A Supply Chain Security Early Warning Method and System Based on Ensemble Learning

Abstract

This invention discloses a supply chain security early warning method and system based on ensemble learning. The method includes: acquiring basic feature vectors of enterprises in the supply chain and constructing a topological network corresponding to the supply chain; iteratively analyzing the supply chain and service chain in the supply chain based on the basic feature vectors of the enterprises to generate supply feature vectors and service feature vectors; concatenating the supply feature vectors and service feature vectors with the basic feature vectors of midstream enterprises to obtain supply aggregation indicators and service aggregation indicators of midstream enterprises; performing same-level aggregation based on the supply aggregation indicators and service aggregation indicators to obtain supply chain feature vectors and service chain feature vectors; training a machine learning model through ensemble learning to obtain an evaluation model; and predicting the input feature vectors through the evaluation model to obtain the supply chain risk level security early warning prediction result.

Description

Technical Field

[0001] This invention relates to the fields of industrial chain and machine learning technology, and in particular to an industrial chain security early warning method and system based on ensemble learning. Background Technology

[0002] The essence of an industry chain is to describe a group of enterprises with certain inherent connections; it is a relatively macro-level concept. Industry chains involve numerous upstream and downstream relationships and the exchange of value. Upstream links supply products or services to downstream links, while downstream links provide feedback to upstream links. Taking the automotive industry chain as an example, the upstream mainly consists of manufacturers providing core components, while midstream vehicle manufacturers use upstream products to produce cars. Consumers, after purchasing a car, still require services provided by downstream companies, such as car insurance and charging services. In this process, if upstream manufacturers encounter risks, it may affect the normal operation of midstream manufacturers, meaning a risk transmission may occur. Poor performance of downstream service companies may also affect the normal operation of midstream companies. Ensemble learning is a machine learning algorithm that, unlike traditional machine learning algorithms, first constructs a series of models and then merges these models to obtain a better-performing model. In ensemble learning, the basic model is the constructed series of models, corresponding to the basic classifier in a classification algorithm. The basic classifier is also called a weak classifier; fusing multiple weak classifiers can result in a strong classifier with better classification performance and stronger generalization ability.

[0003] Existing technologies primarily apply machine learning to risk analysis of individual companies within the industry chain. The general steps are as follows: 1. Construct a dataset consisting of multiple companies, each with a set of features. 2. Manually assign a risk level to each company based on its risk profile. 3. Extract a risk-related feature from each company and train the model using a classifier based on the previously obtained labels. 4. Use the resulting model to predict the risk level of new companies.

[0004] Using a single classifier to categorize individual companies based on their risk profile is a straightforward approach to risk analysis within a supply chain using machine learning. However, this approach has several drawbacks: 1. It relies solely on each company's risk-related characteristics, neglecting the impact of upstream and downstream companies on that company's risk. 2. Current supply chain datasets are limited and difficult to obtain; classification models built using only a single classifier may suffer from uneven data distribution and weak generalization ability, resulting in lower accuracy. 3. Classifying a single company does not reflect the overall situation of the supply chain. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a supply chain security early warning method and system based on ensemble learning. This method can aggregate risk information from upstream and downstream enterprises in the supply chain and use the aggregated results from different enterprises in the supply chain as input to the overall supply chain model. Furthermore, it improves the accuracy of supply chain risk prediction results through ensemble learning even with limited data.

[0006] To achieve the above technical objectives, the present invention provides the following technical solution: a supply chain security early warning method based on ensemble learning, comprising:

[0007] Obtain the basic feature vectors of all enterprises in the industry chain and construct the corresponding topology network of the industry chain, wherein the topology network includes several midstream enterprises and corresponding suppliers and service providers at several levels;

[0008] For suppliers and service providers, the basic feature vectors of the last-level enterprises are aggregated at the same level and across levels in sequence. The cross-level aggregation result is concatenated with the basic feature vector of the corresponding upper-level enterprise to obtain the enterprise feature vector of the upper-level enterprise. The enterprise feature vector is then aggregated at the same level and across levels again. By repeating the process of same-level aggregation, cross-level aggregation and concatenation, all suppliers and service providers are traversed to obtain the supply feature vector and service feature vector.

[0009] For midstream enterprises, the supply feature vector and service feature vector are concatenated with the basic feature vector of the midstream enterprise to obtain the supply aggregation index and service aggregation index of the midstream enterprise. Based on the supply aggregation index and service aggregation index, same-layer aggregation and cross-layer aggregation are performed to obtain the supply chain feature vector and service chain feature vector.

