Supply chain quality early warning method and system based on AI

Abstract

The invention discloses an AI-based supply chain quality early warning method and system, and the method comprises the steps: collecting supply chain data and feedback data of a preset automobile bearing, and carrying out the preprocessing of the supply chain data and the feedback data; performing feature extraction and derivative feature construction according to the supply chain data by adopting feature engineering to obtain supply chain features, and performing adaptive quality anomaly recognition and classification according to the feedback data to obtain classification anomaly rules; using a graph neural network to carry out supply chain correlation analysis on the supply chain features to obtain a supply chain heterogeneous graph, and according to the classification anomaly rule, carrying out anomaly labeling and risk propagation analysis on the supply chain heterogeneous graph to obtain quality early warning data; and constructing a supply chain quality early warning model according to the quality early warning data, inputting data to be subjected to early warning into the supply chain quality early warning model, and outputting a quality early warning result.

Description

Technical Field

[0001] This invention relates to the field of supply chain quality early warning technology, and in particular to an AI-based supply chain quality early warning method and system. Background Technology

[0002] As the automotive industry upgrades towards intelligence and high-end products, automotive bearings, as core transmission components, directly affect the safety and operational stability of the entire vehicle. The bearing supply chain encompasses multiple stages, including raw material procurement, multi-process production, cross-regional logistics, warehousing management, and end-user feedback. It involves heterogeneous data from multiple sources, such as raw material batches, production parameters, testing data, and rework records, resulting in a massive volume of data with complex interrelationships.

[0003] Traditional supply chain quality early warning methods have significant limitations: on the one hand, they rely on fixed thresholds to identify anomalies, making it difficult to adapt to dynamic scenarios such as fluctuations in production processes and changes in raw material characteristics, leading to missed or false alarms; on the other hand, they lack in-depth analysis of the relationships between various links in the supply chain, failing to identify the propagation paths of risks across multiple nodes, resulting in strong early warning lag, often only responding passively after quality problems have occurred. Furthermore, traditional methods are insufficient in processing high-dimensional feedback data, struggling to integrate unstructured information such as customer complaints and industry recalls, and unable to achieve accurate prediction of quality risks across the entire chain. These problems make it difficult to detect potential quality hazards in the bearing supply chain in advance, potentially leading not only to direct losses such as product rework and customer complaints, but also, in severe cases, industry recalls, damaging the company's brand reputation.

[0004] Therefore, there is an urgent need for a supply chain quality early warning method that integrates AI technology, can integrate multi-source data, dynamically identify anomalies, trace the spread of risks, and accurately output early warnings, so as to meet the high-quality management and control needs of the bearing supply chain and improve the timeliness and accuracy of early warnings. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-based supply chain quality early warning method.

[0006] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: The system collects and preprocesses supply chain data and feedback data for pre-defined automotive bearings. The supply chain data includes raw material batch information, production process parameters, online testing data, and flow and warehousing data. The feedback data includes repair records, customer complaints, failure analysis reports, third-party testing reports, supplier credit ratings, anomaly records, and industry recall events. Feature engineering is used to extract and derive features from the supply chain data to obtain supply chain features, and adaptive quality anomaly identification and classification are performed based on the feedback data to obtain classification anomaly rules; A graph neural network is used to perform supply chain correlation analysis on the supply chain features to obtain a supply chain heterogeneity graph. Anomaly labeling and risk propagation analysis are performed on the supply chain heterogeneity graph according to the classification anomaly rules to obtain quality early warning data. A supply chain quality early warning model is constructed based on the quality early warning data. The data to be warned is input into the supply chain quality early warning model, and the quality early warning result is output.

[0007] Furthermore, a method for obtaining supply chain features by using feature engineering to extract and derive features from the supply chain data includes: Basic features are obtained by extracting features from raw material batch information, production process parameters, online detection data, and flow and storage data. These basic features include material composition features, supplier quality features, inspection index features, equipment status features, process parameter features, process efficiency features, dimensional accuracy features, surface quality features, performance features, storage environment features, and turnover efficiency features. Based on fundamental features and the mechanism of bearing quality impact, derived features are constructed: The trend of parameter changes over time is captured by the slope of parameter trends and the fluctuation of sliding windows to identify potential drift and obtain time-series trend features; the interaction between different processes or parameters is reflected by the temperature-hardness correlation ratio and process time matching degree to obtain process correlation features; the cumulative effect of anomalies in multiple stages is quantified by the cumulative number of abnormal parameters and the supplier risk index to obtain quality risk accumulation features; data from multiple stages are integrated based on material and process matching degree and storage and performance degradation coefficients to construct a global quality perspective and obtain cross-data source fusion features; and the deviation from industry standards is quantified by the standard deviation rate and process parameter compliance rate to obtain industry standard deviation features.

[0008] Furthermore, the method for adaptive quality anomaly identification and classification to obtain classification anomaly rules based on the feedback data includes: Logarithmic feedback data is used to calculate dynamic thresholds through a sliding window:

[0009] in Let i be the abnormal threshold for the i-th type of indicator. Let be the average of the i-th indicator across multiple batches. Let be the standard deviation of the i-th indicator. For dynamic adjustment coefficients; When there are no abnormalities in three consecutive batches, the dynamic adjustment coefficient decreases by 0.1; when an abnormality occurs once, the dynamic adjustment coefficient increases by 0.2. The incremental isolated forest algorithm is used for high-dimensional feedback data. The feedback data without anomalies is used as the initial training set. The incremental update is performed using a sliding window. When the sample anomaly score is greater than the anomaly threshold, it is judged as an anomaly and is used as incremental anomaly data. Abnormal data is obtained by triggering anomalies in incremental abnormal data based on preset rules. The preset rules are as follows: when the equipment is down for more than 2 hours, or the parameter fluctuation of the same equipment in 3 consecutive batches is greater than 5%, it is considered an equipment abnormality; when the number of times the test data exceeds the tolerance is greater than 3 times / day, or the same test item exceeds the tolerance in 2 consecutive batches, it is considered a test abnormality; when the transportation temperature exceeds the transportation temperature range for more than 4 hours, or the storage period is greater than 60 days, it is considered a logistics abnormality. Feature extraction is performed on abnormal record data to obtain core features. Based on the bearing quality impact mechanism, a three-level classification system of anomaly source, failure mode, and risk level is constructed. The core features are mapped to anomaly labels to obtain three-level classification labels. Anomaly source: material anomaly, process anomaly, logistics anomaly, and design anomaly, Level 1; therefore, material anomaly is further subdivided into component exceeding standard, inclusion level exceeding standard, and insufficient hardness; process anomaly is further subdivided into dimensional deviation, surface defects, and assembly errors; logistics anomaly is further subdivided into temperature and humidity exceeding standard and delay exceeding time limit; design anomaly is further subdivided into unreasonable structure and parameter matching error, Level 2; Risk level: based on the scope of anomaly impact, severity, and repairability quantification, Level 3; When the discrepancy between the manual classification result and the model classification result of a certain type of anomaly exceeds 15%, a rule update is triggered: the classification rules are dynamically optimized by combining reinforcement learning and expert feedback; manual correction samples from the past 3 months are collected, the classifier is retrained using the XGBoost algorithm, the feature weights are adjusted, and the new rules are used to backtest the historical data; an anomaly classification calibration committee is established to review the classification results of high-risk anomalies every quarter, adjust the classification dimensions, and revise the risk level judgment threshold. The three-level classification labels, feedback data annotation classification results, and record rule adjustment time, reason and effect are output as classification anomaly rules.

