Structured data-based enterprise project management information prediction system and method

Abstract

The application provides an enterprise project operation information prediction system and method based on structured data, relates to the technical field of enterprise project operation information prediction, and comprises the following steps: analyzing multiple source structured data to obtain time series operation data and cross section operation data; calling an SVD decomposition tool to decompose an effective coupling matrix to obtain a project principal component feature vector group; inputting the project principal component feature vector group into three parallel trained LightGBM light prediction models, simultaneously obtaining and analyzing evidence indexes of historical enterprise projects to obtain comprehensive evidence strength, calculating the final operation information prediction probability of the enterprise project based on entropy weight, and evaluating the state of the enterprise project. A two-stage dimension reduction strategy combining singular value decomposition and principal component analysis is adopted, core principal singular values are screened through a cumulative contribution rate threshold, the data dimension is greatly reduced while the main information of the original matrix is retained, and the risk of overfitting is avoided.

Description

Technical Field

[0001] This invention relates to the technical field of enterprise project operation information prediction methods, specifically to an enterprise project operation information prediction system and method based on structured data. Background Technology

[0002] Enterprise project operations are influenced by both internal execution capabilities and changes in the external environment. Internally, this involves structured data such as financial data (revenue, costs, debt-to-equity ratio) and operational data (project schedule, subcontractor management). Externally, it involves structured data such as macroeconomic data (energy prices, PMI index) and industry benchmark data (industry average gross profit margin, default rate). Current technologies for predicting enterprise project operation information primarily employ a "decomposition analysis" model: on the one hand, time-series analysis methods such as ARIMA and exponential smoothing are used to process dynamic data such as daily costs and daily progress to capture fluctuations in the operational process; on the other hand, static comparison methods such as ratio analysis and industry benchmarking are used to process static data such as annual debt-to-equity ratio and industry benchmarks to assess the operational status within the industry.

[0003] However, this traditional approach has significant drawbacks: First, modeling time-series and cross-sectional data separately fails to quantify the coupling relationship between the two types of data, resulting in a predictive model that cannot capture the overall nature of the operating system, leading to biased predictions that fail to reflect the true operating status. Second, traditional methods only employ simple weighted summation or rule-based judgment at the data fusion level, making it difficult to handle the mutual corroboration and conflict between multi-source evidence, and weight settings often rely on subjective experience, lacking objective data support. Third, in terms of high-dimensional feature processing, existing technologies either directly use all features, leading to the curse of dimensionality and model overfitting, or use simple feature selection, resulting in the loss of key information, making it difficult to guarantee the accuracy and robustness of the prediction results. Therefore, how to construct a method for predicting enterprise project operating information that can integrate multi-source structured data, quantify the coupling relationship between dynamic and static factors, objectively determine the weights of each dimension, and achieve accurate risk prediction is a technical problem that urgently needs to be solved in this field.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting enterprise project operation information based on structured data, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The method for predicting enterprise project operation information based on structured data includes the following steps: S1: Collect multi-source structured data on enterprise project operations, and classify and analyze the multi-source structured data through data source classification to obtain time-series operation data and cross-sectional operation data; S2: Divide the time-series and cross-sectional business data into dimensions, construct a single-modal matrix for each dimension, process each single-modal matrix using the matrix correlation alignment method to generate an effective coupling matrix; call the singular value decomposition tool to decompose the effective coupling matrix to extract the principal component feature vector group of the project. S3: Input the project principal component feature vector group into three parallel, trained LightGBM lightweight prediction models respectively, and output the initial probability of cost overrun, initial probability of default risk, and initial probability of compliance violation; simultaneously obtain the historical evidence indicators corresponding to each initial probability, and calculate the comprehensive evidence strength accordingly. S4: Based on the comprehensive evidence strength, each initial probability is corrected to obtain the probability of cost overrun, the probability of default risk, and the probability of compliance violation; the entropy weights of each type of probability after correction are calculated, and the predicted probability of the enterprise project's operating information is determined based on the entropy weights. Finally, the operating status of the enterprise project is evaluated based on the predicted probability of the operating information.

[0007] Furthermore, we acquire time-series and cross-sectional operational data, specifically as follows: The multi-source structured data includes internal enterprise structured data and external enterprise structured data; For the enterprise's internal structured data, if it contains continuous timestamps and is dynamically updated with the business cycle, it is classified as internal time-series candidate data; if it is static data containing only a single timestamp, it is classified as internal cross-sectional candidate data; the business cycle is divided into construction period, operation period and completion period according to the enterprise's project progress percentage. For external structured data of an enterprise, if it has a preset high-frequency update characteristic, it is classified as external time-series candidate data. The high-frequency update characteristic includes daily updates and monthly updates. If the external structured data of the enterprise is a static baseline value, it is classified as external cross-sectional candidate data. The external and internal time-series candidate data are aligned according to the timestamp. Data with the same enterprise project ID and the same timestamp are matched, and data with missing timestamps are removed to generate time-series operational data. The internal and external cross-section candidate data are matched according to the same enterprise project ID to generate cross-section business data.

[0008] Furthermore, the time-series and cross-sectional operating data are dimensionally divided, including operating cost, risk warning, and compliance assessment dimensions. The corresponding time-series and cross-sectional fields under different dimensions are obtained based on the following logic: For the operating cost dimension, fields related to cost fluctuations are extracted from time-series operating data as time-series fields, and fields related to cost benchmarks are extracted from cross-sectional operating data as cross-sectional fields. For risk warning, daily net capital inflow and monthly progress lag days are extracted from time-series operating data as time-series fields, and asset-liability ratio, interest coverage ratio and industry default rate are extracted from cross-sectional operating data as cross-sectional fields. For the compliance assessment dimension, daily compliance inspection scores and rectification completion rates are extracted from time-series operating data as time-series fields, and annual violation counts and industry compliance pass rates are extracted from cross-sectional operating data as cross-sectional fields. Among them, daily net cash inflow equals the total cash inflow of a company's projects on that day minus the total cash outflow on that day; monthly progress lag days refer to the difference between the actual completion date and the planned completion date of a company's projects within the monthly statistical period; compliance inspection score refers to the percentage score given by the company after conducting daily compliance inspections of its projects; rectification completion rate equals the number of issues that have been rectified in a company's projects divided by the total number of issues that should be rectified; annual violation count represents the total number of violations that a company's projects have been formally identified as violations by regulatory agencies in the past year; and industry compliance pass rate equals the proportion of companies that have passed compliance inspections to the total number of companies in the same industry. To obtain the correlation strength between time-series and cross-sectional fields within each dimension, the following steps are taken: For each pair of time-series and cross-sectional fields within each dimension, using the same natural month as the same time period, data samples from the same enterprise project within that natural month are collected. First, the Pearson correlation coefficient between the time-series and cross-sectional fields is calculated, and the weight parameters assigned by expert scoring are obtained. Then, a balancing coefficient is set, and the balancing coefficient is multiplied by the Pearson correlation coefficient to obtain the first product result. The difference between the first and the balancing coefficient is multiplied by the expert scoring weight parameters to obtain the second product result. Finally, the first product result and the second product result are added together, and the sum is the correlation strength between the pair of time-series and cross-sectional fields. The logic for calculating the Pearson correlation coefficient is as follows: For a set of time-series and cross-sectional fields within any dimension, collect multiple samples of the same project within the same time period from the company's historical project database. Calculate the average of all samples in the time-series field and the average of all samples in the cross-sectional field. Then, calculate the deviation of the corresponding value in the time-series field from its average value for each sample pair, multiplied by the deviation of the corresponding value in the cross-sectional field from its average value. Sum these products across all sample pairs to obtain the numerator. Next, calculate the square root of the sum of squares of the deviations in the time-series field and the sum of squares of the deviations in the cross-sectional field, multiplying these two square root results to obtain the denominator. Finally, divide the numerator by the denominator to obtain the Pearson correlation coefficient.

[0009] Furthermore, based on the aforementioned correlation strength, a single-modality matrix for each dimension is constructed, specifically as follows: First, construct a correlation table. Each row in the correlation table records the dimension name, the name of the time-series field, the name of the cross-sectional field, and the correlation strength. Based on the correlation table, construct a unimodal matrix for each dimension using the unimodal matrix construction method: Set the number of enterprise projects to be analyzed as the number of rows in the unimodal matrix, and set the sum of the number of time-series fields and the number of cross-sectional fields as the number of columns in the unimodal matrix. Extract the values ​​of the time-series fields for each enterprise project at each timestamp from the time-series operating data and fill them into the corresponding row and column positions; extract the values ​​of the cross-sectional fields for each enterprise project from the cross-sectional operating data and fill them into the corresponding row and column positions; finally, fill the corresponding association strengths from the association table into the corresponding columns of the unimodal matrix.

[0010] Furthermore, an effective coupling matrix is ​​generated, specifically as follows: Matrix normalization is applied to the unimodal matrix of each dimension to eliminate dimensional differences between different matrices. For each unimodal matrix, Min-Max normalization is used to map all elements to the [0,1] interval, resulting in a normalized unimodal matrix. Based on each normalized unimodal matrix, matrix correlation alignment is used to generate an effective coupling matrix. The standardized unimodal matrix of operating cost dimension is selected as the baseline matrix. Based on the row index of the enterprise project number in the baseline matrix, the rows of the same enterprise project number in the standardized unimodal matrices of risk warning dimension and compliance assessment dimension are matched and aligned in turn to ensure that the data of the same enterprise project is in the same row. Then, the three aligned matrices are horizontally spliced ​​in the column direction. That is, all columns of operating cost dimension are placed first, and then all columns of risk warning dimension and compliance assessment dimension are placed in turn to form a new matrix with the same number of rows and the same number of columns as the sum of the number of columns of the three dimensions. This is the effective coupling matrix.