[0010] A machine learning model is trained using ensemble learning methods to obtain an evaluation model. The evaluation model is then used to predict the feature vectors of the supply chain and the service chain to obtain the safety early warning prediction results for the risk level of the industrial chain.

[0011] Optionally, the process of obtaining the basic feature vector includes:

[0012] Obtain indicator data from enterprises in the industrial chain, including data on innovation capabilities, operating conditions, self-controllability, and the development environment;

[0013] The indicator data is cleaned and processed, and the cleaned and processed data is quantified to obtain the basic feature vector.

[0014] Optionally, the process of constructing the topology network corresponding to the industry chain includes:

[0015] The system statistically analyzes all enterprises in the industry chain and the corresponding relationships between them. It then classifies all enterprises into hierarchical and categorized groups. Based on the classification results, it identifies midstream enterprises, treats all enterprises as network nodes, and uses the corresponding relationships between them as edges. Based on the network nodes and edges, it obtains the topological network corresponding to the industry chain.

[0016] Optionally, the processes of same-layer aggregation and cross-layer aggregation include:

[0017] Aggregate the feature vectors of enterprises at the same level and in the same category to obtain the category feature vector F':

[0018] F' = [max(F1,F2,…,F m ),min(F1,F2,…,F m ),avg(F1,F2,…,F m ),...]

[0019] Wherein F m This represents the enterprise feature vector of the m-th enterprise in the same level and category.

[0020] The category feature vectors are multiplied to obtain the cross-layer aggregation result;

[0021] In the process of same-level aggregation, for the last-level enterprise, the basic feature vector of the last-level enterprise is used as the enterprise feature vector for same-level aggregation. For the midstream enterprises, the supply aggregation index and service aggregation index are used as the enterprise feature vectors of the midstream enterprises for same-level aggregation.

[0022] Optionally, the evaluation model acquisition process:

[0023] Historical data is collected and converted into historical input feature vectors. The historical input feature vectors are then labeled, and the overall input feature vectors and labeling results are integrated into a training set.

[0024] Construct several weak classifiers, and train them sequentially using the Boosting method based on the training set until the number of trained weak classifiers reaches a preset number. Then, perform a weighted sum on the outputs of the trained weak classifiers to obtain the evaluation model.

[0025] To better achieve the above-mentioned technical objectives, the present invention also provides a supply chain security early warning system based on ensemble learning, comprising:

[0026] The system includes a data collection module, an indicator calculation module, a model training module, a model inference module, and a report generation module.

[0027] The data collection module is used to collect indicator data; the indicator calculation module is used to obtain the basic feature vectors of enterprises in the industry chain based on the indicator data, and construct the topology network corresponding to the industry chain, wherein the topology network includes several midstream enterprises and corresponding suppliers and service providers at several levels; for suppliers and service providers, the basic feature vectors of the last-level enterprises are aggregated at the same level and across levels in sequence, and the cross-level aggregation result is concatenated with the basic feature vector of the corresponding upper-level enterprise to obtain the enterprise feature vector of the upper-level enterprise, and the enterprise feature vector is aggregated at the same level and across levels again. By repeating the process of same-level aggregation, cross-level aggregation and concatenation, until all suppliers and service providers are traversed, the supply feature vector is obtained. The system generates supply and service feature vectors. For midstream enterprises, the supply feature vector and service feature vector are concatenated with the basic feature vector of the midstream enterprise to obtain the supply aggregation index and service aggregation index of the midstream enterprise. Based on the supply aggregation index and service aggregation index, same-layer aggregation and cross-layer aggregation are performed to obtain the supply chain feature vector and service chain feature vector. The model training module is used to train a machine learning model through ensemble learning methods to obtain an evaluation model. The model inference module is used to predict the supply chain feature vector and service chain feature vector through the evaluation model to obtain the industrial chain risk level safety early warning prediction result. The report generation module is used to generate a risk report based on the industrial chain risk level safety early warning prediction result.

[0028] Optionally, the indicator data in the data collection module includes innovation capability data, business status data, self-controllability data, and development environment data.

[0029] Optionally, the topology network construction process in the indicator calculation module includes: statistically analyzing the enterprises, enterprise categories, and corresponding relationships between enterprises in the industry chain; using the enterprises as network nodes and the corresponding relationships between enterprises as edges; and obtaining the topology network corresponding to the industry chain based on the network nodes and edges.

