Psychological teaching strategy evolution method based on dredging effect feedback

Abstract

The invention relates to the field of psychological teaching, in particular to a psychological teaching strategy evolution method based on dredging effect feedback. The method comprises the steps that a pressure situation label sequence is generated by obtaining a project teaching event log and combining a pressure situation word list and an event mapping table, meanwhile, reliability evaluation is conducted on a student psychological portrait field set, and an acquisition session configuration package is constructed; performing pre-test snapshot construction and grooming strategy assembly to form an intervention execution package; acquiring and aligning the intervened observation windows, and generating a comparable effect vector packet; carrying out tendency score calculation and dual robust estimation based on the effect vector packet to form a strategy evaluation record; and finally, updating the strategy evolution library according to the strategy evaluation record and generating a strategy analysis report. According to the method, comparable evaluation and iterative evolution of the dredging strategy under different pressure situations and psychological portraits are realized, and improvement of the stability and consistency of strategy adjustment and management in psychological teaching is facilitated.

Description

Technical Field

[0001] This invention relates to the field of psychological teaching, and in particular to an evolutionary method for psychological teaching strategies based on feedback on guidance effects. Background Technology

[0002] In the field of psychological education, existing solutions for assembling and implementing guidance strategies based on project-based learning event logs and student psychological profile field sets typically revolve around event log parsing, collection of pre-test snapshot packages and post-test effect packages, selection of candidate guidance strategy sets, recording of intervention action trajectory packages, and updating of strategy evolution libraries. These solutions have limitations such as difficulty in maintaining consistency of triggering conditions between the stress situation vocabulary and event mapping table in project-based learning event log parsing, potential deviations in the correlation between stress situation label sequences and student and scene round identifiers, and insufficient comparability of strategy evaluation records under a unified time anchor point. Existing methods often rely on direct comparison between pre-test snapshot packages and post-test effect packages to form conclusions related to effectiveness coefficients. However, they have weak constraints on the linkage between observation window acquisition and alignment, time window alignment, missing value repair, and outlier removal. In scenarios where the unified time anchor point and strategy library version coexist, inconsistencies in the alignment granularity between the post-test effect package and the comparable effect vector are likely to occur, and the reliability score of the profile may not synergize with the quality label of the effect vector package. As a result, when the target strategy is reused across scenario rounds, it is difficult to unify the confidence of the strategy evaluation record, making it difficult to achieve stable generation of strategy evaluation records. For the joint processing of stress situation label sequences and psychological profiles, existing technologies generally have shortcomings in the link between tendency score calculation and dual robust estimation, the synchronization constraints between psychological profile clustering and stress situation label sequence clustering, and the consistency of recording standards for effectiveness coefficients and confidence scores. It is difficult to form a continuous process of observation window collection and alignment, time window alignment, quality repair and evaluation record generation among the collection of session configuration packages, intervention execution packages, effect vector packages and strategy evaluation records. As a result, when updating the strategy evolution library, the strategy sorting, freezing or removal rule triggering and gray-scale exploration queue writing lack traceable strategy evaluation record support, which in turn affects the process organization of guidance strategy assembly, intervention execution package generation and strategy analysis report generation in project teaching. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a method for the evolution of psychological teaching strategies based on feedback on guidance effects, comprising: S100: Obtain project teaching event logs and generate stress situation label sequences according to stress situation vocabulary and event mapping table; obtain student psychological profile field set and calculate profile reliability score, register student identifier, scene round identifier, unified time anchor point, strategy library version, and generate collection session configuration package; S200: Input the collection session configuration package into the pre-test snapshot construction and strategy assembly, generate the pre-test snapshot package and filter the candidate guidance strategy set, execute the target strategy according to the action parameter template and risk disabling conditions, and generate the intervention execution package; the intervention execution package includes the pre-test snapshot package and the intervention action trajectory package; S300. Based on the intervention execution package, the observation window is collected and aligned, the post-test effect package is collected and time window alignment, missing value repair, and outlier removal are performed to generate an effect vector package; the effect vector package includes the post-test effect package and a comparable effect vector. S400, based on the effect vector package, performs propensity score calculation and dual robust estimation, calculates the effectiveness coefficient and confidence level according to the stress situation label sequence and mental profile clustering, and generates strategy evaluation records; S500 updates the strategy evolution library based on strategy evaluation records, updates the strategy sorting by effectiveness coefficient and confidence level and executes freeze or removal rules, extracts efficient strategy action fragment patterns, generates new strategy hypotheses according to fragment recombination rules and writes them into the gray-scale exploration queue, and generates a strategy analysis report.

[0004] Furthermore, the stress scenario vocabulary in S100 includes schedule pressure, requirement change pressure, collaboration conflict pressure, review rejection pressure, and resource preemption pressure; the event mapping table includes event type field, trigger condition field, scenario label field, intensity stratification field, and duration window field. After parsing the event log, the event type is matched according to the trigger condition and the corresponding scenario label and intensity stratification are output to form a stress scenario label sequence.

[0005] Furthermore, in S100, the reliability score of the profile is synthesized from the missing rate score, the consistency score, and the fluctuation score. The missing rate score is calculated by the ratio of the number of missing tests of a field to the number of samples. The consistency score is calculated by the difference between short-window retests of the same indicator. The fluctuation score is calculated by the standard deviation of the sliding window. The synthesis rules are written into the acquisition session configuration package.

[0006] Furthermore, the candidate guidance strategy set in S200 is divided into an active queue and a gray-scale exploration queue. When selecting a target strategy, strategies that hit the disabled condition are first screened out, and then selected according to queue priority and random seed. The intervention action trajectory package records the completion mark field and the interruption reason field.

[0007] Furthermore, the observation window in S300 includes an immediate window, a short-term window, and a delayed window. The post-test effect package records the changes in psychological scale, task performance, and behavioral stability in the three windows, respectively, and records the situational disturbance marker field. Time window alignment performs resampling and interpolation under a unified time anchor point, missing value repair uses neighborhood imputation, and outlier removal uses quantile thresholds.

[0008] Furthermore, in S400, the propensity score is output by a logistic regression model, and the input features include stress situation label sequence encoding, psychological profile field set, profile reliability score, and strategy label; the dual robust estimation consists of a weighted outcome model and a weighted processing model, with the weights generated by the propensity score and the confidence level calculated by the resampled confidence interval.

[0009] Furthermore, the policy evaluation record in S500 includes a timestamp field and a decay factor field. The decay factor generates a weight based on the time difference between the timestamp field of the policy evaluation record and the current time. The policy ranking is generated by multiplying the effectiveness coefficient by the weight and combining it with the confidence threshold. The ranking result is written into the policy evolution library version record.

[0010] Furthermore, the action fragment mode in S500 includes a fragment identifier field, a precondition field, a postcondition field, a disabled condition field, and a parameter field; the fragment reorganization rules include precondition matching, postcondition compatibility, disabled condition merging, and parameter field conflict detection. After reorganization, a new strategy hypothesis is generated and a new strategy identifier is generated and written to the grayscale exploration queue.

[0011] Furthermore, the freeze or remove rules in S500 include a condition that the validity coefficient is below the threshold for M consecutive times and the confidence level is above the threshold.

[0012] Furthermore, the strategy analysis report includes a group vulnerability distribution table, a strategy effectiveness coefficient comparison matrix, a risk list for the gray-scale exploration queue, and a situational difficulty gradient parameter table; the group vulnerability distribution table is clustered and aggregated according to the stress situation label sequence and psychological profile, and the strategy effectiveness coefficient comparison matrix is ​​output according to the strategy identifier and stress situation label sequence index.

[0013] The key innovations of this invention include: (1) On the acquisition side, the project teaching event log is generated into a stress situation label sequence through the joint constraint of the stress situation vocabulary and the event mapping table. The portrait reliability score of the student psychological profile field set is registered in the same package with the student identifier, scene round identifier, unified time anchor point and strategy library version to form an acquisition session configuration package that can be directly called by the subsequent main steps, so that the stress situation label sequence and the student psychological profile field set are structurally aligned under the same session caliber.

[0014] (2) The collection session configuration package is used as a unified entry point to perform pre-test snapshot construction and strategy assembly, forming the pre-test snapshot package and filtering the candidate guidance strategy set accordingly. Then, under the constraints of the action parameter template and the risk disabling condition, the target strategy is executed to generate an intervention execution package that simultaneously contains the pre-test snapshot package and the intervention action trajectory package. Based on the intervention execution package, observation window collection and alignment are performed to collect the post-test effect package. Combined with time window alignment, missing value repair and outlier removal processing, an effect vector package containing the post-test effect package and the comparable effect vector is generated, forming a link-based data organization from intervention to comparable representation.