[0010] Furthermore, a method for obtaining a supply chain heterogeneity graph by performing supply chain correlation analysis on the aforementioned supply chain characteristics using graph neural networks includes: The supply chain entities are abstracted into multiple types of nodes, each type of node is associated with corresponding features, and multiple types of directed edges are defined according to business logic to construct a graph neural network. Independent embedding layers are used to obtain embedding vectors for multiple types of nodes. Key features are filtered through mutual information, and the embedding vectors are Euclidean normalized. Independent attention is assigned to directed edges of multiple types, and attention scores for node pairs are calculated.

[0011] in Let be the embedding vector of the b-th node. Let be the embedding vector of the z-th node. Let r be the weight matrix of the r-th edge type. For transpose, Let b be the attention score for the b-th node and z-th node. For embedded dimensions, The bias for the r-th edge type; The attention weights are adjusted by combining historical quality events. A random walk and depth-first strategy is used to sample the 2-hop neighborhood of each node. The degree centrality and betweenness centrality of the nodes are calculated. When the feature similarity between two nodes is greater than the similarity threshold and there is no direct edge in history, potential related edges are automatically generated. The reconstruction probability of normal nodes is learned by graph autoencoder. When the reconstruction error of a node is greater than the error threshold, it is marked as an abnormal node. The abnormal node’s own characteristics and neighborhood anomalies are combined to comprehensively determine the anomaly probability. Starting from the abnormal node, risk propagation paths are searched along edge type priority, the propagation probability of each edge is calculated, and the paths are sorted in descending order by risk value. The top three critical paths are taken as propagation paths. The propagation probability of each edge is the edge weight × historical propagation frequency. Obtain the node anomaly probability, the mean risk of the neighborhood, and historical risk records. Use the weighted sum of the node anomaly probability, the mean risk of the neighborhood, and the historical risk records as the node risk index. Output a supply chain heterogeneous graph based on the node risk index, the propagation path, and the graph neural network.

[0012] Furthermore, the method for anomaly labeling and risk propagation analysis of the supply chain heterogeneity diagram based on the aforementioned anomaly classification rules includes: Anomaly status is determined for each node in the heterogeneous graph based on classification anomaly rules. All nodes in the heterogeneous graph are traversed, node feature values ​​are extracted, and the feature values ​​are compared with the threshold in the classification anomaly rules to determine the anomaly level of each node and add an anomaly label. Anomaly status includes normal, slight anomaly, and severe anomaly. The abnormal propagation capability of the relationship between nodes is quantified to obtain the edge anomaly strength, and the edge weight is dynamically adjusted based on the edge attributes and node anomaly state. The attention scores of adjacent nodes are used as the influence weights of adjacent nodes. The adjacent nodes with the greatest influence on the current node are selected by the influence weights, and the nodes are spatially aggregated according to the influence weights to obtain the spatial aggregation results. The LSTM operator is used to capture the dynamic changes of risk over time in the graph structure at different time snapshots. A time-aggregation result is obtained by integrating information from multiple time points through a time attention mechanism. The expression is as follows:

[0013] in This is the spatial aggregation result at time t. Let be the hidden state of the node at time t. Let t-1 be the hidden state of the node. For LSTM operators; Based on the temporal and spatial aggregation results, a breadth-first search is used to trace the upstream propagation path of abnormal nodes, quantifying the probability of risk propagating from the source node to the target node. The risk propagation probability is obtained by dynamically adjusting the edge risk intensity and path length, expressed as:

[0014] in Let be the edge anomaly strength between node u and its adjacent node c. This is the path length attenuation factor. The number of path edges. For path The probability of risk transmission; The warning level is determined and output based on a comprehensive assessment of the propagation path probability and the anomaly level of the target node.

[0015] Furthermore, the method for constructing a supply chain quality early warning model based on the aforementioned quality early warning data includes: A supply chain quality early warning model is constructed based on graph neural networks for anomaly localization, risk propagation, and early warning output, integrating heterogeneous supply chain graphs and dynamic quality rules; the supply chain quality early warning includes a graph construction layer, a feature fusion layer, a risk reasoning layer, and an early warning output layer; The graph construction layer includes node definition and attribute mapping, edge relationship construction, and heterogeneous graph representation. Node definition and attribute mapping: Supply chain entities are abstracted into multiple types of core nodes, where node attributes are derived from basic features and derived features. Multiple types of directed edges are defined based on supply chain business logic, and edge weights are calculated through a dynamic weighting algorithm. An adjacency list is used to store the graph structure, which is represented as a set of nodes, a set of edges, and a node feature matrix. The feature fusion layer includes a node embedding layer and a temporal feature fusion layer. The node embedding layer uses an improved two-way graph embedding algorithm to map nodes into low-dimensional vectors. The temporal feature fusion layer introduces GraphLSTM units to fuse temporal dimension information to obtain fused temporal features. The risk reasoning layer includes anomaly node identification, risk propagation analysis, and root cause localization; anomaly node identification is performed based on an anomaly detection mechanism based on reconstruction error; risk propagation analysis uses a dynamic message passing mechanism to simulate the anomaly diffusion path; root cause localization: key risk nodes are identified through an attention mechanism. The early warning output layer includes the calculation of early warning indicators and the classification of early warning levels; the early warning indicators include the spread rate, the scope of impact, and the expected losses.

