Artificial intelligence-based automobile parts enterprise supply chain risk early warning method

Abstract

The application belongs to the technical field of automobile parts, and discloses a risk early warning method for automobile parts enterprise supply chain based on artificial intelligence; the method comprises the following steps: collecting and processing supply chain data to construct a graph; evaluating suppliers based on the graph to generate a portrait matrix; based on the portrait matrix, combining production parameters, performing model training, constructing a prediction engine, and generating a parts quality risk prediction result; performing abnormal detection on nodes to form a monitoring network and output a risk assessment result; based on the graph, a supply chain network topology model is constructed, and a risk propagation path analysis is performed to generate a risk transmission graph; a risk assessment model is established, the risk prediction result, the risk assessment result and the risk transmission graph are integrated, the risks of each link of the supply chain are integratedly evaluated, and a scoring system is formed; based on the scoring system, a dynamic risk early warning threshold is generated, a risk response decision tree is constructed, intelligent risk response suggestions are provided, and the efficiency of enterprise risk disposal is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automobile parts, more specifically, the present application relates to a risk early warning method for automobile parts enterprise supply chain based on artificial intelligence. BACKGROUND

[0002] With the rapid development of the automobile industry and the acceleration of the globalization process, the automobile parts supply chain presents the characteristics of large scale, complex structure and extensive cross-domain. According to statistics, an ordinary passenger car contains about 30,000 parts, involving the coordination of hundreds of suppliers. Under this background, supply chain risk management has become a core challenge for the operation of automobile parts enterprises.

[0003] Due to the large and complex automobile parts supply chain system, involving numerous suppliers, logistics nodes and production links, leading to scattered data sources, non-uniform formats, each supplier generates a large amount of heterogeneous data from ERP systems, quality detection systems, logistics tracking systems, etc. every day. These data are scattered in different systems, with different formats and standards, making it impossible to conduct unified analysis and risk monitoring, and making it difficult to effectively integrate and analyze. The traditional risk assessment method relies too much on manual experience, and cannot realize the scientific quantification and dynamic assessment of multi-dimensional risks such as supplier qualification, production capacity and financial status, and the assessment results are often lagging behind the actual risks. When a supplier has a risk event, such as a sudden financial crisis or a sudden drop in production capacity, the traditional method cannot quickly assess the risk transmission range, resulting in the influence of multiple downstream supporting enterprises. In actual production and operation, due to the lack of prediction ability of parts quality fluctuations, quality problems often cannot be discovered and handled in time. The risk monitoring of the logistics link is also weak, and it is difficult to track and warn abnormal conditions in real time. At the same time, the supply chain network structure is complex, and there are complex relationships between nodes, making it difficult for traditional methods to effectively analyze the transmission path and diffusion effect of risks in the supply chain network, resulting in risks often causing chain reactions when discovered. In addition, the existing early warning mechanism generally has the problems of single early warning index, lack of scientific basis for threshold setting, and inaccurate early warning results, and after the risk occurs, it lacks systematic response strategy support, and cannot quickly develop effective risk response plans, causing the enterprise to be passive in response to risk events, affecting the stability of production and operation. These problems may lead to a series of serious consequences such as supply interruption, quality accidents, delivery delay, etc. in actual production, causing significant economic losses to the enterprise.

[0004] In view of this, the present application proposes a risk early warning method for automobile parts enterprise supply chain based on artificial intelligence to solve the above problems. SUMMARY

[0005] To overcome the aforementioned shortcomings of existing technologies and achieve the above objectives, this invention provides the following technical solution: an artificial intelligence-based supply chain risk early warning method for automotive parts enterprises, comprising:

[0006] Step 1: Collect and integrate multi-source heterogeneous data from the supply chain to construct a supply chain knowledge graph;

[0007] Step 2: Based on the supply chain knowledge graph, extract and quantify the multi-dimensional risk characteristics of suppliers to generate a supplier risk profile matrix;

[0008] Step 3: Based on the supplier risk profile matrix and combined with the component production process parameters, train the quality fluctuation prediction model, build a component quality risk prediction engine, and generate component quality risk prediction results.

[0009] Step 4: Conduct spatiotemporal characteristic analysis and anomaly detection on supply chain logistics nodes to form a logistics risk monitoring network and generate logistics node risk assessment results;

[0010] Step 5: Based on the supply chain knowledge graph, construct a supply chain network topology model, perform risk propagation path analysis, and generate a risk transmission diagram;

[0011] Step 6: Establish a multi-agent collaborative risk assessment model, integrate the component quality risk prediction results, logistics node risk assessment results and risk transmission diagram, and conduct integrated assessment of the risks in each link of the supply chain to form a comprehensive risk scoring system;

[0012] Step 7: Based on the comprehensive risk scoring system and combined with multivariate time series analysis, generate dynamic risk warning thresholds and obtain warning results; based on the warning results and the preset historical response strategy library, construct a risk response decision tree and provide intelligent risk response suggestions.

[0013] Furthermore, the process of collecting and fusing multi-source heterogeneous data from the supply chain to construct a supply chain knowledge graph includes:

[0014] Set up a multi-source data acquisition interface to obtain raw data from the enterprise resource planning system, supplier management system, quality management system and external databases to form an initial dataset;

[0015] The initial dataset is cleaned and standardized to obtain a normalized dataset;

[0016] Entity recognition and relation extraction are performed on the normalized dataset to construct an initial semantic network;

[0017] Based on the initial semantic network, an ontology model is constructed to determine entity categories, attributes, and relationship types, forming a supply chain ontology framework;

[0018] The normalized dataset is mapped to the supply chain ontology framework to generate an instantiated knowledge base;

[0019] The instantiated knowledge base is subjected to knowledge reasoning and consistency checks to obtain an optimized supply chain knowledge graph.

[0020] Furthermore, based on the supply chain knowledge graph, the process of extracting and quantifying multi-dimensional risk characteristics of suppliers to generate a supplier risk profile matrix includes:

[0021] The supplier basic information, historical delivery records, quality performance, financial status and geopolitical factors are extracted from the supply chain knowledge graph to form a supplier feature set;

[0022] Principal component analysis was performed on the supplier feature set to determine the key risk dimensions and obtain a dimensionality-reduced feature space.

[0023] Based on the reduced feature space, a supplier risk assessment index system is constructed, and the weight of each index is determined by the analytic hierarchy process, thus forming a weighted index system.

[0024] Suppliers are quantitatively scored across various risk dimensions to obtain a multidimensional risk score vector;

[0025] Cluster analysis is performed on the multidimensional risk scoring vector to identify supplier risk patterns, form risk category divisions, and obtain risk category information; the supplier's multidimensional risk scoring vector is combined with the risk category information to generate a supplier risk profile matrix.

[0026] Furthermore, based on the supplier risk profile matrix and combined with component production process parameters, a quality fluctuation prediction model is trained to construct a component quality risk prediction engine and generate component quality risk prediction results, including:

[0027] Collect process parameters, equipment status, environmental conditions, and testing data during the parts production process to form a production process feature set;

[0028] The production process feature set and the supplier risk profile matrix are fused to construct a quality prediction training dataset;

[0029] Based on the aforementioned quality prediction training dataset, a basic model for quality fluctuation prediction is trained using deep learning methods; transfer learning is then performed on the basic model for quality fluctuation prediction to obtain a group of multimodal prediction models.

[0030] Based on the multimodal prediction model group, an integrated learning framework is constructed, and then a sustainable self-optimizing component quality risk prediction engine is built. The component quality risk prediction engine is used to evaluate the current and future component quality status and generate component quality risk prediction results, including the probability of quality fluctuation, potential defect type, quality risk level, and risk occurrence time prediction.

[0031] Furthermore, the spatiotemporal characteristic analysis and anomaly detection of supply chain logistics nodes form a logistics risk monitoring network, producing logistics node risk assessment results, including:

[0032] The location information, transportation time, transit status and environmental conditions of logistics nodes are collected to form a logistics spatiotemporal dataset; the time series features of the logistics spatiotemporal dataset are extracted to identify the seasonal patterns and periodic changes in logistics operations and obtain a time series feature vector.

[0033] Based on the aforementioned time-series feature vectors, a time-series prediction model for logistics nodes is constructed to generate a baseline for normal logistics behavior. A multivariate anomaly detection algorithm is defined to compare the deviation between the current logistics status and the baseline for normal logistics behavior in real time, identifying potential anomalies. The identified potential anomalies are classified and their risk levels are assessed to form a logistics anomaly risk database. Based on the logistics anomaly risk database, a risk network model is performed on logistics nodes to construct a logistics risk monitoring network. According to the logistics risk monitoring network, a comprehensive risk index for each node is calculated. The exponential smoothing method is used to perform time-series prediction on the comprehensive risk index of each node to obtain a predicted risk index. Based on the predicted risk index, a set of key risk nodes is identified. The set of key risk nodes and their corresponding predicted risk indices constitute the logistics node risk assessment result.

