Real-time data analysis method for enterprise digital operation based on multi-source data fusion

Abstract

The present invention discloses a real-time data analysis method for enterprise digital operations based on multi-source data fusion, which relates to the technical field of data processing and solves the following technical problems: first, it is difficult to analyze and predict the comprehensive benefit index of an enterprise; second, it is difficult to embed a knowledge graph into a vector space and construct a dynamic graph neural network; then, it is difficult to analyze abnormal equipment offsets, abnormal supply chain offsets, and customer activity, construct a risk matrix, and identify the propagation path of a decline in the comprehensive benefit index; it is difficult to dynamically allocate business link resources based on the analysis results of the risk matrix; and finally, it is difficult to visualize the results of data analysis. The present invention collects and pre-processes multi-source enterprise data, constructs a three-dimensional vector space and a dynamic graph neural network, combines a variety of technologies to analyze and predict enterprise benefits and risks, dynamically allocates resources for business links, and visualizes the results of data analysis.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and specifically is a real-time data analysis method for enterprise digital operations based on multi-source data fusion. Background Art

[0002] With the rapid development of the internet, the Internet of Things, and big data technologies, enterprises are generating massive amounts of multi-source data during their operations. However, traditional data processing and analysis technologies struggle to effectively integrate and utilize this multi-source data, failing to meet enterprises' needs for real-time data analysis. By addressing the data processing, analysis, and visualization needs of digital operations, supported by key technologies such as the Apache Flink stream processing framework, multi-source data fusion, knowledge graphs, deep learning, and visualization, we provide enterprises with comprehensive, real-time, and accurate analysis of digital operations data.

[0003] The following problems exist in the existing technology: first, it is difficult to analyze and predict the comprehensive benefit index of an enterprise; second, it is difficult to embed the knowledge graph into the vector space and build a dynamic graph neural network; then, it is difficult to analyze the abnormal offset of equipment, abnormal offset of supply chain and customer activity, build a risk matrix and identify the propagation path of the decline in the comprehensive benefit index; it is difficult to dynamically allocate business link resources based on the analysis results of the risk matrix; finally, it is difficult to visualize the results of data analysis. Summary of the Invention

[0004] To solve the problems existing in the above-mentioned prior art, the first aspect of the present invention provides a real-time data analysis method for enterprise digital operations based on multi-source data fusion, comprising the following modules:

[0005] To solve the above problems, the first aspect of the present invention provides a real-time data analysis method for enterprise digital operations based on multi-source data fusion, comprising the following steps:

[0006] S1: Real-time collection of multi-source operational data during enterprise operations, construction of multi-source operational data sets, and pre-processing;

[0007] S2: Extract features based on the preprocessed operational multi-source dataset and construct a three-dimensional vector space;

[0008] S3: Utilizes three-dimensional vector space and Apache Flink, combined with time series forecasting models, to analyze and predict the comprehensive efficiency index of enterprises. Based on the forecast results, it analyzes and evaluates the trends and potential risks of future enterprise operations. It also embeds knowledge graphs into the three-dimensional vector space and constructs a dynamic graph neural network.

[0009] S4: Analyze abnormal equipment deviations, abnormal supply chain deviations, and customer activity, build a risk matrix, and identify the propagation path of the decline in the comprehensive benefit index; use the analysis results of the risk transmission matrix to dynamically allocate business link resources;

[0010] S5: Use the dashboard to visualize the company's future operating trends, comprehensive benefit index, and the path of the decline in the comprehensive benefit index.

[0011] Preferably, in step S1, the operating multi-source data set includes: equipment operation data, enterprise financial data, customer behavior data, human resources data and external data.

[0012] Preferably, the step S2 comprises the following steps:

[0013] Through Apache Flink, we access pre-processed operational multi-source data sets and extract the maximum, minimum, and average values ​​of equipment operation data, enterprise financial data, customer behavior data, and human resources data as statistical feature values.

[0014] Extract the difference between capital inflow and outflow and capital turnover rate from corporate financial data; extract customer retention rate and customer activity from customer behavior data; extract employee turnover rate from human resources data;

[0015] The feature values ​​extracted from external data include: the busyness and traffic congestion level of the enterprise's geographical area;

[0016] Based on the results of feature extraction of the above multi-source operational data sets, the feature vectors are combined into equipment status feature vectors, enterprise financial feature vectors, customer behavior feature vectors, human resource feature vectors and external feature vectors;

[0017] The five constructed eigenvectors are combined to form a comprehensive eigenvector matrix. Principal component analysis (PCA) is used to reduce the dimensionality of the fused eigenvectors, retaining 95% of the variance of the principal components. The eigenvectors are then normalized and updated in real time within the Apache Flink stream processing framework.

[0018] The comprehensive eigenvector matrix is ​​projected into three-dimensional space, and a right-hand coordinate system is used to construct a three-dimensional space vector, where the x-axis is the equipment status eigenvector, the y-axis is the enterprise financial eigenvector and the customer behavior eigenvector, and the z-axis is the human resources eigenvector and the external eigenvector; three linearly independent vectors are selected as the basis, marked as i, j, and k, and the order of the three directions is determined according to the right-hand rule.

[0019] Preferably, in step S3, the three-dimensional vector space and Apache Flink are used in combination with a time series prediction model to analyze and predict the comprehensive benefit index of the enterprise, including the following steps:

[0020] By leveraging Apache Flink's stream processing capabilities, we can visualize vector changes in a three-dimensional vector space. Based on the position and change trends of vectors in the three-dimensional vector space, we can assess the cost-effectiveness of enterprise operations in real time.

[0021] Formula for calculating comprehensive benefit index:

[0022] Get the comprehensive benefit index Q of the enterprise; among them, is the modulus of the current state vector, which represents the position in the three-dimensional vector space; E(f) represents the information entropy of the vector distribution of the current state vector; f represents the current state eigenvector; S t represents the profit of the enterprise at time t; C t is the enterprise's cost at time t; α, β, and γ represent weight coefficients, which are 0.4, 0.3, and 0.3 respectively;

[0023] Using a long-short-term memory (LSTM) model to predict future comprehensive benefit indices, the multi-source historical operational data set is fed into the LSTM model for training. During training, the data is fed into the LSTM model in chronological order, learning the data characteristics and change patterns at different time points and establishing a mapping relationship from the past to the present and then to the future.