[0030] Optionally, the processes for same-level aggregation and cross-level aggregation in the indicator calculation module include:

[0031] Aggregate the feature vectors of enterprises at the same level and in the same category to obtain the category feature vector F':

[0032] F' = [max(F1,F2,…,F m ),min(F1,F2,…,F m ),avg(F1,F2,…,F m ),...]

[0033] Wherein F m This represents the enterprise feature vector of the m-th enterprise in the same level and category.

[0034] The category feature vectors are multiplied to obtain the cross-layer aggregation result;

[0035] In the process of same-level aggregation, for the last-level enterprise, the basic feature vector of the last-level enterprise is used as the enterprise feature vector for same-level aggregation. For the midstream enterprises, the supply aggregation index and service aggregation index are used as the enterprise feature vectors of the midstream enterprises for same-level aggregation.

[0036] Optionally, the evaluation acquisition process described in the model training module:

[0037] Historical data is collected and converted into historical input feature vectors. The historical input feature vectors are then labeled, and the overall input feature vectors and labeling results are integrated into a training set.

[0038] Construct several weak classifiers, and train them sequentially using the Boosting method based on the training set until the number of trained weak classifiers reaches a preset number. Then, perform a weighted sum on the outputs of the trained weak classifiers to obtain the evaluation model.

[0039] The present invention has the following technical effects:

[0040] This study focuses on midstream enterprises in the industrial chain, aggregating risk information about the current state of the chain by combining basic information, supply relationships, and service relationships of upstream and downstream enterprises. The aggregated results, along with those of midstream enterprises, serve as the overall input for risk analysis. Ensemble learning is then used to predict the risk level of the industrial chain. Based on this prediction, enterprises and management institutions can promptly develop response strategies to prevent localized risks from impacting the entire industrial chain. Ensemble learning is used because data on industrial chains is relatively scarce; it can achieve high prediction results on limited datasets and address the issue of uneven sample distribution within the dataset. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the topology network provided in an embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram of a system provided for an embodiment of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] like Figure 1 As shown, the core idea of ​​this invention is to focus on midstream enterprises in the industrial chain as the primary analysis object. It aggregates risk information about the current state of the industrial chain by combining basic information, supply relationships, and service relationships of upstream and downstream enterprises. The aggregated results, along with the aggregated results of midstream enterprises, serve as the overall input for risk analysis. Then, ensemble learning is used to predict the risk level. Enterprises and management institutions can promptly formulate response strategies based on the prediction results to prevent localized risks from impacting the entire industrial chain. Ensemble learning is used because data on the industrial chain is relatively scarce; it can achieve high prediction results on limited datasets and address the issue of uneven sample distribution within the dataset.

[0047] This invention mainly comprises three parts: risk data collection and feature generation for upstream, midstream, and downstream enterprises in the industrial chain; aggregation and transmission of risk information along the supply chain and service chain; and training of risk warning models using ensemble learning.

[0048] (1) Risk data collection of upstream, midstream and downstream enterprises

[0049] The article "The Connotation and Evaluation System of Industrial Chain Security" defines industrial chain security as: Under open economic conditions, a country's ability to control and compete internationally in key links of the industrial chain, effectively ensuring the survival and sustainable development of its key industries. Based on this, a three-level evaluation system for industrial chain security is presented. Combining viewpoints from other literature, this invention defines the risk assessment index for a single enterprise as:

[0050] Innovation capability (I1): personnel composition, R&D investment ratio, number of core patents, etc.;

[0051] Operating performance (I2): Net profit growth, cost profit margin, market share, etc.

[0052] Self-reliance and controllability (I3): capital structure, supplier country composition, foreign core technology patent control rate, etc.;

[0053] Development environment (I4): natural environment (natural disasters, infectious diseases, etc.) and political environment (war), etc.;

[0054] I1, I2, and I3 can be collected from publicly available data or collected periodically and targeted by governments, third-party organizations, or associations. Data collected from the internet must be cleaned and processed before use; I4 is subjectively scored and evaluated by domain experts based on the company's location, and the information on each company is updated regularly.

[0055] Since the above are all quantitative evaluation indicators, corresponding evaluation results can be generated for each enterprise. Therefore, the quantitative evaluation results of the indicator data will serve as the basic feature vector of the enterprise, denoted as:

[0056] F L =(I 1,1 ,I 1,2 ,…,I 1,n1 ,I 2,1 ,I 2,2 ,…,I 2,n2 ,I 3,1 ,I 3,2 ,…,I 3,n3 ,I 4,1 ,I 4,2 ,…,I 4,n4 (1)

[0057] A company's enterprise feature vector is an aggregated feature vector formed by connecting its own basic feature vector with those of its upstream and downstream suppliers / service providers. In other words, the enterprise feature vector is defined as follows:

[0058] F G =cat(F L ,F”) (2)

[0059] The specific acquisition of the enterprise feature vector F” is obtained by the following formula (4). For the last-level supplier or service provider, F” is all 0.