[0034] Furthermore, the step of constructing a supply chain network topology model based on the supply chain knowledge graph, and performing risk propagation path analysis to generate a risk transmission diagram includes:

[0035] Extract the relationships between suppliers, manufacturers, and logistics service providers from the supply chain knowledge graph to construct an initial network graph;

[0036] The process involves calculating the centrality index and critical path of network nodes to identify key nodes in the supply chain; constructing a network vulnerability assessment model based on these key nodes to quantitatively analyze the stability of the network structure; obtaining the network vulnerability assessment results; defining a risk propagation dynamics model to simulate the diffusion process of risk events in the supply chain network and identify risk cascading effects; performing probability analysis on the propagation paths of different risk events to obtain a risk propagation probability matrix; and constructing a dynamic risk transmission diagram based on the risk propagation probability matrix and the network vulnerability assessment results.

[0037] Furthermore, the establishment of a multi-agent collaborative risk assessment model integrates the component quality risk prediction results, logistics node risk assessment results, and risk transmission diagram to conduct an integrated assessment of risks at each stage of the supply chain, forming a comprehensive risk scoring system, including:

[0038] Define a supplier risk assessment agent, a quality risk assessment agent, a logistics risk assessment agent, and a market risk assessment agent, each responsible for risk assessment in different dimensions;

[0039] Configure expert knowledge bases and machine learning models for each intelligent agent; input the component quality risk prediction results into the quality risk assessment intelligent agent, input the logistics node risk assessment results into the logistics risk assessment intelligent agent, and input the risk transmission diagram into each intelligent agent;

[0040] Construct communication protocols and collaborative decision-making mechanisms among intelligent agents to enable different agents to exchange risk information and coordinate assessment results, forming a self-evolving comprehensive risk scoring system.

[0041] Furthermore, based on the comprehensive risk scoring system and combined with multivariate time series analysis, a dynamic risk early warning threshold is generated, thereby obtaining the early warning result; based on the early warning result and a preset historical response strategy library, a risk response decision tree is constructed to provide intelligent risk response suggestions, including:

[0042] Collect historical risk event data and corresponding risk indicator time series to construct a risk-indicator association dataset; perform time series pattern mining on the risk-indicator association dataset to identify risk warning features and form an early warning indicator library.

[0043] Based on the aforementioned early warning indicator library, a multi-level early warning threshold calculation model is defined to generate differentiated thresholds for different types of components and suppliers.

[0044] A Bayesian network early warning model is constructed, and combined with the comprehensive risk scoring system, the risk probability and early warning level are obtained, i.e., the early warning result.

[0045] A supply chain risk classification system is established, and historical risk events and their corresponding countermeasures are stored in a structured manner to form a historical response strategy library. Based on the historical response strategy library, a case-based reasoning method is used to extract the association rules between risk scenarios and response measures. Using the early warning results and combining critical path analysis of the risk transmission diagram, the current risk situation feature vector is identified. Based on the risk situation feature vector, a risk response decision tree is constructed using the C4.5 algorithm, mapping risk attributes to a set of response strategies. The risk transmission diagram is analyzed to identify the risk propagation path and impact nodes, and the response strategy set in the risk response decision tree is prioritized. A final risk response plan is formed. The risk response plan is transformed into specific execution tasks, generating intelligent risk response suggestions that include action steps, responsible persons, and time nodes.

[0046] The AI-based automotive parts enterprise supply chain risk early warning device includes: a memory and at least one processor, wherein the memory stores instructions.

[0047] The at least one processor invokes the instructions in the memory to cause the AI-based automotive parts enterprise supply chain risk warning device to execute the AI-based automotive parts enterprise supply chain risk warning method as described.

[0048] A computer-readable storage medium storing instructions that, when executed by a processor, implement the AI-based supply chain risk warning method for automotive parts companies as described above.

[0049] The technical effects and advantages of the artificial intelligence-based supply chain risk early warning method for automotive parts enterprises in this invention are as follows:

[0050] This application comprehensively enhances the accuracy, foresight, and decision-making efficiency of supply chain risk management for automotive parts companies. By deeply integrating multi-dimensional data sources and dynamic knowledge graphs, it breaks through traditional information silos, achieving a penetrating insight into risks across the entire chain, from supplier qualifications to production quality fluctuations and logistics anomalies. It accurately identifies potential risk transmission paths and cascading effects, shortening risk warning response time. The constructed multimodal risk assessment system, combined with real-time context adaptive adjustment of warning thresholds, effectively solves the long-standing industry problem of false alarms and missed alarms, improves the accuracy of risk identification at key nodes, and significantly reduces production losses caused by supply chain disruptions. A decision support mechanism based on intelligent inference deeply couples historical experience with real-time situations, automatically generating response strategy combinations matched to the company's resource endowment, driving risk management from passive emergency response to proactive prevention and control, and assisting companies in quickly building resilient supply chain networks in complex market environments. Furthermore, through risk transmission visualization and attribution analysis, it empowers management to accurately locate the root causes of risks, optimize supplier strategic layout and inventory strategies, and systematically reduce supply chain vulnerability. This not only improves the efficiency of enterprise risk handling but also reduces the hidden costs caused by quality defects and logistics delays through preventative control. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the AI-based supply chain risk warning method for automotive parts companies according to the present invention. Detailed Implementation

[0052] This application provides an artificial intelligence-based supply chain risk early warning method for automotive parts companies. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0053] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 In one specific embodiment of the AI-based supply chain risk warning method for automotive parts companies, step 1, "collecting and fusing multi-source heterogeneous data from the supply chain to construct a supply chain knowledge graph," specifically includes the following steps:

[0054] Set up a multi-source data acquisition interface to obtain raw data from the enterprise resource planning system, supplier management system, quality management system and external databases to form an initial dataset;

[0055] The initial dataset is cleaned and standardized to eliminate redundancy and outliers, resulting in a normalized dataset.

[0056] Entity recognition and relation extraction are performed on the normalized dataset to construct an initial semantic network;

[0057] Based on the initial semantic network, an ontology model is constructed to determine entity categories, attributes, and relationship types, forming a supply chain ontology framework.

[0058] The normalized dataset is mapped to the supply chain ontology framework to generate an instantiated knowledge base;

[0059] Knowledge reasoning and consistency checks are performed on the instantiated knowledge base to obtain an optimized supply chain knowledge graph.

[0060] Specifically, the multi-source data acquisition interface design adopts a distributed acquisition architecture, acquiring data from different systems through REST APIs, ODBC connections, and ETL tools. Purchase orders, inventory levels, and production plans are extracted from the Enterprise Resource Planning (ERP) system; basic supplier information, qualification certificates, and evaluation records are obtained from the Supplier Management System; quality inspection reports, non-conforming product records, and quality improvement measures are extracted from the Quality Management System; and market dynamics, macroeconomic indicators, and geopolitical risks are obtained from external databases. All of this data collectively constitutes the initial dataset, providing raw materials for subsequent knowledge graph construction.

[0061] The data cleaning process employs a three-stage approach. First, missing value handling is performed. For data with a missing rate below 5%, mean / mode imputation is used. For data with a missing rate between 5% and 20%, KNN imputation based on similar records is used. Features with a missing rate exceeding 20% ​​are evaluated for importance before retention. Second, outlier detection is performed. Numerical anomalies are identified using the Z-score method, and categorical anomalies are identified using frequency analysis. Finally, data standardization is performed to obtain a normalized dataset; for numerical data, the Min-Max standardization method is used.

[0062] ;in, This represents the standardized value. Represents the original value. and These represent the minimum and maximum values ​​of the feature, respectively.

[0063] A supply chain entity dictionary is constructed, containing specialized terms categorized as suppliers, components, production equipment, and quality indicators. Then, a BiLSTM-CRF model is used for text entity recognition. The model input is a text sequence T = {t1, t2, ..., tn}, and the output is an entity label sequence Y = {y1, y2, ..., yn}.

[0064] The BiLSTM-CRF model consists of BiLSTM layers and CRF layers; the BiLSTM layers encode contextual information to obtain a feature vector sequence H={h1, h2, ..., hn}, and the CRF layers calculate the probability of the optimal label sequence.

[0065] ;in, Represents the label transition matrix (from) Transferred to (probability) Indicates the probability of launch. As the normalization factor, This refers to the indices within the text sequence, the corresponding entity label sequence, and the feature vector sequence; and They belong to the entity label sequence and the feature vector sequence, respectively; The corresponding sequence length;

[0066] Define a set of relationship types in the supply chain domain, R = {supply relationship, cooperation relationship, competition relationship, geographical location relationship, ...}. Then, construct a remote supervised training set based on seed relationship instances and use a graph convolutional network model to learn the relationships between entities. For an entity pair (e1, e2), construct a feature vector using its context representation c and entity representations e1 and e2. Calculate the probability of relation type:

[0067] ;in, This is the weight matrix. This is the bias vector. A relation threshold is set. Determine the relationships between entities:

[0068] ;

[0069] like The corresponding entities are unrelated; thus, the initial semantic network is constructed.

[0070] The ontology model is constructed using a combination of bottom-up and top-down approaches. Based on identified entities and relationships—the initial semantic network—a preliminary class hierarchy is built. Then, domain expert knowledge is incorporated to define core concepts, their attributes, and relationships. Finally, the ontology model is described using the formal language OWL (Web Ontology Language). The ontology framework defines four basic components: (classes), (object attributes), (data attributes), and (individuals), forming a conceptual system for the supply chain domain—the supply chain ontology framework.

