Equipment manufacturing enterprise operation risk assessment method based on Bayesian network

Abstract

The invention provides an equipment manufacturing enterprise operation risk assessment method based on a Bayesian network, and the method comprises the following steps: S1, obtaining the operation data of an equipment manufacturing enterprise, and carrying out the integration of the operation data, and obtaining a multi-source fusion data set; s2, extracting risk factors in the multi-source fusion data set; s3, adopting a recurrent neural network to identify the incidence relation between the risk factors; s4, constructing a business process model, and associating the risk factor with the business process to form a Bayesian network; and S5, based on the Bayesian network, calculating conditional value-at-risk indexes CVaR alpha of the target node under different confidence levels alpha so as to evaluate the risk level of the equipment manufacturing enterprise operation full-service process. And the operation risk cognition elasticity of an equipment manufacturing enterprise in an unstable and uncertain environment is improved.

Description

Technical Field

[0001] The present invention relates to the field of equipment manufacturing enterprise operation risk assessment, and in particular to an equipment manufacturing enterprise operation risk assessment method based on a Bayesian network. Background Art

[0002] Equipment manufacturing enterprises face highly random, complex, dynamic, uncertain, systematic and joint operational risks due to their industry characteristics, such as complex supply chains, high capital investment and fluctuations in market demand, and present a complex situation of multi-source superposition and multiple transmission. As the cornerstone of industrial development, the operational risk assessment of the equipment manufacturing industry is crucial for early warning and prevention of risks in enterprise operations. Therefore, it is necessary to incorporate complex and dynamic risk scenarios into risk assessment considerations and develop operational risk assessment methods for equipment manufacturing enterprises that are adapted to unstable and uncertain environments.

[0003] Existing risk assessment research mainly focuses on safety risk management in specific industries, such as electric power, engineering construction, transportation, and natural disasters. These studies usually focus on identifying risk factors with clear physical damage, and identify risk factors through methods such as fault tree analysis and expert surveys. In addition, researchers have also developed a variety of risk assessment methods, including model-based methods and artificial intelligence methods, to improve the accuracy and reliability of risk assessment. Overall, existing research has the following shortcomings:

[0004] First, many methods rely on a large amount of historical data, which may lead to the loss of accuracy of the assessment results and insufficient generalization ability when the data is incomplete or biased. Secondly, risk assessment methods based on complex models such as artificial intelligence will increase the difficulty of implementation due to the difficulty of explaining the process, which may cause equipment manufacturing companies to react slowly when facing risks. In addition, most of the existing risk accident databases only disclose safety risk factors and cannot be applied to situations where risk factors are complexly and dynamically associated in unstable and uncertain environments. Finally, most of the existing risk assessment methods are aimed at a single risk event, a single safety accident or a specific operating stage of the enterprise, and there is no systematic coverage of the risks of the entire business process of equipment manufacturing enterprises. The multi-source superposition and multi-transmission operational risk characteristics of equipment manufacturing enterprises make it impossible for existing risk assessment methods to meet their dynamic risk assessment needs.

[0005] Therefore, it is necessary to consider the operational risk characteristics of equipment manufacturing enterprises and construct a scientific and comprehensive risk assessment model to improve the cognitive resilience of equipment manufacturing enterprises in operational risks under unstable and uncertain environments. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides an equipment manufacturing enterprise operation risk assessment method based on Bayesian network, comprising the following steps:

[0007] Step S1: After obtaining and integrating the equipment manufacturing enterprise operation data, a multi-source fusion data set is obtained;

[0008] Step S2: extracting risk factors from the multi-source fusion data set;

[0009] Step S3: using a recurrent neural network to identify the correlation between the risk factors;

[0010] Step S4: constructing a business process model, associating the risk factors with the business process to form a Bayesian network;

[0011] Step S5: Based on the Bayesian network, calculate the conditional value at risk indicator CVaR of the target node at different confidence levels α α , in order to assess the risk level of the entire business process of equipment manufacturing enterprises.

[0012] Preferably, step S2 comprises the following steps: extracting keywords and phrases from the multi-source fusion data set as a candidate risk phrase library by using an N-gram model; clustering the candidate risk phrase library by using an LDA model to obtain hierarchical risk factors.

[0013] Preferably, the method of extracting keywords and phrases from the multi-source fusion data set as a candidate risk phrase library by using an N-gram model comprises the following steps:

[0014] Step S21: Set the N value of the N-gram model to 2, 3, and 4, and obtain a set of candidate risk phrases A=A1, A2, A3, ... A including 2-gram, 3-gram, and 4-gram. n , count the number of occurrences of each candidate risk phrase A and calculate the frequency, recorded as FQ(A);

[0015] Step S22: Calculate the comprehensive importance score (A) of the risk phrase A using the following formula:

[0016]

[0017] Where EMI(A) represents the mutual information value, EL(A) represents the left information entropy, ER(A) represents the right information entropy, total_EMI, total_EL and total_ER are the sum of the mutual information values, the sum of the left information entropy and the sum of the right information entropy of the candidate risk phrases respectively;

[0018] Step S23: sorting according to the comprehensive importance and selecting the first M as the candidate risk phrase library.

[0019] Preferably, step S3 includes: using the risk factors as nodes of a recurrent neural network; converting the text data in the multi-source fusion data set into numerical features using a bag-of-words model and word embedding, and then inputting the numerical features into the recurrent neural network to obtain the correlation between the risk factors.