[0060] (2) Aggregation and transmission of risk information along the supply chain and service chain

[0061] Given the complex relationships between upstream, midstream, and downstream enterprises in the industrial chain, and the potential similarities and redundancy in business scope and service targets among multiple enterprises at the same position in the chain, and considering that the operational status of a single enterprise cannot reflect the entire industrial chain—for example, the bankruptcy of one enterprise does not mean that other similar enterprises cannot continue to provide services to the industrial chain—safety early warning for the industrial chain first requires classifying upstream, midstream, and downstream enterprises according to their products, components, and services. The basic steps are as follows:

[0062] 1. For each industry chain requiring risk assessment, suppliers are categorized based on the types of products they provide, and service providers are categorized based on the types of services they offer. Midstream companies are also classified. Taking the automotive industry chain as an example, upstream companies are categorized into different types based on automotive parts suppliers, such as engines, tires, body glass, and ECUs. Midstream companies are categorized into different types, such as passenger vehicles, commercial vehicles, and new energy vehicles. Downstream companies are categorized into different types, such as automotive maintenance, repair, charging, and trading.

[0063] 2. Based on the products and services provided by enterprises and their position in the upstream, midstream, and downstream of the industrial chain, each enterprise in the industrial chain is classified into a corresponding category. An enterprise may appear in different categories.

[0064] 3. For example Figure 2 As shown, the construction of supply and service relationships between enterprises establishes connections between different types of enterprises in the upstream, midstream, and downstream sectors, forming an ecological network, or topological network, of the entire industrial chain. Nodes are different enterprises, and edges correspond to supply or service relationships.

[0065] 4. In the same layer of the topology network, for m enterprises that have a corresponding relationship with the same enterprise in the upper layer and produce the same type of product or provide the same type of service, the same-layer risk aggregation operation is first performed. The calculation method is as follows:

[0066] F' = [max(F1,F2,…,F m ),min(F1,F2,…,F m ),a(F1,F2,…,F m ),…](3)

[0067] Among them, F1, F2 and F m Let F' and F' represent the enterprise feature vectors of each enterprise. These enterprise feature vectors are constructed by connecting the basic feature vectors of the enterprises and the aggregated feature vectors of the next level, and are obtained by formula (2). The enterprise feature vector of the last level is the basic feature vector of the enterprise because its F'' is all 0 during the splicing process. The aggregation method is to calculate the maximum value (max), minimum value (min), and average value (avg) of the enterprise feature vectors of all enterprises in the same layer of the topology network that produce the same type of products or provide the same type of services. The result of the aggregation is a new feature vector F', which is the category feature vector, i.e., the aggregation result of the same layer. max, min, and avg are all aggregate functions that can be used, but they are not limited to the above functions. In practice, the median and other functions can also be used.

[0068] For example, in the automotive supply chain, features of engine manufacturers belonging to the same Tier 1 supplier can be aggregated. If three aggregation functions are used, the length of the aggregated category feature vector F' is three times that of the original feature vector. For ease of processing, all categories should use the same aggregation function for their respective aggregations.

[0069] 5. Aggregate indicators across levels according to product / service relationships, i.e., cross-level aggregation (for example, the indicators of Tier 1 suppliers are aggregated as part of the indicators of midstream car companies, or the indicators of Tier 2 suppliers are aggregated as part of the indicators of Tier 1 suppliers).

[0070] Based on supply or service relationships, aggregation is performed hierarchically from upstream and downstream to midstream. The feature vectors of products / services at the next lower level with corresponding relationships—the aggregated feature vectors—are aggregated into a portion of the feature vector of a specific product / service at the higher level—that is, the enterprise feature vector. For example, in the automotive industry chain, the indicators corresponding to first-tier suppliers such as engines, ECUs, and operating systems are aggregated into a portion of the enterprise feature vector of the midstream automaker's complete vehicle product. The specific calculation method for cross-level aggregation is shown below:

[0071]

[0072] Where F i F' is the category feature vector F' after aggregating the related category enterprises of the i-th next level. That is, the feature vector of the next level product / service category is multiplied element by element to form a new aggregated feature vector F” of the next level product / service. The aggregated feature vector is formed by cross-level aggregation and can form a feature vector that can make up a certain product / service of the next level, which is a part of the enterprise feature vector.

