Intelligent early warning method for flow risk assessment in digital supervision service

Abstract

The invention discloses an intelligent early warning method for flow risk assessment in digital supervision business, which belongs to the technical field of risk early warning of supervision business, and comprises the steps of constructing a risk feature system, constructing a basic assessment decision risk tree, assessing a random forest integration risk and performing intelligent early warning. According to the random forest algorithm, the defect that a single decision-making tree is prone to overfitting is effectively overcome through ensemble learning, decision making is conducted through collective intelligence of multiple trees, and the accuracy of risk assessment is remarkably improved; the risk tree provides a clear white-box model, so that a reasoning path from risk characteristics to a final conclusion can be intuitively displayed; and along with continuous accumulation of new process data and risk cases in the digital supervision service, the algorithm model can be retrained periodically, so that the evaluation performance is continuously optimized, and the adaptability is high.

Description

Technical Field

[0001] This invention relates to the field of risk early warning technology in supervision business, specifically to an intelligent early warning method for process risk assessment in digital supervision business. Background Technology

[0002] With the widespread application of digitalization across various fields, engineering supervision is also undergoing a comprehensive digital transformation. Current risk warning functions in digital supervision are mostly based on simple rule triggers, which have significant shortcomings: 1. Weak ability to trace the causes of risks, making it difficult to provide accurate decision support; 2. Low warning accuracy, prone to generating numerous false alarms, leading to "warning fatigue"; 3. Inability to comprehensively assess the overall risk level from massive, multi-dimensional data. There is an urgent need for an intelligent risk assessment and warning method that can deeply integrate knowledge from the supervision field and possess high precision and interpretability. Therefore, this invention provides an intelligent early warning method for process risk assessment in digital supervision. Summary of the Invention

[0003] To address the aforementioned technical shortcomings, the purpose of this invention is to provide an intelligent early warning method for process risk assessment in digital supervision operations. The random forest algorithm effectively overcomes the overfitting problem of single decision trees through ensemble learning, utilizing the "collective wisdom" of multiple trees for decision-making, thus significantly improving the accuracy of risk assessment. The risk tree provides a clear white-box model that can intuitively demonstrate the reasoning path from risk characteristics to the final conclusion. As new process data and risk cases are continuously accumulated in digital supervision operations, the algorithm model can be periodically retrained, thereby continuously optimizing its assessment performance. This solves the technical problems raised in the background section.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides an intelligent early warning method for process risk assessment in digital supervision business, comprising the following processes:

[0005] A1: Construct a risk feature system: Collect historical data of digital supervision business, extract multi-dimensional features that affect process risks, and construct a risk feature vector;

[0006] A2: Constructing a basic assessment decision risk tree: Based on the risk feature vector, the decision risk tree is constructed using the information gain criterion to achieve risk assessment and causal path location for a single process;

[0007] A3: Assess the risk of random forest ensemble: Integrate multiple decision risk trees to form a random forest, and calculate the comprehensive risk probability of process instances;

[0008] A4: Intelligent Early Warning: Based on the calculated comprehensive risk probability and the causal path provided by the decision risk tree, different levels of early warning are triggered.

[0009] Preferably, the historical data collected in A1 for digital supervision operations includes collecting raw data from various modules of the digital supervision platform, followed by data cleaning and integration, as detailed below:

[0010] Data source:

[0011] Retrieve project plan, WBS, Gantt chart, and actual progress data from the project management module;

[0012] Retrieve material / equipment inspection reports, inspection batch acceptance records, sub-item acceptance records, quality problem notices, and rectification response forms from the quality supervision module;

[0013] Retrieve safety inspection records, safety hazard notices, safety accident reports, and safety logs from the safety supervision module;

[0014] Retrieve supervisor attendance and job records, equipment usage logs, and expense reimbursement records from the resource management module;

[0015] Obtain the minutes of the supervision meeting, supervision work contact sheets, supervision monthly reports, and closed-loop records of instruction sheets from the document and communication module;

[0016] Data preprocessing:

[0017] Data cleaning: handling missing values ​​and outliers;

[0018] Data integration: By using key fields such as "Project ID", "Task ID", and "Timestamp", data from different sources are linked and merged to form a wide table with "Supervision Process Instance" as the basic unit;

[0019] Data transformation: Transforming unstructured data into structured data using natural language processing techniques.