[0020] Preferably, the method for constructing a business process model in step S4 includes: identifying the business activities of each link in the equipment manufacturing enterprise operation system as task nodes in the notation BPMN model; the sequential, conditional, parallel, merged and loop control flows between business processes are represented by connecting lines in BPMN; the relationships between business processes across links and within links are expressed by message flows and data objects in BPMN.

[0021] Preferably, the method of associating the risk factor with the business process model to form a Bayesian network in step S4 comprises the following steps:

[0022] Step S41: Using Dirichlet allocation LDA to obtain the document d in the multi-source fusion data set i The topic distribution and theme t j The vocabulary distribution in

[0023] Step S42: Through each document d i The topic distribution Calculate for each document d i Business activities k Related risk factors j The probability P(r j |d i , a k ):

[0024]

[0025]

[0026] In the formula, P(r j |t j ) indicates that in the subject t j Lower risk factor r j The probability, P(t j |d i , a k ) means in the document d i Medium Theme j and business activities k The joint probability of T represents the set of topics related to all business activities. represents the risk factor r j The corresponding vocabulary set, P(w|tj ) indicates that word w is in topic t j The probability under , R represents the set of risk factors;

[0027] Step S43: The probability P(r) of the risk factor related to the business activity in each document calculated in step S42 j |d i , a k ), calculate the probability P(r j |a k ), the formula is as follows, when the calculation result P(r j |a k ) exceeds a threshold τ, then this risk factor is selected as the key risk factor of this business activity:

[0028]

[0029] Step S44: The key risk factors and business activities are used as nodes of the Bayesian network, and the associations between business activities and risk factors, and the associations between key risk factors connecting business activities are used as edges of the Bayesian network.

[0030] Preferably, when the multi-source fusion data set in step S1 changes, the probability of each risk factor is dynamically updated by the following formula:

[0031]

[0032] Where P t (r j |a k ) means that at time t, in a given target business activity a k Under the premise of j The update probability of t-1 (r j |a k ) indicates that at time t-1, a given target business activity a k Under the premise of j The probability of; α is the smoothing factor; D t represents the new multi-source fusion dataset, d ti Represents the document D at time t t The i-th document in .

[0033] Preferably, step S5 comprises the following steps:

[0034] Step S51: Assume that the loss of risk factor i under risk scenario combination j is L ij, then by dynamically weighting all loss values, update the loss mean μi and variance To adjust the loss function L(i) parameters, L(i)~N(μ,σ 2 ), the update formula of the loss function is:

[0035]

[0036] In the formula, w j (t) is the weight of scene j at time t, satisfying N represents the number of scenes;

[0037] Step S52: Based on the loss function dynamically updated in step S51, combined with the joint probability distribution of the Bayesian network, the loss value of the business activity node to be risk assessed, which is transmitted from the parent node level to the target node, is calculated, and the loss function L(s) of the target node is dynamically updated. The calculation formula is as follows:

[0038]

[0039] Where p represents the number of parent nodes of the target node, q represents the number of parent nodes of the parent node of the target node; β n represents the conditional probability of the nth parent node of the target node; L(S n ) represents the loss function of the nth parent node of the target node; β nj represents the conditional probability of the jth parent node of the nth parent node of the target node; L(s nj ) represents the loss function of the j-th parent node of the n-th parent node of the target node;

[0040] Step S53: Dynamically calculate the VaR value of the target node loss:

[0041] VaR α (L t )=μ t +Φ -1 (α)σ t ;

[0042] Among them, Φ -1 is the inverse cumulative distribution function of the standard normal distribution; α represents the confidence level; L t represents the loss function at time t; σ t Represents the standard deviation of the loss function at time t;

[0043] Step S54: According to VaR α (L t ) Dynamically calculate the CVaR value of the target node at the confidence level α, and the calculation formula is as follows:

[0044]

[0045] Step S55: By dynamically calculating the system CVaR value, the risk level of the entire business process of the equipment manufacturing enterprise under multi-source superimposed risk events is evaluated.

[0046] The beneficial effects of the present invention include at least:

[0047] First, by integrating multimodal data from different sources, various risk factors in corporate operations and the correlation between factors can be more comprehensively captured, thereby improving the comprehensiveness and timeliness of risk assessment.

[0048] Second, by using the N-gram model, LDA model and recurrent neural network (RNN) model, multi-source data can be intelligently mined, which enhances the ability to automatically mine and analyze risk information and reduces dependence on professional knowledge.

[0049] Third, the business process is deconstructed to clarify the dependency between business processes and risk factors, thereby improving the risk assessment level of equipment manufacturing enterprises under the situation of multi-source superposition and multiple transmission, making enterprise risk assessment closer to actual operating conditions.

[0050] Fourth, by dynamically establishing a risk factor loss function and using the conditional value at risk (CVaR) indicator, the operational risk level at different confidence levels is calculated, thereby improving the scientificity and accuracy of risk assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a diagram of the architecture of the present invention for implementing operational risk assessment of equipment manufacturing enterprises using multimodal data mining and dynamic Bayesian networks;

[0052] Figure 2 A flowchart for implementing the risk assessment of equipment manufacturing enterprises’ operations for this risk assessment method;

[0053] Figure 3 is the risk factor association network;

[0054] Figure 4 5 typical business process structures for equipment manufacturing enterprises;

[0055] Figure 5 Provide business process characterization results for equipment manufacturing enterprises;

[0056] Figure 6 Bayesian network structure for risk factors of procurement business process;

[0057] Figure 7 Bayesian network for risk factors of procurement business process. DETAILED DESCRIPTION

[0058] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of the present invention.