[0024] The trained LSTM model is deployed in the Apache Flink environment to perform real-time predictions on newly incoming multi-source operational data. When the real-time multi-source operational data enters the Flink stream processing pipeline, the output prediction result is the future comprehensive benefit index based on the current input data and the comprehensive benefit index formula.

[0025] Preferably, the step S3 of analyzing and evaluating the trend and potential risks of the enterprise's future operations based on the forecast results includes the following steps:

[0026] The comprehensive benefit index at different time points is predicted to be drawn into a time series graph, where the horizontal axis represents time and the vertical axis represents the comprehensive benefit index;

[0027] Using the linear regression method, with time as the independent variable and the comprehensive benefit index as the dependent variable, a straight line is fitted. The slope of the fitted line is used to determine whether the comprehensive benefit index is increasing or decreasing over time. When Q shows an upward trend, it indicates that the company's future operating trend is positive; when Q shows a downward trend, it indicates that the company's future operating trend is negative.

[0028] Based on the historical comprehensive benefit index, different thresholds are set: when Q>0.8, the enterprise's operating status is excellent; when 0.4≤Q≤0.8, the enterprise's operating status is good; when Q<0.4, the enterprise's operating status is dangerous;

[0029] The LSTM model is used to predict the future comprehensive benefit index and determine whether there are potential risks in the future enterprise operation status based on the set threshold. When the future comprehensive benefit index is less than 0.4, the potential risk is identified; otherwise, there is no potential risk.

[0030] Preferably, in step S3, embedding the knowledge graph into a three-dimensional vector space to construct a dynamic graph neural network includes the following steps:

[0031] The knowledge graph is embedded into a three-dimensional vector space. In the knowledge graph, nodes represent entities, including devices, customers, and employees. Edges represent relationships between entities, including the relationship between devices and customers, the relationship between devices and employees, and the relationship between customers and employees.

[0032] The TransE model is used to embed entities into a three-dimensional vector space. The time attribute is introduced, and time is used as a dimension in the graph. The timestamp is converted into a sine or cosine periodic vector. After being concatenated with the three-dimensional space vector, the dimensionality is reduced to three dimensions through a fully connected layer to form a joint spatiotemporal embedding, resulting in a node representation containing spatiotemporal information.

[0033] The EvolveGCN model is used to evolve graph convolutional network parameters through LSTM to capture dynamic changes in graph structure.

[0034] Combined with the Apache Flink stream processing framework, incremental graph structure data is synchronized every 5 minutes. By leveraging Flink's stream processing capabilities, new data is updated to the knowledge graph, and the parameters of the dynamic graph neural network are updated in real time through the EvolveGCN model.

[0035] Preferably, the step S4 analyzes the abnormal device offset, the abnormal supply chain offset, and the customer activity, constructs a risk matrix, and identifies the propagation path of the decline in the comprehensive benefit index, including the following steps:

[0036] The reference vector is obtained by calculating the average value of each characteristic parameter in the normal equipment operation characteristic vector as the reference value; the offset between the current equipment state vector and the reference vector is calculated in real time using the Euclidean distance formula to obtain the equipment abnormality offset;

[0037] Combining the external feature vector, enterprise financial feature vector and human resource feature vector, the supply chain status feature vector is generated. Similarly, the calculation method of equipment abnormal offset is used to obtain the supply chain abnormal offset.

[0038] Obtain client node activity data, calculate the difference between the activity of the previous cycle and the activity of the current cycle, and then divide it by the activity of the previous cycle to obtain the activity decline rate;

[0039] The device anomaly offset, supply chain anomaly offset, customer activity, and timestamp are used as input signals. The graph attention network model is introduced to construct a dynamic graph neural network, and the graph attention mechanism is used to automatically learn the relationship between nodes.

[0040] The input signal is fed into the constructed dynamic graph neural network for training. The graph attention mechanism is used to calculate the impact of the critical path on the potential risk probability and output the potential risk probability.

[0041] Based on the output of the dynamic graph neural network, the risk impact of different nodes and paths is evaluated, and a risk transmission matrix is ​​constructed. The elements in the matrix represent the intensity of risk transmission between different nodes or three business links; the three business links include: equipment operation business link, customer activity business link, and supply chain business link;

[0042] When the comprehensive benefit index shows a downward trend, the risk transmission matrix is ​​used to identify the risk transmission path and the degree of risk impact.

[0043] Preferably, the step S4 utilizes the analysis results of the risk transmission matrix to dynamically allocate business link resources, including the following steps:

[0044] The characteristic vector weighting method is used to obtain the contribution weights of the three business links of equipment operation, customer activity, and supply chain to the comprehensive benefit index;

[0045] According to the intensity of risk transmission between the three business links in the risk transmission matrix, calculate the sensitivity of the three business links in the risk transmission process respectively;

[0046] The graph attention mechanism automatically learns the relationship between nodes, calculates the impact of the critical path on the potential risk probability, and determines the influence ranking of the three business links;

[0047] Combining the contribution weights of the three business links to the comprehensive benefit index, as well as the risk transmission sensitivity and the influence ranking derived from the graph attention network, a resource allocation weight function is constructed to determine the weight coefficient of resource allocation;

[0048] After normalizing the obtained resource allocation weight coefficients of the three business links, the three business links are sorted respectively to determine the resource allocation response priorities of the three business links.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] This invention collects multi-source data sets during enterprise operations in real time and pre-processes them using the Apache Flink stream processing framework and lightweight geocoding architecture. It utilizes three-dimensional vector space and real-time data analysis technology, combined with a time series prediction model, to accurately analyze and predict the comprehensive efficiency index of enterprises, helping enterprises understand future operating trends in advance.

[0051] This method analyzes abnormal equipment and supply chain offsets, as well as customer activity, to construct a risk matrix and identify the paths through which the overall benefit index declines, enabling enterprises to promptly identify potential risks. By dynamically allocating resources across business segments, resource utilization efficiency is improved, ensuring a swift and effective response to risks when they occur, safeguarding the stability and overall effectiveness of enterprise operations.

