A digital government project supervision method and system based on artificial intelligence

Abstract

This invention discloses an artificial intelligence-based method and system for digital government project supervision. The method includes the following steps: first, collecting and preprocessing various supervision data throughout the entire lifecycle of the digital government project; constructing an improved CNN-LSTM risk prediction model and an AHP-fuzzy comprehensive evaluation quality assessment model; inputting a standardized supervision dataset into the two models, combining a source tracing algorithm to perform source analysis of risk points and generate corrective suggestions; employing a dynamic threshold early warning and hierarchical responsibility mechanism for risk early warning and corrective tracking; recording data, analysis results, and corrective action implementation throughout the supervision process to form a supervision file; and optimizing model parameters using a reinforcement learning algorithm. This invention achieves precise and intelligent supervision of the entire lifecycle of digital government projects, significantly improving risk identification capabilities, quality assessment accuracy, and problem correction efficiency.

Description

Technical Field

[0001] This invention relates to the fields of digital government project management and artificial intelligence technology, and in particular to a digital government project supervision method and system based on artificial intelligence. Background Technology

[0002] As the construction of digital government continues to deepen, the scale of various e-government projects is constantly expanding and the technical complexity is gradually increasing, covering multiple core dimensions such as data sharing, business collaboration, and security protection. This places higher demands on the professionalism, real-time performance, and accuracy of project supervision. Currently, digital government project supervision still mainly relies on traditional methods such as manual inspection and document review, which has several prominent pain points: First, supervision efficiency is low. Manual supervision is difficult to achieve parallel management of multiple projects and multiple nodes, and problems such as delayed inspections and omissions are very likely to occur. This is especially true in cross-departmental collaborative projects, where the problem of untimely information transmission is more prominent, directly leading to delayed supervision response. Second, supervision accuracy is insufficient. Supervision judgment relies too much on the experience of staff, which is highly subjective. The identification of various risks such as project quality defects, schedule deviations, and cost overruns is not accurate enough, and misjudgments and omissions occur frequently. Third, there is a lack of full life-cycle closed-loop management capabilities. Existing supervision methods focus more on post-project acceptance and have weak capabilities for dynamic risk warning and real-time correction during project implementation, making it difficult to achieve full-process closed-loop management of "prevention before the event, control during the event, and traceability after the event". Fourth, the phenomenon of data silos is significant. Data from different supervision links and different projects cannot be effectively integrated, making it difficult to form comprehensive and systematic data support, which in turn affects the scientific and rational nature of supervision decisions.

[0003] While some solutions attempt to apply information technology to supervision work, most are limited to simple data recording and online processes, failing to deeply integrate artificial intelligence technology and lacking the ability to intelligently analyze, accurately predict, and adaptively adjust project risks. Furthermore, the model algorithms are relatively simplistic and have not been specifically optimized to suit the unique public welfare and compliance characteristics of digital government projects, thus failing to meet the actual needs of professional and intelligent supervision in digital government projects.

[0004] Therefore, there is an urgent need for a method and system that can break through the limitations of traditional supervision models, deeply integrate artificial intelligence technology, and achieve full life-cycle, precise, and intelligent supervision. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based digital government project supervision method and system, which solves the problems of low efficiency, slow response, insufficient accuracy of risk identification due to reliance on experience, and lack of intelligent prediction and in-process correction capabilities in the existing digital government project supervision. This improves the quality and efficiency of digital government project supervision and ensures the compliant and efficient progress of digital government project construction.

[0006] To achieve the above objectives, this invention provides a digital government project supervision method based on artificial intelligence, comprising the following steps: S1. Collect various types of supervision data throughout the entire lifecycle of digital government projects; S2. Preprocess the collected supervision data to obtain a standardized supervision dataset; S3. Construct an improved CNN-LSTM risk prediction model and an AHP-fuzzy comprehensive evaluation quality assessment model; S4. Input the standardized supervision dataset into the model constructed in S3. Output the project risk level and risk points through the improved CNN-LSTM risk prediction model. Output the project supervision quality score through the AHP-fuzzy comprehensive evaluation quality assessment model. Combine the source tracing algorithm to conduct source tracing analysis on the risk points and generate corrective suggestions. S5. Adopt a dynamic threshold early warning and hierarchical responsibility mechanism to conduct risk early warning and corrective tracking; S6. Record the data, analysis results and corrective actions taken throughout the entire supervision process to form a supervision file, and optimize the model parameters using reinforcement learning algorithms.

[0007] Preferably, in S1, the collected supervision data includes basic project information data, schedule data, quality data, cost data, safety data, compliance data, and collaboration data; Basic project information includes project name, duration, investment, and participating party information; progress data includes completion time of each node, progress deviation, and reasons for node delays; quality data includes code inspection reports, test reports, defect records, and defect rectification status; cost data includes budget execution status, expense details, and cost deviation analysis; security data includes vulnerability scanning reports, security drill records, and security vulnerability rectification records; compliance data includes government data security compliance inspection reports, qualification documents, and compliance review records; and collaboration data includes cross-departmental communication records, problem rectification records, and collaboration node completion status.

[0008] Preferably, in S2, the collected supervision data is cleaned and standardized. Data cleaning includes handling missing values, outliers, and duplicate and noise values. Standardization uses the min-max standardization method to uniformly map all data to the [0,1] interval.

[0009] Preferably, in S3, the improved CNN-LSTM risk prediction model introduces an attention mechanism, including an input layer, a CNN local feature extraction layer, an attention mechanism layer, an LSTM temporal feature extraction layer, a fully connected layer, and an output layer; The convolution operation formula for the local feature extraction layer of a CNN is: ; in, Indicates the first Each convolution outputs features, Represents the ReLU activation function. Indicates the size of the convolution kernel. Indicates the convolution kernel number 1 The weight of each position, Indicates the input data sample number. The first feature dimension and the convolution kernel Data points corresponding to each location, For convolution bias; The weight allocation formula for the attention mechanism layer is: ; in, Indicates the first Attention weights for each local feature. Indicates the first Importance score of each local feature Indicates the number of local features; Indicates the first Importance score of each local feature; The gating operation formulas for the LSTM temporal feature extraction layer include: ; ; ; ; ; ; in, Indicates the output of the forget gate. This represents the sigmoid activation function. This represents the forget gate weight matrix. This represents the output of the hidden layer at the previous time step. This represents the input features at the current time. Indicates the forget gate bias; Indicates the input gate output. This represents the input gate weight matrix. Indicates input gate bias; Indicates the state of candidate cells. Represents the cell state weight matrix. This indicates a cell state bias; This indicates the current state of the cell. This indicates the cell state at the previous moment. Represents element-wise multiplication; Indicates the output gate output. This represents the output gate weight matrix. Indicates output gate bias; This indicates the output of the hidden layer at the current moment.

