Intelligent adaptive training system and platform based on service data closed-loop feedback

Abstract

The application belongs to the technical field of intelligent training, and discloses an intelligent adaptive training system and platform based on business data closed-loop feedback. The system comprises a data acquisition module, a double-channel evaluation module for generating a first decision path graph and a first ability quantitative value represented by a knowledge graph, a deviation analysis and scheme generation module, a secondary acquisition module for generating a second decision path graph and a second ability quantitative value, a drift detection module for calculating a decision path drift amount, a semantic alignment module for triggering a semantic alignment program calibrated by an ability gain amount when the drift amount exceeds a threshold value, and for taking the ability gain amount as the calibrated ability gain amount when the drift amount does not exceed the threshold value, and a closed-loop feedback module for triggering an intervention strategy when the calibrated gain amount is lower than a threshold value. The application eliminates the pollution of the decision path drift caused by training to the ability evaluation scale, and improves the accuracy of ability gain calculation.

Description

Technical Field

[0001] This invention relates to the field of intelligent training technology, and in particular to an intelligent adaptive training system and platform based on closed-loop feedback of business data. Background Technology

[0002] With the deepening of enterprise digital transformation, intelligent adaptive training systems driven by business data are gradually becoming an important tool for improving employee capabilities. In existing technologies, typical adaptive training systems usually construct a closed-loop mechanism of business data collection—capability assessment—training program generation—effect feedback: First, employee workflow data or performance data is obtained from the business system, and a capability profile or capability score is generated based on a preset capability assessment model; then, based on the gap between the capability assessment results and the job capability benchmark, a personalized training program is generated and pushed to the employee; after the employee completes the training, business data is collected again to evaluate the training effect, and the capability assessment model or training recommendation strategy is iteratively updated based on the feedback, forming a continuously optimized intelligent learning closed loop. Some existing technologies also introduce the collection and analysis of operational behavior data, comparing employee operation sequences with standard operating procedures to diagnose employee operational standardization or capability gaps.

[0003] However, the aforementioned existing technologies suffer from a fundamental, long-overlooked technical flaw: they all assume that competency assessments before and after training are conducted under the same benchmark, meaning they assume that training intervention will not change the baseline conditions for competency assessment. In practice, however, training intervention itself alters employees' decision-making paths and operational behavior patterns when completing business tasks—even if an employee's actual competency has not significantly improved, the structural characteristics of their operational sequence will change due to the guidance of the training content. When using static competency quantification models to evaluate business performance data under different behavioral patterns before and after training, the competency assessment benchmark itself has been contaminated by the behavioral pattern changes induced by training, leading to systematic errors in the calculation of competency gains. Existing technologies only focus on how to optimize training programs based on assessment results, failing to recognize the deeper problem of training altering decision-making paths and thus contaminating the assessment benchmark, causing the accuracy of closed-loop feedback to gradually deteriorate with the increase in the training cycle.

[0004] Therefore, this invention proposes an intelligent adaptive training system and platform based on closed-loop feedback of business data. Summary of the Invention

[0005] This invention provides an intelligent adaptive training system and platform based on closed-loop feedback of business data. By detecting the structural drift of employees' decision-making path maps before and after training and triggering a strategy optimization semantic alignment procedure to calibrate the capability gain, it eliminates the contamination of capability assessment scale by behavioral pattern changes caused by training. This transforms the capability gain calculation result from a mixed signal interfered with by behavioral drift into a reliable metric reflecting the net training effect, achieving the technical effect that the accuracy of closed-loop feedback continuously converges with the training cycle rather than gradually deteriorates.

[0006] This invention provides an intelligent adaptive training system based on closed-loop feedback of business data, comprising: The data acquisition module is used to collect employees' business operation sequence logs and business performance result data from the business system within the first time window; The dual-channel evaluation module is used to input business operation sequence logs into the decision path extraction model to generate a first decision path graph represented by a knowledge graph, and to input business performance result data into the capability quantification model to generate a first capability quantification value. The deviation analysis and solution generation module is used to compare the first capability quantification value with the capability benchmark value, locate capability deviation items, generate the first training solution based on the capability deviation items, and push it to the employee terminal. The secondary data acquisition module is used to collect employees' business operation sequence logs and business performance result data again within the second time window, and generate a second decision path diagram and a second capability quantification value represented by a knowledge graph. The drift detection module is used to calculate the decision path drift between the first decision path graph and the second decision path graph. The semantic alignment module is used to calculate the difference between the first capability quantization value and the second capability quantization value as the capability gain. When the decision path drift exceeds the preset drift threshold, the semantic alignment program executed by the policy optimization algorithm is triggered to calibrate the capability gain and output the calibrated capability gain. When the decision path drift does not exceed the preset drift threshold, the capability gain is used as the calibrated capability gain. The closed-loop feedback module is used to trigger an intervention strategy when the calibrated capability gain is lower than a preset gain threshold, generating and pushing an intervention training plan to the employee terminal.

[0007] Furthermore, the semantic alignment module includes: The event acquisition submodule is used to collect system event logs that are time-aligned with the business operation sequence logs from the business system. The system event logs record internal events generated when the business system responds to employee operations. The coupling calculation submodule is used to pair operation nodes in the business operation sequence log with event nodes in the system event log according to time sequence, construct an operation-event coupling pair sequence, calculate the coupling transition probability matrix between adjacent coupling pairs in the operation-event coupling pair sequence, the coupling transition probability matrix represents the dynamic matching pattern between employee operation and system response, and perform difference calculation on the first coupling transition probability matrix corresponding to the first decision path diagram and the second coupling transition probability matrix corresponding to the second decision path diagram to obtain the coupling deviation degree. The compensation calculation submodule is used to apply coupling deviation compensation to the capability gain when the coupling deviation exceeds the preset coupling deviation threshold. The coupling deviation compensation uses the coupling deviation as the independent variable and calculates the compensation amount using a preset compensation function. The compensation output submodule is used to output the compensated capability gain as the calibrated capability gain.

[0008] Furthermore, the drift detection module includes: The Change Event Acquisition Submodule is used to acquire business system change events that occur between the first time window and the second time window. Business system change events include one or more of the following: system version update, business process change, and interface redesign. The drift decomposition submodule is used to construct a structured impact matrix of business system change events on the operation sequence. The structured impact matrix quantifies the impact of business system change events on the addition, deletion, and modification of operation nodes. Based on the structured impact matrix, the training-induced drift component is decomposed from the decision path drift between the first decision path diagram and the second decision path diagram. The training-induced drift component is used as the updated decision path drift to replace the decision path drift in the first embodiment. It is then output to the semantic alignment module as the decision basis for triggering the semantic alignment procedure.

[0009] Furthermore, the semantic alignment module includes: The clustering preprocessing submodule is used to cluster employees in the same position according to the decision path drift amount to obtain multiple drift interval clusters; The hierarchical filtering submodule is used to obtain the job level labels of employees. The job level labels include junior, intermediate and senior levels. After determining the drift interval cluster to which the employee belongs, the module filters out a subset of employees with the same job level labels as the employee from the drift interval cluster, which is used as the constraint calibration subset. The boundary calibration submodule is used to calculate the first statistical distribution characteristics of the capability gain of employees within the constraint calibration subset, and the second statistical distribution characteristics of the capability gain of the subset of employees with adjacent job level labels within the drift interval cluster. The mean of the first statistical distribution characteristics is used as the calibration center value, and the difference between the mean of the first statistical distribution characteristics and the mean of the second statistical distribution characteristics is used as the calibration constraint boundary. When the deviation between the capability gain of an employee and the calibration center value does not exceed the calibration constraint boundary, the calibration center value is used for calibration. When the deviation exceeds the calibration constraint boundary, the calibration constraint boundary is used for amplitude limiting calibration. The capability gain after hierarchical constraint calibration is output as the calibrated capability gain.

[0010] Furthermore, the closed-loop feedback module includes: The abnormal edge location submodule is used to compare the second capability quantization value with the capability benchmark value when the calibrated capability gain is lower than the preset gain threshold, relocate the capability deviation item, and perform graph matching between the operation node subgraph associated with the relocated capability deviation item in the second decision path graph and the preset standard decision path template to locate the abnormal transfer edge that deviates from the standard decision path template. The context extraction submodule is used to extract the set of predecessor operation nodes and the set of successor operation nodes of the abnormal transition edge in the second decision path graph, calculate the first conditional transition probability between the abnormal transition edge and each predecessor operation node in the set of predecessor operation nodes, and the second conditional transition probability between the abnormal transition edge and each successor operation node in the set of successor operation nodes, and construct the context coupling strength vector of the abnormal transition edge using the first conditional transition probability and the second conditional transition probability. The training content library is used to store atomized cognitive components, which are the smallest indivisible knowledge units of the training content. The classification intervention submodule is used to determine that when the magnitude of the context coupling strength vector exceeds a preset coupling strength threshold, the abnormal transition edge is a context-dependent defect. It extracts the atomized cognitive component corresponding to the abnormal transition edge, its predecessor operation node, and its successor operation node from the training content library, and retrieves the training content associated with the ability deviation item from the training content library. The atomized cognitive component and the training content associated with the ability deviation item are then combined to generate a context-dependent intervention training plan. When the magnitude of the context coupling strength vector does not exceed the preset coupling strength threshold, the abnormal transition edge is determined to be an isolated defect. It extracts the atomized cognitive component corresponding to the abnormal transition edge from the training content library, and retrieves the training content associated with the ability deviation item from the training content library. The atomized cognitive component and the training content associated with the ability deviation item are then combined to generate an isolated training intervention training plan.