[0073] The aggregated feature vector F and the basic feature vector F of the current parent company L After connecting according to formula (2), the overall feature vector of the current enterprise is formed. For example, in the automotive industry chain, the feature vectors of the engine, ECU and operating system are multiplied together to obtain a unified feature vector F", and the basic feature vector F of the car company itself is F. L Connecting to "F" forms the enterprise feature vector of the car company.

[0074] 6. After traversing the supply chain and service chain, aggregated feature vectors of first-tier suppliers and first-tier service providers are formed for the above chains, namely supply feature vectors and service feature vectors. The basic feature vectors of midstream enterprises are connected to the above vectors to form enterprise feature vectors of midstream enterprises. For midstream enterprises, the enterprise feature vectors aggregated on the supply chain are used as supply aggregation indicators, and the enterprise feature vectors aggregated on the service chain are used as service aggregation indicators. Finally, the enterprise feature vectors on different chains of midstream enterprises are aggregated again according to the formula (3) according to the formula (3) to form category feature vectors of a certain product category, namely supply chain feature vectors and service chain feature vectors, which are used as input feature vectors for risk analysis and early warning. For example, the enterprise feature vectors of all new energy vehicle companies are aggregated according to the calculation method of formula (3) to form the input feature vector of the new energy vehicle industry.

[0075] (3) Training and use of supply chain security early warning model based on ensemble model

[0076] 1. Training Data Generation: For a specific industry chain, historical input feature vectors are generated for a midstream product category (e.g., different types of car manufacturers) at different development stages. These historical input feature vectors are obtained by collecting relevant data and supplementing it through the above calculations. Experts then score and assess the risk level of each sample. Specific scoring and evaluation methods can refer to existing approaches. The risk level of the industry chain is conventionally divided into 5 levels, where 5 represents a high risk level and 1 represents a low risk level.

[0077] 2. A training set is formed using the calculated historical input feature vectors and the risk levels assessed by experts. The model is then trained using an ensemble learning method based on this training set.

[0078] Considering the small size and imbalance of the automotive industry chain dataset available for model training, we propose using ensemble learning techniques to train the model.

[0079] Generally, ensemble learning can be divided into two steps: first, constructing a series of weak classifiers; and second, merging them to form a stronger classifier with better performance. Traditional ensemble learning methods include Bagging, Boosting, and Stacking. Boosting trains multiple weak models sequentially. After each training iteration, some samples may make errors. In the next training iteration, the weights of these errors are increased, causing the model to focus more on them. This process is repeated until the number of weak models reaches a predetermined number n. In practice, the results of these n weak models are weighted and summed to obtain the final result; generally, later models have higher weights. Boosting effectively reduces model bias. Therefore, this invention chooses to use the Boosting method for model training.

[0080] Model usage: For a new time point, the input feature vector of the industrial chain can be constructed based on the process described above. Then, the risk level can be predicted by using the trained model, i.e., the evaluation model. Based on the prediction results, a certain risk response plan can be formulated using existing methods.

[0081] The following example illustrates the workflow of this supply chain security early warning system:

[0082] 1. Indicator Generation

[0083] (1) First, collect the basic characteristic vector F of all enterprises in China's automotive industry chain from 2011 to 2021. L (Refer to formula (1)).

[0084] (2) At the same layer in the network, m enterprises that produce the same type of products or provide the same type of services first perform risk aggregation on the corresponding enterprise feature vectors to generate F'. The last-level supplier or service provider uses the basic feature vector as the enterprise feature vector (e.g., the ECU indicator F' here). ECU It is by F ECU1 F ECU2 F ECU3 The aggregated product is obtained by referring to formula (3).

[0085] (3) In the supply relationship or service relationship, the category feature vectors of each category of products / services at the next level in step (2) are aggregated across levels into a part of the feature vector of a certain product / service at the next higher level, generating an aggregated feature vector F (refer to formula (4)).

[0086] (4) The basic feature vector F of a certain product / service at the next higher level L Connect the F in step (3) above (e.g., F). 车企ABy F' 发动机 , F' ECU and F' 操作系统 After aggregation according to formula (4), the enterprise feature vector F corresponding to the product or service is formed. G (Refer to formula (2)).

[0087] (5) Repeat the process of steps (2)-(4) on the enterprise feature vector generated in step (4) in order until all suppliers or service providers are traversed, and finally generate the aggregated feature vector of the first-level supplier or first-level service provider, namely the supply feature vector and the service feature vector.

