Purchase supply chain collaborative intelligent management method and system

Abstract

The invention provides a procurement supply chain collaborative intelligent management method and a procurement supply chain collaborative intelligent management system. Belongs to the technical field of supply chain management. Constructing a purchase demand prediction model; dynamically adjusting the purchasing strategy according to the prediction result; the record and the evaluation result are stored on the block chain; the system automatically sends order information to a supplier; the supplier confirms the order through the block chain platform and updates the production progress and the logistics information; the system adjusts an inventory strategy in real time; the logistics state is monitored in real time; the system analyzes data of each link of the supply chain in real time and identifies potential risks. Through dynamic path planning and real-time transportation optimization, the logistics path can be adjusted in real time according to traffic conditions, environmental factors and emergencies, the transportation time is shortened, the transportation cost is reduced, and therefore the overall efficiency of a supply chain is remarkably improved.

Description

Technical Field

[0001] The present invention provides a collaborative intelligent management method and system for a procurement supply chain, belonging to the technical field of supply chain management. Background Art

[0002] With the increasing complexity of the global supply chain, logistics and supply chain management face many challenges, including issues such as transportation route optimization, inventory management, and supply chain coordination. Traditional supply chain management methods usually rely on centralized management systems and manual intervention, making it difficult to cope with rapidly changing market demands and emergencies. In addition, traditional methods lack the full utilization of cross-enterprise cooperation and real-time data streams, resulting in problems such as low supply chain efficiency, inaccurate inventory, and delivery delays. Summary of the Invention

[0003] The present invention provides a collaborative intelligent management method and system for a procurement supply chain to solve the problems mentioned in the above background art: A collaborative intelligent management method for a procurement supply chain proposed by the present invention, the method comprising: S1. Obtain enterprise-related data; S2. Build a procurement demand prediction model; dynamically adjust the procurement strategy according to the prediction result; S3. Store the records and evaluation results on the blockchain; S4. The system automatically sends the order information to the supplier; the supplier confirms the order, updates the production progress and logistics information through the blockchain platform; the system adjusts the inventory strategy in real time; and monitors the logistics status in real time; S5. The system analyzes the data of each link of the supply chain in real time to identify potential risks.

[0004] A collaborative intelligent management system for a procurement supply chain proposed by the present invention includes a memory, a processor, and a computer program stored on the memory and executable on the memory. The processor executes the program to implement the collaborative intelligent management method for a procurement supply chain as described in any one of the above.

[0005] Advantages of the present invention: Through dynamic path planning and real-time transportation optimization, it is possible to adjust the logistics path in real time according to traffic conditions, environmental factors, and emergencies, reducing transportation time and costs, thereby significantly improving the overall efficiency of the supply chain; By adopting an inventory forecasting and risk assessment method based on multi-modal data fusion, it is possible to more accurately predict demand fluctuations and changes in the supply chain environment, avoid overstocking or shortages, ensure the health of inventory, thereby reducing the risk of out-of-stock and inventory backlogs, and improving the accuracy of inventory management; Through smart contracts and edge computing technologies, real-time data sharing and collaborative decision-making can be achieved in all links of the supply chain, enhancing the supply chain's response speed to market demand fluctuations and emergencies, and strengthening the flexibility and adaptability of the supply chain; Based on federated learning technology, collaborative work and data sharing can be realized among different enterprises while protecting data privacy, optimizing global path planning and resource scheduling, reducing information asymmetry, and improving the efficiency of cooperation among all parties; Through the comprehensive integration of intelligent decision-making and real-time data streams, it is possible to better respond to emergencies and risks in the supply chain, such as natural disasters, market fluctuations, etc., improving the resilience and risk response ability of the supply chain; By optimizing transportation routes, reducing inventory backlogs, and improving supply chain collaboration efficiency, it is possible to significantly reduce logistics and operation costs, improve overall profitability, and gain an advantage in the fierce market competition; By introducing advanced technologies such as artificial intelligence and deep learning, the automation and intelligence level of the supply chain can be improved, making supply chain management more efficient, accurate, and flexible, reducing manual intervention, and improving management effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 It is a flowchart of the method steps of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0007] The following describes the preferred embodiments of the present invention with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0008] One embodiment of the present invention is as Figure 1 shown. A collaborative intelligent management method for a procurement supply chain, the method comprising: S1. Obtain enterprise-related data, where the related data includes enterprise internal data (such as historical procurement records, inventory levels, sales plans, etc.) and external data (such as market price fluctuations, industry dynamics, etc.); and clean and fuse the obtained related data; S2. Based on a deep learning algorithm, construct a procurement demand forecasting model with multiple inputs and multiple outputs; the procurement demand model can consider the influence of various factors (such as seasonality, trendiness, periodicity, etc.) on procurement demand; according to the prediction results, combined with the enterprise's strategic goals and market environment, dynamically adjust procurement strategies, such as procurement quantity, procurement timing, and supplier selection; S3. The supplier submits registration information through the blockchain platform. The registration information includes qualification certificates, historical performance, and credit evaluations. After the platform verifies the information, it stores the immutable records on the blockchain. After each transaction is completed, both the buyer and the seller evaluate the transaction, and the evaluation results are recorded on the blockchain. The system dynamically updates the supplier's credit score based on the evaluation results and transaction history. According to the procurement requirements and the supplier's credit score, the system intelligently screens the qualified suppliers and recommends them to the procurement department. S4. After the purchase order is generated, the system automatically sends the order information to the supplier and monitors the order execution status in real time. The supplier confirms the order, updates the production progress, and logistics information through the blockchain platform. The system adjusts the inventory strategy in real time according to the sales forecast, production plan, and inventory level. Combining Internet of Things technology and big data analysis, it monitors the logistics status in real time and optimizes the logistics route and transportation mode. S5. The system analyzes the data of each link in the supply chain in real time to identify potential risks (such as supplier bankruptcy risk, logistics interruption risk, market demand change risk, etc.). Once the risk reaches the warning threshold, the system immediately issues a warning signal. For different types of risks, the system pre - formulates emergency response plans, which include switching suppliers, adjusting procurement strategies, and increasing inventory levels. When a risk occurs, the system automatically executes corresponding measures according to the emergency response plan and monitors the execution effect of the measures in real time.

[0009] The working principle of the above technical solution is as follows: By integrating internal enterprise data (such as historical procurement, inventory, sales plans) and external data (such as market prices, industry trends), a multi-dimensional dataset is constructed; A prediction model with multiple inputs (such as time series, seasonality, trend) and multiple outputs (procurement volume, procurement timing, supplier selection) is adopted, based on the LSTM or Transformer architecture, to capture complex non-linear relationships in the data; Combining enterprise strategic goals (such as cost priority, risk aversion) and market environment (such as demand fluctuations, price changes), real-time procurement strategy suggestions are generated; Suppliers submit qualification certificates, historical performance, and credit evaluations through the blockchain platform. The platform automatically verifies the authenticity of the information through smart contracts and stores the immutable records on the blockchain; Based on transaction history and buyer-seller evaluations (such as on-time delivery rate, quality score), the system dynamically updates the supplier credit score through a weighted algorithm; According to procurement requirements and supplier scores, the system automatically screens eligible suppliers and recommends the optimal options, reducing manual intervention; After the purchase order is generated, the system automatically pushes it to the supplier through the API or blockchain smart contract. The supplier confirms the order and updates the production progress and logistics information in real time; Combining sales forecasts, production plans, and inventory levels, the system adjusts the inventory strategy (such as safety stock threshold, reorder point) in real time through reinforcement learning algorithms; Using Internet of Things devices (such as GPS, RFID) and big data analysis, the system monitors the logistics status in real time and dynamically plans the optimal route and transportation method (such as air / land transportation switch); The system analyzes the data of each link in the supply chain (such as supplier financial health, logistics delay rate, market demand changes) in real time through machine learning algorithms to identify potential risks; When the risk indicators (such as supplier credit score < 50 points, logistics delay exceeding 24 hours) reach the preset threshold, the system automatically triggers a warning signal; For different risk types (such as supplier bankruptcy, logistics interruption), the system pre-formulates emergency plans (such as switching to alternative suppliers, adjusting procurement strategies, increasing inventory buffers), and automatically executes them when the risks occur, while monitoring the execution effect.

[0010] The effects of the above technical solutions are as follows: By using deep learning models and blockchain technology, manual intervention can be reduced, and the procurement cycle can be shortened by 30% - 50%; the blockchain platform enables the transparency of supplier information, and the order confirmation time is shortened from several days to several hours; through real-time path planning and transportation mode selection, logistics costs can be effectively reduced, and the on-time delivery rate can be improved; the dynamic inventory strategy can reduce inventory backlogs and out-of-stock risks, and lower inventory holding costs; the blockchain verification and intelligent screening mechanism reduce the cost of supplier qualification review and shorten the supplier switching time; through path optimization and transportation mode adjustment, logistics costs can be reduced; through blockchain reputation scoring and dynamic monitoring mechanisms, the supplier default rate can be lowered; through real-time early warning and automated emergency response, the logistics interruption recovery time can be shortened; through the multi-factor demand forecasting model, the demand forecasting accuracy can be improved, and the risks of overstock or shortage can be reduced; all transaction records and evaluations are tamper-proof, meeting the requirements of compliance audits; through encryption technology and permission management, sensitive corporate data (such as procurement prices, supplier information) can be protected; through data dashboards and real-time monitoring interfaces, management can quickly obtain the health status and key indicators of the supply chain; the system provides suggestions on procurement strategies, inventory strategies, logistics optimization, etc. based on historical data and real-time analysis to assist decision-making.

