AI intelligent approval method and system based on cross-department data fusion

Abstract

The invention discloses an AI intelligent approval method and system based on cross-department data fusion, and belongs to the field of data processing, and the method comprises the steps: collecting and integrating multi-source heterogeneous data through a cross-department interface, and forming a unified data set; analyzing a data matching degree by using a support vector machine algorithm, and identifying a potential defect region; based on the historical data, predicting risk probability distribution by adopting a random forest algorithm and generating an early warning index; performing time sequence feature extraction and clustering on the early warning indexes to realize defect mode grouping; mining and determining a defect propagation chain by utilizing association rules; when the chain relates to multiple departments, real-time notification and closed-loop log tracking are triggered through a message queue; and finally, iteratively optimizing a model threshold according to the response log to form a dynamic self-adaptive approval process. The system comprises corresponding function modules. According to the method, deep fusion of cross-department data, active early warning of risks and intelligent co-processing are realized, and the accuracy, timeliness and overall risk control capability of an approval process are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of data processing, and in particular relates to an AI-powered intelligent approval method and system based on cross-departmental data fusion. Background Technology

[0002] In today's enterprise management practices, approval processes are a core element in ensuring business compliance and operational efficiency. As organizations grow in size and business complexity increases, cross-departmental collaboration becomes more frequent. Traditional approval methods often rely on human experience and relatively isolated data systems within departments. Existing technologies typically employ automated approval based on fixed rule engines or integrate basic data analysis functions in certain areas, aiming to improve the processing speed of single-point transactions. However, these methods fundamentally fail to break down data silos between departments, making it difficult to achieve deep integration and collaborative analysis of multi-dimensional information, including financial, organizational, and business history data.

[0003] Existing technologies suffer from the following main problems: First, insufficient information integration capabilities. Inconsistent and isolated data standards across departments lead to a lack of holistic perspective in approval decisions, making it difficult to identify deep-seated conflicts arising from mismatched permissions and resource allocation. Second, risk identification is delayed and reactive. Existing systems generally lack the ability to deeply mine and analyze historical data in real time, failing to proactively predict potential risk paths and defect propagation patterns early in the approval process. Third, a lack of cross-departmental collaborative response mechanisms. When issues involve multiple responsible parties, the lack of efficient real-time notification and closed-loop tracking methods results in risk diffusion or inefficient problem resolution. Therefore, achieving intelligent integration of cross-departmental data, proactive risk warning, and efficient collaborative handling has become a key challenge in improving the overall approval efficiency and risk control level of enterprises. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an AI-powered intelligent approval method based on cross-departmental data fusion, comprising: Multi-source heterogeneous data is collected through cross-departmental system interfaces, and the multi-source heterogeneous data is cleaned and integrated to obtain an integrated dataset in a unified format. Based on the integrated dataset, a classification algorithm is applied to analyze the matching degree between information in different dimensions, and potential defect areas are determined by comparing the matching degree results with a preset threshold. Historical data within the potential defect area is obtained, a first machine learning algorithm is applied based on the historical data to predict the risk probability distribution, and high-risk paths are determined according to the risk probability distribution to obtain a set of early warning indicators. Time series features are extracted from the set of early warning indicators, and the time series features are clustered to group similar defect patterns to obtain classified risk groups. Dominant features are selected from the classified risk groups, and the dependency relationships between the dominant features are analyzed using association rule mining methods to determine the defect propagation chain; If the defect propagation chain involves multiple departments, a real-time notification mechanism is triggered to distribute warning signals to the multiple departments through a message queue and obtain response confirmation logs from each department. The integrated dataset is updated based on the response confirmation log, and the classification threshold of the approval process model is adjusted iteratively to obtain an optimized approval process model.

[0005] Optionally, the step of collecting multi-source heterogeneous data through cross-departmental system interfaces, and cleaning and integrating the multi-source heterogeneous data to obtain an integrated dataset in a unified format includes: Financial budget data and organizational structure data are collected through cross-departmental system interfaces to obtain the multi-source heterogeneous data; The multi-source heterogeneous data is cleaned to remove noise and redundancy, resulting in a refined dataset; The refined dataset is then standardized to obtain a unified format dataset. Extract the correspondence between financial budget data and organizational structure data from the unified format dataset, and determine whether their hierarchical mapping structure is consistent; If the hierarchical mapping structure is inconsistent, the correspondence is corrected by rule matching to obtain a dataset with consistent mapping. Based on the dataset with consistent mapping, an association table between the organization and the budget is constructed to form the integrated dataset.

