AI-based project management method for operational risk control

Abstract

The application relates to the technical field of project management, and discloses an AI-based business risk control method for project management, which comprises the following steps: acquiring an AI shallow semantic recognition result of a project text, constructing a plurality of evidence source credibility distributions according to the recognition result, fusing the plurality of evidence sources according to evidence theory to generate a credibility interval reflecting a risk degree, converting the credibility interval into an interval value, performing interval-type cascade propagation in a project task dependency network, calculating a total risk exposure interval, constructing a robust optimization model based on the total risk exposure interval, solving a risk mitigation strategy, dynamically adjusting a monitoring threshold according to a comparison between a project actual execution deviation and the total risk exposure interval, and executing hierarchical intervention and rolling re-optimization when a trigger condition is met. The application automatically extracts multi-dimensional features such as frequency, emotion, novelty and contradiction from a project text by using AI shallow analysis, thereby greatly improving timeliness.

Description

Technical Field

[0001] This application relates to the field of project management technology, and in particular to AI-based operational risk control methods for project management. Background Technology

[0002] Operational risk control in project management is a core management activity to ensure that projects are delivered on time, within budget, and with quality. Traditional project management risk control methods, such as the risk register method based on the project management knowledge system, Monte Carlo simulation, and decision tree analysis, mainly rely on the project manager's subjective experience to assess the probability and impact of identified risks and formulate corresponding response strategies.

[0003] Current risk identification relies heavily on manual methods. The massive amounts of unstructured text information generated by project team members through meetings, emails, weekly reports, etc., contain numerous risk signs. However, traditional methods cannot automatically extract and quantify risk signals from these texts, resulting in the omission or delay in addressing many early risk signs. Existing methods typically treat risk probabilities as precise numerical values ​​for calculation. However, risk information in project management documents is often ambiguous, incomplete, and multi-sourced. The assumption of precise probabilities is seriously inconsistent with reality, leading to distorted risk assessment results. Therefore, how to solve the fundamental problems of relying on manual risk signal extraction and insufficient uncertainty handling capabilities in project management is the problem that this invention aims to address. Summary of the Invention

[0004] To address the shortcomings of manual risk signal extraction and uncertainty handling capabilities in project management, this application provides an AI-based operational risk control method for project management, employing the following technical solution: AI-based project management methods for operational risk control include: Obtain the shallow semantic recognition results of the project text by AI, and construct the confidence allocation of multiple evidence sources based on the recognition results; By integrating multiple sources of evidence using evidence theory, a confidence interval reflecting the degree of risk is generated; The confidence interval is converted into an interval value, and interval-type cascading propagation is performed in the project task dependency network to calculate the total risk exposure interval. A robust optimization model is constructed based on the total risk exposure range to solve for risk mitigation strategies. Based on the comparison between the actual deviation of the project execution and the total risk exposure range, the monitoring threshold is dynamically adjusted, and tiered intervention and rolling re-optimization are performed when the trigger conditions are met.

[0005] Optionally, the process of constructing the reliability allocation for multiple sources of evidence includes: Based on the frequency of mentions in the identification results, confidence levels are assigned to high-risk states and complementary confidence levels are assigned to low-risk states according to the exponential decay law. Based on the emotion polarity values ​​in the identification results, linear reliability assignments are performed for high-risk and low-risk states. Based on the information novelty coefficient and the number of semantic contradictions in the recognition results, a confidence level is assigned to the unknown state; wherein, the confidence level of the unknown state is positively correlated with the exponential decay value of the number of contradictions and positively correlated with the novelty coefficient.

[0006] Optionally, the evidence theory fusion adopts Dempster's synthesis rule, and its normalization factor is determined by the sum of the confidence products of all evidence sources with no intersection. The confidence interval includes the confidence function value and the likelihood function value, wherein the likelihood function value is equal to the sum of the confidence function value and the confidence of the unknown state; The comprehensive risk probability is calculated based on the confidence interval, and the comprehensive risk probability is equal to the sum of the confidence function value and the confidence of half of the unknown states; The width of the confidence interval is the difference between the likelihood function value and the confidence function value.