[0059] like Figure 1 and Figure 2 As shown, an embodiment of the present invention provides an equipment manufacturing enterprise operation risk assessment method based on a Bayesian network, comprising the following steps:

[0060] Step S1, introduce multimodal data mining technology to comprehensively reflect the operational risk status of equipment manufacturing enterprises, and intelligently mine the operational risk factors of the entire business process of the enterprise and the correlation between risk factors.

[0061] Specifically, it includes: using customized scripts and web crawler technology to crawl structured and unstructured data related to equipment manufacturing enterprises; using Azur e Efficiently integrate with the computing platform and preprocess multi-source mixed data to improve data transmission and storage efficiency; use Python's pandas library, Apache NiFi, Tesseract OCR, Google Cloud Speech-to-Text technology, and OpenCV library to process data; identify risk factors in the equipment manufacturing industry based on the N-gram model and LDA model, and mine the potential relationships caused by risks based on the recurrent neural network (RNN) model. Based on the review of technical personnel, a risk factor association network for equipment manufacturing enterprises is proposed.

[0062] 1) In the process of operational risk assessment of equipment manufacturing enterprises, key data sources include unstructured data such as official media reports, annual reports and semi-annual reports of listed companies, and relevant audio and video materials of enterprise interviews. In addition, structured data such as inventory management data and customer order information are also important sources for constructing operational risk Bayesian networks. First, for internal system data, customized scripts are used to extract data from systems such as ERP and MES; for external data sources, web crawler technology is used to obtain information from official websites. Secondly, Python is used to convert unstructured data into text data, and the extracted text data is integrated with the existing structured data; Azure Data Factory is used to integrate data from different systems to achieve the integration and processing of key data sources such as ERP, MES, and SCM. Azure Event Hubs is used as a message queue to ensure real-time data collection and transmission, which enhances the scalability of the system. In terms of data storage, AzureSQL Database is used for data redundancy and load balancing, combined with the automatic expansion function of cloud services to ensure that the data platform has high availability and can meet the storage needs of large amounts of data. Data security uses Azure Key Vault for data encryption and access control, and Azure Policy is used for automated compliance checks to ensure the security of data during transmission and storage. In addition, through Azure Databricks, data virtualization technology is implemented, so users can perform complex operations and analysis without moving data, which improves the flexibility of the system. Edge computing technology is implemented through Azure IoT Edge, which deploys data processing and analysis tasks near the data source, reducing data transmission delays and improving real-time performance.

[0063] 2) The following is an explanation of the steps for processing unstructured data. First, use Python's pandas library combined with automated scripts to clean the data, remove duplicate records, correct errors, and fill in missing values. Second, use Apache NiFi to build a data flow pipeline to unify data in different formats into a standard format for subsequent processing. Third, use TesseractOCR to convert paper documents into digital text; use Google Cloud Speech-to-Text technology to perform speech recognition on recordings; use OpenCV to process image data, and try to extract key visual features through denoising and enhancement. Through 1) and 2), an operational risk data set for equipment manufacturing enterprises is established to identify risk factors and relationships between factors.

[0064] 3) In the embodiment of the present invention, the risk factor identification process includes: wherein the identification and semantic analysis of risk factors are key steps. First, the embodiment of the present invention extracts keywords and phrases as candidate risk factors from a large amount of text data through the N-gram model. Subsequently, the candidate risk factors are clustered using the LDA model to extract deeper risk factors. Finally, the risk factors are modified in combination with expert corrections, and the final risk factors are determined from the output of the deep learning model. Finally, 91 risk factors in 7 categories of equipment manufacturing enterprises were identified. The types and numbers of some risk factors and their classifications are shown in Table 1.

[0065] Table 1 Classification and number of risk factors

[0066]

[0067] The embodiment of the present invention conducts statistical analysis on the risk text after preprocessing and finds that the probability of occurrence of a phrase is directly related to the word frequency of the risk phrase extracted by the N-gram algorithm. Based on this observation, the embodiment of the present invention introduces word frequency as an important statistical indicator and combines it with the mutual information value, left information entropy and right information entropy to form a multi-dimensional indicator phrase mining technology suitable for the risk text of this study. The specific steps are as follows:

[0068] The first step is to calculate multi-dimensional indicators. Use the N-gram algorithm to extract preliminary candidate risk phrases, and set the N value to 2, 3, and 4 to obtain a set of candidate risk phrases A = A1, A2, A3, ... A including 2-gram, 3-gram, and 4-gram. n , count the number of occurrences of each candidate risk phrase A, and calculate the frequency, recorded as FQ(A).

[0069] Mutual information is used to quantify the dependencies between words. A higher mutual information value means that the n-gram words are more closely connected and therefore more likely to form a phrase. The mutual information value calculation formula (1) is as follows:

[0070]

[0071] Where n represents the number of word units in the candidate risk phrase A, FQ(A i ) represents word unit A i Compared with the traditional mutual information index, the square root parameter in formula (1) is changed to n, which avoids the possibility that the mutual information value increases without limit as n increases, and also eliminates the need for additional normalization of different N-gram candidate risk phrases, thereby simplifying the calculation process and reducing potential errors.