[0015] (3) On the evaluation and evolution side, the effect vector package is used as input to perform propensity score calculation and dual robust estimation, and the effectiveness coefficient and confidence are calculated according to the stress situation label sequence and psychological profile group to form a strategy evaluation record. Then, the strategy evolution library is updated based on the strategy evaluation record. The strategy sorting is updated by the effectiveness coefficient and the confidence, and the freeze or removal rules are triggered. At the same time, efficient strategy action fragment patterns are extracted from the intervention action trajectory package associated with the strategy evaluation record and new strategy hypotheses are generated and written into the gray exploration queue according to the fragment recombination rules. Finally, a strategy analysis report is generated to realize the closed loop of strategy iteration driven by the evaluation record.

[0016] The following are its main beneficial effects: (1) In response to the problems of inconsistent trigger conditions between the stress situation vocabulary and the event mapping table in the existing scheme, and the easy deviation of the association between the stress situation tag sequence and the student identifier and the scene round identifier, the stress situation tag sequence and the student psychological profile field set are registered in the same package under the same unified time anchor point and strategy library version by collecting the session configuration package. The profile reliability score is also included in the same session record. This enables the subsequent main steps to refer to the stress situation tag sequence, the student psychological profile field set and the profile reliability score based on a consistent standard when calling the collection session configuration package, thereby reducing the association ambiguity and record break when calling across rounds and versions.

[0017] (2) To address the problem in the existing scheme that the collection and comparison links of the pre-test snapshot package and the post-test effect package are scattered, and the collection and alignment of the observation window and the linkage of the time window are insufficient, resulting in inconsistent granularity between the post-test effect package and the comparable effect vector, the pre-test snapshot package and the intervention action trajectory package are bound and recorded through the intervention execution package, and the post-test effect package and the comparable effect vector are packaged together in the effect vector package. At the same time, time window alignment, missing value repair and outlier removal are carried out in the process of generating the effect vector package, so that the effect vector package can serve as a unified input carrier for subsequent propensity score calculation and dual robust estimation, and support verifiable alignment and comparison of the effect representation of the same target strategy under different observation windows.

[0018] (3) To address the problems of insufficient connection between the propensity score calculation and the dual robust estimation link, inconsistent constraints between psychological profile grouping and stress situation label sequence grouping, and difficulty in supporting the update of the strategy evolution library by the existing technology, a strategy evaluation record containing validity coefficient and confidence level is generated based on the effect vector package. Based on this, the strategy sorting, freezing or removal rules and gray exploration queue are written, so that each adjustment of the strategy evolution library update is traceably associated with the strategy evaluation record. At the same time, by extracting efficient strategy action fragment patterns and generating new strategy hypotheses according to fragment recombination rules, the new strategy has the same evaluation entry as the existing strategy when entering the gray exploration queue. In conjunction with the strategy analysis report, the grouping scope and strategy change process are archived and recorded, thereby alleviating the problem of discontinuous iteration caused by the lack of auditable support in the existing solution. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a psychological teaching strategy evolution method based on feedback of guidance effects, provided as an embodiment of this application. Detailed Implementation

[0020] Example 1: Refer to Figure 1 This is a flowchart illustrating a psychological teaching strategy evolution method based on feedback of guidance effects provided in an embodiment of the present invention. The process may include at least steps S100-S500: S100: Obtain project teaching event logs and generate stress situation label sequences according to stress situation vocabulary and event mapping table; obtain student psychological profile field set and calculate profile reliability score, register student identifier, scene round identifier, unified time anchor point, strategy library version, and generate collection session configuration package; S200: Input the collection session configuration package into the pre-test snapshot construction and strategy assembly, generate the pre-test snapshot package and filter the candidate guidance strategy set, execute the target strategy according to the action parameter template and risk disabling conditions, and generate the intervention execution package; the intervention execution package includes the pre-test snapshot package and the intervention action trajectory package; S300. Based on the intervention execution package, the observation window is collected and aligned, the post-test effect package is collected and time window alignment, missing value repair, and outlier removal are performed to generate an effect vector package; the effect vector package includes the post-test effect package and a comparable effect vector. S400, based on the effect vector package, performs propensity score calculation and dual robust estimation, calculates the effectiveness coefficient and confidence level according to the stress situation label sequence and mental profile clustering, and generates strategy evaluation records; S500 updates the strategy evolution library based on strategy evaluation records, updates the strategy sorting by effectiveness coefficient and confidence level and executes freeze or removal rules, extracts efficient strategy action fragment patterns, generates new strategy hypotheses according to fragment recombination rules and writes them into the gray-scale exploration queue, and generates a strategy analysis report.

[0021] S100: Obtain project teaching event logs and generate stress situation label sequences according to stress situation vocabulary and event mapping table; obtain student psychological profile field set and calculate profile reliability score, register student identifier, scene round identifier, unified time anchor point, strategy library version, and generate collection session configuration package; Specifically, this step is completed collaboratively by the project teaching event log access unit, the context tag generation unit, the profile collection and registration unit, the reliability calculation unit, and the configuration package assembly unit. The project teaching event log access unit is responsible for collecting, verifying, and standardizing the project teaching event logs for storage. The context tag generation unit is responsible for generating stress context tag sequences based on the stress context vocabulary and event mapping table. The profile collection and registration unit is responsible for obtaining the student psychological profile field set and solidifying the field definitions. The reliability calculation unit is responsible for calculating the profile reliability score. The configuration package assembly unit is responsible for registering student identifiers, scene round identifiers, unified time anchors, and strategy library versions, and generating a collection session configuration package. The collection session configuration package serves as the input carrier for the subsequent S200 pre-test snapshot construction and strategy assembly.

[0022] During the acquisition of project teaching event logs, the project teaching event log access unit connects to the event publishing interface of the project management teaching platform, collaboration tools, and course scenario engine. The project teaching event log is defined as a collection of event records occurring within the same teaching scenario cycle, including event type, event occurrence time, event object, event context, and source channel fields. The event type field covers event trigger sources corresponding to schedule pressure, requirement change pressure, collaboration conflict pressure, review rejection pressure, and resource preemption pressure. Specifically, event trigger sources corresponding to schedule pressure include task overdue, milestone extensions, and critical path changes; event trigger sources corresponding to requirement change pressure include new requirements, requirement rollbacks, and adjustments to acceptance criteria; event trigger sources corresponding to collaboration conflict pressure include conflict communication records, collaboration blockages, and disputes over responsibility attribution; event trigger sources corresponding to review rejection pressure include review returns, defect rejections, and failed re-reviews; and event trigger sources corresponding to resource preemption pressure include rejected resource requests, resources being occupied, and scheduling conflicts. When an event arrives, the access unit performs source channel verification and timestamp normalization. For out-of-order events, it rearranges the events according to the event occurrence time field. For duplicate events, it performs deduplication based on the event object field and the event context field. For missing events, it writes a missing flag to the record of the occurrence time field and adds it to the set to be reviewed. At the same time, the above verification and normalization results are recorded as audit log fields. The audit log fields and source channel fields are used together to trace the event access process of the same scenario round in the future.

[0023] During the loading process of the stress scenario vocabulary and event mapping table, the scenario tag generation unit loads versioned entries from the terminology dictionary. The stress scenario vocabulary is defined as a controlled set of stress scenario tags, including five categories of stress scenario tags: schedule pressure, requirement change pressure, collaboration conflict pressure, review rejection pressure, and resource preemption pressure, with each tag having a unique spelling. The event mapping table is defined as a set of mapping rules from event type fields to scenario tag fields, including event type fields, trigger condition fields, scenario tag fields, intensity stratification fields, and duration window fields. The trigger condition field is used to limit the threshold range or state combination that the event context field must satisfy. The intensity stratification field is used to divide the same scenario tag into different intensity levels under different trigger conditions. The duration window field is used to register the coverage range of the scenario tag on the timeline. After loading, the scenario tag generation unit performs consistency checks on the event mapping table, including event type field coverage checks and scenario tag field validity checks. When a missing mapping rule is found, a mapping missing flag is written, and a mapping table maintenance record is triggered. The mapping table maintenance record and the event mapping table version number are jointly written into the data collection session configuration package for subsequent cross-round comparison of mapping rule changes.

[0024] During the generation of the stress scenario label sequence, the scenario label generation unit takes the standardized project teaching event log as input, reads the event type field and event context field one by one, matches the scenario label field with the trigger condition field of the event mapping table, and outputs the corresponding intensity layer field and duration window field, generating a scenario label record stream with time index. The scenario label record stream is then processed by merging and pruning to form the stress scenario label sequence. The merging and pruning process includes merging the same scenario label within adjacent duration windows, marking the coexistence of different scenario labels within the same time segment, and segmenting the intensity layer field as it changes over time. For cases where the same event triggers multiple mapping rules, the scenario label generation unit performs conflict resolution according to the priority field of the trigger condition field and records the conflict resolution log field. For cases where the duration window field crosses the scene round boundary, the scenario label generation unit segments the duration window according to the scene round identifier and writes the boundary marker field at the segmentation point. The stress scenario label sequence is generated in this step and written to the acquisition session configuration package, allowing the S200 to read the current stress scenario label sequence as scenario-side input when building the pre-test snapshot package.