[0007] Optionally, the interval-type cascade propagation process includes: Multiply the confidence function value and likelihood function value in the confidence interval by the preset impact benchmark value of the risk event to obtain the lower and upper bounds of the risk impact intensity interval. The correlation coefficients between project tasks are adjusted up and down by a preset ratio to construct an interval adjacency matrix; Find the lower and upper bounds of the total risk exposure interval after cascading propagation; The system risk interval entropy is constructed based on each component of the total risk exposure interval. The upper and lower bounds of the interval entropy are obtained by weighted logarithmic summation of the total exposure components. When the width of the system risk interval entropy exceeds a preset threshold, a manual review instruction is triggered.

[0008] Optionally, the robust optimization model takes minimizing the auxiliary variable as the optimization objective; The constraints of the robust optimization model include: The sum of the positive values ​​of the upper bound of the total risk exposure range for all risk events minus the minimum efficiency reduction of the selected mitigation measures, plus the implementation cost of the mitigation measures, shall not exceed the auxiliary variable; and the total cost of the selected mitigation measures shall not exceed the preset budget limit. When the conflict coefficient generated during evidence fusion exceeds the preset conflict threshold, a slack variable is introduced into the constraint. The product of the conflict coefficient and the slack variable is added to the constraint as a penalty term, and at least one mitigation measure is required to be selected.

[0009] Optionally, dynamically adjusting the monitoring threshold specifically includes: Construct a conflict-weighted comprehensive index, which is the weighted sum of the ratio of the actual losses incurred in the project to the budget and the ratio of the upper and lower bounds of the total risk exposure range; The index-weighted moving average statistic is calculated recursively based on the conflict-weighted composite index. The primary control upper limit and the secondary control upper limit are determined based on the statistical mean, standard deviation, and conflict coefficient; the primary control upper limit is greater than the secondary control upper limit.

[0010] Optionally, the method further includes a historical evidence feedback correction step: The actual loss deviation of the project in the previous cycle is converted into the basic confidence allocation for the high-risk state and the basic confidence allocation for the unknown state; In the next cycle of evidence fusion, the historical evidence sources will be combined with the multiple evidence sources constructed in the current period again using Dempster synthesis to achieve iterative correction of the confidence interval based on historical facts.

[0011] Optionally, the identification results include at least: a unique identifier of the risk event, the frequency of mention of the risk event in the text, the sentiment polarity value, the information novelty coefficient, and the number of semantic contradictions.

[0012] Optionally, the process of hierarchical intervention and rolling re-optimization includes: When the exponentially weighted moving average statistic exceeds the secondary control limit but does not exceed the primary control limit, an early warning signal is issued and the computational granularity or sampling density of the robust optimization model is increased. When the statistic exceeds the first-level control limit, the project change request is frozen, the robust optimization model solution is forcibly executed, and the resulting mitigation strategy is implemented. When the confidence interval width exceeds the preset ambiguity threshold, a semantic ambiguity alarm is triggered and the automatic optimization process is paused. The evidence fusion step is then re-executed after manual supplementation of textual evidence.

[0013] In summary, this application includes at least one of the following beneficial technical effects: This invention utilizes AI shallow parsing to automatically extract multi-dimensional features such as frequency, sentiment, novelty, and contradiction from project texts, significantly improving timeliness; it employs evidence theory to fuse multi-source evidence, avoiding subjective prior settings, and the conflict coefficient can quantify the degree of evidence inconsistency, enhancing the ability to handle ambiguous information; it compensates for the shortcomings of traditional methods that ignore propagation paths by leveraging the amplification effect in task-dependent networks; it ensures that mitigation strategies remain effective even under worst-case parameters, and introduces a conflict penalty mechanism to force the model to become conservative when information is unreliable; by transforming actual losses into evidence sources and fusing them again, the model possesses self-learning capabilities, continuously improving the accuracy of risk assessment. Attached Figure Description