[0006] Optionally, based on the integrated dataset, a classification algorithm is applied to analyze the matching degree between information in different dimensions, and potential defect areas are determined by comparing the matching degree results with a preset threshold, including: The integrated dataset is trained and classified using the support vector machine algorithm to obtain matching scores that represent the matching degree between information of different dimensions; The matching score is compared with a preset threshold to identify information pairs whose matching score is lower than the preset threshold. The identified information pairs are marked as conflicting permission mismatches, and all marked conflicting information is aggregated to obtain a set of conflicting information. Based on the distribution of multidimensional information in the contradictory information set, potential defect areas are located; Extract the original data subset corresponding to the potential defect region to form a defect subset dataset; The support vector machine algorithm is used to reclassify and verify the defect subset, and the final defect region is determined based on the secondary matching score.

[0007] Optionally, historical data within the potential defect area is obtained, a first machine learning algorithm is applied based on the historical data to predict the risk probability distribution, and high-risk paths are identified according to the risk probability distribution to obtain a set of early warning indicators, including: Historical data within the potential defect area is obtained and cleaned to obtain a basic dataset; Based on the aforementioned basic dataset, the random forest algorithm is applied to calculate the risk probability and generate the risk probability distribution results. Based on the risk probability distribution results, path analysis is performed on high-probability areas to identify a set of high-risk paths; By combining historical anomaly patterns, key features are extracted from the set of high-risk paths to form a set of factors that may trigger risks; Based on the comparison between the occurrence frequency of each factor in the factor set and the frequency threshold, high-risk factors are screened out and a list of high-risk factors is generated. The list of high-risk factors is matched with preset early warning indicators to generate a combination of early warning indicators corresponding to the high-risk path.

[0008] Optionally, time-series features are extracted from the set of early warning indicators, and clustering of these time-series features is performed to group similar defect patterns, resulting in classified risk groups, including: Obtain time-series data from the set of early warning indicators; Statistical features are extracted from each of the time series data to form a feature vector set; The K-means clustering algorithm is used to cluster the feature vector set to obtain preliminary defect groups; Risk scores are calculated based on the distance from each cluster center to the origin, and each cluster is labeled as high-risk or low-risk based on the comparison between the risk scores and a preset distance threshold, resulting in risk-labeled classification groups. For the risk-labeled classification groups, the similarity of defect patterns is determined based on the nearest neighbor distance of their internal time series. All classification groups are sorted according to the risk scores to obtain a risk group priority sequence, and the classified risk group list is output.

[0009] Optionally, dominant features are selected from the classified risk groups, and association rule mining methods are used to analyze the dependencies between the dominant features to determine the defect propagation chain, including: Key features are extracted from each risk group to form a set of key features; The association rule mining method is used to analyze the dependencies between features in the key feature set and determine the strength of the association between features. Construct a preliminary path for defect propagation based on the correlation strength; Paths with association strength below a preset threshold are removed from the initial path to obtain a simplified propagation path. The simplified propagation path is used to locate the core nodes and key features of defect propagation. Based on the roles of the core nodes and key features, a complete defect propagation chain is constructed.

[0010] Optionally, the integrated dataset is updated based on the response confirmation log, and the classification threshold of the approval process model is adjusted iteratively to obtain an optimized approval process model, including: Key fields are extracted from the response confirmation log to obtain structured approval data; The structured approval data is classified to obtain a preliminary category distribution; If the category distribution is unbalanced, data augmentation techniques are used to balance it, resulting in a balanced dataset. Based on the balanced dataset, the logistic regression algorithm is applied to dynamically calculate the classification threshold range; The classification threshold range is adjusted based on the business rules of the approval process to determine the final threshold standard; The decision logic of the approval process is updated based on the final threshold standard to obtain an optimized approval process model; The optimized approval process model is loaded into the system for application, and the approval efficiency is evaluated.

[0011] On the other hand, the present invention also provides an AI-powered intelligent approval system based on cross-departmental data fusion, comprising: The data acquisition and integration module is used to acquire multi-source heterogeneous data through cross-departmental system interfaces, and to clean and integrate the multi-source heterogeneous data to obtain an integrated dataset in a unified format. The matching degree classification and labeling module is used to analyze the matching degree between different dimensions of information based on the integrated dataset using a classification algorithm, and to determine potential defect areas based on the comparison between the matching degree results and a preset threshold. The risk prediction and early warning module is used to acquire historical data within the potential defect area, apply a first machine learning algorithm based on the historical data to predict the risk probability distribution, and determine high-risk paths based on the risk probability distribution to obtain a set of early warning indicators. The time-series feature clustering module is used to extract time-series features from the set of early warning indicators, and to obtain classified risk groups by clustering the time-series features to group similar defect patterns. The association rule mining module is used to select dominant features from the classified risk groups and analyze the dependency relationships between the dominant features using association rule mining methods to determine the defect propagation chain. The real-time notification and distribution module is used to trigger a real-time notification mechanism when the defect propagation chain involves multiple departments, distribute early warning signals to the multiple departments through a message queue, and obtain response confirmation logs from each department. The iterative optimization module is used to update the integrated dataset based on the response confirmation log and iteratively optimize the classification threshold of the approval process model to obtain an optimized approval process model.