[0071] Define a mapping rule set R = {r1, r2, ..., rn}, where each rule defines the transformation relationship from the source data schema to the target ontology. Then, using entity linking technology, the identified entities... With classes in the ontology Perform matching and calculate the matching degree:

[0072] ;in, , , The weighting coefficients and , For string similarity, For context similarity, For attribute similarity, and Entities With class The string, and For entities With class The context indicates that and For entities With class Attributes.

[0073] Preset similarity threshold ,when At that time, the entity Mapping to class This leads to the instantiation of the knowledge base.

[0074] Knowledge reasoning employs a descriptive logic-based reasoning mechanism and a rule-based reasoning method. It utilizes an ontology-based reasoning rule set for subclass reasoning, attribute inheritance, and constraint verification, and a business rule set for deriving domain-specific knowledge.

[0075] It should be explained that the ontology reasoning rule set is a set of logical reasoning rules based on knowledge graphs, used to verify the consistency of knowledge and deduce implicit knowledge. For example: subclass reasoning: if A is a subclass of B, and B has attribute P, then A inherits from P.

[0076] (For example, if "Brake Supplier" is a subclass of "Tier 1 Supplier", and "Tier 1 Supplier" is required to be ISO certified, then "Brake Supplier" will automatically inherit this requirement.)

[0077] Business rule sets are industry-specific rules of experience used to supplement logical reasoning. For example: "If a supplier's delivery delay rate is >15% in the past 3 months, it is marked as high risk" and "Tier 2 suppliers of key components must pass on-site audits".

[0078] Consistency testing mainly consists of two parts: logical consistency testing and business rule consistency testing. Logical consistency testing determines whether there are contradictions in the ontology by describing the logical inference engine; business rule consistency testing verifies whether the knowledge conforms to the business logic through domain constraints, thereby obtaining the optimized supply chain knowledge graph.

[0079] In one specific embodiment, the process of performing step 2, "based on the supply chain knowledge graph, extracting and quantifying the multi-dimensional risk characteristics of suppliers to generate a supplier risk profile matrix," specifically includes the following steps:

[0080] Extract basic supplier information, historical delivery records, quality performance, financial status, and geopolitical factors from the supply chain knowledge graph to form a supplier feature set;

[0081] Principal component analysis was performed on the supplier feature set to identify key risk dimensions and obtain a dimensionality-reduced feature space. Based on the dimensionality-reduced feature space, a supplier risk assessment index system was constructed, and the weight of each index was determined using the analytic hierarchy process to form a weighted index system.

[0082] The fuzzy comprehensive evaluation method is used to quantitatively score suppliers on each risk dimension, resulting in a multidimensional risk score vector. Cluster analysis is then performed on the multidimensional risk score vector to identify supplier risk patterns and form a risk category classification.

[0083] By combining the supplier's multidimensional risk scoring vector with risk category information, a supplier risk profile matrix is ​​generated.

[0084] Specifically, when extracting supplier features from the supply chain knowledge graph, a multi-path query and attribute aggregation method is employed. For supplier entity S, a multi-level query is constructed using the graph query language SPARQL to obtain direct attributes and related entity information. Basic information features include supplier size, establishment time, main business, and technical capability rating; historical delivery record features include historical on-time delivery rate, average delay time, and order response speed; quality performance features include product qualification rate, quality problem resolution cycle, and quality system certification level; financial status features include debt-to-equity ratio, current ratio, and accounts receivable turnover; geopolitical factors include the political stability of the supplier's location, natural disaster risks, and the impact of labor policies. A supplier feature set is constructed through graph queries. ,in to This represents the values ​​of the first feature to the Nth feature.

[0085] Principal component analysis is performed on the supplier feature set to reduce feature dimensionality and identify key risk dimensions. The original feature space of the supplier feature set is N-dimensional. First, the feature covariance matrix C is calculated.

[0086] ;in, For the number of suppliers, The feature mean vector, For the first in the supplier feature set The value of each feature, This is to transpose the content within the parentheses.

[0087] Then, the eigenvalues ​​and corresponding eigenvectors of the covariance matrix C are calculated. The first k eigenvectors are selected to form a projection matrix, where k satisfies the requirement of retaining at least 85% of the information. The first k eigenvectors correspond to the key risk dimensions. The supplier feature set is projected onto a low-dimensional space to obtain the dimensionality-reduced eigenvector F'. That is, in the dimensionality-reduced feature space.

[0088] Based on a reduced-dimensional feature space, a hierarchical supplier risk assessment index system is constructed. The system consists of two levels: primary indicators and secondary indicators. Primary indicators include five dimensions: operational risk, financial risk, quality risk, supply risk, and geopolitical risk. Each primary indicator has 3-5 secondary indicators. The Analytic Hierarchy Process (AHP) is used to determine the weights of the primary and secondary indicators. First, a judgment matrix A is constructed, where each element a_ii' represents the importance of indicator i relative to indicator i'. Then, the maximum eigenvalue λmax and the corresponding eigenvector w of A are calculated. After normalization, the indicator weights are obtained, and these weights are multiplied by the corresponding indicators to obtain the weighted index system.

[0089] A set of evaluation factors, U, and a set of evaluation levels, V, are constructed. The evaluation levels are typically {very low, low, medium, high, very high}. The set of evaluation factors is a set of dimensions for risk assessment, for example: U = {Quality pass rate (Q1), On-time delivery rate (Q2), Financial liquidity ratio (Q3), Geopolitical stability (Q4)}. Then, a fuzzy relation matrix R is constructed, where each element r_ll' represents the membership degree of the l-th factor to the l'-th evaluation level. For quantitative indicators (numerical quantifiable indicators), the membership degree is calculated using a fuzzy membership function; for qualitative indicators (textual or graded indicators), the membership degree is determined through expert evaluation (e.g., using the Delphi method). The comprehensive evaluation result is then obtained. ;in," " indicates fuzzy composition operation, Let M(·, ⊕) be the weight vector of the evaluation factors, representing the relative importance of each factor; a weighted average (M(·, ⊕)) calculation is used. The final score S is calculated using membership degree weighting.

[0090] ;in, For the first The quantitative value of each evaluation level, This indicates the comprehensive evaluation results for the first... Membership degree of each evaluation level This represents the total number of evaluation levels. The above process is repeated for each risk dimension to obtain the supplier's multidimensional risk score vector VS=[s1, s2, ..., sk], where sk represents the supplier's score on the k-th risk dimension; thus, the multidimensional risk score vector is obtained.

[0091] Cluster analysis is performed on the multidimensional risk score vectors to identify supplier risk patterns. The K-means++ clustering algorithm is used. First, the optimal number of clusters K is determined by using the silhouette coefficient method, selecting the K value that maximizes the silhouette coefficient as the optimal number of clusters. Then, the K-means++ algorithm is executed, resulting in K risk categories {C1, C2, ..., CK}, with each supplier assigned to one risk category.

[0092] Finally, the supplier's multidimensional risk score vector is combined with risk category information (generated through cluster analysis (K-means++)) to generate a supplier risk profile matrix. The rows of the supplier risk profile matrix represent different suppliers, and the columns include multidimensional risk scores and risk category identifiers, i.e., the corresponding matrix element M[f]=[VS, f, Cf], where VS, f is the risk score vector of the f-th supplier, and Cf is its risk category. The supplier risk profile matrix provides a comprehensive view of the supplier's risk status, supporting subsequent risk management and decision-making.

[0093] In one specific embodiment, the process of executing step 3, "based on the supplier risk profile matrix and combined with the component production process parameters, training the quality fluctuation prediction model, constructing the component quality risk prediction engine, and generating the component quality risk prediction results," specifically includes the following steps:

[0094] Collect process parameters, equipment status, environmental conditions, and testing data during the parts production process to form a production process feature set;

[0095] The production process feature set and the supplier risk profile matrix are fused to construct a quality prediction training dataset. Based on the quality prediction training dataset, a basic model for quality fluctuation prediction is trained using deep learning methods. The basic model for quality fluctuation prediction is then transferred to adapt to the quality prediction needs of different component types, resulting in a multimodal prediction model group. Based on the multimodal prediction model group, an ensemble learning framework is constructed, and the prediction accuracy is improved through a model weighted voting mechanism.

[0096] By integrating the integrated learning framework with the real-time data processing module, a sustainable and self-optimizing component quality risk prediction engine is built. The component quality risk prediction engine is used to assess the current and future quality status of components and generate component quality risk prediction results, including the probability of quality fluctuation, potential defect types, quality risk level, and risk occurrence time prediction.