[0010] Preferably, in S3, the improved AHP-fuzzy comprehensive evaluation quality assessment model includes an indicator system construction unit, a weight determination unit, a fuzzy evaluation unit, and a comprehensive scoring unit; The indicator system construction unit includes an objective layer, a criterion layer, and an indicator layer. The objective layer is the project supervision quality, the criterion layer includes schedule supervision, quality supervision, cost supervision, safety supervision, compliance supervision, and collaborative supervision, and the indicator layer includes 22 supervision indicators. The weight determination unit adopts the improved AHP method. A judgment matrix is ​​constructed using the 1-9 scaling method, and the weights of each index are obtained by solving the eigenvectors using the sum-product method. The normalization formula for the judgment matrix is: ; The formula for calculating the eigenvector is: ; The consistency test formula is: ; in, Represents the judgment matrix of the first... Line number Column elements, Represents the normalized elements; Indicates the first The weight of each indicator, Indicates the number of indicators. For consistency ratio, As a consistency indicator; ; in, This indicates the determination of the largest eigenvalue of the matrix. Represents the average random consistency index; when When <0.1, the judgment matrix has satisfactory consistency; Fuzzy evaluation unit constructs fuzzy relation matrix : ; in, Indicates the first The first indicator for the first The degree of membership of each evaluation level; the evaluation levels are divided into excellent, good, qualified and unqualified; This indicates the row dimension, corresponding to the number of supervision indicators; This indicates the column dimension and the corresponding number of rating levels. The comprehensive scoring unit uses a weighted average fuzzy operator, the formula of which is: ; in, Represents the comprehensive evaluation vector. Represents the indicator weight vector; Final overall score : ; in, Indicates the first Membership degree of each evaluation level Indicates the first The score for each evaluation level.

[0011] Preferably, in S4, the tracing algorithm adopts the cause-effect graph tracing method, taking the risk point as the target node, and tracing all the antecedent factors that lead to the risk, including the data level, process level, and personnel level. Specifically: At the data level, issues include incomplete data collection by supervisors, missing data, data anomalies, untimely handling of duplicate data, non-compliance with data standardization, insecure data storage, data transmission leaks, missing government compliance data, and untimely data updates. At the process level, issues include missing supervision processes at each stage of the project, non-standard process execution, poor cross-departmental collaboration, broken risk warning-correction-acceptance closed-loop process, and unclear defect rectification process. At the personnel level, issues include insufficient professional competence of supervisory personnel, unclear responsibilities of project participants, untimely response from cross-departmental liaison personnel, non-standard operation by contractor's technical personnel, and failure to assign responsibility for rectification. Construct a risk source map to clarify the impact of each preceding factor; corrective recommendations include rectification measures, division of responsibilities, rectification time limits, and acceptance standards.

[0012] Preferably, in S4, the cause-effect graph tracing method quantifies the influence of each antecedent factor using the following formula: ; in, Indicates the first Antecedent factors and risk points The degree of influence; Indicates the existence of the first Antecedent factors At that time, risk points The probability of occurrence; This indicates that the first [number] does not exist. Antecedent factors At that time, risk points The probability of occurrence.

[0013] Preferably, in S5, a dynamic threshold early warning mechanism is adopted. This mechanism combines project type, project stage, and historical supervision data, and uses an improved CNN-LSTM model to adjust the early warning thresholds for each level in real time. Specifically: The mean and standard deviation of the risk level during the implementation phase are calculated using a sliding window algorithm, and are divided into three threshold levels: The first-level threshold is one standard deviation below the average risk level of the current stage; no warning will be issued at this level. The secondary threshold is the mean risk level of the current stage ± 1 standard deviation, at which a yellow warning is issued. The Level 3 threshold is one standard deviation above the average risk level of the current stage, at which a red alert is issued. The notification priority is assigned according to the warning level, and the notification methods include system messages, SMS, email, and pop-up reminders on the government affairs platform. Specifically: Red alerts are immediately sent to all relevant responsible persons, yellow alerts are sent to the supervising engineer and the contractor's technical head, and no alerts are sent only to the supervision file. The rectification and tracking adopts a hierarchical responsibility mechanism. Red alerts are led by the supervision unit and supervised by the client; yellow alerts are led by the contractor and supervised by the supervision unit. If rectification is not completed on time, the alert will be upgraded and a written rectification explanation will be submitted. It will also be included in the project compliance review record. After rectification is completed, it will be subject to dual acceptance by the supervision unit and the client.

[0014] Preferably, in S6, the reinforcement learning algorithm adopts the Q-learning algorithm, and the update formula is: ; in, Representing state Next action of value, Indicates the learning rate; This indicates a reward signal; Indicates the discount factor; Indicates the execution of an action The next state corresponds to the supervision analysis state after the model parameters are adjusted; Representing state The optimal action is as follows. Representing state Next action The updated value, New State Next, select the optimal action. The maximum expected future that can be obtained value.

