in several consecutive lessons, and the reason for termination is that the wake-up exceeds the upper bound, the system generates an intolerance mark for the meta-operator. The mark indicates that the meta-operator is incompatible with the child's state at the current stage. During the mark's duration, the meta-operator will not enter the orchestration candidate set and will not be scheduled regardless of the severity and fit weight of its target dimension. The removal of intolerance markers can be achieved if either of the following conditions is met: First, the upper bound of the child's tolerance band increases relative to when the marker was generated and exceeds the set threshold, indicating that the child's ability to withstand similar requirements has changed; Second, the severity of the target dimension of the meta-operator remains high even after a set number of cycles have passed since the marker was generated, indicating that the training direction is difficult to avoid and needs to be tried again. The retry after the release is forcibly applied by the weight orchestration module, without re-selection based on the adapted weight, and is not subject to stability constraints: the lower limit of the execution parameter range is taken by the load of the meta-operator instruction, the shortened time block is placed in the middle of a class, and the wake-up precondition is taken in the middle of the tolerance band; if the retry is still terminated by the fast layer, the mark is reactivated and the number of cycles required for release is doubled. The output of the execution control module is an execution log. The execution log is organized by lesson. Each lesson includes the actual execution operator number, block type, execution start and end time, actual personality parameters during execution, the achievement prompt level and result of each trial within the block, the result of no prompt on the first trial, the detection conditions and results of the exploration test, all switching events, the current value of the tolerance limit that supports removal, and the tolerance mark and its removal cycle count. The family terminal is executed by the parents and does not generate personality parameters or switching events. Its execution log includes the task completion status and the results of the exploration test.
[0022] The backflow analysis module, based on the standardized trial record sequence, execution log, defect subtype profile, and scenario-specific meta-operator prescriptions, outputs generalization discrimination results, meta-operator outcome effects, next-period detection conditions, and parent execution deviation prompts after detection condition scheduling, generalization difference calculation by variation dimension, and meta-operator outcome effect statistics. Specifically, the feedback analysis module runs weekly. First, family-based trials are filtered by fidelity, and only valid trials with high fidelity proceed to subsequent calculations. Valid but invalid trials are counted by deviation type. When the proportion of a certain deviation type in the family trials of the week exceeds the third threshold, a corresponding parent execution deviation prompt is generated. For example, if the proportion of insufficient waiting is high, the parent is prompted to wait after the instruction and the purpose of waiting is explained. The deviation prompt does not change the prescription but is pushed to the parent's application. The two ends play different roles in the generalization analysis; the institution end is responsible for condition separation, because only the institution can change one variation dimension while keeping the other two dimensions the same, so the baseline observation layer and the three single variation detection layers at the institution end constitute a comparable set of observations; the family end is responsible for confirming the real conditions and providing evidence of execution, because the training personnel at the family end cannot be the same as the standard training conditions, and the individual influence of materials and context cannot be separated from the influence of personnel at the family end. The values of the baseline observation layer, the single-variation detection layer, and the home-end comprehensive layer are all calculated first according to the training objective and then aggregated to the dimension. The value of a certain training objective under the baseline condition is the training unsupported accuracy calculated according to the objective's caliber, and the calculation rule is the same as that of the profiling side, with all three variation dimensions being the same. The value of the objective under a certain single-variation condition or home-end comprehensive condition is the detection unsupported accuracy calculated according to the objective and the condition's caliber. Both types of values are the achievement proportions under unsupported conditions, with consistent dimensions, and therefore can be directly subtracted. The target-level generalization difference is obtained by subtracting the value of the target under the baseline condition from its value under the corresponding single variation condition. The target-level generalization differences of each related target under the same dimension are weighted and summed according to their respective effective trial counts, which is the dimensional generalization difference of that dimension on that variation dimension. The effective trial count refers to the number of unperturbed and truthfully effective trials or the number of observations without prompting in the first trial in that layer. The comprehensive layer of the family end is obtained by first taking the comprehensive layer generalization difference of each target in the same way and then weighting and summing it to obtain the comprehensive generalization difference of that dimension, which reflects the overall gap between children under real-life conditions and controlled conditions, and is not attributed to any single variation dimension. Both