[0097] Specifically, the parameter acquisition for the component manufacturing process utilizes multi-level sensor networks and Industrial Internet of Things (IIoT) technology. Process parameters include processing temperature, pressure, rotational speed, and feed rate; equipment status includes equipment operating time, vibration frequency, temperature rise, and energy consumption data; environmental conditions include workshop temperature and humidity, air quality, and cleanliness; and detection data includes dimensional deviations, surface roughness, material hardness, and fatigue strength. For continuously produced components, a sliding window method is used to extract temporal features. The window size w is determined based on the production cycle. For each time point u, the feature vector is F(u) = [p(u-w+1), p(u-w+2), ..., p(u)], where p(u) represents the parameter value vector at time point u. For discretely produced components, batch aggregation is used to extract features. For each batch s, the feature vector is:

[0098] F(s) = [mean(p(s)), std(p(s)), min(p(s)), max(p(s)), skew(p(s))]; where p(b) represents the sequence of parameter values ​​for batch b, and mean, std, min, max, and skew represent the mean, standard deviation, minimum, maximum, and skewness, respectively.

[0099] Feature fusion employs a multi-level joint embedding method. First, the production process feature F_prod and the supplier risk feature F_supp are normalized. Then, a nonlinear mapping is used to project the two types of features onto a common feature space, yielding common features. To enhance feature representation while preserving original feature information, residual connections are added to obtain enhanced features:

[0100] ;in, It is the weight matrix of the residual connection. This represents vector concatenation. A training dataset D = {(F_final(h), y(h))} is constructed, where F_final(h) is the h-th training sample (enhanced feature), and y(h) is the corresponding quality fluctuation label, which can be a continuous value (e.g., defect rate) or a discrete value (e.g., quality grade). This yields the quality prediction training dataset.

[0101] The underlying model for predicting quality fluctuations employs a deep neural network architecture. For time-series data, an LSTM network is used to capture temporal dependencies. For non-time-series data, a multilayer perceptron (MLP) is used. The model is trained using the Adam optimizer, employing a mean squared error loss function for regression tasks and a cross-entropy loss function for classification tasks, which are then integrated into the underlying model for predicting quality fluctuations.

[0102] Transfer learning employs domain adaptation techniques to address the distribution differences between different component types. The source domain data distribution and the target domain data distribution are defined. First, a base model for predicting quality fluctuations is trained on the source domain data. Then, it is adapted to the target domain through the following steps:

[0103] 1. The parameters of the feature extraction layer are frozen, and only the task-related layers are fine-tuned;

[0104] 2. Add a neighborhood discriminator d to ensure that the features generated by the feature extractor are distributed consistently in the source and target domains;

[0105] 3. Optimize the joint loss function:

[0106] ;in, It is task-related loss (MSE or CE). For balancing parameters, Domain-based discriminative loss:

[0107] ;in, The number of samples in the source domain is used to normalize the loss of the source domain, ensuring that the contribution of the source domain is proportional to the number of samples. The number of samples in the target domain. Represents the first in the source domain One sample. For the first in the target domain For each sample, f_feat is the feature extractor; The values ​​within parentheses are processed by the domain discriminator d. By applying the above transfer learning process to different component types, a multimodal prediction model group {M_1, M_2, ..., M_G} is obtained, where M_G is the prediction model for the G-th type of component.

[0108] To construct an ensemble learning framework, specifically, define the weights of each prediction model and the calculation formula:

[0109] ;in, It is the first The error of the prediction model for similar parts on the validation set. These are parameters that control the distribution of weights. Weighted predictions are performed based on these weights, meaning the prediction result is multiplied by the weights. To further improve performance, a meta-learner is used to integrate the predictions of the prediction model, which is the ensemble learning framework; the meta-learner can be a simple model such as logistic regression or random forest.

[0110] The component quality risk prediction engine architecture adopts a microservice design, comprising five core components: data access service, feature processing service, model inference service, result feedback service, and self-optimization service. The data access service is responsible for real-time collection of production data and supplier risk data; the feature processing service performs data cleaning, feature extraction, and feature fusion; the model inference service calls appropriate prediction models to generate quality risk prediction results; the result feedback service pushes the prediction results to relevant systems and collects feedback; and the self-optimization service periodically retrains the model and updates its parameters based on the difference between the prediction results and actual quality performance. Self-optimization employs an online learning method. For each batch of new data (X_new, Y_new), X_new represents the new input data, and Y_new represents the corresponding true label or target value.

[0111] Perform incremental update:

[0112] ;in, Indicates at time step The model parameters are a vector containing all the learnable parameters of the model. Indicates at time step Model parameters, It's the learning rate. The gradient of the loss function is used; to prevent catastrophic forgetting, elastic weights are introduced for integration.

[0113] ;in, It is a parameter Fisher's information matrix These are the original parameters. It is the regularization strength. These are the integrated parameters;

[0114] Real-time acquisition of component process parameters and the latest risk status of suppliers in current production; data preprocessing and feature fusion through feature processing services; selection of appropriate prediction models or model combinations based on component type; generation of multi-dimensional quality risk assessment results through model inference services:

[0115] Quality fluctuation probability: This indicates the probability that a component will experience quality abnormalities within a future time window.

[0116] Potential defect types: Predicting the specific defect types that may occur and their probability distribution;

[0117] Quality risk level: The risk is quantified into three levels: high, medium and low, for decision-making reference;

[0118] Risk occurrence time prediction: estimating the time point or time interval in which quality problems may occur.

[0119] The prediction results are pattern-matched with historical quality problem data to enhance the interpretability of the predictions; formatted component quality risk prediction results are output as a key input for subsequent integrated supply chain risk assessment.

[0120] In one specific embodiment, the process of performing step 4, "conducting spatiotemporal characteristic analysis and anomaly detection on supply chain logistics nodes, forming a logistics risk monitoring network, and generating logistics node risk assessment results," specifically includes the following steps:

[0121] The location information, transportation time, transit status and environmental conditions of logistics nodes are collected to form a logistics spatiotemporal dataset; the time series features of the logistics spatiotemporal dataset are extracted to identify the seasonal patterns and periodic changes in logistics operations and obtain a time series feature vector.

[0122] Based on time-series feature vectors, a time-series prediction model for logistics nodes is constructed to generate a baseline for normal logistics behavior; a multivariate anomaly detection algorithm is defined to compare the deviation between the current logistics status and the baseline for normal logistics behavior in real time to identify potential anomalies; the identified potential anomalies are classified and their risk levels are assessed to form a logistics anomaly risk database.

[0123] Based on the logistics anomaly risk database, risk network modeling is performed on logistics nodes to construct a logistics risk monitoring network. According to the logistics risk monitoring network, the comprehensive risk index of the nodes is calculated. The exponential smoothing method is used to perform time-series prediction of the comprehensive risk index of each node to obtain the predicted risk index. Based on the predicted risk index, the set of key risk nodes is identified. The set of key risk nodes and the corresponding predicted risk index are the logistics node risk assessment results.

[0124] Specifically, logistics spatiotemporal data collection employs multi-source sensing technology and IoT devices. Location information is obtained in real-time through GPS / BeiDou positioning systems, acquiring the spatial coordinates of logistics vehicles and goods. Transportation time includes departure time, arrival time, dwell time, and total transportation duration. Transit status includes loading / unloading status, warehousing status, and en route status. Environmental conditions include temperature, humidity, vibration intensity, and air pressure. For critical components, special environmental parameters such as electromagnetic radiation and chemical exposure are also collected. The data collection frequency is dynamically adjusted according to the importance of the logistics links, with high-frequency sampling (1 minute / time) for critical nodes and low-frequency sampling (30 minutes / time) for ordinary nodes. All data is transmitted in real-time to the cloud platform via an IoT gateway, forming a logistics spatiotemporal dataset D1={(p_i, t_i, s_i, e_i)}, where p_i is location information, t_i is time information, s_i is status information, and e_i is environmental information.

[0125] First, the data is segmented based on logistics node type, dividing the time-series data into loading, transportation, transshipment, and unloading segments. Then, statistical features (mean, variance, quantiles, etc.), frequency domain features (spectral features obtained through Fast Fourier Transform), and morphological features (trends, seasonality, periodicity, etc.) are extracted for each segment. For location data, motion features such as speed, acceleration, and dwell point distribution are calculated; for time data, delay features such as average delay time and delay standard deviation are calculated; for state data, state transition frequency and duration distribution are extracted; and for environmental data, exceedance frequency and fluctuation characteristics are calculated.

[0126] Identify seasonal patterns and cyclical changes in logistics operations:

[0127] ;in, It is a trend item. It is a seasonal item. It is the residual term. For time indexing, This involves raw time-series data of logistics operations (logistics performance indicators that change over time, such as transit time, arrival rate, and delay frequency). The STL method is used for decomposition, followed by spectral analysis of the seasonal components to identify major cycles (seasonal patterns and periodic variations). These major cycles are then combined with statistical, frequency domain, and morphological features (vector-level concatenation) to form the time-series feature vector F_time.

[0128] The logistics node time series forecasting model employs an ensemble approach combining the Prophet algorithm and the ARIMA model. The Prophet algorithm is based on an additive model framework.

[0129] ;in, It is a trend function. It is a seasonal function. It is a holiday effect function. This is the error term. The ARIMA model is based on difference equation ensemble, determining model parameters through minimizing information criteria (such as AIC or BIC). The predictions from the two models are fused using a weighted average, and the resulting output constitutes a baseline for normal logistics behavior. ,in It is a predicted value. It is the standard deviation of the forecast.