[0052] The present invention uses a dashboard to visually display important information such as the company's future operating trends, comprehensive benefit index, and the propagation path of the decline in the comprehensive benefit index, so that company managers can understand the company's operating status intuitively and clearly. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0054] Figure 1 is a flow chart of the system method of the present invention;

[0055] Figure 2 Schematic diagram of the prediction model flow in the present invention. DETAILED DESCRIPTION

[0056] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0057] See also Figure 1-Figure 2 The first embodiment of the present invention provides a real-time data analysis method for enterprise digital operations based on multi-source data fusion, comprising the following steps:

[0058] S1: Real-time collection of multi-source operational data during enterprise operations, construction of multi-source operational data sets, and pre-processing;

[0059] S2: Extract features based on the preprocessed operational multi-source dataset and construct a three-dimensional vector space;

[0060] S3: Utilizes three-dimensional vector space and Apache Flink, combined with time series forecasting models, to analyze and predict the comprehensive efficiency index of enterprises. Based on the forecast results, it analyzes and evaluates the trends and potential risks of future enterprise operations. It also embeds knowledge graphs into the three-dimensional vector space and constructs a dynamic graph neural network.

[0061] S4: Analyze abnormal equipment deviations, abnormal supply chain deviations, and customer activity, build a risk matrix, and identify the propagation path of the decline in the comprehensive benefit index; use the analysis results of the risk transmission matrix to dynamically allocate business link resources;

[0062] S5: Use the dashboard to visualize the company's future operating trends, comprehensive benefit index, and the path of the decline in the comprehensive benefit index.

[0063] Specifically, through various sensors and information system interfaces, multi-source operational datasets are collected in real time during enterprise operations, including equipment operation data, corporate financial data, customer behavior data, human resources data, and external data. Apache Flink connectors are used to integrate with these data sources, enabling real-time data access. The collected data is cleaned using the Apache Flink stream processing framework to remove duplicate, erroneous, or missing data. Window operations and other features are used to transform the data, such as unit conversion for equipment operation data. Furthermore, a lightweight geocoding architecture is used to encode data containing geographic location information, unifying the data format and timestamps for subsequent analysis. Feature extraction is performed on the preprocessed multi-source operational datasets, selecting features that reflect key operational information. A three-dimensional vector space is then constructed and mapped to this space, with each dimension representing a different feature or feature combination, thus achieving a multi-dimensional representation of the enterprise's operational data. This constructed three-dimensional vector space, combined with real-time data analysis techniques and time series prediction models, allows for analysis and forecasting of the enterprise's comprehensive efficiency index, providing a preview of the overall operational efficiency trends for the future. Based on the prediction results, the company's future operational trends and potential risks are analyzed and assessed. The knowledge graph is embedded in a vector space to construct a dynamic graph neural network. This network and the graph attention mechanism are used to analyze equipment anomaly offsets and customer activity, construct a risk matrix, identify the propagation paths of declining comprehensive benefit indexes, and determine the transmission paths of risks. The results of the risk transmission matrix analysis are used to analyze resource allocation priorities for the three business segments of equipment, customers, and supply chain. This enables real-time monitoring and optimized resource allocation for the three key business segments of the company's digital operations—equipment operation, customer activity, and supply chain—to improve operational efficiency. Using visualization tools such as dashboards, the analysis results, including future operational trends, comprehensive benefit index, and propagation paths of declining comprehensive benefit indexes, are displayed in intuitive charts and graphs, enabling managers to quickly understand the company's operational status and make timely decisions.

[0064] In one embodiment, in step S1, the operating multi-source data set includes: equipment operation data, enterprise financial data, customer behavior data, human resources data and external data.

[0065] Specifically, real-time multi-source operational data is collected within the current cycle. This multi-source data collection is based on a specific time span, such as hours, days, weeks, months, quarters, or years, determined based on the company's business operations model. Equipment operation data collection involves installing temperature sensors, vibration sensors, production counter sensors, timers, and energy monitors on key locations of production equipment. The collected equipment operation data includes equipment temperature, vibration, production quantity, production time, and energy consumption. Financial data collection involves integrating with the company's existing financial software to obtain real-time financial data such as total sales, sales volume, and cash inflows and outflows through APIs or middleware. A data synchronization mechanism is established to ensure the timeliness and accuracy of financial data. During the data synchronization process, data is verified and compared to avoid duplication or errors. Primary key matching and timestamp comparison are used to check the consistency of sales data across different systems, allowing synchronization issues to be identified and resolved promptly. Historical financial data is also regularly backed up for recovery and analysis when needed. Customer behavior data collection includes: embedding analytical tool code in the company's official website and mobile applications to monitor customer access behavior, and tracking in real time information such as the source of the customer's access, visited pages, dwell time, number of clicks, etc., thereby calculating key indicators such as the number of customer visits, visit duration, and number of interactions. When a customer opens a product page on the company's website, information such as the customer's visit time and the page content viewed is recorded to obtain the customer's dwell time and interaction behavior on that page. Human resources data collection includes: integrating with the company's human resources management system to obtain data such as the number of employees and employee work hours through a data interface; external data collection includes: using geographic information system technology to collect geographic location information of the company's area, including the company's latitude and longitude coordinates, the busyness of the company's geographical area, and the level of traffic congestion.

[0066] In one embodiment, step S2 includes the following steps:

[0067] Through Apache Flink, we access pre-processed operational multi-source data sets and extract the maximum, minimum, and average values ​​of equipment operation data, enterprise financial data, customer behavior data, and human resources data as statistical feature values.

[0068] Extract the difference between capital inflow and outflow and capital turnover rate from corporate financial data; extract customer retention rate and customer activity from customer behavior data; extract employee turnover rate from human resources data;

[0069] The feature values ​​extracted from external data include: the busyness and traffic congestion level of the enterprise's geographical area;

[0070] Based on the results of feature extraction of the above multi-source operational data sets, the feature vectors are combined into equipment status feature vectors, enterprise financial feature vectors, customer behavior feature vectors, human resource feature vectors and external feature vectors;

[0071] The five constructed eigenvectors are combined to form a comprehensive eigenvector matrix. Principal component analysis (PCA) is used to reduce the dimensionality of the fused eigenvectors, retaining 95% of the variance of the principal components. The eigenvectors are then normalized and updated in real time within the Apache Flink stream processing framework.