[0011] Furthermore, the secondary acquisition module includes: The time lag analysis submodule is used to perform time lag cross-correlation analysis on the business operation sequence logs and business performance result data within the first time window, calculate the time lag correlation coefficient curve between the operation sequence change and the performance result change, and extract the peak time lag from the time lag correlation coefficient curve. The peak time lag is the time difference between the operation sequence change and the performance result change. The window adjustment submodule is used to adjust the start time of the second time window based on the peak time delay, so that the start time of the second time window is delayed by the peak time delay relative to the push time of the first training plan. Business operation sequence logs and business performance result data are collected within the adjusted second time window.

[0012] Furthermore, the deviation analysis and solution generation module includes: The co-occurrence analysis submodule is used to obtain the employee's ability deviation items, and at the same time obtain the set of historical ability deviation items accumulated by the employee in the historical training cycle. It constructs a deviation co-occurrence matrix of the ability deviation items, which records the co-occurrence frequency and conditional probability between the ability deviation item and each historical ability deviation item in the historical ability deviation item set. The association aggregation submodule is used to identify associated deviation items whose conditional probability with an employee's ability deviation item exceeds a preset association threshold using the deviation co-occurrence matrix. It then merges the ability deviation items and associated deviation items into a composite deviation item set and generates a first training plan that covers associated ability deficiencies based on the composite deviation item set.

[0013] Furthermore, the deviation analysis and solution generation module includes: The rhythm analysis submodule is used to obtain the business operation rhythm data of employees after the first training plan is generated. The business operation rhythm data includes the distribution of operation interval duration and the distribution of operation intensive periods. Using the business operation rhythm data, the low operation load window of employees is identified. The low operation load window is a continuous time period in which the operation interval duration exceeds the preset interval threshold and is not in the operation intensive period. The timing scheduling submodule is used to schedule the push time of the first training plan to be pushed within the low-load operation window.

[0014] Furthermore, the secondary acquisition module includes: The type identification submodule is used to obtain the training content type of the first training plan. The training content type includes one of knowledge-based training, skills training, and rule training. The cycle determination submodule is used to query a preset type-effective cycle mapping table based on the type of training content to obtain the corresponding effective cycle parameter. The effective cycle parameter represents the average delay time from the completion of the training content type to its effect on business performance. The effective cycle parameter is used as the minimum interval between the second time window and the push time of the first training plan to determine the start time of the second time window, and business operation sequence logs and business performance result data are collected within the determined second time window.

[0015] This invention provides an intelligent adaptive training platform based on closed-loop feedback of business data, which is deployed with any of the above-mentioned intelligent adaptive training systems based on closed-loop feedback of business data.

[0016] The beneficial effects of this invention compared to existing technologies are as follows: By constructing a technical architecture of dual-channel evaluation, drift detection, semantic alignment, and closed-loop feedback, it discovers and solves for the first time a fundamental technical defect that has long been overlooked in existing technologies—namely, the systematic contamination of competency assessment benchmarks caused by employee decision-making path drift triggered by training intervention. Based on the collection of business operation sequence logs and business performance result data, the system uses a knowledge graph to structurally represent the operation sequences, generating a first decision path diagram and a first competency quantification value. After training, data is collected again to generate a second decision path diagram and a second competency quantification value. By calculating the decision path drift between the two decision path diagrams, the degree of structural change in employee operational behavior patterns before and after training is identified. When the drift exceeds a threshold, a semantic alignment procedure executed using a strategy optimization algorithm is triggered to calibrate the competency gain, eliminating the interference of behavioral pattern changes on the assessment results. Finally, the need for further intervention is determined based on the calibrated competency gain. This invention elevates the training effectiveness evaluation from the traditional result feedback optimization paradigm to a dual closed-loop paradigm of process drift calibration plus result feedback. This significantly improves the accuracy of capability gain calculation and the consistency of cross-cycle evaluation, avoiding ineffective or over-training caused by inaccurate evaluation scales, and achieving precise quantification of training effectiveness.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall architecture diagram of the intelligent adaptive training system based on closed-loop feedback of business data in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the overall technical solution of the intelligent adaptive training system based on closed-loop feedback of business data in this embodiment of the invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] like Figure 1 and Figure 2 This invention provides an embodiment of an intelligent adaptive training system based on closed-loop feedback of business data, comprising: The data acquisition module is used to collect employees' business operation sequence logs and business performance result data from the business system within the first time window; The dual-channel evaluation module is used to input business operation sequence logs into the decision path extraction model to generate a first decision path graph represented by a knowledge graph, and to input business performance result data into the capability quantification model to generate a first capability quantification value. The deviation analysis and solution generation module is used to compare the first capability quantification value with the capability benchmark value, locate capability deviation items, generate the first training solution based on the capability deviation items, and push it to the employee terminal. The secondary data acquisition module is used to collect employees' business operation sequence logs and business performance result data again within the second time window, and generate a second decision path diagram and a second capability quantification value represented by a knowledge graph. The drift detection module is used to calculate the decision path drift between the first decision path graph and the second decision path graph. The semantic alignment module is used to calculate the difference between the first capability quantization value and the second capability quantization value as the capability gain. When the decision path drift exceeds the preset drift threshold, the semantic alignment program executed by the policy optimization algorithm is triggered to calibrate the capability gain and output the calibrated capability gain. When the decision path drift does not exceed the preset drift threshold, the capability gain is used as the calibrated capability gain. The closed-loop feedback module is used to trigger an intervention strategy when the calibrated capability gain is lower than a preset gain threshold, generating and pushing an intervention training plan to the employee terminal.

[0022] In this embodiment, the first time window refers to a preset time period before the training plan is pushed out, such as the past 30 days. The first time window is used to collect operational behavior data and performance output data of employees in normal business work status, as the source of baseline data before training.

[0023] In this embodiment, the business system refers to the information system used by enterprise employees in their daily work, such as a customer relationship management system, enterprise resource planning system, office automation system, or customer service ticketing system. The business system records operation logs and performance data generated by employees when completing business tasks.

[0024] In this embodiment, the business operation sequence log is a sequence of operation steps recorded chronologically when an employee performs operations on the business system. Each operation step includes the operation type, the operation object, and the operation timestamp. The business performance result data is the quantitative result of the business indicators completed by the employee within a preset assessment period, including any combination of task completion rate, customer evaluation score, and operation compliance rate. The business operation sequence log reflects the employee's decision-making path and operating habits in completing tasks, while the business performance result data reflects the quality of the employee's task output.

[0025] In this embodiment, the data acquisition module retrieves employee business operation sequence logs and business performance result data within the start and end time range of the first time window through the application programming interface or log acquisition agent provided by the business system. The acquisition process associates data by employee ID and timestamp to ensure that the business operation sequence logs and business performance result data of the same employee correspond in the time dimension.

[0026] In this embodiment, the dual-channel evaluation module receives the business operation sequence logs and business performance result data output by the data acquisition module, and sends them to two parallel processing channels. The first channel inputs the business operation sequence logs into the decision path extraction model and outputs a first decision path graph represented by a knowledge graph. The second channel inputs the business performance result data into the capability quantification model and outputs a first capability quantification value. The two channels execute in parallel and are independent of each other.

[0027] In this embodiment, the decision path extraction model is constructed using an inductive mining algorithm based on process mining. The model construction method is as follows: First, the business operation sequence log is preprocessed to extract the operation step chain for each complete task; then, operation nodes are identified according to the order in which the operation steps appear, and transition edges are identified according to the sequential relationship between adjacent operation steps; finally, the frequency of each transition edge is counted as the edge weight. Specifically, for branch selection points appearing in the business operation sequence, decision nodes are identified by analyzing the conditional dependencies between adjacent operation steps: when multiple different subsequent operation steps exist in the historical log after a certain operation step, and the occurrence of each subsequent operation step depends on different input parameters, operation options, or execution results of that operation step, that operation step is identified as a decision node, and the directed edges from that decision node to its subsequent operation steps are marked as conditional relationships. Conditional parameters that trigger the branch can be further labeled on the conditional relationships. Decision nodes and conditional relationship edges together represent the branch decision logic of employees in the business operation process. The input of the decision path extraction model is the business operation sequence log, and the output is a decision path graph represented using a knowledge graph.

[0028] In this embodiment, the first decision path graph, represented by a knowledge graph, is a knowledge representation of the employee's operational decision-making process using a graph data structure. The knowledge graph's node types include operation nodes and decision nodes. Operation nodes represent single-function operations performed by the employee on the business system, while decision nodes represent branch choices made by the employee during the operation. The knowledge graph's edge types include temporal relationships and conditional relationships. Temporal relationships represent the execution order between two operation nodes, while conditional relationships represent the conditional dependencies between decision nodes and operation nodes. The first decision path graph represents the complete operational decision-making path of an employee completing a business task before training.