[0020] Preferably, multi-dimensional features affecting process risk are extracted from A1, as follows:

[0021] Progress-related features:

[0022] Critical path task delay rate = (number of delayed tasks on the current critical path / total number of tasks on the critical path) × 100%; Overall schedule deviation rate = (planned workload - actual workload) / planned workload × 100%; Milestone urgency = (current time - planned date of the most recent future milestone) / remaining time;

[0023] Quality characteristics:

[0024] Acceptance pass rate = (Number of inspection batches that passed acceptance on the first attempt / Total number of inspection batches submitted for inspection) × 100%; Quality hazard rectification rate = (Number of quality problems that have been closed / Total number of quality problem notices issued) × 100%; Frequency of non-conformity in key processes: Statistics on the number of non-conformity reports issued for key processes;

[0025] Security features:

[0026] Safety hazard density per unit time = (number of safety hazards discovered this week / month / actual number of construction days this week / month); Percentage of major hazards = (number of safety hazards judged as "major" / total number of safety hazards) × 100%; Delay rate in the implementation of safety measures = (overdue safety rectification orders / total number of safety rectification orders issued) × 100%;

[0027] Resource characteristics:

[0028] The attendance rate of core supervision personnel = (actual attendance days of chief supervisor / professional supervision engineer / planned attendance days) × 100%; the availability rate of testing equipment = (normal standby time of equipment / total time) × 100%; the deviation in the execution of supervision cost budget = (actual supervision cost / budgeted supervision cost - 1) × 100%;

[0029] Compliance and document-related features:

[0030] Supervision instruction closure rate = (number of closed supervision instructions / total number of supervision instructions issued) × 100%; Supervision log timely reporting rate = (number of logs submitted on time / total number of logs to be submitted) × 100%; Document quality index: based on text analysis, assess the completeness and standardization of supervision monthly reports, minutes and other documents.

[0031] Preferably, the risk feature vector is constructed in A1 as follows:

[0032] Constructing a feature vector: For a supervision process instance to be evaluated, calculate its values ​​on all selected features from the fused data to form a fixed-length numerical vector X = [x1, x2, x3, ..., xn];

[0033] Label definition: Define a risk label Y for the aforementioned supervision process instance.

[0034] Preferably, the process of constructing the decision risk tree in A2 using the information gain criterion is as follows:

[0035] The recursive steps are as follows:

[0036] A21: Feature Selection and Information Gain Calculation: For all samples at the current node, iterate through each feature F; for each feature F, calculate the information gain obtained when using it as a splitting rule; the information gain calculation is based on information entropy; for a node D containing K risk categories, its entropy is defined as:

[0037]

[0038] Among them, P k It represents the proportion of samples of the k-th class in node D;

[0039] If feature F has V possible values, then the dataset D is split into V subsets {D}. 1 D 2 ,...,D V}; Calculate the information gain obtained by partitioning node D using feature F:

[0040]

[0041] A22: Branch Node Decision: From all candidate features, select the feature with the largest information gain as the decision feature for the current node, and establish a branch based on the value type of the decision feature:

[0042] For discrete features: create a branch directly for each value;

[0043] For continuous features: find the optimal split point to discretize them;

[0044] A23: Recursive splitting: For all child nodes generated in A22, recursively call the procedures A21 and A22, that is, from the beginning of the step, treat the child node as the new "current node";

[0045] A24: Termination condition: Termination occurs when any of the following conditions are met;

[0046] All samples in the node belong to the same risk category;

[0047] The current node contains fewer samples than the preset minimum value;

[0048] All features have been used on this path, or the information gain of the remaining features is below a preset threshold.

[0049] The path depth of this node has reached the preset maximum depth.

[0050] Preferably, the risk assessment and causal path identification for a single process in A2 are as follows:

[0051] Vector input: Input the feature vector X = [x_1, x_2, ..., x_n] of the process instance to be evaluated from the root node of the tree;

[0052] Path traversal: Starting from the root node, according to the decision rule of the current node, test the feature value corresponding to the instance; based on the test result, select the corresponding branch and move to the next child node; repeat the above process until a leaf node is reached.

