Working hour management system and method for AI confidence evaluation and adaptive approval

Abstract

The invention discloses a man-hour management system and method for AI confidence evaluation and adaptive approval. Comprising a data acquisition module for acquiring a demand text and a code submission record; the confidence evaluation module is used for outputting predicted working hours and confidence by adopting Monte Carlo Dropout; the grading examination and approval module is used for executing automatic passing, manual rechecking or automatic rejection according to the confidence coefficient, and allocating rechecking persons according to dominant uncertainty; the deviation monitoring module calculates the deviation between the actual working hours and the predicted working hours; the threshold value self-adaptive module is used for performing Pareto optimal threshold value search by adopting an NSGA-II algorithm and taking an automatic passing rate and a re-checking accuracy rate as double targets; and the model self-optimization module analyzes a positioning deviation root cause through an anti-fact root cause, and performs directional increment updating by adopting an elastic weight consolidation algorithm. Through five-layer closed-loop cooperative control, the problems of low prediction credibility, poor examination and approval efficiency and incapability of continuously optimizing the model are solved, and intelligent closed loop of man-hour management is realized.

Description

Technical Field

[0001] This invention relates to the fields of software development and DevOps technology, and in particular to a time management system and method for AI confidence assessment and adaptive approval. Background Technology

[0002] In modern DevOps software development processes, time reporting, approval, and resource analysis are core components of project cost accounting, R&D efficiency measurement, and resource scheduling. As R&D toolchains become increasingly complex, heterogeneous tools such as requirement management platforms (e.g., Jira, ZenTao) and code hosting repositories (e.g., GitLab, GitHub) support different stages, resulting in significant "information gaps" and "management silos" in time data. Traditional time management solutions suffer from the following fundamental technical limitations:

[0003] 1. AI predictions lack an uncertainty quantification mechanism: Existing technologies have attempted to predict work hours based on historical work hour data using machine learning models, but these models only output point estimates and cannot measure the uncertainty of the predictions. This technical deficiency prevents downstream approval processes from quantifying the reliability of the prediction results, creating a dilemma for approvers: blindly trusting low-quality predictions may distort work hour data; manually reviewing all predictions negates the efficiency advantage of automated data entry. Current technologies lack a means to quantify prediction uncertainty and transmit it to downstream processes.

[0004] 2. Disconnect between Approval Strategy and Predictive Uncertainty: In existing technologies, the approval process is based on static rules independent of AI predictions. Even if the AI ​​model can output confidence levels, existing systems lack the technical mechanism to differentiate approvals based on confidence levels, failing to achieve adaptive triage of "automatic approval for high confidence levels, manual intervention for medium confidence levels, and automatic rejection for low confidence levels." This disconnect causes the approval process to be unable to respond to changes in the reliability of prediction results, resulting in wasted management costs and missed detections of low-quality predictions.

[0005] 3. The model cannot be continuously optimized using approval results: When there is a discrepancy between AI predictions and actual working hours, the existing system cannot automatically capture the deviation data and feed it back into model training. Model performance stagnates in the initial training phase and cannot continuously improve with changes in business scenarios, forming a one-way "prediction-application-abandonment" process. This technical deficiency causes the model to be unable to adapt to dynamic changes in team development models, technology stacks, or requirement types, and prediction accuracy gradually declines over time.

[0006] 4. Threshold adjustment uses single-objective optimization, failing to balance conflicting objectives: Existing systems mostly use fixed values ​​or simple rule adjustments for approval thresholds, failing to consider the synergistic optimization of the conflicting objectives of automatic approval rate and review accuracy. Increasing the automatic approval rate may lead to low-quality predictions flowing into the production system, while increasing the review accuracy requires reducing the automatic approval rate and increasing labor costs. Traditional single-objective optimization methods cannot handle this conflicting relationship, preventing the system from achieving an optimal balance between efficiency and accuracy.

[0007] In summary, the fundamental technical flaw of the existing solution lies in the fact that the three stages of prediction, approval, and optimization are independent and cannot form a closed loop. Specifically, this manifests as follows: 1. The forecasting process only outputs point estimates, which cannot provide quantitative evidence of forecast reliability, causing downstream approval processes to "dare not use" the forecasts. 2. The approval process cannot adaptively adjust strategies based on the reliability of predictions, leading to a dilemma between efficiency and accuracy: blind trust results in data distortion, while full review leads to low efficiency. 3. The optimization process cannot utilize approval feedback data, causing model performance to stagnate at the initial training stage, and prediction accuracy to decline over time.

