CN-121998509-A - AI-based supply chain quality early warning method and system

CN121998509ACN 121998509 ACN121998509 ACN 121998509ACN-121998509-A

Abstract

The invention discloses an AI-based supply chain quality early warning method and system, which comprise the steps of collecting supply chain data and feedback data of a preset automobile bearing, preprocessing the supply chain data and the feedback data, adopting feature engineering to perform feature extraction and derivative feature construction according to the supply chain data to obtain supply chain features, performing self-adaptive quality anomaly recognition and classification according to the feedback data to obtain a classification anomaly rule, adopting a graph neural network to perform supply chain association analysis on the supply chain features to obtain a supply chain abnormal graph, performing anomaly labeling and risk propagation analysis on the supply chain abnormal graph according to the classification anomaly rule to obtain quality early warning data, constructing a supply chain quality early warning model according to the quality early warning data, inputting data to be early warned into the supply chain quality early warning model, and outputting a quality early warning result.

Inventors

Lai Fanghua
WEI FANGFANG
ZHAO HAIGANG
YANG QINGPING
YAN HAIYANG

Assignees

万向钱潮股份公司

Dates

Publication Date: 20260508
Application Date: 20260206

Claims (7)

1. The AI-based supply chain quality early warning method is characterized by comprising the following steps: The method comprises the steps of collecting supply chain data and feedback data of a preset automobile bearing, and preprocessing the supply chain data and the feedback data, wherein the supply chain data comprises raw material batch information, production process parameters, online detection data, flow and storage data, and the feedback data comprises a repair record, a customer complaint, a failure analysis report, a third party detection report, a supplier credit rating, abnormal record information and an industry recall event; Performing feature extraction and derivative feature construction by adopting feature engineering according to the supply chain data to obtain supply chain features, and performing self-adaptive quality anomaly identification and classification according to the feedback data to obtain classification anomaly rules; Performing supply chain association analysis on the supply chain characteristics by adopting a graph neural network to obtain a supply chain heterogram, and performing anomaly labeling and risk propagation analysis on the supply chain heterogram according to the classification anomaly rule to obtain quality early warning data; And constructing a supply chain quality early-warning model according to the quality early-warning data, inputting the data to be early-warned into the supply chain quality early-warning model, and outputting a quality early-warning result.
2. The AI-based supply chain quality warning method of claim 1, wherein the method for obtaining supply chain features by feature extraction and derivative feature construction from the supply chain data using feature engineering comprises: Carrying out characteristic extraction on raw material batch information, production process parameters, online detection data, stream and storage data to obtain basic characteristics, wherein the basic characteristics comprise material component characteristics, supplier weight quality characteristics, inspection index characteristics, equipment state characteristics, process parameter characteristics, process efficiency characteristics, scale precision characteristics, surface quality characteristics, performance characteristics, storage environment characteristics and turnover efficiency characteristics; The method comprises the steps of constructing derivative features based on basic features and a bearing quality influence mechanism, identifying potential drift according to a parameter trend slope and a variation trend of sliding window fluctuation capturing parameters along with time, obtaining time sequence trend features, reflecting interaction among different procedures or parameters according to a temperature-hardness association ratio and a procedure time matching degree, obtaining process association features, quantifying multi-link abnormal accumulation effects according to abnormal parameter accumulation numbers and provider risk indexes, obtaining quality risk accumulation features, integrating multi-link data according to material and process matching degree and storage and performance attenuation coefficients, constructing global quality view angles, obtaining cross-data source fusion features, and obtaining industry standard deviation features according to standard deviation rates, process parameter compliance rate quantification and industry standard deviation degrees.
3. The AI-based supply chain quality warning method of claim 1, wherein the method for adaptively identifying and classifying quality anomalies according to the feedback data to obtain classification anomaly rules comprises: Calculating a dynamic threshold value for the numerical feedback data through a sliding window: ; Wherein the method comprises the steps of Is an anomaly threshold value for the i-th class of indicators, Is the average value of the ith index of a plurality of batches, Is the standard deviation of the i-th index, Is a dynamic adjustment coefficient; When the continuous 3 batches have no abnormality, the dynamic adjustment coefficient is reduced by 0.1, and when 1 abnormality occurs, the dynamic adjustment coefficient is increased by 0.2; Adopting an increment isolated forest algorithm to the high-dimensional feedback data, taking the non-abnormal feedback data as an initial training set, adopting a sliding window to update increment, and judging that the sample is abnormal and taking the sample as increment