[0011] Furthermore, the principal component feature vector sets of the project are extracted, specifically as follows: The effective coupling matrix is ​​decomposed by calling the singular value decomposition tool into a left singular orthogonal matrix, a diagonal matrix, and a right singular orthogonal matrix. All non-zero singular values ​​are extracted from the diagonal of the diagonal matrix to obtain a sequence of singular values ​​arranged in descending order. Starting from the first largest singular value in the sequence, the values ​​are accumulated sequentially. When the sum of the accumulated values ​​reaches or exceeds 85% of the total sum of all singular values ​​for the first time, all singular values ​​from the first singular value to the current singular value are identified as core principal singular values. Each column of the left singular orthogonal matrix corresponding to the core principal singular value is labeled as a left singular vector. These left singular vectors are combined according to their columns to form the principal component eigenvector group of the project.

[0012] Further, the overall strength of evidence is calculated as follows: Three parallel LightGBM lightweight prediction models are constructed. Each lightweight prediction model is trained. The project principal component feature vector group is input into the three parallel trained LightGBM lightweight prediction sub-models, and the initial probabilities of cost overrun, default risk, and compliance violation are output respectively. We statistically analyzed the evidence indicators corresponding to the three initial probabilities for the enterprise's projects over the past three consecutive and complete calendar months. The evidence indicators corresponding to the initial probability of cost overrun were the cost consumption acceleration rate and the number of budget warnings triggered. The evidence indicators corresponding to the initial probability of default risk were the number of days of schedule deviation and the frequency of requirement changes. The evidence indicators corresponding to the initial probability of compliance and non-compliance were the number of audit issues and the number of contract penalty warnings triggered. After marking cost overruns, default risk events, and compliance violations as risk events, the dataset of each evidence indicator for all projects in the three months prior to the occurrence of the risk event is extracted from the enterprise's historical project database. First, the median of each evidence indicator in the dataset is calculated as the adjustment threshold parameter for that evidence indicator, and the interquartile range of each evidence indicator is also calculated. Then, for each evidence indicator, the original value is subtracted from the adjustment threshold parameter to obtain the deviation value. The quotient obtained by dividing 4 by the interquartile range is used as the scaling factor. The scaling factor is multiplied by the deviation value to obtain the scaled deviation. Then, the exponential function value with the natural constant as the base and the exponent as the scaled deviation is calculated. The denominator is obtained by adding 1 to the exponential function value. Finally, the standardized evidence strength of the evidence indicator is obtained by dividing 1 by the denominator. Meanwhile, using standardized evidence strength as input and actual occurrence of risk events as output labels, a logistic regression model is constructed and trained. After training, the absolute values ​​of the obtained feature regression coefficients are normalized to obtain the corresponding weight coefficients for each evidence indicator. For each type of initial probability, including the initial probability of cost overrun, the initial probability of default risk, and the initial probability of compliance violation, first determine the set of evidence indicators corresponding to that type of initial probability. Then, for each evidence indicator in the set, multiply the standardized evidence strength of the evidence indicator by its corresponding weight coefficient to obtain the weighted evidence strength of the evidence indicator. Finally, add the weighted evidence strengths of all evidence indicators in the set, and the sum is the comprehensive evidence strength of that type of initial probability.

[0013] Furthermore, the probability of predicting the business information of the enterprise project is determined, specifically as follows: For each type of initial probability, firstly, obtain the comprehensive strength of evidence corresponding to that type of initial probability, and calculate the Beta function value with the comprehensive strength of evidence as a parameter when the risk event occurs, and the Beta function value with the comprehensive strength of evidence as a parameter when the risk event does not occur; then, multiply the type of initial probability by the Beta function value when the risk event occurs to obtain the first product result, and multiply the difference between the first and second initial probabilities by the Beta function value when the risk event does not occur to obtain the second product result; finally, divide the first product result by the sum of the first and second product results, and the quotient is the corrected probability of that type of initial probability; For each type of probability—cost overrun probability, default risk probability, and compliance violation probability—first, calculate the information entropy of that type of probability. Specifically, multiply that type of probability by its base-2 logarithm, add one minus the difference of that type of probability multiplied by its base-2 logarithm, and then take the negative of the sum to obtain the information entropy of that type of probability. Next, calculate the difference between one and the information entropy to obtain the entropy weight numerator of that type of probability. Finally, divide the entropy weight numerator of that type of probability by the sum of the entropy weight numerators of all types of probabilities, and the quotient is the entropy weight of that type of probability. For each type of probability—cost overrun probability, default risk probability, and compliance violation probability—the project stage weight is first determined based on the current operational cycle stage of the enterprise's project: If in the construction phase, the project stage weight for cost overrun probability is 60%, default risk probability is 30%, and compliance violation probability is 10%; if in the operation phase, the project stage weight for cost overrun probability is 20%, default risk probability is 50%, and compliance violation probability is 30%; if in the finalization phase, the project stage weight for cost overrun probability is 10%, default risk probability is 30%, and compliance violation probability is 60%. A balance coefficient is defined. Multiplying this coefficient by the project stage weight for that probability class yields the first product. Multiplying the difference between the first and second products by the entropy weight for that probability class yields the second product. Adding the first and second products gives the numerator of the fusion weight for that probability class. Adding the numerators of the fusion weights for all three probability classes gives the denominator of the fusion weight. Dividing the numerator by the denominator gives the fusion weight for that probability class. Finally, multiplying the cost overrun probability by its fusion weight gives the cost overrun weighted value; multiplying the default risk probability by its fusion weight gives the default risk weighted value; and multiplying the compliance / non-compliance probability by its fusion weight gives the compliance / non-compliance weighted value. Adding these three weighted values ​​gives the final predicted probability of the enterprise project's operational information. If the predicted probability of business information is greater than a preset threshold, an operational risk warning is generated; otherwise, the business project is deemed to be operating normally.

[0014] The present invention also provides an enterprise project operation information prediction system based on structured data, the prediction system being used to execute the above prediction method, comprising: The data acquisition module is used to collect multi-source structured data of enterprise project operations, and to divide and analyze the multi-source structured data through the data source classification method to obtain time-series operation data and cross-sectional operation data; The matrix decomposition module is used to divide time-series and cross-sectional business data into dimensions, construct unimodal matrices for each dimension, process each unimodal matrix using matrix correlation alignment, and generate an effective coupling matrix; the singular value decomposition tool is called to decompose the effective coupling matrix to extract the principal component feature vector group of the project. The evidence strength calculation module is used to input the project principal component feature vector group into three parallel, trained LightGBM lightweight prediction models, and output the initial probability of cost overrun, initial probability of default risk, and initial probability of compliance violation. It also simultaneously acquires historical evidence indicators corresponding to each initial probability and calculates the comprehensive evidence strength accordingly. The information prediction module is used to correct each initial probability based on the comprehensive evidence strength to obtain the probability of cost overrun, the probability of default risk, and the probability of compliance violation; calculate the entropy weight of each type of probability after correction, and determine the predicted probability of the enterprise project's operating information based on the entropy weight; and finally evaluate the operating status of the enterprise project based on the predicted probability of the operating information.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By constructing a multimodal coupling matrix and introducing coupling coefficients, time-series and cross-sectional operational data are unified into the same matrix framework, explicitly quantifying the correlation strength between dynamic execution data and static benchmark data. A two-stage dimensionality reduction strategy combining singular value decomposition and principal component analysis is employed, using a cumulative contribution rate threshold to filter core principal singular values. This significantly reduces data dimensionality while preserving the main information of the original matrix, avoiding the risk of overfitting. This scheme constructs three parallel LightGBM lightweight prediction models, each outputting initial probabilities for three dimensions, and introduces historical evidence indicators, significantly improving prediction accuracy and robustness. The entropy weighting method is used to objectively weight the corrected probabilities, measuring the dispersion of probabilities in each dimension through information entropy. The data itself determines the weight allocation, completely avoiding the bias caused by subjective weight setting in traditional methods. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a graph showing the relationship between the original values ​​of the evidence indicators and their corresponding standardized strengths of evidence. Figure 3 This is a schematic diagram of the overall system structure of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0018] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0019] Example: Please see Figures 1-2 The present invention provides a technical solution: The method for predicting enterprise project operation information based on structured data includes the following steps: S1: Collect multi-source structured data on enterprise project operations, and classify and analyze the multi-source structured data through data source classification to obtain time-series operation data and cross-sectional operation data; The specific steps for obtaining time-series and cross-sectional business data are as follows: The multi-source structured data includes internal enterprise structured data and external enterprise structured data; For the enterprise's internal structured data, if it contains continuous timestamps and is dynamically updated with the business cycle, it is classified as internal time-series candidate data; if it is static data containing only a single timestamp, it is classified as internal cross-sectional candidate data; the business cycle is divided into construction period, operation period and completion period according to the enterprise's project progress percentage. For external structured data of an enterprise, if it has a preset high-frequency update characteristic, it is classified as external time-series candidate data. The high-frequency update characteristic includes daily updates and monthly updates. If the external structured data of the enterprise is a static baseline value, it is classified as external cross-sectional candidate data. In the above process, the multi-source structured data refers to enterprise project operation-related data that comes from different business systems and different collection channels and has clear field definitions and format specifications. Specifically, it includes two main categories: internal enterprise structured data and external enterprise structured data. Among them, internal structured data refers to data automatically generated and stored by the business systems operated by the enterprise itself, such as enterprise resource planning systems, financial accounting systems, project management systems, and human resource systems, during the operation process. This data is centered on the enterprise's own projects and includes standard fields such as enterprise project number, timestamp, amount, and quantity. Typical examples include daily cost records, monthly cash flow statements, project completion progress, and subcontractor performance records. Its core feature is that it is bound to a specific enterprise project ID and reflects the execution status of the project itself. External structured data refers to standardized data obtained by authoritative third-party platforms such as Wind Financial Database, industry association statistical platforms, government regulatory agencies' official websites, and commodity exchanges through application programming interfaces (APIs) or data collection interfaces. This data is organized by industry or market and includes environmental fields such as data release time, industry classification, and regional identifiers. Typical examples include daily raw material futures prices, monthly purchasing managers' index, industry average default rate, quarterly industry cost benchmarks, and annual compliance standards. Its core characteristic is that it reflects the external environmental conditions of the project rather than the project's own execution status. By dividing internal structured data into time-series and cross-sectional categories based on time-dimensional characteristics, and external structured data into time-series and cross-sectional categories based on update frequency characteristics, accurate identification and classification of multi-source data is achieved. This ensures that dynamic and static data can play their respective roles in the subsequent construction of the multimodal coupling matrix, avoiding the problems of dynamic information being submerged by static noise and static benchmarks being distorted by dynamic fluctuations due to data type mixing. This significantly improves the coupling matrix's ability to characterize the overall business system. This classification and categorization is based on the fundamental differences between internal and external data in terms of business meaning and statistical characteristics, necessitating differentiated classification standards. For internal structured data, its core value lies in reflecting the execution status of a project; therefore, the classification standard focuses on "continuity in the time dimension," meaning data containing continuous timestamps and dynamically updated with the business cycle. This data can depict the dynamic evolution of the business process and is suitable for trend analysis and fluctuation capture. In contrast, static data containing only a single timestamp can only reflect a snapshot of the state at a specific point in time, suitable for benchmark comparison and status assessment. For external structured data of an enterprise, its core value lies in reflecting the environmental conditions in which the project is located. Therefore, the classification criteria focus on the "update frequency", that is, data with high-frequency update characteristics on a daily or monthly basis. This is the core environmental variable that drives the dynamic changes in the business status and needs to be analyzed in conjunction with internal time-series candidate data. On the other hand, the static benchmark value released at a low frequency is only used to provide a reference scale for horizontal comparison and needs to be compared with internal cross-sectional candidate data.