[0008] The aforementioned deficiencies are not isolated but rather systemic problems that are interconnected and mutually causal: unreliable predictions lead to untrustworthy approval processes; rigid approval processes prevent feedback on deviations; lack of feedback prevents model optimization; and model degradation further exacerbates unreliable predictions. Current technologies lack a solution that can bridge these three stages and achieve synergistic optimization.

[0009] The core difference between this invention and existing technologies lies in the construction of a five-layer closed-loop collaborative control system that uses confidence assessment as the link and connects the three major links of prediction, approval, and optimization. Through deep collaboration of confidence decomposition, hierarchical approval, deviation monitoring, counterfactual root cause analysis, targeted incremental learning, and multi-objective threshold adaptation, it achieves intelligent closed-loop management of the entire process, which is reliable in prediction, efficient in approval, and self-evolving in model. Summary of the Invention

[0010] This invention provides a work hour management system and method based on AI confidence assessment and adaptive approval. The system aims to solve core technical problems in the R&D process, such as the lack of reliability in AI work hour predictions, the disconnect between the approval process and prediction, the inability to continuously optimize models, and the single-objective nature of threshold adjustments. Ultimately, it achieves intelligent closed-loop management of the entire process of work hour reporting, approval, and analysis.

[0011] The technical solution adopted in this invention is as follows:

[0012] In a first aspect, the present invention provides a work hour management system based on AI confidence assessment and adaptive approval. The system includes: a data acquisition module, a confidence assessment module, a tiered approval module, a deviation monitoring module, a threshold adaptation module, and a model self-optimization module; wherein... The data acquisition module is used to connect to the requirement management platform and the code hosting repository to automatically collect requirement texts and code submission records associated with the target task; The confidence assessment module is used to input the requirement text and code submission record into a pre-trained work hour prediction model and output the predicted work hours and the corresponding confidence level; wherein, the confidence level is calculated based on the uncertainty metric output by the work hour prediction model and is used to characterize the reliability of the prediction result; The tiered approval module is used to execute different approval strategies based on the comparison results between the confidence level and preset first and second thresholds, wherein the first threshold is greater than the second threshold. If the confidence level is greater than or equal to the first threshold, the approval is automatically granted and approved work hour data is generated. If the confidence level is between the second threshold and the first threshold, a manual review process is triggered, and the reason for the low confidence level is marked on the review interface. If the confidence level is less than or equal to the second threshold, the approval will be automatically rejected and the reason for rejection will be recorded. The deviation monitoring module is used to collect actual working hour data after approval and calculate the deviation between the actual working hour data and the predicted working hour. The threshold adaptive module is used to dynamically adjust the first threshold and the second threshold using the Pareto optimization algorithm with the dual objectives of maximizing the automatic pass rate and maximizing the review accuracy. The model self-optimization module is used to trigger the incremental update of the working time prediction model by using the feature data of the target task and the actual working time data as incremental training samples when the deviation exceeds a preset deviation threshold.

[0013] In some embodiments, the data acquisition module includes a requirement data acquisition unit, a code behavior acquisition unit, and a data normalization unit. The requirement data acquisition unit is used to connect to a requirement management platform via API to acquire requirement data, which includes: requirement ID, title, description, requirement type, creation time, and handler information. The code behavior acquisition unit is used to connect to a code hosting repository via Webhook or API to acquire code behavior data from code commit records, which includes: committer, commit time, list of modified files, number of modified lines, and commit message information. The data normalization unit is used to unify the format of the requirement data and the code behavior data, perform deduplication, and establish a unique association mapping between requirements and code commits based on the requirement ID in the commit message or an association model trained based on historical data.