[0012] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0013] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: This invention achieves automatic collection, cleaning, and deep fusion of multi-source heterogeneous data through cross-departmental system interfaces, constructing a unified, standardized dataset to lay the foundation for global analysis. Applying machine learning algorithms such as Support Vector Machines and Random Forests, the system can automatically identify permission conflicts and potential defect areas, and predict high-risk paths, shifting from passive response to proactive early warning. Furthermore, by extracting temporal features and performing cluster analysis on early warning indicators, it can accurately group similar defect patterns and reveal defect propagation chains using association rule mining. When multiple departments are involved, the system triggers real-time notifications and closed-loop log tracking through message queues to ensure collaborative efficiency. Finally, based on response feedback data, iterative optimization of model thresholds forms a dynamically adaptive intelligent approval process. This invention effectively solves problems such as cross-departmental data silos, delayed risk identification, difficulty in tracing defect propagation, and inefficient collaborative response, significantly improving the intelligence level of the approval process, risk control capabilities, and overall operational efficiency. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Detailed Implementation It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0016] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0017] Example 1 like Figure 1 As shown, this embodiment provides an AI-powered intelligent approval method based on cross-departmental data fusion, including: Multi-source heterogeneous data is collected through cross-departmental system interfaces, and the multi-source heterogeneous data is cleaned and integrated to obtain an integrated dataset in a unified format. Based on the integrated dataset, a classification algorithm is applied to analyze the matching degree between information in different dimensions, and potential defect areas are determined by comparing the matching degree results with a preset threshold. Historical data within the potential defect area is obtained, a first machine learning algorithm is applied based on the historical data to predict the risk probability distribution, and high-risk paths are determined according to the risk probability distribution to obtain a set of early warning indicators. Time series features are extracted from the set of early warning indicators, and the time series features are clustered to group similar defect patterns to obtain classified risk groups. Dominant features are selected from the classified risk groups, and the dependency relationships between the dominant features are analyzed using association rule mining methods to determine the defect propagation chain; If the defect propagation chain involves multiple departments, a real-time notification mechanism is triggered to distribute warning signals to the multiple departments through a message queue and obtain response confirmation logs from each department. The integrated dataset is updated based on the response confirmation log, and the classification threshold of the approval process model is adjusted iteratively to obtain an optimized approval process model.

[0018] S101. Collect multi-source data such as financial budgets and organizational structures through cross-departmental system interfaces, and use data cleaning methods to remove noise and redundancy to obtain an integrated dataset in a unified format.

[0019] Financial budget data and organizational structure data are collected through cross-departmental and system interfaces to obtain multi-source data. Data cleaning methods are used to remove noise, resulting in a preliminary cleaned dataset. Redundant data is removed from the preliminary cleaned dataset to obtain a refined dataset. Format integration is performed on the refined dataset to determine a unified format dataset. The correspondence between financial budget data and organizational structure data is extracted from the unified format dataset to determine the hierarchical mapping structure. If inconsistencies exist in the hierarchical mapping structure, rule matching is used to correct the correspondence, resulting in a consistent mapping dataset. An organizational budget association table is constructed using the consistent mapping dataset to obtain the integrated dataset.

[0020] When collecting multi-source data such as financial budgets and organizational structures through cross-departmental system interfaces, automated data extraction can be achieved using internal enterprise APIs. For example, annual budget data can be retrieved from the financial system interface. Assuming a company's total budget for 2023 is 50 million yuan, the detailed allocation to each department can be returned in JSON format via the interface. Simultaneously, organizational structure data, including the department affiliation and hierarchical relationships of 1000 employees, can be retrieved from the human resources system interface. This data is stored in XML format. Next, data cleaning methods are used to remove noise and redundancy. For instance, for duplicate records in the financial budget data, a Python script is written using the `drop_duplicates()` function of the Pandas library to remove duplicate rows. Assuming there are 50 duplicate records in the original data, this is reduced to 4950 valid records after cleaning. Furthermore, missing values ​​in the organizational structure data are addressed. It was found that 5% of the employee data lacked department information. By associating with historical data tables and using rule-based inference algorithms (such as determining the department based on the prefix of the employee ID), the missing values ​​are filled in, achieving a completion rate of 98%. Subsequently, the multi-source data was integrated into a unified dataset in a standardized format. Specifically, a standardized data model was designed to map financial budget data and organizational structure data to a unified CSV format. Fields included department ID, budget amount, number of employees, etc. Assuming the integrated dataset contained records from 100 departments, with an average budget of 500,000 yuan per department and an average number of 10 employees per department, the data was grouped and statistically analyzed using SQL scripts. The analysis showed that the correlation coefficient between budget and number of employees was 0.85, indicating a high correlation between the two. This analysis result can provide a basis for subsequent resource allocation optimization.

[0021] S102. Based on the integrated dataset, apply the support vector machine algorithm to classify the matching degree between multidimensional information. If the matching degree is lower than the preset threshold, mark it as a permission mismatch contradiction and determine the potential defect area.