[0072] The comprehensive eigenvector matrix is ​​projected into three-dimensional space, and a right-hand coordinate system is used to construct a three-dimensional space vector, where the x-axis is the equipment status eigenvector, the y-axis is the enterprise financial eigenvector and the customer behavior eigenvector, and the z-axis is the human resources eigenvector and the external eigenvector; three linearly independent vectors are selected as the basis, marked as i, j, and k, and the order of the three directions is determined according to the right-hand rule.

[0073] Specifically, Apache Flink's data source interfaces, such as KafkaConsumer and SocketStream, are used to access preprocessed real-time data streams, including equipment operation data, corporate financial data, customer behavior data, human resources data, and external data. Using Flink's built-in or custom functions, the maximum, minimum, and average values ​​of data such as equipment temperature, vibration, production quantity, production time, and energy consumption are calculated in real time. These statistical feature values ​​are then combined into a device status feature vector. Similarly, Flink functions are used to calculate the maximum, minimum, and average values ​​of total sales, sales quantity, and cash inflow and outflow. Furthermore, the capital turnover rate is calculated by dividing total sales by the difference between cash inflow and outflow, divided by 2. This difference, the capital turnover rate, and the statistical feature values ​​are combined to form a corporate financial feature vector. The maximum, minimum, and average values ​​of the number of customer visits, visit duration, and number of interactions are extracted. In addition, the customer retention rate is calculated by calculating the difference between the total number of customers at the end of the current cycle and the total number of new customers added during the current cycle, which is equal to the total number of customers at the end of the previous cycle. Customer activity is calculated using the RFE model based on the sum of a customer's visit time R, visit frequency F, and number of interactions E at the end of the current cycle. The statistical eigenvalues, customer retention rate, and customer activity are combined to form a customer behavior feature vector. The maximum, minimum, and average values ​​of the number of employees and employee working hours are calculated. The employee turnover rate is calculated by dividing the number of employee resignations by the total number of employees and combined with the statistical eigenvalues ​​to form a human resource feature vector. Eigenvalues ​​such as latitude and longitude coordinates, area code, busyness, and traffic congestion levels are extracted from the geographic location information and combined to form an external feature vector. The constructed device status feature vector, enterprise financial feature vector, customer behavior feature vector, human resource feature vector, and external feature vector are combined to form a comprehensive feature vector matrix. In Flink, the PCA algorithm in the FlinkML library can be used for dimensionality reduction. To retain 95% of the principal components, the number of principal components to retain is determined based on the size of the eigenvalues ​​and their cumulative contribution. Normalize the fused feature vectors to keep them within the range [0, 1]. This can be done using methods such as MinMaxScaler. In the Apache Flink stream processing framework, real-time feature vector updates can be achieved by periodically recalculating the feature vectors and performing dimensionality reduction and normalization during data stream processing. In Flink, these feature vectors can be collected and combined using functions such as CollectList. Principal component analysis (PCA) is used for dimensionality reduction, projecting the combined feature vectors into three-dimensional space. First, the covariance matrix of the combined feature vectors is calculated. The eigenvalues ​​and eigenvectors of this matrix are then calculated. The eigenvectors corresponding to the three largest eigenvalues ​​are selected as principal components. The original data is projected onto these three principal components to obtain a three-dimensional vector.A three-dimensional vector space is constructed using a right-handed coordinate system. Three linearly independent vectors are selected as the basis, labeled i, j, and k. Three main, mutually independent characteristic directions can be selected from the three-dimensional space vectors after dimensionality reduction as the basis, and then the order of the three directions can be determined according to the right-hand rule to ensure the orthogonality and directional consistency of the coordinate system. Assume that the first principal component direction of the equipment state feature vector after dimensionality reduction is the main direction, and use it as the x-axis direction, and select the corresponding basis vector i; then, select a direction that is linearly independent of the x-axis from the combined direction of the enterprise financial feature vector and the customer behavior feature vector as the y-axis direction, and the basis vector is j; finally, determine a direction that is linearly independent of both the x-axis and y-axis from the combined direction of the human resources feature vector and the external feature vector as the z-axis direction, and the basis vector is k. According to the right-hand rule, the order of the three directions is determined to be x, y, and z, and a right-handed three-dimensional vector space with i, j, and k as the basis is constructed.

[0074] In one embodiment, step S3 utilizes a three-dimensional vector space and Apache Flink in combination with a time series prediction model to analyze and predict the comprehensive benefit index of an enterprise, including the following steps:

[0075] By leveraging Apache Flink's stream processing capabilities, we can visualize vector changes in a three-dimensional vector space. Based on the position and change trends of vectors in the three-dimensional vector space, we can assess the cost-effectiveness of enterprise operations in real time.

[0076] Formula for calculating comprehensive benefit index:

[0077] Get the comprehensive benefit index Q of the enterprise; among them, is the modulus of the current state vector, which represents the position in the three-dimensional vector space; E(f) represents the information entropy of the vector distribution of the current state vector; f represents the current state eigenvector; S t represents the profit of the enterprise at time t; C t is the enterprise's cost at time t; α, β, and γ represent weight coefficients, which are 0.4, 0.3, and 0.3 respectively;

[0078] Using a long-short-term memory (LSTM) model to predict future comprehensive benefit indices, the multi-source historical operational data set is fed into the LSTM model for training. During training, the data is fed into the LSTM model in chronological order, learning the data characteristics and change patterns at different time points and establishing a mapping relationship from the past to the present and then to the future.

[0079] The trained LSTM model is deployed in the Apache Flink environment to perform real-time predictions on newly incoming multi-source operational data. When the real-time multi-source operational data enters the Flink stream processing pipeline, the output prediction result is the future comprehensive benefit index based on the current input data and the comprehensive benefit index formula.