[0088] (6) The supply feature vector and service feature vector are concatenated with the basic feature vector of midstream enterprises to generate enterprise feature vectors in the supply chain and service chain of midstream enterprises.

[0089] (7) Determine the characteristic vector F of enterprises in the supply chain and service chain of midstream enterprises according to the category of midstream enterprises. G Aggregate according to formula (3) to form the feature vector of a certain product category (for example, aggregate the feature vectors of all new energy vehicle companies to form the feature vector of the new energy vehicle industry).

[0090] In light of the above, combined with Figure 2 To elaborate: The topology network includes second-tier suppliers Sensor 1, Sensor 2, Sensor 3, Processor 1, and Processor 2; first-tier suppliers Engine 1, Engine 2, ECU 1, ECU 2, ECU 3, Operating System 1, and Operating System 2; midstream companies include car manufacturers A and B; and first-tier service providers include car charging service, repair service 1, repair service 2, repair service 3, car transaction service 1, and car transaction service 2.

[0091] In the above-mentioned industrial chain, the second-level suppliers are the lowest-level suppliers. The basic feature vectors corresponding to their sensor and processor suppliers are obtained by formula (1). The basic feature vectors of the sensor and processor suppliers are aggregated at the same level by formula (3) to generate the category feature vectors of the second-level suppliers. The aggregated feature vectors of the sensor and processor suppliers at the same level are aggregated across levels by formula (4) to generate the aggregated feature vectors of the second-level suppliers. For example, for the correspondence of ECU1, the sensor categories of sensor 1 and sensor 2 are aggregated at the same level and across levels to generate the cross-level aggregation result corresponding to ECU1.

[0092] After aggregation, the cross-level aggregation results of different ECU suppliers are concatenated with the basic feature vector of the ECU supplier corresponding to the first-level supplier using formula (2) to generate the enterprise feature vector of the ECU enterprise. The enterprise vectors of the engine and operating system are also obtained based on the feature vectors of the second-level suppliers through the above process.

[0093] After obtaining the enterprise feature vectors of engine, ECU and operating system suppliers, the enterprise feature vectors of the above-mentioned first-level suppliers are calculated by formula (3) and formula (4) respectively to generate the aggregated feature vector of the first-level suppliers, i.e. the supply feature vector. The supply feature vector is then spliced ​​with the basic feature vectors of different car companies by formula (2) to generate the enterprise feature vector of the car company's supply chain, i.e. the supply aggregation index.

[0094] For service providers, after obtaining the basic feature vector through formula (1), the same-layer aggregation and cross-layer aggregation are performed through formula (3) and formula (4). After aggregation through formula (2), the basic feature vector of the car company is correspondingly spliced ​​to generate the enterprise feature vector on the service chain of the car company, i.e., the service aggregation index.

[0095] Formula (3) is used to aggregate the supply aggregation index and the service aggregation index at the same level, generating the aggregated feature vector on the automotive enterprise supply chain and the aggregated feature vector on the service chain, namely the supply chain feature vector and the service chain feature vector, as input feature vectors.

[0096] 2. Data labeling and model training

[0097] (1) The feature vector generated in step 1 (5) is used as a training sample. Experts need to score each sample to determine its risk level, which is divided into 5 levels, where 5 indicates a relatively high risk level and 1 indicates a relatively low risk level.

[0098] (2) Train the model using the ensemble learning method described above. Train multiple weak learners, and the final output of the model is the weighted sum of the multiple weak learners.

[0099] 3. Model inference and report generation

[0100] For example, to predict the risk level of the new energy vehicle industry chain in 2022, we can collect industry chain data at a new time point (e.g., 2022) and generate a feature vector through step 1. This feature vector is then fed into the previously trained model to generate the predicted risk level and a risk assessment report for the industry chain.

[0101] To better achieve the above-mentioned technical objectives, the present invention also provides a supply chain security early warning system.

[0102] A supply chain security early warning system comprises modules for data collection, indicator calculation, model training, model inference, and report generation. The data collection module includes both external and internal data collection. Indicator calculation includes objective risk indicator calculation, subjective risk indicator assessment, and risk indicator aggregation (feature vector calculation). Model training includes both manual risk level assessment and ensemble learning model training.

[0103] Data collection module, indicator calculation module, model training module, model inference module, and report generation module.

[0104] The data collection module is used to collect indicator data;

[0105] The indicator calculation module is used to obtain the basic feature vectors of enterprises in the industry chain based on indicator data, construct the topology network corresponding to the industry chain, calculate the basic feature vectors based on the topology network to obtain the aggregated feature vector, connect the aggregated feature vector and the basic feature vector to obtain the overall feature vector, and aggregate the overall feature vector again to obtain the input feature vector.