the baseline observation layer and the comprehensive family layer are subject to the same effective number of trials. Both must be at least the sixth threshold to enter the discrimination process. If the threshold is not met, it is considered insufficient evidence, and the profile maintains the generalization of the previous period. The reason is the same as the lower limit of the effective number of trials for the single variant layer, that is, the proportion is unreliable when the number of observations is too small. The discrimination is carried out in three stages. When the value of the baseline observation layer is lower than the fourth threshold, it is judged as not yet learned. No generalization discrimination or generalization detection is performed in this cycle because the skill has not yet been established under controlled conditions. When the baseline observation layer exceeds the threshold and the overall generalization difference is not higher than the fifth threshold, it is judged as generalized. At this time, single variant detection is no longer consumed because the child can complete it under real conditions, indicating that all three variant dimensions have been passed. When the baseline observation layer exceeds the threshold and the overall generalization difference is higher than the fifth threshold, it is judged that there is a generalization obstacle and the cause is undetermined, and it enters single variant localization; when the generalization difference of a certain variant dimension is higher than the fifth threshold and the number of effective trials for that layer is not lower than the sixth threshold, it is judged that the variant dimension has not generalized; when the generalization difference is not higher than the fifth threshold and the number of effective trials is also up to standard, it is judged that the variant dimension has generalized and the corresponding high generalization value is written according to the aforementioned conversion, and that layer will no longer occupy detection opportunities in subsequent cycles; when the number of effective trials is lower than the sixth threshold, that layer is recorded as insufficient evidence, and the profile maintains the generalization degree of the previous period; the reason for setting a lower limit for the number of effective trials is that there are only a few opportunities for detection each week, and a few accidental failures are enough to make the proportion estimate deviate from the true value; Since single-mutation detection opportunities are limited, it is necessary to select which layer to deploy in the current period among the conditional layers where no conclusions have been reached. The system calculates the detection priority for each combination of the undetermined dimension and the mutated dimension, and the calculation is performed in the following steps: First, take the correct and incorrect counts from the existing valid trials of the conditional layer, and estimate the width of the value interval according to the binomial ratio; Second, take the severity of the defect dimension to which the combination belongs; Third, take the current adaptation weight of the generalization class operator whose generalization type corresponds to the mutated dimension and whose target dimension is consistent with the combination; Fourth, take the implementation cost coefficient of the mutated dimension; Finally, multiply the first three items and divide by the implementation cost coefficient to obtain the detection priority of the combination. Interval width prioritizes conditional layers with no conclusions yet, and the fewer the effective trials, the wider the interval. Conditional layers with stable conclusions are no longer given repeated opportunities. Severity biases priority towards dimensions with more severe defects. Adaptation weight biases priority towards conditions that can be adopted by prescription once they are determined to be non-generalized. If the corresponding generalization class operator will not be selected under the current profile, the conclusion of the condition cannot be converted into training arrangements for the time being. The implementation cost coefficient reflects that the actual cost of obtaining a single probe is not the same across the three variation dimensions, which directly determines which layer to explore first under the same amount of information; material variation only requires the use of a substitute when presenting the target, hardly requiring any additional arrangements, so the cost coefficient is set to the lowest level; person variation requires another therapist to be present at that time, and the cost is reflected in the scheduling, so it is set to the middle level; situation variation requires taking the child away from the training table to the play area or daily life situation, which takes up class time and also requires bearing the state fluctuations brought about by the situation change. For autistic children, the change itself is a burden, so it is set to the high level; After the cost coefficient is used as a deductible item, the system prioritizes the condition layer with lower cost. Only when all of these conditions have been concluded but the cause has not yet been located will the detection that requires scheduling or relocation be activated. The combination with the highest priority value is returned in turn, and the number does not exceed the sum of the detection blocks of each class in this week. The level of the cost coefficient is set by the supervisor according to the organization's scheduling and site conditions. The discrimination results are passed to the weight orchestration module through the profile and correspond to different prescription responses; if the severity of a dimension is not yet learned, it will remain high while the generalization value is not written. Therefore, the weight orchestration module will continue to increase the adaptation weight of the non-generalization class meta-operator of that dimension; if a certain variant dimension