[0072] Left information entropy and right information entropy are used to measure the instability of the left and right boundaries of a candidate risk phrase. In information theory, information entropy is used to describe the uncertainty of a random variable. Therefore, if the left / right information entropy of a candidate phrase is higher, it is more likely to be a suitable phrase. The formulas for left information entropy and right information entropy are as follows:

[0073]

[0074] l and r are the left and right adjacent words of the candidate risk phrase, respectively; L represents the set of all possible adjacent words on the left of A, and R represents the set of all possible adjacent words on the right of A. p(lA|A) represents the conditional probability that l and A occur simultaneously when A occurs; p(Ar|A) represents the conditional probability that A and r occur simultaneously when A occurs. When EL(A) and ER(A) are high, it means that the left and right adjacent words of the candidate risk phrase A are more diverse, and A is more likely to be considered as a complete phrase.

[0075] The second step is phrase mining. The score value represents the comprehensive importance of the phrase, which is calculated by the sum of the word frequency, normalized mutual information value, left information entropy and right information entropy. The calculation formula is as follows:

[0076]

[0077] total_EMI, total_EL, and total_ER are the sum of the mutual information values, the sum of the left information entropy, and the sum of the right information entropy of the candidate risk phrases, respectively. e The scoring values ​​are used to rank the candidate risk phrases in order of importance, and domain experts perform manual assisted judgment and screening to extract reusable equipment manufacturing industry risk factors.

[0078] 4) The risk factor relationship mining process is specifically explained. First, the text data in the multi-source data set is converted into numerical feature data using the Bag of Words (BoW) model and word embeddings, and then input into the recurrent neural network to mine the association relationship. Then, by using the risk factors as nodes, the data is trained using the recurrent neural network (RNN) model to capture the time dependency and dynamic changes between risk factors. Furthermore, by calculating the similarity between hidden states at different time steps, the association between risk factors is identified. After collecting new multi-source data, the above data mining method is used to dynamically adjust the risk factors and the association between factors in the operational risk Bayesian network.

[0079] Finally, the analysis results of the graph neural network are visualized using the Gephi tool. Through visualization, risk factors and their interrelationships are clearly displayed, allowing decision makers to quickly identify key risk points and potential risk transmission paths. The present invention inputs 3976 binary risk factor association relationships into the graph analysis tool Gephi, classifies risk factors into 7 levels according to the structure of risk factors, and constructs a risk factor association network. Some risk factor association networks are shown in Figure 2. Figure 3 shown.

[0080] In an embodiment of the present invention, the SHAP value is used to enhance the interpretability of the recurrent neural network (RNN) model in the risk assessment of the equipment manufacturing industry. The SHAP (SHapley Additive exPlanations) value is based on the Shapley value in game theory, which can be used to quantify the contribution of each feature to the model prediction results. After model training and prediction, the SHAP library is used to calculate the contribution of each risk factor to the model prediction results. The specific operation is: first, for each text data sample in the data warehouse, the changes in the model output are observed by slightly changing the feature value of the sample. Secondly, the average of these changes is calculated to obtain the SHAP value of each feature. Finally, a detailed explanatory report is created in combination with the recognition results of the SHAP value.

[0081] Step S2, constructing a dynamic operational risk Bayesian network structure based on the entire business process of equipment manufacturing enterprises.

[0082] Specifically, it includes: analyzing the key links in the business process, standardizing business process modeling based on the Business Process Model and Notation (BPMN), and modeling and analyzing the entire business process of equipment manufacturing enterprises, including procurement, production, marketing, and operation and maintenance; dynamically identifying the key risk factors of each business activity through text mining technology, and constructing an operational risk Bayesian network structure based on business processes according to the correlation between business activities.

[0083] 1) In the embodiment of the present invention, the business process model and notation (BPMN) is used to deconstruct and characterize the business process of the equipment manufacturing enterprise. Specifically, based on the typical business process structure, the uncertain risks existing on the process nodes, and the data flow and information transmission between the processes, the business process of the equipment manufacturing enterprise is modeled and characterized, forming a full business process operation system for the operation of equipment manufacturing enterprises covering the four major business links of procurement, production, marketing, and operation and maintenance, and five typical business process structures of sequential processes, conditional branches, parallel gateways, merged gateways, and loop structures, realizing the effective deconstruction and characterization of the equipment manufacturing enterprise operation system based on the multi-layer heterogeneous network model.

[0084] It should be noted that, by referring to BPMN for business process modeling, the process pattern of the business activity layer can be expressed as a triple BP=(Activity, Flow, Gateway), where Activity represents each specific task in the process; Flow represents the connection between process nodes, and defines the execution order and dependency relationship between each activity, such as sequential flow, parallel flow, etc.; Gateway describes the logical relationship between different connection edges, such as the mutual influence of conditional gateways and parallel gateways.

[0085] The full business process system of equipment manufacturing enterprises contains complex and diverse business process structures and forms. Five typical business process structures can be used as reference Figure 4 shown.