[0025] During the process of acquiring the student psychological profile field set, the profile collection and registration unit connects with the psychological assessment terminal, the classroom behavior collection terminal, and the course interaction record database to extract the student psychological profile field set corresponding to the scene round identifier. The student psychological profile field set is defined as a set of structured fields describing the student's psychological state and classroom behavior stability in the current scene round. The field definitions are fixed in the terminology dictionary and include the resilience baseline field, the anxiety and depression risk dimension field, the sleep fatigue self-report field, and the classroom behavior stability field. The resilience baseline field is obtained from the baseline assessment records before the start of the scene round, the anxiety and depression risk dimension field is obtained from the micro-measurement scale and event response sampling records, the sleep fatigue self-report field is obtained from the student self-report collection records, and the classroom behavior stability field is obtained from the statistical records of attendance, interaction, submission delay, and classroom interaction rhythm. When extracting fields, the image acquisition and registration unit performs unified field naming and solidification of unit dimensions. For fields with missing data, it writes missing data markers and retains the original source channel fields. For abnormal jump records, it writes jump markers and associates them with the corresponding project teaching event log fragment index. The project teaching event log fragment index is used in the acquisition session configuration package to support subsequent steps in tracing the source of mixed disturbances.

[0026] During the calculation of the profile reliability score, the reliability calculation unit takes the student psychological profile field set as input, calculates the missing rate score, consistency score, and volatility score for each field, and synthesizes the profile reliability score. The missing rate score is generated by statistically analyzing the number of missing test markers and the number of samplings in the student psychological profile field set, and then maps the missing rate record to the missing rate score. The consistency score is generated by reading short-window retest records of the same indicator, calculating the retest difference, and then mapping the consistency record to the consistency score. The volatility score is generated by calculating the standard deviation of field values ​​within a sliding window after alignment with a unified time anchor point, and then mapping the volatility record to the volatility score. The reliability calculation unit loads the synthesis rules when synthesizing the profile reliability score. The synthesis rules are fixed in the terminology dictionary and written into the acquisition session configuration package. During the synthesis process, a reliability downgrade flag is written for samples with a missing field ratio exceeding the threshold, a consistency anomaly flag is written for samples with a short window retest difference exceeding the threshold, and a fluctuation anomaly flag is written for samples with a sliding window standard deviation exceeding the threshold. The above flags and the profile reliability score are written into the acquisition session configuration package for S200 to call when screening candidate guidance strategy sets and executing risk disabling conditions.

[0027] During the registration of student identifiers, scenario round identifiers, unified time anchors, and strategy library versions, the configuration package configuration unit generates the primary key field of the current round of data collection session and completes version association. The student identifier is defined as a unique identifier consistent between the teaching platform and the psychological assessment terminal. The scenario round identifier is defined as the identifier of a single simulation round published by the same course context engine. The unified time anchor is defined as the baseline time record that aligns the project teaching event log with the student psychological profile field set. The unified time anchor is generated by the course context engine at the start of the scenario round and written into the round metadata corresponding to the scenario round identifier. The strategy library version is defined as the version record number of the strategy library to which the candidate guidance strategy set belongs. The strategy library version is generated by the strategy library publisher during version switching and written into the data collection session configuration package by the configuration package configuration unit. When configuring the package configuration unit to write the above fields, it also writes the stress scenario vocabulary version number and the event mapping table version number, and archives the mapping missing flag, conflict resolution log field, boundary flag field, missing test flag, jump flag, reliability degradation flag, consistency anomaly flag and fluctuation anomaly flag to the audit field field of the collection session configuration package, so as to maintain the traceability and reproducibility of the same scenario round in subsequent steps.

[0028] The following describes the operation process of this step using an engineering example. In a project sprint scenario, the project management course's scenario engine generates scenario round identifiers and publishes unified time anchors. The project teaching event log access unit accesses three types of event records from the collaboration tool: task overdue, new requirement, and review rejection, and performs random ordering and deduplication. The scenario tag generation unit loads the stress scenario vocabulary and event mapping table, mapping task overdue to schedule pressure, new requirement to requirement change pressure, and review rejection to review rejection pressure. It then matches the intensity stratification field and duration based on the duration and impact scope fields in the event context. The window field forms a stress situation label sequence with a time index; the profile acquisition and registration unit acquires the student psychological profile field set under the same scenario round and completes the field caliber solidification; the reliability calculation unit synthesizes the missing rate score, consistency score and fluctuation score to generate a profile reliability score; the configuration package assembly unit writes the student identifier, scenario round identifier, unified time anchor point, strategy library version, stress situation label sequence, student psychological profile field set and profile reliability score into the acquisition session configuration package, and outputs the acquisition session configuration package to the pre-test snapshot construction and strategy assembly steps of S200 as the input source.

[0029] Summary of the technical effects of this step: This step completes the session registration of project teaching event logs and student psychological profile field sets, forming a unified output standard for stress situation label sequences and profile reliability scores, and writes mapping rules and audit fields into the data collection session configuration package. The data collection session configuration package structurally solidifies the association between student identifiers, scene round identifiers, unified time anchors, and strategy library versions, enabling subsequent steps to read inputs from both the situation and profile sides under the same benchmark. The data collection session configuration package serves as an S200 input, establishing a traceable data baseline that connects across steps.

[0030] S200: Input the collection session configuration package into the pre-test snapshot construction and strategy assembly, generate the pre-test snapshot package and filter the candidate guidance strategy set, execute the target strategy according to the action parameter template and risk disabling conditions, and generate the intervention execution package; the intervention execution package includes the pre-test snapshot package and the intervention action trajectory package; This step is completed collaboratively by the pre-test snapshot construction unit, the candidate guidance strategy screening unit, the action parameter template assembly unit, the risk disabling condition determination unit, and the intervention execution orchestration unit, running on the teaching ecosystem data bus and strategy library service. Specifically, this step uses the collection session configuration package output by S100 as the sole input carrier. The collection session configuration package has registered student identifiers, scenario round identifiers, unified time anchors, and strategy library versions, and has written stress situation label sequences, student psychological profile field sets, profile reliability scores, and corresponding audit fields. Based on this, this step completes the generation of the pre-test snapshot package and the strategy assembly and execution, and outputs the intervention execution package for S300 as the input source for observation window collection and alignment.

[0031] In the pre-test snapshot construction chain, the pre-test snapshot construction unit reads the student identifier, scene round identifier, and unified time anchor from the collection session configuration package. It locates the current pre-intervention sampling section on the timeline according to the unified time anchor, and extracts the situation label fragments intersecting with this section from the stress situation label sequence to form the current stress situation label record. Simultaneously, it extracts the field values ​​corresponding to this section from the student psychological profile field set and retains the field caliber information. The pre-test snapshot construction unit writes the current stress situation label record, the field values, the profile reliability score, and the strategy library version into the pre-test snapshot package. The primary key fields of the pre-test snapshot package reuse the student identifier, scene round identifier, and unified time anchor, and append a snapshot generation time field and a snapshot source field. The snapshot source field points to the version association information of the collection session configuration package, thus forming a traceable link reference in subsequent steps. Furthermore, the pre-test snapshot construction unit merges the missing test flags, jump flags, reliability degradation flags, consistency anomaly flags, and fluctuation anomaly flags in the acquisition session configuration package, and writes the merging result into the snapshot audit field of the pre-test snapshot package, which is used by the subsequent risk disabling condition determination unit to perform boundary constraints on intervention execution.

[0032] In the candidate guidance strategy set screening process, the candidate guidance strategy screening unit accesses the strategy library service according to the strategy library version and loads the strategy index table. Each strategy entry in the strategy index table corresponds to a strategy identifier, an applicable domain label, an action parameter template index, and a risk disabling condition index. The candidate guidance strategy screening unit uses the pre-test snapshot package as screening input, matches the applicable domain labels based on the current stress situation label record to generate a situation matching set, performs profile clustering mapping based on the field values ​​of the trainee psychological profile field set to generate a cluster matching set, and then performs an intersection operation on the situation matching set and the cluster matching set to obtain the candidate guidance strategy set. Understandably, the profile segmentation mapping is implemented through a segmentation rule table, which consists of field names, segmentation thresholds, and segmentation identifiers. The segmentation rule table is associated with the strategy library version, loaded, and written into the execution log field of this step. When the segmentation rule table does not have a matching item under the strategy library version, the candidate guidance strategy filtering unit writes a segmentation missing flag and sets the candidate guidance strategy set corresponding to the student identifier to an empty set. At the same time, the reason for the empty set is recorded in the filtering exception log field to prevent the subsequent intervention execution orchestration unit from entering an undefined strategy call state.