[0015] This invention also provides an artificial intelligence-based digital government project supervision system, including a data acquisition module, a data preprocessing module, a model building module, an analysis and tracing module, an early warning push module, and an archive management and model optimization module; The data acquisition module includes an interface acquisition unit for calling government system interfaces to collect various types of supervision data, a manual input unit for manually inputting supervision data that cannot be automatically collected, an equipment acquisition unit for collecting equipment operation data during project implementation, and a document parsing unit for parsing project-related documents and extracting supervision data. The data preprocessing module includes a data cleaning unit and a standardization unit. The data cleaning unit is used to handle missing values, outliers, duplicate values, and noise in the supervision data. The standardization unit uses the min-max standardization method to unify the format and units of the cleaned data to obtain a standardized supervision dataset. The model building module includes a risk prediction model building unit and a quality assessment model building unit. The risk prediction model building unit is used to build an improved CNN-LSTM risk prediction model that incorporates an attention mechanism; the quality assessment model building unit is used to build an improved AHP-fuzzy comprehensive evaluation quality assessment model with an optimized index system. The analysis and source tracing module includes a model inference unit, a risk source tracing unit, and a corrective suggestion generation unit. The model inference unit is used to input the standardized supervision dataset into the dual model and output the project risk level, risk points, and supervision quality score. The risk source tracing unit is used to trace the risk points through the entire chain using an improved cause-effect graph source tracing method and quantify the influence of antecedent factors. The corrective suggestion generation unit is used to generate targeted corrective suggestions based on the source tracing results. The early warning push module includes a threshold dynamic adjustment unit, an early warning generation unit, and a push tracking unit. The threshold dynamic adjustment unit adjusts the early warning thresholds for each level in real time using a sliding window algorithm combined with historical supervision data. The early warning generation unit determines the early warning level based on the dynamic thresholds and generates corresponding early warning information. The push tracking unit is used to allocate push priorities according to the early warning level. The document management and model optimization module includes a document generation unit, a document storage unit, and a model optimization unit. The document generation unit is used to record data, analysis results, and corrective actions throughout the entire supervision process, forming standardized supervision documents. The document storage unit is used to encrypt and store supervision documents and perform keyword retrieval. The model optimization unit uses the Q-learning algorithm to optimize the parameters of the two models.

[0016] Therefore, the present invention employs the above-mentioned artificial intelligence-based digital government project supervision method and system, and the beneficial effects are as follows: (1) This invention utilizes an improved AHP-fuzzy comprehensive evaluation model to construct a multi-dimensional and multi-level quality assessment system for digital government projects, effectively overcoming the problems of traditional manual supervision relying on experience, strong subjectivity, and high rates of omission and misjudgment, and greatly improving the scientificity and efficiency of supervision work.

[0017] (2) This invention uses the cause-effect graph tracing method to accurately locate the root cause of risk, and combines dynamic threshold early warning and hierarchical responsibility mechanism to ensure that problems can be discovered and accurately located in a timely manner, thus solving the pain point of traditional supervision lacking process control capabilities.

[0018] (3) This invention breaks down data silos by comprehensively collecting and integrating multi-source heterogeneous data such as project foundation, schedule, quality, cost, safety, compliance and collaboration. On this basis, the Q-learning algorithm is used to continuously optimize the parameters of the core model based on historical supervision files and the effect of corrective actions, so that the entire supervision system has the ability to learn and evolve on its own and can adapt to the supervision needs of different project types and different stages, thus ensuring the compliance and efficient progress of digital government project construction.

[0019] 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

[0020] Figure 1 This is an overall flowchart of an artificial intelligence-based digital government project supervision method according to the present invention; Figure 2 This is an overall structural block diagram of an artificial intelligence-based digital government project supervision system according to the present invention; Figure 3 This is a structural block diagram of the data acquisition module of a digital government project supervision system based on artificial intelligence, according to the present invention. Figure 4 This is a structural block diagram of the data preprocessing module of an artificial intelligence-based digital government project supervision system according to the present invention; Figure 5 This is a block diagram of the model building module structure of an artificial intelligence-based digital government project supervision system according to the present invention; Figure 6 This is a structural block diagram of the analysis and tracing module of an artificial intelligence-based digital government project supervision system according to the present invention; Figure 7 This is a structural block diagram of the early warning push module of an artificial intelligence-based digital government project supervision system according to the present invention; Figure 8 This is a structural block diagram of the file management and model optimization module of an artificial intelligence-based digital government project supervision system according to the present invention; Figure 9This is a performance comparison chart of an embodiment of a digital government project supervision method and system based on artificial intelligence according to the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0023] like Figure 1 As shown, an artificial intelligence-based digital government project supervision method includes the following steps: S1. Collect various types of supervision data throughout the entire lifecycle of digital government projects. The collection scope covers all stages of project initiation, implementation, acceptance, and operation and maintenance. The collection methods include four types: interface calls, manual input, equipment collection, and document parsing, to ensure the comprehensiveness and timeliness of the data.

[0024] The supervision data collected by this invention includes basic project information data, progress data, quality data, cost data, safety data, compliance data, and collaboration data.

[0025] The project's basic information data includes the project name, duration, investment, and participating party information; progress data includes the completion time of each node, progress deviations, and reasons for node delays; quality data includes code inspection reports, test reports, defect records, and defect rectification status; cost data includes budget execution status, expense details, and cost deviation analysis; security data includes vulnerability scanning reports, security drill records, and security vulnerability rectification records; compliance data includes government data security compliance inspection reports, qualification documents, and compliance review records; and collaboration data includes cross-departmental communication records, problem rectification records, and collaboration node completion status.

[0026] S2. Preprocess the collected supervision data to obtain a standardized supervision dataset. Specifically: The collected supervision data underwent data cleaning and standardization. Data cleaning included handling missing values, outliers, and duplicates and noise. For missing values, the mean was used to fill in numerical data, and the mode was used to fill in categorical data. Outlier handling used a 3x3 algorithm. Outliers are identified and removed in principle. Duplicate values ​​are removed using hash verification. Noise is reduced using the moving average method.

[0027] Simultaneously, the min-max standardization method is used for standardization, and the standardization formula is: ; in, This represents the standardized data. Represents the original data. This represents the minimum value of the data in this dimension. This represents the maximum value of the data in this dimension, ensuring that all data are uniformly mapped to the [0,1] interval, thus solving the problem of inconsistent dimensions of data in different dimensions.

[0028] S3. Construct an improved CNN-LSTM risk prediction model and an AHP-fuzzy comprehensive evaluation quality assessment model. The improved CNN-LSTM risk prediction model introduces an attention mechanism to capture the local and temporal features of the supervision data and output the project risk level and risk points; the AHP-fuzzy comprehensive evaluation quality assessment model is used to output the comprehensive score of project supervision quality.