[0014] In some embodiments, the confidence assessment module includes a feature extraction unit, a prediction unit, and a confidence estimation unit; wherein, the feature extraction unit is used to extract semantic feature vectors of the requirement text through a pre-trained language model, extract code structure feature vectors of the code submission records through abstract syntax tree parsing, and fuse the semantic feature vectors and code structure feature vectors; the prediction unit is used to output the predicted working time based on the fused feature vectors; the confidence estimation unit is used to calculate the variance of the prediction distribution through multiple forward propagations using the Monte Carlo Dropout method or a deep ensemble learning method, and determine the confidence level based on the variance; the confidence estimation unit is also used to decompose the confidence level into a requirement uncertainty component, a code uncertainty component, and a historical uncertainty component, wherein, the requirement uncertainty component is calculated based on the semantic perplexity and information entropy of the requirement text; the code uncertainty component is calculated based on the coefficient of variation of the cyclomatic complexity distribution and the dispersion of the number of modified lines of code submission records; the historical uncertainty component is calculated based on the standard deviation of the prediction error distribution of similar historical tasks.

[0015] In some embodiments, the hierarchical approval module includes a review management unit, used to dynamically select a review approver based on the type of the dominant uncertainty component when a manual review is triggered, wherein the dominant uncertainty component is the component with the largest value among the requirement uncertainty component, code uncertainty component, and historical uncertainty component; if the type of the dominant uncertainty component is the requirement uncertainty component, the requirement analyst is selected as the review approver; if the type of the dominant uncertainty component is the code uncertainty component, the technical lead is selected as the review approver; if the type of the dominant uncertainty component is the historical uncertainty component, the project manager is selected as the review approver.

[0016] In some embodiments, the deviation monitoring module includes a deviation calculation unit, a threshold comparison unit, and a sample storage unit; wherein, the deviation calculation unit is used to collect actual working hour data after approval and calculate the relative deviation δ =|actual working hours - predicted working hours| / max(actual working hours, predicted working hours); the threshold comparison unit is used to compare the relative deviation with a preset deviation threshold to determine whether to trigger sample collection; the sample storage unit is used to store the feature data, predicted working hours, actual working hours, deviation, confidence level, and timestamp of the target task into an incremental sample library when the relative deviation exceeds the preset deviation threshold.

[0017] In some embodiments, the threshold adaptive module includes a performance statistics unit, a Pareto optimization unit, and a threshold update unit; wherein, the performance statistics unit is used to periodically calculate the automatic pass rate and review accuracy corresponding to each confidence interval within a historical time period; the Pareto optimization unit is used to perform Pareto optimal threshold search using the non-dominated sorting genetic algorithm NSGA-II, with the dual objectives of maximizing the automatic pass rate and maximizing the review accuracy, and output the optimal threshold combination on the Pareto front; the threshold update unit is used to select a set of thresholds from the Pareto front according to the current business preferences and update the first threshold and the second threshold.

[0018] In some embodiments, the model self-optimization module includes an incremental sample caching unit, a directional weight adjustment unit, and an incremental training unit; wherein, the incremental sample caching unit is used to store samples whose deviation exceeds a preset deviation threshold into an incremental sample library; the directional weight adjustment unit is used to amplify the weight adjustment magnitude of the corresponding features in the work time prediction model according to the counterfactual root cause analysis results; the counterfactual root cause analysis is achieved by calculating the causal contribution of each feature to the deviation; the incremental training unit is used to perform incremental updates using an elastic weight consolidation algorithm, with the loss function being: L(θ) = L_new(θ) + λ / 2 · Σ_i (F_i / η_i) (θ_i - θ_i^*)^2, where η_i is the normalized value of the root cause contribution of feature i.

[0019] Secondly, this invention provides a time management method for AI confidence assessment and adaptive approval, comprising the following steps: Step 1: Connect the requirements management platform and the code hosting repository to automatically collect the requirement text and code submission records associated with the target task; Step 2: Input the required text and code submission record into the work hour prediction model, and output the predicted work hours and the corresponding confidence level; Step 3: Execute different approval strategies based on the comparison results between the confidence level and the preset first threshold and second threshold, including automatically approving the approval, triggering manual review and marking the reason for the low confidence level, or automatically rejecting the approval. Step 4: Collect actual working hour data after approval, and calculate the deviation between the actual working hour data and the predicted working hours; Step 5: With the dual objectives of maximizing the automatic pass rate and maximizing the review accuracy, the Pareto optimization algorithm is used to dynamically adjust the first threshold and the second threshold. Step 6: When the deviation exceeds the preset deviation threshold, the feature data of the target task and the actual working time data are used as incremental training samples to trigger the incremental update of the working time prediction model.