[0020] The external and internal time-series candidate data are aligned according to the timestamp. Data with the same enterprise project ID and the same timestamp are matched, and data with missing timestamps are removed to generate time-series operational data. The internal and external cross-section candidate data are matched according to the same enterprise project ID to generate cross-section business data.

[0021] The above process, through a dual association mechanism of "timestamp alignment + project ID matching", generates time-series operational data and cross-sectional operational data respectively, ensuring the synchronization of dynamic data in the time dimension and the comparability of static data in the business dimension.

[0022] S2: Divide the time-series and cross-sectional business data into dimensions, construct a single-modal matrix for each dimension, process each single-modal matrix using the matrix correlation alignment method to generate an effective coupling matrix; call the singular value decomposition tool to decompose the effective coupling matrix to extract the principal component feature vector group of the project. The time-series and cross-sectional operating data are dimensionally divided, including operating cost, risk warning, and compliance assessment dimensions. The corresponding time-series and cross-sectional fields under each dimension are then obtained, based on the following logic: For the operating cost dimension, fields related to cost fluctuations are extracted from time-series operating data as time-series fields, and fields related to cost benchmarks are extracted from cross-sectional operating data as cross-sectional fields. For risk warning, daily net capital inflow and monthly progress lag days are extracted from time-series operating data as time-series fields, and asset-liability ratio, interest coverage ratio and industry default rate are extracted from cross-sectional operating data as cross-sectional fields. For the compliance assessment dimension, daily compliance inspection scores and rectification completion rates are extracted from time-series operating data as time-series fields, and annual violation counts and industry compliance pass rates are extracted from cross-sectional operating data as cross-sectional fields. In the above process, the operating cost dimension directly corresponds to the project's profitability and cost control level, and is a core indicator for measuring the project's economic benefits; the risk warning dimension corresponds to the project's ability to perform its obligations and financial security, reflecting major risks such as default, delays, or cash flow disruptions that may occur during the project's implementation; the compliance assessment dimension corresponds to the project's legality and compliance, involving adherence to regulatory requirements, contractual obligations, and industry standards; these three dimensions together constitute the overall picture of the enterprise's project operating risks. Decomposing the forecasting task according to these three dimensions makes the forecasting objectives of each dimension more focused, and also makes the forecasting results more interpretable. In terms of operating costs, extracting fields related to cost fluctuations, such as daily net cash flow, from time-series operating data can capture the dynamic changes in project cash flow in real time, reflecting the true rhythm and abnormal fluctuations of cost expenditures; extracting fields related to cost benchmarks, such as debt-to-equity ratio and interest coverage ratio, from cross-sectional operating data can assess the project's cost-bearing capacity and risk resistance boundaries from a financial structure perspective. The combination of the two means that the assessment of cost overrun risk has both sensitive indicators at the dynamic execution level and constrained benchmarks at the static financial level. In the risk warning dimension, extracting daily net cash inflow and monthly progress lag days as time-series fields enables dynamic monitoring of project cash turnover efficiency and schedule execution. In other words, a persistently negative net cash inflow or a persistently delayed schedule is often a precursor to default risk. Extracting asset-liability ratio, interest coverage ratio, and industry default rate as cross-sectional fields enables assessment of the probability of default from the perspective of the overall financial health of the enterprise and industry benchmarking. This allows risk warning to have both the dynamic performance of individual projects and horizontal reference to the industry environment. In the compliance assessment dimension, daily compliance inspection scores and rectification completion rates are extracted as time-series fields, which can dynamically track the compliance performance and problem rectification efficiency in the daily operation of the project. A continuous decline in scores or a long-term insufficient rectification rate indicates that compliance risks are accumulating. Annual violation counts and industry compliance pass rates are extracted as cross-sectional fields, which can provide a benchmark for the current compliance status of the project from the perspectives of historical violation records and industry compliance levels. This makes the assessment of compliance risks have both a vertical historical trajectory and a horizontal industry positioning.

[0023] In summary, this dimensional division and corresponding data extraction strategy enable this technical solution to comprehensively depict the operational status of enterprise projects from three core dimensions: cost, risk, and compliance.

[0024] Among them, daily net cash inflow equals the total cash inflow of a company's projects on that day minus the total cash outflow on that day; monthly progress lag days refer to the difference between the actual completion date and the planned completion date of a company's projects within the monthly statistical period; compliance inspection score refers to the percentage score given by the company after conducting daily compliance inspections of its projects; rectification completion rate equals the number of issues that have been rectified in a company's projects divided by the total number of issues that should be rectified; annual violation count represents the total number of violations that a company's projects have been formally identified as violations by regulatory agencies in the past year; and industry compliance pass rate equals the proportion of companies that have passed compliance inspections to the total number of companies in the same industry. To obtain the correlation strength between time-series and cross-sectional fields within each dimension, the following steps are taken: For each pair of time-series and cross-sectional fields within each dimension, using the same natural month as the same time period, data samples from the same enterprise project within that natural month are collected. First, the Pearson correlation coefficient between the time-series and cross-sectional fields is calculated, and the weight parameters assigned by expert scoring are obtained. Then, a balancing coefficient is set, and the balancing coefficient is multiplied by the Pearson correlation coefficient to obtain the first product result. The difference between the first and the balancing coefficient is multiplied by the expert scoring weight parameters to obtain the second product result. Finally, the first product result and the second product result are added together, and the sum is the correlation strength between the pair of time-series and cross-sectional fields. The formula upon which the above process is based is: in, Indicates the balance coefficient; Indicates the strength of the association; This represents the weighting parameters assigned to the expert scores; This represents the Pearson correlation coefficient between time-series and cross-sectional fields. In the above process, the weighting parameters assigned to expert scores are determined by inviting multiple experts with extensive project management experience within the company, typically including cost control experts, risk management experts, compliance audit experts, and senior project managers. The specific scoring process is as follows: First, each expert is provided with detailed definitions and business meanings of time-series and cross-sectional fields within each dimension, and the scoring criteria are clarified, for example, using a scale of 1 to 9 points, where 1 point represents "almost no impact," 3 points represent "minor impact," 5 points represent "moderate impact," 7 points represent "significant impact," and 9 points represent "decisive impact." Then, each expert independently gives their assessment of the importance of each pair of time-series and cross-sectional field combinations in the risk assessment of the corresponding dimension. Finally, all expert scores are collected, and for each pair of field combinations, the arithmetic mean of all expert scores is calculated and normalized to the [0,1] interval to obtain the final weighting parameter W. Balance coefficient The relative contribution of the Pearson correlation coefficient and the weighting parameters assigned by expert scores in the calculation of association strength is used to adjust the relative contribution. Their values ​​typically range from 0 to 1. During the initialization phase, the expert group directly sets an initial value based on their level of trust in the data quality and their business experience. For example, if they believe statistical correlation is more reliable, they set... =0.7, if expert experience is considered more important, then set =0.7. =0.3, if both are equally important, then set =0.5; The logic for calculating the Pearson correlation coefficient is as follows: For a set of time-series and cross-sectional fields within any dimension, collect multiple samples of the same project within the same time period from the company's historical project database. Calculate the average of all samples in the time-series field and the average of all samples in the cross-sectional field. Then, calculate the deviation of the corresponding value in the time-series field from its average value for each sample pair, multiplied by the deviation of the corresponding value in the cross-sectional field from its average value. Sum these products across all sample pairs to obtain the numerator. Next, calculate the square root of the sum of squares of the deviations in the time-series field and the sum of squares of the deviations in the cross-sectional field, multiplying these two square root results to obtain the denominator. Finally, divide the numerator by the denominator to obtain the Pearson correlation coefficient.