[0080] Specifically, Apache Flink's stream processing capabilities are used to access real-time data streams from data sources such as Kafka and Sockets. Flink's DataStream API is used to perform transformations and aggregations on real-time data, calculating changes in vectors in three-dimensional space. This includes calculating the values ​​of device state vectors and enterprise financial vectors at various points in time. Using Flink's Table API or SQL API, the calculation results are output to visualization tools such as Grafana and Echarts to visualize the position and trend of vectors in three-dimensional space. A three-dimensional line chart can be drawn to display the trajectory of vector changes over time, visually reflecting the real-time cost-effectiveness of enterprise operations. In Flink, the comprehensive benefit index Q is calculated using the comprehensive benefit index formula. Here, w1, w2, and w3 are weighted coefficients of 0.4, 0.3, and 0.3, respectively, and are dynamically adjusted based on actual application conditions and historical data. Historical operational data sets from multiple sources, including equipment operation data, enterprise financial data, and customer behavior data, are collected, cleaned, normalized, and converted into sample or time series formats suitable for LSTM model input. Take the data from the past n time steps as input and predict the comprehensive benefit index for the n+1th time step. Use a deep learning framework such as TensorFlow, Keras, or PyTorch to build an LSTM model. Input the prepared data into the LSTM model for training. During training, the data is input in chronological order to learn the data characteristics and change patterns at different time points, establishing a mapping relationship from the past to the present and to the future. Use appropriate optimization algorithms, such as Adam, and suitable loss functions, such as mean squared error, to optimize the model training process. Export the trained LSTM model and deploy it in the Apache Flink environment. The LSTM model can be invoked using Flink's machine learning library or through custom functions. When real-time operational multi-source data enters the Flink stream processing pipeline, it is preprocessed. Then, the current input data and the comprehensive benefit index formula are input as features into the LSTM model to predict the future comprehensive benefit index. The predicted result is the future comprehensive benefit index.

[0081] In one embodiment, the step S3 of analyzing and evaluating the trends and potential risks of the enterprise's future operations based on the forecast results includes the following steps:

[0082] The comprehensive benefit index at different time points is predicted to be drawn into a time series graph, where the horizontal axis represents time and the vertical axis represents the comprehensive benefit index;

[0083] Using the linear regression method, with time as the independent variable and the comprehensive benefit index as the dependent variable, a straight line is fitted. The slope of the fitted line is used to determine whether the comprehensive benefit index is increasing or decreasing over time. When Q shows an upward trend, it indicates that the company's future operating trend is positive; when Q shows a downward trend, it indicates that the company's future operating trend is negative.

[0084] Based on the historical comprehensive benefit index, different thresholds are set: when Q>0.8, the enterprise's operating status is excellent; when 0.4≤Q≤0.8, the enterprise's operating status is good; when Q<0.4, the enterprise's operating status is dangerous;

[0085] The LSTM model is used to predict the future comprehensive benefit index and determine whether there are potential risks in the future enterprise operation status based on the set threshold. When the future comprehensive benefit index is less than 0.4, the potential risk is identified; otherwise, there is no potential risk.

[0086] Specifically, organize the predicted comprehensive benefit index at different time points into a dataset in chronological order, with time on the horizontal axis and the comprehensive benefit index on the vertical axis. Use data visualization tools such as Matplotlib, Seaborn, or Plotly to plot the time series graph. In Flink, export the prediction results to an external storage system such as HDFS or MySQL. Then, use a scripting language such as Python to read and plot the data. Observe the trend of the time series graph. If the curve shows an overall upward trend, meaning the Q value gradually increases over time, it indicates a positive future for the company's operations. If the curve shows a downward trend, meaning the Q value gradually decreases over time, it indicates a negative future for the company's operations. Collect historical time series data on the comprehensive benefit index and organize it into a dataset containing time as an independent variable and Q value as a dependent variable. Use a linear regression algorithm, such as the LinearRegression model from the scikit-learn library, with time as the independent variable and Q value as the dependent variable, to fit the data. A straight line is fitted, representing the linear trend of the Q value over time. Obtain the slope k of the fitted line. If k > 0, Q is on an upward trend, indicating a positive future for the company's operations. If k < 0, Q is on a downward trend, indicating a negative future for the company's operations. Based on historical comprehensive benefit index data, the company's operating status is categorized into three levels: When Q > 0.8, the company's operating status is excellent. When 0.4 ≤ Q ≤ 0.8, the company's operating status is good. When Q < 0.4, the company's operating status is critical. The thresholds are dynamically adjusted based on historical data, experimental analysis, or practical application. The trained LSTM model is used to predict the future comprehensive benefit index. The predicted future Q value is compared with the set threshold. If the predicted future comprehensive benefit index is less than 0.4 (i.e., Q < 0.4), a potential risk is identified, indicating that the company's future operating status may be critical. Otherwise, no potential risk exists, and the company's operating status is relatively good.

[0087] In one embodiment, step S3 embeds the knowledge graph into a three-dimensional vector space to construct a dynamic graph neural network, including the following steps:

[0088] The knowledge graph is embedded into a three-dimensional vector space. In the knowledge graph, nodes represent entities, including devices, customers, and employees. Edges represent relationships between entities, including the relationship between devices and customers, the relationship between devices and employees, and the relationship between customers and employees.

[0089] The TransE model is used to embed entities into a three-dimensional vector space. The time attribute is introduced, and time is used as a dimension in the graph. The timestamp is converted into a sine or cosine periodic vector. After being concatenated with the three-dimensional space vector, the dimensionality is reduced to three dimensions through a fully connected layer to form a joint spatiotemporal embedding, resulting in a node representation containing spatiotemporal information.

[0090] The EvolveGCN model is used to evolve graph convolutional network parameters through LSTM to capture dynamic changes in graph structure.

[0091] Combined with the Apache Flink stream processing framework, incremental graph structure data is synchronized every 5 minutes. By leveraging Flink's stream processing capabilities, new data is updated to the knowledge graph, and the parameters of the dynamic graph neural network are updated in real time through the EvolveGCN model.

[0092] Specifically, a knowledge graph is constructed, where nodes represent entities, including devices, customers, and employees; edges represent relationships between entities, including those between devices and customers, between devices and employees, and between customers and employees. The TransE model is used to embed entities into a three-dimensional vector space. The core idea of ​​the TransE model is to map entities and relationships in the knowledge graph into a low-dimensional vector space, representing the semantic connection between entities and relationships through the distance between vectors. A time attribute is introduced to form a joint spatiotemporal embedding. Time is used as a dimension in the graph, and timestamps are converted into sine or cosine periodic vectors. After concatenating them with spatial vectors, the dimensionality is reduced to three dimensions through a fully connected layer to form a joint spatiotemporal embedding, resulting in a node representation that incorporates spatiotemporal information. The EvolveGCN model is used to evolve the parameters of the graph convolutional network (GCN), and LSTM is used to evolve the parameters to capture dynamic changes in the graph structure. The EvolveGCN model adapts the training of the GCN model along the time dimension, using a RNN to obtain the GCN parameters, thereby capturing the dynamic information of the graph sequence. Specifically, at each time step, the LSTM is used to adjust the GCN model parameters to improve model adaptability. Integrating with the Apache Flink stream processing framework, we synchronize incremental graph data every 5 minutes. Leveraging Flink's stream processing capabilities, we update new data into the knowledge graph and use the EvolveGCN model to update the graph neural network parameters in real time.