[0130] The system defines point anomaly detection rules based on the 3σ principle; then it defines context anomaly detection, calculated using the Local Outlier Factor (LOF) algorithm; and finally it defines set anomaly detection, implemented using the Subsequence Density Distribution Change Detection Algorithm (DPARSAD). By combining point anomaly detection rules, context anomaly detection, and set anomaly detection, an anomaly score (a weighted sum of the three anomaly types) is calculated. When the anomaly score is greater than a preset anomaly threshold, it is marked as an anomaly point, i.e., a potential anomaly.

[0131] Define an anomaly category set C1 = {delay anomaly, path deviation, environmental exceedance, state anomaly, ...}. Classify the detected potential anomalies using a decision tree model:

[0132] ;in, These are the abnormal feature vectors, and f_DT is the decision tree model. It is an exception belonging to the category The probability of an anomaly is then determined. The risk level of each anomaly is then assessed, considering three factors: anomaly severity S1, impact range R1, and duration D2. The risk score Ris is calculated using the following formula:

[0133] ;in, , , All values ​​are normalized to the [0, 1] interval. Based on risk scores, anomalies are categorized into five risk levels: extremely low (0-0.2), low (0.2-0.4), medium (0.4-0.6), high (0.6-0.8), and extremely high (0.8-1.0). All detected anomalies, their classifications, and risk levels are stored in the logistics anomaly risk database L1={(ai, ci, ri)}, where ai is the anomaly characteristic, ci is the anomaly category, and ri is the risk level.

[0134] Based on a logistics anomaly risk database, a logistics node network G1=(V1,E1) is constructed, where V1 is the set of nodes and E1 is the set of edges. Nodes represent entities in the logistics system (such as suppliers, warehouses, distribution centers, customers, etc.), and edges represent logistics paths. Then, based on the logistics anomaly risk database, a risk index is calculated for each node v1.

[0135] ;in, λ1 is the set of anomalies associated with node v1, t_a1 is the occurrence time of anomaly a1, t_now is the current time, and λ1 is the time decay factor. Calculate the risk propagation probability P_risk(e) for each edge e=(u1,v1), where u1 and v1 are the nodes connected by edge e;

[0136] ;in, , and It is a weighting coefficient and , It represents the historical risk level of edge e (obtained based on machine learning). The risk index is u1.

[0137] A logistics risk monitoring network G_risk=(V1, E1, R2, P2) is constructed, where R2 is the risk index mapping of nodes and P2 is the edge risk propagation probability mapping. The logistics risk monitoring network is updated and displayed in real time using visualization technology, with different risk levels marked by different colors, and the risk propagation probability represented by the thickness of the edges.

[0138] Based on the constructed logistics risk monitoring network, the comprehensive risk index of nodes is calculated, which is the weighted sum of node centrality index, node vulnerability index and risk index; the set of key risk nodes is identified by the threshold method; that is, a high risk threshold is set, usually set as the upper quartile of the risk distribution, and nodes whose comprehensive risk index is greater than the high risk threshold are key risk nodes.

[0139] In one specific embodiment, the process of performing step 5, "constructing a supply chain network topology model based on the supply chain knowledge graph, performing risk propagation path analysis, and generating a risk transmission diagram," specifically includes the following steps:

[0140] Extract the relationships between suppliers, manufacturers, and logistics service providers from the supply chain knowledge graph to construct an initial network graph;

[0141] By employing complex network analysis methods, we calculate the centrality index and critical path of network nodes to identify key nodes in the supply chain.

[0142] Based on key nodes, a network vulnerability assessment model is constructed to quantitatively analyze the stability of the network structure, thereby obtaining the network vulnerability assessment results.

[0143] Define a risk propagation dynamics model to simulate the diffusion process of risk events in the supply chain network and identify risk cascading effects;

[0144] Monte Carlo simulations were used to perform probabilistic analysis on the propagation paths of different risk events, resulting in a risk propagation probability matrix.

[0145] Based on the risk propagation probability matrix and network vulnerability assessment results, a dynamic risk transmission diagram is constructed to visualize risk propagation.

[0146] Specifically, supply chain network relationship extraction employs knowledge graph query and relationship mining techniques. Direct relationships between entities are extracted from the supply chain knowledge graph, including supply relationships, contractual relationships, and ownership relationships; simultaneously, implicit relationships are mined, such as shared geographical locations, common customers, and technological dependencies. First-order relationships are directly obtained through SPARQL queries.

[0147] For higher-order relationships, multi-hop paths connecting two entities are identified using path analysis algorithms such as Floyd-Warshall. All extracted relationships constitute an edge set E={(u2, v2, r2, w2)}, where u2 and v2 are entities, r2 is the relationship type, and w2 is the relationship strength. Relationship strength is calculated (weighted sum) based on relationship type, historical transaction frequency, and importance score. Relationship type is quantified (e.g., supplier relationship = 1.0, cooperation relationship = 0.8), historical transaction frequency is normalized to [0, 1] (e.g., 5 transactions per month → 0.7), and importance score is assigned (e.g., key supplier = 1.0).

[0148] In the initial network graph G2=(V3,E3), V3 is the set of entities, including suppliers, manufacturers, logistics service providers, etc., and E3 is the set of extracted relationships.

[0149] Four types of centrality metrics are calculated, including degree centrality, proximity centrality, betweenness centrality, and eigenvector centrality. Degree centrality is the number of connections a node has, proximity centrality is the reciprocal of the average shortest path length from a node to other nodes, betweenness centrality is the frequency of a node being on the shortest path, and eigenvector centrality is a centrality measure that takes into account the importance of its neighbors.

[0150] Identify the critical path in the network graph:

[0151] Where P3 is the set of all paths, and Duration(p2) is the total duration of path p2. Combining centrality metrics and critical paths, a set K3 of critical nodes in the supply chain is identified. Specifically, this is filtered by defining thresholds; for example, if the centrality metric is >0.7 and the critical path is on ≥3 critical paths, it is marked as a critical node.

[0152] Define node removal strategies, including random removal, degree-first removal, and centrality-first removal. For each strategy, remove nodes from G2 incrementally and calculate the size of the largest connected component G2(p4), where p4 is the proportion of nodes retained. The percolation threshold p_c is defined as the critical point that causes G2(p4) to abruptly change from 0 to a positive value; a larger p_c indicates a more fragile network. Define network robustness metrics:

[0153] ;in, It is the number of sampling points. The number of nodes to retain; , A higher value indicates a more robust network. For the set of critical nodes in the supply chain, a vulnerability coefficient is defined:

[0154] ;in, The size of the large connected component when all nodes are retained. This indicates the relative decrease in network connectivity after removing all critical nodes. The stability of the network structure is quantitatively analyzed using the aforementioned network robustness indicators and vulnerability coefficients. Specifically, thresholds are defined for screening; for example, if the network robustness indicator is <0.5 and the vulnerability coefficient is >0.6, the network structure is considered vulnerable. A heatmap is generated to show the changes in connectivity after node removal, forming the network vulnerability assessment result.

[0155] The risk propagation dynamics model adopts a variant of the SIR model. The set of node states is defined as S3 = {healthy (H), exposed (E), affected (I), recovered (R)}. For any time t, the state S3_v1(t) ∈ S of node v1 is determined by both the node's own attributes and the states of its neighbors.

[0156] ;in, For a moment The state transition probability of node v1 Risk originates from the node spread to The probability, Let v1 be the set of its neighboring nodes; It is an indicator function, when the node The function value is 1 when the state is affected (I) at time t; otherwise, it is 0. Let t be the node The state; It is the product of the corresponding edge weight and the vulnerability coefficient of node v1.

[0157] By iteratively simulating the state transition process described above, the propagation of risk in the network is simulated, and risk cascading paths and the set of affected nodes are identified, i.e., the risk cascading effect.

[0158] To perform probabilistic analysis on the propagation paths of different risk events, a set of risk events is first defined. Each event in the risk event set has an initial outbreak node set Oi and risk characteristic parameters θi. For each risk event ei, N3 independent simulations are performed (e.g., N=1000), recording the risk propagation path and affected nodes in each simulation. For any two nodes u1 and v1, the risk propagation probability is defined as the number of times the risk propagates from u1 to v1 in N3 simulations divided by N3. A preliminary risk propagation probability matrix Ti is constructed, where the element Ti[u1, v1] is the risk propagation probability between nodes u1 and v1. For all risk events, the risk propagation probability matrix is ​​calculated, which is to multiply each risk propagation probability by its corresponding weight.

[0159] The risk transmission graph employs multi-layer network visualization technology. The basic network layer G_base consists of supply chain entities and relationships; the risk assessment layer G_risk includes node risk scores and edge propagation probabilities; and the time-dynamic layer G_time displays the evolution of risk over time. A dynamic risk transmission graph G_cascade is constructed, including node sets, edge sets, node risk mappings, edge propagation probability mappings (risk propagation probability matrices), and time evolution functions (such as Gaussian functions). Interactive visualization technology enables the dynamic display of risk propagation, including:

[0160] 1. Node color coding: Gradually changes from green (low) to red (high) according to the risk level;

[0161] 2. Edge width encoding: Adjust the edge thickness according to the propagation probability;

[0162] 3. Node size encoding: Adjust the size of nodes according to their importance;

[0163] 4. Timeline Control: View the spread of risk at different points in time by sliding the timeline;

[0164] 5. Risk Profile Diagram: Shows the risk propagation path and probability distribution between any two points;

[0165] By using risk transmission diagrams, decision-makers can intuitively understand the possible paths, key nodes, and speed of risk transmission, providing decision support for risk prevention and control.