[0025] The above process is based on the following formula: Suppose a set of time-series fields within a certain dimension and section class fields Collected from the company's historical project database Group sample data, obtain sample pairs ; First, calculate and Sample mean: in, , These represent the sample means of the time-series field and the cross-sectional field, respectively. Indicates the total number of sample pairs; , They represent the first A time-series type field and a section type field; Then Pearson correlation coefficient The calculation formula is: in, This represents the Pearson correlation coefficient; Based on the correlation strength, a single-modality matrix is ​​constructed for each dimension, specifically as follows: First, construct a correlation table. Each row in the correlation table records the dimension name, the name of the time-series field, the name of the cross-sectional field, and the correlation strength. Based on the correlation table, construct a unimodal matrix for each dimension using the unimodal matrix construction method: Set the number of enterprise projects to be analyzed as the number of rows in the unimodal matrix, and set the sum of the number of time-series fields and the number of cross-sectional fields as the number of columns in the unimodal matrix. Extract the values ​​of the time-series fields for each enterprise project at each timestamp from the time-series operating data and fill them into the corresponding row and column positions; extract the values ​​of the cross-sectional fields for each enterprise project from the cross-sectional operating data and fill them into the corresponding row and column positions; finally, fill the corresponding association strengths from the association table into the corresponding columns of the unimodal matrix.

[0026] In the above process, for each time series field, extract the values ​​corresponding to all timestamps in the last three consecutive and complete natural months, calculate the arithmetic mean of these values, and fill the matrix with the representative value of the field; for example, if the daily net cash inflow of a project has 90 values ​​in the past 90 days, add these 90 values ​​together and divide by 90 to get the average value. Using the enterprise project number as the row index of the matrix ensures that all data for each project can be completely recorded in the same row. Using the sum of time-series fields, cross-sectional fields, and correlation strength as the column number ensures that the matrix's column dimensions include both the original feature fields and the quantified coupling relationship fields, achieving synchronous storage of feature and relationship information. Extracting the time-series field values ​​for each project at each timestamp from the time-series operational data and filling them into the corresponding positions preserves the time-series sensitivity and individual differences of dynamic data. Extracting the cross-sectional field values ​​for each project from the cross-sectional operational data and filling them into the corresponding positions provides a horizontal comparison benchmark for static data. Filling the pre-generated correlation strength from the correlation table into the corresponding columns explicitly encodes the degree of coupling between dynamic and static fields into the matrix. This approach enables subsequent matrix standardization and matrix correlation alignment operations to be performed efficiently on a unified data structure, avoiding information loss or alignment errors caused by scattered data storage or inconsistent formats. This is a key preliminary step in constructing a high-quality, multimodal, and effectively coupled matrix.

[0027] Generate an effective coupling matrix, specifically as follows: Matrix normalization is applied to the unimodal matrix of each dimension to eliminate dimensional differences between different matrices. For each unimodal matrix, Min-Max normalization is used to map all elements to the [0,1] interval, resulting in a normalized unimodal matrix. Based on each normalized unimodal matrix, matrix correlation alignment is used to generate an effective coupling matrix. The standardized unimodal matrix of operating cost dimension is selected as the baseline matrix. Based on the row index of the enterprise project number in the baseline matrix, the rows of the same enterprise project number in the standardized unimodal matrices of risk warning dimension and compliance assessment dimension are matched and aligned in turn to ensure that the data of the same enterprise project is in the same row. Then, the three aligned matrices are horizontally spliced ​​in the column direction. That is, all columns of operating cost dimension are placed first, and then all columns of risk warning dimension and compliance assessment dimension are placed in turn to form a new matrix with the same number of rows and the same number of columns as the sum of the number of columns of the three dimensions. This is the effective coupling matrix.

[0028] In the above process, Min-Max normalization maps all elements in each unimodal matrix to the [0,1] interval, eliminating the dimensional differences caused by different field units, such as net capital inflow in RMB ten thousand, progress delay days in days, and compliance inspection scores in points. This avoids information weight imbalance caused by significant differences in numerical magnitude during subsequent matrix decomposition. The standardized unimodal matrix of operating cost dimension is chosen as the benchmark matrix because operating cost is the core focus dimension of enterprise project operation. Using it as the benchmark ensures that the primary and secondary order of matrix alignment conforms to business logic. Row matching and alignment are performed using the enterprise project number as the index, ensuring that cost, risk, and compliance data of the same project can be accurately associated with the same row, avoiding data interference caused by misaligned project numbers. Horizontal splicing is performed in the order of operating cost → risk warning → compliance assessment, so that the column dimensions of the effectively coupled matrix present a clear business hierarchy structure, which facilitates subsequent matrix decomposition operations and enhances the interpretability of the matrix results.

[0029] Extract the principal component feature vector set of the project, specifically: The effective coupling matrix is ​​decomposed by calling the singular value decomposition tool into a left singular orthogonal matrix, a diagonal matrix, and a right singular orthogonal matrix. All non-zero singular values ​​are extracted from the diagonal of the diagonal matrix to obtain a sequence of singular values ​​arranged in descending order. Starting from the first largest singular value in the sequence, the values ​​are accumulated sequentially. When the sum of the accumulated values ​​reaches or exceeds 85% of the total sum of all singular values ​​for the first time, all singular values ​​from the first singular value to the current singular value are identified as core principal singular values. Each column of the left singular orthogonal matrix corresponding to the core principal singular value is labeled as a left singular vector. These left singular vectors are combined according to their columns to form the principal component eigenvector group of the project.

[0030] In the above process, singular value decomposition breaks down the original matrix into three sub-matrices with clear mathematical meaning. The left singular orthogonal matrix reflects the correlation strength between the enterprise project and the core features. The singular values ​​in the diagonal matrix quantify the information content in each core feature direction. The right singular orthogonal matrix reflects the weight relationship between the original feature fields and the core features. By calculating the sum of all singular values ​​as a measure of the total information, and accumulating the singular values ​​from largest to smallest, when the cumulative sum first reaches or exceeds 85%, it means that the selected core principal singular values ​​have contained at least 85% of the information in the original matrix. The remaining singular values ​​discarded mainly represent noise or secondary factors. This threshold screening method based on cumulative contribution rate not only ensures the full preservation of information, but also avoids the subjectivity and blindness of artificially setting a fixed number of principal components. The set 85% cumulative contribution rate threshold is a balanced value derived from experimental analysis and experience summarization of historical data from a large number of enterprise projects. Based on the information compression characteristics of singular value decomposition, when the cumulative contribution rate reaches 85%, the dimension of the original matrix can usually be compressed to one-fifth to one-tenth of its original size, for example, from 100 columns to 5 to 10 columns, while retaining most of the effective information. The less than 15% of information discarded is mainly noise and secondary factors. This avoids the loss of key information due to excessive dimensionality reduction, and also prevents the curse of dimensionality and model overfitting caused by insufficient dimensionality reduction. Secondly, it refers to the common practice of principal component analysis in the field of machine learning. In a large number of academic studies and industrial practices, 85% to 95% cumulative contribution rate is the most common range for principal component retention. Among them, 85% as the lower limit threshold can ensure sufficient information retention in most application scenarios. The column vectors of the left singular orthogonal matrix corresponding to the core principal singular values ​​are combined column by column to form the project principal component feature vector group. This reduces the feature vector dimension of each enterprise project from hundreds of columns in the original matrix to the number of core principal singular values, significantly reducing the dimensionality of the feature space. At the same time, these left singular vectors are orthogonal, meaning that different column vectors are independent of each other and have no information redundancy. This fundamentally eliminates the adverse effects of multicollinearity on subsequent prediction models, providing the LightGBM lightweight prediction model with high-quality input features that are low in dimensionality, rich in information, and free of redundancy.

[0031] S3: Input the project principal component feature vector group into three parallel, trained LightGBM lightweight prediction models respectively, and output the initial probability of cost overrun, initial probability of default risk, and initial probability of compliance violation; simultaneously obtain the historical evidence indicators corresponding to each initial probability, and calculate the comprehensive evidence strength accordingly. The overall strength of evidence is calculated as follows: Three parallel LightGBM lightweight prediction models are constructed, and each lightweight prediction model is trained. The project principal component feature vector group is input into the three parallel trained LightGBM lightweight prediction models, and the initial probability of cost overrun, initial probability of default risk, and initial probability of compliance violation are output respectively. In the above process, the LightGBM lightweight prediction model is trained. The training process includes: The model extracts principal component feature vectors of completed enterprise projects from the enterprise's historical project database as input features, and extracts the actual operating results of the corresponding projects as training labels. Specifically, the label for the cost overrun prediction model is whether a cost overrun occurred, the label for the default risk prediction model is whether a default risk occurred, and the label for the compliance and non-compliance prediction model is whether a compliance and non-compliance occurred. The labeled dataset is divided into training and validation sets proportionally. The training set is used for the model to learn the mapping relationship between input features and output labels, while the validation set is used to monitor the model's generalization ability. The model is iteratively trained on the training set, and evaluated on the validation set after each round of training. The model is considered complete when the prediction accuracy on the validation set stabilizes and no longer shows significant improvement after multiple rounds of training. After training, the three independent LightGBM lightweight prediction sub-models can respectively receive the principal component feature vectors of new projects as input and output the initial probabilities of cost overrun, default risk, and compliance and non-compliance.