[0011] In one embodiment of the present invention, S1 includes: S11: Real-time access to systems such as ERP, MES, and WMS through API interfaces to collect data such as production plans, inventory dynamics, and logistics trajectories; use natural language processing (NLP) technology to parse text data such as procurement contracts and supplier emails, and extract key information (such as delivery dates, price terms). S12: Use WebScraping technology to capture data such as price fluctuations and inventory levels on e-commerce platforms and industry reports; use sentiment analysis technology to monitor industry trends, policy changes, and supplier reputations on social media. S13: Predict missing data based on time series models (such as ARIMA), and combine the isolation forest algorithm with domain knowledge to identify and correct outliers. S14: Uniformly map text, numerical, and time series data to a semantic space (such as Word2Vec + numerical normalization); construct a multi-dimensional graph model of suppliers - products - time, and integrate historical procurement data with external market dynamics.

[0012] The working principle of the above technical solution is as follows: Real-time access to enterprise resource planning (ERP), manufacturing execution system (MES), warehouse management system (WMS), etc. through standardized API interfaces, automatically collecting structured data such as production plans, inventory dynamics, and logistics trajectories; using pre-trained language models (such as BERT, GPT) to parse unstructured texts such as procurement contracts and supplier emails, and extracting key information (such as delivery dates, price terms, quality standards); automatically annotating entities such as supplier names, product models, and amounts in contracts through named entity recognition (NER) technology; combining context semantic analysis to solve ambiguity problems (such as whether "delivery within 30 days" refers to working days or natural days); using tools such as Selenium or Playwright to simulate user behavior and scrape dynamic data such as prices and inventory on e-commerce platforms; using technologies such as proxy IP pools and request header disguises to avoid anti-crawler mechanisms; using BERTopic or LDA topic models to extract key topics from industry reports (such as "chip shortage"); judging the impact of policies on the supply chain (positive / negative) through keyword matching (such as "tariff adjustment") and sentiment polarity analysis; combining data such as social media comments and news reports, and quantifying supplier reputation through sentiment scoring (such as VADER, TextBlob); predicting missing values at future moments based on the time dependence of historical data (such as the inventory level of a certain product in the next week); using LSTM or Transformer models to capture complex time patterns and improve prediction accuracy; identifying outliers by randomly dividing the data space (such as a sudden 300% extension of a supplier's delivery time); combining industry experience (such as "the standard delivery period of a certain product is 7 days") to perform secondary verification on the algorithm results; using Word2Vec or BERT to map text data such as supplier names and product models into high-dimensional vectors; scaling numerical data such as inventory levels and prices to the [0,1] interval to eliminate dimensional differences; using suppliers, products, and time as nodes, and purchase quantities, prices, and delivery dates as edge weights; propagating information in the graph structure through a message passing mechanism to capture the complex relationships between suppliers-products-time; real-time fusing new data (such as the latest purchase records and market price fluctuations), and dynamically adjusting the graph model structure.

[0013] The effects of the above technical solutions are as follows: By integrating multi-source data such as enterprise internal systems, external platforms, and social media, a 360-degree supply chain view can be constructed, improving the degree of data visualization; by replacing manual review with NLP technology, the accuracy of key information extraction can be increased to over 95%; by correcting outliers and reducing the interference of "dirty data" on decision-making, the quality of model training data can be improved by 30%-50%; by implementing minute-level data updates through API and WebScraping technologies, real-time decision-making can be supported; by using time series models to predict future trends (such as price fluctuations and inventory gaps), risks can be warned 30-90 days in advance; by sentiment analysis, policy changes and industry crises can be discovered in advance to avoid sudden risks; by graph model association analysis, implicit relationships between suppliers - products - time can be revealed (such as "the delivery delay rate of a certain supplier increases in a specific season"); based on the graph model and prediction results, the system can automatically generate procurement strategies (such as "increase backup suppliers" and "adjust procurement batches"); by displaying supply chain vulnerabilities through graph networks (such as "a certain product depends on a single supplier"), management can be assisted in optimizing the structure; by automated data collection and parsing, the cost of data collation can be reduced; the risk of inventory backlog or out-of-stock can be reduced, and inventory turnover can be improved; by reputation assessment and historical cooperation data, high-cost-effective suppliers can be screened to reduce procurement costs.

[0014] In one embodiment of the present invention, S13 includes: For time series data (such as prices and inventory), an ARIMA-LSTM hybrid model is adopted. ARIMA is used to capture linear trends, and LSTM is used to capture non-linear fluctuations. The prediction results are fused through weighted averaging; and an attention mechanism is introduced to dynamically adjust the weights of different historical time windows. Construct a text-time series association model. For example, using the "delivery date" text information in the procurement contract and combining historical logistics track data, the missing logistics time nodes are predicted through a graph neural network. Develop a supplier-product-time triple knowledge graph, and use knowledge reasoning to fill in the missing procurement quantity data; for geographical distribution data (such as warehouse inventory), use a spatial autoregressive model, combine the inventory levels of surrounding warehouses, and predict missing values through a spatial weight matrix. Introduce domain knowledge constraints, such as "price fluctuations should not exceed ±20% of the historical average", and dynamically adjust the outlier threshold of IsolationForest through a rule engine; combine time series features (such as volatility and autocorrelation) to construct a time series-spatial joint anomaly detection model to improve the response speed to sudden anomalies (such as supply chain disruptions). Perform GMM clustering on multi-dimensional data (such as price, inventory, delivery time), and identify abnormal clusters that deviate from the mainstream distribution; Embed abnormal rules in the supplier-product-time knowledge graph, and quickly locate abnormal nodes through graph traversal algorithms (such as random walk); Combined with causal reasoning, analyze the abnormal propagation path and identify the root cause; Train a generative adversarial network (GAN) to generate a normal data distribution, and identify abnormal samples with significant differences from the generated data through a discriminator; Define a domain knowledge rule base and correct abnormal data in real time through a rule engine; Combine with an expert system to provide correction suggestions for complex anomalies (such as price fluctuations caused by policy changes); Construct a multi-objective optimization model, with data quality metrics (such as missing rate, anomaly rate) and business objectives (such as cost minimization) as constraints, and generate an optimal correction plan through genetic algorithms or particle swarm optimization.

[0015] The working principle of the above technical solution is as follows: Based on autoregressive (AR), differencing (I), and moving average (MA) components, capture the linear trend and periodicity of time series; through the gating mechanism (input gate, forget gate, output gate), learn the non-linear fluctuations and long-term dependence relationships; dynamically adjust the prediction weights of ARIMA and LSTM (for example, tend to ARIMA when historical data is stable and tend to LSTM during abnormal fluctuations); introduce time window attention weights (for example, recent data has higher weights) to enhance the response ability to emergencies; associate the text information of "delivery date" in the purchase contract (such as "delivery in 30 days") with historical logistics trajectory data (such as transportation time, delay records); construct a multi-dimensional graph of supplier-logistics node-time, and predict the missing logistics time nodes through the message passing mechanism (such as "transportation time from supplier A to warehouse B"); fill in the missing purchase quantity data based on historical cooperation data (such as "supplier A has cooperated with product B 10 times, with an average purchase quantity of 1000 pieces"); transform domain knowledge (such as "the purchase quantity fluctuation of long-term cooperative suppliers should be less than 15%") into constraint conditions in the graph; define the geographical proximity between warehouses (such as distance, traffic connectivity), and construct a spatial lag term; combine the inventory levels of surrounding warehouses and predict the inventory of the target warehouse through the spatial autoregressive equation (such as "inventory of warehouse A = β1 inventory of warehouse B + β2 inventory of warehouse C + ε"); define rules, such as "price fluctuations should not exceed ±20% of the historical mean", and dynamically adjust the anomaly threshold of IsolationForest; combine time series features (such as volatility, autocorrelation) and spatial features (such as geographical distribution) to construct a joint anomaly detection model; perform Gaussian mixture model (GMM) clustering on data such as price, inventory, and delivery date to identify anomaly clusters (such as "abnormally high price and extremely low inventory"); embed rules in the graph (such as "if the on-time delivery rate of supplier A is lower than 80% and there is negative public opinion recently, then mark it as high risk"), and quickly locate anomaly nodes through the random walk algorithm; based on the causal graph model (such as the DoWhy library), identify the root cause of anomalies (such as "raw material shortage causes supplier B to delay delivery"); train the generator to generate a normal data distribution, and identify anomaly samples with significant differences from the generated data through the discriminator; perform rule base and real-time correction, and define rules, such as "if the delivery date of supplier C is delayed by more than 5 days, then automatically trigger the evaluation of alternative suppliers"; combine the experience of domain experts to provide correction suggestions; use data quality indicators (such as missing rate, anomaly rate) and business goals (such as cost minimization) as constraints; through genetic algorithm or particle swarm optimization (PSO), generate the optimal correction plan (such as "adjust the purchase quantity of supplier A to 800 pieces and increase the purchase quantity of alternative supplier B at the same time").