[0014] Figure 1 This is a flowchart of the AI-based project management risk control method in this invention. Detailed Implementation

[0015] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0016] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0017] Please see Figure 1 This application discloses an AI-based method for controlling operational risks in project management, comprising: 01: Obtain the shallow semantic recognition results of the project text by AI, and construct the reliability allocation of multiple evidence sources based on the recognition results; 02: Integrate multiple sources of evidence using evidence theory to generate a confidence interval that reflects the degree of risk; 03: Convert the confidence interval into an interval value, perform interval-type cascading propagation in the project task dependency network, and calculate the total risk exposure interval; 04: Construct a robust optimization model based on the total risk exposure range and solve for risk mitigation strategies; 05: Based on the comparison between the actual execution deviation of the project and the total risk exposure range, dynamically adjust the monitoring threshold, and perform tiered intervention and rolling re-optimization when the trigger conditions are met.

[0018] This invention utilizes AI shallow parsing to automatically extract multi-dimensional features such as frequency, sentiment, novelty, and contradiction from project texts, significantly improving timeliness; it employs evidence theory to fuse multi-source evidence, avoiding subjective prior settings, and the conflict coefficient can quantify the degree of evidence inconsistency, enhancing the ability to handle ambiguous information; it compensates for the shortcomings of traditional methods that ignore propagation paths by leveraging the amplification effect in task-dependent networks; it ensures that mitigation strategies remain effective even under worst-case parameters, and introduces a conflict penalty mechanism to force the model to become conservative when information is unreliable; by transforming actual losses into evidence sources and fusing them again, the model possesses self-learning capabilities, continuously improving the accuracy of risk assessment.

[0019] The process of obtaining the shallow semantic recognition results of the project text is achieved by using an artificial intelligence semantic recognition model to perform entity recognition, sentiment analysis and coreference resolution on the project text, and outputting structured risk tuples, which include at least: a unique identifier of the risk event, the frequency of mention of the risk event in the text f, the sentiment polarity value s, the information novelty coefficient v, and the number of semantic contradictions c.

[0020] The process of constructing a reliability allocation for multiple sources of evidence includes: Based on the mention frequency *f* in the identification results, reliability is assigned to high-risk states and complementary reliability to low-risk states according to the exponential decay law; the reliability of high-risk states based on mention frequency is... The reliability of low-risk states mentioned frequently is The reliability of the unknown state in terms of frequency of mention is .

[0021] Based on the emotion polarity value 's' in the identification results, linear reliability assignments are performed for high-risk and low-risk states; the reliability for the high-risk state based on the emotion polarity value is... The low-risk state reliability of the emotional polarity value is The unknown-state reliability of the emotional polarity value is .

[0022] Based on the information novelty coefficient v and the number of semantic contradictions c in the identification results, a confidence level is assigned to the unknown state to reflect insufficient information reliability. The confidence level of the unknown state with novel contradictions is positively correlated with the exponential decay value of the number of contradictions and also positively correlated with the novelty coefficient, as expressed as: The reliability of novel and contradictory high-risk states is The reliability of novel and contradictory low-risk states is .

[0023] Then, the three sources of evidence , , Using Dempster's composition rule, the normalization factor is the sum of the confidence products of all evidence sources where the intersection is empty. Determined; the normalization factor is expressed as: ,in, Here is the conflict coefficient; the global basic probability allocation is: Where B represents the source of evidence. Any subset that can be taken, where C represents the source of evidence. Any subset that can be taken, where D represents the source of evidence. Any subset that can be taken, namely {H}, {L} and .

[0024] According to the fusion For high-risk state H, construct a confidence interval. Including trust function value with likelihood function value Trust function value , The likelihood function value is equal to the trust function value. Reliability with unknown states The sum of .