[0093] In one embodiment, the step S4 analyzes the abnormal device offset, the abnormal supply chain offset, and the customer activity, constructs a risk matrix, and identifies the propagation path of the decline in the comprehensive benefit index, including the following steps:

[0094] The reference vector is obtained by calculating the average value of each characteristic parameter in the normal equipment operation characteristic vector as the reference value; the offset between the current equipment state vector and the reference vector is calculated in real time using the Euclidean distance formula to obtain the equipment abnormality offset;

[0095] Combining the external feature vector, enterprise financial feature vector and human resource feature vector, the supply chain status feature vector is generated. Similarly, the calculation method of equipment abnormal offset is used to obtain the supply chain abnormal offset.

[0096] Obtain client node activity data, calculate the difference between the activity of the previous cycle and the activity of the current cycle, and then divide it by the activity of the previous cycle to obtain the activity decline rate;

[0097] The device anomaly offset, supply chain anomaly offset, customer activity, and timestamp are used as input signals. The graph attention network model is introduced to construct a dynamic graph neural network, and the graph attention mechanism is used to automatically learn the relationship between nodes.

[0098] The input signal is fed into the constructed dynamic graph neural network for training. The graph attention mechanism is used to calculate the impact of the critical path on the potential risk probability and output the potential risk probability.

[0099] Based on the output of the dynamic graph neural network, the risk impact of different nodes and paths is evaluated, and a risk transmission matrix is ​​constructed. The elements in the matrix represent the intensity of risk transmission between different nodes or three business links; the three business links include: equipment operation business link, customer activity business link, and supply chain business link;

[0100] When the comprehensive benefit index shows a downward trend, the risk transmission matrix is ​​used to identify the risk transmission path and the degree of risk impact.

[0101] Specifically, the characteristic parameter data in the normal equipment operation characteristic vector are collected, the average value of each characteristic parameter is calculated, and these average values ​​are combined into a vector as the reference vector. During the equipment operation process, the current equipment state vector is obtained in real time. This vector contains the same characteristic parameters as the reference vector. Then, the Euclidean distance formula is used to calculate the distance between the current equipment state vector and the reference vector, which is the equipment abnormality offset. The Euclidean distance formula is: let the reference vector be (a1, a2, ..., an) and the current equipment state vector be (b1, b2, ..., bn), then the equipment abnormality offset is Data such as external feature vectors, enterprise financial feature vectors, and human resource feature vectors are collected, integrated, and processed to generate a supply chain status feature vector. Once the supply chain status feature vector is obtained, the supply chain anomaly offset is calculated using the same Euclidean distance formula. Customer node activity data is obtained, and the activity values ​​for the previous and current cycles are tallied. The difference between the activity values ​​of the two cycles is calculated and then divided by the activity value of the previous cycle. The result is the activity decline rate, which measures the degree of change in customer activity. The calculated device anomaly offset, supply chain anomaly offset, and customer activity, along with the corresponding timestamps, are used as input signals. The graph attention network model automatically learns complex relationships and interactions between nodes. By constructing a dynamic graph neural network and incorporating the graph attention mechanism, the model can better capture the dynamic connections and risk transmission paths between different business links and nodes. These input signals are fed into the constructed dynamic graph neural network for training. During training, the graph attention mechanism enables the model to automatically learn and calculate the impact of critical paths on potential risk probabilities, ultimately outputting potential risk probabilities and providing a quantitative basis for risk assessment. Based on the output of the dynamic graph neural network, the risk impact of different nodes and paths is assessed to determine which nodes and paths play a key role in risk propagation. A matrix is ​​constructed, with elements representing the intensity of risk transmission between different nodes or three business segments. These segments include equipment operation, customer activity, and supply chain. This matrix visually illustrates the transmission of risk across various business segments and nodes. When the comprehensive benefit index shows a downward trend, the constructed risk transmission matrix analyzes the risk transmission intensity of each node and path in the matrix to determine the direction of risk propagation, critical paths, and the degree of risk impact, thereby helping companies to proactively implement effective risk control measures. Critical paths are those paths within the risk transmission matrix constructed by the dynamic graph neural network that have a significant impact on the probability of potential risks. The nodes and edges along these paths represent key links in risk transmission and have a direct or significant impact on the decline in the comprehensive benefit index.

[0102] In one embodiment, the step S4 dynamically allocates business link resources using the analysis results of the risk transmission matrix, including the following steps:

[0103] The characteristic vector weighting method is used to obtain the contribution weights of the three business links of equipment operation, customer activity, and supply chain to the comprehensive benefit index;

[0104] According to the intensity of risk transmission between the three business links in the risk transmission matrix, calculate the sensitivity of the three business links in the risk transmission process respectively;

[0105] The graph attention mechanism automatically learns the relationship between nodes, calculates the impact of the critical path on the potential risk probability, and determines the influence ranking of the three business links;

[0106] Combining the contribution weights of the three business links to the comprehensive benefit index, as well as the risk transmission sensitivity and the influence ranking derived from the graph attention network, a resource allocation weight function is constructed to determine the weight coefficient of resource allocation;

[0107] After normalizing the obtained resource allocation weight coefficients of the three business links, the three business links are sorted respectively to determine the resource allocation response priorities of the three business links.