[0016] Secondly, an AI-based supply chain quality early warning system includes: Data acquisition and processing module: used to collect preset automotive bearing supply chain data and feedback data, and preprocess the supply chain data and feedback data; the supply chain data includes raw material batch information, production process parameters, online testing data, and flow and warehousing data; the feedback data includes repair records, customer complaints, failure analysis reports, third-party testing reports, supplier credit ratings, anomaly record information, and industry recall events; Feature extraction and anomaly identification and classification module: used to extract and construct features from the supply chain data using feature engineering to obtain supply chain features, and to perform adaptive quality anomaly identification and classification based on the feedback data to obtain classification anomaly rules; The correlation analysis and propagation analysis module is used to perform supply chain correlation analysis on the supply chain features using a graph neural network to obtain a supply chain heterogeneity graph, and to perform anomaly labeling and risk propagation analysis on the supply chain heterogeneity graph according to the classification anomaly rules to obtain quality early warning data. Model building and output module: used to build a supply chain quality early warning model based on the quality early warning data, input the data to be warned into the supply chain quality early warning model, and output the quality early warning results.

[0017] The beneficial effects of this invention are: This invention is an AI-based supply chain quality early warning method and system. Compared with existing technologies, this invention has the following technical advantages: This invention integrates multi-source heterogeneous data from the supply chain, constructs full-dimensional features through feature engineering to break down data silos; adopts dynamic thresholds and incremental algorithms to adaptively identify anomalies, solving the problem of poor adaptability of fixed thresholds; uses graph neural networks to mine node associations and accurately trace the risk propagation path; combines reinforcement learning and expert feedback to dynamically optimize rules and improve classification accuracy; realizes anomaly location, root cause analysis and hierarchical early warning, predicts risks in advance, overcomes the shortcomings of traditional early warnings such as lag and lack of global perspective, and significantly improves the efficiency and accuracy of supply chain quality control. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of an AI-based supply chain quality early warning method according to the present invention. Detailed Implementation

[0019] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0020] The present invention provides an AI-based supply chain quality early warning method and system, comprising the following steps: like Figure 1 As shown, this embodiment includes the following steps: The system collects and preprocesses supply chain data and feedback data for pre-defined automotive bearings. The supply chain data includes raw material batch information, production process parameters, online testing data, and flow and warehousing data. The feedback data includes repair records, customer complaints, failure analysis reports, third-party testing reports, supplier credit ratings, anomaly records, and industry recall events. In the actual assessment, the supply chain data included: Raw material batch information: GCr15 bearing steel provided by supplier S003, batch B20240512, carbon content 0.98%, chromium content 1.45%, inclusion grade 2; Production process parameters: quenching temperature 850℃, holding time 90min, stamping pressure 120MPa, grinding speed 3000r / min; Online inspection data: inner diameter 50.003mm (standard 50±0.005mm), outer diameter 62.002mm (standard 62±0.008mm), surface roughness Ra0.025μm; Logistics and warehousing data: transportation temperature 18-22℃ (duration 24 hours), warehousing environment temperature 20℃, humidity 45%, storage time 15 days; Feedback Data: Repair Records: A total of 5 units of the same model were repaired in the previous 3 batches, including 3 units with dimensional deviations and 2 units with surface defects; Customer Complaints: In April 2024, customer C002 complained that the bearing made abnormal noise after 3 months of use (related to batch B20240308); Third-Party Testing Report: The pass rate of key indicators for batch B20240308 was 97.2%, and the proportion of surface hardness failures was 1.8%; Supplier Credit Rating: Supplier S003 has a credit rating of A for the past 6 months, with a historical pass rate of 98.5%; Industry Recall Event: In March 2024, a competitor recalled 1,000 vehicles due to excessive bearing material composition; Data Preprocessing: 1. Data Cleaning: Two duplicate records were deleted. The missing chromium content value of a certain batch of supplier S003 was filled using the KNN algorithm (based on the chromium content of 1.43%, 1.45%, and 1.46% of the three adjacent batches, it was filled to 1.45%). The abnormal stamping pressure value of 150MPa (interquartile range 30MPa, upper limit 145MPa) was identified by the IQR rule and corrected to 145MPa. 2. Data Alignment: Based on batch number B20240512, the raw material data → quenching process parameters → inner diameter detection results → storage records were linked to form a complete data chain. Feature engineering is used to extract and derive features from the supply chain data to obtain supply chain features, and adaptive quality anomaly identification and classification are performed based on the feedback data to obtain classification anomaly rules; A graph neural network is used to perform supply chain correlation analysis on the supply chain features to obtain a supply chain heterogeneity graph. Anomaly labeling and risk propagation analysis are performed on the supply chain heterogeneity graph according to the classification anomaly rules to obtain quality early warning data. A supply chain quality early warning model is constructed based on the quality early warning data. The data to be warned is input into the supply chain quality early warning model, and the quality early warning result is output.