[0053] Output and positioning:

[0054] Risk assessment result: The risk level marked by this leaf node;

[0055] Risk causation path location: The path from the root node to the leaf node is the decision logic chain for this risk assessment.

[0056] Preferably, multiple decision risk trees are integrated in A3 to form a random forest, and the specific process is as follows:

[0057] A31: Bootstrap Sampling to Generate Training Subsets: From the original training dataset D containing N historical process instances, perform T Bootstrap samplings; each sampling generates a training subset D of the same size as the original training set. t ;

[0058] A32: Parallel Construction of Decision Risk Trees: For the t-th training subset D_t, independently train a decision risk tree. t Randomly select a subset of features, and from this random subset, select the feature with the largest information gain for node splitting;

[0059] A33: Constructing a Random Forest: Repeat the above process T times to generate T decision risk trees {Tree1, Tree2, ..., Tree...} T The T trees together constitute a random forest model for risk assessment of digital supervision business processes.

[0060] Preferably, in A3, the comprehensive risk probability of the process instance is calculated precisely, and the specific process is as follows:

[0061] Vector Input: The feature vector X of the process instance to be evaluated is input into each decision risk tree in the random forest. t ;

[0062] Single tree prediction: Each tree t Based on its internal independent decision rules, the vector X is traversed until a leaf node is reached; each tree outputs a prediction result; depending on the task type, the output format is as follows:

[0063] Classification task: Output a specific risk category label for each tree;

[0064] Regression task: Output a continuous probability value for each tree;

[0065] Integrated decision-making mechanism:

[0066] For classification tasks: a "voting method" is used: the prediction results of all T trees in the forest for the feature vector X are counted; the risk level with the most votes is taken as the final comprehensive prediction result of the random forest, which is expressed as: final risk level Where c is the risk category, I() is the indicator function, and when y t The value is 1 when the expression is equal to c, and 0 otherwise.

[0067] For regression tasks: the "averaging method" is used: the probability values ​​output by all T trees in the forest are averaged to obtain the final comprehensive risk probability, which can be expressed as a formula:

[0068] Final risk probability

[0069] Where T is the total number of models or predictions; and y is... t The risk probability value given by the t-th model.

[0070] Preferably, the specific process for A4 is as follows:

[0071] Multi-level early warning triggering mechanism:

[0072] Warning level definition:

[0073] Level 1 Warning: P low ≤P <P medium Furthermore, the decision-making process does not contain any major anomalies such as "veto power";

[0074] Level II Warning: P medium ≤P <P high Or, even if P does not reach the intermediate level, a serious anomaly occurs in one key feature of the decision-making path;

[0075] Level 3 Warning: P ≥ P high Or, serious anomalies with ≥2 key features or "veto" features appear in the decision-making path;

[0076] Threshold setting: Threshold P low P medium P high It is determined by historical data backtesting, model performance evaluation, and combined with management's risk preferences.

[0077] Preferably, the risk assessment report generated in A4 contains the following: Digital Supervision Business Process Risk Assessment Report: Report Number, Generation Time, Assessment Object, Comprehensive Risk Level, and Comprehensive Risk Probability.

[0078] The beneficial effects of this invention are as follows:

[0079] The random forest algorithm of this invention effectively overcomes the overfitting problem of a single decision tree through ensemble learning, and makes decisions by utilizing the "collective wisdom" of multiple trees, which significantly improves the accuracy of risk assessment. The risk tree provides a clear white-box model that can intuitively show the reasoning path from risk characteristics to the final conclusion. As new process data and risk cases are continuously accumulated in digital supervision business, the algorithm model can be retrained periodically to continuously optimize its assessment performance, making it highly adaptable. Attached Figure Description

[0080] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0081] Figure 1 A flowchart of the intelligent early warning method for process risk assessment in digital supervision business provided in this embodiment of the invention.

[0082] Figure 2 This is a flowchart of constructing the decision risk tree using the information gain criterion in A2 of the present invention.