[0108] Specifically, the eigenvector weighting method is used to calculate the contribution weights of the three business links (equipment operation, customer activity, and supply chain) to the comprehensive benefit index, thereby determining the importance of each business link in the comprehensive benefit index. Historical data related to the three business links of equipment operation, customer activity, and supply chain are collected and cleaned to remove outliers, missing values, and duplicate values. Data of different dimensions and orders of magnitude are then normalized and mapped to the same scale range, such as the [0,1] interval. The preprocessed historical data of the three business links and the comprehensive benefit index data are constructed into a data matrix, in which each row represents a time point or sample, and each column corresponds to the indicator data of equipment operation, customer activity, and the three supply chain business links, as well as the comprehensive benefit index. The covariance matrix between the indicator data of the three business links and the comprehensive benefit index in the data matrix is ​​calculated to measure the strength of the linear relationship and the direction of the correlation between them. The covariance matrix is ​​subjected to eigenvalue decomposition to obtain eigenvalues ​​and corresponding eigenvectors. The eigenvectors are sorted in descending eigenvalue order, and the first three eigenvectors are selected to form an eigenvector matrix. The original data matrix is ​​projected onto the eigenvector matrix to obtain the principal component score matrix. The contribution weight of each business segment to the comprehensive benefit index is determined based on the variance contribution rate of each principal component (that is, the ratio of the eigenvalue to the total eigenvalue). The contribution weight is equal to the sum of the squares of the coefficients of the corresponding business segment in each principal component multiplied by the variance contribution rate of the principal component. The contribution weights of the three business segments are then normalized so that their total is 1.

[0109] Based on the intensity of risk transmission between the three business links in the risk transmission matrix, calculate the sensitivity of the three business links in the risk transmission process. For each business link, calculate the average value of the risk transmission intensity between it and the other two business links. Use the eigenvector weighting method to obtain the contribution weight of each business link to the comprehensive benefit index, that is, the weight value calculated in the previous step. The formula for calculating the risk transmission sensitivity coefficient is equal to the average value of the risk transmission intensity multiplied by the business link weight multiplied by the risk impact coefficient of the business link. Calculate the risk transmission sensitivity coefficient of each business link;

[0110] The graph attention mechanism automatically learns the relationships between nodes and calculates the impact of the critical path on the potential risk probability. Based on the importance and role of the business links along the critical path, the influence ranking of the three business links is determined, clarifying the priority of each business link in the risk transmission network. Based on the feature representation of each node output by the graph attention network, an importance score is calculated for each node. For example, a pooling operation or readout function can be used to map the node features to a scalar value representing the node's influence in the entire graph. The three business links are then ranked based on their influence scores to determine their criticality in the risk transmission network.

[0111] A resource allocation weight function is constructed by combining the contribution weights of the three business segments to the comprehensive benefit index, their risk transmission sensitivity, and the influence rankings derived from the graph attention network. This function comprehensively considers all factors and determines the resource allocation weight coefficient for each business segment, quantifying its relative importance in resource allocation. A resource allocation weight function can be constructed by, for example, multiplying the contribution weight, risk transmission sensitivity coefficient, and influence ranking score by their respective influence coefficients and then adding them together to obtain the comprehensive resource allocation weight for each business segment. The influence coefficients for the contribution weight, risk transmission sensitivity coefficient, and influence ranking score are 0.4, 0.4, and 0.2, respectively. These influence coefficients can be adjusted based on the company's emphasis on different factors, or derived through optimization algorithms or expert experience. The resulting resource allocation weight coefficients for the three business segments are normalized to the same scale and range for easier comparison and allocation. The three business segments are then ranked, with those with the largest resource allocation weights receiving priority, ensuring that resources are allocated rationally according to priority. Develop an adaptive scheduling engine to automatically trigger emergency plans and dynamically adjust resource allocation based on real-time risk changes and resource usage.

[0112] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A real-time data analysis method for enterprise digital operations based on multi-source data fusion, characterized in that: The following steps are involved: S1: Real-time collection of multi-source operational data during enterprise operations, construction of multi-source operational data sets, and pre-processing; S2: Extract features based on the preprocessed operational multi-source dataset and construct a three-dimensional vector space; S3: Utilizes 3D vector space and Apache Flink, combined with a time series forecasting model, to analyze and predict the comprehensive efficiency index of an enterprise. Based on the forecast results, it analyzes and evaluates the trends and potential risks of future operations. It also embeds the knowledge graph into the 3D vector space and constructs a dynamic graph neural network. S4: Analyze abnormal equipment deviations, abnormal supply chain deviations, and customer activity, build a risk matrix, and identify the propagation path of the decline in the comprehensive benefit index; use the analysis results of the risk transmission matrix to dynamically allocate business link resources; S5: Use the dashboard to visualize the company's future operating trends, comprehensive benefit index, and the path of the decline in the comprehensive benefit index; In step S3, the three-dimensional vector space and Apache Flink are used in combination with the time series prediction model to analyze and predict the comprehensive benefit index of the enterprise, including the following steps: By leveraging Apache Flink's stream processing capabilities, we can visualize vector changes in a three-dimensional vector space. Based on the position and change trends of vectors in the three-dimensional vector space, we can assess the cost-effectiveness of enterprise operations in real time. Formula for calculating comprehensive benefit index: Get the comprehensive benefit index Q of the enterprise; among them, is the modulus of the current state vector, which represents the position in the three-dimensional vector space; represents the information entropy of the vector distribution of the current state vector; f represents the current state feature vector; represents the revenue of the enterprise at time t; is the enterprise’s cost at time t; 、 and denote weight coefficients, which are 0.4, 0.3, and 0.3, respectively; Using a long-short-term memory (LSTM) model to predict future comprehensive benefit indices, the multi-source historical operational data set is fed into the LSTM model for training. During training, the data is fed into the LSTM model in chronological order, learning the data characteristics and change patterns at different time points and establishing a mapping relationship from the past to the present and then to the future. The trained LSTM model is deployed in the Apache Flink environment to perform real-time predictions on newly incoming multi-source operational data. When the real-time multi-source operational data enters the Flink stream processing pipeline, the output is a prediction of the future comprehensive benefit index based on the current input data and the comprehensive benefit index formula. In step S3, the knowledge graph is embedded in the three-dimensional vector space to construct a dynamic graph neural network, which includes the following steps: The knowledge graph is embedded into a three-dimensional vector space. In the knowledge graph, nodes represent entities, including devices, customers, and employees. Edges represent relationships between entities, including the relationship between devices and customers, the relationship between devices and employees, and the relationship between customers and employees. Use the TransE model to embed entities into a three-dimensional vector space. Introduce the time attribute, treat time as a dimension in the graph, convert the timestamp into a sine or cosine periodic vector, concatenate it with the three-dimensional space vector, and then reduce the dimensionality to three dimensions through a fully connected layer to form a joint spatiotemporal embedding, obtaining a node representation containing spatiotemporal information. Adopting the ‌EvolveGCN model‌, the LSTM is used to evolve the graph convolutional network parameters to capture the dynamic changes of the graph structure. Combined with the Apache Flink stream processing framework, incremental graph structure data is synchronized every 5 minutes. Utilizing Flink's stream processing capabilities, new data is updated to the knowledge graph, and the parameters of the dynamic graph neural network are updated in real time through the EvolveGCN model.