[0029] The CNN-LSTM risk prediction model of this invention includes an input layer, a CNN local feature extraction layer, an attention mechanism layer, an LSTM temporal feature extraction layer, a fully connected layer, and an output layer. The input layer is used to input a standardized supervision dataset; the CNN feature extraction layer is used to extract key local features from the supervision data; the attention mechanism layer is used to assign weights to the extracted local features to highlight key risk features; the LSTM temporal feature extraction layer is used to capture the temporal variation patterns of the supervision data; the fully connected layer is used to integrate local and temporal features; and the output layer is used to output the project risk level and corresponding risk points.

[0030] Specifically, the convolution operation formula for the local feature extraction layer of a CNN is as follows: ; in, Indicates the first Each convolution outputs features; This represents the ReLU activation function, used to solve the vanishing gradient problem; Indicates the size of the convolution kernel. Indicates the convolution kernel number 1 The weight of each position, This is the kernel position index, with values ​​ranging from 1 to... ; Indicates the input data sample number. The first feature dimension and the convolution kernel Data points corresponding to each location, This is the convolution bias.

[0031] The weight allocation formula for the attention mechanism layer is: ; in, Indicates the first Attention weights for each local feature. Indicates the first Importance score of each local feature Indicates the number of local features; Indicates the first Importance scores for each local feature.

[0032] The gating operation formulas for the LSTM temporal feature extraction layer include: ; ; ; ; ; ; in, Indicates the output of the forget gate. This represents the sigmoid activation function. This represents the forget gate weight matrix. This represents the output of the hidden layer at the previous time step. This represents the input features at the current time. Indicates the forget gate bias; Indicates the input gate output. This represents the input gate weight matrix. Indicates input gate bias; Indicates the state of candidate cells. Represents the cell state weight matrix. This indicates a cell state bias; This indicates the current state of the cell. This indicates the cell state at the previous moment. Represents element-wise multiplication; Indicates the output gate output. This represents the output gate weight matrix. Indicates output gate bias; This indicates the output of the hidden layer at the current moment.

[0033] The improved AHP-fuzzy comprehensive evaluation quality assessment model includes an indicator system construction unit, a weight determination unit, a fuzzy evaluation unit, and a comprehensive scoring unit; The indicator system is constructed from three layers: the target layer, the criteria layer, and the indicator layer. The target layer focuses on project supervision quality; the criteria layer includes six dimensions: schedule supervision, quality supervision, cost supervision, safety supervision, compliance supervision, and collaborative supervision; and the indicator layer comprises 22 supervision indicators, namely: 1. Schedule deviation rate; 2. On-time completion rate of nodes; 3. Timely rectification rate of schedule delays; 4. Functional test pass rate; 5. Defect detection rate; 6. Defect rectification qualification rate; 7. Code standard compliance rate; 8. Budget execution rate; 9. Cost deviation rate; 10. Expense compliance rate; 11. High-risk vulnerability detection rate; 12. Timely rectification rate of security vulnerabilities; 13. Security drill pass rate; 14. Compliance rate of government data security; 15. Completeness rate of qualification documents; 16. Compliance review pass rate; 17. Timely response rate of cross-departmental communication; 18. Completion rate of collaborative nodes; 19. Problem rectification closure rate; 20. Timely submission rate of supervision reports; 21. Completeness rate of project document archiving; 22. Satisfaction rating of all parties.

[0034] The weight determination unit adopts the improved AHP method. A judgment matrix is ​​constructed using the 1-9 scaling method, and the weights of each index are obtained by solving the eigenvectors using the sum-product method. The normalization formula for the judgment matrix is: ; The formula for calculating the eigenvector is: ; The consistency test formula is: ; in, Represents the judgment matrix of the first... Line number Column elements, Represents the normalized elements; Indicates the first The weight of each indicator, Indicates the number of indicators. For consistency ratio, As a consistency indicator; ; in, This indicates the determination of the largest eigenvalue of the matrix. Represents the average random consistency index; when When <0.1, the judgment matrix has satisfactory consistency.

[0035] Fuzzy evaluation unit constructs fuzzy relation matrix : ; in, Indicates the first The first indicator for the first The degree of membership of each evaluation level; the evaluation levels are divided into four levels: excellent (90-100 points), good (80-89 points), satisfactory (60-79 points) and unsatisfactory (<60 points); This indicates the row dimension, corresponding to the number of supervision indicators; This indicates the column dimension and the number of corresponding evaluation levels.

[0036] The comprehensive scoring unit uses a weighted average fuzzy operator, the formula of which is: ; in, Represents the comprehensive evaluation vector. This represents the indicator weight vector.

[0037] Final overall score : ; in, Indicates the first The score for each evaluation level, Indicates the first The membership degree of each evaluation level is obtained by weighting and summing the membership degrees of the 22 supervision indicators to that evaluation level according to their respective weights, and directly corresponding one-to-one with the four evaluation levels. The values ​​correspond exactly to the evaluation levels. =1 corresponds to "excellent" =2 corresponds to "good" =3 corresponds to "qualified" =4 corresponds to "unqualified". The closer the value is to 1, the better the project meets the requirements of the corresponding evaluation level. The value is multiplied by the corresponding evaluation level score as a weight and then summed to obtain the comprehensive score of project supervision quality.

[0038] S4. Input the standardized supervision dataset into the model constructed in S3. Output the project risk level and risk points through the improved CNN-LSTM risk prediction model. Output the project supervision quality score through the AHP-fuzzy comprehensive evaluation quality assessment model. Combine the source tracing algorithm to conduct source tracing analysis on the risk points and generate corrective suggestions.

[0039] This invention's source tracing algorithm employs a cause-effect graph approach, using risk points as target nodes to trace all antecedent factors leading to the risk, including data, process, and personnel aspects. Specifically: At the data level, issues include incomplete data collection by supervisors, missing data, data anomalies, untimely processing of duplicate data, substandard data standardization, insecure data storage, data transmission leaks, missing government compliance data, and untimely data updates.

[0040] At the process level, issues include missing supervision processes at each stage of the project, non-standard process execution, poor cross-departmental collaboration, broken risk warning-correction-acceptance closed-loop process, and unclear defect rectification process.

[0041] At the personnel level, issues include insufficient professional competence of supervisory personnel, unclear responsibilities of project participants, untimely response from cross-departmental liaisons, non-standard operation by contractor technical personnel, and failure to assign responsibility for rectification.