[0083] Figure 3 This is a flowchart of the process of integrating multiple decision risk trees to form a random forest in A3 of the present invention. Detailed Implementation

[0084] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0085] Example:

[0086] like Figures 1 to 3 As shown, this invention provides an intelligent early warning method for process risk assessment in digital supervision operations, comprising the following steps:

[0087] A1: Construct a risk feature system: Collect historical data of digital supervision business, extract multi-dimensional features that affect process risks, and construct a risk feature vector;

[0088] A2: Constructing a basic assessment decision risk tree: Based on the risk feature vector, the decision risk tree is constructed using the information gain criterion to achieve risk assessment and causal path location for a single process;

[0089] A3: Assess the risk of random forest ensemble: Integrate multiple decision risk trees to form a random forest, and calculate the comprehensive risk probability of process instances;

[0090] A4: Intelligent Early Warning: Based on the calculated comprehensive risk probability and the causal path provided by the decision risk tree, different levels of early warning are triggered, and a risk assessment report is generated.

[0091] In this embodiment, the random forest algorithm effectively overcomes the disadvantage of overfitting of a single decision tree through ensemble learning, and makes decisions by utilizing the "collective wisdom" of multiple trees, which significantly improves the accuracy of risk assessment. The risk tree provides a clear white-box model that can intuitively show the reasoning path from risk characteristics to the final conclusion. As new process data and risk cases are continuously accumulated in digital supervision business, the algorithm model can be retrained periodically to continuously optimize its assessment performance and has strong adaptability.

[0092] Furthermore, the historical data collection for digital supervision operations in A1 includes collecting raw data from various modules of the digital supervision platform, and performing data cleaning and integration, as detailed below:

[0093] Data sources: Project plans, WBS, Gantt charts, and actual progress data are obtained from the project management module; material / equipment inspection records, inspection batch acceptance records, sub-item acceptance records, quality problem notices, and rectification response forms are obtained from the quality supervision module; safety inspection records, safety hazard notices, safety accident reports, and safety logs are obtained from the safety supervision module; attendance and job records of supervisory personnel, equipment usage ledgers, and expense reimbursement records are obtained from the resource management module; and minutes of supervisory meetings, supervisory work contact forms, supervisory monthly reports, and closed-loop records of instruction forms are obtained from the document and communication module.

[0094] Data preprocessing:

[0095] Data cleaning: handling missing values ​​by imputing with the mean or model prediction, and identifying and correcting outliers based on business rules;

[0096] Data integration: By using key fields such as "Project ID", "Task ID", and "Timestamp", data from different sources are linked and merged to form a wide table with "Supervision Process Instance" as the basic unit;

[0097] Data transformation: Transforming unstructured data into structured data using natural language processing techniques.

[0098] Furthermore, A1 extracts multi-dimensional features that affect process risk, as follows:

[0099] Schedule-related characteristics: Critical path task delay rate = (number of delayed tasks on the current critical path / total number of tasks on the critical path) × 100%, reflecting the health of the core schedule; Overall schedule deviation rate = (planned workload - actual workload) / planned workload × 100%, quantifying the degree of overall schedule lag; Milestone achievement urgency = (current time - planned date of the most recent future milestone) / remaining time, the smaller this value, the greater the time pressure;

[0100] Quality characteristics: First-time acceptance pass rate = (Number of inspection batches that passed the first acceptance / Total number of inspection batches submitted for inspection) × 100%, directly reflecting the stability of construction quality; Quality hidden danger rectification rate = (Number of quality problems that have been closed / Total number of quality problem notices issued) × 100%, measuring the efficiency of correcting quality problems; Frequency of non-conformity in key processes: Statistics on the number of non-conformity reports issued for key processes.