[0029] In this embodiment, the capability quantification model adopts a multi-dimensional quantification model based on weighted summation. The model is constructed as follows: first, the business performance indicator dimensions required for capability assessment are determined, such as task completion rate, customer evaluation score, and operational compliance rate; then, preset capability weight coefficients are assigned to each indicator dimension; finally, the capability quantification value is defined as the result of the weighted sum of the values ​​of each indicator dimension according to the capability weight coefficients. The input to the capability quantification model is business performance result data, and the output is the capability quantification value.

[0030] In this embodiment, the first capability quantification value is a value output by the capability quantification model after calculating the employee's business performance results data within the first time window, representing the employee's comprehensive capability level before training.

[0031] In this embodiment, the deviation analysis and solution generation module first reads the first capability quantification value and the preset capability benchmark value, and calculates the difference between the two. When the difference exceeds a preset deviation range, the capability dimension corresponding to the difference is marked as a capability deviation item. Then, using the capability deviation item as a search condition, the module queries the training content library for training content associated with the capability deviation item, and combines the queried training content into a first training solution. Finally, the module calls the push interface to send the first training solution to the employee terminal.

[0032] In this embodiment, the competency benchmark is a preset quantitative standard value for each position, representing the minimum competency level that employees in that position should achieve. The competency benchmark can be set by business experts based on the job competency model, or it can be determined based on the statistical average of the quantitative competency values ​​of a group of high-performing employees in the same position.

[0033] In this embodiment, the capability deviation item is a capability dimension identifier in which there is a significant gap between the employee's capability quantification value and the capability benchmark value, used to indicate the specific capability weaknesses that the employee needs to receive training for.

[0034] In this embodiment, the secondary data acquisition module is activated after the first training plan is pushed out. After waiting for the second time window to arrive, it uses the same acquisition method as the data acquisition module to retrieve the employee's business operation sequence logs and business performance result data from the business system again. The secondary data acquisition module inputs the re-collected business operation sequence logs into the decision path extraction model to generate a second decision path graph represented by a knowledge graph; and inputs the re-collected business performance result data into the capability quantification model to generate a second capability quantification value.

[0035] In this embodiment, the second time window refers to a preset time period after the first training program is completed, such as the 7th to 37th day after the first training program is completed. The second time window does not overlap with the first time window and is used to collect operational behavior data and performance output data of employees in actual business work after receiving training.

[0036] In this embodiment, the second decision path graph, represented by a knowledge graph, is a graph structure representation of the operational decision paths of employees when completing the same business task after training. Its node types, edge types, and construction methods are completely consistent with the first decision path graph. The second decision path graph represents the operational decision paths of employees after training.

[0037] In this embodiment, the second capability quantification value is a value output by the capability quantification model after calculating the employee's business performance results data within the second time window, representing the employee's comprehensive capability level after training.

[0038] In this embodiment, the drift detection module receives a first decision path graph output by the dual-channel evaluation module and a second decision path graph output by the secondary acquisition module. It then uses a graph edit distance algorithm to calculate the structural difference between the two graphs, and uses this structural difference as the decision path drift. The graph edit distance is defined as the minimum number of operation steps required to transform the first decision path graph into the second decision path graph. These operation steps include node addition, node deletion, edge addition, and edge deletion. A larger decision path drift indicates a more significant change in the operational decision path when employees complete the same business task before and after training.

[0039] In this embodiment, the semantic alignment module receives the decision path drift amount output by the drift detection module and compares it with a preset drift threshold. When the decision path drift amount exceeds the preset drift threshold, the semantic alignment module triggers a semantic alignment procedure executed using a near-end policy optimization algorithm to calibrate the capability gain. When the decision path drift amount does not exceed the preset drift threshold, the semantic alignment module directly subtracts the first capability quantization value from the second capability quantization value to obtain the capability gain, and outputs this capability gain as the calibrated capability gain.

[0040] In this embodiment, the preset drift threshold is an empirical threshold determined based on the statistical distribution of decision path drift among employees in the same position during a period without training intervention; for example, it is taken as the 95th percentile of this statistical distribution. The preset drift threshold is used to distinguish between normal operational fluctuations and significant changes in decision paths caused by training.

[0041] In this embodiment, the policy optimization algorithm is not limited to the proximal policy optimization algorithm, and can be replaced by any of the trust region policy optimization algorithm, evolutionary policy algorithm, or Bayesian optimization algorithm, as long as it can output a calibration policy based on the decision path drift. The specific implementation of the semantic alignment procedure executed by the policy optimization algorithm is to use the proximal policy optimization algorithm. The proximal policy optimization algorithm is constructed as follows: the calibration operation of the capability gain is modeled as a Markov decision process, where the state space includes the decision path drift and the capability gain, the action space includes the calibration amplitude and calibration direction, and the reward function is defined as the negative of the standard deviation of the calibrated capability gain over multiple consecutive evaluation periods. The inputs of the semantic alignment procedure are the decision path drift, the first capability quantization value, and the second capability quantization value, and the output is the calibrated capability gain. In one specific implementation, the policy network adopts a three-layer fully connected structure, with two neurons in the input layer corresponding to the decision path drift and the initial capability gain, 64 neurons in the hidden layer using the ReLU activation function, and two neurons in the output layer outputting the calibration amplitude and calibration direction. Training was performed using the Adam optimizer with a learning rate of 0.001. After each training cycle, 32 samples were randomly sampled from the experience replay buffer for training, and the target network was updated every 100 steps.

[0042] In this embodiment, the semantic alignment procedure calibrates the capability gain as follows: First, the difference between the first capability quantization value and the second capability quantization value is calculated as the initial capability gain; then, based on the magnitude and direction of the decision path drift, the calibration amplitude and calibration direction are determined by the policy output by the near-end policy optimization algorithm; finally, the initial capability gain and calibration amplitude are adjusted according to the calibration direction to obtain the calibrated capability gain. The purpose of calibration is to remove the spurious gain signal introduced by the change in the decision path from the initial capability gain, retaining the net gain portion that reflects the true capability improvement.

[0043] In this embodiment, to address the delayed feedback issue where reward signals for training effectiveness evaluation require multiple evaluation cycles to obtain, an experience replay mechanism is introduced during the training process of the near-end policy optimization algorithm. The state, action, and actual reward obtained after each calibration operation are stored in an experience replay buffer. After each training loop closure, batch data is randomly sampled from the buffer for offline policy updates. Simultaneously, to accelerate policy convergence, an auxiliary reward function based on real-time technical signals is introduced. The auxiliary reward function is defined as the reduction in decision path drift after a single calibration, reflecting the immediate contribution of the calibration behavior itself to reducing evaluation scale contamination, thereby providing dense feedback signals for the training process. In the specific training implementation of the near-end policy optimization algorithm, the main reward function and the auxiliary reward function are fused through a weighted summation. The fusion weights are dynamically adjusted according to the training stage: in the early stages of training, the auxiliary reward function has a higher weight, enabling the policy to converge quickly in a direction that effectively reduces drift; in the later stages of training, the weight of the main reward function is gradually increased to ensure that the final converged policy aims to improve cross-cycle evaluation stability.

[0044] In this embodiment, the preset gain threshold is the minimum capability gain required to determine whether the training has produced a substantial effect, for example, it is set to 0.1. When the calibrated capability gain is lower than the preset gain threshold, it indicates that the training has not produced the expected capability improvement effect and further intervention is required.

[0045] In this embodiment, the closed-loop feedback module receives the calibrated capability gain output from the semantic alignment module and compares it with a preset gain threshold. When the calibrated capability gain is lower than the preset gain threshold, the closed-loop feedback module triggers an intervention strategy, generates an interventional training plan based on the specific characteristics of the capability deviation and decision path drift, and sends the interventional training plan to the employee terminal via a push interface. When the calibrated capability gain is not lower than the preset gain threshold, the closed-loop feedback module does not trigger intervention, and the training loop ends.

[0046] In this embodiment, the intervention strategy is a set of preset business rules used to determine whether to trigger intervention and the type of intervention content based on the calibrated capability gain, capability deviation, and decision path drift. The judgment logic of the intervention strategy is as follows: when the calibrated capability gain is lower than a preset gain threshold, if the decision path drift exceeds a preset drift threshold, an intervention training program mainly focused on operational path correction is generated; if the decision path drift does not exceed the preset drift threshold but the capability deviation still exists, an intervention training program mainly focused on capability enhancement is generated.

[0047] In this embodiment, the intervention training plan is a supplementary training plan generated by the closed-loop feedback module when it determines that the training effect has not met the standards. Its content types include operational path correction training and capability enhancement training. The data structure of the intervention training plan is consistent with the first training plan; both retrieve training content associated with capability deviation items from the training content library and combine them to generate the plan. After the closed-loop feedback module pushes the intervention training plan to the employee's terminal, it initiates a new round of data collection and evaluation.