[0042] Based on this, a risk source map is constructed to clarify the degree of influence of each antecedent factor. In practical applications, the cause-and-effect diagram source tracing method quantifies the degree of influence of each antecedent factor using the following formula: ; in, Indicates the first Antecedent factors and risk points The degree of influence, with a value range of [-1, 1]. The closer the value is to 1, the greater the impact of this factor on the risk point. The closer to -1, the smaller the impact; There is no effect when the value is 0.

[0043] Indicates the existence of the first Antecedent factors At that time, risk points The probability of occurrence; This indicates that the first [number] does not exist. Antecedent factors At that time, risk points The probability of occurrence.

[0044] After quantitative analysis using this formula, targeted corrective suggestions are generated. These suggestions include rectification measures, division of responsibilities, rectification timelines, and acceptance standards to ensure that they can be implemented.

[0045] S5. This invention employs a dynamic threshold early warning and hierarchical responsibility mechanism to conduct risk early warning and correction tracking.

[0046] A dynamic threshold early warning mechanism is adopted, which combines project type, project stage, and historical supervision data. An improved CNN-LSTM model is used to adjust the early warning thresholds for each level in real time. Specifically: The mean and standard deviation of the risk level during the implementation phase are calculated using a sliding window algorithm, and are divided into three threshold levels: The first-level threshold is one standard deviation below the average risk level of the current stage; no warning will be issued at this level.

[0047] The secondary threshold is the mean risk level of the current stage ± 1 standard deviation, at which a yellow warning is issued.

[0048] The Level 3 threshold is one standard deviation above the average risk level of the current stage, at which a red alert is issued.

[0049] The notification priority is assigned according to the warning level, and the notification methods include system messages, SMS, email, and pop-up reminders on the government affairs platform. Specifically: Red alerts are immediately sent to all relevant responsible persons, yellow alerts are sent to the supervising engineer and the contractor's technical manager, and no alerts are sent only to the supervision file.

[0050] The corrective action tracking adopts a tiered responsibility mechanism to clarify the responsible parties for different warning levels. Red warnings are led by the supervision unit, with the client providing assistance and supervision; yellow warnings are led by the contractor, with the supervision unit providing oversight; if rectification is not completed on time, the warning will be upgraded and a written rectification explanation will be submitted, which will be included in the project compliance review record. After rectification is completed, it will be subject to dual acceptance by the supervision unit and the client.

[0051] S6. Record the data, analysis results and corrective actions taken throughout the entire supervision process to form a supervision file, and optimize the model parameters using reinforcement learning algorithms.

[0052] The reinforcement learning algorithm employs the Q-learning algorithm, using model prediction accuracy, early warning accuracy, and correction effect as reward signals to continuously optimize the parameters of the two models. The update formula is as follows: ; in, Representing state Next action of value; Represents the learning rate, and ; This indicates a reward signal; Represents the discount factor, and ; Indicates the execution of an action The next state corresponds to the supervision analysis state after the model parameters are adjusted; Representing state The optimal action is as follows. Representing state Next action The updated value, New State Next, select the optimal action. The maximum expected future that can be obtained The value represents an estimate of the optimal future reward. Supervision files are stored in encrypted form and support keyword searches by project name, time, risk level, etc., facilitating traceability and querying.

[0053] like Figure 2As shown, an AI-based digital government project supervision system is used to implement the above methods. It includes a data acquisition module, a data preprocessing module, a model building module, an analysis and tracing module, an early warning push module, and an archive management and model optimization module. The modules work together to achieve intelligent supervision of the entire life cycle of digital government projects.

[0054] Among them, such as Figure 3 As shown, the data acquisition module is used to execute step S1, including an interface acquisition unit for calling the government system interface to collect various types of supervision data, a manual input unit for manually inputting supervision data that cannot be automatically collected, an equipment acquisition unit for collecting equipment operation data during project implementation, and a document parsing unit for parsing project-related documents and extracting supervision data. This module supports importing and updating data in multiple formats.

[0055] like Figure 4 As shown, the data preprocessing module is used to perform step S2, including a data cleaning unit and a standardization unit. The data cleaning unit is used to handle missing values, outliers, duplicate values ​​and noise in the supervision data. The standardization unit uses the min-max standardization method to unify the format and units of the cleaned data to obtain a standardized supervision dataset.

[0056] like Figure 5 As shown, the model building module is used to execute step S3, including a risk prediction model building unit and a quality assessment model building unit. The risk prediction model building unit is used to build an improved CNN-LSTM risk prediction model that incorporates an attention mechanism; the quality assessment model building unit is used to build an improved AHP-fuzzy comprehensive evaluation quality assessment model with an optimized index system.

[0057] like Figure 6 As shown, the analysis and tracing module is used to execute step S4, including a model inference unit, a risk tracing unit, and a corrective suggestion generation unit. The model inference unit is used to input the standardized supervision dataset into the dual model and output the project risk level, risk points, and supervision quality score. The risk tracing unit is used to use the improved cause-effect graph tracing method to trace the risk points through the entire chain and quantify the influence of antecedent factors. The corrective suggestion generation unit is used to generate targeted corrective suggestions based on the tracing results.

[0058] like Figure 7 As shown, the early warning push module is used to execute step S5, including a threshold dynamic adjustment unit, an early warning generation unit, and a push tracking unit. The threshold dynamic adjustment unit adjusts the early warning thresholds of each level in real time by combining the sliding window algorithm with historical supervision data; the early warning generation unit determines the early warning level based on the dynamic threshold and generates corresponding early warning information; the push tracking unit is used to allocate push priorities according to the early warning level.

[0059] like Figure 8 As shown, the document management and model optimization module includes a document generation unit, a document storage unit, and a model optimization unit. The document generation unit is used to record data, analysis results, and corrective actions throughout the entire supervision process, forming standardized supervision documents. The document storage unit is used to encrypt and store supervision documents and perform keyword retrieval. The model optimization unit uses the Q-learning algorithm to optimize the parameters of the two models.