[0086] The business activities of each link in the operation system of the equipment manufacturing enterprise are identified as task nodes in the BPMN model, and the control flows such as sequence, condition, and parallelism between business processes are represented by connecting lines in BPMN. The relationships between business processes across links and within links, such as information transmission, data flow, resource exchange, decision impact, and risk transmission, are expressed through message flows and data objects in BPMN. Based on the analysis of the processed multi-source fusion data set of equipment manufacturing enterprises, combined with interview surveys, business process tracking and other methods, the four links of procurement, production, marketing, and operation and maintenance are determined, which contain 9, 15, 10, and 14 business processes respectively, with a total of 10 cross-link message flow transmission relationships. Taking the procurement and production business processes as an example, the characterization results can be seen in Figure 5 .

[0087] 2) Establish the dependency relationship between business processes and risk factors through text mining technology and construct a dynamic Bayesian network.

[0088] Specifically, a text mining technology based on topic model, namely Latent Dirichlet Allocation (LDA), is used to identify the potential topics of risk documents in multi-source fusion data sets and establish the relationship between business activities and risk factors.

[0089] First, in the multi-source fusion equipment manufacturing enterprise operation risk data set in S1, there is a risk document set D = {d1, d2, ..., d n}, each document consists of a vocabulary w. The LDA model identifies a set of topics T = {t1, t2, ..., t k} and determine the distribution of these topics in each document. Through LDA, we can obtain the document d i The topic distribution in and theme t j The vocabulary distribution in The LDA model is expressed as:

[0090]

[0091] Among them, P(w|t j ) indicates that in the subject t j The probability of the next word w, P(t j |d i ) indicates document d i Medium Theme j probability.

[0092] By training the LDA model, we can obtain the topic distribution of each document. and the vocabulary distribution for each topic To convert business activity A = {a1, a2, ..., a m}Corresponding to the topic T, the following formula is defined to calculate each document d through the topic distribution of each document i Business activities k Related risk factors j Probability:

[0093]

[0094] Among them, P(r j |t j ) indicates that in the subject t j Lower risk factor r j The probability, P(t j |d i , a k ) means in the document d i Medium Theme j and business activities k The joint probability of .

[0095] Based on the 91 identified risk factors R = {r1, r2, ..., r 91}, for each risk factor r j Build it in the theme j The probability distribution under is the risk factor r j The corresponding vocabulary set can be calculated by the following formula:

[0096]

[0097] Among them, P(w|t j ) is the word w in the subject t j The probability of is the risk factor r j A collection of words.

[0098] Then, the probability P(r j |d i , a k ), calculate the probability P(r j |a k ), the formula is as follows, when the calculation result P(r j |a k ) exceeds a threshold τ, then this risk factor is selected as the key risk factor of this business activity:

[0099]

[0100] In the embodiment of the present invention, a probability update mechanism based on a time window is introduced to achieve dynamic update and identification. Assume that at time t, a new stream data set D t ={d t1 , d t2 , ..., d tn}, and dynamically update the probability of each risk factor through the following formula:

[0101]

[0102] Among them, P t (r j |a k ) means that at time t, in a given target business activity a k Under the premise of j The update probability of t-1 (r j |a k ) indicates that at time t-1, a given target business activity a k Under the premise of j The probability of; α is the smoothing factor, which is used to balance the influence of previous probability and new observation data; d ti Represents the document D at time t t The i-th document in .

[0103] Using the above method, the key risk factors associated with each business activity can be effectively and dynamically extracted from the 91 identified risk factors. Taking the procurement business process as an example, the key risk factors of business activities such as market demand assessment errors and delayed raw material delivery time are determined. After the key risk factors of the business activities are determined according to the above process, the key risk factors and business activities are used as nodes of the Bayesian network, and the association between business activities and risk factors is established as one part of the edge of the Bayesian network. Then, because the relationship between business activities has been determined in the previous BPMN, the association between their key risk factors can be connected according to this business activity relationship as another part of the edge of the Bayesian network, that is, the structure of the Bayesian network is determined. Figure 6 As shown in the figure, for example, the risk of substandard product quality may affect the risk of shortage of key raw materials, and this influence relationship is represented by directed edges in the network. In this way, the Bayesian network can capture the conditional dependencies between risk factors and provide a structured representation method for risk assessment.

[0104] It should be noted that the input data required for this step are the multi-source fused equipment manufacturing enterprise operational risk data set and the identified equipment manufacturing enterprise operational risk factors in S1, and the business process structure constructed in S2. The output of this step is the operational risk Bayesian network structure based on the business process.

[0105] Step S3, construct the risk factor loss function and evaluate the conditional value at risk (CVaR) indicator. Specifically, it includes: fitting the risk factor value parameters under various risk scenarios into a normal distribution function, recorded as the loss function of the risk factor; combining the joint probability distribution formula of the Bayesian network, calculating the loss value transmitted from the parent node level to the target node, thereby obtaining the loss function of the target node; calculating the CVaR of the target node under different confidence levels α. α , in order to assess the risk level of the entire business process of equipment manufacturing enterprises.

[0106] 1) Based on the risk value parameters of risk factors when multiple sources of risk occur concurrently, the loss function of risk factors is dynamically constructed.

[0107] Specifically: According to the Central Limit Theorem m ), when the total loss of the system is accumulated by a large number of independent and identically distributed small losses, its total loss tends to be normally distributed. The normal distribution has simple mathematical properties and good asymptotic properties in statistics, which can effectively capture the mean and variance of the loss distribution. Therefore, the normal distribution model is used to fit the possible loss of the risk factor.