[0021] In this embodiment, a method for obtaining supply chain features by feature engineering based on the supply chain data through feature extraction and derived feature construction includes: Basic features are obtained by extracting features from raw material batch information, production process parameters, online detection data, and flow and storage data. These basic features include material composition features, supplier quality features, inspection index features, equipment status features, process parameter features, process efficiency features, dimensional accuracy features, surface quality features, performance features, storage environment features, and turnover efficiency features. Based on fundamental features and the mechanism of bearing quality impact, derived features are constructed: The trend of parameter changes over time is captured by the slope of parameter trends and the fluctuation of sliding windows to identify potential drift and obtain time-series trend features; the interaction between different processes or parameters is reflected by the temperature-hardness correlation ratio and process time matching degree to obtain process correlation features; the cumulative effect of anomalies in multiple stages is quantified by the cumulative number of abnormal parameters and the supplier risk index to obtain quality risk accumulation features; data from multiple stages are integrated based on material and process matching degree and storage and performance degradation coefficients to construct a global quality perspective and obtain cross-data source fusion features; and the degree of deviation from industry standards is quantified by the standard deviation rate and process parameter compliance rate to obtain industry standard deviation features. In actual evaluation, preprocessing includes: 1. Data cleaning: deleting duplicate records, filling missing values ​​with the KNN algorithm, and identifying and correcting outliers using the IQR rule; 2. Data alignment: associating multi-source data based on batch number / timestamp (a batch of raw materials → corresponding production process parameters → online test results → storage records). Five batches are used as windows; the parameter trend slope is calculated by linear regression of the stamping temperature data of 30 consecutive batches of a certain equipment; the sliding window fluctuation is calculated by the standard deviation of the dimensional deviation within the window of five batches; the temperature-hardness correlation ratio is (actual quenching temperature / standard quenching temperature) × (actual hardness / standard hardness); the process time matching degree is the ratio of the processing time of the preceding process to the processing time of the following process; the cumulative number of abnormal parameters is the number of parameters exceeding the threshold in the raw materials, production, and testing stages of a certain batch; the supplier risk index is a comprehensive score calculated based on the supplier's historical pass rate and the number of complaints in the past 3 months; the material and process matching degree is the raw material carbon content × quenching temperature / standard carbon content × standard temperature; the storage and performance degradation coefficient is the storage residence time × storage temperature fluctuation × surface roughness; the standard tolerance deviation rate is (measured size - standard size) / standard tolerance band width; the process parameter compliance rate is the number of parameters that meet the process requirements / the total number of parameters. Basic characteristics: Material composition: Carbon content 0.98%, Chromium content 1.45%, Inclusion grade 2; Supplier quality characteristics: Supplier S003 pass rate 98.5%, Credit rating A; Inspection indicators: Inner diameter 50.003mm, outer diameter 62.002mm, surface roughness Ra 0.025μm; Equipment status characteristics: Stamping equipment M005 has been running for 800 hours with no fault records; Storage environment characteristics: Storage temperature 20℃, humidity 45%, storage time 15 days; Derived characteristics: Time series trend characteristics: Taking the stamping temperatures of 30 consecutive batches from equipment M005 (845℃, 847℃, ..., 850℃), the trend slope was calculated using linear regression, with a slope of 0.15℃ / batch; Using 5 batches as a window, the standard deviation of the sliding window fluctuation of dimensional deviations (0.001mm, 0.002mm, 0.003mm, 0.002mm, 0.001mm) was 0.0007mm; Process correlation characteristics: Temperature-hardness correlation ratio (850 / 840) × (62HRC / 60HRC) = 1.06; Process time matching degree (quenching 90min / grinding 60min) = 1.5; Quality risk accumulation characteristics. The cumulative number of abnormal parameters in this batch is 2 (stamping pressure exceeded the threshold, and the inner diameter deviation is close to the upper limit); the supplier's S003 risk index is 0.8 (calculated based on a historical pass rate of 98.5% and 1 complaint in the past 3 months); cross-data source fusion characteristics: material and process matching degree (0.98%×850℃) / (1.0%×840℃)=0.99; storage and performance degradation coefficient (15 days×2℃ fluctuation×0.025μm)=0.75; industry standard deviation characteristics: inner diameter standard deviation deviation rate (50.003-50) / 0.005=0.6; process parameter compliance rate (4 / 5)=80% (stamping pressure exceeded the threshold, and the other 4 items are compliant).

[0022] In this embodiment, the method for adaptive quality anomaly identification and classification to obtain classification anomaly rules based on the feedback data includes: Logarithmic feedback data is used to calculate dynamic thresholds through a sliding window:

[0023] in Let i be the abnormal threshold for the i-th type of indicator. Let be the average of the i-th indicator across multiple batches. Let be the standard deviation of the i-th indicator. For dynamic adjustment coefficients; When there are no abnormalities in three consecutive batches, the dynamic adjustment coefficient decreases by 0.1; when an abnormality occurs once, the dynamic adjustment coefficient increases by 0.2. The incremental isolated forest algorithm is used for high-dimensional feedback data. The feedback data without anomalies is used as the initial training set. The incremental update is performed using a sliding window. When the sample anomaly score is greater than the anomaly threshold, it is judged as an anomaly and is used as incremental anomaly data. Abnormal data is obtained by triggering anomalies in incremental abnormal data based on preset rules. The preset rules are as follows: when the equipment is down for more than 2 hours, or the parameter fluctuation of the same equipment in 3 consecutive batches is greater than 5%, it is considered an equipment abnormality; when the number of times the test data exceeds the tolerance is greater than 3 times / day, or the same test item exceeds the tolerance in 2 consecutive batches, it is considered a test abnormality; when the transportation temperature exceeds the transportation temperature range for more than 4 hours, or the storage period is greater than 60 days, it is considered a logistics abnormality. Feature extraction is performed on abnormal record data to obtain core features. Based on the bearing quality impact mechanism, a three-level classification system of anomaly source, failure mode, and risk level is constructed. The core features are mapped to anomaly labels to obtain three-level classification labels. Anomaly sources are categorized into material anomalies, process anomalies, logistics anomalies, and design anomalies, at level one. Material anomalies are further subdivided into component exceedance, inclusion grade exceedance, and insufficient hardness. Process anomalies are subdivided into dimensional deviation, surface defects, and assembly errors. Logistics anomalies are subdivided into temperature and humidity exceedance and excessive storage time. Design anomalies are subdivided into unreasonable structure and parameter matching errors, at level two. Risk level is quantified based on the anomaly's impact range, severity, and repairability, at level three. When the discrepancy between the manual classification result and the model classification result of a certain type of anomaly exceeds 15%, a rule update is triggered: the classification rules are dynamically optimized by combining reinforcement learning and expert feedback; manual correction samples from the past 3 months are collected, the classifier is retrained using the XGBoost algorithm, the feature weights are adjusted, and the new rules are used to backtest historical data to ensure that the accuracy is improved by more than or equal to 5%; an anomaly classification calibration committee is established to review the classification results of high-risk anomalies every quarter, adjust the classification dimensions, and revise the risk level judgment threshold. The three-level classification labels, feedback data annotation classification results, and record rule adjustment time, reason and effect are output as classification anomaly rules; In actual evaluation, the preprocessing steps are as follows: 1. Data standardization: Textual data is converted into 768-dimensional vectors using the BERT model; numerical data is normalized to the [0,1] interval using min-max; categorical data is converted into binary vectors using One-Hot encoding; 2. Data association: Multi-source feedback data is associated based on the three elements of batch number, timestamp, and supplier ID. The sliding window has a window period of 30 days and an anomaly threshold of 0.7. Material anomalies include excessive composition, excessive inclusion level, and insufficient hardness. Process anomalies include out-of-tolerance dimensions, surface defects, and assembly errors. The risk levels are divided into: Level 1 is a single batch anomaly affecting <10 products; Level 2 is a single batch anomaly affecting 10-50 products; Level 3 is multiple batch anomalies affecting 50-200 products, or customer complaints at or above Level 3; Level 4 is anomalies affecting >200 products, or causing production line shutdowns for customers, or third-party testing failing to meet key indicators; Level 5 may trigger a recall, or the failure mode involves safety performance. Dynamic threshold calculation: Taking inner diameter deviation as the i-th type of indicator, the mean μ of multiple batches is 0.001mm, the standard deviation is 0.0005mm, and the initial dynamic adjustment coefficient is 2.0; since the first 3 batches had no abnormalities, k was adjusted to 1.9, and the abnormality threshold TR was calculated as TR = 0.001 + 1.9 × 0.0005 = 0.00195mm; Incremental anomaly identification: 1000 feedback data without abnormalities were selected as the initial training set, with a sliding window of 200 data points, and an anomaly score threshold of 0.7; the feedback data of batch B20240512 had an anomaly score of 0.68, which was judged as normal; Anomaly triggering: no equipment downtime exceeded 2 hours in this batch, and the inspection... The number of out-of-tolerance tests was 3 times / day, and no abnormal records were triggered; Level 3 classification label: Assuming that a certain batch B20240401 has a carbon content of 1.1% (exceeding the 0.95%-1.05% standard), it is marked as material abnormality - component exceeding the standard - risk level 3 (affecting 80 products, which can be repaired); 5. Rule update: In Q2 2024, the inconsistency rate between manual and model classification of process abnormality - dimensional out-of-tolerance category was 18% (exceeding 15%). 200 manually corrected samples were collected in the past 3 months, and the classifier was retrained through XGBoost. The dimensional deviation weight was adjusted from 0.3 to 0.4, and the backtracking verification accuracy improved by 7%.