[0016] The effects of the above technical solutions are as follows: by combining linear and nonlinear prediction capabilities through ARIMA-LSTM, the time series prediction error can be reduced; by dynamically adjusting the time window weight, the response speed to emergencies (such as logistics disruptions caused by the epidemic) can be improved; the accuracy of filling missing data is increased to more than 85%, which can reduce the need for manual intervention; combined with geographical distribution data, the RMSE of predicted warehouse inventory is reduced by 15%-25%; by combining GMM clustering with graph traversal algorithms, the recall rate of anomaly detection can be effectively improved; by quickly locating the root cause of the anomaly, the false alarm rate can be reduced (for example, attributing "delayed delivery by supplier B" to "shortage of raw materials" rather than "insufficient supplier capacity"); by correcting abnormal data in real time, decision delays can be reduced (for example, the response time of "alternative supplier evaluation" is shortened from hours to minutes); by generating correction plans that take into account both data quality and business goals, supply chain costs can be reduced; by automatically identifying high-risk suppliers (for example, "on-time delivery rate is less than 80% and negative public opinion"), potential risks can be warned in advance.

[0017] In one embodiment of the present invention, the S14 includes: The numerical data (such as price and inventory) are binned and discretized, and the distribution features are extracted by combining the Gaussian mixture model (GMM) and mapped into high-dimensional sparse vectors. The domain ontology library is introduced (for example, "price > 1,000 yuan - high price tag"), and the numerical features are converted into semantic tags through the rule engine. The BERT-BiLSTM-CRF model is used to extract entities (such as "delivery date" and "default clause") and relationships (such as "supplier-supply-product") in the procurement contract to generate structured semantic triples. The BERT model is compressed into a lightweight embedding model through knowledge distillation and deployed on edge devices for real-time semantic parsing. The domain adaptation module based on the Generative Adversarial Network (GAN) trains the text encoder and the numeric encoder to generate a shared semantic space and minimize the distribution difference between modalities. Perform wavelet transform on time series data (such as price fluctuations) to extract multi-scale features, combine the self-attention mechanism to capture periodicity and trend, and generate time series semantic fingerprints; align the time series semantic fingerprints with the text / numeric semantic vectors through dynamic time warping (DTW) to construct a spatiotemporal semantic association matrix; Define the supply chain domain ontology (e.g., "supplier-product-time" triples) based on OWL2 ontology language, and perform cross-system data association through SPARQL query; use entity linking technology (e.g., BERT-NER+relation extraction) to link entities in unstructured text (e.g., "XX Company") to standardized nodes in the knowledge graph; Define dynamic rule templates (e.g., "If supplier A has 3 consecutive late deliveries, its risk level is upgraded to high"), and use a rule inference engine (e.g., Drools) to update the properties of the knowledge graph in real time; Combined with a time series database, convert time series data into temporal attributes in the knowledge graph; construct a temporal heterogeneous graph with the dimensions of supplier - product - time, where the node types include suppliers, products, and warehouses, and the edge types include supply relationships, logistics relationships, and price associations; Introduce timestamp encoding and temporal edge weights to enhance the temporal awareness of the graph model; adopt a temporal graph convolutional network, and use a gating mechanism (e.g., GRU) to fuse the historical graph structure and the current graph snapshot to predict the future state of the supply chain network; Combined with the graph attention network (GAT), dynamically adjust the importance weights of different nodes (e.g., "high - risk suppliers") to improve the accuracy of anomaly detection; through a joint training objective, simultaneously optimize triple tasks of node classification (e.g., supplier risk level prediction), edge prediction (e.g., future cooperation probability), and graph - level regression (e.g., overall supply chain cost prediction); Embed a causal graph structure in the knowledge graph, and simulate the impact of different policies on the supply chain through causal intervention; construct a supply chain risk propagation model, and simulate the diffusion path of risks (e.g., "supplier bankruptcy") in the supply chain network through the graph neural network propagation algorithm; combine Monte Carlo simulation to generate a risk probability distribution to provide quantitative support for procurement decisions.

[0018] The working principle of the above technical solution is as follows: continuous numerical values (such as prices, inventory) are divided into discrete intervals (such as "0 - 500 yuan", "501 - 1000 yuan"); the numerical distribution is fitted by a Gaussian mixture model, and statistical features such as mean and variance are extracted and mapped into high-dimensional sparse vectors (such as "the price is in the range of 501 - 1000 yuan, and the variance is 120"); combined with ontology rules (such as "price > 1000 yuan - high price label"), the numerical features are transformed into semantic labels (such as "high price", "low inventory"); the procurement contract text is encoded by BERT, BiLSTM captures context dependencies, and the CRF layer outputs entity boundaries (such as "delivery date: 2023 - 12 - 31"); the relationships between entities are identified through an attention mechanism (such as "supplier A - supplies - product B") to generate structured triples; a large BERT model is used as the teacher model to train a lightweight embedding model (such as DistilBERT) as the student model to retain key semantic information; the lightweight model is deployed on edge devices (such as IoT gateways) to achieve real-time semantic parsing; a text encoder (such as BERT) and a numerical encoder (such as MLP) are trained, and a shared semantic space is generated through GAN; the distribution difference between the text and numerical modalities (such as KL divergence) is minimized to achieve cross-modal semantic alignment; time series data (such as price fluctuations) are decomposed by wavelet transform to extract high-frequency (short-term fluctuations) and low-frequency (long-term trends) features; the importance of different time steps is dynamically adjusted through attention weights to generate a temporal semantic fingerprint (such as "the price increased significantly in Q3 of 2023"); the temporal semantic fingerprint is aligned with the text / numerical semantic vectors to construct a spatio-temporal semantic association matrix (such as the temporal correlation between "price increase" and "supplier's late delivery"); a supply chain ontology (such as the "supplier - product - time" triple) is defined, and cross-system data association is achieved through SPARQL queries (such as "query all transaction records of supplier A in 2023"); named entity recognition (NER) is used to extract entities from unstructured text (such as "XX company") and link them to standardized nodes in the knowledge graph through relation extraction; such as "if supplier A delays delivery continuously for 3 times, then its risk level is upgraded to high"; the attributes of the knowledge graph are updated in real time through a rule engine such as Drools (such as "the risk level of supplier A: high"); the nodes include suppliers, products, and warehouses, and the edges include supply relationships, logistics relationships, and price associations; timestamp encoding (PositionalEncoding) and temporal edge weights (such as "cooperation frequency in the past 30 days") are introduced to enhance the temporal awareness ability of the graph model; the historical graph structure and the current graph snapshot are fused through GRU to predict the future supply chain network state (such as "the probability of cooperation with suppliers in Q1 of 2024"); the attention weights are adjusted according to the node importance (such as "high-risk suppliers") to improve the accuracy of anomaly detection;Optimize node classification (such as supplier risk level prediction), edge prediction (such as future cooperation probability), and graph-level regression (such as overall supply chain cost prediction) simultaneously; simulate the impact of policy changes (such as tariff adjustments) on the supply chain through Do-Calculus (such as "a 10% increase in tariffs leads to a 5% increase in supplier quotes"); simulate the diffusion path of risks (such as "supplier bankruptcy") in the supply chain network; generate risk probability distributions (such as "the probability of supplier A going bankrupt is 20%") to provide quantitative support for procurement decisions.

[0019] The effects of the above technical solutions are as follows: By semanticizing numerical data, the feature dimension can be increased by 3 - 5 times, enhancing the model interpretability; the semantic alignment accuracy between text and numerical modalities is increased to over 90%, and multi-source data fusion analysis is supported; by adopting a lightweight model, the inference speed can be increased by 4 - 6 times, meeting the real-time requirements of edge devices; the GAN domain adaptation module reduces the distribution differences between modalities, and the cross-modal task performance is improved by 15% - 25%; by combining wavelet transform and self-attention mechanism, the accuracy of temporal feature extraction is improved by 20% - 30%; the DTW algorithm achieves precise alignment between temporal semantic fingerprints and text / numerical vectors, improving the efficiency of constructing the correlation matrix; the combination of OWL2 ontology language and SPARQL query reduces the response time for cross-system data association to the second level; the BERT-NER + relation extraction technology improves the accuracy of entity linking; the combination of dynamic rule templates and rule inference engines can make the delay of knowledge graph attribute updates less than 1 second; the combination of temporal heterogeneous graphs and timestamp encoding can improve the temporal prediction accuracy of graph models; under the joint training objective, the performance of node classification, edge prediction, and graph-level regression tasks can be improved by 12%, 18%, and 20% respectively; the GAT dynamic weight adjustment mechanism improves the recall rate of anomaly detection; the combination of the GraphSAGE propagation algorithm and Monte Carlo simulation can increase the F1 score of risk diffusion path prediction to over 80%.

[0020] In one embodiment of the present invention, S2 includes: S21. Construct a multi-modal Transformer model, where the input layer fuses text (policy text), numerical (price), and time series (inventory) data, and outputs multi-dimensional predictions such as procurement quantity and procurement timing; S22. Introduce adversarial samples (such as simulating price shocks) during the training process; use historical procurement task data to train the model's environmental adaptability, for example, quickly adapt to the new market environment (such as policy changes); the Actor network outputs procurement strategies (such as procurement quantity), and the Critic network evaluates the long-term benefits of the strategies, and optimizes the strategies by interacting with the environment; S23. Simulate the performance of the supply chain under different procurement strategies in a virtual environment for assisting in decision-making.