abnormal data when the sample abnormal score is greater than an abnormal threshold; The method comprises the steps of carrying out exception triggering on incremental exception data based on preset rules to obtain exception record data, wherein the preset rules are that when equipment is stopped, the equipment is longer than 2 hours, or the fluctuation range of parameters of 3 batches which are continuous with the equipment is larger than 5%, the equipment is abnormal, when the number of times of detecting data out-of-tolerance is larger than 3 times/day, or the number of times of detecting data out-of-tolerance is continuous with 2 batches, the equipment is abnormal, when the transportation temperature exceeds the transportation temperature range, the equipment is longer than 4 hours, or storage retention is larger than 60 days, and the equipment is logistical abnormality; The method comprises the steps of carrying out feature extraction on abnormal record data to obtain core features, constructing an abnormal source, a fault mode and a risk level three-level classification system based on a bearing quality influence mechanism, mapping the core features into abnormal labels to obtain three-level classification labels, wherein the abnormal source comprises material abnormality, process abnormality, logistics abnormality and design abnormality, wherein the material abnormality is subdivided into components exceeding a standard, inclusion level exceeding a standard and insufficient hardness, the process abnormality is subdivided into dimensions exceeding a standard, surface defects and assembly errors, the logistics abnormality is subdivided into humiture exceeding a standard and retention overtime, the design abnormality is subdivided into unreasonable structure, parameter matching errors and two-level, and the risk level is quantized and three-level based on an abnormal influence range, severity and repairability; When the inconsistency rate of the manual classification result and the model classification result of a certain type of abnormality is more than 15%, triggering rule updating, namely dynamically optimizing the classification rule by adopting a mode of combining reinforcement learning and expert feedback, collecting manual correction samples of nearly 3 months, retraining a classifier by adopting a XGBoost algorithm, adjusting feature weights, performing retrospective verification on historical data by using the new rule, establishing an abnormality classification calibration committee, checking the classification result of high risk abnormality in quarterly, adjusting classification dimension, and revising a risk grade judgment threshold; and outputting the three-level classification labels, the feedback data labeling classification results and the recording rule adjustment time, reasons and effects as classification abnormal rules.
4. The AI-based supply chain quality warning method of claim 1, wherein the method for obtaining a supply chain abnormal pattern by performing supply chain association analysis on the supply chain characteristics using a graph neural network comprises the following steps: Abstract the supply chain entity into multi-class nodes, associating corresponding characteristics with each class of nodes, and defining multi-class directed edges according to service logic to construct a graph neural network; The method comprises the steps of obtaining embedded vectors for multiple classes of nodes by adopting independent embedded layers, carrying out Euclidean normalization on the embedded vectors through mutual information screening key features, distributing independent attention to multiple classes of directed edges, and calculating the attention score of node pairs: ; Wherein the method comprises the steps of Is the embedded vector of the b-th node, Is the embedded vector of the z-th node, For the weight matrix of the r-th edge type, For the purpose of the transposition, The attention scores for the b-th and z-th nodes, In order to embed the dimensions in-line, Bias for the r-th edge type; Adjusting attention weight by combining with historical quality events, sampling 2-hop neighborhood for each node by adopting random walk and depth priority strategies, calculating degree centrality and medium centrality of the nodes, and automatically generating potential associated edges when the feature similarity of the two nodes is larger than a similarity threshold value and the history has no direct edges; Learning the reconstruction probability of the normal node through a graph self-encoder, marking the node as an abnormal node when the reconstruction error of the node is larger than an error threshold value, and comprehensively judging the abnormal probability by combining the self-characteristics of the abnormal node and the neighborhood abnormality; Searching risk propagation paths along the edge type priority by taking an abnormal node as a starting point, calculating the propagation probability of each edge, and ordering in a descending order through path risk values, wherein the propagation probability of each edge is edge weight multiplied by historical propagation frequency; Acquiring node anomaly probability, neighborhood risk mean value and historical risk record, taking weighted sum of the node anomaly probability, neighborhood risk mean value and historical risk record as node risk index, and outputting a supply chain abnormal pattern according to the node risk index, the propagation path and the graph neural network.