is not generalized, the generalization value of the corresponding type of that dimension will be written to a low value, the adaptation weight of the matching generalization class meta-operator will be increased, and the switching margin will be half value so that it can be included in the prescription within the same week. The institution will add a training block of this type, and the family will arrange the corresponding task; if it has been generalized, the generalization value will be written to a high value, the meta-operator weight of that dimension will fall back, and the training will turn to maintenance; execution deviation does not change the profile, only pushes a prompt; In one embodiment of the present invention, the outcome effect is used to measure the actual effect of a certain meta-operator on its target dimension. It is statistically analyzed according to the rolling bi-weekly evaluation units, without waiting for the meta-operator to stop scheduling. The attribution falls on the attribution relationship recorded in the execution log. Each trial and each time block clearly belongs to a certain meta-operator, and its effect is directly calculated from the records during its own execution. The effect of a skill-based meta-operator is determined by the decreasing trend of the cue level on the trained objective, with lessons as the statistical unit. For each training objective executed by the meta-operator in a lesson, the number of trials executed by the meta-operator without interruption are arranged chronologically. The cue level achieved in each trial is assigned a decreasing sequence number in order from assisted completion to independent completion. Trials that are not completed are considered to be one level higher than assisted completion. The linear regression slope of the sequence number as the number of trials changes is taken, and the movement towards independent completion is recorded as positive. If the number of trials is less than two, no slope is provided for that objective in that lesson. The slope of each objective is weighted by the number of trials, which is the objective level effect of the meta-operator in that lesson. Using the downward trend of the prompt level as the source of the effect is consistent with the judgment criteria for whether the required support decreases step by step; Meta-operators are categorized into skill-based and non-skill-based based on whether their trial counts record the cue level. This categorization is not on the same axis as the skill-based and state-based categorization on the dimension side. Meta-operators under the same dimension may belong to different types. The training trial counts of differential reinforcement alternative behaviors record the cue level and are calculated according to the skill type. Non-skill-type meta-operators do not have a cue level, and their effects are also statistically measured in lessons, taking the state changes during execution. Regulation-type meta-operators take the negative value of the change in the arousal index from the beginning to the end of the block within their time block, plus the proportion of time slices in which the arousal index remains within the tolerance band after the end of the block, and the two are combined according to the coefficient set by the supervisor. Meta-operators targeting stereotyped behaviors take the change in the interval between two consecutive occurrences of the target behavior during its execution, with a longer interval being counted as positive. If the behavior occurs less than twice in a lesson, the lesson does not provide this effect. When the same meta-operator is executed in multiple time blocks in the same lesson, the average value of each block is taken as the value of the lesson. The two types of values are normalized to the interval with the same downward trend slope as the cue level according to the upper and lower quantiles of the historical distribution. The effect needs to be subtracted from the common component; the overall state of children in the same class will affect the record of all meta-operators in that class. If it is not subtracted, each meta-operator in a class with a good state will get a higher value; therefore, the baseline is the mean of the target layer effect of the other meta-operators in the same class. The net effect of the meta-operator in that class is its target layer effect minus the baseline of the class. Using the remaining meta-operators in the same lesson as a baseline, training is not required to stop in any dimension, nor is a blank control needed. This is still possible under the constraint that the institution must cover all high-severity dimensions. If only one meta-operator is executed in a lesson, no baseline is provided for that lesson; its records are merged into the calculation of adjacent lessons. The outcome effect is given by the following formula:
[0023] in, For example operators The outcome effect, For example operators The target layer effect of each lesson within this assessment unit is calculated as a weighted average based on the number of time blocks occupied by the corresponding meta-operator within that lesson. The average of the baselines for the corresponding lessons, weighted equally. The scaling factor. The tolerance penalty coefficient, For example operators The intolerance marker takes values of 1 if marked and 0 if unmarked. This indicates that the quantity within the parentheses is limited to between positive and negative one; The target layer effect is applied by taking the weighted average of the downward trend of the prompt level for skill-type meta-operators and the normalized value of the state change for non-skill-type meta-operators. The two are already in the same range. Weighting by the number of time blocks within a class period makes the classes with more sufficient observation account for a larger proportion in the average. Taking the average has eliminated the influence of the total amount of arrangement, so there is no need to divide by the dose separately. The scaling factor is taken as