[0048] Furthermore, the semantic alignment module includes: The event acquisition submodule is used to collect system event logs that are time-aligned with the business operation sequence logs from the business system. The system event logs record internal events generated when the business system responds to employee operations. The coupling calculation submodule is used to pair operation nodes in the business operation sequence log with event nodes in the system event log according to time sequence, construct an operation-event coupling pair sequence, calculate the coupling transition probability matrix between adjacent coupling pairs in the operation-event coupling pair sequence, the coupling transition probability matrix represents the dynamic matching pattern between employee operation and system response, and perform difference calculation on the first coupling transition probability matrix corresponding to the first decision path diagram and the second coupling transition probability matrix corresponding to the second decision path diagram to obtain the coupling deviation degree. The compensation calculation submodule is used to apply coupling deviation compensation to the capability gain when the coupling deviation exceeds the preset coupling deviation threshold. The coupling deviation compensation uses the coupling deviation as the independent variable and calculates the compensation amount using a preset compensation function. The compensation output submodule is used to output the compensated capability gain as the calibrated capability gain.

[0049] In this embodiment, the event acquisition submodule retrieves system event logs that are time-aligned with the business operation sequence logs through the system log interface or message queue provided by the business system. The system event logs are underlying event records automatically generated by the business system in response to employee operations, containing any combination of database query events, interface call events, cache update events, and business rule trigger events. Each system event log record includes the event type, event timestamp, and the operation identifier associated with the event. Time alignment refers to matching the event timestamps of the system event logs with the operation timestamps of the business operation sequence logs within a preset time window, ensuring that each operation node can be associated with the corresponding system response event.

[0050] In this embodiment, the operation nodes in the business operation sequence log are single-function operation units extracted from the log. Each operation node includes an operation type and an operation timestamp. An operation node represents a complete operation performed by an employee on the business system, such as clicking the submit button, entering query conditions, selecting a dropdown option, or opening a details page.

[0051] In this embodiment, an event node in the system event log is a single system event unit extracted from the system event log. Each event node contains an event type and an event timestamp. An event node represents an internal processing event generated when the business system responds to an employee's operation, such as executing a database query, returning data from an interface, writing to a cache, or passing a business rule validation.

[0052] In this embodiment, the process of constructing the operation-event coupling pair sequence by the coupling calculation submodule is as follows: First, sort the operation nodes in the business operation sequence log by operation timestamp; then sort the event nodes in the system event log by event timestamp; next, for each operation node, search in the event node sequence for all event nodes whose event timestamp matches the operation timestamp and whose event-associated operation identifier corresponds to the operation node; finally, pair the successfully matched operation nodes and event nodes into an operation-event coupling pair, arrange all coupling pairs in the order of operation timestamp, and form the operation-event coupling pair sequence.

[0053] In this embodiment, the coupling calculation submodule calculates the coupling transition probability matrix as follows: First, it traverses the sequence of operation-event coupling pairs, counting the number of transitions from the event type of the preceding coupling pair to the event type of the following coupling pair in each adjacent coupling pair; then, it groups the pairs by the event type of the preceding coupling pair, calculating the frequency of transitions from that event type to each subsequent event type; finally, using all event types as row and column indices, and the transition frequencies as matrix elements, it constructs the coupling transition probability matrix. Each row of the coupling transition probability matrix represents the probability distribution of transitions from one system response event type to other system response event types. When the system response event transition pattern caused by employee operations changes before and after training, the numerical distribution of the coupling transition probability matrix will also change accordingly.

[0054] In this embodiment, the first coupling transition probability matrix corresponding to the first decision path diagram is a coupling transition probability matrix obtained by constructing and calculating the above coupling pairs from the business operation sequence logs and system event logs within the first time window, representing the dynamic matching pattern between employee operations and system responses before training.

[0055] In this embodiment, the second coupling transition probability matrix corresponding to the second decision path diagram is a coupling transition probability matrix obtained by constructing and calculating the above coupling pairs from the business operation sequence logs and system event logs within the second time window, representing the dynamic matching pattern between employee operations and system responses after training.

[0056] In this embodiment, the process of the coupling calculation submodule performing difference calculations on the first and second coupling transition probability matrices is as follows: Elements at the same row and column positions in the first and second coupling transition probability matrices are subtracted one by one, and the absolute value of each difference is taken. The sum of all absolute values ​​of the differences is used as the coupling deviation. The coupling deviation measures the degree of change in the system response event transfer pattern triggered by employee operations before and after training.

[0057] In this embodiment, the preset coupling deviation threshold is an empirical threshold determined based on the statistical distribution of the coupling deviation between the operation and system response of employees in the same position during a period without training intervention. For example, it is the 90th percentile of this statistical distribution. The preset coupling deviation threshold is used to distinguish between normal coupling pattern fluctuations and significant coupling pattern changes caused by training.

[0058] In this embodiment, the preset compensation function is a piecewise linear function with coupling deviation as the independent variable. When the coupling deviation is less than or equal to twice the preset coupling deviation threshold, the compensation amount is the coupling deviation multiplied by a first compensation coefficient. When the coupling deviation is greater than twice the preset coupling deviation threshold, the compensation amount is the preset coupling deviation threshold multiplied by a second compensation coefficient, where the second compensation coefficient is greater than the first compensation coefficient. The specific values ​​of each compensation coefficient in the preset compensation function are determined by fitting an empirical relationship between coupling deviation and the degree of overestimation of capability gain in historical training data.

[0059] In this embodiment, the process of applying coupling deviation compensation to the capability gain by the compensation calculation submodule is as follows: First, the coupling deviation degree is input into a preset compensation function to calculate the coupling deviation compensation amount; then, the compensation direction is determined according to the direction of the coupling deviation degree. If the transition mode complexity of the second coupling transition probability matrix is ​​lower than that of the first coupling transition probability matrix, the compensation direction is downward adjustment; otherwise, the compensation direction is upward adjustment; finally, the coupling deviation compensation amount is subtracted or added to the initial capability gain amount according to the compensation direction to obtain the compensated capability gain amount.

[0060] In this embodiment, the compensation output submodule uses the compensated capability gain output by the compensation calculation submodule as the calibrated capability gain output by the semantic alignment module, and passes it to the closed-loop feedback module for subsequent gain threshold judgment.

[0061] Furthermore, the drift detection module includes: The Change Event Acquisition Submodule is used to acquire business system change events that occur between the first time window and the second time window. Business system change events include one or more of the following: system version update, business process change, and interface redesign. The drift decomposition submodule is used to construct a structured impact matrix of business system change events on the operation sequence. The structured impact matrix quantifies the impact of business system change events on the addition, deletion, and modification of operation nodes. Based on the structured impact matrix, the training-induced drift component is decomposed from the decision path drift between the first decision path diagram and the second decision path diagram. The training-induced drift component is used as the updated decision path drift to replace the decision path drift in the first embodiment. It is then output to the semantic alignment module as the decision basis for triggering the semantic alignment procedure.

[0062] In this embodiment, a system version update refers to a change in the software version number of the business system, such as upgrading from version 2.1 to version 2.2. A business process change refers to an adjustment to the preset operation steps, approval nodes, or workflow rules in the business system. A user interface redesign refers to a change in the layout, control positions, or interaction methods of the business system's user interface. All three events are external environmental changes that may occur between the first and second time windows and are independent of training interventions.

[0063] In this embodiment, the operation sequence refers to the sequence of operation nodes arranged chronologically after an employee performs a series of operation steps on the business system. Each operation node corresponds to a single functional operation. The operation sequence is the original data source for constructing the decision path map and is also the object of the structured influence matrix.

[0064] In this embodiment, the process of constructing the structured impact matrix by the drift decomposition submodule is as follows: First, all business system change events occurring between the first and second time windows are identified, and each change event is labeled with its event type and occurrence time. Then, the change logs corresponding to the change events are traversed to extract a list of operation nodes affected by the changes. Being affected by a change means that the operation function corresponding to the operation node has been added, deleted, or modified. Next, a matrix is ​​constructed with operation nodes as row indices and change event types as column indices. Each element in the matrix takes the value of 1, -1, or 0. A value of 1 indicates that the operation node was added due to a change event of the corresponding type; a value of -1 indicates that the operation node was deleted due to a change event of the corresponding type; and a value of 0 indicates that the operation node was not affected by a change event of the corresponding type. The structured impact matrix quantifies the impact of business system change events on the addition, deletion, and modification of each operation node in the operation sequence.

[0065] In this embodiment, the process by which the drift decomposition submodule decomposes the training-induced drift component from the decision path drift using the structured influence matrix is ​​as follows: Graph difference is performed between the first decision path graph and the second decision path graph to identify the nodes and edges that have changed; the changed nodes and edges are matched with the structured influence matrix, and the changed parts that can be matched with records of additions, deletions, and modifications in the structured influence matrix are marked as environmental attribution changes, while the remaining unmatched changed parts are marked as potential training-induced changes; based on the first decision path graph, graph editing is performed only using potential training-induced changes, and the minimum number of editing operation steps required is the training-induced drift component.

[0066] In this embodiment, the training-induced drift component refers to the portion remaining after excluding changes in environmental attribution from the decision path drift. It represents the degree of structural change in the employee's operational decision path caused solely by training intervention, excluding the impact of changes in the business system itself on the operational sequence.

[0067] In this embodiment, the drift decomposition submodule assigns the training-induced drift component to the decision path drift variable, which is then passed to the semantic alignment module as the updated decision path drift. The semantic alignment module uses the updated decision path drift as a criterion and compares it with a preset drift threshold to determine whether to trigger the semantic alignment procedure.