[0101] Safety-related characteristics: Safety hazard density per unit time = (Number of safety hazards discovered this week / month / Actual number of construction days this week / month), dynamically measuring the on-site safety situation; Percentage of major hazards = (Number of safety hazards judged as "major" / Total number of safety hazards) × 100%, reflecting the severity of the risk; Delay rate in the implementation of safety measures = (Overdue safety rectification orders / Total number of safety rectification orders issued) × 100%, measuring the execution of safety management;

[0102] Resource-related characteristics: Core supervision personnel attendance rate = (actual attendance days of chief supervisor / professional supervision engineer / planned attendance days) × 100%, assessing human resource support; Testing equipment availability rate = (normal standby time of equipment / total time) × 100%, ensuring the effectiveness of supervision tools; Supervision cost budget execution deviation = (actual supervision cost / budgeted supervision cost - 1) × 100%, monitoring financial risks;

[0103] Compliance and Documentation Characteristics: Supervision Instruction Closure Rate = (Number of closed supervision instructions / Total number of supervision instructions issued) × 100%, measuring the authority of supervision and the closed-loop management; Supervision Log Timeliness Rate = (Number of logs submitted on time / Total number of logs to be submitted) × 100%, reflecting the standardization and discipline of supervision work; Document Quality Index: Based on text analysis, assessing the completeness and standardization of supervision monthly reports, minutes, and other documents.

[0104] Furthermore, the risk feature vector is constructed in A1 as follows:

[0105] Constructing a feature vector: For a supervision process instance to be evaluated, calculate its values ​​on all selected features from the fused data to form a fixed-length numerical vector X = [x1, x2, x3, ..., xn];

[0106] Tag definition: Define a risk tag Y for the aforementioned supervision process instance;

[0107] Category labels: For example, Y∈{low risk, medium risk, high risk}; the definition criteria can be: "low risk" is defined as: no quality or safety incidents and schedule deviation <5%; "medium risk" is defined as: minor but controllable problems; "high risk" is defined as: major quality / safety / schedule problems.

[0108] Regression label: For example, Y∈[0,1] is a continuous risk probability value, which can be retrospectively assessed by experts based on the severity of the problem and the loss.

[0109] Furthermore, the process of constructing the decision risk tree in A2 using the information gain criterion is as follows:

[0110] The recursive steps are as follows:

[0111] A21: Feature Selection and Information Gain Calculation: For all samples at the current node, iterate through each feature F; for each feature F, calculate the information gain that can be obtained when it is used as a splitting rule; the calculation of information gain is based on information entropy, which is an indicator of the purity of a sample set; for a node D containing K risk categories, its entropy is defined as:

[0112]

[0113] Among them, P k It represents the proportion of samples of the k-th class in node D;

[0114] If feature F has V possible values, then the dataset D is split into V subsets {D}. 1 D 2 ,...,D V}; Calculate the information gain obtained by partitioning node D using feature F:

[0115] The larger the information gain Gain(D, F), the greater the purity of the child nodes is improved after splitting using feature F;

[0116] A22: Branch Node Decision: From all candidate features, select the feature with the largest information gain as the decision feature for the current node, and establish a branch based on the value type of the decision feature:

[0117] For discrete features: create a branch directly for each value;

[0118] For continuous features: find the best split point to discretize it; after sorting all the values ​​of the feature, try all possible split points and select the point that produces the maximum information gain as the split threshold, thus generating two branches, "yes" and "no".

[0119] A23: Recursive splitting: For all child nodes generated in A22, recursively call the procedures A21 and A22, that is, from the beginning of the step, treat the child node as the new "current node";

[0120] A24: Termination condition: Termination occurs when any of the following conditions are met;

[0121] All samples in the node belong to the same risk category;

[0122] The current node contains fewer samples than the preset minimum value;

[0123] All features have been used on this path, or the information gain of the remaining features is below a preset threshold.

[0124] The path depth of this node has reached the preset maximum depth.

[0125] Furthermore, A2 details the risk assessment and causal path identification for a single process, as follows:

[0126] Vector input: Input the feature vector X = [x_1, x_2, ..., x_n] of the process instance to be evaluated from the root node of the tree;

[0127] Path traversal: Starting from the root node, according to the decision rule of the current node, test the feature value corresponding to the instance; based on the test result, select the corresponding branch and move to the next child node; repeat the above process until a leaf node is reached.

[0128] Output and positioning:

[0129] Risk assessment result: The risk level marked by this leaf node;

[0130] Risk causation path location: The path from the root node to the leaf node is the decision logic chain for this risk assessment;

[0131] Example: A path might be: "IF schedule deviation rate > 10% AND acceptance pass rate < 85% AND number of safety hazards > 3 THEN High risk".