[0032] We statistically analyzed the evidence indicators corresponding to the three initial probabilities for the enterprise's projects over the past three consecutive and complete calendar months. The evidence indicators corresponding to the initial probability of cost overrun were the cost consumption acceleration rate and the number of budget warnings triggered. The evidence indicators corresponding to the initial probability of default risk were the number of days of schedule deviation and the frequency of requirement changes. The evidence indicators corresponding to the initial probability of compliance and non-compliance were the number of audit issues and the number of contract penalty warnings triggered. In the above process, three consecutive and complete calendar months were selected as the statistical window because this time window can capture the latest trends in the project's operational status and accumulate sufficient data samples to smooth out occasional fluctuations and avoid misjudgments caused by abnormal data on a single day. For the initial probability of cost overruns, the cost consumption acceleration rate and the number of budget warning triggers were selected as evidentiary indicators because these two indicators reflect the precursory signals of cost runaway from the perspectives of actual cost consumption speed and abnormal budget execution frequency, respectively. For the initial probability of default risk, the number of days of schedule deviation and the frequency of requirement changes were selected as evidentiary indicators because schedule deviation directly reflects the ability to meet the project schedule, while the frequency of requirement changes indirectly reflects the stability of the project scope; together, they constitute the core early warning signal of default risk. For the initial probability of compliance violations, the number of audit findings and the number of contract penalty warnings were selected as evidentiary indicators because the number of audit findings reflects weaknesses in compliance management, while the number of contract penalty warnings directly reflects the severity of contractual violations; both, from different perspectives, jointly characterize the level of compliance risk exposure. After marking cost overruns, default risk events, and compliance violations as risk events, the dataset of each evidence indicator for all projects in the three months prior to the occurrence of the risk event is extracted from the enterprise's historical project database. First, the median of each evidence indicator in the dataset is calculated as the adjustment threshold parameter for that evidence indicator, and the interquartile range of each evidence indicator is also calculated. Then, for each evidence indicator, the original value is subtracted from the adjustment threshold parameter to obtain the deviation value. The quotient obtained by dividing 4 by the interquartile range is used as the scaling factor. The scaling factor is multiplied by the deviation value to obtain the scaled deviation. Then, the exponential function value with the natural constant as the base and the exponent as the scaled deviation is calculated. The denominator is obtained by adding 1 to the exponential function value. Finally, the standardized evidence strength of the evidence indicator is obtained by dividing 1 by the denominator. The formula upon which the above process is based is: in, Indicates the first Adjustment threshold parameters for each evidence indicator; An index representing the evidence indicators; Indicates the first Interquartile range of each evidence indicator; Indicates the first The original values ​​of each evidence indicator; Indicates the first Standardized strength of evidence for each evidence indicator; In the above process, the dependent variable Indicates the first The standardized strength of evidence for a given evidence indicator means that the original value of that evidence indicator is used to determine the strength of evidence. After nonlinear transformation, it is mapped to The closer the score is to 1, the stronger the indicative power of the evidence indicator to the risk event; the closer the score is to 0, the weaker the indicative power of the evidence indicator to the risk event. The technical effect is to unify various evidence indicators with different dimensions and distribution characteristics to the same numerical scale, making it possible to calculate the comprehensive evidence strength across indicators. At the same time, the smoothing properties of the Sigmoid function avoid the drastic disturbance of the fusion result by extreme outliers in the original data. The independent variable in the formula is This refers to the difference between the original value of the evidence indicator and the adjusted threshold parameter. This independent variable influences the standardized evidence strength because the greater the deviation of the original value of the evidence indicator from its historical median (adjusted threshold parameter), the more significant the difference between the current project's performance and its historical normal level. This significant difference is a crucial early warning signal of an impending risk event; therefore, this deviation is designed as the core variable driving changes in standardized evidence strength. The interquartile range in the denominator... As a scaling factor, it is used to automatically adjust the steepness of the Sigmoid function according to the dispersion of historical data of different evidence indicators, so that the variation of standardized evidence strength is adapted to the fluctuation characteristics of each indicator. Regarding positive and negative correlations, the strength of standardized evidence With independent variable A positive correlation is observed: when the raw values ​​of the evidence indicators... Greater than the adjustment threshold parameter hour, The value is positive, and the larger the positive value, the stronger the standardized evidence. The closer the indicator is to 1, the stronger the evidence points to a risk event; conversely, the closer the original value of the indicator is to 1, the stronger the evidence. Less than the adjustment threshold parameter hour, The stronger the standardized evidence, the greater the absolute value of the negative value. The closer the value is to 0, the weaker or even reversed the evidence that the indicator points to a risk event.

[0033] In the above embodiments, 15 sets of data for the original values ​​of the evidence indicators and the corresponding standardized evidence strength are given to reflect the changes in the standardized evidence strength as the original values ​​of the evidence indicators change, as shown in Table 1: Table 1: Relationship between raw values ​​of evidence indicators and corresponding standardized evidence strengths In Table 1 above, the adjustment threshold parameters are set. And given the interquartile range You can see when the original value of the evidence indicator... Greater than the adjustment threshold parameter Standardized strength of evidence The closer the value of the evidence indicator is to 1, the stronger the evidence points to a risk event; when the original value of the evidence indicator... Less than the adjustment threshold parameter Standardized strength of evidence The closer a value is to 0, the weaker the evidence that the indicator points to a risk event.

[0034] Meanwhile, using standardized evidence strength as input and actual occurrence of risk events as output labels, a logistic regression model is constructed and trained. After training, the absolute values ​​of the obtained feature regression coefficients are normalized to obtain the corresponding weight coefficients for each evidence indicator. In the above process, when constructing the logistic regression model, the input and output of the model are first clarified: the input is the standardized evidence strength of each evidence indicator, and the output is whether the risk event corresponding to that evidence indicator actually occurred, i.e., a binary label, with a value of 0 or 1; specifically, for the cost overrun dimension, the input is the standardized evidence strength of the cost consumption acceleration rate and the standardized evidence strength of the number of budget warning triggers, and the output is whether the cost overrun event actually occurred; for the default risk dimension, the input is the standardized evidence strength of the schedule deviation days and the standardized evidence strength of the requirement change frequency, and the output is whether the default risk event actually occurred; for the compliance and non-compliance dimension, the input is the standardized evidence strength of the number of audit findings and the standardized evidence strength of the number of contract penalty warnings triggered, and the output is whether the compliance and non-compliance event actually occurred.

[0035] During the data preparation phase, all sample data extracted from the enterprise's historical project database are randomly divided into training and validation sets in a 7:3 or 8:2 ratio. The training set is used for model parameter learning, while the validation set is used to monitor the model's generalization performance. During model training, the logistic regression model continuously updates the feature regression coefficients on the training set using an iterative optimization algorithm. After each round of training, the model makes predictions on the validation set and calculates evaluation metrics. The model is considered complete when the loss function value on the validation set no longer decreases for multiple rounds (e.g., ten rounds), or when the prediction accuracy on the validation set no longer improves for multiple rounds. At this point, the model has learned the optimal mapping relationship between input features and output labels. After training, the feature regression coefficients corresponding to each evidence indicator are extracted from the trained logistic regression model. The absolute value of the coefficient is taken and normalized to obtain the weight coefficient of each evidence indicator.

[0036] For each type of initial probability, including the initial probability of cost overrun, the initial probability of default risk, and the initial probability of compliance violation, first determine the set of evidence indicators corresponding to that type of initial probability. Then, for each evidence indicator in the set, multiply the standardized evidence strength of the evidence indicator by its corresponding weight coefficient to obtain the weighted evidence strength of the evidence indicator. Finally, add the weighted evidence strengths of all evidence indicators in the set, and the sum is the comprehensive evidence strength of that type of initial probability.

[0037] The formula upon which the above process is based is as follows: in, Indicates the first The set of evidence indicators corresponding to the initial probability of a class, where D is the index of the initial probability class. And take integers, D=1, 2, 3 represent the initial probability of cost overrun, the initial probability of default risk, and the initial probability of compliance violation, respectively; Indicates the first The overall strength of evidence for the initial probability of a class; Indicates the first The weighting coefficients corresponding to each evidence indicator.

[0038] In the above process, by weighting and summing the standardized evidentiary strengths of multiple evidentiary indicators within the same dimension with their respective weighting coefficients, the scattered, multi-dimensional evidentiary information is integrated into a single comprehensive evidentiary strength. This achieves the intensive expression and quantitative assessment of multi-source evidence under each risk dimension. Specifically, for the cost overrun dimension, the standardized evidentiary strength of the cost consumption acceleration rate and the standardized evidentiary strength of the number of budget warning triggers are multiplied by their respective weighting coefficients and then summed. This allows the two evidentiary indicators to contribute different amounts of information based on their respective importance to the cost overrun risk—indicators with higher importance receive greater weight and dominate the comprehensive evidentiary strength, while indicators with lower importance receive less weight and only serve as auxiliary references. Similarly, for the default risk dimension, the standardized evidentiary strength of the number of days of schedule deviation and the standardized evidentiary strength of the frequency of requirement changes are weighted and summed, allowing the risk signals from the two dimensions of project execution and requirement stability to be organically integrated. For the compliance and non-compliance dimension, the standardized evidentiary strength of the number of audit findings and the standardized evidentiary strength of the number of contract penalty warnings are weighted and summed, allowing the compliance risks from both internal audit findings and external contract defaults to be comprehensively assessed. This weighted summation method preserves the original evidence strength information of each evidence indicator, while also achieving differentiated treatment of the importance of different evidence indicators through weighting coefficients.