[0068] Furthermore, the semantic alignment module includes: The clustering preprocessing submodule is used to cluster employees in the same position according to the decision path drift amount to obtain multiple drift interval clusters; The hierarchical filtering submodule is used to obtain the job level labels of employees. The job level labels include junior, intermediate and senior levels. After determining the drift interval cluster to which the employee belongs, the module filters out a subset of employees with the same job level labels as the employee from the drift interval cluster, which is used as the constraint calibration subset. The boundary calibration submodule is used to calculate the first statistical distribution characteristics of the capability gain of employees within the constraint calibration subset, and the second statistical distribution characteristics of the capability gain of the subset of employees with adjacent job level labels within the drift interval cluster. The mean of the first statistical distribution characteristics is used as the calibration center value, and the difference between the mean of the first statistical distribution characteristics and the mean of the second statistical distribution characteristics is used as the calibration constraint boundary. When the deviation between the capability gain of an employee and the calibration center value does not exceed the calibration constraint boundary, the calibration center value is used for calibration. When the deviation exceeds the calibration constraint boundary, the calibration constraint boundary is used for amplitude limiting calibration. The capability gain after hierarchical constraint calibration is output as the calibrated capability gain.

[0069] In this embodiment, the clustering preprocessing submodule clusters employees in the same position according to their decision path drift as follows: First, it collects the decision path drift values ​​generated by all employees in the current position during this training loop; then, it uses the K-means clustering algorithm to cluster these values, with a pre-set number of clusters of 3, corresponding to low drift intervals, medium drift intervals, and high drift intervals respectively; after clustering, each employee is assigned to a drift interval cluster, and each drift interval cluster contains a group of employees with similar decision path drift values.

[0070] In this embodiment, the hierarchical screening submodule retrieves employee job level tags from the enterprise's human resources system or job competency management system. Job level tags are pre-configured employee career development level identifiers, categorized into three types: junior, intermediate, and senior. Employees at different levels within the same job position exhibit systematic differences in the maturity of their operational decision-making paths and their potential for skill improvement when performing the same business tasks. Therefore, it is necessary to differentiate between levels when calibrating competency gain.

[0071] In this embodiment, the hierarchical screening submodule determines the constraint calibration subset as follows: First, based on the employee's decision path drift value, the drift interval cluster to which the employee was assigned in the clustering step is determined; then, all employees within this drift interval cluster are traversed, and employees with the same job level label as the current employee are selected, forming the constraint calibration subset. Employees in the constraint calibration subset share similar decision path drift levels with the current employee and are at the same job level; therefore, the statistical characteristics of their capability gain can serve as a reliable reference for calibrating the current employee's capability gain.

[0072] In this embodiment, the boundary calibration submodule calculates the first and second statistical distribution features as follows: the first statistical distribution feature is the mean and standard deviation of the capability gain of all employees within the constraint calibration subset; the second statistical distribution feature is the mean and standard deviation of the capability gain of the subset of employees within the drift interval cluster that have adjacent job level labels to the current employee. Adjacent job level labels refer to a level one level higher or lower than the current employee's job level; for example, if the current employee is at the intermediate level, the adjacent levels would be junior and senior.

[0073] In this embodiment, the boundary calibration submodule determines the calibration center value and calibration constraint boundary as follows: The mean of the first statistical distribution characteristic is used as the calibration center value, representing the reasonable expected level of capability gain within the current drift interval and at the current level; the difference between the mean of the first statistical distribution characteristic and the mean of the second statistical distribution characteristic is used as the calibration constraint boundary, representing the average difference in capability gain between adjacent employee groups within the same drift interval. The purpose of the calibration constraint boundary is to set an upper limit for the calibration amplitude, preventing deviation from the reasonable range of capability gain for the current employee's level during calibration.

[0074] In this embodiment, the boundary calibration submodule uses the calibration center value as the calibration logic: It compares the employee's capability gain with the calibration center value and calculates the deviation. When the deviation is less than or equal to the calibration constraint boundary, it indicates that although the employee's capability gain deviates from the center within the group, the deviation is still within the normal fluctuation range of the same level. In this case, the calibration center value is directly used to replace the employee's capability gain as the calibration result. The calibration center value represents the stable expected value of the group within this drift range and level.

[0075] In this embodiment, the boundary calibration submodule uses the following judgment logic for calibration constraint boundary limiting calibration: when the deviation between an employee's capability gain and the calibration center value is greater than the calibration constraint boundary, it indicates that the employee's capability gain has exceeded the normal fluctuation range of the same level. In this case, instead of directly replacing it with the calibration center value, the employee's original capability gain is adjusted towards the calibration center value, but the adjustment range does not exceed the calibration constraint boundary. Limiting calibration preserves the individual extreme characteristic information in the employee's capability gain while pulling it back to a reasonable range for the hierarchical group.

[0076] In this embodiment, the boundary calibration submodule takes the calibrated capability gain as the capability gain after hierarchical constraint calibration. This value is the calibrated capability gain finally output by the semantic alignment module and is passed to the closed-loop feedback module for gain threshold judgment.

[0077] Furthermore, the closed-loop feedback module includes: The abnormal edge location submodule is used to compare the second capability quantization value with the capability benchmark value when the calibrated capability gain is lower than the preset gain threshold, relocate the capability deviation item, and perform graph matching between the operation node subgraph associated with the relocated capability deviation item in the second decision path graph and the preset standard decision path template to locate the abnormal transfer edge that deviates from the standard decision path template. The context extraction submodule is used to extract the set of predecessor operation nodes and the set of successor operation nodes of the abnormal transition edge in the second decision path graph, calculate the first conditional transition probability between the abnormal transition edge and each predecessor operation node in the set of predecessor operation nodes, and the second conditional transition probability between the abnormal transition edge and each successor operation node in the set of successor operation nodes, and construct the context coupling strength vector of the abnormal transition edge using the first conditional transition probability and the second conditional transition probability. The training content library is used to store atomized cognitive components, which are the smallest indivisible knowledge units of the training content. The classification intervention submodule is used to determine that when the magnitude of the context coupling strength vector exceeds a preset coupling strength threshold, the abnormal transition edge is a context-dependent defect. It extracts the atomized cognitive component corresponding to the abnormal transition edge, its predecessor operation node, and its successor operation node from the training content library, and retrieves the training content associated with the ability deviation item from the training content library. The atomized cognitive component and the training content associated with the ability deviation item are then combined to generate a context-dependent intervention training plan. When the magnitude of the context coupling strength vector does not exceed the preset coupling strength threshold, the abnormal transition edge is determined to be an isolated defect. It extracts the atomized cognitive component corresponding to the abnormal transition edge from the training content library, and retrieves the training content associated with the ability deviation item from the training content library. The atomized cognitive component and the training content associated with the ability deviation item are then combined to generate an isolated training intervention training plan.

[0078] In this embodiment, the subgraph of operation nodes associated with capability deviation items in the second decision path graph refers to the subgraph formed by the operation nodes and their connected edges that are mapped to capability deviation items and extracted from the second decision path graph. The mapping relationship between capability deviation items and operation nodes is defined by a preset capability deviation item and operation node mapping table, which records which operation nodes correspond to each capability deviation item. The operation node subgraph is extracted as follows: starting from the operation node associated with the capability deviation item, extend outward by one hop along the temporal and conditional relationship edges of the knowledge graph, and include the extended nodes and edges into the subgraph to obtain the operation node subgraph.

[0079] In this embodiment, the preset standard decision path template is a graph structure template of ideal operational decision paths pre-constructed for each position and business task. The standard decision path template is constructed as follows: Business operation sequence logs of high-performing employees in the same position are collected during a period without training intervention. After generating a decision path graph using a decision path extraction model, the operation nodes and transition edges that appear most frequently in each decision path graph are selected as template elements. These elements are then reviewed and confirmed by business experts before being permanently stored. The nodes and edges in the standard decision path template represent the optimal operational decision path for completing the business task under that position.

[0080] In this embodiment, the graph matching and abnormal transition edge localization process of the abnormal edge localization submodule is as follows: The subgraph of the operation node isomorphic is matched with a preset standard decision path template to identify operation nodes in the operation node subgraph that do not have matching nodes in the standard decision path template, and transition edges whose transition directions are inconsistent with those in the standard decision path template. Transition edges with inconsistent directions are marked as abnormal transition edges. If an operation node exists in the operation node subgraph that is not in the standard decision path template, then any transition edge originating from that operation node is also marked as an abnormal transition edge.

[0081] In this embodiment, the process of the context extraction submodule extracting the predecessor operation node set and the successor operation node set is as follows: In the second decision path graph, all operation nodes directly connected to the starting node of the abnormal transition edge are collected as the predecessor operation node set, and all operation nodes directly connected to the ending node of the abnormal transition edge are collected as the successor operation node set. The predecessor operation node set represents the operation context performed by the employee before executing the abnormal operation step, and the successor operation node set represents the operation context performed by the employee after executing the abnormal operation step.