[0106] The model training module is used to train machine learning models using ensemble learning methods.

[0107] The model inference module is used to predict the input feature vector using a trained machine learning model, and obtain the safety early warning prediction result of the industrial chain risk level.

[0108] The report generation module is used to generate risk reports based on the safety early warning and prediction results of the industrial chain risk level.

[0109] The data collection module includes an external data collection module and an internal data collection module;

[0110] The external data collection module is used to collect and acquire data on innovation capabilities, operational status, and independent controllability.

[0111] The internal data collection module is used to collect and acquire development environment data.

[0112] The indicator calculation module consists of three parts: an objective risk indicator calculation module, a subjective risk indicator assessment module, and a risk indicator aggregation module.

[0113] The objective risk indicator calculation module is used to clean and process innovation capability data, operational status data, and self-controllable data, and to quantitatively score and evaluate the cleaned and processed data to obtain the first basic feature vector.

[0114] The subjective risk indicator assessment module is used to assess and score development environment data to obtain the second basic feature vector;

[0115] Risk indicator aggregation is used to aggregate the first basic feature vector and the second basic feature vector to generate a basic feature vector; and to obtain the input feature vector based on the basic feature vector.

[0116] The model training process includes two parts: manual risk level assessment and ensemble learning model training.

[0117] The risk level manual assessment module is used to obtain historical input feature vectors, label the historical feature vectors with risk levels, and generate a training set based on the historical feature vectors and risk level labeling results.

[0118] The ensemble learning model training module is used to train a machine learning model using ensemble learning methods based on the training set, resulting in a trained machine learning model.

[0119] The relationships between modules and submodules are as follows: Figure 3 As shown, this content corresponds to the above-mentioned method and technical solution, and will not be elaborated further here.

[0120] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An industry chain security early warning method based on ensemble learning, characterized in that, The application relates to an industrial chain risk early warning method and device. The method comprises the following steps: obtaining the basic characteristic vectors of all enterprises in the industrial chain and constructing a topological network corresponding to the industrial chain, wherein the topological network comprises a plurality of midstream enterprises and corresponding suppliers and service providers of a plurality of levels; for the suppliers and service providers, based on the corresponding relationship between the suppliers or service providers, the basic characteristic vectors of the last level enterprises are sequentially subjected to same-level aggregation and cross-level aggregation, the cross-level aggregation result is spliced with the basic characteristic vectors of the corresponding upper level enterprises to obtain the enterprise characteristic vectors of the upper level enterprises, and the enterprise characteristic vectors are subjected to same-level aggregation and cross-level aggregation again, and the process of same-level aggregation, cross-level aggregation and splicing is repeated until all the suppliers and service providers are traversed to obtain the supply characteristic vectors and service characteristic vectors; for the midstream enterprises, the supply characteristic vectors and service characteristic vectors are spliced with the basic characteristic vectors of the midstream enterprises to obtain the supply aggregation indexes and service aggregation indexes of the midstream enterprises, and the supply chain characteristic vectors and service chain characteristic vectors are obtained through same-level aggregation based on the supply aggregation indexes and service aggregation indexes; an evaluation model is obtained by training a machine learning model through an ensemble learning method, and the supply chain characteristic vectors and service chain characteristic vectors are predicted through the evaluation model to obtain an industrial chain risk grade safety early warning prediction result. The construction process of the topological network corresponding to the industrial chain comprises the following steps: all enterprises in the industrial chain and the corresponding correlation between the enterprises are counted, the enterprises are divided into levels and categories, and based on the division results of the levels and categories, the midstream enterprises are determined, the enterprises are taken as network nodes, the corresponding correlation between the enterprises is taken as edges, and based on the network nodes and edges, the topological network corresponding to the industrial chain is obtained; the process of same-level aggregation and cross-level aggregation comprises the following steps: the enterprise characteristic vectors of the enterprises of the same level and the same category are subjected to same-level aggregation to obtain a same-level aggregation result, namely a category characteristic vector F': F' = [max (F1, F2, …, Fm), min (F1, F2, …, Fm), a𝑣𝑔 (F1, F2, …, Fm), …] wherein Fm represents the enterprise characteristic vector of the mth enterprise in the same level and the same category, the category characteristic vector is subjected to multiplication calculation to obtain a cross-level aggregation result; in the same-level aggregation process, for the last level enterprises, the basic characteristic vectors of the last level enterprises are taken as the enterprise characteristic vectors for same-level aggregation, and for the midstream enterprises, the supply aggregation indexes and service aggregation indexes are taken as the enterprise characteristic vectors of the midstream enterprises for same-level aggregation; the evaluation model acquisition process comprises the following steps: collecting historical data, converting the historical data into historical input characteristic vectors, labeling the historical input characteristic vectors, integrating the input overall characteristic vectors and the labeling results into a training set; a plurality of weak classifiers are constructed, the plurality of weak classifiers are sequentially trained based on the training set through a Boosting method until the trained weak classifiers reach a preset number, and the outputs of the trained weak classifiers are weighted and summed to obtain the evaluation model.