[0020] The advantages or beneficial effects of the above technical solutions include at least the following: 1. Introducing Confidence Assessment to Enhance Prediction Reliability: This invention outputs confidence scores using Monte Carlo Dropout or deep ensemble learning methods, providing a quantitative basis for subsequent approval. Experiments show that the confidence score output can effectively distinguish between high-reliability predictions (confidence score > 0.85) and low-reliability predictions (confidence score < 0.55), avoiding the dilemma of "either fully believe or not believe at all" in existing technologies, enabling approvers to adopt differentiated processing strategies based on the confidence score.

[0021] 2. Implement tiered approval and optimize management efficiency: This invention automatically triages applications based on confidence level. High-confidence results are automatically approved, medium-confidence results are precisely assigned to reviewers, and low-confidence results are automatically rejected. Compared with the traditional single approval process, the automatic approval rate can reach 60%-80%, the workload of manual review is reduced by more than 50%, and the review accuracy rate is improved by 30%.

[0022] 3. Multi-objective Pareto threshold optimization to balance conflicting objectives: This invention uses the NSGA-II algorithm for multi-objective optimization, dynamically selecting the optimal threshold on the Pareto front of automatic pass rate and review accuracy, solving the problem that traditional single-objective optimization cannot handle conflicting objectives, and achieving the optimal balance between efficiency and accuracy in the system.

[0023] 4. Establishing a bias feedback loop to achieve model self-optimization: This invention feeds back the approved actual working hour data to the model and performs incremental updates through the Elastic Weight Consolidation (EWC) algorithm, which both absorbs new knowledge and prevents catastrophic forgetting. Experiments show that after 5 rounds of incremental updates, the model prediction error (MAE) is reduced by about 25%, and the prediction accuracy for historical types of tasks does not degrade.

[0024] 5. Counterfactual Root Cause Analysis and Targeted Optimization for Improved Efficiency: This invention locates the root causes of bias through counterfactual reasoning and adjusts the weights of high-contribution features accordingly. Compared with traditional incremental learning, the model convergence speed is improved by 30%, and the prediction accuracy is significantly enhanced, achieving a deep integration of interpretability and model optimization.

[0025] 6. Five-layer closed-loop collaborative control produces a "1+1>2" technical effect. The five modules of this invention—confidence assessment, hierarchical approval, deviation monitoring, threshold adaptation, and model self-optimization—are not simply connected in series, but form a deeply collaborative closed-loop control system: Confidence decomposition provides a refined basis for the "source of uncertainty" in hierarchical approval, enabling review tasks to be accurately assigned to the most suitable approvers; the actual results of hierarchical approval (approval / modification / rejection) provide real data for deviation monitoring, forming a deviation sample library; the samples captured by deviation monitoring undergo counterfactual root cause analysis to locate the causal contribution of specific features and output an interpretable root cause report; the root cause contribution is used for targeted incremental learning (the Fisher information matrix value of high-contribution features in the EWC loss function is reduced), making model optimization more targeted; after model optimization, the confidence distribution changes, and the threshold adaptation module (NSGA-II Pareto optimization) dynamically adjusts the approval criteria accordingly, ensuring that the system continuously operates at the optimal balance between efficiency and accuracy. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the described drawings are only a part of the embodiments of the present invention, and not all of them. Based on the drawings in the present invention, those skilled in the art can obtain other related drawings without creative effort.

[0027] Figure 1 This is a schematic diagram of the structure of a work hour management system with AI confidence assessment and adaptive approval according to the present invention; Figure 2 This is a flowchart illustrating a time management method based on AI confidence assessment and adaptive approval according to the present invention. Detailed Implementation

[0028] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of this application. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0029] As attached Figure 1As shown, this invention provides a work hour management system based on AI confidence assessment and adaptive approval. The system includes a data acquisition module, a confidence assessment module, a hierarchical approval module, a deviation monitoring module, a threshold adaptive module, and a model self-optimization module. The following provides a detailed description of each module:

[0030] 1. Data Acquisition Module The data acquisition module interfaces with the requirements management platform and the code hosting repository, automatically collecting requirement texts and code commit records associated with the target task. This module includes a requirement data acquisition unit, a code behavior collection unit, and a data normalization unit. (1) Demand data acquisition unit, used to connect to the demand management platform via API to acquire demand data; the demand data includes: demand ID, title, description, demand type, creation time, and handler information; (2) Code behavior collection unit, used to connect to the code hosting repository via Webhook or API to obtain code behavior data in the code submission record; the code behavior data includes: submitter, submission time, list of modified files, number of modified lines, and commit message information; (3) Data normalization unit, used to unify the format of the requirement data and the code behavior data, deduplicate them, and establish a unique association mapping between the requirement and the code submission based on the requirement ID in the commit message or the association model trained based on historical data.