[0082] In this embodiment, the process by which the context extraction submodule calculates the first conditional transition probability and the second conditional transition probability is as follows: The first conditional transition probability is calculated by counting the frequency of each predecessor operation node transferring to the starting node of the abnormal transition edge in the second decision path graph, dividing this frequency by the total outgoing edge frequency of the predecessor operation node in the second decision path graph, to obtain the conditional transition probability from the predecessor operation node to the abnormal transition edge. The second conditional transition probability is calculated by counting the frequency of the terminating node of the abnormal transition edge transferring to each successor operation node in the second decision path graph, dividing this frequency by the total outgoing edge frequency of the terminating node in the second decision path graph, to obtain the conditional transition probability from the abnormal transition edge to the successor operation node.

[0083] In this embodiment, the process of constructing the context coupling strength vector by the context extraction submodule is as follows: The first conditional transition probabilities of each predecessor operation node are arranged according to the order of their appearance in the second decision path graph, forming a first probability sequence; the second conditional transition probabilities of each successor operation node are arranged according to the order of their appearance in the second decision path graph, forming a second probability sequence; the first and second probability sequences are concatenated into a vector, which is the context coupling strength vector of the abnormal transition edge. The context coupling strength vector comprehensively reflects the degree of association between the abnormal transition edge and its operation context.

[0084] In this embodiment, an atomized cognitive component refers to the smallest indivisible knowledge unit formed by breaking down the training content stored in the training content library. Each atomized cognitive component corresponds to a micro-operational skill point. Each piece of training content in the training content library is broken down into multiple atomized cognitive components upon entry into the library, and each atomized cognitive component is accompanied by its associated operation node identifier. For example, in a customer service training scenario, the opening remarks for responding to a customer complaint are one atomized cognitive component, and the operation step of confirming the customer's identity information is another atomized cognitive component.

[0085] In this embodiment, the preset coupling strength threshold is a discrimination boundary value used to determine whether an abnormal transition edge belongs to a context-dependent defect or an isolated defect; for example, it is set to 0.5. The preset coupling strength threshold is obtained by training on samples with labeled defect types in historical training data, and the magnitude of the context coupling strength vector that maximizes the defect classification accuracy is taken as the threshold.

[0086] In this embodiment, the classification intervention submodule determines that an abnormal transition edge is a context-dependent defect based on the following condition: the magnitude of the context coupling strength vector is greater than a preset coupling strength threshold. A context-dependent defect indicates that the occurrence of the abnormal operation step is strongly correlated with the operation steps before and after it; that is, the employee only exhibits this operational deviation under a specific operational context. This type of defect needs to be corrected through overall training that incorporates the operational context; correcting the abnormal operation step itself in isolation is ineffective.

[0087] In this embodiment, the process of the classification intervention submodule generating a context-dependent intervention training program is as follows: First, the atomic cognitive components corresponding to the abnormal transfer edge, the predecessor operation node of the abnormal transfer edge, and the successor operation node of the abnormal transfer edge are determined and retrieved from the training content library respectively; then, according to the actual business operation sequence of the predecessor operation node, the abnormal transfer edge, and the successor operation node, the three atomic cognitive components are arranged into a coherent micro-scenario simulation exercise. The micro-scenario simulation exercise includes the reproduction of the operation scenario before the abnormal operation step occurs, the demonstration of the correction of the abnormal operation step, and the demonstration of the normal connection of the subsequent operation steps after the abnormal operation step is corrected, so that employees can understand and correct the operation deviation in a complete operation scenario.

[0088] In this embodiment, the classification intervention submodule determines that an abnormal transition edge is an isolated defect based on the following condition: the magnitude of the context coupling strength vector is less than or equal to a preset coupling strength threshold. An isolated defect indicates that the occurrence of the abnormal operation step is weakly related to the operation steps before and after it, meaning that the employee may produce this operational deviation in any operational situation, which belongs to insufficient mastery of a single operation skill.

[0089] In this embodiment, the process of the classification intervention submodule generating an interventional training program for isolated training is as follows: Atomized cognitive components corresponding to abnormal transition edges are extracted from the training content library; simultaneously, training content associated with competency deviation items is retrieved from the training content library; the atomized cognitive components are converted into training exercises for single operational skills; and the training exercises are combined with the training content associated with competency deviation items to generate an interventional training program for isolated training. The training exercises focus on the correct operation method of the abnormal operational step itself, without needing to reproduce the preceding and following operational scenarios. Repeated training enables employees to master the correct execution method of the operational step. The training content associated with competency deviation items is used to compensate for the deep-seated competency deficiencies that lead to the operational deviation.

[0090] Furthermore, the secondary acquisition module includes: The time lag analysis submodule is used to perform time lag cross-correlation analysis on the business operation sequence logs and business performance result data within the first time window, calculate the time lag correlation coefficient curve between the operation sequence change and the performance result change, and extract the peak time lag from the time lag correlation coefficient curve. The peak time lag is the time difference between the operation sequence change and the performance result change. The window adjustment submodule is used to adjust the start time of the second time window based on the peak time delay, so that the start time of the second time window is delayed by the peak time delay relative to the push time of the first training plan. Business operation sequence logs and business performance result data are collected within the adjusted second time window.

[0091] In this embodiment, the time-delay analysis submodule performs time-delay cross-correlation analysis as follows: First, the first time window is divided into multiple consecutive time sub-windows with a preset step size, each sub-window lasting, for example, 7 days. Then, within each sub-window, the change in the operation sequence and the change in performance results are calculated. The change in the operation sequence is defined as the month-on-month change rate of the total number of operation nodes within that sub-window, and the change in performance results is defined as the month-on-month change rate of the task completion rate in the business performance data within that sub-window. Next, the sequence of changes in the operation sequence is used as the first signal, and the sequence of changes in performance results is used as the second signal. The cross-correlation coefficient between the first signal and the second signal is calculated at different lag times, with the lag time ranging from 0 to 10 time sub-windows. Finally, the cross-correlation coefficients corresponding to all lag times are connected to form a curve, resulting in the time-delay correlation coefficient curve. The cross-correlation coefficient is calculated using the Pearson correlation coefficient.

[0092] In this embodiment, the process of extracting the peak time lag in the time lag analysis submodule is as follows: The point with the largest absolute value of the cross-correlation coefficient is found on the time lag correlation coefficient curve. If the cross-correlation coefficient corresponding to this point is positive, then the lag time corresponding to this point is the peak time lag. The peak time lag indicates how many time sub-windows are needed after an employee's operational behavior changes before the corresponding change effect is observed in the business performance results data. For example, a peak time lag of 2 indicates that the change in the operational sequence will only be reflected in the performance results data after approximately 14 days.

[0093] In this embodiment, the process of adjusting the start time of the second time window by the window adjustment submodule is as follows: The push time of the first training plan is obtained as the baseline time point. The baseline time point is then delayed by the number of days corresponding to the peak time lag. The delayed time point is used as the start time of the second time window. The duration of the second time window is consistent with the duration of the first time window. The adjusted second time window avoids the window period before changes in operational behavior after training have been transmitted to performance results, ensuring that the secondary collected business performance data can fully reflect the true impact of training on business performance.

[0094] Furthermore, the deviation analysis and solution generation module includes: The co-occurrence analysis submodule is used to obtain the employee's ability deviation items, and at the same time obtain the set of historical ability deviation items accumulated by the employee in the historical training cycle. It constructs a deviation co-occurrence matrix of the ability deviation items, which records the co-occurrence frequency and conditional probability between the ability deviation item and each historical ability deviation item in the historical ability deviation item set. The association aggregation submodule is used to identify associated deviation items whose conditional probability with an employee's ability deviation item exceeds a preset association threshold using the deviation co-occurrence matrix. It then merges the ability deviation items and associated deviation items into a composite deviation item set and generates a first training plan that covers associated ability deficiencies based on the composite deviation item set.

[0095] In this embodiment, the co-occurrence analysis submodule obtains the capability deviation items and the historical capability deviation item set as follows: It reads the capability deviation items located in the current training loop from the deviation analysis and solution generation module, and simultaneously queries the training history database for all capability deviation items generated in all historical training cycles completed by the employee before this training. These historical capability deviation items are then summarized into a historical capability deviation item set. Each historical capability deviation item in the historical capability deviation item set is accompanied by the historical training cycle number in which it occurred.

[0096] In this embodiment, the historical training cycle refers to all previous training closed-loop cycles prior to the current training closed loop. Each historical training cycle includes a complete process of data collection, capability assessment, training plan generation, training execution, and effect feedback. Capability deviations identified in the historical training cycles are persistently stored in the training history database for use in deviation correlation analysis of subsequent training cycles.

[0097] In this embodiment, the co-occurrence analysis submodule constructs the deviation co-occurrence matrix as follows: First, duplicate capability deviation items are removed from the historical capability deviation item set to obtain a list of all historically occurring capability deviation item types. Then, using this type list as the row index and the capability deviation item type list as the column index, a square matrix is ​​constructed as the deviation co-occurrence matrix. The co-occurrence frequency of each element in the deviation co-occurrence matrix is ​​the number of historical training periods in which the capability deviation item corresponding to the row index and the capability deviation item corresponding to the column index appear simultaneously. The conditional probability is the co-occurrence frequency divided by the total number of times the capability deviation item corresponding to the row index appears in all historical training periods. The deviation co-occurrence matrix records the correlation information of both the co-occurrence frequency and conditional probability dimensions.