[0132] Furthermore, A3 integrates multiple decision risk trees to form a random forest, the specific process of which is as follows:

[0133] A31: Bootstrap Sampling to Generate Training Subsets: From the original training dataset D containing N historical process instances, perform T Bootstrap samplings; each sampling generates a training subset D of the same size as the original training set. t Because sampling with replacement is used, approximately 63.2% of the original samples in each subset will be selected, with the remainder being duplicate samples. Meanwhile, approximately 36.8% of the samples will not be selected. These "out-of-bag samples" can be used for subsequent evaluation of model performance.

[0134] A32: Parallel Construction of Decision Risk Trees: For the t-th training subset D_t, independently train a decision risk tree. t A subset of features is randomly selected, and the feature with the highest information gain is selected from this random subset for node splitting. This process ensures that each tree in the forest is learned under slightly different data and feature perspectives, so each tree has high variance but low bias and is different from one another.

[0135] A33: Constructing a Random Forest: Repeat the above process T times to generate T decision risk trees {Tree1, Tree2, ..., Tree...} T The T trees together constitute a random forest model for risk assessment of digital supervision business processes.

[0136] Furthermore, A3 performs a precise calculation of the overall risk probability for process instances, as detailed below:

[0137] Vector Input: The feature vector X of the process instance to be evaluated is input into each decision risk tree in the random forest. t ;

[0138] Single tree prediction: Each tree t Based on its internal independent decision rules, the vector X is traversed until a leaf node is reached; each tree outputs a prediction result; depending on the task type, the output format is as follows:

[0139] Classification task: Output a specific risk category label for each tree;

[0140] Regression task: Output a continuous probability value for each tree;

[0141] Integrated decision-making mechanism:

[0142] For classification tasks: a "voting method" is used: the prediction results of all T trees in the forest for the feature vector X are counted; the risk level with the most votes is taken as the final comprehensive prediction result of the random forest, which is expressed as: final risk level Where c is the risk category, I() is the indicator function, and when yt The value is 1 when the expression is equal to c, and 0 otherwise.

[0143] For regression tasks: the "averaging method" is used: the probability values ​​output by all T trees in the forest are averaged to obtain the final comprehensive risk probability, which can be expressed as a formula:

[0144] Final risk probability

[0145] Where T is the total number of models or predictions; and y is... t The risk probability value given by the t-th model.

[0146] Furthermore, the specific process for A4 is as follows:

[0147] Multi-level early warning triggering mechanism:

[0148] Warning level definition:

[0149] Level 1 Warning: P low ≤P <P medium Furthermore, the decision-making process does not contain any major anomalies such as "veto power";

[0150] Level II Warning: P medium ≤P <P high Or, even if P does not reach the intermediate level, a serious anomaly occurs in one key feature of the decision-making path;

[0151] Level 3 Warning: P ≥ P high Or, serious anomalies with ≥2 key features or "veto" features appear in the decision-making path;

[0152] Threshold setting: Threshold P low P medium P high It is determined by historical data backtesting, model performance evaluation, and combined with management's risk preferences.

[0153] Furthermore, the risk assessment report generated in A4 contains the following content: Digital Supervision Business Process Risk Assessment Report:

[0154] Report Number: RISK-20231027-008;

[0155] Generation time: 2025-06-27 15:30:02;

[0156] Assessment object: Project A - Supervision process during the main structure construction phase;

[0157] Overall risk level: High risk;

[0158] Overall risk probability: 86%.

[0159] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for intelligent early warning of process risk assessment in digital supervision operations, characterized in that, The process includes the following: A1: Construct a risk feature system: Collect historical data of digital supervision business, extract multi-dimensional features that affect process risks, and construct a risk feature vector; A2: Constructing a basic assessment decision risk tree: Based on the risk feature vector, the decision risk tree is constructed using the information gain criterion to achieve risk assessment and causal path location for a single process; A3: Assess the risk of random forest ensemble: Integrate multiple decision risk trees to form a random forest, and calculate the comprehensive risk probability of process instances; A4: Intelligent Early Warning: Based on the calculated comprehensive risk probability and the causal path provided by the decision risk tree, different levels of early warning are triggered.