[0021] The working principle of the above technical solution is as follows: construct a multi-modal Transformer model, which has the ability to process various types of data; input different modal data such as text (policy text), numerical values (prices), and time series (inventory) into the input layer of the model. The model comprehensively processes and analyzes these multi-modal data through internal structures such as the self-attention mechanism, and finally outputs multi-dimensional prediction results such as the purchase quantity and purchase timing; during the model training process, adversarial samples are introduced, such as simulating price shocks. These adversarial samples can simulate various unexpected situations that may occur in the actual procurement process, increasing the diversity and complexity of the training data; use historical procurement task data to train the environmental adaptability of the model. By continuously exposing to different types of data and scenarios, the model can quickly adapt to the new market environment, such as policy changes, etc.; adopt the Actor-Critic framework for policy optimization. The Actor network is responsible for outputting procurement policies, such as the purchase quantity; the Critic network evaluates the policies output by the Actor network and predicts their long-term benefits. The Actor network continuously adjusts and optimizes its own policies according to the evaluation results of the Critic network, and gradually finds the optimal procurement policy through interaction with the simulated environment; in the virtual environment, simulate the performance of the supply chain under different procurement policies. The virtual environment can be constructed according to the characteristics and rules of the actual supply chain, simulating various possible supply chain scenarios; by observing and analyzing the performance of different procurement policies in the virtual environment, provide assistance for actual procurement decisions. Decision-makers can evaluate the advantages and disadvantages of different procurement policies based on the simulation results and select the most suitable procurement policy for the current situation.

[0022] The effects of the above technical solutions are as follows: By integrating data of multiple modalities such as text, numerical values, and time series, the model can obtain more comprehensive and rich information, thereby improving the accuracy of predictions such as procurement volume and procurement timing; introducing adversarial samples for training enables the model to better handle various emergencies in the actual procurement process, enhancing the robustness and generalization ability of the model, and further improving the reliability of predictions; using historical procurement task data to train the environmental adaptability of the model enables the model to quickly adapt to changes in the new market environment, such as policy changes. This helps enterprises adjust procurement strategies in a complex and changing market environment in a timely manner and reduce procurement risks; the Actor-Critic framework continuously optimizes the procurement strategy, enabling the model to automatically adjust procurement behavior according to different market environments, improving the scientificity and rationality of procurement decisions; simulating different procurement strategies in a virtual environment can evaluate the effects of various strategies in advance and provide references for actual decisions. This helps decision-makers avoid blind decisions and reduce decision-making risks; by simulating the performance of the supply chain, procurement strategies can be evaluated from multiple dimensions, such as cost, inventory level, supply stability, etc. Decision-makers can formulate better procurement decisions by comprehensively considering various factors based on these evaluation results.

[0023] In one embodiment of the present invention, the S3 includes: S31. Verify the qualification certificates submitted by the supplier through ZKP; based on a preset qualification standard (such as ISO certification), the smart contract automatically reviews the certification documents submitted by the supplier; S32. Use cross-chain technology (such as CosmosIBC) to synchronize the supplier credit score to multiple industry alliance chains to expand the scope of credit evaluation; S33. Combine the time decay factor and the trading volume weight to dynamically adjust the supplier credit score to avoid excessive weight of historical evaluations; S34. On the premise of protecting enterprise privacy, jointly train a supplier recommendation model with the data of multiple enterprises and deploy a lightweight recommendation model on local edge devices to respond to procurement requirements in real time.

[0024] The working principle of the above technical solution is as follows: The zero-knowledge proof (ZKP) technology is used to verify the qualification certificates submitted by suppliers. Zero-knowledge proof allows the verifier to confirm that the supplier has a certain qualification without obtaining the specific content of the supplier's qualification certificate, thus protecting the privacy of the supplier while ensuring the accuracy of verification; Based on preset qualification standards, such as ISO certification, etc., the smart contract automatically reviews the proof documents submitted by the suppliers. The smart contract checks the authenticity, validity, etc. of the proof documents according to the pre-set rules and logic and gives the review results; Using cross-chain technology, such as Cosmos IBC (Inter-Blockchain Communication Protocol), the supplier credit score is synchronized to multiple industry alliance chains. Through cross-chain technology, data interaction and sharing can be achieved between different blockchains, thus expanding the scope of credit evaluation and enabling more enterprises and institutions to obtain the credit information of suppliers; Combining the time decay factor and the trading volume weight, the supplier credit score is dynamically adjusted. The time decay factor makes the weight of historical evaluations gradually decrease over time, avoiding the excessive influence of historical evaluations on the current credit score; The trading volume weight, on the other hand, weights the evaluations of different transactions according to the trading volume of the supplier, more accurately reflecting the actual credit status of the supplier; On the premise of protecting the privacy of enterprises, the data of multiple enterprises are combined to train the supplier recommendation model. By adopting privacy protection technologies, such as federated learning, etc., multiple enterprises can jointly participate in the training of the model without sharing the original data, improving the accuracy and generalization ability of the model; The trained lightweight recommendation model is deployed on local edge devices.

[0025] The effects of the above technical solutions are as follows: The combination of zero-knowledge proof and smart contract ensures the accuracy and automation of supplier qualification verification. The smart contract can execute the preset audit rules quickly and accurately, avoiding the interference of human factors and improving the efficiency and quality of verification; The zero-knowledge proof technology protects the privacy of suppliers, enabling them not to worry about the leakage of the specific content of qualification certificates and increasing their enthusiasm for participating in cooperation; The application of cross-chain technology enables the synchronization of supplier credit scores among multiple industry alliance chains, expanding the scope of credit assessment. More enterprises and institutions can obtain the credit information of suppliers, thereby more comprehensively evaluating the credit status of suppliers and reducing procurement risks; Dynamically adjusting the supplier credit score by combining the time decay factor and transaction volume weight can more accurately reflect the actual credit status of suppliers. Avoiding the distortion of the score caused by the excessive weight of historical evaluations makes the credit score more timely and valuable for reference; Training the supplier recommendation model by combining the data of multiple enterprises makes full use of the data resources of different enterprises and improves the accuracy and generalization ability of the model. The model can better understand the procurement needs of enterprises and the characteristics of suppliers and provide more personalized recommendation services; Deploying the lightweight recommendation model on local edge devices realizes real-time response to procurement needs. Enterprises can quickly obtain supplier recommendation results, make procurement decisions in a timely manner, and improve procurement efficiency.

[0026] In one embodiment of the present invention, step S4 includes: S41. Automatically execute operations such as order confirmation and delivery date change through the smart contract; Real-time process sensor data (such as temperature and humidity) through the edge devices deployed at logistics nodes, and optimize the transportation route based on the processing results; S42. Build a digital twin model of physical inventory and virtual inventory, simulate inventory changes in real time, and predict out-of-stock risks; Combine the genetic algorithm (GA) and deep reinforcement learning (DRL) to optimize the replenishment strategy of multi-level inventory (such as central warehouse and regional warehouse); S43. Use the federated learning framework to combine the data of multiple logistics enterprises to train a dynamic path planning model to cope with sudden traffic conditions.

[0027] The working principle of the above technical solution is as follows: A smart contract is an automatically executable program based on blockchain technology. In scenarios such as order confirmation and delivery date change, when the preset conditions are met, the smart contract will automatically trigger corresponding operations. For example, when the supplier confirms the order, the smart contract will automatically update the order status to "confirmed" and notify the purchaser; if the delivery date needs to be changed due to special circumstances and both parties reach an agreement, the smart contract will automatically adjust the delivery date information to ensure the automation and transparency of the entire transaction process; various sensors, such as temperature sensors and humidity sensors, are deployed on the edge devices at logistics nodes. These sensors collect data on the transportation environment in real time and transmit the data to the edge devices for processing. The edge devices use built-in algorithms and models to analyze the sensor data and determine whether the current transportation environment is suitable for transporting goods. If an abnormal environment is detected, such as too high temperature or too high humidity, the edge device will issue an alarm in a timely manner and optimize the transportation route according to the preset rules. For example, select a route that avoids high-temperature or high-humidity areas to ensure the quality and safety of goods during transportation; constructing a digital twin model of physical inventory and virtual inventory is to digitally map the actual inventory system in a virtual environment. Through technologies such as sensors and the Internet of Things, various data of physical inventory, such as inventory quantity, location, and status, are collected in real time and synchronized to the virtual inventory model. The virtual inventory model will be updated in real time according to the changes in the actual inventory, so as to realize the real-time simulation of the inventory status; based on the digital twin model, combined with historical data and real-time data, prediction algorithms are used to predict inventory changes, so as to discover the risk of out-of-stock in advance. For example, by analyzing historical sales data and current inventory levels, predict the demand in the next period of time. When it is predicted that the inventory may not be able to meet the demand, issue a warning in a timely manner. At the same time, combining genetic algorithm (GA) and deep reinforcement learning (DRL) to optimize the replenishment strategy for multi-level inventory (such as central warehouse and regional warehouse). The genetic algorithm can search for the optimal replenishment plan by simulating natural selection and genetic mechanisms; deep reinforcement learning can continuously learn and adjust the replenishment strategy through interaction with the environment to achieve the goal of minimizing inventory costs and maximizing service levels; Federated learning is a distributed machine learning method that allows multiple participating parties to jointly train a global model without sharing the original data. In the logistics field, the data of multiple logistics enterprises are combined, and a dynamic path planning model is trained using the federated learning framework. Each logistics enterprise trains its local data locally and then uploads the model parameters obtained from the training to the central server for aggregation to form a global model. In this way, the data resources of each logistics enterprise can be fully utilized to improve the accuracy and generalization ability of the model; the trained dynamic path planning model can dynamically adjust the transportation route according to real-time traffic information, such as road conditions and traffic accidents.When sudden traffic conditions occur, the model can respond quickly, plan the optimal driving route for logistics vehicles, reduce transportation time and costs, and improve logistics efficiency.