[0098] In this embodiment, the process of identifying associated deviation items by the association aggregation submodule is as follows: The row corresponding to the current capability deviation item is found in the deviation co-occurrence matrix; the conditional probability values ​​of all elements in that row are traversed; capability deviation items corresponding to column indices whose conditional probabilities exceed a preset association threshold are selected; and these capability deviation items are marked as associated deviation items. Associated deviation items indicate that, during historical training cycles, whenever the current capability deviation item appears, these other capability deviation items also have a high probability of appearing simultaneously, indicating an inherent capability correlation between them.

[0099] In this embodiment, the preset association threshold is the minimum conditional probability value for determining whether there is a significant correlation between two capability deviation items, for example, set to 0.6. When the conditional probability between a certain capability deviation item and the current capability deviation item exceeds 0.6, it indicates that these two capability deviation items have appeared simultaneously in more than 60% of the training cycles in historical training data, indicating a strong correlation. The preset association threshold can be adjusted based on the statistical analysis results of actual training data.

[0100] In this embodiment, the process of merging the composite deviation item set in the association aggregation submodule is as follows: the current capability deviation item and all identified associated deviation items are placed into the same set, and duplicate items are removed to obtain the composite deviation item set. Each capability deviation item in the composite deviation item set corresponds to a capability defect dimension that needs to be covered.

[0101] In this embodiment, the process by which the association aggregation submodule generates a first training plan covering associated capability deficiencies is as follows: using each capability deviation item in the composite deviation item set as a search condition, training content associated with each capability deviation item is queried from the training content library. All retrieved training content is then aggregated, deduplicated, and combined into the first training plan. This first training plan simultaneously covers the capability deficiencies corresponding to the current capability deviation item and its associated deviation items, avoiding the omission of associated capability shortcomings due to training only on a single capability deviation item, and reducing the probability of associated deviation items reappearing in subsequent training cycles.

[0102] Furthermore, the deviation analysis and solution generation module includes: The rhythm analysis submodule is used to obtain the business operation rhythm data of employees after the first training plan is generated. The business operation rhythm data includes the distribution of operation interval duration and the distribution of operation intensive periods. Using the business operation rhythm data, the low operation load window of employees is identified. The low operation load window is a continuous time period in which the operation interval duration exceeds the preset interval threshold and is not in the operation intensive period. The timing scheduling submodule is used to schedule the push time of the first training plan to be pushed within the low-load operation window.

[0103] In this embodiment, the operation interval duration distribution refers to the statistical distribution of the time interval between two adjacent operations performed by an employee on the business system. Typically, the operation interval duration over several consecutive workdays is used as a sample to statistically analyze its distribution frequency across different duration intervals. The operation-intensive period distribution refers to the distribution of the number of operations performed by an employee within a single workday across several consecutive time periods. Typically, the number of operations within each time period is statistically analyzed in hourly units.

[0104] In this embodiment, the process by which the rhythm analysis submodule acquires employee business operation rhythm data is as follows: After the deviation analysis and solution generation module generates the first training solution, the rhythm analysis submodule pulls the timestamp data of all operations performed by the employee within the most recent several working days from the business system's operation logs. It calculates the difference between adjacent operation timestamps to obtain the operation interval duration sequence, and statistically analyzes the operation frequency by time period to obtain the operation density for each time period. The statistical distribution of the operation interval duration sequence is used as the operation interval duration distribution, and time periods with operation density exceeding a preset density threshold are merged to form the operation-intensive time period distribution. The operation interval duration distribution and the operation-intensive time period distribution together constitute the business operation rhythm data.

[0105] In this embodiment, the process of identifying low-load operation windows by the rhythm analysis submodule is as follows: First, using the employee's daily working hours as the scanning range, continuous time periods are scanned segment by segment at a preset step size, for example, 15 minutes. Then, for each scanned continuous time period, it is determined whether it simultaneously meets two conditions. Condition one is that the median of the operation interval duration distribution within the continuous time period exceeds a preset interval threshold. Condition two is that the continuous time period does not overlap with any operation-intensive period in the distribution of operation-intensive periods. Continuous time periods that simultaneously meet both conditions are marked as low-load operation windows. Low-load operation windows are idle periods when employees have a low frequency of operation on the business system and a long operation interval.

[0106] In this embodiment, the preset interval threshold is a critical duration value for determining whether the operation interval has reached an idle state, for example, it is set to 120 seconds. When the median operation interval within a continuous time period exceeds 120 seconds, it indicates that the employee's operation frequency has significantly decreased during that time period, and the employee is in a relatively idle state.

[0107] In this embodiment, a period of high operational activity is a continuous period of time in which employees perform operations significantly more frequently than in other periods within a single workday. The determination of a period of high operational activity involves counting the number of operations in each period on an hourly basis, calculating the average number of operations across all periods, and merging and marking consecutive periods with an operational frequency exceeding 1.5 times the average as high operational activity periods.

[0108] In this embodiment, the timing scheduling submodule schedules the push time of the first training plan to be executed within a low-load window as follows: First, it selects the window closest to the current time from all identified low-load windows, prioritizing windows whose duration is longer than the estimated learning time of the first training plan; then, it uses the start time of this window as the push execution time of the first training plan; finally, when the push execution time arrives, it calls the push interface to send the first training plan to the employee terminal. Scheduling the push time to a low-load window allows employees to receive and learn the training content during their free time, avoiding interruptions to their normal work processes by pushing training plans during peak operational periods.

[0109] Furthermore, the secondary acquisition module includes: The type identification submodule is used to obtain the training content type of the first training plan. The training content type includes one of knowledge-based training, skills training, and rule training. The cycle determination submodule is used to query a preset type-effective cycle mapping table based on the type of training content to obtain the corresponding effective cycle parameter. The effective cycle parameter represents the average delay time from the completion of the training content type to its effect on business performance. The effective cycle parameter is used as the minimum interval between the second time window and the push time of the first training plan to determine the start time of the second time window, and business operation sequence logs and business performance result data are collected within the determined second time window.

[0110] In this embodiment, the process by which the type identification submodule obtains the training content type of the first training plan is as follows: Preset training content type tags are extracted from the first training plan generated by the deviation analysis and plan generation module. The training content type tags are determined by the attribute fields of the corresponding training content in the training content library when the first training plan is generated. Each piece of training content is labeled as one of knowledge-based training, skills-based training, or rules-based training when it is entered into the library.

[0111] In this embodiment, knowledge-based training refers to training content that primarily imparts declarative knowledge such as concepts, principles, and process descriptions, such as product knowledge introductions, interpretation of company policies, and explanation of industry regulations. Skills-based training refers to training content that primarily trains procedural skills such as operational techniques, tool usage, and scenario handling, such as system operation drills, customer communication simulations, and equipment operation training. Rule-based training refers to training content that primarily explains rule-based knowledge such as business rules, judgment standards, and processing procedures, such as explanations of approval conditions, discount authority rules, and compliance judgment standards.

[0112] In this embodiment, the preset type-to-effectiveness period mapping table is a mapping relationship table pre-stored in the training system configuration library, used to record the effectiveness period parameter corresponding to each type of training content. The effectiveness period parameter is determined based on the average delay days between the completion of training and the significant change in employee business performance indicators for each type of training content in historical training data. For example, the effectiveness period parameter for knowledge-based training is 3 days, for skills-based training it is 14 days, and for rule-based training it is 7 days. Different types of training content have different speeds at which their effects are transmitted to business performance after employees complete the training. Knowledge-based training usually shows results faster, while skills-based training requires a longer period of practice and consolidation before it is reflected in performance indicators.

[0113] In this embodiment, the process of the period determination submodule querying the type-effective period mapping table is as follows: using the training content type output by the type identification submodule as the query key, a matching record is searched in the type-effective period mapping table, and the effective period parameter value corresponding to that record is read. If the first training plan contains multiple training content types, the maximum value among the effective period parameters corresponding to each type is taken as the final effective period parameter.

[0114] In this embodiment, the effective period parameter is a value in days, representing the average delay required from the completion of training for a specific type of training content to the stable observability of training effects in business performance data. This parameter is used to determine the minimum waiting time for secondary data collection, avoiding premature data collection before the training effects have been fully reflected in business performance.

[0115] In this embodiment, the process by which the period determination submodule determines the start time of the second time window is as follows: using the completion time of the first training plan's push as the baseline time point, the effective period parameter as the minimum interval duration, and taking the date resulting from adding the minimum interval duration to the baseline time point as the start time of the second time window. The duration of the second time window is consistent with the duration of the first time window. This method of determining the start time ensures that there is a sufficiently long effect transmission waiting period between the second and first data collection.

[0116] In this embodiment, after the start time of the second time window determined by the period determination submodule arrives, the secondary acquisition module starts the data acquisition process, pulls the business operation sequence logs and business performance result data of employees within the second time window from the business system, and passes them to the dual-channel evaluation module or drift detection module for subsequent processing.

[0117] This invention provides an embodiment of an intelligent adaptive training platform based on closed-loop feedback of business data, including an application server, a database server, and a network interface. The application server is deployed with any of the above-mentioned intelligent adaptive training systems based on closed-loop feedback of business data. The database server is used to store business operation sequence logs, business performance result data, training content library, and capability benchmark values. The network interface is used for data communication with business systems and employee terminals.