[0039] S4: Based on the comprehensive evidence strength, each initial probability is corrected to obtain the probability of cost overrun, the probability of default risk, and the probability of compliance violation; the entropy weights of each type of probability after correction are calculated, and the predicted probability of the enterprise project's operating information is determined based on the entropy weights. Finally, the operating status of the enterprise project is evaluated based on the predicted probability of the operating information.

[0040] The probability of predicting the business information of a company's project is determined as follows: For each type of initial probability, firstly, obtain the comprehensive strength of evidence corresponding to that type of initial probability, and calculate the Beta function value with the comprehensive strength of evidence as a parameter when the risk event occurs, and the Beta function value with the comprehensive strength of evidence as a parameter when the risk event does not occur; then, multiply the type of initial probability by the Beta function value when the risk event occurs to obtain the first product result, and multiply the difference between the first and second initial probabilities by the Beta function value when the risk event does not occur to obtain the second product result; finally, divide the first product result by the sum of the first and second product results, and the quotient is the corrected probability of that type of initial probability; The formula used in the above process is: in, in, Indicates the first Initial probability of class , , These represent the initial probability of cost overrun, the initial probability of default, and the initial probability of compliance violation, respectively. and They represent respectively with The parameter represents the value of the Beta function when the risk event occurs and when it does not occur; , These represent the first and second shape parameters of the Beta function when each type of risk event occurs and does not occur, respectively. , These represent the first and second shape parameters of the Beta function when each type of risk event does not occur, respectively, and are obtained through the method of moments. express The probability obtained after correcting the initial probability of the class. , , These represent the probability of cost overrun, the probability of default, and the probability of compliance violations, respectively. In the above process, firstly, two sets of data corresponding to each type of risk event are extracted from the enterprise's historical project database. The first set is the comprehensive evidence strength value corresponding to all projects that actually experienced this type of risk event, and the second set is the comprehensive evidence strength value corresponding to all projects that did not experience this type of risk event. Then, the parameters of the Beta distribution are estimated independently for these two sets of data. Given a set of comprehensive evidence strength sample data, first calculate the arithmetic mean of the sample, then calculate the variance of the sample. The arithmetic mean of the sample is denoted as the mean. Then calculate the variance of this sample group and denote it as . The mathematical properties of the Beta distribution show that the mean of the distribution is related to the shape parameter. and Existence Relationship: The variance of the distribution is related to its shape parameters. Using these two simultaneous relationships, the estimated values ​​of α and β can be derived in reverse: by calculating the comprehensive evidence strength samples for the risk event occurrence group and the risk event non-occurrence group respectively, the corresponding shape parameters can be obtained. , as well as , ; In the formula and The text describes the likelihood of the current overall strength of evidence occurring under conditions of both the occurrence and non-occurrence of the risk event—when the overall strength of evidence is high, The value will be significantly greater than This makes the corrected probability Greater than the initial probability That is, to correct upwards; conversely, when the overall strength of evidence is low, The value will be significantly smaller than This makes the corrected probability Less than the initial probability This means downward correction; when the overall strength of evidence is at a moderate level, the ratio of the two probabilities is close to 1, the correction magnitude is small, and the corrected probability is basically the same as the initial probability. The technical advantage of this Bayesian correction mechanism is that the initial probability of the model... This represents a baseline judgment based on historical big data, equivalent to a "prior belief"; while the overall strength of evidence... This represents the actual observed signal based on the recent performance of the current project, equivalent to a "likelihood evidence"; the two are multiplied by Bayes' theorem and normalized to obtain the final corrected probability output. This approach achieves a dynamic balance between historical patterns and real-time evidence. When recent evidence is strong, the revised probability deviates significantly from the initial probability to reflect the latest risk changes; when recent evidence is weak or consistent with historical patterns, the revised probability relies more on the initial probability to maintain prediction stability. This approach effectively addresses the technical shortcomings of traditional prediction methods, such as static model outputs and inability to respond to recent dynamic changes, significantly improving the timeliness and accuracy of risk prediction.

[0041] For each type of probability—cost overrun probability, default risk probability, and compliance violation probability—first, calculate the information entropy of that type of probability. Specifically, multiply the probability by its base-2 logarithm, add one minus the difference of that type of probability multiplied by its base-2 logarithm, and then take the negative of the sum to obtain the information entropy of that type of probability. Then, calculate the difference between one and the information entropy to obtain the entropy weight numerator of that type of probability. Finally, divide the entropy weight numerator of that type of probability by the sum of the entropy weight numerators of all types of probabilities, and the quotient is the entropy weight of that type of probability. The formula upon which the above process is based is: in, in, express The information entropy of the probability obtained after correcting the initial probability of the class; express The entropy weight of the probability obtained after correcting the initial probability of the class; In the above process, the dependent variable Indicates the first The entropy weight of the modified probability is the objective weight of the probability in the whole risk assessment system. The larger the entropy weight, the greater the amount of information contained in the probability and the higher the contribution to the prediction probability of the final business information. Its technical effect is that it realizes the complete data-driven weight allocation through the information entropy theory, avoiding the bias and instability caused by the reliance on subjective experience to set weights in traditional methods. when Information entropy approaches 0 or approaches 1. A value approaching 0 indicates a high degree of certainty and a large amount of information regarding the probability of that type; when... When it approaches 0.5, the information entropy... Reaching a maximum value of 1 indicates that the uncertainty of this type of probability is very high and the information content is very low; regarding positive and negative correlations, entropy weight... With information entropy There is a negative correlation: when the information entropy of a certain type of corrected probability is... When the value is relatively small, it indicates that the distribution of this type of probability is relatively certain and contains a large amount of information. The value is large, which makes the calculated entropy weight A larger information entropy indicates that this type of probability receives a higher weight in the final fusion; conversely, a smaller information entropy indicates a lower probability after modification. A larger value indicates that the distribution of this type of probability is relatively ambiguous and contains less information. The value of is small, which makes the calculated entropy weight The lower the weight, the more important the risk dimension is in the final fusion. This negative correlation ensures that the risk dimension with greater information and higher certainty plays a more important role in the final decision, which is consistent with the business logic of risk prediction.

[0042] A balance coefficient is defined. Multiplying this coefficient by the project stage weight for that probability class yields the first product. Multiplying the difference between the first and second products by the entropy weight for that probability class yields the second product. Adding the first and second products gives the numerator of the fusion weight for that probability class. Adding the numerators of the fusion weights for all three probability classes gives the denominator of the fusion weight. Dividing the numerator by the denominator gives the fusion weight for that probability class. Finally, multiplying the cost overrun probability by its fusion weight gives the cost overrun weighted value; multiplying the default risk probability by its fusion weight gives the default risk weighted value; and multiplying the compliance / non-compliance probability by its fusion weight gives the compliance / non-compliance weighted value. Adding these three weighted values ​​gives the final predicted probability of the enterprise project's operational information. The formula upon which the above process is based is: in, in, Indicates the probability of predicting business information; express The fusion weights of the probabilities obtained after correcting the initial class probabilities; Indicating the project phase, The entropy weights of the probabilities obtained after adjusting the initial probabilities are as follows: If the enterprise project is in the construction phase, the entropy weight of the cost overrun probability is 60%, the entropy weight of the default risk probability is 30%, and the entropy weight of the compliance / non-compliance probability is 10%; if the enterprise project is in the operation phase, the entropy weight of the cost overrun probability is 20%, the entropy weight of the default risk probability is 50%, and the entropy weight of the compliance / non-compliance probability is 30%; if the enterprise project is in the final stage, the entropy weight of the cost overrun probability is 10%, the entropy weight of the default risk probability is 30%, and the entropy weight of the compliance / non-compliance probability is 60%. This represents the balance coefficient between entropy weight and fusion weight. ; In the above process, The range of values ​​is That is, λ is any real number in the closed interval [0,1]. At that time, the fusion weights are entirely determined by the entropy weights. Decision: Weighting of experience at each project phase It has no effect; at this point, the determination of the weights relies entirely on data-driven information entropy, suitable for scenarios with sufficient and high-quality historical data, where it is desirable to completely avoid subjective experience bias. At that time, the fusion weight was entirely determined by the experience weight of the project phase. Decision, entropy weight It has no effect at this stage; the determination of weights relies entirely on the experience and judgment of business experts, and is suitable for the initial deployment phase when historical data is insufficient or data quality is poor. At this time, entropy weight and project phase experience weight each account for half, achieving a balanced integration of data-driven approaches and business experience. In practical applications, it is recommended to use this approach during initial deployment. As a default setting, after accumulating sufficient historical data, a grid search method can be used to traverse the range from 0 to 1 with a step size of 0.1. The value of λ is chosen to maximize the historical prediction accuracy as the final value. Dependent variable This represents the predicted probability of an enterprise's project's operational information. It means that the final risk score is obtained by combining the corrected probabilities of three dimensions—cost overrun, default risk, and compliance violation—with their respective weights. The score ranges from 0 to 1. The closer it is to 1, the higher the overall operational risk of the project, and the closer it is to 0, the lower the overall operational risk of the project. Entropy weight With the corrected probability The connection is established through information entropy—the probability after a certain dimension is corrected. When the entropy approaches 0 or 1, the information entropy is relatively small, and the entropy weight... A larger value indicates that the dimension contains a lot of information and should be given a higher weight; when When the entropy approaches 0.5, the information entropy is relatively large, and the entropy weight... A smaller value indicates high uncertainty in that dimension, and therefore it should receive a lower weight. Project phase experience weight. Directly related to the project's operational cycle stage, it reflects the business importance of each risk dimension at different stages. Balance coefficient Used to adjust the entropy weight and the project stage experience weight in the fusion weight. The relative proportion in; Business Information Prediction Probability Corrected probabilities for each dimension There is a positive correlation, and the weights are integrated. With entropy weight Weighting of experience in project phases They are all positively correlated, that is, when the entropy weight or project stage experience weight of a certain dimension is larger, its fusion weight is also larger, and the influence of that dimension in the final probability is also greater. During the construction phase, the main task of the project is to complete the physical construction of the engineering works. Significant funds are continuously invested in purchasing materials, paying for labor, and leasing equipment. Cost control pressure is greatest, and the risk of cost overruns is highest. Therefore, the weight of the probability of cost overruns is set to the highest at 60%. During the construction phase, payments are typically made in stages according to the contract, making the risk of default relatively controllable. Compliance issues have not yet been fully exposed; therefore, the weights for default risk and compliance violation probability are set at 30% and 10%, respectively. During the operation phase, the project has entered a stable operating stage, the cost structure tends to be fixed, and the risk of cost overruns is relatively reduced. However, project payment collection, cash flow stability, and contract fulfillment become the core concerns, and the risk of default increases significantly. Therefore, the weight of the probability of default risk is set to the highest at 50%, while the weights for cost overruns and compliance violation probabilities are adjusted to 20% and 30%, respectively. In the final stages, as the project nears delivery and acceptance, most costs have been incurred, and the risk of cost overruns is minimized. However, the final acceptance, documentation, and final payment stages involve numerous compliance requirements, making compliance risk a primary concern. Therefore, the weight of the compliance probability is set to the highest level of 60%, while the weights for cost overruns and default risks are reduced to 10% and 30%, respectively. This dynamic weighting strategy improves the probability of accurate final business information predictions. It can adaptively adjust the level of attention to each risk dimension according to different stages of the project, achieving a precise match between risk prediction and the actual needs of project management.