[0022] A Support Vector Machine (SVM) algorithm is used to train and classify the integrated dataset, obtaining matching scores between multidimensional information. The matching scores are compared with a preset threshold to identify information pairs with scores below the threshold. If a matching score is below the threshold, the corresponding information pair is marked as having a permission mismatch. All classification results below the preset threshold are aggregated using these permission mismatch markers to obtain a set of conflicting information. The location of potential defective regions is determined based on the distribution of multidimensional information in the conflicting information set. A subset of the original multidimensional information corresponding to each potential defective region is extracted to obtain a defective subset dataset. The SVM algorithm is then used to reclassify and validate the defective subset dataset, obtaining a secondary matching score and determining the final defective region.

[0023] In the implementation of multidimensional information matching degree classification by integrating datasets and applying the support vector machine algorithm, the user permission data, access record data, and resource attribute data in the multidimensional information dataset are first standardized through data preprocessing. For example, the permission level values ​​are mapped to the range of 0 to 1. Assuming a user's permission level is 3, the standardized value is 0.6. At the same time, the access frequency data is normalized to the range of 0 to 1 based on the maximum value of 100 times. For example, a record with 50 times is 0.5. Next, a feature vector is constructed. The standardized permission level, access frequency, and resource sensitivity (assuming sensitivity is 0.8) are combined into a three-dimensional vector [0.6, 0.5, 0.8], and input into the support vector machine model for training. A radial basis function kernel (RBF) is used, with a penalty parameter C of 1.0 and a kernel parameter gamma of 0.1. The model is optimized through cross-validation, and the classification accuracy is 85%. Subsequently, a matching score is calculated, using the probability value output by a support vector machine as the matching degree. Assuming a user's matching degree with a resource is 0.3, and the preset threshold is 0.5, since 0.3 is lower than 0.5, it is marked as a permission mismatch. Further analysis of potential defect areas is conducted by clustering data points with matching degrees below the threshold using the K-means algorithm (K=3), dividing the data points into three categories. The centroid features of each category are calculated; for example, if the centroids of a certain category are [0.2, 0.3, 0.9], the analysis reveals that this category mainly focuses on access to highly sensitive resources, suggesting that the defect area may be due to an overly strict permission allocation strategy for highly sensitive resources. Finally, the analysis results are combined with business rules to automatically generate permission adjustment suggestions, such as reducing the permission level requirement for highly sensitive resources from 0.7 to 0.6, and storing these suggestions in the database for future system optimization.

[0024] S103. Obtain historical data within the potential defect area, use the random forest algorithm to predict the risk probability distribution, identify high-risk paths, and obtain a set of early warning indicators.

[0025] Step 1: Obtain historical data from the defective area in the target system. Clean the collected records, removing outliers and missing values ​​to obtain a cleaned basic dataset. Step 2: Using the cleaned basic dataset, calculate the risk probability using the random forest algorithm, analyze the probability distribution of each area, and determine the risk probability distribution result. Step 3: Based on the risk probability distribution result, perform path analysis on areas with high probabilities to identify potential high-risk paths, obtaining a set of high-risk paths. Step 4: Using the high-risk path set and combining it with anomaly patterns in historical data, extract key features related to the paths to determine the set of factors that may trigger risks. Step 5: After obtaining the factor set, if the frequency of a certain factor exceeds a preset threshold, it is classified as a high-risk factor, determining the high-risk factor list. Step 6: For the high-risk factor list, associate it with corresponding indicators in the early warning indicator set to generate early warning rules matching the high-risk paths, obtaining the final early warning indicator combination.

[0026] First, potential defect areas are identified through database queries and image recognition algorithms. For example, in manufacturing pipeline systems, historical sensor data is used to screen areas with a corrosion rate exceeding 0.15 as potential defect points. 10,000 sample data points are extracted from records of the past 5 years, including 15 features such as temperature, pressure, material thickness, and historical defect occurrence rate. After data cleaning and removal of missing values, the data is divided into a training set of 8,000 data points and a test set of 2,000 data points. Secondly, a random forest algorithm was used to construct the model. Using the scikit-learn library, the number of trees was set to n_estimators=500, max_depth=20, and training was performed using the sqrt method with random feature subsets. Each tree was sampled using bootstrap and split nodes using the Gini index. The model achieved an accuracy of 0.92 on the test set. Feature importance analysis showed that pressure contributed 0.28, thickness contributed 0.25, and temperature contributed 0.18. The predict_proba method was used to output the risk probability distribution for each sample. For example, a certain path had a probability of high risk (0.75), medium risk (0.20), and low risk (0.05), forming a continuous probability distribution curve with a mean of 0.68 and a variance of 0.12. Next, based on the probability distribution, high-risk paths were identified, with a threshold of probability greater than 0.7 defined as high risk. Three high-risk paths were identified out of ten. Combined with path connectivity analysis, these paths involved key nodes accounting for 60%, indicating a high potential for risk propagation. Finally, a set of early warning indicators was extracted. From the top five features of importance, pressure exceeding 150 Pa, thickness less than 2.5 mm, temperature fluctuation standard deviation greater than 5.0, historical defect rate greater than 0.10, and corrosion index greater than 0.20 were selected as early warning indicators. The reliability was verified by the model's OOB error rate of 0.08.