[0025] In order to use the confidence intervals for subsequent numerical calculations, the overall risk probability is calculated based on the confidence intervals. The overall risk probability is equal to the sum of the trust function value and the confidence level of half the unknown states, i.e. This composite risk probability is only used to convert confidence intervals into point estimates required for subsequent interval propagation, and it does not change the basic probability distribution of evidence fusion.

[0026] The width of the confidence interval is the difference between the likelihood function value and the confidence function value. This is used to quantify the degree of ambiguity in semantic recognition results.

[0027] It should be noted that, The value and same, This represents the level of ignorance regarding the evidence. This represents the hesitation zone at the decision-making level, hence the difference between the two.

[0028] The interval-type cascading propagation process includes: Multiplying the confidence function value and likelihood function value in the confidence interval by the pre-defined impact benchmark value of the risk event, respectively, yields the lower and upper bounds of the risk impact intensity interval; risk impact intensity interval The pre-set impact benchmark value of the risk event This represents the absolute impact benchmark of the risk event if it occurs; that is, the maximum potential loss that the risk will cause to the project's operating objectives (cost or schedule) without considering the probability of occurrence or uncertainty.

[0029] The correlation coefficients between project tasks are adjusted upwards and downwards by a preset percentage to construct an interval adjacency matrix. Its elements are represented as: ,in, The floating preset ratio is set according to the standardized settings of historical data; in this embodiment, it is set to 10%. This indicates the strength of the dependency of task k on task j. The Pearson correlation coefficient between two project tasks is determined by reading historical actual execution data of completed work packages in the current project. For task pairs (j,k) with dependencies, the actual schedule deviation rate of these two tasks is retrieved over several rolling cycles (e.g., the past 4-6 weeks), and the Pearson correlation coefficient between the two deviation sequences is calculated. , .

[0030] Solving for the lower and upper bounds of the total risk exposure interval after cascading propagation is achieved using the Leontief inverse matrix, through the formula... Total risk exposure range after cascading propagation Where I is the n×n standard identity matrix, For the interval adjacency matrix ( for (elements) A column vector of intervals of n×1 ( for (element), t represents the t-th monitoring period, and thus obtain The upper realm and the lower realm N is the number of samples.

[0031] This scheme uses the interval Leontief inverse matrix to quantify the amplification effect of risk in task-dependent networks, thus overcoming the shortcomings of traditional methods that ignore the propagation path.

[0032] Next, a system risk interval entropy is constructed based on each component of the total risk exposure interval to assess the degree of risk diffusion throughout the network. The upper and lower bounds of this interval entropy are obtained by weighted logarithmic summation of the total exposure components. Therefore, the system risk interval entropy is expressed as: The width of the system risk interval entropy is: When the width of the system risk interval entropy exceeds a preset threshold, it means that the predicted exposure fluctuates greatly, and the risk distribution pattern is extremely uncertain, triggering a manual review instruction. The preset threshold is determined based on the project task network size, which is characterized by the total number of nodes *u*. The preset threshold value is then set to... .

[0033] Then, the robust optimization model is used to minimize the auxiliary variables. To optimize the objective, it is represented as: The constraints of the robust optimization model include: The sum of the positive values ​​of the total risk exposure interval for all risk events minus the minimum efficiency reduction of the selected mitigation measures, plus the implementation cost of the mitigation measures, shall not exceed the auxiliary variable. Furthermore, when the conflict coefficient generated during evidence fusion exceeds a preset conflict threshold, a slack variable is introduced into the constraints. The product of the conflict coefficient and the slack variable is added as a penalty term to the constraints, and at least one mitigation measure must be selected. in, For optional mitigation measures, m = 1, ..., M. , To mitigate the costs associated with the measures, Let K be the minimum efficiency reduction amount for the i-th risk event, K be the conflict coefficient, and f be a conditional function. When K is greater than a preset conflict threshold, ... ;otherwise, The preset conflict threshold is set to 0.3 based on the experience of Dempster-Shafer theory. When the conflict coefficient is less than or equal to 0.3, the judgment directions of each source of evidence are basically consistent or there are only slight differences. When the conflict coefficient is greater than 0.3, it indicates that there is a significant contradiction between the evidence. As slack variables, In this plan The initial value is set to 1. When the constraint is activated, the solver will operate within the range of values. It automatically finds the optimal solution.