the upper quantile of the absolute value of the net effect in the historical data, so that the outcome effect of most meta-operators falls between positive and negative one without being truncated, thus preserving the difference between values; the intolerance penalty causes meta-operators that are repeatedly terminated by the fast layer to receive a lower outcome effect even if the target layer effect is positive, because their records mostly come from the fragments before termination, and repeated termination itself indicates that the meta-operator is incompatible with the current state; if the effect of a labeled meta-operator is reduced to a negative value as a result, the outcome term will have its adaptation weight reduced in the next fine-tuning, which is consistent with the treatment of excluding it from the candidate set during the arrangement stage; The outcome effects of the institutional and family ends are calculated separately and fed back with scene labels; the family end only uses the record of the valid trials with high fidelity, and the target layer effect of its skill-type meta-operator is calculated separately by the achievement prompt level identified in the family end trials. This makes the weight orchestration module learn the actual effect under the parent execution conditions in the family end scene vector, rather than the replication of the effect in the institutional end. The feedback process is conducted weekly; the generalization discrimination results are output to the profile building module to write the generalization degree and insufficient evidence markers; the outcome effects of each meta-operator are output to the weight arrangement module as outcome item samples for the next fine-tuning; the condition with the highest detection priority and the number not exceeding the total upper limit of the detection blocks this week is output to the weight arrangement module as the detection condition for the next period to arrange detection blocks; parent execution deviation prompts are pushed to the parent application; the five links are thus connected end to end, and the prescription is continuously adjusted according to the child's actual performance at both ends. Each adjustment can be traced back to the specific trial records, detection results, switching events and discrimination conclusions.
[0024] In one embodiment of the present invention, the above system is applied to a rehabilitation institution for institutional and family collaborative training of a preschool boy. The child receives training three times a week at the institution, which is carried out by a rehabilitation therapist in conjunction with a virtual tutor on a tablet. At the family end, the mother and father take turns to perform push tasks and record at night. An example of standardized trial records collected in the third week of system operation is shown in Table 1, with each record corresponding to one trial. Table 1, Example of standardized trial record;
[0025] The profile provided by the profile building module at the third week boundary shows that the child's command obedience severity is moderate, low support accuracy is in the middle range, joint attention severity is high, low support accuracy is low, emotion regulation volatility is high, and stereotyped behavior severity is high; the state tolerance band is narrow and the upper bound is low, the state type is avoidant, and the three generalizations are marked as unusable. The weighted arrangement module arranges prescriptions accordingly; on the institutional side, emotional acceptance is placed at the beginning of each lesson, followed by step-by-step instruction training with low instruction density. Stereotyped behaviors are replaced by differential reinforcement combined with tactile modulation activities; behavior extinction is not selected because it has a mutually exclusive edge with interactive following and the upper bound of the applicable band is lower than the child's expected arousal state; on the family side, after the mask, joint attention games and visual flowcharts are arranged for execution. Behavior extinction cannot be executed at home and therefore does not appear, and a comprehensive layer detection block is arranged; Because the three generalization factors are unavailable and the child's own sample size is limited, the coverage is below the coverage threshold, and the contraction is based on a three-level retreat target. The system retrieves several snapshots from the historical child profiles that have similar severity distributions and tolerance band widths and similar training weeks. The adoption frequency of each meta-operator in the corresponding week is taken and weighted according to the outcome effect at that time as the contraction target for the first two weeks. Therefore, the prescriptions for the first two weeks reflect the previous arrangements for children with similar situations rather than the default order of the rule table. Only the sensory regulation dimension is reverted to the default priority of the rule table because the adjacent snapshots are insufficient to meet the snapshot threshold. During the execution phase, the virtual tutor started with a gentle emotional expression intensity and a long waiting time. Due to the low support accuracy, the starting point of the prompt level was set to the demonstration prompt level. However, the first trial of two-step instruction obedience in each lesson was still executed without prompts. This resulted in one unsupported observation per lesson, and the unsupported accuracy of instruction obedience training was continuously calculated. In a lesson in the second week, the arousal index in the step-by-step instruction training block exceeded the upper bound and persisted. The fast layer stopped the meta-operator and switched to the reserved tactile adjustment activity. The instruction density was reduced by one level. After the arousal returned, the original meta-operator was restored and the switching event was recorded. In subsequent lessons, the slow layer found that the