[0033] In the action parameter template and risk disabling condition assembly link, the action parameter template assembly unit reads the action parameter template index one by one from the candidate guidance strategy set and loads the action parameter template from the template repository. The action parameter template is defined as an executable configuration structure of the target strategy, including an action sequence field, a duration field, a guidance dialogue task card index field, and a risk disabling condition reference field. The action sequence field is composed of action fragment identifiers arranged in sequence and bound to the execution channel, which covers the classroom interaction channel, the mobile prompt channel, and the homework task channel. The duration field is used to constrain the time span of the action fragment after a unified time anchor point. The guidance dialogue task card index field points to the content item number in the dialogue library and the task card library, and the content item number is consistent with the strategy library version. The risk disabling condition determination unit reads the profile reliability score, snapshot audit field, and current stress situation label record from the pre-test snapshot package, and reads the risk disabling condition index from the strategy index table. It then determines whether each disabling condition is matched according to the disabling condition trigger table, which is loaded in S200 as part of the exploration constraint rules. This table includes the disabling condition number, trigger field name, trigger threshold, trigger situation label, and disposal marker field. The disposal marker field is used to mark disabling strategy entries or to indicate manual intervention. Specifically, when the profile reliability score falls into the low confidence interval defined by the disabling condition trigger table, the risk disabling condition determination unit writes a reliability disabling marker and removes the corresponding strategy entry from the candidate guidance strategy set. When the snapshot audit field has a consistency anomaly marker or a fluctuation anomaly marker and the trigger field name corresponds to it, the risk disabling condition determination unit writes an audit disabling marker and records the trigger source field. When the current stress situation label record matches the trigger situation label in the disabling condition trigger table, the risk disabling condition determination unit writes a situation disabling marker and records the boundary marker field of the duration window field for subsequent interpretative tracing of the intervention action trajectory package. Furthermore, the risk disabling condition determination unit generates a disabling determination log field for the removal operation. The disabling determination log field includes a policy identifier, a disabling condition number, and a trigger field name, and is associated with the student identifier, scenario round identifier, and unified time anchor point and written into the audit field field of the intervention execution package.

[0034] In the target strategy selection and intervention execution orchestration chain, the intervention execution orchestration unit obtains an executable candidate set after determining the risk disabling conditions, and then executes the target strategy selection according to the exploration constraint rules. The exploration constraint rules include an upper limit on the number of gray-scale strategies, disabling condition triggering rules, and manual intervention marking rules. The upper limit on the number of gray-scale strategies limits the proportion of gray-scale exploration queue strategies entering the executable candidate set. The manual intervention marking rules trigger manual intervention and stop automatic execution when the execution channel is unavailable, the script library entry is missing, or the task card library entry is missing. The intervention execution orchestration unit decomposes the executable candidate set into an active queue and a gray-scale exploration queue. It first sorts the active queue according to the strategy priority field, and then samples the target strategy based on the sorted sequence using a random seed field. The random seed field is generated by the scenario round identifier and the student identifier and written to the operation log field, thereby maintaining a reproducible selection process within the same scenario round. When the active queue is empty and the grayscale exploration queue meets the upper limit constraint on the number of grayscale policies, the intervention execution orchestration unit selects a target policy from the grayscale exploration queue and writes a grayscale selection flag; when both the active queue and the grayscale exploration queue are empty, the intervention execution orchestration unit writes a flag indicating that no policy can be executed and generates an empty intervention action trajectory package. The interruption reason field of the empty intervention action trajectory package is recorded as an association reference between the candidate set being empty and the disable judgment log field.

[0035] During the execution of the target strategy, the intervention execution orchestration unit assembles the action parameter template into an execution plan. The execution plan includes action segment identifiers, execution channels, start time fields, end time fields, and content item numbers, and then distributes the execution plan to the control interface of the corresponding execution channel. The classroom interaction channel pushes prompts and interactive tasks through the teaching terminal; the mobile prompt channel pushes guiding scripts through the mobile terminal; and the homework task channel distributes task cards through the task publishing interface. After execution, each channel returns an execution status field and a receipt time field. The intervention execution orchestration unit performs consistency checks on the returned status and writes a receipt missing marker for action segments that have not been received within the time limit. Furthermore, the intervention execution orchestration unit records the causes of interruption events that occur during execution. Interruption events include channel unavailability, student rejection, missing content items, and secondary hits of disabled conditions. A secondary hit of disabled conditions refers to a situation where the profile reliability score update during execution triggers a disabled condition. In this case, the intervention execution orchestration unit writes an execution stop marker and records the trigger field name in the interruption reason field.

[0036] This step generates the intervention execution package as the output after completing the pre-test snapshot construction and strategy assembly, and executing the target strategy. The intervention execution package includes a pre-test snapshot package and an intervention action trajectory package. The intervention action trajectory package records the strategy identifier, action sequence field, duration field, guiding dialogue task card index field, execution status field, completion marker field, and interruption reason field, and associates with audit information such as a disable judgment log field and a grayscale selection marker. The pre-test snapshot package maintains the same primary key field as the acquisition session configuration package in the intervention execution package and carries a snapshot generation time field. Understandably, when outputting, the intervention execution package is written to the session cache by the intervention execution orchestration unit and synchronously stored in the database. The stored records include the student identifier, scene round identifier, unified time anchor point, and strategy library version. The intervention execution package is subsequently called in S300 as the input carrier for observation window acquisition and alignment. The pre-test snapshot package is used to define the alignment benchmark of the post-test effect package, and the intervention action trajectory package is used to define the start and end of the observation window and the acquisition range of the execution channel.

[0037] In summary, this step achieves the following technical effects: Based on the aforementioned data collection session configuration package, it solidifies the pre-intervention snapshot package and matches, disables, and assembles the candidate guidance strategy set from the strategy library version, forming an executable target strategy and execution plan. The intervention action trajectory package records action segments, execution channels, execution status, and interruption reasons in a session-based manner, ensuring consistent alignment and audit references for subsequent steps. The intervention execution package serves as an S300 input, establishing a cross-step connection between the intervention behavior and post-test data collection.

[0038] S300. Based on the intervention execution package, the observation window is collected and aligned, the post-test effect package is collected and time window alignment, missing value repair, and outlier removal are performed to generate an effect vector package; the effect vector package includes the post-test effect package and a comparable effect vector. This step involves the collaborative operation of an observation window acquisition unit, a time window alignment unit, a missing value repair unit, an outlier removal unit, and an effect vector construction unit. These units are deployed on the application server of the psychological teaching strategy evolution system and connected to the execution channel of the S200 via an event bus. They persistently write the post-test effect package and effect vector package through the session storage area, and simultaneously interact with the scale acquisition terminal, project-based teaching platform, and classroom interaction system through an interface gateway. Specifically, this step receives the intervention execution package output by the S200 as input. The intervention execution package includes a pre-test snapshot package and an intervention action trajectory package. The pre-test snapshot package includes student identifier, scene round identifier, unified time anchor, strategy library version, and stress situation label sequence snapshot fields. The intervention action trajectory package includes a strategy identifier, completion marker field, and interruption reason field. The observation window acquisition unit is triggered upon receiving the intervention execution package on the event bus. It creates an acquisition task record bound to the student identifier and scene round identifier in the session storage area, writes the unified time anchor to the time reference field, and writes the strategy library version to the version association field. The acquisition task record serves as the index entry for subsequent post-test effect packages and effect vector packages.

[0039] During the observation window acquisition and alignment scheduling phase, the observation window acquisition unit reads the action trajectory sequence from the intervention action trajectory package, parses the action start time, action end time, action channel identifier, and receipt status fields, and generates a window boundary record after aligning it with the unified time anchor point. The window boundary record includes a window type field, a window start field, a window end field, and an interruption flag field. The window type field is limited to immediate windows, short-term windows, and delayed windows. The window start field for immediate windows is taken as the action end time, and the window end field is calculated from the duration window field defined by the acquisition session configuration package. Short-term windows and delayed windows use the window end field of the previous window as the window start field and recursively derive their respective window end fields. The unified time anchor point, the three types of window type fields, and the duration window field constitute the minimum set of operating parameters for this step. If any one of these is missing, the acquisition task record writes the manual intervention flag field and stops subsequent acquisition. Understandably, when the completion flag field indicates an execution interruption, the observation window acquisition unit still generates a window boundary record and writes the interruption status into the interruption flag field. At the same time, it writes the interruption reason field, receipt status field, and interruption occurrence time field into the interruption audit field field for the missing test result repair unit to reference in the missing test result field.