[0108] 1) The risk value parameters of risk factors under different risk scenarios can be dynamically fitted into a normally distributed loss function L(i), L(i)~N(μ,σ 2 ) to quantify the possible loss of risk factors. Let the loss of risk factor i under risk scenario combination j be L ij , then by dynamically weighting all loss values, update the loss mean μi and variance Define its update formula as:

[0109]

[0110] Among them, w j (t) is the weight of scene j at time t, satisfying Through the above dynamic weighting and updating mechanism, the loss function L(i) parameters are adjusted in real time according to new data to quantify the possible loss of risk factors. Table 2 shows the parameters μ and σ of the loss function of some risk factor nodes in the Bayesian network for operational risk assessment of equipment manufacturing enterprises constructed using the procurement business process as an example. 2 .

[0111] Table 2 Risk factor loss function parameter table

[0112]

[0113] It should be noted that the input data required for this step is the risk factor risk value parameter dynamically updated in S3, and the output of this step is the loss function of the risk factor.

[0114] 2) Evaluate the risk level of the entire business process of enterprise operations under multi-source superimposed risk events. Specifically, in a Bayesian network, the CVaR of each node can be calculated separately according to the calculation formula, but the CVaR of the entire system cannot be simply obtained by adding the CVaR values ​​of each node. CVaR is a risk metric that considers losses in extreme situations, which may occur when multiple nodes influence each other.

[0115] In a Bayesian network, each node is not completely independent, but has a certain correlation between them. For each node i, the loss function L(i) is dynamically updated. Combined with the joint probability distribution of the Bayesian network, the loss value of the business activity node to be risk assessed, which is passed from its parent node level to the target node, can be calculated, and the loss function L(s) of the target node is dynamically updated. The calculation formula is as follows:

[0116]

[0117] Where p represents the number of parent nodes of the target node, q represents the number of parent nodes of the parent node of the target node; β n represents the conditional probability of the nth parent node of the target node; L(S n ) represents the loss function of the nth parent node of the target node; β nj represents the conditional probability of the jth parent node of the nth parent node of the target node; L(s nj ) represents the loss function of the j-th parent node of the n-th parent node of the target node.

[0118] On this basis, the VaR value of the target node loss is dynamically calculated:

[0119] VaR α (L t )=μ t +Φ -1 (α)σ t (13)

[0120] Among them, Φ -1 is the inverse cumulative distribution function of the standard normal distribution, L t represents the loss function at time t; σ t Represents the standard deviation of the loss function at time t.

[0121] According to VaR α (L t ) Dynamically calculate the CVaR value of the target node at the confidence level α, and the calculation formula is as follows:

[0122]

[0123] Among them, VaR u (L t ) represents the VaR value at the confidence level u.

[0124] In risk management, a higher confidence level means that more extreme loss situations are considered when calculating CVaR. As the confidence level increases, the CVaR value usually increases. Through the above steps, the dynamic calculation of the system CVaR value can be achieved, thereby evaluating the risk level of the entire business process of equipment manufacturing enterprises under multi-source superimposed risk events.

[0125] The embodiment of the present invention also provides the following method, which dynamically adjusts the structure, network node conditional probability and risk value parameters based on historical data and real-time data, so that the Bayesian network evaluation model has self-learning and self-adaptive properties.

[0126] Specifically, it includes: deriving the node parameters of the operational risk Bayesian network based on the business process operational risk Bayesian network structure and improved Bayesian formula, and cross-validating based on the prior probability and posterior probability of the risk factors. Dynamically adjusting the risk factors and the correlation between factors of the operational risk Bayesian network based on the collected multi-source stream data makes it self-learning and adaptive.

[0127] 1) Based on the improved Bayesian formula, the node parameters of the operational risk Bayesian network are derived. Specifically, under the set specific risk scenario, the node parameters of the operational risk Bayesian network are derived based on the operational risk Bayesian network structure of the business process and the improved Bayesian formula.

[0128] In the embodiment of the present invention, the Bayesian network calculation formula considering the time factor is used to calculate the node probability. Assume that there are three risk factors A, B and C in the Bayesian network, and the possible values ​​of these three events are Y or N, indicating that the risk factor "occurs" or "does not occur". The values ​​of random variables A and B are a i and b j , then when A and B are in an AND relationship, the probability P(C=Y, t) of the random variable C taking the value Y in time t is calculated as shown in formula (15), where P(C=Y|A=a i，t B=b j，t ) means at time t, given A = a i and B = b j Under the condition of , the probability of C taking the value of Y (i.e. the risk occurring) is:

[0129]

[0130] The above formula is used to calculate the probability of event C occurring at time t given that risk factors A and B occur simultaneously (ie, an AND relationship).

[0131] Since there are not only AND relationships but also OR relationships in the case of multiple branches in the business process, in this case, the probability P(C=Y) of the random variable C taking the value Y is calculated as shown in formula (16):

[0132]

[0133] The above formula is used to calculate the probability of event C occurring at time t given that risk factor A or B occurs (ie, an OR relationship). 1-[1-P(C=Y|A=a i，t )][1-P(C=Y|B=b j，t)] represents the probability that event C does not occur at time t, given that at least one of A and B occurs. Then take the complement of this probability, which is the probability of event C occurring. Through the above formula, the conditional probability of the risk factor node in the Bayesian network of equipment manufacturing enterprise operation risk can be calculated.