unsupported accuracy rate for instruction obedience training was higher than the upper bound of the target interval. Therefore, it lowered the starting level of the cue level by one level and shortened the waiting time by one level. After the adjustment, the frequency of arousal out-of-bounds increased in the lesson. The slow layer then reversed the adjustment by one level and recorded the upper limit of support removal tolerance. The concurrent outcome effect statistics showed that the achievement cue level for step-by-step instruction training on this target shifted from physical assistance to gestural cue, and remained positive after subtracting the baseline of the same lesson, thus increasing its outcome effect. The achievement cue level for differential reinforcement of alternative behaviors was still mostly physical assistance, and remained negative after subtracting the baseline, thus decreasing its outcome effect. At the end of the third week, the reflux analysis module first performs a fidelity screening. If the proportion of insufficient waiting in the trials executed by the father exceeds the threshold, an execution deviation prompt is generated. Then, it calculates and aggregates the command compliance dimension according to the training objective. The value of the baseline observation layer has exceeded the fourth threshold, and the overall generalization difference has exceeded the fifth threshold, which is judged to indicate that there is a generalization obstacle and the cause is yet to be determined. At this time, each of the three single mutation layers has zero to two valid trials, and the interval width is similar. The material layer has the lowest cost coefficient because it only needs to use a substitute, so it obtains the highest detection priority and becomes the detection condition for the fourth week. In the fourth week, the mechanism detects two-step command compliance under different material conditions. Both detections are correct, and the material generalization difference is lower than the fifth threshold. This layer is judged as generalized and written with a high value, and will not occupy any more opportunities in the future. Among the remaining two layers, the corresponding generalization class operator adaptation weight of the character layer is higher than that of the context layer, and the cost of introducing a second trainer is lower than that of the transition. The character layer becomes the detection condition for the fifth week. If the cost is not considered, the system will directly arrange the context detection that requires a transition in the fourth week. In the fifth week, the institution arranged for the second trainer to perform single mutation detection for the target. Both detections were incorrect. The range of the character layer narrowed and the value was too low. The character's generalization poor exceeded the fifth threshold and was judged as not generalizing. The system wrote the low value of the character's generalization into the profile. The weight arrangement module increased the adaptation weight of the character's generalization class operator. The switching margin was set to half the value of the exception so that it would be included in the prescription in the same week. On the institutional side, a training block is added for a second trainer to perform the goal. On the family side, the father is arranged to perform the same task after the mother demonstrates, and a waiting interval prompt is sent to the father. If the division of labor between the two ends is not distinguished, the influence of materials and context will be mixed with the influence of people and cannot be separated. If the limited exploration opportunities are not scheduled according to priority, they may be spent on the condition layer where the conclusion has been stable or the conclusion cannot be included in the prescription.
[0026] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.
Claims
1. An adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights, characterized in that, include: The data collection and standardization module receives audio and video data from the institution and the family, training terminal records, and parent selection records. After behavioral event recognition, trial segmentation, execution fidelity labeling, variation condition labeling, and state sequence calculation, it outputs a standardized trial record sequence and a child state sequence. The profile construction module calculates multidimensional progress indicators, writes generalization degree, estimates state tolerance band and updates cumulatively according to the upper limit of step size based on the standardized trial record sequence, child state sequence and generalization discrimination results, and outputs defect subtype profile, state tolerance band and profile displacement. The weighted orchestration module, based on defect subtype profiles, state tolerance bands, profile displacement, child state sequences, execution logs, meta-operator outcome effects, and next-period detection conditions, outputs scenario-specific meta-operator prescriptions through meta-operator attribute encoding, profile dimension weighted aggregation, relationship graph correction, coverage density shrinkage, and constrained time block orchestration. The execution control module receives scenario-specific meta-operator prescriptions, defect subtype profiles, state tolerance bands, and child state sequences, completes personality parameter initialization, fast layer state switching, and slow layer support quantity adjustment, and outputs execution logs. The reflux analysis module, based on the standardized trial record sequence, execution log, defect subtype profile, and scenario-specific meta-operator prescriptions, outputs generalization discrimination results, meta-operator outcome effects, next-period detection conditions, and parent execution deviation prompts after detection condition scheduling, generalization difference calculation by variation dimension, and meta-operator outcome effect statistics.