[0024] In this embodiment, the method for obtaining a heterogeneous supply chain graph by performing supply chain correlation analysis on the supply chain features using a graph neural network includes: The supply chain entities are abstracted into multiple types of nodes, each associated with corresponding features. Multiple types of directed edges are defined according to business logic to construct a graph neural network. The multiple types of nodes include raw material batch nodes, production equipment nodes, testing data nodes, supplier nodes, and customer nodes. The multiple types of directed edges include supply relationship edges, processing relationship edges, testing relationship edges, storage relationship edges, complaint relationship edges, collaboration relationship edges, and rectification relationship edges. The edge weights are initialized to the historical interaction frequency. Independent embedding layers are used to obtain embedding vectors for multiple types of nodes. Key features are filtered through mutual information, and the embedding vectors are Euclidean normalized. Independent attention is assigned to directed edges of multiple types, and attention scores for node pairs are calculated.

[0025] in Let be the embedding vector of the b-th node. Let be the embedding vector of the z-th node. Let r be the weight matrix of the r-th edge type. For transpose, Let b be the attention score for the b-th node and z-th node. For embedded dimensions, The bias for the r-th edge type; The attention weights are adjusted by combining historical quality events. A random walk and depth-first strategy is used to sample the 2-hop neighborhood of each node. The degree centrality and betweenness centrality of the nodes are calculated. When the feature similarity between two nodes is greater than the similarity threshold and there is no direct edge in history, potential related edges are automatically generated. The reconstruction probability of normal nodes is learned by graph autoencoder. When the reconstruction error of a node is greater than the error threshold, it is marked as an abnormal node. The abnormal node’s own characteristics and neighborhood anomalies are combined to comprehensively determine the anomaly probability. Starting from the abnormal node, risk propagation paths are searched along edge type priority, the propagation probability of each edge is calculated, and the paths are sorted in descending order by risk value. The top three critical paths are taken as propagation paths. The propagation probability of each edge is the edge weight × historical propagation frequency. Obtain the node anomaly probability, the mean risk of the neighborhood, and historical risk records. Use the weighted sum of the node anomaly probability, the mean risk of the neighborhood, and the historical risk records as the node risk index. Output a supply chain heterogeneous graph based on the node risk index, the propagation path, and the graph neural network. In actual evaluation, key features are screened using mutual information: features strongly correlated with quality anomalies are retained; the similarity threshold is 0.85; and the error threshold is 0.15. Adjust attention weights based on historical quality events: If supplier A has caused a quality incident due to material issues, the weight of supplier A's supply relationship edge will increase by 20% in the following 6 months; Node and edge definitions: Abstract nodes include raw material batch B20240512, production equipment M005, inspection node T012, supplier S003, and warehouse node W004; directed edges include B20240512→M005 (supply relationship, weight 0.8), M005→T012 (processing relationship, weight 0.9), and T012→W004 (inspection relationship, weight 0.7). Attention score calculation: Let the embedding vector of B20240512 be h_b=[0.1,0.2,0.3], the embedding vector of M005 be h_z=[0.2,0.3,0.4], the edge type weight matrix be [[0.1,0.2],[0.3,0.4],[0.5,0.6]], the bias be [0.1], and the embedding dimension be 3; the attention score of the node pair is 0.65; Abnormal node identification: The reconstruction probability of normal nodes is learned through graph autoencoder. The reconstruction error of batch B20240401 (excessive carbon content) is 0.19>0.15, so it is marked as an abnormal node with an abnormal probability of 0.85. Risk propagation path: Starting from B20240401, the search follows the edge priority. The top 3 critical paths are B20240401→M005→T012→W004, B20240401→S003→B20240402→M005, and B20240401→W004→Customer C001. The anomaly probability of B20240401 is 0.85, the average neighborhood risk is 0.6, the historical risk is 0.7, and the node risk index is 0.735.

[0026] In this embodiment, the method for anomaly labeling and risk propagation analysis of the supply chain heterogeneity diagram based on the classification anomaly rules includes: Anomaly status is determined for each node in the heterogeneous graph based on classification anomaly rules. All nodes in the heterogeneous graph are traversed, node feature values ​​are extracted, and the feature values ​​are compared with the threshold in the classification anomaly rules to determine the anomaly level of each node and add an anomaly label. Anomaly status includes normal, slight anomaly, and severe anomaly. The anomaly propagation capability of the relationships between nodes is quantified to obtain the edge anomaly strength. Edge weights are dynamically adjusted based on edge attributes and node anomaly states. The calculation of edge anomaly strength is then performed. :

[0027] in The impact of node anomalies on weights, The anomaly level of node u. These are the standardized values ​​of the edge attributes between source node u and node c; The attention scores of adjacent nodes are used as their influence weights. These influence weights are then used to select the adjacent nodes that have the greatest impact on the current node. Spatial aggregation is then performed on the nodes based on their influence weights to obtain the spatial aggregation result, expressed as:

[0028] in It is the sigmoid activation function. Let u be the influence weight of its neighboring node c. , The weight matrix is ​​a learnable matrix. For learnable bias vectors, Let u be the feature vector of node u. Let be the feature vector of the adjacent node c; The LSTM operator is used to capture the dynamic changes of risk over time in the graph structure at different time snapshots. A time-aggregation result is obtained by integrating information from multiple time points through a time attention mechanism. The expression is as follows:

[0029] in This is the spatial aggregation result at time t. Let be the hidden state of the node at time t. Let t-1 be the hidden state of the node. For LSTM operators; Based on the temporal and spatial aggregation results, a breadth-first search is used to trace the upstream propagation path of abnormal nodes, quantifying the probability of risk propagating from the source node to the target node. The risk propagation probability is obtained by dynamically adjusting the edge risk intensity and path length, expressed as:

[0030] in Let be the edge anomaly strength between node u and its adjacent node c. This is the path length attenuation factor. The number of path edges. For path The probability of risk transmission; The warning level is determined and output based on a comprehensive assessment of the propagation path probability and the anomaly level of the target node; In actual assessments, the anomaly criteria for raw material nodes are as follows: carbon content outside the range of 0.95%-1.05% is considered a minor anomaly when it deviates by 10%-20% and a severe anomaly when it deviates by more than 20%. For production equipment nodes, the anomaly criteria are: stamping temperature fluctuation coefficient greater than 0.05 is considered a minor anomaly when the fluctuation coefficient is between 0.05 and 0.08 and a severe anomaly when it is greater than 0.08. For inspection report nodes, the anomaly criteria are: absolute value of inner diameter deviation greater than 0.01 mm is considered a minor anomaly when the inner diameter deviation is between 0.01 and 0.015 mm and a severe anomaly when it is greater than 0.015 mm. A red alert is issued when the risk propagation probability is greater than 1.2 and the target node is classified as severely abnormal; a yellow alert is issued when the risk propagation probability is greater than 0.8 and the target node is classified as slightly abnormal; and a blue alert is issued when the risk propagation probability is less than or equal to 0.8 and the target node is classified as normal but with potential risks. Anomaly labeling: B20240401 has a carbon content of 1.1% (exceeding the standard by 20%), and is marked as a serious anomaly; M005 is marked as a minor anomaly due to a processing defect. Edge anomaly strength: The weight of node anomaly impact is 0.6, the anomaly level of B20240401 is 3 (severe anomaly), the standardized value is 0.8, and the edge anomaly strength is 2.12; Spatiotemporal aggregation: The spatial aggregation result at time t is 0.7 (based on the influence weight of adjacent nodes), the hidden state at time t-1 is 0.6, and the node hidden state at time t is 0.65; Risk propagation probability: The EdgeR values ​​for path B20240401→M005→T012 are 2.12 and 1.8 respectively, the path length is 2, the path length decay factor for path B20240401→M005→T012 is 0.6703, and the risk propagation probability is 2.12×1.8×0.6703≈2.56; Warning level: If the propagation probability 2.56>1.2 and the target node T012 is severely abnormal, a red warning is output.

[0031] In this embodiment, the method for constructing a supply chain quality early warning model based on the quality early warning data includes: A supply chain quality early warning model is constructed based on graph neural networks for anomaly localization, risk propagation, and early warning output, integrating heterogeneous supply chain graphs and dynamic quality rules; the supply chain quality early warning includes a graph construction layer, a feature fusion layer, a risk reasoning layer, and an early warning output layer; The graph construction layer includes node definition and attribute mapping, edge relationship construction, and heterogeneous graph representation. Node definition and attribute mapping: Supply chain entities are abstracted into multiple types of core nodes, where node attributes are derived from basic features and derived features. Multiple types of directed edges are defined based on supply chain business logic, and edge weights are calculated through a dynamic weighting algorithm. An adjacency list is used to store the graph structure, which is represented as a set of nodes, a set of edges, and a node feature matrix. The feature fusion layer includes a node embedding layer and a temporal feature fusion layer. The node embedding layer uses an improved two-way graph embedding algorithm to map nodes into low-dimensional vectors. The temporal feature fusion layer introduces GraphLSTM units to fuse temporal dimension information to obtain fused temporal features. The risk reasoning layer includes anomaly node identification, risk propagation analysis, and root cause localization; anomaly node identification is performed based on an anomaly detection mechanism based on reconstruction error; risk propagation analysis uses a dynamic message passing mechanism to simulate the anomaly diffusion path; root cause localization: key risk nodes are identified through an attention mechanism. The early warning output layer includes the calculation of early warning indicators and the classification of early warning levels; the early warning indicators include the spread rate, the scope of impact, and the expected losses. In actual assessment, multiple core nodes include raw material nodes, production equipment nodes, testing data nodes, logistics and warehousing nodes, supplier nodes, and product batch nodes; multiple directed edges include supply relationships, production relationships, testing relationships, warehousing relationships, and feedback relationships. To obtain fused temporal features, a GraphLSTM unit is introduced to fuse temporal dimension information: GCN and LSTM are combined, and a graph convolutional layer is added before the traditional LSTM input. Training is performed on nearly 30 batches of node embedding sequences to capture parameter drift trends. The expression is as follows:

[0032] in It is an adjacency matrix. Let t be the LSTM hidden state. The fused temporal features at time t; Abnormal node identification: Construct a Seq2Seq+VAE model, reconstruct the node embedding sequence into a feature distribution under normal conditions, calculate the global abnormal score and component abnormal score, and trigger an alert when the global abnormal score is greater than 0.8 or any component abnormal score is greater than 0.9. Risk propagation analysis: Initial risk is assigned to nodes, and multiple rounds of message passing are performed on the nodes. Based on the stable risk value after multiple rounds of propagation, the nodes are divided into high risk (risk value greater than 0.7), medium risk (risk value between 0.3 and 0.7), and low risk (risk value less than 0.3). The diffusion rate is the ratio of risk value to duration; the scope of impact is the proportion of affected nodes to the total number of nodes; the estimated loss is calculated based on historical failure cost data. Graph construction layer: Node feature matrices are B20240512[0.98,0.8,0.003], M005[0.9,0.7,0.002], T012[0.85,0.6,0.001], and edge set weight matrix is ​​[[0.8,0,0],[0,0.9,0],[0,0,0.7]]. Feature fusion layer: B20240512 is mapped to a low-dimensional vector [0.15, 0.23, 0.31] using an improved bidirectional graph embedding algorithm; a GraphLSTM unit is introduced to fuse nearly 30 batches of node embedding sequences, resulting in a fused temporal feature of 0.35. Risk reasoning layer: The Seq2Seq+VAE model calculates the global anomaly score of B20240512 as 0.6 < 0.8, and the highest component anomaly score is 0.7 < 0.9, indicating no anomaly was triggered; root cause localization identifies S003 as a potential risk node (attention weight 0.4). Early warning output layer: Early warning indicators include a diffusion rate of 0.3 / 24≈0.0125, an impact range of 3 / 50=6%, and an estimated loss of 30 items × 500 yuan / item = 15,000 yuan; Input the data to be warned for batch B20240512 into the model and output the warning result: Blue warning (risk propagation probability 0.6 < 0.8, target node is normal but there is potential risk).