[0040] During the post-test results collection phase, the observation window collection unit, based on window boundary records, collected micro-test questionnaire scores from the scale collection terminal, task submission records, review records, rework records, and project teaching event logs from the project teaching platform, and attendance records, interaction response records, and homework delay records from the classroom interaction system. These were then aggregated according to student identifiers and scenario round identifiers to generate a post-test results package. Within the three window partitions, the post-test results package registered fields for psychological scale changes, task performance changes, behavioral stability changes, and situational disturbance markers, as well as fields for data collection source, collection time, and data quality markers. The psychological scale change field is obtained by subtracting the micro-test questionnaire score from the baseline value of the same caliber field in the pre-test snapshot package. The scale audit field records the item number, scoring caliber, completion time field, and terminal identifier field. The task performance change field is obtained by aligning the performance event sequence generated from task submission records, review records, and rework records with the task baseline records. Duplicate event elimination uses a deduplication key generation rule, which is generated by combining the event type field, task identifier field, and collection time field. The first event is retained and the remaining events are written into the duplicate audit field. The behavior stability change field is obtained by aligning the behavior sequence generated from class attendance records, interaction response records, and homework delay records with the behavior baseline records in the pre-test snapshot package. The calculation of the situation disturbance label field depends on the stress situation vocabulary and event mapping table. The stress situation vocabulary includes schedule pressure, requirement change pressure, collaboration conflict pressure, review rejection pressure, and resource grabbing pressure. The event mapping table includes event type field, trigger condition field, situation label field, intensity stratification field, and duration window field. The observation window acquisition unit parses the project teaching event log within each window time range, extracts the event type field and trigger condition field, and matches the trigger condition field with the event mapping table to obtain the context label field and intensity stratification field. Then, it generates a window context label fragment based on the duration window field. The observation window acquisition unit compares the window context label fragment with the stress context label sequence snapshot field. When a new context label field or an intensity stratification field change occurs, the context disturbance marker field is set to a disturbed state, and the event type field, trigger condition field, context label field, intensity stratification field, and duration window field are written to the disturbance audit field field. When no such situation occurs, the context disturbance marker field is set to a non-disturbed state, and the comparison time field and comparison caliber field are written to the disturbance audit field field.

[0041] During the time window alignment, missing value repair, and outlier removal stages, the time window alignment unit generates a window alignment time axis based on a unified time anchor point. It performs resampling and interpolation on the scale collection sequence, performance event sequence, and behavior sequence to generate aligned window sequences. The resampling caliber field records the time granularity definition; continuous fields use linear interpolation and record the interpolation interval field; event fields use nearest neighbor endpoints and record the event endpoint field. The missing value repair unit performs neighborhood interpolation on the aligned window sequences and writes the interpolation source range field and interpolation count field into the repair record field. When the window sample size field is below the sampling lower limit or the interruption flag field is in an interrupted state, the missing window flag field is set to a missing state and the missing reason field is written. The outlier removal unit performs quantile threshold determination on the aligned window sequences that have completed missing value repair. Outliers are written into the removal index field, and the threshold caliber field records the threshold source. When the context disturbance flag field is in a disturbed state, the outlier removal unit enables the disturbance retention flag field within the duration window field range marked in the disturbance audit field. The outlier removal action only applies to data points outside the duration window field range.

[0042] During the effect vector package generation phase, the effect vector construction unit maps the cleaned post-test effect package into comparable effect vectors. The field order of the comparable effect vectors is limited by the vector caliber table and bound to the strategy library version, writing the version association field. The vector header field writes the window type field, data quality marker field, missing test window marker field, removal index field, and interruption marker field. The vector body field writes the psychological scale change field, task performance change field, behavior stability change field, and situational disturbance marker field according to the field position. Subsequently, the effect vector package is generated, which contains the post-test effect package and comparable effect vectors. The student identifier, scenario round identifier, unified time anchor point, and strategy identifier are written into the effect vector package index field. The effect vector package serves as the input carrier for S400, which is called and read for propensity score calculation and dual robust estimation. At the same time, the output timestamp field and the writing channel identifier field are written into the session storage area. Understandably, in another embodiment, the observation window acquisition unit accesses the physiological sensing sequence and writes it into the extended feature field field. The extended feature field field is bound to the extended position definition of the vector caliber table. The effect vector construction unit completes the writing of the extended field of the comparable effect vector according to the extended position definition.

[0043] In one engineering implementation, the project-based learning platform is deployed in the iterative practice environment of a project management course. The classroom interaction system records interaction responses and assignment delays, while the scale collection terminal pushes micro-test questionnaires and returns scoring results. All three are bound to a unified account system for member identification. In a sprint review round corresponding to a scenario round identifier, the event bus delivers an intervention execution package to the observation window collection unit. The observation window collection unit generates window boundary records for immediate, short-term, and delayed windows based on the intervention action trajectory package and initiates the collection task. The scale collection terminal returns micro-test questionnaire scoring results in the immediate window, the project-based learning platform returns performance event sequences and project-based learning event logs in the short-term window, the classroom interaction system returns behavioral sequences in the delayed window, and the observation window collection unit generates a post-test effect package and updates the situational disturbance marker field by generating window situational label fragments according to the event mapping table. The time window alignment unit generates an alignment window sequence, the missing value repair unit performs neighborhood interpolation and writes it into the repair record field, the outlier removal unit performs quantile threshold determination and writes it into the removal index field, and the effect vector construction unit generates comparable effect vectors and encapsulates them into an effect vector package, which is then delivered to S400 as input.

[0044] Summary of the technical effects of this step: Under a unified time anchor constraint, this step completes the collection, alignment, and cleaning of post-test results for three types of observation windows, forming a post-test result package. The scenario perturbation label field is generated and written into the perturbation audit field field, driven by the stress scenario vocabulary and event mapping table, supporting the processing of perturbation samples in the subsequent estimation stage. The effect vector package uniformly encapsulates the post-test result package and comparable effect vectors and delivers them to S400, forming a direct input for the policy evaluation record.

[0045] S400, based on the effect vector package, performs propensity score calculation and dual robust estimation, calculates the effectiveness coefficient and confidence level according to the stress situation label sequence and mental profile clustering, and generates strategy evaluation records; This step is jointly executed by the propensity score calculation unit, feature encoding unit, dual robust estimation unit, cluster calculation unit, and evaluation record generation unit. This step is performed on the evaluation service node of the psychological teaching strategy evolution system. The evaluation service node maintains communication connections with the session storage area, model parameter repository, and audit log repository, and is triggered by the event bus upon receiving the effect vector package output from S300. Specifically, this step takes the effect vector package as input. The effect vector package includes a post-test effect package and a comparable effect vector. The post-test effect package includes at least the following fields: psychological scale change, task performance change, behavioral stability change, situational disturbance marker, student identifier, scene round identifier, unified time anchor, and strategy identifier. The comparable effect vector includes the vector caliber table binding field, window type field, data quality marker field, missing test window marker field, and removal index field. This step simultaneously reads the registered stress situation label sequence, psychological profile field set, profile reliability score, strategy library version, and unified time anchor from the collection session configuration package generated by S100, and reads the strategy identifier and completion mark fields from the intervention execution package index field of S200 as the basis for constructing an index linking propensity score calculation input features and assessment records.

[0046] In the feature encoding and alignment stage, the feature encoding unit converts the stress scenario label sequence into a stress scenario label sequence encoding. This stress scenario label sequence encoding is defined as a structured input feature set for a logistic regression model, containing the order features of the scenario label occurrence sequence, intensity hierarchical distribution features, and duration window aggregation features. The scenario label set comes from a stress scenario vocabulary, which includes pressure related to project schedule, requirement changes, collaboration conflicts, review rejections, and resource contention. Specifically, the feature encoding unit extracts the scenario label field and intensity hierarchical field from the stress scenario label sequence, maps the window boundaries to the time axis index using a unified time anchor point, performs window aggregation based on the duration window field, generates a window-level label statistical vector, and binds the window-level label statistical vector to the scenario round identifier, writing it into the encoding index field. Furthermore, the feature encoding unit performs field alignment and missing test labeling on the psychological profile field set. The psychological profile field set is defined as a multi-dimensional set of fields describing the trainee's psychological state and sensitivity to guidance. The field types include scale baseline fields, behavioral baseline fields, and learning situation preference fields, which, together with the profile reliability score, constitute the covariate reliability constraint term. When the profile reliability score is lower than the reliability threshold registered in the collection session configuration package, the feature encoding unit writes the low-reliability label field into the feature metadata field and writes the sample into the low-reliability sample queue for the clustering calculation unit to process using a conservative clustering caliber. Understandably, the minimum set of core parameters required for this step includes stress situation label sequence encoding, psychological profile field set, profile reliability score, strategy identifier, and comparable effect vector. If any one of these is missing, the evaluation record generation unit writes an evaluation failure label field and records the missing field name field, while excluding the sample from the training and estimation input of subsequent dual robust estimation.