[0060] Example 1: This embodiment uses a municipal-level digital government "one-stop online service" upgrade project as an application scenario. The project has a 12-month cycle and an investment of 5 million yuan. Participants include the government service bureau, a technology development company, and a supervision unit. The method and system of this invention are verified to verify their effectiveness. The specific implementation steps are as follows: 1. Data Acquisition: The data acquisition module collects supervision data throughout the entire lifecycle of this "One-Stop Government Service" upgrade project, including: (1) Basic information data of the project: The project name is "City-level One-stop Service Upgrade Project", the cycle is 12 months, the investment is 5 million yuan, and the participants are Party A Municipal Government Service Bureau, Party B Technology Development Company and Party B Supervision Unit.

[0061] (2) Progress data: the planned completion time, actual completion time, and progress deviation of each node (requirement research, system development, testing and launch, operation and maintenance optimization). The required research node was planned for 2 months, but was actually completed in 2.5 months, with a progress deviation of 25%.

[0062] (3) Quality data: The code inspection report detected 3 syntax vulnerabilities, the test report showed a 92% pass rate for functional tests, and the defect record included 5 functional defects and their rectification status.

[0063] (4) Cost data: The budget execution status shows that RMB 3 million has been spent, accounting for 60% of the budget. The details of the expenses and costs are accurate.

[0064] (5) Security data: The vulnerability scan report shows 2 high-risk vulnerabilities, the security drill record shows 1 security drill with a pass rate of 85%, and the security hazard rectification record.

[0065] (6) Compliance data: The government data security compliance test report complies with the "Government Data Security Management Measures", the contractor's qualification documents are complete, and the compliance review records are complete.

[0066] (7) Collaborative data: 12 cross-departmental coordination meetings, 6 out of 8 issues have been rectified.

[0067] 2. Data preprocessing: (1) Data cleaning: The mean filling method was used to fill in one missing cost expenditure detail data, the 3σ principle was used to remove one abnormal progress deviation data, the error was 120%, the hash check was used to remove two duplicate communication records, and the moving average method was used to reduce noise in the safety exercise pass rate data.

[0068] (2) Standardization processing: The min-max standardization method is used to map all data to the interval [0,1]. For example, the schedule deviation of 25% is standardized to 0.25, and the functional test pass rate of 92% is standardized to 0.92, thus obtaining the standardized supervision dataset.

[0069] 3. Model Building: (1) Improved CNN-LSTM risk prediction model: Construct an input layer, a CNN feature extraction layer (convolutional kernel size 3, number 16), an attention mechanism layer, an LSTM temporal feature extraction layer (hidden layer number 64), a fully connected layer, and an output layer; the activation function is the ReLU function, the learning rate is set to 0.01, and the number of iterations is 100.

[0070] (2) Improved AHP-Fuzzy Comprehensive Evaluation Quality Assessment Model: A three-level evaluation system is constructed. The target layer is project supervision quality, the criterion layer includes six dimensions: schedule supervision, quality supervision, cost supervision, safety supervision, compliance supervision, and collaborative supervision, and the indicator layer includes the above 22 supervision indicators. The improved AHP method is used to determine the weight of each indicator, with quality supervision weighted at 0.25, safety supervision weighted at 0.20, compliance supervision weighted at 0.18, schedule supervision weighted at 0.15, cost supervision weighted at 0.12, and collaborative supervision weighted at 0.10. Consistency verification is performed. =0.08<0.1, the judgment matrix has satisfactory consistency; the fuzzy evaluation level is divided into four levels: excellent, good, qualified, and unqualified.

[0071] 4. Analysis and Source Tracing: Input the standardized supervision dataset into the dual model to obtain: (1) Risk prediction model output: The project risk level is 3, which is a medium risk. The risk points are "delay in demand survey", "high-risk security vulnerabilities in the system" and "not all functional defects have been rectified".

[0072] (2) Quality assessment model output: The project supervision quality comprehensive score of 82 points belongs to the good level.

[0073] (3) Source analysis: By using the improved cause-effect graph source analysis algorithm, the causes of risk points are traced: the delay in the progress of demand survey is due to the untimely cross-departmental demand coordination, the high-risk security vulnerability is due to the non-standard security development by the contractor, and the functional defects have not been rectified because the rectification time limit is not clear. The impact of each antecedent factor is calculated by the core formula, of which the impact of "untimely cross-departmental demand coordination" is 0.62, the impact of "non-standard security development by the contractor" is 0.78, and the impact of "non-standard rectification time limit" is 0.45.

[0074] Generating corrective action recommendations: ① Establish a cross-departmental liaison team and hold a requirements coordination meeting once a week to expedite requirements confirmation; ② The contractor shall complete the rectification of high-risk vulnerabilities within 15 days, with the supervisor overseeing the entire process; ③ Clarify the rectification deadline (within 10 days) and responsible persons for the remaining two functional defects.

[0075] 5. Early warning push and corrective tracking: A dynamic threshold early warning mechanism is adopted. Based on historical risk level data of similar "one-stop online service" projects, a sliding window algorithm is used to dynamically determine the three-level threshold by setting the average risk level of the implementation phase to 2.8 and the standard deviation to 0.5 over a 30-day window. Level 1 threshold is below 2.3, no warning; Level 2 threshold is 2.3-3.3, yellow warning; Level 3 threshold is above 3.3, red warning; this project has a risk level of 3, which is within the Level 2 threshold range, so a yellow warning is issued and pushed to the supervising engineer and the contractor's technical manager via system message, SMS and government office platform pop-up.

[0076] A tiered responsibility mechanism is adopted, with Party B taking the lead in rectification and Party A supervising. If rectification is not completed on time, a written rectification explanation must be submitted and the rectification will be included in the compliance review. The implementation of the rectification will be tracked: Party B shall complete the rectification of high-risk vulnerabilities within 15 days, complete the rectification of remaining functional defects within 10 days, ensure that the cross-departmental coordination team is operating normally, resolve the problem of delayed demand surveys, and pass the dual acceptance by Party A and Party A to complete the rectification closed loop.

[0077] 6. Record Management and Model Optimization: Record all data, analysis results, and corrective action implementation throughout the supervision process to create standardized supervision files. These files are stored encrypted and support keyword retrieval. The Q-learning algorithm is used to optimize the parameters of the dual-model system, with a learning rate of [missing information]. =0.05, discount factor =0.9, with risk identification accuracy, quality assessment error and correction completion rate as reward signals. After optimization, the model's risk identification accuracy increased to 94% and the quality assessment error decreased to 3%.