[0031] 2. Confidence Assessment Module The confidence assessment module is used to input the requirement text and code submission record into a pre-trained work hour prediction model, and output the predicted work hours and the corresponding confidence score; wherein, the confidence score is calculated based on the uncertainty measure output by the work hour prediction model, and is used to characterize the reliability of the prediction result; this module includes a feature extraction unit, a prediction unit, and a confidence score estimation unit, wherein, (1) Feature extraction unit, used to extract the semantic feature vector of the requirement text through a pre-trained language model, extract the code structure feature vector of the code submission record through abstract syntax tree parsing, and fuse the semantic feature vector with the code structure feature vector; (2) Prediction unit, used to output predicted working time based on the fused feature vector; (3) Confidence estimation unit, used to calculate the variance of the prediction distribution through multiple forward propagations using the Monte Carlo Dropout method or deep ensemble learning method, and determine the confidence level based on the variance; the confidence estimation unit is also used to decompose the confidence level into a demand uncertainty component, a code uncertainty component and a historical uncertainty component, wherein the demand uncertainty component is calculated based on the semantic perplexity and information entropy of the demand text; the code uncertainty component is calculated based on the coefficient of variation of the cyclomatic complexity distribution of the code submission record and the dispersion of the number of modified lines; the historical uncertainty component is calculated based on the standard deviation of the prediction error distribution of similar historical tasks.

[0032] 3. Tiered Approval Module The tiered approval module executes different approval strategies based on the comparison results between the confidence level and preset first and second thresholds, where the first threshold is greater than the second threshold: if the confidence level is greater than or equal to the first threshold, approval is automatically granted, and approved work hour data is generated; if the confidence level is between the second and first thresholds, a manual review process is triggered, and the reason for low confidence is marked on the review interface; if the confidence level is less than or equal to the second threshold, approval is automatically rejected, and the reason for rejection is recorded. This module includes a review management unit, used to dynamically select the review approver based on the type of dominant uncertainty component when a manual review is triggered. (1) The dominant uncertainty component is the component with the largest value among the demand uncertainty component, code uncertainty component and historical uncertainty component; (2) If the dominant uncertainty component is the demand uncertainty component, select a demand analyst as the reviewer and approver; (3) If the type of the dominant uncertainty component is a code uncertainty component, select the technical person in charge as the reviewer and approver; (4) If the dominant uncertainty component is a historical uncertainty component, select the project manager as the reviewer and approver.

[0033] 4. Deviation Monitoring Module The deviation monitoring module is used to collect actual working hour data after approval and calculate the deviation between the actual working hour data and the predicted working hour; this module includes a deviation calculation unit, a threshold comparison unit, and a sample storage unit; wherein... (1) Deviation calculation unit, used to collect actual working hours data after approval, and calculate the relative deviation δ = |actual working hours - predicted working hours| / max(actual working hours, predicted working hours); (2) Threshold comparison unit, used to compare the relative deviation with a preset deviation threshold to determine whether to trigger sample collection; (3) Sample entry unit, used to store the feature data, predicted working hours, actual working hours, deviation, confidence level and timestamp of the target task into the incremental sample library when the relative deviation exceeds the preset deviation threshold.

[0034] 5. Threshold Adaptive Module The threshold adaptive module is used to dynamically adjust the first and second thresholds using a Pareto optimization algorithm, with the dual objectives of maximizing the automatic pass rate and maximizing the review accuracy. This module includes a performance statistics unit, a Pareto optimization unit, and a threshold update unit. (1) Performance statistics unit, used to periodically calculate the automatic pass rate and review accuracy rate corresponding to each confidence interval within a historical time period; (2) Pareto optimization unit, used to perform Pareto optimal threshold search with the non-dominated sorting genetic algorithm NSGA-II as the dual objective of maximizing automatic pass rate and maximizing verification accuracy, and output the optimal threshold combination on the Pareto front. (3) Threshold update unit, used to select a set of thresholds from the Pareto frontier according to the current business preference and update the first threshold and the second threshold.