[0033] Secondly, an AI-based supply chain quality early warning system includes: Data acquisition and processing module: used to collect preset automotive bearing supply chain data and feedback data, and preprocess the supply chain data and feedback data; the supply chain data includes raw material batch information, production process parameters, online testing data, and flow and warehousing data; the feedback data includes repair records, customer complaints, failure analysis reports, third-party testing reports, supplier credit ratings, anomaly record information, and industry recall events; Feature extraction and anomaly identification and classification module: used to extract and construct features from the supply chain data using feature engineering to obtain supply chain features, and to perform adaptive quality anomaly identification and classification based on the feedback data to obtain classification anomaly rules; The correlation analysis and propagation analysis module is used to perform supply chain correlation analysis on the supply chain features using a graph neural network to obtain a supply chain heterogeneity graph, and to perform anomaly labeling and risk propagation analysis on the supply chain heterogeneity graph according to the classification anomaly rules to obtain quality early warning data. Model building and output module: This module is used to build a supply chain quality early warning model based on the quality early warning data, input the data to be warned into the supply chain quality early warning model, and output the quality early warning results. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI-based supply chain quality early warning method, characterized in that, Includes the following steps: The system collects and preprocesses supply chain data and feedback data for pre-defined automotive bearings. The supply chain data includes raw material batch information, production process parameters, online testing data, and flow and warehousing data. The feedback data includes repair records, customer complaints, failure analysis reports, third-party testing reports, supplier credit ratings, anomaly records, and industry recall events. Feature engineering is used to extract and derive features from the supply chain data to obtain supply chain features, and adaptive quality anomaly identification and classification are performed based on the feedback data to obtain classification anomaly rules; A graph neural network is used to perform supply chain correlation analysis on the supply chain features to obtain a supply chain heterogeneity graph. Anomaly labeling and risk propagation analysis are performed on the supply chain heterogeneity graph according to the classification anomaly rules to obtain quality early warning data. A supply chain quality early warning model is constructed based on the quality early warning data. The data to be warned is input into the supply chain quality early warning model, and the quality early warning result is output.

2. The AI-based supply chain quality early warning method according to claim 1, characterized in that, A method for obtaining supply chain features by using feature engineering to extract and derive features from the supply chain data includes: Basic features are obtained by extracting features from raw material batch information, production process parameters, online detection data, and flow and storage data. These basic features include material composition features, supplier quality features, inspection index features, equipment status features, process parameter features, process efficiency features, dimensional accuracy features, surface quality features, performance features, storage environment features, and turnover efficiency features. Based on fundamental features and the mechanism of bearing quality impact, derived features are constructed: The trend of parameter changes over time is captured by the slope of parameter trends and the fluctuation of sliding windows to identify potential drift and obtain time-series trend features; the interaction between different processes or parameters is reflected by the temperature-hardness correlation ratio and process time matching degree to obtain process correlation features; the cumulative effect of anomalies in multiple stages is quantified by the cumulative number of abnormal parameters and the supplier risk index to obtain quality risk accumulation features; data from multiple stages are integrated based on material and process matching degree and storage and performance degradation coefficients to construct a global quality perspective and obtain cross-data source fusion features; and the deviation from industry standards is quantified by the standard deviation rate and process parameter compliance rate to obtain industry standard deviation features.

3. The AI-based supply chain quality early warning method according to claim 1, characterized in that, A method for adaptive quality anomaly identification and classification to obtain classification anomaly rules based on the feedback data includes: Logarithmic feedback data is used to calculate dynamic thresholds through a sliding window: ； in Let i be the abnormal threshold for the i-th type of indicator. Let be the average of the i-th indicator across multiple batches. Let be the standard deviation of the i-th indicator. For dynamic adjustment coefficients; When there are no abnormalities in three consecutive batches, the dynamic adjustment coefficient decreases by 0.1; when an abnormality occurs once, the dynamic adjustment coefficient increases by 0.

2. The incremental isolated forest algorithm is used for high-dimensional feedback data. The feedback data without anomalies is used as the initial training set. The incremental update is performed using a sliding window. When the sample anomaly score is greater than the anomaly threshold, it is judged as an anomaly and is used as incremental anomaly data. Abnormal data is obtained by triggering anomalies in incremental abnormal data based on preset rules. The preset rules are as follows: when the equipment is down for more than 2 hours, or the parameter fluctuation of the same equipment in 3 consecutive batches is greater than 5%, it is considered an equipment abnormality; when the number of times the test data exceeds the tolerance is greater than 3 times / day, or the same test item exceeds the tolerance in 2 consecutive batches, it is considered a test abnormality; when the transportation temperature exceeds the transportation temperature range for more than 4 hours, or the storage period is greater than 60 days, it is considered a logistics abnormality. Feature extraction is performed on abnormal record data to obtain core features. Based on the bearing quality impact mechanism, a three-level classification system of anomaly source, failure mode, and risk level is constructed. The core features are mapped to anomaly labels to obtain three-level classification labels. Anomaly sources are categorized into material anomalies, process anomalies, logistics anomalies, and design anomalies, at level one. Material anomalies are further subdivided into component exceedance, inclusion grade exceedance, and insufficient hardness. Process anomalies are subdivided into dimensional deviation, surface defects, and assembly errors. Logistics anomalies are subdivided into temperature and humidity exceedance and excessive storage time. Design anomalies are subdivided into unreasonable structure and parameter matching errors, at level two. Risk level is quantified based on the anomaly's impact range, severity, and repairability, at level three. When the discrepancy between the manual classification result and the model classification result of a certain type of anomaly exceeds 15%, a rule update is triggered: the classification rules are dynamically optimized by combining reinforcement learning and expert feedback; manual correction samples from the past 3 months are collected, the classifier is retrained using the XGBoost algorithm, the feature weights are adjusted, and the new rules are used to backtest the historical data; an anomaly classification calibration committee is established to review the classification results of high-risk anomalies every quarter, adjust the classification dimensions, and revise the risk level judgment threshold. The three-level classification labels, feedback data annotation classification results, and record rule adjustment time, reason and effect are output as classification anomaly rules.