[0134] Specifically, the embodiment of the present invention uses the improved Choquet fuzzy integral to calculate the risk value of the risk factor node in the Bayesian network of the equipment manufacturing enterprise operation risk. The traditional Choquet fuzzy integral does not conform to the reality that the joint harm degree between multiple risk factors exceeds the harm degree of a single risk factor. Therefore, in order to reflect the mutual reinforcement effect between risk factors, the embodiment of the present invention believes that the fuzzy measure μ satisfies the following specific conditions:

[0135] ①

[0136] ②

[0137] where μ is a set function on the power set P(X); X = {x1, x2, ..., x n} represents a non-empty finite set of risk factors at the risk superposition point; |X| represents the number of risk factors in the set X.

[0138] As the set X expands, the number of fuzzy measures grows exponentially. The superposition risk calculation method based on the improved Choquet fuzzy integral is:

[0139]

[0140] Y is the risk factor x1, x2…x n The total risk of interaction; f(x i ) represents the risk factor x i The risk generated satisfies 0≤f(x1)≤f(x2)≤…≤f(x n ), f(x0) = 0. According to the formula, the risk value parameters of risk factors under different risk scenario combinations can be calculated.

[0141] Self-learning and self-adaptation of the Bayesian network assessment model for operational risk. Specifically, it includes: dynamically mining historical data and real-time data generated by equipment manufacturing, automatically updating the evolution of risk factors and the correlation between risk factors into the Bayesian network, that is, dynamically adjusting the Bayesian network structure of operational risk through real-time collection of multi-source stream data mining. Further, by introducing a reinforcement learning algorithm that considers time factors, incremental learning is performed based on data, thereby dynamically adjusting weights to increase adaptability to the environment.

[0142] It should be noted that a reinforcement learning algorithm that considers the importance of time nodes is introduced to dynamically adjust the Bayesian network. In order to take into account the importance of nodes before this moment, the present invention uses a weighted strategy to adjust the reward signal, giving nodes at earlier time steps greater weights. At the same time, new multi-source data is collected and the model is updated using an incremental learning method. The time range of historical data is limited by sliding window technology to maintain the real-time and adaptability of the model. Finally, the hyperparameters are tuned according to actual conditions to obtain the best weight control effect, including the setting of time weights and learning rates. The adjusted operational risk Bayesian network can more accurately reflect the impact of historical data and improve the accuracy of risk assessment.

[0143] It should be noted that the Bayesian network can realize dynamic adjustment of node parameters by integrating the time factor. Specifically, it realizes dynamic adjustment with an improved reinforcement learning algorithm and realizes periodic updates with full-text data. When constructing and applying the operational risk Bayesian network, considering the time factor t enables the operational risk network to capture and reflect the risk status and probability distribution at a specific time point. By collecting newly added multi-source stream data, the operational risk Bayesian network can respond to the latest changes in the status of risk factors in a timely manner and automatically adjust its internal probability distribution, thereby realizing self-adaptation and self-learning. For example, if real-time data indicates that risk factor A has occurred, then the probability P (A = a i，t ) will be updated to 1, ensuring the timeliness and accuracy of the network. According to the hierarchical relationship of the operational risk Bayesian network, the business activity nodes in the network are set as primary indicators, and the risk factor nodes are set as secondary indicators. Then, through historical data simulation and expert discussion and correction, the conditional probability between the risk factor node and the business activity node is obtained, and finally the reference Figure 7 The operational risk Bayesian network shown is based on the procurement business process as an example.

[0144] It should be noted that this step is based on the business process-based operational risk factor transmission network constructed by S2. The initial risk transmission probability and distribution function of each business process are calculated through the structured data and survey data in the equipment manufacturing enterprise operational risk data set.

[0145] Through dynamic model self-learning and self-adaptation, new data from multiple sources of equipment manufacturing enterprises are collected to dynamically adjust the risk factors and relationships between factors of the operational risk Bayesian network; reinforcement learning algorithms that consider time factors and integrated data are used to dynamically adjust node parameters to improve the applicability and timeliness of risk assessment.

[0146] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. Only the preferred embodiments of the present invention are expressed. The description is more specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. As long as there is no contradiction in the combination of these technical features, they should be considered as the scope recorded in this specification.

[0147] It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for assessing operational risk of equipment manufacturing enterprises based on Bayesian networks, characterized in that: The following steps are involved: Step S1: After obtaining and integrating the equipment manufacturing enterprise operation data, a multi-source fusion data set is obtained; Step S2: extracting risk factors from the multi-source fusion data set; Step S3: using a recurrent neural network to identify the correlation between the risk factors; Step S4: constructing a business process model, associating the risk factors with the business process to form a Bayesian network; Step S5: Based on the Bayesian network, calculate the conditional value at risk indicator CVaR of the target node at different confidence levels α α , in order to assess the risk level of the entire business process of equipment manufacturing enterprises.

2. According to claim 1, a method for assessing operational risk of equipment manufacturing enterprises based on Bayesian networks is characterized by: Step S2 includes the following steps: extracting keywords and phrases from the multi-source fusion data set as a candidate risk phrase library through an N-gram model; clustering the candidate risk phrase library using an LDA model to obtain a hierarchical risk factor.