5. The AI-based supply chain quality warning method of claim 1, wherein the method for anomaly labeling and risk propagation analysis of the supply chain heterogeneous map according to the classification anomaly rules comprises: judging the abnormal state of each node in the abnormal composition based on the classification abnormal rule, traversing all nodes in the abnormal composition, extracting node characteristic values, comparing the characteristic values with a threshold value in the classification abnormal rule, determining the abnormal grade of each node and adding an abnormal label, wherein the abnormal state comprises normal, slight abnormal and serious abnormal; quantifying the abnormal conduction capacity of the association relation between the nodes to obtain the edge abnormal strength, and dynamically adjusting the edge weight based on the edge attribute and the abnormal state of the nodes; taking the attention score of the adjacent node as the influence weight of the adjacent node, screening the adjacent node with the greatest influence on the current node by the influence weight, and carrying out space aggregation on the node according to the influence weight to obtain a space aggregation result; Capturing dynamic changes of risks along with time by adopting an LSTM operator for graph structures of different time snapshots, and integrating multi-time information through a time attention mechanism to obtain a time aggregation result, wherein the expression is as follows: ; Wherein the method comprises the steps of As a result of the spatial aggregation at the instant t, The state is hidden for the node at time t, The state is hidden for the node at time t-1, Is an LSTM operator; Based on the time space aggregation result and the space aggregation result, the upstream propagation path of the abnormal node is traced by adopting breadth-first search, the probability of the risk propagation from the source node to the target node is quantized, and the risk propagation probability is obtained by combining the dynamic adjustment of the edge risk intensity and the path length, wherein the expression is as follows: ; Wherein the method comprises the steps of For the edge anomaly strengths of node u and adjacent node c, As a path length attenuation factor, As the number of edges of the path, Is a path Risk propagation probability of (a); and comprehensively evaluating and outputting the early warning grade based on the propagation path probability and the abnormal grade of the target node.
6. The AI-based supply chain quality warning method of claim 1, wherein the method of constructing a supply chain quality warning model from the quality warning data comprises: the method comprises the steps of constructing a supply chain quality early warning model for abnormal positioning, risk propagation and early warning output based on a graph neural network, and fusing supply chain abnormal composition and dynamic quality rules, wherein the supply chain quality early warning comprises a graph construction layer, a feature fusion layer, a risk reasoning layer and an early warning output layer; The graph construction layer comprises node definition and attribute mapping, side relation construction and heterogeneous graph representation, wherein the node definition and attribute mapping comprises the steps of abstracting a supply chain entity into multi-class core nodes, wherein node attributes are derived from basic features and derivative features, defining multi-class directed edges based on supply chain business logic, calculating edge weights through a dynamic weighting algorithm, and storing a graph structure by adopting an adjacency list, wherein the graph structure is represented as a node set, an edge set and a node feature matrix; The feature fusion layer comprises a node embedding layer and a time sequence feature fusion layer, wherein the node embedding layer adopts an improved node bipartite graph embedding algorithm to map nodes into low-dimensional vectors; the risk reasoning layer comprises abnormal node identification, risk propagation analysis, root cause positioning, wherein the abnormal node identification is carried out based on an abnormal detection mechanism of a reconstruction error; the early warning output layer comprises early warning index calculation and early warning grade division, and the early warning indexes comprise diffusion speed, influence range and predicted loss.
7. An AI-based supply chain quality warning system for performing the method of any of claims 1-6, comprising: The system comprises a data acquisition and processing module, a data processing module and a data processing module, wherein the data acquisition and processing module is used for acquiring supply chain data and feedback data of a preset automobile bearing and preprocessing the supply chain data and the feedback data, wherein the supply chain data comprises raw material batch information, production process parameters, online detection data, stream and storage data, and the feedback data comprises a repair record, a customer complaint, a failure analysis report, a third party detection report, a supplier credit rating, abnormal record information and an industry recall event; The characteristic extraction and anomaly identification classification module is used for carrying out characteristic extraction and derivative characteristic construction according to the supply chain data by adopting characteristic engineering to obtain supply chain characteristics, and carrying out self-adaptive quality anomaly identification and classification according to the feedback data to obtain a classification anomaly rule; The association analysis and propagation analysis module is used for carrying out supply chain association analysis on the supply chain characteristics by adopting a graph neural network to obtain supply chain heterograms, and carrying out anomaly labeling and risk propagation analysis on the supply chain heterograms according to the classification anomaly rules to obtain quality early warning data; And the model construction and output module is used for constructing a supply chain quality early-warning model according to the quality early-warning data, inputting the data to be early-warned into the supply chain quality early-warning model and outputting a quality early-warning result.