2. The industry chain security early warning method based on ensemble learning according to claim 1, characterized in that: the process of obtaining the basic feature vector comprises: obtaining index data of enterprises in the industry chain, wherein the index data comprises innovation capability data, business status data, self-controllable data and development environment data; cleaning and processing the index data, and quantifying the data after cleaning and processing to obtain the basic feature vector. 3.The industry chain security early warning method based on ensemble learning system according to any one of claims 1-2, characterized in that, comprises: a data collection module, an index calculation module, a model training module, a model inference module and a report generation module; wherein the data collection module is used to collect index data; the index calculation module is used to obtain the basic feature vector of the enterprises in the industry chain according to the index data, and construct a corresponding topological network of the industry chain, wherein the topological network comprises a plurality of midstream enterprises and a plurality of corresponding levels of suppliers and service providers; for the suppliers and service providers, the basic feature vector of the last level enterprise is sequentially aggregated within the same layer and across layers, the cross-layer aggregation result is spliced with the basic feature vector of the corresponding upper level enterprise to obtain the enterprise feature vector of the upper level enterprise, and the enterprise feature vector is aggregated within the same layer and across layers again, and the process of repeating the same layer aggregation, cross-layer aggregation and splicing is performed until all the suppliers and service providers are traversed to obtain the supply feature vector and the service feature vector; for the midstream enterprises, the supply feature vector and the service feature vector are spliced with the basic feature vector of the midstream enterprise respectively to obtain the supply aggregation index and the service aggregation index of the midstream enterprise, and the supply chain feature vector and the service chain feature vector are obtained by aggregating within the same layer and across layers based on the supply aggregation index and the service aggregation index; the model training module is used to train a machine learning model by an ensemble learning method to obtain an evaluation model, the model inference module is used to predict the supply chain feature vector and the service chain feature vector by the evaluation model to obtain an industry chain risk level security early warning prediction result; and the report generation module is used to generate a risk report according to the industry chain risk level security early warning prediction result.

4. The industry chain security early warning system based on ensemble learning according to claim 3, characterized in that: the index data in the data collection module comprises innovation capability data, business status data, self-controllable data and development environment data.

5. The industry chain security early warning system based on ensemble learning according to claim 3, characterized in that: the process of constructing the topological network in the index calculation module comprises: counting the enterprises in the industry chain, the enterprise categories and the corresponding correlation between the enterprises, taking the enterprises as network nodes and the corresponding correlation between the enterprises as edges, and obtaining the topological network corresponding to the industry chain based on the network nodes and the edges.

6. The industry chain security early warning system based on ensemble learning according to claim 3, characterized in that: the process of aggregating within the same layer and across layers in the index calculation module comprises: aggregating the enterprise feature vectors of the enterprises of the same category within the same layer to obtain the same layer aggregation result, i.e. the category feature vector F’: F’=[max(F1, F2, …, F m ), min(F1, F2, …, F m ), a𝑣𝑔(F1, F2, …, F m ), …] wherein the F m represents the enterprise feature vector of the mth enterprise in the same category of the same level, multiplying the category feature vectors to obtain the cross-layer aggregation result; In the same layer aggregation process, the basic feature vector of the last stage enterprise is taken as the enterprise feature vector for the last stage enterprise, and the supply aggregation index and the service aggregation index are taken as the enterprise feature vector of the midstream enterprise for the midstream enterprise.

7. The industry chain security early warning system based on ensemble learning according to claim 3, characterized in that: The evaluation obtaining process in the model training module: Collect historical data, convert the historical data into historical input feature vectors, label the historical input feature vectors, and integrate the input overall feature vectors and the labeling results into a training set; A plurality of weak classifiers are constructed, and the plurality of weak classifiers are sequentially trained based on the training set by a Boosting method until the trained weak classifiers reach a preset number, and the outputs of the trained weak classifiers are weighted and summed to obtain an evaluation model.

Images