2. The intelligent early warning method for process risk assessment in digital supervision operations as described in claim 1, characterized in that, The historical data collection for digital supervision operations in A1 includes collecting raw data from various modules of the digital supervision platform, followed by data cleaning and integration, as detailed below: Data source: Retrieve project plan, WBS, Gantt chart, and actual progress data from the project management module; Retrieve material / equipment inspection reports, inspection batch acceptance records, sub-item acceptance records, quality problem notices, and rectification response forms from the quality supervision module; Retrieve safety inspection records, safety hazard notices, safety accident reports, and safety logs from the safety supervision module; Retrieve supervisor attendance and job records, equipment usage logs, and expense reimbursement records from the resource management module; Obtain the minutes of the supervision meeting, supervision work contact sheets, supervision monthly reports, and closed-loop records of instruction sheets from the document and communication module; Data preprocessing: Data cleaning: handling missing values ​​and outliers; Data integration: By using key fields such as "Project ID", "Task ID", and "Timestamp", data from different sources are linked and merged to form a wide table with "Supervision Process Instance" as the basic unit; Data transformation: Transforming unstructured data into structured data using natural language processing techniques.

3. The intelligent early warning method for process risk assessment in digital supervision operations as described in claim 1, characterized in that, The multi-dimensional features affecting process risk are extracted from A1, as follows: Progress-related features: Critical path task delay rate = (number of delayed tasks on the current critical path / total number of tasks on the critical path) × 100%; Overall schedule deviation rate = (planned workload - actual workload) / planned workload × 100%; Milestone urgency = (current time - planned date of the most recent future milestone) / remaining time; Quality characteristics: Acceptance pass rate = (Number of inspection batches that passed acceptance on the first attempt / Total number of inspection batches submitted for inspection) × 100%; Quality hazard rectification rate = (Number of quality problems that have been closed / Total number of quality problem notices issued) × 100%; Frequency of non-conformity in key processes: Statistics on the number of non-conformity reports issued for key processes; Security features: Safety hazard density per unit time = (number of safety hazards discovered this week / month / actual number of construction days this week / month); Percentage of major hazards = (number of safety hazards judged as "major" / total number of safety hazards) × 100%; Delay rate in the implementation of safety measures = (overdue safety rectification orders / total number of safety rectification orders issued) × 100%; Resource-related characteristics: The attendance rate of core supervision personnel = (actual attendance days of chief supervisor / professional supervision engineer / planned attendance days) × 100%; the availability rate of testing equipment = (normal standby time of equipment / total time) × 100%; the deviation in the execution of supervision cost budget = (actual supervision cost / budgeted supervision cost - 1) × 100%; Compliance and document-related features: Supervision instruction closure rate = (number of closed supervision instructions / total number of supervision instructions issued) × 100%; Supervision log timely reporting rate = (number of logs submitted on time / total number of logs to be submitted) × 100%; Document quality index: based on text analysis, assess the completeness and standardization of supervision monthly reports, minutes and other documents.

4. The intelligent early warning method for process risk assessment in digital supervision operations as described in claim 1, characterized in that, The risk feature vector is constructed in A1 as follows: Constructing a feature vector: For a supervision process instance to be evaluated, calculate its values ​​on all selected features from the fused data to form a fixed-length numerical vector X = [x1, x2, x3, ..., xn]; Label definition: Define a risk label Y for the aforementioned supervision process instance.

5. The intelligent early warning method for process risk assessment in digital supervision operations as described in claim 1, characterized in that, The process of constructing the decision risk tree using the information gain criterion in A2 is as follows: The recursive steps are as follows: A21: Feature Selection and Information Gain Calculation: For all samples at the current node, iterate through each feature F; for each feature F, calculate the information gain obtained when using it as a splitting rule; the information gain calculation is based on information entropy; for a node D containing K risk categories, its entropy is defined as: Among them, P k It represents the proportion of samples of the k-th class in node D; If feature F has V possible values, then the dataset D is split into V subsets {D}. 1 D 2 ,...,D V }; Calculate the information gain obtained by partitioning node D using feature F: A22: Branch Node Decision: From all candidate features, select the feature with the largest information gain as the decision feature for the current node, and establish a branch based on the value type of the decision feature: For discrete features: create a branch directly for each value; For continuous features: find the optimal split point to discretize them; A23: Recursive splitting: For all child nodes generated in A22, recursively call the processes A21 and A22, that is, from the beginning of the step, treat the child node as the new "current node"; A24: Termination condition: Termination occurs when any of the following conditions are met; All samples in the node belong to the same risk category; The current node contains fewer samples than the preset minimum value; All features have been used on this path, or the information gain of the remaining features is below a preset threshold. The path depth of this node has reached the preset maximum depth.