[0047] In the propensity score calculation stage, the propensity score calculation unit calls the logistic regression model to output the propensity score. The logistic regression model is a binary classification probability output model. The model parameters are loaded from the model parameter warehouse according to the model version field. The model input features include stress situation label sequence encoding, psychological profile field set, profile reliability score, and strategy identifier. The model output is the propensity score field, which represents the probability that a sample will be assigned to the guidance strategy corresponding to the strategy identifier under given covariates. Specifically, the propensity score calculation unit first selects the model parameter set according to the strategy library version and model version field matching, and reads the version-related feature dictionary and field standardization rules; then, it expands the stress situation label sequence encoding and psychological profile field set to fixed field positions according to the feature dictionary, and completes numerical range pruning, missing test placeholder encoding, and low confidence label fusion according to the field standardization rules; then, it writes the strategy identifier as a processing indicator field into the processing label field; after completing the above alignment, the logistic regression model calculates the propensity score field and writes it into the sample-level propensity score record. Furthermore, to handle samples where the context disturbance label field is in a disturbed state, the propensity score calculation unit writes the context disturbance label field into the disturbance covariate field and records the disturbance audit field index in the feature metadata field. Subsequently, the dual robust estimation unit reads this index to complete the weight constraints of the disturbed samples when constructing the weighted model. The sample-level propensity score record includes at least the student identifier, scenario round identifier, unified time anchor, policy identifier, propensity score field, model version field, policy library version, and data quality label field, and serves as the direct input for the dual robust estimation unit to generate weights.

[0048] In the dual robust estimation phase, the dual robust estimation unit constructs a weighted outcome model and a weighted processing model and outputs validity coefficients. The dual robust estimation is defined as an estimation process that simultaneously utilizes propensity score and outcome modeling information, including weighted correction for processing allocation bias and modeling the expected outcome. Specifically, the dual robust estimation unit first generates a sample weight field from the propensity score field. The sample weight field is jointly determined by the propensity score field, the policy identifier, and the queue caliber. The queue caliber is registered by the collection session configuration package and bound to the policy library version. When the propensity score field approaches the boundary probability, the sample weight field triggers a weight pruning rule and is written into the pruning flag field. The pruning threshold is written into the weight configuration field and used as an audit basis. Subsequently, the dual robust estimation unit trains a weighted outcome model under the influence of the sample weight field. This model is defined as a regression model predicting the expected comparable effect vector from stress situation label sequence encoding, psychological profile field set, profile reliability score, and strategy identifier. Its output is the outcome prediction vector field. Simultaneously, a weighted processing model is trained under the same weighting caliber. This model is defined as a probabilistic model fitting the processing label domain under a weighted sample distribution, and its output is the weighted processing prediction field. The dual robust estimation unit maps the psychological scale change field, task performance change field, and behavioral stability change field from the post-test effect package to the comparable effect vector according to the vector caliber table, forming the observation result vector field. The observation result vector field and the outcome prediction vector field are then subjected to residual calculation and aggregated under the constraint of the sample weight field to obtain the strategy-level effectiveness coefficient field. The effectiveness coefficient field serves as a characterization of the comprehensive effect of the strategy identifier under the current covariate distribution and situational perturbation labeling caliber. The aggregation process is recorded in the aggregation caliber field, and the participation of the window type field is also recorded. Understandably, in a preferred embodiment, the dual robust estimation unit generates window validity sub-coefficient fields for different window type fields, and performs window weighted synthesis when the policy-level validity coefficient fields are formed. The window weights are written into the window weight field and loaded by the acquisition session configuration package. In another embodiment, the dual robust estimation unit adopts a conservative weight upper limit rule for the low-confidence sample queue and records the low-confidence weight label field as the stratification basis for subsequent confidence calculation.

[0049] In the clustering calculation and confidence generation stage, the clustering calculation unit performs clustering according to the stress situation label sequence and mental profile, and outputs the clustering effectiveness coefficient and confidence score. Clustering is defined as the process of applying a stable grouping index to the sample set. The clustering index is generated driven by a clustering rule table, which includes situation label grouping rule fields, mental profile grouping rule fields, and reliability threshold fields, and is bound to the policy library version in the version association field. Specifically, the clustering calculation unit reads the stress situation label sequence encoding and generates a situation clustering identifier field according to the situation label grouping rule fields; it reads the mental profile field set and profile reliability score and generates a profile clustering identifier field according to the mental profile grouping rule fields; and forms a clustering unit index at the level of combining the clustering identifier fields. For each clustering unit index, the dual robust estimation unit reuses the sample weight field and the result prediction vector field to generate a clustering effectiveness coefficient subfield, and writes the sample size field, the proportion of the missing test window marker field, and the proportion of the weight pruning marker field into the clustering quality field field. The confidence level calculation is implemented through a resampling confidence interval process. The confidence level is defined as a quantitative characterization of the stability of the validity coefficient estimation. The resampling confidence interval process is triggered by the evaluation record generation unit according to a preset resampling count field. The resampling count field is read from the acquisition session configuration package and written to the audit log repository. Specifically, the evaluation record generation unit performs stratified sampling to generate resampling batches within the same cluster unit index according to the student identifier. It repeatedly calls the dual robust estimation unit to output a set of resampling validity coefficients and generates confidence interval and confidence level fields based on this set. When a resampling batch triggers the lower limit constraint of sample size or the proportion of the missing test window marker field exceeds the threshold, the evaluation record generation unit writes the confidence level downgrade marker field and records the trigger condition field. The relevant samples are still retained in the policy evaluation record for audit reference in S500's freeze or removal rules. The above trigger conditions, threshold caliber, and downgrade marker fields are all written into the audit field field and form a traceable association with the model version field, policy library version, and unified time anchor.

[0050] During the strategy evaluation record generation and output phase, the evaluation record generation unit encapsulates the strategy-level and cluster-level evaluation products to generate strategy evaluation records. These records include at least the following fields: strategy identifier, stress scenario label sequence encoding index, psychological profile cluster identifier, scenario cluster identifier, effectiveness coefficient, cluster effectiveness coefficient subfield, propensity score aggregation field, sample weight configuration field, confidence level field, confidence interval field, model version field, strategy library version, unified time anchor, scenario round identifier, evaluation timestamp field, data quality summary field, and audit field. The strategy evaluation records are written to the evaluation record table, generating a primary key field. This primary key field, along with the strategy identifier and strategy library version, forms a joint index, which is used as input for the S500 strategy evolution library update. The effectiveness coefficient and confidence level fields serve as core input fields for S500 to update strategy sorting and execute freeze or removal rules, while the audit field serves as a traceability input field for S500 to generate strategy analysis reports. Understandably, this step also writes the sample-level propensity score records and resampling batch records into the audit log repository, and writes an audit reference field into the evaluation record table. The audit reference field contains the propensity score record index and the resampling batch index, which are used for subsequent consistency checks and recalculation calls in version rollback scenarios.

[0051] In one engineering implementation, the evaluation service node is deployed on the course project server of the psychological teaching platform. The project teaching event log comes from the event stream related to review rejection pressure and demand change pressure of the project teaching platform. The psychological profile field set comes from the summary of behavioral sequences from the course registration questionnaire and classroom interaction system. The effect vector package comes from the post-test effect package and comparable effect vector generated by S300 in the immediate window, short window and delayed window. The event bus delivers the effect vector package after the review corresponding to each scenario round is completed. The propensity score calculation unit loads the logistic regression model parameters that match the strategy library version and outputs the sample-level propensity score record. The dual robust estimation unit generates the strategy-level effectiveness coefficient field under the constraint of the weight configuration field and simultaneously generates the window effectiveness sub-coefficient field. The cluster calculation unit generates the cluster effectiveness coefficient sub-field according to the stress situation label sequence encoding and the psychological profile cluster identifier field. The evaluation record generation unit completes the resampling batch construction and outputs the confidence field and confidence interval field. Finally, the strategy evaluation record is written into the evaluation record table and delivered as input to the S500 strategy evolution library update process.

[0052] This step's technical effects can be summarized as follows: Under the input condition of the effect vector package, this step completes the linked calculation chain of propensity score calculation and dual robust estimation, forming a strategy evaluation record. The strategy evaluation record synchronously registers the cluster effectiveness coefficient and confidence level under the policy identifier, policy library version, and unified time anchor caliber, and writes them into the audit field. This strategy evaluation record serves as a direct input to the S500 system, supporting subsequent policy ranking updates, freezing or removal rules, and strategy analysis report generation processes.

[0053] S500: Based on the strategy evaluation record, update the strategy evolution library, update the strategy sorting according to the effectiveness coefficient and confidence level and execute the freeze or removal rules, extract efficient strategy action fragment patterns, generate new strategy hypotheses according to fragment recombination rules and write them into the gray-scale exploration queue, and generate a strategy analysis report. This step is jointly executed by the strategy evolution library management unit, sorting and updating unit, freeze or remove judgment unit, action fragment extraction unit, fragment recombination unit, gray-scale exploration queue management unit, and strategy analysis report generation unit. This step is performed on the strategy evolution service node, which is connected in parallel to the evaluation record table, strategy evolution library index table, and audit log repository, and maintains a strategy index synchronization channel with the strategy assembly service of S200. Specifically, this step takes the strategy evaluation record output by S400 as input. The strategy evaluation record includes at least the strategy identifier field, effectiveness coefficient field, confidence field, timestamp field, and decay factor field, and contains associated information such as strategy library version, unified time anchor point, stress situation label sequence encoding index field, psychological profile cluster identifier field, situation cluster identifier field, and audit field. The effectiveness coefficient field is defined as the estimated effect of the strategy identifier under the current cluster caliber; the confidence field is defined as a characterization of the stability of the effectiveness coefficient field; the timestamp field describes the time when the strategy evaluation record was generated; and the decay factor field describes the weight decay caliber of the strategy evaluation record over time. The minimum set of core parameters required for this step includes the policy identifier field, validity coefficient field, confidence level field, timestamp field, policy library version, and unified time anchor. When any of the above fields are missing from the policy evaluation record, the policy evolution library management unit writes the evolution failure flag field and registers the missing field name field. At the same time, the record is written to the exception record queue and an audit field is generated for subsequent review and retrieval.