[0078] The method and system of this invention were compared with traditional manual supervision methods and existing simple information-based supervision methods. The experimental period was 12 months for the "One-Stop Government Service" upgrade project. The experimental indicators included supervision efficiency, risk identification accuracy, quality assessment error, project overdue rate, cost overrun rate, and supervision response time. The experimental data are shown in Table 1 below: Table 1. Comparison of various experimental indicators for different supervision methods

[0079] From Table 1 and Figure 9 It can be seen that the present invention is significantly superior to traditional manual supervision and existing simple information-based supervision methods in all indicators. The accuracy of risk identification is increased by 12-22 percentage points, the quality assessment error is reduced by 5-9 percentage points, the project cycle overdue rate and cost overrun rate are reduced by more than 60%, and the supervision response time is shortened by 66.7%-83.3%.

[0080] This fully verifies the effectiveness and superiority of the present invention, which can effectively solve the pain points of existing digital government project supervision, improve supervision quality and efficiency, and ensure the compliant and efficient progress of projects. In particular, after adopting the dynamic threshold early warning + hierarchical responsibility closed-loop scheme, the early warning accuracy is improved by 25%, and the completion rate of the correction closed loop is improved by 30%, further demonstrating the adaptability and superiority of the solution.

[0081] Therefore, the present invention adopts the above-mentioned digital government project supervision method and system based on artificial intelligence, and constructs a full life cycle, precise and adaptive intelligent supervision system, which significantly improves the quality and efficiency of digital government project supervision and ensures the safe, compliant and efficient progress of project construction.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A digital government project supervision method based on artificial intelligence, characterized in that, Includes the following steps: S1. Collect various types of supervision data throughout the entire lifecycle of digital government projects; S2. Preprocess the collected supervision data to obtain a standardized supervision dataset; S3. Construct an improved CNN-LSTM risk prediction model and an AHP-fuzzy comprehensive evaluation quality assessment model; S4. Input the standardized supervision dataset into the model constructed in S3. Output the project risk level and risk points through the improved CNN-LSTM risk prediction model. Output the project supervision quality score through the AHP-fuzzy comprehensive evaluation quality assessment model. Combine the source tracing algorithm to conduct source tracing analysis on the risk points and generate corrective suggestions. S5. Adopt a dynamic threshold early warning and hierarchical responsibility mechanism to conduct risk early warning and corrective tracking; S6. Record the data, analysis results and corrective actions implemented throughout the entire supervision process to form a supervision file, and optimize the model parameters using reinforcement learning algorithms.

2. The method for supervising digital government projects based on artificial intelligence according to claim 1, characterized in that, In S1, the collected supervision data includes basic project information data, schedule data, quality data, cost data, safety data, compliance data, and collaboration data; Basic project information includes project name, duration, investment, and participating party information; progress data includes completion time of each node, progress deviation, and reasons for node delays; quality data includes code inspection reports, test reports, defect records, and defect rectification status; cost data includes budget execution status, expense details, and cost deviation analysis; security data includes vulnerability scanning reports, security drill records, and security vulnerability rectification records; compliance data includes government data security compliance inspection reports, qualification documents, and compliance review records; and collaboration data includes cross-departmental communication records, problem rectification records, and collaboration node completion status.

3. The method for supervising digital government projects based on artificial intelligence according to claim 2, characterized in that, In S2, the collected supervision data is cleaned and standardized. Data cleaning includes handling missing values, outliers, and duplicate and noise values. Standardization uses the min-max standardization method to map all data to the [0,1] interval.

4. The method for supervising digital government projects based on artificial intelligence according to claim 3, characterized in that, In S3, the improved CNN-LSTM risk prediction model introduces an attention mechanism, including an input layer, a CNN local feature extraction layer, an attention mechanism layer, an LSTM temporal feature extraction layer, a fully connected layer, and an output layer. The convolution operation formula for the local feature extraction layer of a CNN is: ； in, Indicates the first Each convolution outputs features, Represents the ReLU activation function. Indicates the size of the convolution kernel. Indicates the convolution kernel number 1 The weight of each position, Indicates the input data sample number. The first feature dimension and the convolution kernel Data points corresponding to each location, For convolution bias; The weight allocation formula for the attention mechanism layer is: ； in, Indicates the first Attention weights for each local feature. Indicates the first Importance score of each local feature Indicates the number of local features; Indicates the first Importance score of each local feature; The gating formulas for the LSTM temporal feature extraction layer include: ； ； ； ； ； ； in, Indicates the output of the forget gate. This represents the sigmoid activation function. This represents the forget gate weight matrix. This represents the output of the hidden layer at the previous time step. This represents the input features at the current time. Indicates the forget gate bias; Indicates the input gate output. This represents the input gate weight matrix. Indicates input gate bias; Indicates the state of candidate cells. Represents the cell state weight matrix. This indicates a cell state bias; This indicates the current state of the cell. This indicates the cell state at the previous moment. Represents element-wise multiplication; Indicates the output gate output. This represents the output gate weight matrix. Indicates output gate bias; This indicates the output of the hidden layer at the current moment.

5. The method for supervising digital government projects based on artificial intelligence according to claim 3, characterized in that, In S3, the improved AHP-fuzzy comprehensive evaluation quality assessment model includes an indicator system construction unit, a weight determination unit, a fuzzy evaluation unit, and a comprehensive scoring unit; The indicator system construction unit includes an objective layer, a criterion layer, and an indicator layer. The objective layer is the project supervision quality, the criterion layer includes schedule supervision, quality supervision, cost supervision, safety supervision, compliance supervision, and collaborative supervision, and the indicator layer includes 22 supervision indicators. The weight determination unit adopts the improved AHP method. A judgment matrix is ​​constructed using the 1-9 scaling method, and the weights of each index are obtained by solving the eigenvectors using the sum-product method. The normalization formula for the judgment matrix is: ； The formula for calculating the eigenvector is: ； The consistency test formula is: ； in, Represents the judgment matrix of the first... Line number Column elements, Represents the normalized elements; Indicates the first The weight of each indicator, Indicates the number of indicators. For consistency ratio, As a consistency indicator; ； in, This indicates the determination of the largest eigenvalue of the matrix. Represents the average random consistency index; when When <0.1, the judgment matrix has satisfactory consistency; Fuzzy evaluation unit constructs fuzzy relation matrix : ； in, Indicates the first The first indicator for the first The degree of membership of each evaluation level; the evaluation levels are divided into excellent, good, satisfactory and unsatisfactory; This indicates the row dimension, corresponding to the number of supervision indicators; This indicates the column dimension and the corresponding number of rating levels. The comprehensive scoring unit uses a weighted average fuzzy operator, the formula of which is: ； in, Represents the comprehensive evaluation vector. Represents the indicator weight vector; Final overall score : ； in, Indicates the first Membership degree of each evaluation level Indicates the first The score for each evaluation level.