[0166] In one specific embodiment, the process of executing step 6, "establishing a multi-agent collaborative risk assessment model, integrating component quality risk prediction results, logistics node risk assessment results, and risk transmission diagrams, conducting integrated assessments of risks at each stage of the supply chain, and forming a comprehensive risk scoring system," specifically includes the following steps:

[0167] Define a supplier risk assessment agent, a quality risk assessment agent, a logistics risk assessment agent, and a market risk assessment agent, each responsible for risk assessment in different dimensions;

[0168] Configure expert knowledge bases and machine learning models for each intelligent agent to achieve semi-supervised risk learning capabilities;

[0169] Input the component quality risk prediction results into the quality risk assessment intelligent agent, and input the logistics node risk assessment results into the logistics risk assessment intelligent agent;

[0170] The risk transmission diagram is input into each agent to assess the potential impact of risk propagation on each stage;

[0171] Construct communication protocols and collaborative decision-making mechanisms among intelligent agents to enable different agents to exchange risk information and coordinate assessment results, forming a self-evolving comprehensive risk scoring system.

[0172] Specifically, the multi-agent system architecture adopts a layered distributed design. The bottom layer is the data acquisition layer, including various data interfaces and sensor networks; the middle layer is the agent layer, containing multiple specialized agents; and the top layer is the collaborative decision-making layer, responsible for communication and result fusion between agents. Each agent is defined as a triple A3=(P4, K4, M4), where P4 is the perception module, K4 is the knowledge base, and M4 is the evaluation model. The supplier risk assessment agent A_sup is responsible for assessing risks such as the supplier's financial status, production capacity, and management level; the quality risk assessment agent A_qua is responsible for assessing risks such as product quality fluctuations, defect rates, and reliability; the logistics risk assessment agent A_log is responsible for assessing risks such as transportation delays, inventory shortages, and delivery anomalies; and the market risk assessment agent A_mkt is responsible for assessing risks such as demand fluctuations, price changes, and competitive landscape.

[0173] Each agent has a specific mechanism for processing input data: the quality risk assessment agent A_qua receives the component quality risk prediction results generated in step 3, uses them as the main input data for assessing quality risk, and combines them with historical quality problem data to perform risk enhancement analysis; the logistics risk assessment agent A_log receives the logistics node risk assessment results produced in step 4, integrates them with historical logistics performance data, and forms a complete logistics risk view; all agents receive the risk transmission diagram generated in step 5, and analyze the potential impact of the risk propagation path and cascading effect in the network on their respective areas of responsibility.

[0174] The expert knowledge base employs a combination of ontology reasoning and case-based reasoning. The ontology knowledge base K_ont uses OWL-DL to describe domain concepts and rules; the case base K_case stores historical risk events and corresponding countermeasures; and the rule base K_rule contains the IF-THEN rule set for risk assessment. The machine learning model uses a semi-supervised learning method to address the problem of scarce labeled data. Model training utilizes federated learning, enabling agents to collaboratively train the model without sharing raw data. Each agent A_i trains model M_i on local data D_i, then uploads the model parameters θ_i to the coordinator for aggregation (weighted sum). The aggregated global model parameters are then distributed to all agents for the next round of training.

[0175] The application mechanisms of risk transmission diagrams include:

[0176] Node impact analysis: Each agent analyzes the vulnerability indicators of nodes related to its responsibilities in the risk transmission graph;

[0177] Path dependency modeling: Calculate the propagation probability and time delay of risk events based on the topology of the risk transmission graph;

[0178] Cross-impact assessment: Analyze the mutual impact between different risk areas, such as the transmission effect of supplier risk on logistics and quality;

[0179] Critical Link Monitoring: Identify critical links in the risk transmission diagram and prioritize the assessment of the risk status on these links.

[0180] The inter-agent communication protocol adopts an asynchronous messaging system based on a publish / subscribe pattern. A message type set M5 is defined as {risk warning, state update, model parameters, collaboration request, ...}, where each message includes a sender ID, receiver ID, message type, content, and timestamp. A message bus is established to implement a many-to-many message routing and subscription mechanism. To handle communication latency and loss issues, message acknowledgment and retransmission mechanisms are introduced, and a message history queue is maintained to support message backtracking and analysis.

[0181] The collaborative decision-making mechanism employs a two-layer game model. The bottom layer represents the internal decision-making process of the agents, modeled using a Markov Decision Process (MDP).

[0182] MDP=(S5, A5, P5, R5, γ5)

[0183] Where S5 is the state space, A5 is the action space, P5 is the state transition probability, R5 is the reward function, and γ5 is the discount factor. The agent solves for the optimal policy π* using dynamic programming.

[0184] The top layer involves collaboration among intelligent agents, employing an integrated voting mechanism to consolidate the evaluation results of each agent. When disagreements arise, a consensus is reached through a negotiation mechanism.

[0185] 1. Each agent o submits the evaluation result r1_o and the confidence level c1_o;

[0186] 2. Preliminary Integration Results ;

[0187] 3. Calculate the degree of divergence ;

[0188] 4. If the degree of divergence If the value exceeds the preset threshold, a negotiation mechanism is triggered: each agent shares the evaluation criteria, re-evaluates and submits updated results, thereby forming a comprehensive risk scoring system.

[0189] In one specific embodiment, the process of executing step 7, "based on a comprehensive risk scoring system and combined with multivariate time series analysis, generating dynamic risk warning thresholds, and then obtaining warning results to achieve differentiated risk warnings; based on the warning results and the historical response strategy database, constructing a risk response decision tree to provide intelligent risk response suggestions," specifically includes the following steps:

[0190] Collect historical risk event data and corresponding risk indicator time series to construct a risk-indicator association dataset; perform time series pattern mining on the risk-indicator association dataset to identify risk warning features and form an early warning indicator library.

[0191] Based on the early warning indicator library, a multi-level early warning threshold calculation model is defined to generate differentiated thresholds for different types of parts and suppliers.

[0192] A Bayesian network early warning model is constructed, and combined with a comprehensive risk scoring system, the risk probability and early warning level are obtained, i.e., the early warning result.

[0193] Establish a supply chain risk classification system, structurally store historical risk events and their corresponding countermeasures to form a historical response strategy library; based on this library, use case-based reasoning to extract the association rules between risk scenarios and countermeasures; utilize early warning results and critical path analysis of the risk transmission diagram to identify the current risk situation feature vector; based on this feature vector, use the C4.5 algorithm to construct a risk response decision tree, mapping risk attributes to a set of response strategies; analyze the risk propagation path and impact nodes according to the risk transmission diagram, and prioritize the response strategies in the risk response decision tree; formulate the final risk response plan; transform the risk response plan into specific execution tasks, generating intelligent risk response suggestions that include action steps, responsible persons, and time nodes.

[0194] Historical risk event data includes detailed information on specific risk events such as supply chain disruptions, supplier defaults, quality issues, and logistics delays, including the time of occurrence, scope of impact, duration, and extent of loss. The risk indicator time series contains various indicator data corresponding to these events, such as supplier financial ratios, quality pass rates, logistics timeliness indicators, and market volatility indices. By associating risk events with indicator data for corresponding time periods, a risk-indicator association dataset is constructed, providing foundational data for subsequent time series pattern mining.

[0195] Temporal pattern mining is performed on risk-indicator association datasets to identify risk precursor features and form a warning indicator library. The temporal pattern mining employs a Long Short-Term Memory (LSTM) network structure, which effectively captures long-term dependencies and temporal patterns in indicator sequences, i.e., risk precursor features. It can identify typical indicator change patterns before risk events occur, thereby determining warning precursor indicators. These precursor indicators are ranked according to their predictive ability for different risk types, forming the warning precursor indicator library.

[0196] Based on a leading early warning indicator library, a multi-level early warning threshold calculation model is defined to generate differentiated thresholds for different types of components and suppliers. The multi-level early warning threshold model employs a combination of decision trees and fuzzy logic. Based on factors such as the importance level of components (critical, important, general), the risk category of suppliers (high risk, medium risk, low risk), and the criticality of production lines (main line, auxiliary line), customized early warning thresholds are generated for different combinations. The decision tree structure is used to divide different hierarchical categories, while each leaf node corresponds to a fuzzy logic system used to generate specific threshold parameters. The membership function of the fuzzy logic system is:

[0197] ;in, This is a risk indicator value. With the center point, and To control parameters, the shape of the membership function is determined. Through a multi-level early warning threshold calculation model, the system can generate differentiated early warning thresholds for different types of components and suppliers, achieving refined risk management.