3. The method for assessing operational risk of equipment manufacturing enterprises based on Bayesian networks according to claim 2 is characterized by: The method for extracting keywords and phrases from the multi-source fusion data set as a candidate risk phrase library by using an N-gram model comprises the following steps: Step S21: Set the N value of the N-gram model to 2, 3, and 4 to obtain a set of candidate risk phrases A=A1, A2, A3, ... A including 2-gram, 3-gram, and 4-gram. n , count the number of occurrences of each candidate risk phrase A and calculate the frequency, recorded as FQ(A); Step S22: Calculate the comprehensive importance score (A) of the risk phrase A using the following formula: Where EMI(A) represents the mutual information value, EL(A) represents the left information entropy, ER(A) represents the right information entropy, total_EMI, total_EL and total_ER are the sum of the mutual information values, the sum of the left information entropy and the sum of the right information entropy of the candidate risk phrases respectively; Step S23: sorting according to the comprehensive importance and selecting the first M as the candidate risk phrase library.

4. According to claim 1, a method for assessing operational risk of equipment manufacturing enterprises based on Bayesian networks is characterized by: Step S3 includes: using the risk factors as nodes of a recurrent neural network; using a bag-of-words model and word embedding to convert text data in the multi-source fusion data set into numerical features, and then inputting the features into the recurrent neural network to obtain the correlation between the risk factors.

5. The method for assessing operational risk of equipment manufacturing enterprises based on Bayesian networks according to claim 1 is characterized in that: The method for constructing a business process model in step S4 includes: identifying the business activities of each link in the equipment manufacturing enterprise operation system as task nodes in the notation BPMN model; the sequence, condition, parallel, merge and loop control flows between business processes are represented by connecting lines in BPMN; the relationship between cross-link and intra-link business processes is expressed by message flows and data objects in BPMN.

6. The method for assessing operational risk of equipment manufacturing enterprises based on Bayesian networks according to claim 1 is characterized by: The method of associating the risk factor with the business process model to form a Bayesian network in step S4 includes the following steps: Step S41: Using Dirichlet allocation LDA to obtain the document d in the multi-source fusion data set i The topic distribution and theme t j The vocabulary distribution in Step S42: Through each document d i The topic distribution of each document d i Business activities k Related risk factors j The probability P(r j ∣d i ,a k ): In the formula, P(r j ∣t j ) indicates that in the subject t j Lower risk factor r j The probability, P(t j ∣d i ,a k ) means in the document d i Medium Theme j and business activities k The joint probability of T represents the set of topics related to all business activities. represents the risk factor r j The corresponding vocabulary set, P(w|t j ) is the word w in the subject t j The probability under , R represents the set of risk factors; Step S43: The probability P(r) of the risk factor related to the business activity in each document calculated in step S42 j ∣d i ,a k ), calculate the probability P(r j ∣a k ), the formula is as follows, when the calculation result P(r j ∣a k ) exceeds a threshold τ, then this risk factor is selected as the key risk factor of this business activity: Step S44: The key risk factors and business activities are used as nodes of the Bayesian network, and the associations between business activities and risk factors, and the associations between key risk factors connecting business activities are used as edges of the Bayesian network.

7. The method for assessing operational risk of equipment manufacturing enterprises based on Bayesian networks according to claim 6 is characterized by: When the multi-source fusion data set described in step S1 changes, the probability of each risk factor is dynamically updated by the following formula: Where P t (r j ∣a k ) means that at time t, in a given target business activity a k Under the premise of j The update probability of t-1 (r j ∣a k ) indicates that at time t-1, a given target business activity a k Under the premise of j The probability of; α is the smoothing factor; D t represents the new multi-source fusion dataset, d ti Represents the document D at time t t The i-th document in .

8. The method for assessing operational risk of equipment manufacturing enterprises based on Bayesian networks according to claim 1 is characterized by: Step S5 includes the following steps: Step S51: Assume that the loss of risk factor i under risk scenario combination j is L ij , then by dynamically weighting all loss values, update the loss mean μ i and variance To adjust the loss function L(i) parameters, L(i)~N(μ,σ 2 ), the update formula of the loss function is: In the formula, w j (t) is the weight of scene j at time t, satisfying N represents the number of scenes; Step S52: Based on the loss function dynamically updated in step S51, combined with the joint probability distribution of the Bayesian network, the loss value of the business activity node to be risk assessed, which is transmitted from the parent node level to the target node, is calculated, and the loss function L(s) of the target node is dynamically updated. The calculation formula is as follows: Where p represents the number of parent nodes of the target node, q represents the number of parent nodes of the parent node of the target node; β n represents the conditional probability of the nth parent node of the target node; L(S n ) represents the loss function of the nth parent node of the target node; β nj represents the conditional probability of the jth parent node of the nth parent node of the target node; L(s nj ) represents the loss function of the jth parent node of the nth parent node of the target node; Step S53: Dynamically calculate the VaR value of the target node loss: VaR α (L t )=μ t +F -1 (a)s t ; Among them, Φ -1 is the inverse cumulative distribution function of the standard normal distribution; α represents the confidence level; L t represents the loss function at time t; σ t Represents the standard deviation of the loss function at time t; Step S54: According to VaR α (L t ) Dynamically calculate the CVaR value of the target node at the confidence level α, and the calculation formula is as follows: Step S55: By dynamically calculating the system CVaR value, the risk level of the entire business process of the equipment manufacturing enterprise under multi-source superimposed risk events is evaluated.

Images