[0034] The total cost of the selected mitigation measures shall not exceed the preset budget limit B, that is: The optimal mitigation strategy is obtained by solving this mixed-integer linear programming problem using the branch and bound method. .

[0035] This scheme uses robust linear programming based on interval numbers to ensure that the mitigation strategy remains effective even in the worst-case scenario of parameters, and introduces a conflict penalty mechanism to force the model to become conservative when information is unreliable.

[0036] It should be noted that the branch and bound method converges to the optimal solution by enumerating all feasible integer combinations and pruning them. The specific process can be implemented using existing solvers and will not be described in detail here.

[0037] In addition, dynamically adjusting monitoring thresholds specifically includes: Constructing a conflict-weighted comprehensive index The conflict-weighted composite index is the weighted sum of the proportion of actual project losses to the budget and the ratios of the upper and lower bounds of the total risk exposure range, expressed as: in, This represents the actual losses already incurred in the project during the current t-period. This represents the budget for the project in period t. These are weighting coefficients. The weighting coefficient takes different values ​​at different project stages. In this embodiment, the value is 0.2 for the first 20% of the project, 0.5 for 20% to 80% of the project, and 0.80 for 80% to 100% of the project.

[0038] The index-weighted moving average statistic is calculated recursively based on the conflict-weighted composite index. ; in, This is a smoothing coefficient, and based on experience, its value is typically between 0.2 and 0.3. The initial value is 1.

[0039] Based on the statistical mean Standard deviation The conflict coefficient determines the upper limit of the primary control level and the upper limit of the secondary control level; if the upper limit of the primary control level is greater than the upper limit of the secondary control level, the upper limit of the primary control level... Level 2 control upper limit ,in, and Based on historical data estimates, To pre-determine the conflict amplification factor, this scheme uses a value of 1.5. This represents the average conflict coefficient over the most recent three periods.

[0040] This scheme achieves adaptive adjustment of thresholds by combining dynamic control limits with actual deviation feedback, and forms a complete closed loop through hierarchical intervention and rolling re-optimization.

[0041] The process of tiered intervention and rolling re-optimization includes: When the exponentially weighted moving average statistic exceeds the secondary control limit but does not exceed the primary control limit, i.e. Issue early warning signals and increase the computational granularity or sampling density of the robust optimization model; When the statistic exceeds the first-level control limit, i.e. Freeze project change requests, enforce robust optimization model solution and implement the resulting mitigation strategy; When the confidence interval width exceeds the preset ambiguity threshold, the ambiguity threshold is set for the confidence interval width. The set upper limit is used to determine whether the ambiguity of the AI ​​semantic recognition result has exceeded the acceptable range of the mathematical model; in this scheme, the ambiguity threshold is set at 0.6, which is the generally accepted ambiguity threshold in evidence theory; that is... When this happens, a semantic ambiguity alarm is triggered and the automatic optimization process is paused. The evidence fusion step (steps 01-02) is re-executed after manual supplementation of textual evidence.

[0042] The operational risk control method in this application also includes a historical evidence feedback and correction step: Deviation of actual losses from the previous period's projects Converting to basic confidence assignments for high-risk and unknown states, thus obtaining the fourth source of evidence. , , , ,in, To predetermine a normalization constant, in international project management standards, a project cost / schedule deviation rate exceeding 5% (i.e., 0.05) is considered a significant deviation. Therefore, in this plan... The value is 0.05 times the total project budget.