[0027] S104. Extract time series features from the set of early warning indicators, group similar defect patterns using clustering methods, and obtain classified risk groups.

[0028] Acquire all time series data in the early warning indicator set. Extract sequence features such as mean, standard deviation, slope, and number of peaks from each time series using statistical methods to obtain a feature vector set. Use the K-means clustering algorithm to cluster the feature vector set, determining the cluster centers of each time series and obtaining preliminary defect groups. Calculate group risk scores based on the distance from each cluster center to the origin. If the distance is greater than a preset threshold, the cluster is labeled as a high-risk group; otherwise, it is labeled as a low-risk group, resulting in risk-labeled classification groups. For all time series within a high-risk group, calculate the nearest neighbor distance. If the nearest neighbor distance is less than a preset threshold, the defect patterns within the group are considered highly similar, and the group label is retained. Sort all classification groups in descending order based on their group risk scores to obtain a risk group priority sequence. Output a list of classified risk groups based on the risk group priority sequence.

[0029] First, multidimensional time series data, such as vibration amplitude, temperature, and current, are collected from the set of equipment early warning indicators. Each series has 1000 sampling points. The tsfresh library is used to automatically extract comprehensive time series features, including mean, standard deviation, skewness, kurtosis, autocorrelation coefficient, and Fourier transform coefficient, generating a total of 794 feature vectors. Among them, about 200 highly correlated features are selected and retained through hypothesis testing. These features capture the statistical characteristics, frequency domain information, and nonlinear dynamics of the series, avoiding interference from the original high-dimensional noise. The extracted feature matrices were then standardized to eliminate dimensional differences. K-means clustering was then applied to group similar defect patterns, with K=5 clusters. The optimal K value was determined using the elbow rule, and the Euclidean distance was iteratively calculated to minimize the sum of squares within each cluster. Analysis revealed that one cluster, characterized by a high vibration standard deviation of 0.85 and low autocorrelation of 0.32, corresponded to bearing wear patterns. Another cluster, with a high temperature mean of 45.6 degrees Celsius and kurtosis of 3.8, corresponded to overheating defects. Other clusters represented large current fluctuations and multi-frequency peak patterns, thus identifying the inherent correlation between similar defects, such as wear often being accompanied by a 0.72-fold increase in vibration kurtosis. Finally, a risk score was calculated for each cluster based on the clustering results. For example, the Mahalanobis distance from the cluster center to the normal cluster was used as a risk indicator, resulting in low-risk groups (distance < 2.5, 40%), medium-risk groups (distance 2.5-5.0, 35%), and high-risk groups (distance > 5.0, 25%). The high-risk group corresponded to a historical failure rate of 85%.

[0030] S105. Select dominant features from the classified risk groups, use association rules to analyze the dependencies between features, and determine the defect propagation chain.

[0031] Classification data is obtained from risk groups, and features are extracted for each group to obtain a set of key features. Association rule mining is used to analyze the dependencies within the key feature set and determine the strength of associations between features. Based on the association strength data, the dependencies between features are analyzed to construct a preliminary path for defect propagation. Relationship analysis is performed on the preliminary path; if the association strength in a path is below a preset threshold, that path is discarded, resulting in a simplified propagation path. Based on the simplified propagation path, the core nodes of defect propagation are located, and the role of key features in the path is determined. Using the data on the roles of core nodes and key features, a complete chain of defect propagation is constructed, determining the final defect localization result.

[0032] After classifying the risk groups, the dominant feature is first selected from the classification results. Suppose we extract features from three risk groups in a software defect dataset through data analysis. We find that the "code complexity" feature has the highest weight in group A, reaching 0.85, while other features such as "test coverage" have only 0.3. Using a weighted scoring algorithm (weight calculation formula: feature importance = frequency * influence), "code complexity" is determined to be the dominant feature of group A. Next, we use association rule mining to analyze the dependencies between features. Using the Apriori algorithm, we set the minimum support to 0.2 and the minimum confidence to 0.6. We analyze the relationship between "code complexity" and "error rate" and "module coupling" in group A. We find that the rule "high code complexity → high error rate" has a support of 0.25 and a confidence of 0.75, indicating a strong correlation between the two. The mining process is completed by traversing frequent itemsets to generate candidate rules and verifying the confidence. Subsequently, the defect propagation chain was determined. Based on the association rule results and combined with causal analysis algorithms (such as Bayesian networks), the propagation path was derived: high code complexity (probability 0.8) leads to an increased error rate (probability 0.7), which in turn affects module coupling (probability 0.65), forming the chain "code complexity → error rate → module coupling". Historical data verified that this chain holds true in 80% of defect cases. To enhance logical coherence, business correlation was introduced, assuming that high module coupling further affects system maintenance costs (correlation coefficient 0.6). Regression analysis calculated that the increase in maintenance costs was 15%.