2. The method for real-time data analysis of enterprise digital operations based on multi-source data fusion according to claim 1 is characterized in that: In step S1, the multi-source data set for operation includes: equipment operation data, enterprise financial data, customer behavior data, human resources data and external data.

3. The method for real-time data analysis of enterprise digital operations based on multi-source data fusion according to claim 1 is characterized in that: The step S2 comprises the following steps: Through Apache Flink, we access pre-processed operational multi-source data sets and extract the maximum, minimum, and average values ​​of equipment operation data, enterprise financial data, customer behavior data, and human resources data as statistical feature values. Extract the difference between capital inflow and outflow and capital turnover rate from corporate financial data; extract customer retention rate and customer activity from customer behavior data; extract employee turnover rate from human resources data; The feature values ​​extracted from external data include: the busyness and traffic congestion level of the enterprise's geographical area; Based on the results of feature extraction of the above multi-source operational data sets, the feature vectors are combined into equipment status feature vectors, enterprise financial feature vectors, customer behavior feature vectors, human resource feature vectors and external feature vectors; The five constructed eigenvectors are combined to form a comprehensive eigenvector matrix. Principal component analysis (PCA) is used to reduce the dimensionality of the fused eigenvectors, retaining 95% of the variance of the principal components. The eigenvectors are then normalized and updated in real time within the Apache Flink stream processing framework. The comprehensive eigenvector matrix is ​​projected into three-dimensional space, and a right-hand coordinate system is used to construct a three-dimensional space vector, where the x-axis is the equipment status eigenvector, the y-axis is the enterprise financial eigenvector and the customer behavior eigenvector, and the z-axis is the human resources eigenvector and the external eigenvector. Three linearly independent vectors are selected as the basis, labeled i, j, and k, and the order of the three directions is determined according to the right-hand rule.

4. The method for real-time data analysis of enterprise digital operations based on multi-source data fusion according to claim 1 is characterized in that: The step S3 analyzes and evaluates the trend and potential risks of the enterprise's future operations based on the forecast results, including the following steps: The comprehensive benefit index at different time points is predicted to be drawn into a time series graph, where the horizontal axis represents time and the vertical axis represents the comprehensive benefit index; Using the linear regression method, with time as the independent variable and the comprehensive benefit index as the dependent variable, a straight line is fitted. The slope of the fitted line is used to determine whether the comprehensive benefit index is increasing or decreasing over time. When Q shows an upward trend, it indicates that the company's future operating trend is positive; when Q shows a downward trend, it indicates that the company's future operating trend is negative. According to the historical comprehensive benefit index, different thresholds are divided: when Q When Q is 0.8, it indicates that the enterprise's operating status is excellent; when 0.4≤Q≤0.8, it indicates that the enterprise's operating status is good; when Q When it is 0.4, it means that the enterprise’s operating status is dangerous; The LSTM model is used to predict the future comprehensive benefit index and determine whether there are potential risks in the future enterprise operation status based on the set threshold. When the future comprehensive benefit index is less than 0.4, the potential risk is identified; otherwise, there is no potential risk.

5. The method for real-time data analysis of enterprise digital operations based on multi-source data fusion according to claim 1 is characterized in that: The step S4 analyzes the abnormal device offset, the abnormal supply chain offset, and the customer activity, constructs a risk matrix, and identifies the propagation path of the decline in the comprehensive benefit index, including the following steps: The reference vector is obtained by calculating the average value of each characteristic parameter in the normal equipment operation characteristic vector as the reference value; the offset between the current equipment state vector and the reference vector is calculated in real time using the Euclidean distance formula to obtain the equipment abnormality offset; Combining the external feature vector, enterprise financial feature vector and human resource feature vector, the supply chain status feature vector is generated. Similarly, the calculation method of equipment abnormal offset is used to obtain the supply chain abnormal offset. Obtain client node activity data, calculate the difference between the activity of the previous cycle and the activity of the current cycle, and then divide it by the activity of the previous cycle to obtain the activity decline rate; The device anomaly offset, supply chain anomaly offset, customer activity, and timestamp are used as input signals. The graph attention network model is introduced to construct a dynamic graph neural network, and the graph attention mechanism is used to automatically learn the relationship between nodes. The input signal is fed into the constructed dynamic graph neural network for training. The graph attention mechanism is used to calculate the impact of the critical path on the potential risk probability and output the potential risk probability. Based on the output of the dynamic graph neural network, the risk impact of different nodes and paths is evaluated, and a risk transmission matrix is ​​constructed. The elements in the matrix represent the intensity of risk transmission between different nodes or three business links; the three business links include: equipment operation business link, customer activity business link, and supply chain business link; When the comprehensive benefit index shows a downward trend, the risk transmission matrix is ​​used to identify the risk transmission path and the degree of risk impact.

6. The method for real-time data analysis of enterprise digital operations based on multi-source data fusion according to claim 5 is characterized in that: In step S4, the analysis results of the risk transmission matrix are used to dynamically allocate business link resources, including the following steps: The characteristic vector weighting method is used to obtain the contribution weights of the three business links of equipment operation, customer activity, and supply chain to the comprehensive benefit index; According to the intensity of risk transmission between the three business links in the risk transmission matrix, calculate the sensitivity of the three business links in the risk transmission process respectively; The graph attention mechanism automatically learns the relationship between nodes, calculates the impact of the critical path on the potential risk probability, and determines the influence ranking of the three business links; Combining the contribution weights of the three business links to the comprehensive benefit index, as well as the risk transmission sensitivity and the influence ranking derived from the graph attention network, a resource allocation weight function is constructed to determine the weight coefficient of resource allocation; After normalizing the obtained resource allocation weight coefficients of the three business links, the three business links are sorted respectively to determine the resource allocation response priorities of the three business links.

Images