[0028] The effects of the above technical solutions are as follows: Through smart contracts, manual intervention is reduced, the efficiency and accuracy of order processing and delivery date changes are improved, and the risk of human errors is reduced; By processing sensor data in real time and optimizing the transportation route, the safety and quality of goods during transportation can be ensured, and damage and deterioration of goods caused by environmental factors can be reduced; The digital twin model enables real-time monitoring and prediction of inventory status, can detect out-of-stock risks in advance, and avoid sales losses and decreased customer satisfaction caused by out-of-stock; Combining genetic algorithms and deep reinforcement learning to optimize the replenishment strategy of multi-level inventory can reduce inventory costs, improve inventory turnover rate, and achieve refined and intelligent inventory management; Using the data of multiple logistics enterprises to train the dynamic path planning model improves the accuracy and adaptability of the model, and can better handle various complex traffic conditions; Through dynamic path planning in the event of sudden traffic conditions, the optimal path can be quickly planned, the transportation efficiency of logistics vehicles can be improved, and the emergency response ability of logistics enterprises can be enhanced. At the same time, federated learning promotes data sharing and collaborative cooperation among logistics enterprises and drives the overall development of the logistics industry.

[0029] In one embodiment of the present invention, S41 includes: Construct a consortium chain smart contract, integrate the three-party nodes of the purchaser, supplier, and logistics provider, and realize the decentralized execution of operations such as order confirmation and delivery date change; And introduce a time lock mechanism to ensure that order changes take effect within a preset time window and avoid supply chain disruptions caused by delayed confirmation; Through the state synchronization oracle, synchronize the order status in the smart contract (such as "confirmed", "in transit") to the ERP and WMS systems in real time; Combine with a rule engine to automatically trigger a compensation process (such as fine calculation, alternative supplier evaluation) for abnormal status (such as delivery date delay exceeding the threshold); Embed the Stackelberg game model in the smart contract to simulate the interest game between the purchaser and the supplier in the delivery date change, automatically generate an acceptable delivery date adjustment plan for both parties, and introduce a reputation mechanism to incorporate the supplier's historical performance into the game parameters to encourage long-term cooperation; Construct a multi-objective optimization model with the goals of minimizing delivery date deviation, minimizing cost, and maximizing customer satisfaction, and generate the optimal delivery date adjustment plan through the non-dominated sorting genetic algorithm (NSGA-II); Deploy multi-modal sensors at logistics nodes (such as warehouses, transport vehicles) to collect environmental data such as temperature, humidity, light, and vibration in real time, as well as transportation data such as vehicle location, speed, and load; Based on the edge computing resource scheduling model, according to the real-time requirements of sensor data (for example, temperature data requires millisecond-level response), computing resources (such as CPUs, GPUs) are dynamically allocated; a lightweight path planning algorithm is deployed on the edge device, and in combination with real-time traffic data (such as congestion index, accident information), the transportation path is dynamically adjusted. The working principle of the above technical solution is as follows: Build a consortium blockchain composed of three-party nodes of the purchaser, supplier, and logistics provider, and deploy smart contracts on the blockchain. The smart contract defines the rules and processes for operations such as order confirmation and delivery date change. When the preset conditions are met, the corresponding operations are automatically executed to achieve decentralized execution, ensuring that all participating parties reach a consensus and complete the transaction without a third-party intermediary; Set a time lock in the smart contract, stipulating that the order change operation must be confirmed within the preset time window. For example, when a delivery date change request is made, the purchaser and the supplier need to confirm within the specified time (such as 24 hours), otherwise the change request will automatically expire. This can avoid supply chain disruptions caused by delayed confirmation and ensure the stability and efficiency of the supply chain; The oracle, as a bridge between the blockchain and external systems (such as ERP and WMS systems), synchronizes the order status in the smart contract (such as "confirmed" and "in transit") to these systems in real time. Ensure data consistency among various internal systems of the enterprise, facilitating unified supply chain management and decision-making for the enterprise; Set a series of rules. When the order status is abnormal (such as the delivery date is delayed beyond the set threshold), the rule engine automatically triggers a compensation process. For example, calculate the fine amount based on the delay time, or start the evaluation process of alternative suppliers to find suppliers who can deliver on time to reduce the losses caused by the delivery date delay; Introduce the Stackelberg game model into the smart contract to simulate the interest game between the purchaser and the supplier in the delivery date change. The purchaser, as the leader, first proposes the requirement for the delivery date change, and the supplier, as the follower, responds according to the purchaser's requirement. The model automatically analyzes the interest demands of both parties and generates a delivery date adjustment plan acceptable to both parties; Incorporate the supplier's historical performance (such as on-time delivery rate, quality qualification rate, etc.) into the game parameters. The higher the supplier's reputation, the more favorable conditions it may obtain in the game, thus motivating the supplier to maintain good performance in the long term and promoting the establishment of a long-term stable cooperation relationship between the two parties; With the goal of minimizing delivery date deviation, minimizing cost, and maximizing customer satisfaction, construct a multi-objective optimization model. Minimizing delivery date deviation can ensure that goods arrive on time and meet customer needs; Minimizing cost can reduce the operating costs of the enterprise; Maximizing customer satisfaction can improve the market competitiveness of the enterprise; Use the non-dominated sorting genetic algorithm (NSGA-II) to solve the multi-objective optimization model. This algorithm can generate a set of non-dominated solutions, that is, among these solutions, no single objective can be further optimized without sacrificing other objectives. The enterprise can select the optimal delivery date adjustment plan from these non-dominated solutions according to the actual situation; Deploy various types of sensors at logistics nodes (such as warehouses and transport vehicles) to collect environmental data (such as temperature, humidity, light, vibration) and transport data (such as vehicle location, speed, load) in real time.These data can comprehensively reflect various situations in the logistics process, providing a basis for subsequent decision-making; according to the real-time requirements of sensor data, computing resources are dynamically allocated. For example, temperature data requires millisecond-level response, so more computing resources (such as CPUs, GPUs) are allocated to tasks processing temperature data. A lightweight path planning algorithm is deployed on edge devices, combined with real-time traffic data (such as congestion index, accident information), to dynamically adjust the transportation route, ensuring that goods can be transported along the optimal path and improving logistics efficiency.

[0030] The effects of the above technical solutions are as follows: The consortium chain smart contract realizes the decentralized execution of operations such as order confirmation and delivery date change, reducing intermediate links and human intervention, and improving the transparency and efficiency of transactions. Each participating party can view the transaction status in real time, ensuring the openness and sharing of information; the time lock mechanism avoids supply chain interruptions caused by delayed confirmation, ensuring the stability and reliability of the supply chain. The purchaser and the supplier can complete the transaction within the specified time, improving the response speed of the supply chain; the state synchronization oracle realizes the real-time data synchronization between the smart contract and the ERP and WMS systems, promoting the collaborative work among various internal systems of the enterprise. The enterprise can adjust the production and logistics plans in a timely manner according to the order status, improving the flexibility of the supply chain; the compensation process is automatically triggered based on the rule engine, which can quickly respond to abnormal situations and reduce losses caused by delivery date delays. The enterprise can choose appropriate compensation measures according to the actual situation, improving the adaptability of the supply chain; by simulating the interest game between both parties, a delivery date adjustment plan acceptable to both parties is generated, and a reputation mechanism is introduced to encourage long-term cooperation between the supplier and the purchaser. In order to obtain a better reputation and more favorable cooperation conditions, the supplier will strive to improve its performance, thus promoting the establishment of a long-term stable cooperation relationship between both parties; considering multiple objectives such as delivery date, cost, and customer satisfaction, an optimal delivery date adjustment plan is generated. The enterprise can choose the optimal plan according to the actual situation, reducing logistics costs and improving customer satisfaction; computing resources are dynamically allocated according to the real-time requirements of sensor data, and the transportation route is adjusted in real time, improving the efficiency and accuracy of logistics transportation. The enterprise can respond to changes in traffic conditions in a timely manner, select the optimal transportation route, and reduce transportation time and costs; various data in the logistics process are collected in real time, enabling potential safety hazards and quality problems to be discovered in a timely manner. For example, the storage temperature of goods can be monitored through temperature sensors to prevent goods from being damaged due to excessive or too low temperature; real-time processing of sensor data on edge devices can quickly respond to abnormal situations and take corresponding measures to ensure the safety and reliability of the logistics process.