[0033] S106. If the defect propagation chain involves multiple departments, a real-time notification mechanism will be triggered to distribute early warning signals through a message queue and obtain response confirmation logs.

[0034] The system monitoring module captures the data flow of the defect propagation chain and performs real-time analysis on propagation paths involving multiple departments to obtain a preliminary assessment of the propagation scope. If the propagation scope assessment indicates involvement of multiple departments, a real-time notification mechanism is triggered, using message queue technology to send warning signals to relevant departments and determine the signal distribution status. Based on the signal distribution status, the response confirmation information of each department is recorded, obtaining the response time and confirmation content to form an initial response confirmation dataset. For this initial response confirmation dataset, the response efficiency of departmental collaboration is analyzed. If the response time exceeds a preset threshold, a reminder signal is sent again via message queue to determine if the reminder has been received. The reception status of the reminder signal is obtained, and combined with the confirmation log module, all response confirmation and reminder reception data are archived to determine the complete log record content. Based on the log record content, the collaboration of each department in the defect propagation chain is analyzed, and a support vector machine algorithm is used to classify and evaluate collaboration efficiency, obtaining the efficiency level of departmental collaboration. Based on the efficiency level, the distribution priority of the warning signal is adjusted, increasing the signal sending frequency for low-efficiency departments, and recording the adjusted distribution logs to complete the closed-loop processing of the entire notification process.

[0035] When the system detects a defect propagation chain involving multiple departments, it first queries the defect tracking log database for change records associated with the defect ID, counting the number of departments involved. If more than two departments are involved, a real-time notification mechanism is triggered. In practice, a graph traversal algorithm is used to construct the defect propagation graph. Defect reports are used as nodes, and code changes or data flows are used as directed edges. A depth-first search (DFS) algorithm traverses the set of reachable nodes starting from the initial defect node, counting the number of unique department identifiers (such as department codes). When the number is ≥3, it is determined to be cross-department propagation. The propagation risk score is calculated as the number of departments multiplied by the defect severity (severity values ​​range from 1 to 5). If the score is ≥10, a notification is immediately activated. The notification mechanism is implemented based on RabbitMQ message queues, encapsulating the warning signal into a JSON format message, including fields such as defect ID, list of involved departments, risk score, and timestamp. This message is pushed to the topic exchange "defect.alert" and bound to multiple department-specific queues (such as queue.deptA and queue.deptB) to ensure highly reliable message distribution. Consumer services are deployed in microservice instances across various departments. Upon receiving a message, they immediately execute an automated response process, including calling an internal API to record the response timestamp, generating an acknowledgment hash (using the SHA-256 algorithm to calculate the defect ID + department code + timestamp), and writing the acknowledgment log back to the central log cluster via the Kafka topic "defect.ack". The system monitors the arrival of acknowledgment logs in real time. If no acknowledgment is received from a department within 30 seconds, a secondary push is automatically triggered, and the acknowledgment rate (number of acknowledging departments / total number of departments involved) is calculated. When the acknowledgment rate is less than 0.8, it is escalated to a high-priority alarm and pushed to the operations center.

[0036] S107. Update the integrated dataset based on the response confirmation log, and use iterative optimization to adjust the classification threshold to obtain an optimized approval process model.

[0037] Initial information is obtained from response logs, and key fields are extracted using automated tools to obtain structured data. Based on this structured data, approval-related records are categorized to determine a preliminary category distribution. If an imbalance exists in the category distribution, data augmentation techniques are used to supplement a few categories, resulting in a balanced dataset. Based on this balanced dataset, a logistic regression algorithm is used to dynamically calculate classification thresholds and determine a suitable threshold range. The calculated threshold range is then matched and adjusted according to the business rules of the approval process to determine the final threshold standard. Based on this final threshold standard, the decision logic of the approval process is updated to obtain an optimized process configuration. This optimized process configuration is then loaded into the system for real-time application, and the improvement in approval efficiency is assessed.