[0035] 6. Model self-optimization module The model self-optimization module is used to trigger incremental updates of the work time prediction model by using the feature data of the target task and the actual work time data as incremental training samples when the deviation exceeds a preset deviation threshold. This module includes an incremental sample caching unit, a directional weight adjustment unit, and an incremental training unit. (1) Incremental sample caching unit, used to store samples whose deviation exceeds a preset deviation threshold into the incremental sample library; (2) Targeted weight adjustment unit, used to amplify the weight adjustment range of the corresponding features in the time prediction model according to the counterfactual root cause analysis results; the counterfactual root cause analysis is achieved by calculating the causal contribution of each feature to the deviation; (3) Incremental training unit, used to perform incremental updates using the elastic weight consolidation algorithm, with the loss function being: L(θ)= L_new(θ) + λ / 2 · Σ_i (F_i / η_i) (θ_i - θ_i^*)^2, where η_i is the normalized value of the root cause contribution of feature i.

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0037] Example: Time Management in the Development of the "Cross-Project Requirement Association" Function in a Project Management System

[0038] This embodiment uses the development task of the "cross-project requirement association" function of a project management system as an example to illustrate the complete workflow of the system.

[0039] 1. Data Collection Developer Zhang San was responsible for developing requirement PM-2024-001, "Implement cross-project requirement association functionality," which includes subtask TASK-001, "Develop a requirement association engine." The system collected requirement information via the Jira API: the requirement title was "Cross-project requirement association functionality," and the description was "Support the establishment of requirement dependencies between different projects, including three types: pre-dependencies, post-dependencies, and parallel dependencies. It needs to implement functions such as visualization of association relationships, impact scope analysis, and change notifications." Code was captured via GitLab Webhook and committed to Commit-abc123, with the commit message "feat:#TASK-001 Completed the core logic of the requirement association engine." Modified files included RequirementLinkEngine.java, LinkGraphService.java, and DependencyAnalyzer.java, with +234 / -45 lines changed, involving code changes to three core modules.

[0040] 2. Confidence assessment The system inputs the requirement text and code submission records into the work-hour prediction model, outputting a predicted work-hour of 8.5 hours. Monte Carlo Dropout calculation yields a confidence score of c=0.68. Specifically, the requirement uncertainty component c_demand=0.82 (clear requirement description, well-defined functionalities), the code uncertainty component c_code=0.55 (high code complexity, involving interaction logic across multiple modules), and the historical uncertainty component c_history=0.72 (sufficient samples of similar functional tasks). The overall confidence score after fusion is 0.68, with the code uncertainty component being the dominant uncertainty component.

[0041] 3. Tiered approval The value of c=0.68 falls between T2=0.55 and T1=0.85, triggering manual review. Based on the dominant uncertainty component being code uncertainty, the system assigns the review task to the technical lead, Li Si. Li Si sees a predicted working time of 8.5 hours on the review interface. Considering the high technical complexity due to the interaction logic involving multiple modules, he modifies the working time to 12 hours before confirming approval.

[0042] 4. Deviation monitoring After Zhang San actually completed the task, he reported the actual working hours as 11 hours. The system calculated the deviation δ = |11-8.5| / 11 = 0.227 > 0.2, and the sample was stored in the incremental sample library.

[0043] 5. Threshold Adaptation The system updates the Pareto front monthly using the NSGA-II algorithm. Currently, at the front, Option A (T1=0.82, T2=0.52) has an automatic approval rate of 73% and a review accuracy rate of 93%; Option B (T1=0.88, T2=0.58) has an automatic approval rate of 63% and a review accuracy rate of 95%. Considering the tight project delivery schedule, the project manager selected Option A, and the system automatically updated the approval threshold.

[0044] 6. Model self-optimization When the incremental sample library accumulates to 100 entries, an EWC incremental update is triggered. Counterfactual root cause analysis shows that the "cross-module interaction complexity" feature contributes the most (0.52). This feature has a large η_i, which reduces the Fisher information matrix value in the EWC loss function, allowing for a larger adjustment range. The updated model's prediction accuracy on "cross-project requirement association" tasks improved from 65% to 81%.