[0198] A Bayesian network early warning model is constructed, combined with a comprehensive risk scoring system, to achieve risk probability inference and early warning level determination. The Bayesian network early warning model consists of a graph structure composed of nodes (representing risk factors) and directed edges (representing causal relationships). Each node has a conditional probability table, representing the probability distribution of that node under different states of its parent node. The posterior probability of a risk event is calculated using Bayes' theorem. Through inference calculation, the system can obtain the risk probability of different risk events and determine the early warning level (normal, attention, warning, emergency) based on the risk probability value. The criteria for classifying early warning levels are based on a comprehensive consideration of risk probability and potential impact, ensuring that the system can issue appropriate levels of early warning for different risk situations.

[0199] The supply chain risk classification system adopts a multi-level classification structure. The primary classification includes supplier risk, quality risk, logistics risk, and market risk; the secondary classification is further refined, such as subdividing supplier risk into financial risk, capacity risk, and management risk; the tertiary classification is down to specific risk event types, such as subdividing financial risk into cash flow difficulties and credit rating downgrades. For each type of risk event, a standardized description template is established, including attribute fields such as risk code, risk description, triggering conditions, scope of impact, severity, duration, and frequency.

[0200] The historical response strategy repository employs a storage method combining relational databases and knowledge graphs. The relational database stores structured information, including basic risk information tables, response measure tables, execution record tables, and effect evaluation tables. The knowledge graph stores complex relationships, including causal relationships between risks, matching relationships between risks and response measures, and synergistic relationships between response measures. Each response strategy record includes fields such as strategy ID, applicable risk type, preconditions, execution steps, resource requirements, expected results, actual results, and lessons learned. Historical response strategies are sourced from internal corporate risk response records, industry best practices, expert experience summaries, and academic research findings.

[0201] The case-based reasoning method is based on similarity measurement and knowledge transfer principles. First, a framework for calculating risk scenario similarity is defined, comprehensively considering similarity in risk type, scope of impact, severity, and environmental conditions. Different similarity calculation methods are used for different attribute types: categorical attributes use tree-structured semantic distance; numerical attributes use normalized Euclidean distance; and textual attributes use semantic vector cosine similarity. The case database retrieval uses the KNN algorithm to retrieve the most similar historical cases from a historical response strategy database.

[0202] Association rule extraction employs an improved version of the Apriori algorithm. Risk scenario attribute sets are used as antecedent itemsets, and response measure attribute sets are used as consequent itemsets, mining association rules in the form of "In risk scenario X, response measure Y should be taken." To improve rule quality, three evaluation metrics are introduced: support, confidence, and lift. Only rules that simultaneously satisfy the minimum support threshold, minimum confidence threshold, and minimum lift threshold are retained. For rule conflicts, a rule ranking mechanism is used, with ranking based on confidence, historical success rate, and expert confirmation. The final result is an association rule base, providing a rule foundation for decision tree construction.

[0203] The construction process of risk scenario feature vectors includes two stages: feature selection and vector generation. Feature selection employs the information gain method, extracting the most discriminative attributes from early warning results and risk transmission maps. Commonly used features include risk type, risk level, scope of impact, urgency, propagation speed, damage level of key nodes, and feasibility of alternative solutions. The critical path analysis provided by the risk transmission map includes risk source node identification, risk diffusion path prediction, affected node assessment, and network vulnerability analysis. Feature vector generation uses a hybrid encoding method: one-hot encoding for categorical features, normalization for numerical features, and ordinal encoding for sequential features, ultimately forming a fixed-dimensional feature vector representing the complete characteristics of the current risk scenario.

[0204] The C4.5 decision tree construction is based on the principles of information entropy and information gain ratio. First, a training dataset is constructed, with samples in the form of <feature vector, response strategy label>, extracted from a historical response strategy database. Then, the information gain ratio of each feature is calculated, and the feature with the highest gain ratio is selected as the root node for splitting. Subtrees are recursively constructed until a termination condition is met. To handle continuous and missing feature values, the C4.5 algorithm uses the principle of minimizing information entropy to determine the optimal split point and introduces a weight allocation mechanism to handle missing values. The constructed decision tree is then pruned and optimized, including pre-pruning and post-pruning. Pre-pruning limits tree growth by using an early stopping condition, while post-pruning removes unreliable branches through error estimation, balancing the model's complexity and accuracy.

[0205] The decision tree mapping process starts from the root node and propagates the input risk scenario feature vector along the decision path to the leaf node, obtaining the corresponding set of coping strategies. For cases with incomplete feature values, a multi-path exploration method is used to perform weighted decisions based on the probability distribution of different feature values. For new scenarios not covered by the decision tree, the nearest neighbor interpolation method is used to find the most similar known scenario in the feature space and borrow its coping strategies.

[0206] Prioritizing response strategies takes into account multiple factors. First, the risk propagation path is analyzed using a risk transmission diagram, calculating the risk exposure and time urgency at each node. Second, the effectiveness indicators of each response strategy are evaluated, including implementation cost, implementation time, expected results, and probability of success. Third, corporate resource constraints and business objectives are considered, balancing short-term response with long-term prevention. Finally, a comprehensive scoring model is established using the analytic hierarchy process (AHP) to rank the alternative response strategies, forming a priority queue.

[0207] The risk response plan development process follows the principles of systematicity and synergy. First, based on the systemic impact of the risk, response strategies are categorized into three types: direct intervention strategies, mitigation strategies, and backup strategies. Direct intervention strategies target the source of the risk, such as replacing problematic suppliers; mitigation strategies address the impact of the risk, such as adjusting production plans; and backup strategies provide security guarantees, such as activating alternative suppliers. Then, different strategies are combined and optimized to resolve dependencies and conflicts between them, forming a collaborative response plan. Finally, the plan undergoes feasibility verification, including resource adequacy checks, time constraint checks, and risk reassessment, to ensure its executability and effectiveness.

[0208] The intelligent risk response recommendation generation method combines structured templates with natural language generation technology. First, the risk response plan is broken down into specific execution tasks. Each task includes attributes such as task description, execution steps, resource requirements, responsible department, collaborating departments, start time, completion time, priority, and evaluation criteria. Then, based on the company's organizational structure and personnel responsibilities, task leaders are automatically assigned, and reasonable timelines are set. Finally, natural language generation technology is used to transform the structured task information into clear and concise text descriptions, generating a standardized risk response recommendation report. The report includes sections on risk overview, impact assessment, response strategies, task breakdown, resource allocation, timeline, and effectiveness monitoring.

[0209] The intelligent push notification strategy for risk response recommendations employs a multi-channel, tiered approach. Different notification methods are selected based on the urgency and importance of the risk: high-urgency risks receive instant messaging and telephone notifications, while general risks receive email and system message notifications. Information content is pushed at different levels based on user roles and permissions: management-level users receive decision-making information, while execution-level users receive operational guidance.

[0210] The risk response closed-loop management mechanism comprises three stages: execution monitoring, effectiveness evaluation, and knowledge updating. Execution monitoring tracks the execution status of each response task through a task management system, promptly identifying and resolving execution obstacles. Effectiveness evaluation comprehensively assesses the actual effectiveness of response measures using qualitative and quantitative indicators, including the degree of risk elimination, resource consumption, and business impact. Knowledge updating feeds new response experiences into the historical response strategy database, enriching the case library, optimizing association rules, improving the accuracy of decision trees, and forming a continuously improving intelligent risk management system.

[0211] This application also provides an AI-based automotive parts enterprise supply chain risk warning device. The AI-based automotive parts enterprise supply chain risk warning device includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor performs the steps of the AI-based automotive parts enterprise supply chain risk warning method in the above embodiments.

[0212] This application also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the artificial intelligence-based automotive parts enterprise supply chain risk warning method.

[0213] This application comprehensively enhances the accuracy, foresight, and decision-making efficiency of supply chain risk management for automotive parts companies. By deeply integrating multi-dimensional data sources and dynamic knowledge graphs, it breaks through traditional information silos, achieving a penetrating insight into risks across the entire chain, from supplier qualifications to production quality fluctuations and logistics anomalies. It accurately identifies potential risk transmission paths and cascading effects, shortening risk warning response time. The constructed multimodal risk assessment system, combined with real-time context adaptive adjustment of warning thresholds, effectively solves the long-standing industry problem of false alarms and missed alarms, improves the accuracy of risk identification at key nodes, and significantly reduces production losses caused by supply chain disruptions. A decision support mechanism based on intelligent inference deeply couples historical experience with real-time situations, automatically generating response strategy combinations matched to the company's resource endowment, driving risk management from passive emergency response to proactive prevention and control, and assisting companies in quickly building resilient supply chain networks in complex market environments. Furthermore, through risk transmission visualization and attribution analysis, it empowers management to accurately locate the root causes of risks, optimize supplier strategic layout and inventory strategies, and systematically reduce supply chain vulnerability. This not only improves the efficiency of enterprise risk handling but also reduces the hidden costs caused by quality defects and logistics delays through preventative control.