2. The adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights according to claim 1, characterized in that, The trial segmentation uses the instruction-type events output by the behavior event recognition as the anchor point to extract the response window; For the first training trial with the same training objective in the same lesson, the time period after the instruction-type event ends, which is the waiting time specified by the scenario-specific meta-operator prescription, is taken out as a no-prompt observation window. The results corresponding to the child's reaction-type events in the window are recorded in the first no-prompt result. When a prompt-type event is identified in the window, it is recorded as disturbed and is not included in the training no-support accuracy rate. The state sequence calculation outputs the arousal index and engagement index according to a fixed-length time slice.
3. The adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights according to claim 1, characterized in that, The execution fidelity label checks the integrity of the delivery instructions, the waiting interval, and the consistency of the consequences in the family trial. Trials in which prompt events or corrective consequences occur are recorded as probe disturbances. Trials that pass all three checks and do not experience probe disturbances are recorded as fidelity valid. The feedback analysis module generates a parent execution deviation prompt when the proportion of failed checks in the family trials this week exceeds the third threshold. The variation condition labels are marked according to three variation dimensions: the trainee, the training materials, and the context, indicating the similarities and differences between the trial and the standard training conditions at the institution.
4. The adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights according to claim 1, characterized in that, The progress indicators for each dimension are defined and normalized to the same interval according to the type of defect dimension. The skill-type dimension takes the low support accuracy rate, which is the ratio of the number of trials in the institutional training trials in which the prompt level is achieved as independent completion or gesture prompt to the total number of institutional training trials. The state-type dimension takes the negative normalized value of the intensity of the negative state event. The generalization degree is normalized according to the generalization difference of each dimension and its complement value is written. The defect subtype profile is updated at the weekly boundary. When the absolute value of the difference between the newly calculated index and the current value of the profile is multiplied by the merging coefficient and exceeds the step size limit, the value is taken according to the step size limit and the sign is retained before being added to the current value of the profile. When the difference exceeds the first threshold and the direction is consistent for two consecutive weeks, the step size limit is released.
5. The adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights according to claim 1, characterized in that, The meta-operator attribute encoding maps the meta-operator attribute vector into a meta-operator embedding and query vector through a two-layer feedforward network. The image dimension weighted aggregation concatenates the values of each defect dimension in the defect subtype image with its availability flag, and then maps them to key vectors and value vectors through two layers of feedforward network. The value vectors are weighted and summed according to the attention obtained by normalizing the dot product of the query vector and each key vector to obtain the image context vector. The image context vector is concatenated with the meta-operator embedding and scene identifier vector and mapped to the initial adaptation score. The relationship graph is corrected by making a round of corrections to the initial fit score along the relationship graph. The coefficients of mutually exclusive edges are negative, and the coefficients of successive affinities are positive. When the support accuracy of the preceding dependent edge in the target dimension of the preceding meta-operator has not reached the threshold, the subsequent meta-operator is suppressed with a negative coefficient to obtain the corrected fit score.
6. The adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights according to claim 1, characterized in that, The coverage density shrinkage uses the number of samples in the training samples whose distance from the current image is less than the neighborhood radius as the coverage quantity. When the coverage quantity is lower than the coverage threshold, the modified adaptation score is pulled towards the shrinkage target score according to the gap ratio relative to the coverage threshold and converted into adaptation weight. The shrinkage target score is obtained by weighting the frequency of adoption of each meta-operator in the snapshots of historical child images that are closest to the current image and have similar training weeks, based on the meta-operator outcome effect at that time. When the number of retrieved snapshots is lower than the snapshot threshold, the value corresponding to the default priority in the supervision rule table is taken.