6. The intelligent early warning method for process risk assessment in digital supervision operations as described in claim 1, characterized in that, The risk assessment and causal path identification for a single process in A2 are as follows: Vector input: Input the feature vector X = [x_1, x_2, ..., x_n] of the process instance to be evaluated from the root node of the tree; Path traversal: Starting from the root node, according to the decision rules of the current node, test the feature values ​​corresponding to the instance; based on the test results, select the appropriate branch and move to the next child node; Repeat the above process until a leaf node is reached; Output and positioning: Risk assessment result: The risk level marked by this leaf node; Risk causation path location: The path from the root node to the leaf node is the decision logic chain for this risk assessment.

7. The intelligent early warning method for process risk assessment in digital supervision operations as described in claim 1, characterized in that, A3 integrates multiple decision risk trees to form a random forest, and the specific process is as follows: A31: Bootstrap Sampling to Generate Training Subsets: From the original training dataset D containing N historical process instances, perform T Bootstrap samplings; each sampling generates a training subset D of the same size as the original training set. t ; A32: Parallel Construction of Decision Risk Trees: For the t-th training subset D_t, independently train a decision risk tree. t Randomly select a subset of features, and from this random subset, select the feature with the largest information gain for node splitting; A33: Constructing a Random Forest: Repeat the above process T times to generate T decision risk trees {Tree1, Tree2, ..., Tree...} T The T trees together constitute a random forest model for risk assessment of digital supervision business processes.

8. The intelligent early warning method for process risk assessment in digital supervision operations as described in claim 1, characterized in that, A3 performs a precise calculation of the overall risk probability for process instances, as detailed below: Vector Input: The feature vector X of the process instance to be evaluated is input into each decision risk tree in the random forest. t ; Single tree prediction: Each tree t Based on its internal independent decision rules, the vector X is traversed until a leaf node is reached; each tree outputs a prediction result; depending on the task type, the output format is as follows: Classification task: Output a specific risk category label for each tree; Regression task: Output a continuous probability value for each tree; Integrated decision-making mechanism: For classification tasks: a "voting method" is used: the prediction results of all T trees in the forest for the feature vector X are counted; the risk level with the most votes is taken as the final comprehensive prediction result of the random forest, which is expressed as: final risk level Where c is the risk category, I() is the indicator function, and when y t The value is 1 when the expression is equal to c, and 0 otherwise. For regression tasks: the "averaging method" is used: the probability values ​​output by all T trees in the forest are averaged to obtain the final comprehensive risk probability, which can be expressed as follows: Final risk probability Where T is the total number of models or predictions; and y is... t The risk probability value given by the t-th model.

9. The intelligent early warning method for process risk assessment in digital supervision operations as described in claim 1, characterized in that, The specific process for A4 is as follows: Multi-level early warning triggering mechanism: Warning level definition: Level 1 Warning: P low ≤P <P medium Furthermore, the decision-making process does not contain any major anomalies such as "veto power"; Level 2 Warning: P medium ≤P <P high Or, even if P does not reach the intermediate level, a serious anomaly occurs in one key feature of the decision-making path; Level 3 Warning: P ≥ P high Or, serious anomalies with ≥2 key features or "veto" features appear in the decision-making path; Threshold setting: Threshold P low P medium P high It is determined by historical data backtesting, model performance evaluation, and combined with management's risk preferences.

10. The intelligent early warning method for process risk assessment in digital supervision operations as described in claim 1, characterized in that, The risk assessment report generated in A4 contains the following: Digital Supervision Business Process Risk Assessment Report: Report Number, Generation Time, Assessment Object, Comprehensive Risk Level, and Comprehensive Risk Probability.

Images