[0045] As another embodiment, the present invention provides a time management method for AI confidence assessment and adaptive approval, as shown in the appendix. Figure 2 As shown, the process includes the following steps: S101: Connects to the requirements management platform and code hosting repository to automatically collect requirement texts and code submission records associated with target tasks; S102: Input the demand text and code submission record into the work hour prediction model, and output the predicted work hours and the corresponding confidence level; S103: Execute different approval strategies based on the comparison results between the confidence level and the preset first threshold and second threshold, including automatically approving the approval, triggering manual review and marking the reason for the low confidence level, or automatically rejecting the approval. S104: Collect actual working hour data after approval, and calculate the deviation between the actual working hour data and the predicted working hour; S105: With the dual objectives of maximizing the automatic pass rate and maximizing the review accuracy, the Pareto optimization algorithm is used to dynamically adjust the first threshold and the second threshold. S106: When the deviation exceeds a preset deviation threshold, the feature data of the target task and the actual working time data are used as incremental training samples to trigger the incremental update of the working time prediction model.

[0046] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A work hour management system with AI confidence assessment and adaptive approval, characterized in that, The system includes: a data acquisition module, a confidence assessment module, a tiered approval module, a deviation monitoring module, a threshold adaptive module, and a model self-optimization module; wherein... The data acquisition module is used to connect to the requirement management platform and the code hosting repository to automatically collect requirement texts and code submission records associated with the target task; The confidence assessment module is used to input the requirement text and code submission record into a pre-trained work hour prediction model and output the predicted work hours and the corresponding confidence level; wherein, the confidence level is calculated based on the uncertainty metric output by the work hour prediction model and is used to characterize the reliability of the prediction result; The tiered approval module is used to execute different approval strategies based on the comparison results between the confidence level and preset first and second thresholds, wherein the first threshold is greater than the second threshold. If the confidence level is greater than or equal to the first threshold, the approval is automatically granted and approved work hour data is generated. If the confidence level is between the second threshold and the first threshold, a manual review process is triggered, and the reason for the low confidence level is marked on the review interface. If the confidence level is less than or equal to the second threshold, the approval will be automatically rejected and the reason for rejection will be recorded. The deviation monitoring module is used to collect actual working hour data after approval and calculate the deviation between the actual working hour data and the predicted working hour. The threshold adaptive module is used to dynamically adjust the first threshold and the second threshold using the Pareto optimization algorithm with the dual objectives of maximizing the automatic pass rate and maximizing the review accuracy. The model self-optimization module is used to trigger the incremental update of the working time prediction model by using the feature data of the target task and the actual working time data as incremental training samples when the deviation exceeds a preset deviation threshold.

2. The work hour management system with AI confidence assessment and adaptive approval according to claim 1, characterized in that, The data acquisition module includes a requirement data acquisition unit, a code behavior acquisition unit, and a data normalization unit; wherein... The requirement data acquisition unit is used to connect to the requirement management platform via API to acquire requirement data; the requirement data includes: requirement ID, title, description, requirement type, creation time, and handler information; The code behavior collection unit is used to connect to the code hosting repository via Webhook or API to obtain code behavior data from the code commit records; the code behavior data includes: committer, commit time, list of modified files, number of modified lines, and commit message information. The data normalization unit is used to unify the format and deduplicate the requirement data and the code behavior data, and establish a unique association mapping between requirements and code submissions based on the requirement ID in the commit message or an association model trained based on historical data.

3. The work hour management system with AI confidence assessment and adaptive approval according to claim 1, characterized in that, The confidence assessment module includes a feature extraction unit, a prediction unit, and a confidence estimation unit; wherein... The feature extraction unit is used to extract the semantic feature vector of the requirement text through a pre-trained language model, extract the code structure feature vector of the code submission record through abstract syntax tree parsing, and fuse the semantic feature vector with the code structure feature vector. The prediction unit is used to output the predicted working time based on the fused feature vector; The confidence estimation unit is used to calculate the variance of the predicted distribution through multiple forward propagations using the Monte Carlo Dropout method or a deep ensemble learning method, and to determine the confidence level based on the variance. The confidence estimation unit is further configured to decompose the confidence level into a demand uncertainty component, a code uncertainty component, and a historical uncertainty component, wherein... The demand uncertainty component is calculated based on the semantic perplexity and information entropy of the demand text; The code uncertainty component is calculated based on the coefficient of variation of the cyclomatic complexity distribution of the code commit record and the dispersion of the number of modified lines. The historical uncertainty component is calculated based on the standard deviation of the prediction error distribution of similar historical tasks.