[0043] In the next cycle of evidence fusion, historical evidence sources will be used. The model is then synthesized again using Dempster with multiple evidence sources constructed in the current period to achieve iterative correction of the confidence interval based on historical facts, so that the model gradually converges to the true risk level.

[0044] This approach transforms actual losses into sources of evidence and then re-integrates them, enabling the model to have self-learning capabilities and continuously improve the accuracy of risk assessment.

[0045] Example 1 This embodiment takes an infrastructure construction project as an example. The project text includes weekly reports, meeting minutes and emails. The AI ​​model outputs risk events, and its metadata is: f=5, s=-0.6, v=0.8, c=2, and the impact benchmark value I=1 million yuan.

[0046] The value is 0.5, and the calculation yields... , ; , ; , ; The reliability after global fusion is Confidence interval The width of the confidence interval Overall risk probability .

[0047] Impact intensity range The task network consists of 15 nodes, with a correlation coefficient fluctuating by 10%, and an interval adjacency matrix is ​​constructed. Using the interval Leontief inverse solution, the lower bound of the total open interval is 1.682 million yuan, and the upper bound is 2.054 million yuan. The system interval entropy width is 0.31 < 0.5, requiring no manual verification.

[0048] Optional mitigation measures include: increasing alternative suppliers (cost 150,000, reducing efficiency minimum by 0.4), increasing safety stock (cost 80,000, reducing by 0.3), and adjusting the plan (cost 50,000, reducing by 0.1); budget B = 300,000, conflict coefficient K = 0.367 > 0.3, penalty constraints are activated, and the robust optimization solution is obtained. , , , (RMB 10,000), In this embodiment, due to semantic conflict, the decision tends to be conservative, and an additional cost of RMB 50,000 is added, but the upper bound of the worst loss is reduced from RMB 2,054,000 to RMB 1,234,000, a decrease of about 40%.

[0049] The actual loss ratio is 0.05, and the ratio of the upper and lower bounds of the total exposure range is 1.221. The conflict-weighted composite index is calculated. Through recursion using conflict-weighted comprehensive indicators, Based on the average conflict coefficient of the last three cycles, and with a preset conflict amplification factor of 1.5, the upper limit of the first-level control is calculated. Level 2 control upper limit , The system determines that the current overall performance is under control and does not trigger an alarm.

[0050] Example 2 This embodiment also uses an infrastructure construction project as an example, showing the deviation of the actual loss of the project in the previous cycle. 10,000 yuan, with a preset normalization constant. The value is 500,000 yuan. , ; Source of historical evidence By combining the evidence with the three sources of evidence from the current period, historical correction can be achieved.

[0051] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. An AI-based project management method for operational risk control, characterized in that, include: Obtain the shallow semantic recognition results of the project text by AI, and construct the confidence allocation of multiple evidence sources based on the recognition results; By integrating multiple sources of evidence using evidence theory, a confidence interval reflecting the degree of risk is generated; The confidence interval is converted into an interval value, and interval-type cascading propagation is performed in the project task dependency network to calculate the total risk exposure interval. A robust optimization model is constructed based on the total risk exposure range to solve for risk mitigation strategies. Based on the comparison between the actual deviation of the project execution and the total risk exposure range, the monitoring threshold is dynamically adjusted, and tiered intervention and rolling re-optimization are performed when the trigger conditions are met.

2. The AI-based project management risk control method according to claim 1, characterized in that, The process of constructing a reliability allocation for multiple sources of evidence includes: Based on the frequency of mentions in the identification results, confidence levels are assigned to high-risk states and complementary confidence levels are assigned to low-risk states according to the exponential decay law. Based on the emotion polarity values ​​in the identification results, linear reliability assignments are performed for high-risk and low-risk states. Based on the information novelty coefficient and the number of semantic contradictions in the recognition results, a confidence level is assigned to the unknown state; wherein, the confidence level of the unknown state is positively correlated with the exponential decay value of the number of contradictions and positively correlated with the novelty coefficient.