[0054] During the triggering and session loading phase, the strategy evolution library management unit initiates this step upon receiving a new strategy evaluation record write event or reaching the batch processing trigger threshold. The batch processing trigger threshold is jointly defined by the record count threshold field and the time interval threshold field and written into the runtime configuration. Specifically, the strategy evolution library management unit reads the set of strategy evaluation records within the processing window and merges them according to the strategy identifier field and strategy library version to generate an evolution input set. It then performs deduplication on the evolution input set, based on the record primary key field and timestamp field. Duplicate records are written to the duplicate marker field, and the record corresponding to the latest timestamp is retained for subsequent processing. Further, the strategy evolution library management unit loads the strategy evolution library index table and generates an evolution transaction identifier field. This evolution transaction identifier field is bound to the strategy library version and written to the version association field. Subsequently, it locks the write channel for the strategy evolution library version records and completes sorting updates, state transitions, fragment extraction, gray-scale queuing, and report storage within the same evolution transaction identifier field. Each write action in the process records an audit field and writes it to the audit log repository. The audit field includes at least the operator identifier field, the running node identifier field, the runtime configuration version field, and the input record range field.

[0055] In the decay factor generation and ranking update stage, the ranking update unit generates ranking weights based on the timestamp field and decay factor field as described in the claims, and uses these weights together with the effectiveness coefficient field and confidence field for policy ranking updates. Specifically, the ranking update unit reads the timestamp field from the policy evaluation record and generates a time difference field with the current time difference. It then calculates the decay factor field based on the time difference field and the decay rule table. The decay rule table is bound to the policy library version and includes a decay interval field, a decay multiplier field, and a pruning threshold field. When the time difference field falls into different decay intervals, the corresponding decay multiplier field is used to generate the decay factor field, and a pruning flag field is written when the time difference field exceeds the pruning threshold field. Subsequently, the ranking update unit generates a ranking score field, which is formed by the product of the effectiveness coefficient field and the decay factor field. A confidence threshold field is introduced for filtering. Policies with a confidence field lower than the confidence threshold field are written into a low-confidence candidate flag field and enter the observation queue, not participating in the main ranking update. For strategy identifiers that meet the confidence threshold field, the sorting update unit generates sorting results from high to low according to the sorting score field and writes them into the strategy sorting result field. At the same time, the sorting results and the strategy library version are written into the strategy evolution library version record. The strategy evolution library version record includes at least a version number field, a generation time field, a sorting result summary field, and an input record range field. The version number field is written into the version association field and is used by the acquisition session configuration package of S100 to register the strategy library version and is used by the candidate guidance strategy set assembly of S200 to read the strategy sorting result field.

[0056] During the freeze or remove determination phase, the freeze or remove determination unit migrates the policy state according to the freeze or remove rules and writes it into the policy evolution library index table. The freeze or remove rules include a consecutive count threshold field, a validity threshold field, and a confidence threshold field. The determination conditions include that the validity coefficient field within the consecutive count threshold field is lower than the validity threshold field and the confidence field is higher than the confidence threshold field. Specifically, the freeze or remove determination unit reads the most recent window record sequence from the strategy evaluation record by the strategy identifier field, sorts it by the timestamp field, generates a continuous inefficiency count field, and compares the continuous inefficiency count field with the continuous number of times threshold field. When the determination condition is met, the freeze or remove determination unit further reads the strategy status field and historical operation field from the strategy evolution library index table. If the strategy status field is active, it writes the freeze status and writes the strategy identifier to the freeze queue, and writes the freeze reason field, freeze time field, and audit field. If the strategy status field is already frozen and the determination condition is continuously met, it writes the remove status and deletes the corresponding index record from the strategy evolution library index table, and generates a removal audit field. The removal audit field includes at least the removal operator identifier field, the removal trigger condition field, the removal sorting position field, and the associated evaluation record range field. Furthermore, to maintain consistency with the candidate guidance strategy set of S200, the freeze or removal decision unit pushes a status change notification on the index synchronization channel. The notification payload includes a strategy identifier field, a strategy status field, a freeze queue flag field, and a removal audit field index. When constructing the candidate guidance strategy set, S200 reads this notification and excludes frozen status strategies from the active queue. Removed status strategies no longer appear in the strategy assembly index.

[0057] In the action fragment pattern extraction stage, the action fragment extraction unit extracts efficient strategy action fragment patterns from the intervention action trajectory package associated data and writes them into the action fragment pattern library. The action fragment pattern is defined as a reusable action sub-sequence structure, including a fragment identifier field, a precondition field, a postcondition field, a disable condition field, and a parameter field. The fragment identifier field uniquely identifies the action fragment; the precondition field describes the contextual conditions that must be met before the fragment begins; the postcondition field describes the state change that occurs after the fragment ends; the disable condition field describes the risk constraints that prevent the fragment from being triggered; and the parameter field describes the legal range of action parameters. Specifically, the action fragment extraction unit reads the audit field from the strategy evaluation record, parses the intervention action trajectory package index from the audit field, and locates the intervention action trajectory package corresponding to the strategy identifier field in the trajectory storage area. The intervention action trajectory package at least includes an action type field, an action start and end time field, an action parameter field, a completion marker field, and an interruption reason field. The action segment extraction unit segments the action sequence according to the completion marker field, generating a candidate segment set, and writes the risk segment marker field for segments with the interruption reason field as risk interruption. Subsequently, the action segment extraction unit combines the situational disturbance marker field of S300 with the stress situation label sequence encoding index field of S100 to generate a context summary field, which serves as the input for generating the precondition field and the disabling condition field. It also combines the pretest snapshot package index to generate a strategy start state summary field, which is written into the precondition field. For the candidate segment set, the action segment extraction unit merges the action parameter fields to generate parameter domain fields, and performs conflict marking processing on different parameter domain fields. The conflict marker field is written into the parameter domain audit field. Further, the action segment extraction unit extracts the psychological scale change field, task performance change field, and behavioral stability change field that are close to the segment end time from the posttest effect package to generate a post-state summary field, which is written into the post-condition field. It also establishes an association index between the segment and the validity coefficient field and the confidence field to form a segment score record field. The segment score record field is written into the action segment pattern library and is available for use by the segment reconstruction unit. In a preferred embodiment, the action segment extraction unit prioritizes extracting segments based on strategy identifiers with high confidence fields and high rankings, and the priority rule is written into the extraction scheduling field; in another embodiment, for strategy identifiers corresponding to low confidence candidate marker fields, only segments with risk segment marker fields that are not risky are retained and added to the pattern library, and the relevant filtering rules are written into the audit field.

[0058] During the fragment recombination and new strategy hypothesis generation phase, the fragment recombination unit generates new strategy hypotheses according to fragment recombination rules and writes them into the grayscale exploration queue. The fragment recombination rules include precondition matching, postcondition compatibility, disabled condition merging, and parameter domain conflict detection. Precondition matching determines whether two fragments can be connected under the same context summary field. Postcondition compatibility determines whether the postcondition field of the preceding fragment is consistent with the precondition field of the following fragment. Disabled condition merging performs a union fusion of the disabled condition fields of the fragments and resolves conflicting terms. Parameter domain conflict detection determines conflicts in the range of action parameter fields and generates conflict resolution records. Specifically, the fragment reorganization unit reads the fragments ranked highest in the fragment score record field from the action fragment pattern library, clusters them according to the precondition field, and generates candidate fragment chains. For each candidate fragment chain, the fragment reorganization unit sequentially verifies the compatibility of the postconditions, and writes the link breakpoint field and truncates it to generate a short chain candidate when incompatible. For candidate fragment chains that pass the compatibility verification, the fragment reorganization unit performs disabled condition merging and generates a merged disabled condition field. At the same time, it performs parameter field conflict detection on the parameter field. When a conflict is detected, a parameter field conflict detection record is generated and the candidate fragment chain is marked as a field requiring manual intervention or an automatically downgraded field. The marking rules are loaded by the runtime configuration. Subsequently, the fragment recombination unit encapsulates the verified candidate fragment chain into a new strategy hypothesis. This new strategy hypothesis includes a new strategy identifier field, a fragment chain description field, a merge disabling condition field, a parameter field, a generation time field, and a source fragment index field. A new strategy identifier is then generated and written to the grayscale exploration queue. The grayscale exploration queue consists of a queue index field, an enqueue time field, a risk level field, and an exploration constraint marker field. The risk level field is obtained by mapping the merge disabling condition field to a risk dictionary. The exploration constraint marker field is obtained by comparing the grayscale strategy number limit field with the current grayscale queue length. When the grayscale strategy number limit field is triggered, the grayscale exploration queue management unit executes the queue elimination rule. The elimination rule is determined by both the enqueue time field and the most recent evaluation time field, and the elimination strategy is written to the elimination audit field. After the new strategy hypothesis is enqueued, it is used by S200 as a source for the grayscale exploration queue when screening the candidate guidance strategy set. During the target strategy selection phase, the new strategy identifier field is sampled and scheduled under the constraint of the disabling condition trigger, and the completion marker field and interruption reason field are recorded in the intervention action trajectory package for subsequent S300 and S400 backflow evaluation.