[0118] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. An intelligent adaptive training system based on closed-loop feedback of business data, characterized in that, include: The data acquisition module is used to collect employees' business operation sequence logs and business performance result data from the business system within the first time window; The dual-channel evaluation module is used to input business operation sequence logs into the decision path extraction model to generate a first decision path graph represented by a knowledge graph, and to input business performance result data into the capability quantification model to generate a first capability quantification value. The deviation analysis and solution generation module is used to compare the first capability quantification value with the capability benchmark value, locate capability deviation items, generate the first training solution based on the capability deviation items, and push it to the employee terminal. The secondary data acquisition module is used to collect employees' business operation sequence logs and business performance result data again within the second time window, and generate a second decision path diagram and a second capability quantification value represented by a knowledge graph. The drift detection module is used to calculate the decision path drift between the first decision path graph and the second decision path graph. The semantic alignment module is used to calculate the difference between the first capability quantization value and the second capability quantization value as the capability gain. When the decision path drift exceeds the preset drift threshold, a semantic alignment procedure executed by the policy optimization algorithm is triggered to calibrate the capability gain and output the calibrated capability gain. When the decision path drift does not exceed the preset drift threshold, the capability gain is used as the calibrated capability gain. The closed-loop feedback module is used to trigger an intervention strategy when the calibrated capability gain is lower than a preset gain threshold, generating and pushing an intervention training plan to the employee terminal.

2. The intelligent adaptive training system based on closed-loop feedback of business data according to claim 1, characterized in that, The semantic alignment module includes: The event acquisition submodule is used to collect system event logs that are time-aligned with the business operation sequence logs from the business system. The system event logs record internal events generated when the business system responds to employee operations. The coupling calculation submodule is used to pair operation nodes in the business operation sequence log with event nodes in the system event log according to time sequence, construct an operation-event coupling pair sequence, calculate the coupling transition probability matrix between adjacent coupling pairs in the operation-event coupling pair sequence, the coupling transition probability matrix represents the dynamic matching pattern between employee operation and system response, and perform difference calculation on the first coupling transition probability matrix corresponding to the first decision path diagram and the second coupling transition probability matrix corresponding to the second decision path diagram to obtain the coupling deviation degree. The compensation calculation submodule is used to apply coupling deviation compensation to the capability gain when the coupling deviation exceeds the preset coupling deviation threshold. The coupling deviation compensation uses the coupling deviation as the independent variable and calculates the compensation amount using a preset compensation function. The compensation output submodule is used to output the compensated capability gain as the calibrated capability gain.

3. The intelligent adaptive training system based on closed-loop feedback of business data according to claim 1, characterized in that, The drift detection module includes: The Change Event Acquisition Submodule is used to acquire business system change events that occur between the first time window and the second time window. Business system change events include one or more of the following: system version update, business process change, and interface redesign. The drift decomposition submodule is used to construct a structured impact matrix of business system change events on the operation sequence. The structured impact matrix quantifies the impact of business system change events on the addition, deletion, and modification of operation nodes. Based on the structured impact matrix, the training-induced drift component is decomposed from the decision path drift between the first decision path diagram and the second decision path diagram. The training-induced drift component is used as the updated decision path drift to replace the decision path drift described in claim 1, and output to the semantic alignment module as the decision basis for triggering the semantic alignment procedure.

4. The intelligent adaptive training system based on closed-loop feedback of business data according to claim 3, characterized in that, The semantic alignment module includes: The clustering preprocessing submodule is used to cluster employees in the same position according to the decision path drift amount to obtain multiple drift interval clusters; The hierarchical filtering submodule is used to obtain the job level labels of employees. The job level labels include junior, intermediate and senior levels. After determining the drift interval cluster to which the employee belongs, the module filters out a subset of employees with the same job level labels as the employee from the drift interval cluster, which is used as the constraint calibration subset. The boundary calibration submodule is used to calculate the first statistical distribution characteristics of the capability gain of employees within the constraint calibration subset, and the second statistical distribution characteristics of the capability gain of the subset of employees with adjacent job level labels within the drift interval cluster. The mean of the first statistical distribution characteristics is used as the calibration center value, and the difference between the mean of the first statistical distribution characteristics and the mean of the second statistical distribution characteristics is used as the calibration constraint boundary. When the deviation between the capability gain of an employee and the calibration center value does not exceed the calibration constraint boundary, the calibration center value is used for calibration. When the deviation exceeds the calibration constraint boundary, the calibration constraint boundary is used for amplitude limiting calibration. The capability gain after hierarchical constraint calibration is output as the calibrated capability gain.

5. The intelligent adaptive training system based on closed-loop feedback of business data according to claim 1, characterized in that, The closed-loop feedback module includes: The abnormal edge location submodule is used to compare the second capability quantization value with the capability benchmark value when the calibrated capability gain is lower than the preset gain threshold, relocate the capability deviation item, and perform graph matching between the operation node subgraph associated with the relocated capability deviation item in the second decision path graph and the preset standard decision path template to locate the abnormal transfer edge that deviates from the standard decision path template. The context extraction submodule is used to extract the set of predecessor operation nodes and the set of successor operation nodes of the abnormal transition edge in the second decision path graph, calculate the first conditional transition probability between the abnormal transition edge and each predecessor operation node in the set of predecessor operation nodes, and the second conditional transition probability between the abnormal transition edge and each successor operation node in the set of successor operation nodes, and construct the context coupling strength vector of the abnormal transition edge using the first conditional transition probability and the second conditional transition probability. The training content library is used to store atomized cognitive components, which are the smallest indivisible knowledge units of the training content. The classification intervention submodule is used to determine that when the magnitude of the context coupling strength vector exceeds a preset coupling strength threshold, the abnormal transition edge is a context-dependent defect. It extracts the atomized cognitive component corresponding to the abnormal transition edge, its predecessor operation node, and its successor operation node from the training content library, and retrieves the training content associated with the ability deviation item from the training content library. The atomized cognitive component and the training content associated with the ability deviation item are then combined to generate a context-dependent intervention training plan. When the magnitude of the context coupling strength vector does not exceed the preset coupling strength threshold, the abnormal transition edge is determined to be an isolated defect. It extracts the atomized cognitive component corresponding to the abnormal transition edge from the training content library, and retrieves the training content associated with the ability deviation item from the training content library. The atomized cognitive component and the training content associated with the ability deviation item are then combined to generate an isolated training intervention training plan.

6. The intelligent adaptive training system based on closed-loop feedback of business data according to claim 1, characterized in that, The secondary acquisition module includes: The time lag analysis submodule is used to perform time lag cross-correlation analysis on the business operation sequence logs and business performance result data within the first time window, calculate the time lag correlation coefficient curve between the operation sequence change and the performance result change, and extract the peak time lag from the time lag correlation coefficient curve. The peak time lag is the time difference between the operation sequence change and the performance result change. The window adjustment submodule is used to adjust the start time of the second time window based on the peak time delay, so that the start time of the second time window is delayed by the peak time delay relative to the push time of the first training plan. Business operation sequence logs and business performance result data are collected within the adjusted second time window.

7. The intelligent adaptive training system based on closed-loop feedback of business data according to claim 1, characterized in that, The deviation analysis and solution generation module includes: The co-occurrence analysis submodule is used to obtain the employee's ability deviation items, and at the same time obtain the set of historical ability deviation items accumulated by the employee in the historical training cycle. It constructs a deviation co-occurrence matrix of the ability deviation items, which records the co-occurrence frequency and conditional probability between the ability deviation item and each historical ability deviation item in the historical ability deviation item set. The association aggregation submodule is used to identify associated deviation items whose conditional probability with an employee's ability deviation item exceeds a preset association threshold using the deviation co-occurrence matrix. It then merges the ability deviation items and associated deviation items into a composite deviation item set and generates a first training plan that covers associated ability deficiencies based on the composite deviation item set.

8. The intelligent adaptive training system based on closed-loop feedback of business data according to claim 1, characterized in that, The deviation analysis and solution generation module includes: The rhythm analysis submodule is used to obtain the business operation rhythm data of employees after the first training plan is generated. The business operation rhythm data includes the distribution of operation interval duration and the distribution of operation intensive periods. Using the business operation rhythm data, the low operation load window of employees is identified. The low operation load window is a continuous time period in which the operation interval duration exceeds the preset interval threshold and is not in the operation intensive period. The timing scheduling submodule is used to schedule the push time of the first training plan to be pushed within the low-load operation window.

9. The intelligent adaptive training system based on closed-loop feedback of business data according to claim 1, characterized in that, The secondary acquisition module includes: The type identification submodule is used to obtain the training content type of the first training plan. The training content type includes one of knowledge-based training, skills training, and rule training. The cycle determination submodule is used to query a preset type-effective cycle mapping table based on the type of training content to obtain the corresponding effective cycle parameter. The effective cycle parameter represents the average delay time from the completion of the training content type to its effect on business performance. The effective cycle parameter is used as the minimum interval between the second time window and the push time of the first training plan to determine the start time of the second time window, and business operation sequence logs and business performance result data are collected within the determined second time window.

10. An intelligent adaptive training platform based on closed-loop feedback of business data, characterized in that, The system comprises any one of claims 1 to 9, which is an intelligent adaptive training system based on closed-loop feedback of business data.

Images