7. The adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights according to claim 1, characterized in that, The constrained time block arrangement starts from the end of the lesson, reserving a probe block according to the maximum number of probe blocks, and the remaining time blocks are training blocks; Candidate meta-operators are masked according to the scenario and those with intolerance markers are removed. The adjustment-type meta-operator with the largest adaptation weight is selected as the reserved meta-operator for this lesson. Among the candidates whose wake-up applicable band covers their expected wake-up state, whose prerequisites are satisfied, and whose meta-operators have no mutually exclusive edges with the selected meta-operators, the meta-operator with the largest adaptation weight is selected for each training block. The meta-operators selected in the previous week are used until the adaptation weight of the largest candidate in this week exceeds the sum of its and the switching margin. The switching margin increases as the image displacement decreases.
8. The adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights according to claim 1, characterized in that, The personality parameters are initialized by setting initial values for each dimension according to the defect subtype profile and the state tolerance band. The smaller the width and the smaller the upper bound of the state tolerance band, the smaller the initial values of emotional expression intensity and instruction density, and the larger the initial values of follow-up ratio and waiting time. When the wake-up index exceeds the upper limit of the state tolerance band and continues to exceed the second duration threshold, the fast layer state switching stops the current meta-operator, switches to the reserved meta-operator of this lesson, reduces the instruction density, and extends the waiting time and the prompt progression interval. When the wake-up index falls back into the state tolerance band and continues to exceed the third duration threshold, the stopped meta-operator is restored.
9. The adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights according to claim 1, characterized in that, The adjustment of the slow layer support is carried out by comparing the training unsupported accuracy rate of each meta-operator target dimension with the target interval after each lesson. The training unsupported accuracy rate is the proportion of independent completion under the condition of no support at the institution. If it is higher than the upper limit of the target interval, the starting point of the prompt level is adjusted to the independent completion side and the waiting time and prompt progression interval are shortened. If it is lower than the lower limit of the target interval, the adjustment is reversed. If the comparison results are inconsistent, the adjustment direction is determined by weighted voting based on the number of training trials of each dimension in this lesson. If the frequency of wake-up exceeding the limit increases relative to the previous period and exceeds the threshold within several lessons after the adjustment is made to independently complete one side, the adjustment is rolled back by one level, and the support amount before the rollback is recorded as the upper limit of the support removal tolerance. When the same element operator is terminated by fast layer state switching in several consecutive lessons and the reason for the termination is that the wake-up index exceeds the upper limit of the state tolerance band, an intolerance mark is generated.
10. The adaptive closed-loop training system with dynamic fusion of multiple evidence-based weights according to claim 1, characterized in that, The generalization difference of the sub-variation dimension is calculated by taking the difference between the training no-support accuracy under the baseline condition and the detection no-support accuracy under the single-variation condition. The difference is weighted and summarized by the number of effective trials for each training objective. The number of effective trials is the number of trials that are undisturbed and valid or the number of observations without prompts on the first trial. When the value under the baseline condition is lower than the fourth threshold, the generalization judgment result is taken as not yet learned. When the sub-variation dimension generalization difference is higher than the fifth threshold and the number of effective trials is not lower than the sixth threshold, the corresponding variation dimension is not generalized. The detection condition scheduling takes the combination of undetermined defect dimension and variation dimension, the effective number of trials, the interval width estimated by the binomial ratio interval, the severity of the defect dimension to which the combination belongs, and the current adaptation weight of the corresponding generalization class operator, and divides it by the implementation cost coefficient that increases in order of changing training materials, changing training personnel, and changing the situation to obtain the detection priority, and takes the combination with the largest value as the detection condition for the next period. The outcome effect of the meta-operator is calculated by taking the slope of the linear regression of the achievement prompt level moving toward independent completion with each trial as the target level effect. The outcome effect of the meta-operator is the target level effect minus the mean of the target level effects of other meta-operators in the same lesson, normalized by the scaling coefficient and limited, and deducted by the penalty coefficient when there is an intolerance mark.