4. The AI-based supply chain quality early warning method according to claim 1, characterized in that, A method for obtaining a supply chain heterogeneity graph by performing supply chain correlation analysis on the aforementioned supply chain features using graph neural networks includes: The supply chain entities are abstracted into multiple types of nodes, each type of node is associated with corresponding features, and multiple types of directed edges are defined according to business logic to construct a graph neural network. Independent embedding layers are used to obtain embedding vectors for multiple types of nodes. Key features are filtered through mutual information, and the embedding vectors are Euclidean normalized. Independent attention is assigned to directed edges of multiple types, and attention scores for node pairs are calculated. ； in Let be the embedding vector of the b-th node. Let be the embedding vector of the z-th node. Let r be the weight matrix of the r-th edge type. For transpose, Let b be the attention score for the b-th node and z-th node. For embedded dimensions, The bias for the r-th edge type; The attention weights are adjusted by combining historical quality events. A random walk and depth-first strategy is used to sample the 2-hop neighborhood of each node. The degree centrality and betweenness centrality of the nodes are calculated. When the feature similarity between two nodes is greater than the similarity threshold and there is no direct edge in history, potential related edges are automatically generated. The reconstruction probability of normal nodes is learned by graph autoencoder. When the reconstruction error of a node is greater than the error threshold, it is marked as an abnormal node. The abnormal node’s own characteristics and neighborhood anomalies are combined to comprehensively determine the anomaly probability. Starting from the abnormal node, risk propagation paths are searched along edge type priority, the propagation probability of each edge is calculated, and the paths are sorted in descending order by risk value. The top three critical paths are taken as propagation paths. The propagation probability of each edge is the edge weight × historical propagation frequency. Obtain the node anomaly probability, the mean risk of the neighborhood, and historical risk records. Use the weighted sum of the node anomaly probability, the mean risk of the neighborhood, and the historical risk records as the node risk index. Output a supply chain heterogeneous graph based on the node risk index, the propagation path, and the graph neural network.

5. The AI-based supply chain quality early warning method according to claim 1, characterized in that, A method for anomaly labeling and risk propagation analysis of the supply chain heterogeneity diagram based on the aforementioned anomaly classification rules includes: Anomaly status is determined for each node in the heterogeneous graph based on classification anomaly rules. All nodes in the heterogeneous graph are traversed, node feature values ​​are extracted, and the feature values ​​are compared with the threshold in the classification anomaly rules to determine the anomaly level of each node and add an anomaly label. Anomaly status includes normal, slight anomaly, and severe anomaly. The abnormal propagation capability of the relationship between nodes is quantified to obtain the edge anomaly strength, and the edge weight is dynamically adjusted based on the edge attributes and node anomaly state. The attention scores of adjacent nodes are used as the influence weights of adjacent nodes. The adjacent nodes with the greatest influence on the current node are selected by the influence weights, and the nodes are spatially aggregated according to the influence weights to obtain the spatial aggregation results. The LSTM operator is used to capture the dynamic changes of risk over time in the graph structure at different time snapshots. A time-aggregation result is obtained by integrating information from multiple time points through a time attention mechanism. The expression is as follows: ； in This is the spatial aggregation result at time t. Let be the hidden state of the node at time t. Let t-1 be the hidden state of the node. For LSTM operators; Based on the temporal and spatial aggregation results, a breadth-first search is used to trace the upstream propagation path of abnormal nodes, quantifying the probability of risk propagating from the source node to the target node. The risk propagation probability is obtained by dynamically adjusting the edge risk intensity and path length, expressed as: ； in Let be the edge anomaly strength between node u and its adjacent node c. This is the path length attenuation factor. The number of path edges. For path The probability of risk transmission; The warning level is determined and output based on a comprehensive assessment of the propagation path probability and the anomaly level of the target node.

6. The AI-based supply chain quality early warning method according to claim 1, characterized in that, The method for constructing a supply chain quality early warning model based on the aforementioned quality early warning data includes: A supply chain quality early warning model is constructed based on graph neural networks for anomaly localization, risk propagation, and early warning output, integrating heterogeneous supply chain graphs and dynamic quality rules; the supply chain quality early warning includes a graph construction layer, a feature fusion layer, a risk reasoning layer, and an early warning output layer; The graph construction layer includes node definition and attribute mapping, edge relationship construction, and heterogeneous graph representation. Node definition and attribute mapping: Supply chain entities are abstracted into multiple types of core nodes, where node attributes are derived from basic features and derived features. Multiple types of directed edges are defined based on supply chain business logic, and edge weights are calculated through a dynamic weighting algorithm. An adjacency list is used to store the graph structure, which is represented as a set of nodes, a set of edges, and a node feature matrix. The feature fusion layer includes a node embedding layer and a temporal feature fusion layer. The node embedding layer uses an improved two-way graph embedding algorithm to map nodes into low-dimensional vectors. The temporal feature fusion layer introduces GraphLSTM units to fuse temporal dimension information to obtain fused temporal features. The risk reasoning layer includes anomaly node identification, risk propagation analysis, and root cause localization; anomaly node identification is performed based on an anomaly detection mechanism based on reconstruction error; risk propagation analysis uses a dynamic message passing mechanism to simulate the anomaly diffusion path; root cause localization: key risk nodes are identified through an attention mechanism. The early warning output layer includes the calculation of early warning indicators and the classification of early warning levels; the early warning indicators include the spread rate, the scope of impact, and the expected losses.

7. An AI-based supply chain quality early warning system, used to execute the method according to any one of claims 1-6, characterized in that, include: Data acquisition and processing module: used to collect supply chain data and feedback data of preset automotive bearings, and to preprocess the supply chain data and feedback data; The supply chain data includes raw material batch information, production process parameters, online testing data, and flow and warehousing data; the feedback data includes rework records, customer complaints, failure analysis reports, third-party testing reports, supplier credit ratings, anomaly record information, and industry recall events. Feature extraction and anomaly identification and classification module: used to extract and construct features from the supply chain data using feature engineering to obtain supply chain features, and to perform adaptive quality anomaly identification and classification based on the feedback data to obtain classification anomaly rules; The correlation analysis and propagation analysis module is used to perform supply chain correlation analysis on the supply chain features using a graph neural network to obtain a supply chain heterogeneity graph, and to perform anomaly labeling and risk propagation analysis on the supply chain heterogeneity graph according to the classification anomaly rules to obtain quality early warning data. Model building and output module: used to build a supply chain quality early warning model based on the quality early warning data, input the data to be warned into the supply chain quality early warning model, and output the quality early warning results.

Images