[0214] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0215] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0216] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0217] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. An AI-based method for early warning of supply chain risks in automotive parts enterprises, characterized in that: include: Step 1: Collect and integrate multi-source heterogeneous data from the supply chain to construct a supply chain knowledge graph, including: Set up a multi-source data acquisition interface to obtain raw data from the enterprise resource planning system, supplier management system, quality management system and external databases to form an initial dataset; The initial dataset is cleaned and standardized to obtain a normalized dataset; Entity recognition and relation extraction are performed on the normalized dataset to construct an initial semantic network; Based on the initial semantic network, an ontology model is constructed to determine entity categories, attributes, and relationship types, forming a supply chain ontology framework; The normalized dataset is mapped to the supply chain ontology framework to generate an instantiated knowledge base; The instantiated knowledge base is subjected to knowledge reasoning and consistency checks to obtain an optimized supply chain knowledge graph; Step 2: Based on the supply chain knowledge graph, extract and quantify the multi-dimensional risk characteristics of suppliers to generate a supplier risk profile matrix; Step 3: Based on the supplier risk profile matrix and combined with the component production process parameters, train the quality fluctuation prediction model, build a component quality risk prediction engine, and generate component quality risk prediction results. Step 4: Conduct spatiotemporal characteristic analysis and anomaly detection on supply chain logistics nodes to form a logistics risk monitoring network and generate logistics node risk assessment results; Step 5: Based on the supply chain knowledge graph, construct a supply chain network topology model, perform risk propagation path analysis, and generate a risk transmission diagram; Step 6: Establish a multi-agent collaborative risk assessment model, integrate the component quality risk prediction results, logistics node risk assessment results and risk transmission diagram, and conduct integrated assessment of the risks in each link of the supply chain to form a comprehensive risk scoring system; Step 7: Based on the comprehensive risk scoring system and combined with multivariate time series analysis, generate dynamic risk warning thresholds and obtain warning results; based on the warning results and the preset historical response strategy library, construct a risk response decision tree and provide intelligent risk response suggestions.

2. The method for early warning of supply chain risks for automotive parts enterprises based on artificial intelligence according to claim 1, characterized in that, Based on the supply chain knowledge graph, the supplier risk profile matrix is ​​generated by extracting and quantifying multi-dimensional risk characteristics and assessing them. The supplier basic information, historical delivery records, quality performance, financial status and geopolitical factors are extracted from the supply chain knowledge graph to form a supplier feature set; Principal component analysis was performed on the supplier feature set to determine the key risk dimensions and obtain a dimensionality-reduced feature space. Based on the reduced feature space, a supplier risk assessment index system is constructed, and the weight of each index is determined by the analytic hierarchy process, thus forming a weighted index system. Suppliers are quantitatively scored across various risk dimensions to obtain a multidimensional risk score vector; Cluster analysis is performed on the multidimensional risk scoring vector to identify supplier risk patterns, form risk category divisions, and obtain risk category information; the supplier's multidimensional risk scoring vector is combined with the risk category information to generate a supplier risk profile matrix.

3. The method for early warning of supply chain risks for automotive parts enterprises based on artificial intelligence according to claim 2, characterized in that, The process involves training a quality fluctuation prediction model based on the supplier risk profile matrix and combining it with component production process parameters to construct a component quality risk prediction engine and generate component quality risk prediction results, including: Collect process parameters, equipment status, environmental conditions, and testing data during the parts production process to form a production process feature set; The production process feature set and the supplier risk profile matrix are fused to construct a quality prediction training dataset; Based on the aforementioned quality prediction training dataset, a basic model for quality fluctuation prediction is trained using deep learning methods; transfer learning is then performed on the basic model for quality fluctuation prediction to obtain a group of multimodal prediction models. Based on the multimodal prediction model group, an integrated learning framework is constructed, and then a sustainable self-optimizing component quality risk prediction engine is built. The component quality risk prediction engine is used to evaluate the current and future component quality status and generate component quality risk prediction results, including the probability of quality fluctuation, potential defect type, quality risk level, and risk occurrence time prediction.

4. The method for early warning of supply chain risks for automotive parts enterprises based on artificial intelligence according to claim 3, characterized in that, The process involves analyzing the spatiotemporal characteristics and detecting anomalies at supply chain logistics nodes to form a logistics risk monitoring network, generating risk assessment results for each logistics node, including: The location information, transportation time, transit status and environmental conditions of logistics nodes are collected to form a logistics spatiotemporal dataset; the time series features of the logistics spatiotemporal dataset are extracted to identify the seasonal patterns and periodic changes in logistics operations and obtain a time series feature vector. Based on the aforementioned time-series feature vectors, a time-series prediction model for logistics nodes is constructed to generate a baseline for normal logistics behavior. A multivariate anomaly detection algorithm is defined to compare the deviation between the current logistics status and the baseline for normal logistics behavior in real time, identifying potential anomalies. The identified potential anomalies are classified and their risk levels are assessed to form a logistics anomaly risk database. Based on the logistics anomaly risk database, a risk network model is performed on logistics nodes to construct a logistics risk monitoring network. According to the logistics risk monitoring network, a comprehensive risk index for each node is calculated. The exponential smoothing method is used to perform time-series prediction on the comprehensive risk index of each node to obtain a predicted risk index. Based on the predicted risk index, a set of key risk nodes is identified. The set of key risk nodes and their corresponding predicted risk indices constitute the logistics node risk assessment result.

5. The method for early warning of supply chain risks for automotive parts enterprises based on artificial intelligence according to claim 4, characterized in that, The process of constructing a supply chain network topology model based on the supply chain knowledge graph, analyzing risk propagation paths, and generating a risk transmission diagram includes: Extract the relationships between suppliers, manufacturers, and logistics service providers from the supply chain knowledge graph to construct an initial network graph; The process involves calculating the centrality index and critical path of network nodes to identify key nodes in the supply chain; constructing a network vulnerability assessment model based on these key nodes to quantitatively analyze the stability of the network structure; obtaining the network vulnerability assessment results; defining a risk propagation dynamics model to simulate the diffusion process of risk events in the supply chain network and identify risk cascading effects; performing probability analysis on the propagation paths of different risk events to obtain a risk propagation probability matrix; and constructing a dynamic risk transmission diagram based on the risk propagation probability matrix and the network vulnerability assessment results.

6. The method for early warning of supply chain risks for automotive parts enterprises based on artificial intelligence according to claim 5, characterized in that, The establishment of a multi-agent collaborative risk assessment model integrates the component quality risk prediction results, logistics node risk assessment results, and risk transmission diagrams to conduct an integrated assessment of risks at each stage of the supply chain, forming a comprehensive risk scoring system, including: Define a supplier risk assessment agent, a quality risk assessment agent, a logistics risk assessment agent, and a market risk assessment agent, each responsible for risk assessment in different dimensions; Configure expert knowledge bases and machine learning models for each intelligent agent; input the component quality risk prediction results into the quality risk assessment intelligent agent, input the logistics node risk assessment results into the logistics risk assessment intelligent agent, and input the risk transmission diagram into each intelligent agent; Construct communication protocols and collaborative decision-making mechanisms among intelligent agents to enable different agents to exchange risk information and coordinate assessment results, forming a self-evolving comprehensive risk scoring system.

7. The method for early warning of supply chain risks for automotive parts enterprises based on artificial intelligence according to claim 6, characterized in that, Based on the comprehensive risk scoring system, combined with multivariate time series analysis, a dynamic risk warning threshold is generated, and then the warning result is obtained; Based on the early warning results and a pre-set historical response strategy database, a risk response decision tree is constructed to provide intelligent risk response suggestions, including: Collect historical risk event data and corresponding risk indicator time series to construct a risk-indicator association dataset; perform time series pattern mining on the risk-indicator association dataset to identify risk warning features and form an early warning indicator library. Based on the aforementioned early warning indicator library, a multi-level early warning threshold calculation model is defined to generate differentiated thresholds for different types of components and suppliers. A Bayesian network early warning model is constructed, and combined with the comprehensive risk scoring system, the risk probability and early warning level are obtained, i.e., the early warning result. A supply chain risk classification system is established, and historical risk events and their corresponding countermeasures are stored in a structured manner to form a historical response strategy library. Based on the historical response strategy library, a case-based reasoning method is used to extract the association rules between risk scenarios and response measures. Using the early warning results and combining critical path analysis of the risk transmission diagram, the current risk situation feature vector is identified. Based on the risk situation feature vector, a risk response decision tree is constructed using the C4.5 algorithm, mapping risk attributes to a set of response strategies. The risk transmission diagram is analyzed to identify the risk propagation path and impact nodes, and the response strategy set in the risk response decision tree is prioritized. A final risk response plan is formed. The risk response plan is transformed into specific execution tasks, generating intelligent risk response suggestions that include action steps, responsible persons, and time nodes.

8. An AI-based supply chain risk early warning device for automotive parts companies, characterized in that: The AI-based automotive parts enterprise supply chain risk early warning device includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the AI-based automotive parts enterprise supply chain risk warning device to execute the AI-based automotive parts enterprise supply chain risk warning method as described in any one of claims 1-7.

9. A computer-readable storage medium storing instructions thereon, characterized in that, When the instruction is executed by the processor, it implements the AI-based supply chain risk warning method for automotive parts companies as described in any one of claims 1-7.

Images