3. The AI-based project management risk control method according to claim 1, characterized in that, The evidence theory fusion adopts Dempster's synthesis rule, and its normalization factor is determined by the sum of the confidence products of all evidence sources where the intersection is empty; The confidence interval includes the confidence function value and the likelihood function value, wherein the likelihood function value is equal to the sum of the confidence function value and the confidence of the unknown state; The comprehensive risk probability is calculated based on the confidence interval, and the comprehensive risk probability is equal to the sum of the confidence function value and the confidence of half of the unknown states; The width of the confidence interval is the difference between the likelihood function value and the confidence function value.

4. The AI-based project management risk control method according to claim 1, characterized in that, The interval-type cascaded propagation process includes: Multiply the confidence function value and likelihood function value in the confidence interval by the preset impact benchmark value of the risk event to obtain the lower and upper bounds of the risk impact intensity interval. The correlation coefficients between project tasks are adjusted up and down by a preset ratio to construct an interval adjacency matrix; Find the lower and upper bounds of the total risk exposure interval after cascading propagation; The system risk interval entropy is constructed based on each component of the total risk exposure interval. The upper and lower bounds of the interval entropy are obtained by weighted logarithmic summation of the total exposure components. When the width of the system risk interval entropy exceeds a preset threshold, a manual review instruction is triggered.

5. The AI-based project management risk control method according to claim 1, characterized in that, The robust optimization model aims to minimize the auxiliary variables. The constraints of the robust optimization model include: The sum of the positive values ​​of the upper bound of the total risk exposure range for all risk events minus the minimum efficiency reduction of the selected mitigation measures, plus the implementation cost of the mitigation measures, shall not exceed the auxiliary variable; and the total cost of the selected mitigation measures shall not exceed the preset budget limit. When the conflict coefficient generated during evidence fusion exceeds the preset conflict threshold, a slack variable is introduced into the constraint. The product of the conflict coefficient and the slack variable is added to the constraint as a penalty term, and at least one mitigation measure is required to be selected.

6. The AI-based project management risk control method according to claim 1, characterized in that, Dynamically adjusting monitoring thresholds specifically includes: Construct a conflict-weighted comprehensive index, which is the weighted sum of the ratio of the actual losses incurred in the project to the budget and the ratio of the upper and lower bounds of the total risk exposure range; The index-weighted moving average statistic is calculated recursively based on the conflict-weighted composite index. The primary control upper limit and the secondary control upper limit are determined based on the statistical mean, standard deviation, and conflict coefficient; the primary control upper limit is greater than the secondary control upper limit.

7. The AI-based project management risk control method according to claim 1, characterized in that, The method also includes a historical evidence feedback and correction step: The actual loss deviation of the project in the previous cycle is converted into the basic confidence allocation for the high-risk state and the basic confidence allocation for the unknown state; In the next cycle of evidence fusion, the historical evidence sources will be combined with the multiple evidence sources constructed in the current period again using Dempster synthesis to achieve iterative correction of the confidence interval based on historical facts.

8. The AI-based project management risk control method according to claim 1, characterized in that, The identification results include at least: the unique identifier of the risk event, the frequency of mention of the risk event in the text, the sentiment polarity value, the information novelty coefficient, and the number of semantic contradictions.

9. The AI-based project management risk control method according to claim 1, characterized in that, The process of hierarchical intervention and rolling re-optimization includes: When the exponentially weighted moving average statistic exceeds the secondary control limit but does not exceed the primary control limit, an early warning signal is issued and the computational granularity or sampling density of the robust optimization model is increased. When the statistic exceeds the first-level control limit, the project change request is frozen, the robust optimization model solution is forcibly executed, and the resulting mitigation strategy is implemented. When the confidence interval width exceeds the preset ambiguity threshold, a semantic ambiguity alarm is triggered and the automatic optimization process is paused. The evidence fusion step is then re-executed after manual supplementation of textual evidence.

Images