6. The method for supervising digital government projects based on artificial intelligence according to claim 3, characterized in that, In S4, the source tracing algorithm uses a cause-effect graph approach, taking the risk point as the target node to trace all antecedent factors that led to the risk, including data, process, and personnel aspects. Specifically: At the data level, issues include incomplete data collection by supervisors, missing data, data anomalies, untimely handling of duplicate data, non-compliance with data standardization, insecure data storage, data transmission leaks, missing government compliance data, and untimely data updates. At the process level, issues include missing supervision processes at each stage of the project, non-standard process execution, poor cross-departmental collaboration, broken risk warning-correction-acceptance closed-loop process, and unclear defect rectification process. At the personnel level, issues include insufficient professional competence of supervisory personnel, unclear responsibilities of project participants, untimely response from cross-departmental liaison personnel, non-standard operation by contractor's technical personnel, and failure to assign responsibility for rectification. Construct a risk source map to clarify the impact of each preceding factor; corrective recommendations include rectification measures, division of responsibilities, rectification time limits, and acceptance standards.

7. The method for supervising digital government projects based on artificial intelligence according to claim 6, characterized in that, In S4, the cause-effect graph approach quantifies the influence of each antecedent factor using the following formula: ； in, Indicates the first Antecedent factors and risk points The degree of influence; Indicates the existence of the first Antecedent factors At that time, risk points The probability of occurrence; This indicates that the first [number] does not exist. Antecedent factors At that time, risk points The probability of occurrence.

8. The method for supervising digital government projects based on artificial intelligence according to claim 7, characterized in that, In S5, a dynamic threshold early warning mechanism is adopted. Combining project type, project stage, and historical supervision data, the early warning thresholds for each level are adjusted in real time using an improved CNN-LSTM model. Specifically: The mean and standard deviation of the risk level during the implementation phase are calculated using a sliding window algorithm, and are divided into three threshold levels: The first-level threshold is one standard deviation below the average risk level of the current stage; no warning will be issued at this level. The secondary threshold is the mean risk level of the current stage ± 1 standard deviation, at which a yellow warning is issued. The Level 3 threshold is one standard deviation above the average risk level of the current stage, at which a red alert is issued. Push notifications are prioritized based on the warning level, and the methods include system messages, SMS, emails, and pop-up reminders on government office platforms. Specifically: Red alerts are immediately sent to all relevant responsible persons, yellow alerts are sent to the supervising engineer and the contractor's technical head, and no alerts are sent only to the supervision file. The corrective action tracking adopts a hierarchical responsibility mechanism, with red alerts being led by the supervision unit and supervised by the client. The yellow alert will be led by Party B and supervised by the supervisor; if rectification is not completed on time, the alert will be upgraded and a written rectification explanation will be submitted, which will be included in the project compliance review record. After rectification is completed, it will be subject to dual acceptance by the supervisor and Party A.

9. The method for supervising digital government projects based on artificial intelligence according to claim 8, characterized in that, In S6, the reinforcement learning algorithm uses the Q-learning algorithm, and the update formula is: ； in, Representing state Next action of value, Indicates the learning rate; This indicates a reward signal; Indicates the discount factor; Indicates the execution of an action The next state corresponds to the supervision analysis state after the model parameters are adjusted. Representing state The optimal action is as follows. Representing state Next action The updated value, New State Next, select the optimal action. The maximum expected future that can be obtained value.

10. An artificial intelligence-based digital government project supervision system, used to implement the method described in any one of claims 1-9, characterized in that, It includes a data acquisition module, a data preprocessing module, a model building module, an analysis and tracing module, an early warning and push module, and a file management and model optimization module; The data acquisition module includes an interface acquisition unit for calling government system interfaces to collect various types of supervision data, a manual input unit for manually inputting supervision data that cannot be automatically collected, an equipment acquisition unit for collecting equipment operation data during project implementation, and a document parsing unit for parsing project-related documents and extracting supervision data. The data preprocessing module includes a data cleaning unit and a standardization unit. The data cleaning unit is used to handle missing values, outliers, duplicate values, and noise in the supervision data. The standardization unit uses the min-max standardization method to unify the format and units of the cleaned data to obtain a standardized supervision dataset. The model building module includes a risk prediction model building unit and a quality assessment model building unit. The risk prediction model building unit is used to build an improved CNN-LSTM risk prediction model that incorporates an attention mechanism; the quality assessment model building unit is used to build an improved AHP-fuzzy comprehensive evaluation quality assessment model with an optimized index system. The analysis and source tracing module includes a model inference unit, a risk source tracing unit, and a corrective suggestion generation unit. The model inference unit is used to input the standardized supervision dataset into the dual model and output the project risk level, risk points, and supervision quality score. The risk tracing unit is used to trace the risk points across the entire chain using the improved cause-effect graph tracing method and quantify the influence of antecedent factors; the corrective suggestion generation unit is used to generate targeted corrective suggestions based on the tracing results; The early warning push module includes a threshold dynamic adjustment unit, an early warning generation unit, and a push tracking unit. The threshold dynamic adjustment unit adjusts the early warning thresholds for each level in real time by combining a sliding window algorithm with historical supervision data. The early warning generation unit determines the early warning level based on the dynamic thresholds and generates corresponding early warning information. The push tracking unit is used to allocate push priorities according to the warning level; The document management and model optimization module includes a document generation unit, a document storage unit, and a model optimization unit. The document generation unit is used to record data, analysis results, and corrective actions throughout the entire supervision process, forming standardized supervision documents. The document storage unit is used to encrypt and store supervision documents and perform keyword retrieval. The model optimization unit uses the Q-learning algorithm to optimize the parameters of the two models.

Images