[0038] First, data is extracted from the response confirmation logs. For example, 1000 approval records are collected, including response scores (simulated uniform distribution ranging from 0 to 1), normal distribution values ​​of feature variables such as feature1 and feature2, and actual approval results (pass rate of 70%). The dataset is read and updated using the pandas library, and the cumulative average response score is calculated as the cumulative mean of the new feature cumulative_score=response_score to integrate historical log information and enhance data stability. Next, an iterative optimization loop was entered. The dataset was divided into an 80% training set and a 20% test set using sklearn's `train_test_split` function. The `response_score` was used as the initial prediction probability. A grid search was performed with a threshold search range of 0.30 to 0.75 and a step size of 0.05. At each threshold, the predicted label was calculated (a probability greater than or equal to the threshold was considered a pass). The F1 score for binary classification was calculated using the `precision_recall_f1_support` function (averaged in binary form). The analysis showed that the F1 score at the initial threshold of 0.50 was approximately 0.82. After iterative comparison, the optimal threshold was adjusted to 0.55, corresponding to the highest F1 score of 0.8542, which improved the model performance by 3.2%. This avoided the problem of too many false positives caused by low thresholds and too many false negatives caused by high thresholds, thus balancing the precision (0.81) and recall (0.89). The final optimized approval process model was obtained. The optimal threshold of 0.55 was applied to classify the entire dataset. After optimization, the approval pass rate was reduced to 62.3%, which is more accurate in matching the actual business risk control needs compared to the initial threshold of 68.5%. This ensures that the model can achieve efficient approval decisions under the automated processing of information technology.

[0039] On the other hand, this embodiment provides an AI-powered intelligent approval system based on cross-departmental data fusion, which mainly includes: The data acquisition and integration module is used to collect multi-source data such as financial budgets and organizational structures through cross-departmental system interfaces, and uses data cleaning methods to remove noise and redundancy to obtain an integrated dataset in a unified format. The matching degree classification and labeling module is used to classify the matching degree between multidimensional information based on the integrated dataset using the support vector machine algorithm. If the matching degree is lower than the preset threshold, it is marked as a permission mismatch contradiction to identify potential defect areas. The risk prediction and early warning module is used to obtain historical data within potential defect areas, use the random forest algorithm to predict the risk probability distribution, identify high-risk paths, and obtain a set of early warning indicators. The time-series feature clustering module is used to extract time-series features from the early warning indicator set, and group similar defect patterns through clustering methods to obtain classified risk groups; The association rule mining module is used to select dominant features from the classified risk groups, analyze the dependencies between features using association rule mining, and determine the defect propagation chain. The real-time notification and distribution module is used to trigger a real-time notification mechanism if the defect propagation chain involves multiple departments, distribute warning signals through a message queue, and obtain response confirmation logs. The iterative optimization module is used to update the integrated dataset based on the response confirmation logs, and to iteratively adjust the classification threshold to obtain an optimized approval process model.

[0040] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0041] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0042] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI-powered intelligent approval method based on cross-departmental data fusion, characterized in that, include: Multi-source heterogeneous data is collected through cross-departmental system interfaces, and the multi-source heterogeneous data is cleaned and integrated to obtain an integrated dataset in a unified format. Based on the integrated dataset, a classification algorithm is applied to analyze the matching degree between information in different dimensions, and potential defect areas are determined by comparing the matching degree results with a preset threshold. Historical data within the potential defect area is obtained, a first machine learning algorithm is applied based on the historical data to predict the risk probability distribution, and high-risk paths are determined according to the risk probability distribution to obtain a set of early warning indicators. Time series features are extracted from the set of early warning indicators, and the time series features are clustered to group similar defect patterns to obtain classified risk groups. Dominant features are selected from the classified risk groups, and the dependency relationships between the dominant features are analyzed using association rule mining methods to determine the defect propagation chain; If the defect propagation chain involves multiple departments, a real-time notification mechanism is triggered to distribute warning signals to the multiple departments through a message queue and obtain response confirmation logs from each department. The integrated dataset is updated based on the response confirmation log, and the classification threshold of the approval process model is adjusted iteratively to obtain an optimized approval process model.

2. The method according to claim 1, characterized in that, The process involves collecting multi-source heterogeneous data through cross-departmental system interfaces, cleaning and integrating the multi-source heterogeneous data to obtain an integrated dataset in a unified format, including: Financial budget data and organizational structure data are collected through cross-departmental system interfaces to obtain the multi-source heterogeneous data; The multi-source heterogeneous data is cleaned to remove noise and redundancy, resulting in a refined dataset; The refined dataset is then standardized to obtain a unified format dataset. Extract the correspondence between financial budget data and organizational structure data from the unified format dataset, and determine whether their hierarchical mapping structure is consistent; If the hierarchical mapping structure is inconsistent, the correspondence is corrected by rule matching to obtain a dataset with consistent mapping. Based on the dataset with consistent mapping, an association table between the organization and the budget is constructed to form the integrated dataset.

3. The method according to claim 1, characterized in that, Based on the integrated dataset, a classification algorithm is applied to analyze the matching degree between information in different dimensions, and potential defect areas are determined by comparing the matching degree results with preset thresholds, including: The integrated dataset is trained and classified using the support vector machine algorithm to obtain matching scores that represent the matching degree between information of different dimensions; The matching score is compared with a preset threshold to identify information pairs whose matching score is lower than the preset threshold. The identified information pairs are marked as conflicting permission mismatches, and all marked conflicting information is aggregated to obtain a set of conflicting information. Based on the distribution of multidimensional information in the contradictory information set, potential defect areas are located; Extract the original data subset corresponding to the potential defect region to form a defect subset dataset; The support vector machine algorithm is used to reclassify and verify the defect subset, and the final defect region is determined based on the secondary matching score.