Description

AI-based supply chain quality early warning method and system Technical Field The invention relates to the technical field of supply chain quality early warning, in particular to an AI-based supply chain quality early warning method and system. Background With the development of the automobile industry to intelligent and high-end, the quality of the automobile bearing serving as a core transmission component directly relates to the safety and the running stability of the whole automobile. The bearing supply chain covers a plurality of links such as raw material purchase, multi-station production, cross-regional logistics, warehouse management, terminal feedback and the like, and relates to multi-source heterogeneous data such as raw material batches, production parameters, detection data, repair records and the like, and the data volume is huge and the association is complex. The traditional supply chain quality early warning method has the remarkable limitations that on one hand, abnormal judgment is dependent on a fixed threshold value, dynamic scenes such as production process fluctuation, raw material characteristic change and the like are difficult to adapt to, missing report or false report is easy to occur, on the other hand, deep mining on the association relation of each link of the supply chain is lacking, the propagation path of risks among multiple nodes cannot be recognized, early warning hysteresis is high, and passive response is often carried out after quality problems are exploded. In addition, the traditional method has insufficient processing capability on high-dimensional feedback data, unstructured information such as customer complaints, industry recalls and the like is difficult to integrate, and accurate pre-judgment of the quality risk of the whole chain cannot be achieved. The problems cause that the hidden danger of the quality of the bearing supply chain is difficult to check in advance, so that not only can direct losses of product repair, customer complaints and the like be possibly caused, but also an industry recall event can be caused when serious, and the brand reputation of enterprises is damaged. Therefore, a fusion AI technology is needed to integrate the supply chain quality early warning method of multisource data, dynamically identify anomalies, trace back risk propagation and accurately output early warning, meet the high quality control requirement of the bearing supply chain, and improve early warning timeliness and accuracy. Disclosure of Invention The invention aims to provide an AI-based supply chain quality early warning method. In order to achieve the above purpose, the invention is implemented according to the following technical scheme: The invention comprises the following steps: The method comprises the steps of collecting supply chain data and feedback data of a preset automobile bearing, and preprocessing the supply chain data and the feedback data, wherein the supply chain data comprises raw material batch information, production process parameters, online detection data, flow and storage data, and the feedback data comprises a repair record, a customer complaint, a failure analysis report, a third party detection report, a supplier credit rating, abnormal record information and an industry recall event; Performing feature extraction and derivative feature construction by adopting feature engineering according to the supply chain data to obtain supply chain features, and performing self-adaptive quality anomaly identification and classification according to the feedback data to obtain classification anomaly rules; Performing supply chain association analysis on the supply chain characteristics by adopting a graph neural network to obtain a supply chain heterogram, and performing anomaly labeling and risk propagation analysis on the supply chain heterogram according to the classification anomaly rule to obtain quality early warning data; And constructing a supply chain quality early-warning model according to the quality early-warning data, inputting the data to be early-warned into the supply chain quality early-warning model, and outputting a quality early-warning result. Further, the method for obtaining the supply chain characteristics by carrying out characteristic extraction and derivative characteristic construction according to the supply chain data by adopting characteristic engineering comprises the following steps: Carrying out characteristic extraction on raw material batch information, production process parameters, online detection data, stream and storage data to obtain basic characteristics, wherein the basic characteristics comprise material component characteristics, supplier weight quality characteristics, inspection index characteristics, equipment state characteristics, process parameter characteristics, process efficiency characteristics, scale precision characteristics, surface quality characteristics, performance characteristics,