[0031] In one embodiment of the present invention, the S42 includes: Based on Internet of Things sensors (such as RFID, UWB positioning tags), real-time collect inventory data such as the location, quantity, and status of inventory (such as shelf life, temperature sensitivity); based on multi-modal data fusion algorithms, combine external data such as historical order data and supplier delivery cycles to construct a time-series feature space of inventory data; Based on the digital twin engine, construct a high-fidelity virtual mirror of the physical inventory to achieve real-time mapping and visualization of the inventory status; introduce digital twin health assessment, and dynamically adjust the model parameters by comparing the deviations between physical and virtual inventory data; Construct a causal graph model (such as a Bayesian network) to analyze the causal relationships between inventory changes and events such as demand fluctuations, supplier delays, and transportation disruptions; based on an event-driven simulator, simulate the dynamic change process of inventory by inputting external events (such as promotional activities, epidemic lockdowns); Introduce a probabilistic graph model (such as a Markov chain) to model uncertain factors such as inventory demand and delivery cycles, and generate multi-scenario inventory forecasts; combine Monte Carlo simulation to evaluate inventory risks (such as stockout probability, overstocking cost) under different scenarios; Decompose the stockout risk into time scales (such as days, weeks, months) and space scales (such as central warehouses, regional warehouses, stores), and construct prediction models respectively; based on spatio-temporal convolutional neural networks (ST-CNN), fuse historical demand data and geospatial information to jointly predict multi-scale stockout risks; Construct a stockout risk index (SOI), comprehensively consider factors such as stockout probability, stockout duration, and stockout cost, quantitatively evaluate the risk, and based on the risk classification strategy, divide the risk into three levels: low, medium, and high according to the SOI value, and trigger different levels of early warning responses; Train a conditional generative adversarial network (cGAN) to learn the normal inventory change pattern, and detect abnormal events (such as sudden demand surges, supplier defaults) by comparing the differences between actual data and generated data; construct a supply chain resilience assessment model to evaluate the risk resistance ability of the supply chain by simulating the impact of different abnormal events (such as natural disasters, policy mutations) on inventory; Combine optimization algorithms to generate strategies to enhance supply chain resilience (such as multi-source procurement, safety stock redundancy); encode replenishment strategies (such as reorder point, replenishment quantity, replenishment frequency) as chromosomes, and construct a fitness function based on minimizing the total cost; Construct a cross-level inventory information sharing platform to achieve real-time synchronization of inventory data among central warehouses, regional warehouses, and stores; based on the inter-level replenishment strategy coordination model, dynamically adjust the replenishment plan according to the inventory status and demand forecast of each level.

[0032] The working principle of the above technical solution is as follows: Using Internet of Things sensors such as RFID and UWB positioning tags to obtain data on the location, quantity, shelf life, temperature sensitivity, etc. of the inventory in real time. These sensors can accurately track the dynamic changes of the inventory, providing basic data support for inventory management; Combining external data such as historical order data and supplier delivery cycles, and using multi-modal data fusion algorithms to construct a temporal feature space of inventory data. By integrating data from different sources and types, potential correlations between data are mined to provide richer information for subsequent analysis and prediction; Based on the digital twin engine, a high-fidelity virtual mirror of the physical inventory is created. This virtual mirror can map the state of the physical inventory in real time and present it in a visual way, enabling managers to intuitively understand the distribution and changes of the inventory; Introducing a digital twin health assessment mechanism, and dynamically adjusting model parameters by comparing the deviations between physical and virtual inventory data. If a large deviation is found, it indicates that the model may be inaccurate and needs to be adjusted in a timely manner to ensure that the virtual mirror can accurately reflect the actual situation of the physical inventory; Constructing causal graph models such as Bayesian networks to analyze the causal relationships between inventory changes and events such as demand fluctuations, supplier delays, and transportation disruptions. By establishing a causal relationship model, the driving factors of inventory changes can be deeply understood, providing a basis for inventory management decisions; Based on an event-driven simulator, input external events (such as promotional activities, epidemic lockdowns), and simulate the dynamic change process of the inventory. By simulating the impact of different events on the inventory, the changing trends of the inventory can be predicted in advance, and corresponding countermeasures can be formulated; Introducing probabilistic graph models such as Markov chains to model uncertain factors such as inventory demand and delivery cycles. By establishing a probability model, the distribution characteristics and changing rules of these uncertain factors can be described; Combining Monte Carlo simulation to generate multi-scenario inventory forecasts and evaluate the inventory risks under different scenarios (such as stockout probability, overstocking cost). By analyzing and evaluating multiple scenarios, the risks faced by the inventory can be comprehensively understood, providing decision support for risk management; Decompose the stockout risk into time scales (such as days, weeks, months) and space scales (such as central warehouses, regional warehouses, stores), and construct prediction models respectively. By considering the influencing factors in different time and space dimensions, the accuracy of stockout risk prediction is improved; Based on the spatio-temporal convolutional neural network (ST-CNN), fuse historical demand data and geospatial information to jointly predict multi-scale stockout risks. The spatio-temporal convolutional neural network can process time and space information simultaneously, mining spatio-temporal features in the data, thereby improving the prediction accuracy; Construct a stockout risk index (SOI), comprehensively considering factors such as stockout probability, stockout duration, and stockout cost, and quantitatively evaluate the risk. By calculating the SOI value, the magnitude of the stockout risk can be intuitively understood; Based on the risk grading strategy, divide the risk into three levels: low, medium, and high according to the SOI value, and trigger different levels of early warning responses.When the risk reaches the corresponding level, the system will automatically issue a warning signal to remind the management to take corresponding measures; train a conditional generative adversarial network (cGAN) to learn the normal inventory change pattern, and detect abnormal events (such as sudden demand surges, supplier defaults) by comparing the differences between actual data and generated data. The cGAN can generate samples similar to normal data, and abnormal situations can be found by comparing the differences between actual data and generated data; build a supply chain resilience assessment model to evaluate the risk resistance ability of the supply chain by simulating the impacts of different abnormal events (such as natural disasters, policy mutations) on inventory. By evaluating the performance of the supply chain under different abnormal events, the weak links of the supply chain can be identified, and corresponding measures can be taken to improve the resilience of the supply chain; combine optimization algorithms to generate strategies for enhancing supply chain resilience (such as multi-source procurement, safety stock redundancy). The optimization algorithm can find the optimal solution according to the goals and constraints of the supply chain to improve the risk resistance ability of the supply chain; encode replenishment strategies (such as reorder point, replenishment quantity, replenishment frequency) as chromosomes, and build a fitness function based on the minimization of the total cost. Through optimization algorithms such as genetic algorithms, search for the optimal replenishment strategy to reduce inventory costs; build a cross-level inventory information sharing platform to achieve real-time synchronization of inventory data in the central warehouse, regional warehouses, and stores. Through the information sharing platform, each level can timely understand the distribution and changes of inventory, providing data support for collaborative replenishment; based on the inter-level replenishment strategy coordination model, dynamically adjust the replenishment plan according to the inventory status and demand forecast of each level. Through collaborative replenishment, reasonable allocation of inventory can be achieved, inventory turnover can be improved, and inventory costs can be reduced.

[0033] The effects of the above technical solutions are as follows: The Internet of Things sensors can collect inventory data in real time to ensure the accuracy and timeliness of inventory information. Managers can understand the latest inventory situation at any time and make decisions in a timely manner; the virtual mirror created by the digital twin engine can map the physical inventory status in real time and present it in a visual way, making inventory management more intuitive and efficient; the causal diagram model can analyze the causal relationship between inventory changes and various events, providing a scientific basis for inventory management decisions. Managers can formulate corresponding strategies based on the causal relationship to improve the effectiveness of inventory management; the probabilistic graph model and Monte Carlo simulation can generate multi-scenario inventory forecasts and evaluate inventory risks under different scenarios. By analyzing and evaluating multiple scenarios, managers can formulate more reasonable inventory strategies to reduce inventory risks; cGAN can detect abnormal events and send out early warning signals in a timely manner. Managers can take corresponding measures based on the early warning signals to reduce the impact of abnormal events on the supply chain; the supply chain resilience assessment model can evaluate the risk resistance ability of the supply chain and generate strategies to enhance the supply chain resilience. By adopting these strategies, the stability and reliability of the supply chain can be improved; the spatio-temporal convolutional neural network can conduct joint prediction of multi-scale out-of-stock risks, helping managers arrange inventory reasonably and reduce out-of-stock risks and overstocking costs; the optimization algorithm can generate the optimal replenishment strategy to reduce the total cost. By adopting the optimal replenishment strategy, the inventory turnover rate can be increased and the inventory cost can be reduced; the cross-level inventory information sharing platform can realize the real-time synchronization of inventory data of the central warehouse, regional warehouses, and stores, promoting information sharing and collaborative decision-making among different levels; the inter-level replenishment strategy coordination model can dynamically adjust the replenishment plan according to the inventory status and demand forecast of each level to achieve reasonable allocation and efficient utilization of inventory.

[0034] In one embodiment of the present invention, the S43 includes: Adopt horizontal federated learning, combine the historical path data of multiple logistics enterprises (such as vehicle trajectories, transportation times, traffic events), and construct a logistics data federation shared across enterprises.

[0035] Based on the multi-modal data alignment algorithm, unify the data formats of different enterprises (such as timestamps, geographical coordinate systems) to achieve spatio-temporal alignment of cross-enterprise data; Introduce a knowledge graph, associate logistics data (such as traffic events, weather information) with geographical entities (such as roads, nodes), and construct a logistics knowledge graph; based on the incentive mechanism of blockchain, automatically allocate the contribution degree of federated learning (such as data volume, model accuracy) through smart contracts to encourage enterprises to participate in data sharing.