4. The method according to claim 1, characterized in that, Historical data within the potential defect area is obtained; a first machine learning algorithm is applied based on the historical data to predict the risk probability distribution; and high-risk paths are identified based on the risk probability distribution to obtain a set of early warning indicators, including: Historical data within the potential defect area is obtained and cleaned to obtain a basic dataset; Based on the aforementioned basic dataset, the random forest algorithm is applied to calculate the risk probability and generate the risk probability distribution results. Based on the risk probability distribution results, path analysis is performed on high-probability areas to identify a set of high-risk paths; By combining historical anomaly patterns, key features are extracted from the set of high-risk paths to form a set of factors that may trigger risks; Based on the comparison between the occurrence frequency of each factor in the factor set and the frequency threshold, high-risk factors are screened out and a list of high-risk factors is generated. The list of high-risk factors is matched with preset early warning indicators to generate a combination of early warning indicators corresponding to the high-risk path.

5. The method according to claim 1, characterized in that, Time-series features are extracted from the aforementioned set of early warning indicators, and clustering of these features is performed to group similar defect patterns, resulting in classified risk groups, including: Obtain time-series data from the set of early warning indicators; Statistical features are extracted from each of the time series data to form a feature vector set; The K-means clustering algorithm is used to cluster the feature vector set to obtain preliminary defect groups; Risk scores are calculated based on the distance from each cluster center to the origin, and each cluster is labeled as high-risk or low-risk based on the comparison between the risk scores and a preset distance threshold, resulting in risk-labeled classification groups. For the risk-labeled classification groups, the similarity of defect patterns is determined based on the nearest neighbor distance of their internal time series. All classification groups are sorted according to the risk scores to obtain a risk group priority sequence, and the classified risk group list is output.

6. The method according to claim 1, characterized in that, From the classified risk groups, dominant features are selected, and association rule mining methods are used to analyze the dependencies between the dominant features to determine the defect propagation chain, including: Key features are extracted from each risk group to form a set of key features; The association rule mining method is used to analyze the dependencies between features in the key feature set and determine the strength of the association between features. Construct a preliminary path for defect propagation based on the correlation strength; Paths with association strength below a preset threshold are removed from the initial path to obtain a simplified propagation path. The simplified propagation path is used to locate the core nodes and key features of defect propagation. Based on the roles of the core nodes and key features, a complete defect propagation chain is constructed.

7. The method according to claim 1, characterized in that, The integrated dataset is updated based on the response confirmation log, and the classification threshold of the approval process model is adjusted iteratively to obtain an optimized approval process model, including: Key fields are extracted from the response confirmation log to obtain structured approval data; The structured approval data is classified to obtain a preliminary category distribution; If the category distribution is unbalanced, data augmentation techniques are used to balance it, resulting in a balanced dataset. Based on the balanced dataset, the logistic regression algorithm is applied to dynamically calculate the classification threshold range; The classification threshold range is adjusted based on the business rules of the approval process to determine the final threshold standard; The decision logic of the approval process is updated based on the final threshold standard to obtain an optimized approval process model; The optimized approval process model is loaded into the system for application, and the approval efficiency is evaluated.

8. An AI-powered intelligent approval system based on cross-departmental data fusion, characterized in that: include: The data acquisition and integration module is used to acquire multi-source heterogeneous data through cross-departmental system interfaces, and to clean and integrate the multi-source heterogeneous data to obtain an integrated dataset in a unified format. The matching degree classification and labeling module is used to analyze the matching degree between different dimensions of information based on the integrated dataset using a classification algorithm, and to determine potential defect areas based on the comparison between the matching degree results and a preset threshold. The risk prediction and early warning module is used to acquire historical data within the potential defect area, apply a first machine learning algorithm based on the historical data to predict the risk probability distribution, and determine high-risk paths based on the risk probability distribution to obtain a set of early warning indicators. The time-series feature clustering module is used to extract time-series features from the set of early warning indicators, and to obtain classified risk groups by clustering the time-series features to group similar defect patterns. The association rule mining module is used to select dominant features from the classified risk groups and analyze the dependency relationships between the dominant features using association rule mining methods to determine the defect propagation chain. The real-time notification and distribution module is used to trigger a real-time notification mechanism when the defect propagation chain involves multiple departments, distribute early warning signals to the multiple departments through a message queue, and obtain response confirmation logs from each department. The iterative optimization module is used to update the integrated dataset based on the response confirmation log and iteratively optimize the classification threshold of the approval process model to obtain an optimized approval process model.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-7.

Images