4. The work hour management system with AI confidence assessment and adaptive approval according to claim 1, characterized in that, The hierarchical approval module includes a review management unit, which dynamically selects the review approver based on the type of the dominant uncertainty component when a manual review is triggered. The dominant uncertainty component is the component with the largest value among the requirement uncertainty component, code uncertainty component, and historical uncertainty component; if the type of the dominant uncertainty component is the requirement uncertainty component, the requirement analyst is selected as the reviewer and approver; if the type of the dominant uncertainty component is the code uncertainty component, the technical lead is selected as the reviewer and approver; if the type of the dominant uncertainty component is the historical uncertainty component, the project manager is selected as the reviewer and approver.

5. The work hour management system with AI confidence assessment and adaptive approval according to claim 1, characterized in that, The deviation monitoring module includes a deviation calculation unit, a threshold comparison unit, and a sample storage unit; wherein... The deviation calculation unit is used to collect actual working hour data after approval and calculate the relative deviation δ = |actual working hours - predicted working hours| / max(actual working hours, predicted working hours); The threshold comparison unit is used to compare the relative deviation with a preset deviation threshold to determine whether to trigger sample collection. The sample entry unit is used to store the feature data, predicted working hours, actual working hours, deviation, confidence level, and timestamp of the target task into the incremental sample library when the relative deviation exceeds a preset deviation threshold.

6. The work hour management system with AI confidence assessment and adaptive approval according to claim 1, characterized in that, The threshold adaptive module includes a performance statistics unit, a Pareto optimization unit, and a threshold update unit; wherein... The performance statistics unit is used to periodically calculate the automatic pass rate and review accuracy rate corresponding to each confidence interval within a historical time period. The Pareto optimization unit is used to perform Pareto optimal threshold search with the non-dominated sorting genetic algorithm NSGA-II as the dual objective of maximizing automatic pass rate and maximizing verification accuracy, and output the optimal threshold combination on the Pareto front. The threshold update unit is used to select a set of thresholds from the Pareto front according to the current business preferences and update the first threshold and the second threshold.

7. The work hour management system with AI confidence assessment and adaptive approval according to claim 1, characterized in that, The model self-optimization module includes an incremental sample caching unit, a directional weight adjustment unit, and an incremental training unit; wherein... The incremental sample caching unit is used to store samples whose deviation exceeds a preset deviation threshold into the incremental sample library. The directional weight adjustment unit is used to amplify the weight adjustment range of the corresponding features in the time prediction model based on the counterfactual root cause analysis results; the counterfactual root cause analysis is achieved by calculating the causal contribution of each feature to the deviation. The incremental training unit is used to perform incremental updates using the elastic weight consolidation algorithm, with the loss function being: L(θ) = L_new(θ) + λ / 2 · Σ_i (F_i / η_i) (θ_i - θ_i^*)^2, where η_i is the normalized value of the root cause contribution of feature i.

8. A time management method based on AI confidence assessment and adaptive approval, applied to the system described in any one of claims 1-7, characterized in that, Includes the following steps: Step 1: Connect the requirements management platform and the code hosting repository to automatically collect the requirement text and code submission records associated with the target task; Step 2: Input the required text and code submission record into the work hour prediction model, and output the predicted work hours and the corresponding confidence level; Step 3: Execute different approval strategies based on the comparison results between the confidence level and the preset first threshold and second threshold, including automatically approving the approval, triggering manual review and marking the reason for the low confidence level, or automatically rejecting the approval. Step 4: Collect actual working hour data after approval, and calculate the deviation between the actual working hour data and the predicted working hours; Step 5: With the dual objectives of maximizing the automatic pass rate and maximizing the review accuracy, the Pareto optimization algorithm is used to dynamically adjust the first threshold and the second threshold. Step 6: When the deviation exceeds the preset deviation threshold, the feature data of the target task and the actual working time data are used as incremental training samples to trigger the incremental update of the working time prediction model.

Images