[0036] Build a Spatio-Temporal Graph Neural Network (ST-GNN) to model the spatio-temporal correlations of traffic data, capture the temporal evolution patterns and spatial propagation characteristics of traffic events, and based on a traffic event clustering algorithm, automatically identify the types (such as congestion, accidents) and impact scopes of traffic events, and generate a traffic risk heat map; Construct a Traffic Risk Index (TRI), comprehensively consider factors such as traffic event types, durations, and impact scopes, and quantitatively evaluate traffic risks; through a risk grading strategy, divide the risks into low, medium, and high levels according to the TRI value, and trigger different levels of path adjustment strategies; Define the state space (such as current location, traffic risk, time window) and action space (such as path selection, speed adjustment); build a deep neural network (such as A3C, PPO) as the backbone model of the policy network and value network; use logistics enterprises as federated learning nodes, and each node independently trains a local path planning model, and aggregates the global model through the Federated Averaging algorithm (FedAvg); Construct a multi-objective optimization algorithm (such as NSGA-III), balance multiple objectives such as path length, transportation time, and risk exposure, and generate a set of Pareto optimal paths; develop a path selection strategy, and select the optimal path from the set of Pareto optimal paths according to the real-time traffic conditions and user preferences (such as time priority, cost priority); Based on the real-time path adjustment mechanism, when the traffic risk changes, quickly generate a new path planning scheme through the federated learning model and push it to the logistics vehicles in real time.

[0037] The working principle of the above technical solution is as follows: The Horizontal Federated Learning (HorizontalFL) technology is adopted to combine the historical path data (such as vehicle trajectories, transportation times, traffic events) of multiple logistics enterprises. In this way, different logistics enterprises can jointly construct a cross-enterprise shared logistics data federation to realize the joint utilization of data without sharing the original data; Based on the multi-modal data alignment algorithm, the data formats of different enterprises are unified (such as timestamps, geographic coordinate systems). Since the data of different enterprises may adopt different standards and formats, the data alignment algorithm can convert these data into a unified format to achieve the spatio-temporal alignment of cross-enterprise data, providing a basis for subsequent data analysis and modeling; The knowledge graph technology is introduced to associate logistics data (such as traffic events, weather information) with geographical entities (such as roads, nodes) to construct a logistics knowledge graph. The knowledge graph can intuitively display the relationships between logistics data, providing richer information support for logistics decision-making; Based on the blockchain-based incentive mechanism, the contribution degrees of federated learning (such as data volume, model accuracy) are automatically allocated through smart contracts. Enterprises participating in data sharing can obtain corresponding contribution degree rewards, thus motivating enterprises to actively participate in data sharing and improving the quality and scale of the data federation; A Spatio-Temporal Graph Neural Network (ST-GNN) is constructed to perform spatio-temporal correlation modeling on traffic data. The ST-GNN can capture the temporal evolution law and spatial propagation characteristics of traffic events. By analyzing the spatio-temporal characteristics of traffic data, a more accurate model for traffic risk analysis is provided; Based on the traffic event clustering algorithm, the types (such as congestion, accidents) and influence ranges of traffic events are automatically identified, and a traffic risk heat map is generated. The heat map can intuitively display the distribution of traffic risks, helping logistics enterprises plan routes in advance and avoid risk areas; A Traffic Risk Index (TRI) is constructed to comprehensively consider factors such as traffic event types, durations, and influence ranges to quantitatively evaluate traffic risks. Through the risk grading strategy, the risks are divided into three levels: low, medium, and high according to the TRI value, triggering different levels of route adjustment strategies; The state space (such as current location, traffic risk, time window) and action space (such as route selection, speed adjustment) are defined. The state space describes various state information of logistics vehicles during transportation, and the action space defines various actions that the vehicle can take; A deep neural network (such as A3C, PPO) is constructed as the backbone model of the policy network and value network. The policy network is used to generate route planning strategies, and the value network is used to evaluate the advantages and disadvantages of the strategies; Logistics enterprises are used as federated learning nodes. Each node independently trains a local route planning model and aggregates the global model through the Federated Averaging algorithm (FedAvg). Through federated learning, the models of different enterprises can learn from each other and fuse, improving the generalization ability and accuracy of the models; A multi-objective optimization algorithm (such as NSGA-III) is constructed to balance multiple objectives such as route length, transportation time, and risk exposure, generating a Pareto optimal route set.The Pareto optimal path set contains path solutions that perform well under different objectives, providing more choices for logistics enterprises; develop path selection strategies, and select the optimal path from the Pareto optimal path set according to real-time traffic conditions and user preferences (such as time priority, cost priority). By comprehensively considering various factors, select the path solution most suitable for the current situation; based on the real-time path adjustment mechanism, when the traffic risk changes, quickly generate a new path planning solution through the federated learning model and push it to the logistics vehicles in real time. Logistics vehicles can adjust their driving routes in a timely manner according to the new path planning solution, avoid risk areas, and ensure the smooth completion of transportation tasks.

[0038] The effects of the above technical solutions are as follows: Horizontal federated learning realizes data sharing among multiple logistics enterprises, breaks data silos, and enables the full utilization of data from different enterprises. Through joint modeling and analysis, more valuable information can be mined to provide more comprehensive support for logistics decision-making; The multi-modal data alignment algorithm unifies the data formats of different enterprises, solves the problem of inconsistent data formats, and improves the usability and operability of data; The spatio-temporal graph neural network and traffic event clustering algorithm can accurately identify the types and influence ranges of traffic events and generate a traffic risk heat map. Logistics enterprises can plan routes in advance according to the heat map, avoid risk areas, and reduce transportation risks; The traffic risk index quantitatively evaluates traffic risks, enabling logistics enterprises to more intuitively understand the magnitude of traffic risks. The risk grading strategy can take corresponding measures according to different risk levels to improve the ability to respond to traffic risks; The federated learning path planning model combines the data and experience of multiple logistics enterprises, and has stronger generalization ability and accuracy. Through the learning and optimization of deep neural networks, the model can generate more reasonable path planning solutions; The multi-objective optimization algorithm weighs among multiple objectives and generates a Pareto optimal path set. The path selection strategy can select the path solution most suitable for the current situation according to real-time traffic conditions and user preferences, improving the flexibility and personalization of path planning; The real-time path adjustment mechanism can adjust the driving route in a timely manner according to the change of traffic risk, ensuring that logistics vehicles can reach the destination quickly and safely. This helps to improve the efficiency of logistics transportation, reduce transportation time and costs; By considering user preferences (such as time priority, cost priority) and selecting the optimal path solution, it can better meet the needs of users and improve the quality of logistics services and user satisfaction; The blockchain incentive mechanism automatically allocates the contribution degree of federated learning through smart contracts, motivating enterprises to actively participate in data sharing. This helps to promote cooperation among logistics enterprises and form a good data sharing ecosystem.

[0039] In one embodiment of the present invention, the S5 includes: S51. Combine text (policy text), numerical values (market prices), and image (satellite remote sensing image) data to predict risks such as supplier bankruptcy and logistics interruption; S52. Train a risk assessment model with the data of multiple enterprises, use GAN to generate extreme market environments (such as epidemic lockdowns, natural disasters), and test the effectiveness of emergency response plans; S53. Automatically generate emergency response plans for different risks based on the historical case library and reinforcement learning; when a risk occurs, the smart contract automatically triggers emergency measures (such as switching suppliers, adjusting inventory); S54. Real-time simulate the emergency response effect in a virtual environment and dynamically adjust measures.

[0040] The working principle of the above technical solution is as follows: Integrate different types of data such as text (policy text), numerical values (market prices), and images (satellite remote sensing images). Policy texts may contain information such as regulations and policy orientations that affect the operations of suppliers and the logistics industry; market price data reflects the supply and demand relationship and cost changes in the market; satellite remote sensing images can provide intuitive information about the infrastructure status and natural environment changes (such as flood and earthquake affected areas) at the location of the supplier. Through comprehensive analysis of these multi-source data, use machine learning algorithms (such as decision trees, neural networks, etc.) to establish a prediction model to predict risks such as supplier bankruptcy and logistics interruption. Market price fluctuations may reflect the stability issues of the supply chain; abnormal situations in satellite remote sensing images (such as signs of floods around a factory) may indicate the risk of logistics interruption; Combine the data of multiple enterprises to train the risk assessment model. The data of different enterprises contains rich risk scenarios and experiences. By integrating these data, a more comprehensive and accurate risk assessment model can be trained. For example, the data of one enterprise may focus on the impact of market price fluctuations on suppliers, while the data of another enterprise may be more concerned about the impact of policy changes on logistics. Combining these data can comprehensively evaluate risks; Use a generative adversarial network (GAN) to generate extreme market environments, such as pandemic lockdowns, natural disasters, etc. A GAN consists of a generator and a discriminator. The generator is responsible for generating simulated extreme market environment data, and the discriminator determines whether the generated data is real. Through continuous training, the generator can generate realistic extreme market environment data. Then, use the generated data to test the effectiveness of the emergency response plan. For example, in the scenario of a pandemic lockdown, test whether the enterprise can adjust the supply chain in a timely manner to ensure the supply of materials; Collect and organize historical risk event cases, including the handling processes and results of events such as supplier bankruptcy and logistics interruption. These cases form a historical case library, providing a reference for the generation of emergency plans; Based on the historical case library and reinforcement learning algorithms, automatically generate emergency plans for different risks. Reinforcement learning enables an agent to interact with the environment and continuously adjust the emergency plan according to the feedback from the environment (such as the losses after the occurrence of a risk), making it more adaptable to the actual risk scenario; When a risk occurs, the smart contract automatically triggers emergency measures, such as switching suppliers and adjusting inventory. A smart contract is an automatically executable contract based on blockchain technology that can automatically execute corresponding operations when preset conditions are met, improving the efficiency and accuracy of emergency response; Real-time simulate the emergency response effect in a virtual environment. By constructing a virtual environment similar to the actual business scenario, test the generated emergency plan in the virtual environment and observe its execution effect; Dynamically adjust emergency measures according to the simulation results in the virtual environment. For example, if it is found that a certain emergency plan will result in too high transportation costs in the virtual environment, the plan can be adjusted to optimize the emergency response process.