[0059] During the strategy analysis report generation and output phase, the strategy analysis report generation unit generates a strategy analysis report based on the strategy evolution library version record, strategy evaluation record, and gray-scale exploration queue, and writes it to the report storage area. The strategy analysis report includes a group vulnerability distribution table, a strategy effectiveness coefficient comparison matrix, a gray-scale exploration queue risk list, and a situational difficulty gradient parameter table. The group vulnerability distribution table is obtained by aggregating the stress situation label sequence encoding index field and the psychological profile group identifier field, and written to the distribution table index field. The strategy effectiveness coefficient comparison matrix is ​​formed by cross-indexing the strategy identifier field and the stress situation label sequence encoding index field, and written to the matrix index field. The gray-scale exploration queue risk list is formed by summarizing the risk level field, the merged disabled condition field, and the exploration constraint marker field in the gray-scale exploration queue, and written to the risk list index field. The situational difficulty gradient parameter table is obtained by statistically analyzing the distribution of the situational group identifier field and the effectiveness coefficient field in the strategy evaluation record, and written to the gradient parameter index field. Specifically, the strategy analysis report generation unit encapsulates the above content to generate a report primary key field and writes it into the report time field, strategy library version field, and input record range field. The report primary key field is also written into the strategy evolution library version record as a version attachment index. At the same time, the strategy analysis report generation unit pushes a report summary to the index synchronization channel of the strategy assembly service. The summary payload includes a version number field, a strategy sorting result summary field, a frozen queue size field, a gray-scale exploration queue size field, and a risk level distribution field, which are displayed and called by the teaching management interface and used for subsequent rounds of review.

[0060] In one engineering embodiment, the strategy evolution service node is deployed in the backend cluster of the enterprise internal training psychological teaching platform. After each scenario round is identified, S400 writes the strategy evaluation record and triggers the new record writing event. This step starts batch processing on the same day according to the time interval threshold field. The strategy evolution library management unit loads the set of strategy evaluation records in the pending window and generates the evolution transaction identifier field. The sorting update unit generates the decay factor field based on the timestamp field and writes it into the strategy sorting result field. The freeze or removal judgment unit executes the freeze state of inefficient and high-confidence strategies according to the continuous number threshold field and writes it to the strategy assembly index of S200. The action fragment extraction unit searches back the trajectory storage area according to the intervention action trajectory package index and generates action fragment pattern library records. The fragment recombination unit combines high-scoring fragment chains to generate new strategy hypotheses and writes them into the gray-scale exploration queue. The strategy analysis report generation unit completes the storage of the group vulnerability distribution table and the strategy effectiveness coefficient comparison matrix and generates the report primary key field. When the next teaching round starts, S100 reads the updated policy library version field when generating the collection session configuration package, and S200 reads the policy sorting result field when building the candidate dredging policy set and combines the gray-scale exploration queue and the disabled condition trigger table to perform target policy selection, forming a cross-step closed-loop iterative link.

[0061] This step's technical effects can be summarized as follows: Under the input criteria of the strategy evaluation record, this step updates the strategy evolution library and generates version records, and writes the strategy sorting results along with the frozen or removed state transitions into a unified index. Through the linkage of action fragment pattern extraction and fragment recombination rules, new strategy hypotheses are generated and entered into the gray-scale exploration queue, forming an iterative link with subsequent rounds of strategy assembly. The strategy analysis report is aggregated and output under the same version number field, retaining audit fields to support subsequent recalculation and version rollback scenarios.

Claims

1. A method for the evolution of psychological teaching strategies based on feedback on guidance effects, characterized in that, include: S100. Obtain the project teaching event log and generate a stress situation label sequence according to the stress situation vocabulary and event mapping table; Obtain the student psychological profile field set and calculate the profile reliability score, register the student identifier, scene round identifier, unified time anchor point, strategy library version, and generate the data collection session configuration package; S200: Input the collection session configuration package into the pre-test snapshot construction and strategy assembly, generate the pre-test snapshot package and filter the candidate guidance strategy set, execute the target strategy according to the action parameter template and risk disabling conditions, and generate the intervention execution package; the intervention execution package includes the pre-test snapshot package and the intervention action trajectory package; S300: Based on the intervention execution package, perform observation window acquisition and alignment, acquire post-test effect package and perform time window alignment, missing value repair, outlier removal processing, and generate effect vector package; The effect vector package includes a post-test effect package and comparable effect vectors; S400, based on the effect vector package, performs propensity score calculation and dual robust estimation, calculates the effectiveness coefficient and confidence level according to the stress situation label sequence and mental profile clustering, and generates strategy evaluation records; S500 updates the strategy evolution library based on strategy evaluation records, updates the strategy sorting by effectiveness coefficient and confidence level and executes freeze or removal rules, extracts efficient strategy action fragment patterns, generates new strategy hypotheses according to fragment recombination rules and writes them into the gray-scale exploration queue, and generates a strategy analysis report.

2. The method according to claim 1, characterized in that, The stress scenario vocabulary in S100 includes schedule pressure, requirement change pressure, collaboration conflict pressure, review rejection pressure, and resource preemption pressure; the event mapping table includes event type field, trigger condition field, scenario tag field, intensity stratification field, and duration window field. After parsing the event log, the event type is matched according to the trigger condition and the corresponding scenario tag and intensity stratification are output to form a stress scenario tag sequence.

3. The method according to claim 1, characterized in that, In S100, the reliability score of the profile is composed of the missing rate score, consistency score, and volatility score. The missing rate score is calculated by the ratio of the number of missing tests of a field to the number of samples. The consistency score is calculated by the difference between short-window retests of the same indicator. The volatility score is calculated by the standard deviation of the sliding window. The synthesis rules are written into the acquisition session configuration package.

4. The method according to claim 1, characterized in that, In S200, the candidate guidance strategy set is divided into an active queue and a gray-scale exploration queue. When selecting a target strategy, strategies that hit the disabled condition are first screened out, and then selected according to queue priority and random seed. The intervention action trajectory package records the completion mark field and the interruption reason field.

5. The method according to claim 1, characterized in that, The S300 observation window includes an immediate window, a short-term window, and a delayed window. The post-test effect package records changes in psychological scales, task performance, and behavioral stability in the three windows, and also records the situational disturbance marker field. Time window alignment performs resampling and interpolation at a unified time anchor point. Missing data is repaired using neighborhood imputation, and outlier removal uses quantile thresholds.

6. The method according to claim 1, characterized in that, In S400, the propensity score is output by a logistic regression model, and the input features include stress situation label sequence encoding, psychological profile field set, profile reliability score, and strategy label; the dual robust estimation consists of a weighted outcome model and a weighted processing model, with the weights generated by the propensity score and the confidence level calculated by the resampled confidence interval.

7. The method according to claim 1, characterized in that, In S500, the policy evaluation record includes a timestamp field and a decay factor field. The decay factor generates a weight based on the time difference between the timestamp field of the policy evaluation record and the current time. The policy ranking is generated by multiplying the effectiveness coefficient by the weight and combining it with the confidence threshold. The ranking result is written into the version record of the policy evolution library.

8. The method according to claim 1, characterized in that, In S500, the action fragment pattern includes a fragment identifier field, a precondition field, a postcondition field, a disabled condition field, and a parameter field. The fragment reorganization rules include precondition matching, postcondition compatibility, disabled condition merging, and parameter field conflict detection. After reorganization, a new policy hypothesis is generated and a new policy identifier is generated and written to the grayscale exploration queue.

9. The method according to claim 1, characterized in that, The freeze or remove rules in S500 include a condition that the validity coefficient is below the threshold for M consecutive times and the confidence level is above the threshold.

10. The method according to claim 1, characterized in that, The strategy analysis report includes a group vulnerability distribution table, a strategy effectiveness coefficient comparison matrix, a risk list for the gray-scale exploration queue, and a situational difficulty gradient parameter table. The group vulnerability distribution table is clustered and aggregated according to the stress situation label sequence and psychological profile, and the strategy effectiveness coefficient comparison matrix is ​​output according to the strategy identifier and stress situation label sequence index.

Images