[0043] If the predicted probability of business information is greater than a preset threshold, an operational risk warning is generated; otherwise, the business project is deemed to be operating normally.

[0044] The preset threshold mentioned above refers to a risk assessment threshold used to determine the operational status of a company's projects. A risk warning is triggered when the predicted probability of operational information exceeds this threshold; otherwise, the operational status is considered normal. This preset threshold can be determined based on industry benchmarks, i.e., by referring to the risk tolerance standards commonly used by companies in the same industry. For example, setting the threshold to 0.7 means that a warning is triggered when the risk probability exceeds 70%. The specific value can be fine-tuned according to the company's own risk appetite; that is, risk-averse companies can appropriately lower the threshold, while companies with a stronger risk tolerance can appropriately increase it.

[0045] Please see Figure 3 The present invention also provides an enterprise project operation information prediction system based on structured data, the prediction system being used to execute the above-described prediction method, including: The data acquisition module is used to collect multi-source structured data of enterprise project operations, and to divide and analyze the multi-source structured data through the data source classification method to obtain time-series operation data and cross-sectional operation data; The matrix decomposition module is used to divide time-series and cross-sectional business data into dimensions, construct unimodal matrices for each dimension, process each unimodal matrix using matrix correlation alignment, and generate an effective coupling matrix; the singular value decomposition tool is called to decompose the effective coupling matrix to extract the principal component feature vector group of the project. The evidence strength calculation module is used to input the project principal component feature vector group into three parallel, trained LightGBM lightweight prediction models, and output the initial probability of cost overrun, initial probability of default risk, and initial probability of compliance violation. It also simultaneously acquires historical evidence indicators corresponding to each initial probability and calculates the comprehensive evidence strength accordingly. The information prediction module is used to correct each initial probability based on the comprehensive evidence strength to obtain the probability of cost overrun, the probability of default risk, and the probability of compliance violation; calculate the entropy weight of each type of probability after correction, and determine the predicted probability of the enterprise project's operating information based on the entropy weight; and finally evaluate the operating status of the enterprise project based on the predicted probability of the operating information.

[0046] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0047] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0048] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0049] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes 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.

Claims

1. A method for predicting enterprise project operation information based on structured data, characterized in that, The specific steps include: S1: Collect multi-source structured data on enterprise project operations, and classify and analyze the multi-source structured data through data source classification to obtain time-series operation data and cross-sectional operation data; S2: Divide the time-series and cross-sectional business data into dimensions, construct a single-modal matrix for each dimension, process each single-modal matrix using the matrix correlation alignment method to generate an effective coupling matrix; call the singular value decomposition tool to decompose the effective coupling matrix to extract the principal component feature vector group of the project. S3: Input the project principal component feature vector group into three parallel, trained LightGBM lightweight prediction models respectively, and output the initial probability of cost overrun, initial probability of default risk, and initial probability of compliance violation; simultaneously obtain the historical evidence indicators corresponding to each initial probability, and calculate the comprehensive evidence strength accordingly. S4: Based on the comprehensive evidence strength, each initial probability is corrected to obtain the probability of cost overrun, the probability of default risk, and the probability of compliance violation; the entropy weights of each type of probability after correction are calculated, and the predicted probability of the enterprise project's operating information is determined based on the entropy weights. Finally, the operating status of the enterprise project is evaluated based on the predicted probability of the operating information.

2. The method for predicting enterprise project operation information based on structured data according to claim 1, characterized in that, The specific steps for obtaining time-series and cross-sectional business data are as follows: The multi-source structured data includes internal enterprise structured data and external enterprise structured data; For the enterprise's internal structured data, if it contains continuous timestamps and is dynamically updated with the business cycle, it is classified as internal time-series candidate data; if it is static data containing only a single timestamp, it is classified as internal cross-sectional candidate data; the business cycle is divided into construction period, operation period and completion period according to the enterprise's project progress percentage. For external structured data of an enterprise, if it has a preset high-frequency update characteristic, it is classified as external time-series candidate data. The high-frequency update characteristic includes daily updates and monthly updates. If the external structured data of the enterprise is a static baseline value, it is classified as external cross-sectional candidate data. The external and internal time-series candidate data are aligned according to the timestamp. Data with the same enterprise project ID and the same timestamp are matched, and data with missing timestamps are removed to generate time-series operational data. The internal and external cross-section candidate data are matched according to the same enterprise project ID to generate cross-section business data.

3. The method for predicting enterprise project operation information based on structured data according to claim 1, characterized in that, The time-series and cross-sectional operating data are dimensionally divided, including operating cost, risk warning, and compliance assessment dimensions. The corresponding time-series and cross-sectional fields under each dimension are then obtained, based on the following logic: For the operating cost dimension, fields related to cost fluctuations are extracted from time-series operating data as time-series fields, and fields related to cost benchmarks are extracted from cross-sectional operating data as cross-sectional fields. For risk warning, daily net capital inflow and monthly progress lag days are extracted from time-series operating data as time-series fields, and asset-liability ratio, interest coverage ratio and industry default rate are extracted from cross-sectional operating data as cross-sectional fields. For the compliance assessment dimension, daily compliance inspection scores and rectification completion rates are extracted from time-series operating data as time-series fields, and annual violation counts and industry compliance pass rates are extracted from cross-sectional operating data as cross-sectional fields. Among them, daily net cash inflow equals the total cash inflow of a company's projects on that day minus the total cash outflow on that day; monthly progress lag days refer to the difference between the actual completion date and the planned completion date of a company's projects within the monthly statistical period; compliance inspection score refers to the percentage score given by the company after conducting daily compliance inspections of its projects; rectification completion rate equals the number of issues that have been rectified in a company's projects divided by the total number of issues that should be rectified; annual violation count represents the total number of violations that a company's projects have been formally identified as violations by regulatory agencies in the past year; and industry compliance pass rate equals the proportion of companies that have passed compliance inspections to the total number of companies in the same industry. To obtain the correlation strength between time-series and cross-sectional fields within each dimension, the following steps are taken: For each pair of time-series and cross-sectional fields within each dimension, using the same natural month as the same time period, data samples from the same enterprise project within that natural month are collected. First, the Pearson correlation coefficient between the time-series and cross-sectional fields is calculated, and the weight parameters assigned by expert scoring are obtained. Then, a balancing coefficient is set, and the balancing coefficient is multiplied by the Pearson correlation coefficient to obtain the first product result. The difference between the first and the balancing coefficient is multiplied by the expert scoring weight parameters to obtain the second product result. Finally, the first product result and the second product result are added together, and the sum is the correlation strength between the pair of time-series and cross-sectional fields. The logic for calculating the Pearson correlation coefficient is as follows: For a set of time-series and cross-sectional fields within any dimension, collect multiple samples of the same project within the same time period from the company's historical project database. Calculate the average of all samples in the time-series field and the average of all samples in the cross-sectional field. Then, calculate the deviation of the corresponding value in the time-series field from its average value for each sample pair, multiplied by the deviation of the corresponding value in the cross-sectional field from its average value. Sum these products across all sample pairs to obtain the numerator. Next, calculate the square root of the sum of squares of the deviations in the time-series field and the sum of squares of the deviations in the cross-sectional field, multiplying these two square root results to obtain the denominator. Finally, divide the numerator by the denominator to obtain the Pearson correlation coefficient.