[0041] The effects of the above technical solution are as follows: By integrating multi-source data such as text, numerical values, and images, it is possible to comprehensively evaluate risks such as supplier bankruptcy and logistics interruption from different perspectives, avoiding the limitations of a single data source and improving the accuracy of risk prediction. For example, policy texts may foreshadow industry risks in advance, while market price data can reflect market changes in real time. The combination of the two can more accurately predict risks. Through accurate risk prediction, enterprises can take measures in advance, such as negotiating with suppliers to adjust contract terms and increasing inventory buffers, to reduce the likelihood and impact of risks. Using GAN to generate extreme market environments for testing enables enterprises to make preparations in advance and improve their emergency response capabilities in real extreme situations. For example, in the simulation test of the epidemic lockdown, enterprises can optimize the supply chain layout and establish a backup supplier network to ensure a quick adjustment of the supply chain in the event of a real epidemic. The intelligent contract automatically triggers emergency measures, reducing the time and error of human intervention and improving the efficiency and accuracy of emergency response. For example, when the risk of supplier bankruptcy occurs, the intelligent contract can automatically switch to a backup supplier to ensure the continuity of logistics operations. The emergency response plan generated based on the historical case library and reinforcement learning is more in line with the actual business scenario, with stronger operability and pertinence. Reinforcement learning can continuously optimize the plan according to different risk scenarios, improving the quality of the emergency response plan. Real-time simulation of the emergency response effect in a virtual environment and dynamic adjustment of measures can make the emergency response plan more flexible and adaptable, and can adjust strategies in a timely manner according to the actual situation to improve the effect of emergency response. Training the risk assessment model by combining the data of multiple enterprises promotes data sharing and cooperation among enterprises, forming a more comprehensive risk assessment system. Different enterprises can give play to their respective data advantages and jointly improve the accuracy of risk assessment. This technical solution promotes the application of new technologies such as machine learning, GAN, reinforcement learning, and intelligent contracts in the logistics industry, promotes technological innovation and development in the logistics industry, and improves the overall competitiveness of the logistics industry.

[0042] An embodiment of the present invention, a procurement supply chain collaborative intelligent management system, includes a memory, a processor, and a computer program stored on the memory and executable on the memory. The processor executes the program to implement the procurement supply chain collaborative intelligent management method as described in any one of the above.

[0043] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these changes and modifications.

Claims

1. A procurement supply chain collaborative intelligent management method, characterized in that: The method comprises: S1. Obtain enterprise-related data; S2. Build a procurement demand forecasting model; dynamically adjust procurement strategies based on forecasting results; S3, storing the records and evaluation results on the blockchain; S4. The system automatically sends order information to suppliers; suppliers confirm orders, update production progress and logistics information through the blockchain platform; the system adjusts inventory strategies in real time; and monitors logistics status in real time; S5. The system analyzes data from all links of the supply chain in real time to identify potential risks.

2. The procurement supply chain collaborative intelligent management method according to claim 1 is characterized in that: Said S1 comprises: S11, parse text data and extract key information; S12, crawling data; S13, identify and correct outliers; S14. Build a multi-dimensional graph model of supplier-product-time, integrating historical procurement data with external market dynamics.

3. The procurement supply chain collaborative intelligent management method according to claim 1 is characterized in that: The S2 comprises: S21. Build a multimodal Transformer model. S22. Use historical procurement task data to train the model's environmental adaptability; S23. Simulate the supply chain performance under different procurement strategies in a virtual environment.

4. The procurement supply chain collaborative intelligent management method according to claim 1 is characterized in that: The S3 includes: S31. Verify the qualification certificates submitted by suppliers; S32. Synchronize supplier reputation scores to multiple industry alliance chains; S33. Combine the time decay factor and transaction volume weight to dynamically adjust the supplier reputation score; S34. Deploy lightweight recommendation models on local edge devices to respond to procurement needs in real time.

5. The procurement supply chain collaborative intelligent management method according to claim 1 is characterized in that: The S4 comprises: S41. Automatically execute operations through smart contracts; process sensor data in real time through edge devices deployed at logistics nodes, and optimize transportation routes based on the processing results; S42. Build digital twin models of physical and virtual inventory, simulate inventory changes in real time, and predict out-of-stock risks; combine genetic algorithms with deep reinforcement learning to optimize replenishment strategies for multi-level inventory; S43. Use the federated learning framework to combine data from multiple logistics companies and train a dynamic path planning model.

6. The procurement supply chain collaborative intelligent management method according to claim 5 is characterized in that: The S41 includes: Build a consortium chain smart contract and introduce a time lock mechanism; Through the state synchronization oracle, the order status in the smart contract is synchronized to the ERP and WMS systems in real time; combined with the rule engine, the compensation process is automatically triggered for abnormal status; The Stackelberg game model is embedded in the smart contract to simulate the interest game between the purchaser and the supplier in the change of delivery date, and automatically generate a delivery date adjustment plan acceptable to both parties; Construct a multi-objective optimization model and generate the optimal delivery adjustment plan through a non-dominated sorting genetic algorithm; deploy multimodal sensors at logistics nodes to collect environmental data and transportation data in real time; Based on the edge computing resource scheduling model, computing resources are dynamically allocated according to the real-time requirements of sensor data. A lightweight path planning algorithm is deployed on edge devices, and the transportation route is dynamically adjusted in combination with real-time traffic data.

7. The procurement supply chain collaborative intelligent management method according to claim 5 is characterized in that: The S42 includes: Based on IoT sensors, inventory data is collected in real time. Based on multimodal data fusion algorithms and combined with external data, the time series feature space of inventory data is constructed. Based on the digital twin engine, a high-fidelity virtual image of the physical inventory is constructed, and the health assessment of the digital twin is introduced. By comparing the deviation between the physical and virtual inventory data, the model parameters are dynamically adjusted; Construct a causal graph model; based on an event-driven simulator, simulate the dynamic change process of inventory by inputting external events; Introduce probabilistic graphical models to model uncertainty factors and generate multi-scenario inventory forecasts; combine Monte Carlo simulation to evaluate inventory risks under different scenarios; Decompose the out-of-stock risk into time scale and space scale, and construct prediction models for each. Based on spatiotemporal convolutional neural network, integrate historical demand data and geographic spatial information to conduct joint prediction of multi-scale out-of-stock risk. Construct a stock-out risk index to quantitatively assess the risk. Based on the risk grading strategy, divide the risk into three levels: low, medium, and high according to the SOI value, and trigger different levels of early warning responses; Train conditional generative adversarial networks to detect abnormal events by comparing the differences between actual data and generated data; build a supply chain resilience assessment model to evaluate the risk resistance of the supply chain by simulating the impact of different abnormal events on inventory; Combined with optimization algorithms, strategies to improve supply chain resilience are generated; replenishment strategies are encoded as chromosomes, and a fitness function based on total cost minimization is constructed; Build a cross-level inventory information sharing platform to achieve real-time synchronization of central warehouse, regional warehouse and store inventory data; based on the inter-level replenishment strategy collaborative model, dynamically adjust the replenishment plan according to the inventory status and demand forecast of each level.

8. The procurement supply chain collaborative intelligent management method according to claim 5 is characterized in that: The S43 includes: Adopt horizontal federated learning, combine the historical path data of multiple logistics companies, and build a cross-enterprise shared logistics data federation; based on the multimodal data alignment algorithm, unify the data formats of different companies and achieve the spatiotemporal alignment of cross-enterprise data; Introduce knowledge graphs, associate logistics data with geographic entities, and build logistics knowledge graphs; Construct a spatiotemporal graph neural network to model the spatiotemporal correlation of traffic data, capture the temporal evolution and spatial propagation characteristics of traffic events, automatically identify the type and impact range of traffic events based on the traffic event clustering algorithm, and generate a traffic risk heat map; Construct a traffic risk index to quantitatively assess traffic risks; through risk grading strategies, divide risks into three levels: low, medium, and high according to the TRI value, triggering different levels of route adjustment strategies; Define the state space and action space; use logistics enterprises as federated learning nodes, each node independently trains the local path planning model, and aggregates the global model through the federated averaging algorithm; Construct a multi-objective optimization algorithm to generate a Pareto optimal path set; select the optimal path from the Pareto optimal path set based on real-time traffic conditions and user preferences; Based on the real-time path adjustment mechanism, when traffic risks change, new path planning plans are quickly generated through the federated learning model and pushed to logistics vehicles in real time.

9. The procurement supply chain collaborative intelligent management method according to claim 1 is characterized in that: The S5 comprises: S51. Predict supplier risks; S52. Use GAN to generate extreme market environments; S53. Generate emergency plans for different risks; S54. Simulate emergency response effects in real time in a virtual environment and adjust measures dynamically.

10. A procurement supply chain collaborative intelligent management system, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the memory, wherein the processor executes the program to implement the procurement supply chain collaborative intelligent management method as described in any one of claims 1 to 9.

Images