4. The method for predicting enterprise project operation information based on structured data according to claim 3, characterized in that, Based on the correlation strength, a single-modality matrix is ​​constructed for each dimension, specifically as follows: First, construct a correlation table. Each row in the correlation table records the dimension name, the name of the time-series field, the name of the cross-sectional field, and the correlation strength. Based on the correlation table, construct a unimodal matrix for each dimension using the unimodal matrix construction method: Set the number of enterprise projects to be analyzed as the number of rows in the unimodal matrix, and set the sum of the number of time-series fields and the number of cross-sectional fields as the number of columns in the unimodal matrix. Extract the values ​​of the time-series fields for each enterprise project at each timestamp from the time-series operating data and fill them into the corresponding row and column positions; extract the values ​​of the cross-sectional fields for each enterprise project from the cross-sectional operating data and fill them into the corresponding row and column positions; finally, fill the corresponding association strengths from the association table into the corresponding columns of the unimodal matrix.

5. The method for predicting enterprise project operation information based on structured data according to claim 1, characterized in that, Generate an effective coupling matrix, specifically as follows: Matrix normalization is applied to the unimodal matrix of each dimension to eliminate dimensional differences between different matrices. For each unimodal matrix, Min-Max normalization is used to map all elements to the [0,1] interval, resulting in a normalized unimodal matrix. Based on each normalized unimodal matrix, matrix correlation alignment is used to generate an effective coupling matrix. The standardized unimodal matrix of operating cost dimension is selected as the baseline matrix. Based on the row index of the enterprise project number in the baseline matrix, the rows of the same enterprise project number in the standardized unimodal matrices of risk warning dimension and compliance assessment dimension are matched and aligned in turn to ensure that the data of the same enterprise project is in the same row. Then, the three aligned matrices are horizontally spliced ​​in the column direction. That is, all columns of operating cost dimension are placed first, and then all columns of risk warning dimension and compliance assessment dimension are placed in turn to form a new matrix with the same number of rows and the same number of columns as the sum of the number of columns of the three dimensions. This is the effective coupling matrix.

6. The method for predicting enterprise project operation information based on structured data according to claim 1, characterized in that, Extract the principal component feature vector set of the project, specifically: The effective coupling matrix is ​​decomposed by calling the singular value decomposition tool into a left singular orthogonal matrix, a diagonal matrix, and a right singular orthogonal matrix. All non-zero singular values ​​are extracted from the diagonal of the diagonal matrix to obtain a sequence of singular values ​​arranged in descending order. Starting from the first largest singular value in the sequence, the values ​​are accumulated sequentially. When the sum of the accumulated values ​​reaches or exceeds 85% of the total sum of all singular values ​​for the first time, all singular values ​​from the first singular value to the current singular value are identified as core principal singular values. Each column of the left singular orthogonal matrix corresponding to the core principal singular value is labeled as a left singular vector. These left singular vectors are combined according to their columns to form the principal component eigenvector group of the project.

7. The method for predicting enterprise project operation information based on structured data according to claim 1, characterized in that, The overall strength of evidence is calculated as follows: Three parallel LightGBM lightweight prediction models are constructed, and each lightweight prediction model is trained. The project principal component feature vector group is input into the three parallel trained LightGBM lightweight prediction models, and the initial probability of cost overrun, initial probability of default risk, and initial probability of compliance violation are output respectively. We statistically analyzed the evidence indicators corresponding to the three initial probabilities for the enterprise's projects over the past three consecutive and complete calendar months. The evidence indicators corresponding to the initial probability of cost overrun were the cost consumption acceleration rate and the number of budget warnings triggered. The evidence indicators corresponding to the initial probability of default risk were the number of days of schedule deviation and the frequency of requirement changes. The evidence indicators corresponding to the initial probability of compliance and non-compliance were the number of audit issues and the number of contract penalty warnings triggered. After marking cost overruns, default risk events, and compliance violations as risk events, the dataset of each evidence indicator for all projects in the three months prior to the occurrence of the risk event is extracted from the enterprise's historical project database. First, the median of each evidence indicator in the dataset is calculated as the adjustment threshold parameter for that evidence indicator, and the interquartile range of each evidence indicator is also calculated. Then, for each evidence indicator, the original value is subtracted from the adjustment threshold parameter to obtain the deviation value. The quotient obtained by dividing 4 by the interquartile range is used as the scaling factor. The scaling factor is multiplied by the deviation value to obtain the scaled deviation. Then, the exponential function value with the natural constant as the base and the exponent as the scaled deviation is calculated. The denominator is obtained by adding 1 to the exponential function value. Finally, the standardized evidence strength of the evidence indicator is obtained by dividing 1 by the denominator. Meanwhile, using standardized evidence strength as input and actual occurrence of risk events as output labels, a logistic regression model is constructed and trained. After training, the absolute values ​​of the obtained feature regression coefficients are normalized to obtain the corresponding weight coefficients for each evidence indicator. For each type of initial probability, including the initial probability of cost overrun, the initial probability of default risk, and the initial probability of compliance violation, first determine the set of evidence indicators corresponding to that type of initial probability. Then, for each evidence indicator in the set, multiply the standardized evidence strength of the evidence indicator by its corresponding weight coefficient to obtain the weighted evidence strength of the evidence indicator. Finally, add the weighted evidence strengths of all evidence indicators in the set, and the sum is the comprehensive evidence strength of that type of initial probability.

8. The method for predicting enterprise project operation information based on structured data according to claim 7, characterized in that, The probability of predicting the business information of a company's project is determined as follows: For each type of initial probability, firstly, obtain the comprehensive strength of evidence corresponding to that type of initial probability, and calculate the Beta function value with the comprehensive strength of evidence as a parameter when the risk event occurs, and the Beta function value with the comprehensive strength of evidence as a parameter when the risk event does not occur; then, multiply the type of initial probability by the Beta function value when the risk event occurs to obtain the first product result, and multiply the difference between the first and second initial probabilities by the Beta function value when the risk event does not occur to obtain the second product result; finally, divide the first product result by the sum of the first and second product results, and the quotient is the corrected probability of that type of initial probability; For each type of probability—cost overrun probability, default risk probability, and compliance violation probability—first, calculate the information entropy of that type of probability. Specifically, multiply the probability by its base-2 logarithm, add one minus the difference of that type of probability multiplied by its base-2 logarithm, and then take the negative of the sum to obtain the information entropy of that type of probability. Then, calculate the difference between one and the information entropy to obtain the entropy weight numerator of that type of probability. Finally, divide the entropy weight numerator of that type of probability by the sum of the entropy weight numerators of all types of probabilities, and the quotient is the entropy weight of that type of probability. For each type of probability—cost overrun probability, default risk probability, and compliance violation probability—the project stage weight is first determined based on the current operational cycle stage of the enterprise's project: If in the construction phase, the project stage weight for cost overrun probability is 60%, default risk probability is 30%, and compliance violation probability is 10%; if in the operation phase, the project stage weight for cost overrun probability is 20%, default risk probability is 50%, and compliance violation probability is 30%; if in the finalization phase, the project stage weight for cost overrun probability is 10%, default risk probability is 30%, and compliance violation probability is 60%. A balance coefficient is set. The balance coefficient is multiplied by the project stage weight for that probability type to obtain the first product. The difference between the first and second products is multiplied by the entropy weight for that probability type to obtain the second product. The first and second products are added together to obtain the numerator of the fusion weight for that probability type. The numerators of the fusion weights for all three probability types are added together to obtain the denominator of the fusion weight. The quotient of the numerator and denominator of the fusion weight is the fusion weight for that probability type. Finally, the cost overrun probability is multiplied by its fusion weight to obtain the cost overrun weighted value, the default risk probability is multiplied by its fusion weight to obtain the default risk weighted value, and the compliance / non-compliance probability is multiplied by its fusion weight to obtain the compliance / non-compliance weighted value. The sum of these three weighted values ​​is the final predicted probability of the enterprise project's operating information. If the predicted probability of business information is greater than a preset threshold, an operational risk warning is generated; otherwise, the business project is deemed to be operating normally.

9. A structured data-based enterprise project operation information prediction system, wherein the structured data-based enterprise project operation information prediction system is used to implement the structured data-based enterprise project operation information prediction method according to any one of claims 1-8, characterized in that: The data acquisition module is used to collect multi-source structured data of enterprise project operations, and to divide and analyze the multi-source structured data through the data source classification method to obtain time-series operation data and cross-sectional operation data; The matrix decomposition module is used to divide time-series and cross-sectional business data into dimensions, construct unimodal matrices for each dimension, process each unimodal matrix using matrix correlation alignment, and generate an effective coupling matrix; the singular value decomposition tool is called to decompose the effective coupling matrix to extract the principal component feature vector group of the project. The evidence strength calculation module is used to input the project principal component feature vector group into three parallel, trained LightGBM lightweight prediction models, and output the initial probability of cost overrun, initial probability of default risk, and initial probability of compliance violation. It also simultaneously acquires historical evidence indicators corresponding to each initial probability and calculates the comprehensive evidence strength accordingly. The information prediction module is used to correct each initial probability based on the comprehensive evidence strength to obtain the probability of cost overrun, the probability of default risk, and the probability of compliance violation, respectively. The entropy weights of various probabilities after correction are calculated, and the predicted probability of business information for enterprise projects is determined based on the entropy weights. Finally, the business status of enterprise projects is evaluated based on the predicted probability of business information.

Images