A supply chain collaborative material management system and method

Abstract

The application provides a supply chain collaborative material management system and method, and relates to the technical field of material management.The method comprises the following steps: receiving a multi-enterprise heterogeneous material data stream, constructing an industry knowledge graph based on field-level semantic mapping of a dynamic ontology library, and generating a standardized data table; injecting the standardized data into a distributed message queue for parallel processing, creating a dynamic cache partition according to a material object label and internally storing a timestamp index, scanning the data by using a sliding window mechanism, and executing multi-strategy resolution on conflicting data in the window; jointly analyzing the cleaned data, generating an enhanced analysis report, activating a case transfer learning mechanism to update the model when the similarity between new data and historical cases is higher than a threshold value; constructing a material state transition matrix for abnormal data, simulating and correcting operations in a digital twin environment, and triggering automatic correction; solving the problems of multi-enterprise heterogeneous data integration, dynamic data processing, intelligent analysis and the like, and improving the collaborative efficiency and management level of the supply chain.

Description

Technical Field

[0001] This application relates to the field of materials management technology, and in particular to a supply chain collaborative materials management system and method. Background Technology

[0002] With the deepening of global economic integration and the widespread application of digital technology, supply chain management has become a core area for enterprises to enhance their competitiveness. In the modern supply chain system, material management involves the collaborative operation of multiple enterprises and multiple links, and faces many challenges such as strong data heterogeneity, complex information interaction, and high real-time requirements.

[0003] Traditional supply chain materials management models have significant limitations. On the one hand, differences in business systems, data standards, and coding rules among different enterprises lead to heterogeneous material data flows, making efficient integration and sharing difficult. This data silo phenomenon severely hinders information flow between various links in the supply chain and reduces collaboration efficiency. For example, the inconsistent data formats of suppliers, manufacturers, and logistics companies make it impossible to accurately match and correlate key data such as material categories, supplier information, and logistics routes, increasing the difficulty and cost of materials management.

[0004] On the other hand, existing materials management methods lack effective strategies for dealing with dynamically changing supply chain environments. The entire lifecycle of materials involves multiple stages such as procurement, production, transportation, and warehousing, and the status and data of each stage are constantly changing. Traditional methods struggle to capture and analyze this dynamic information in real time, making it difficult to promptly identify potential risks and anomalies. Furthermore, in areas such as conflict data processing, anomaly identification, and correction, traditional methods often rely on human experience and simple rules, lacking intelligent and automated processing capabilities, and thus failing to meet the demands of efficient and precise management in modern supply chains.

[0005] Furthermore, with the continuous expansion and increasing complexity of the supply chain, higher demands are placed on the interpretability and decision support capabilities of materials management. Enterprises need to clearly understand the basis and logic of materials management decisions in order to make scientific and reasonable decisions. However, existing technologies are insufficient in providing interpretability analysis and intelligent decision support, and cannot provide enterprises with comprehensive and in-depth materials management solutions.

[0006] Therefore, there is an urgent need for an innovative supply chain collaborative material management method that can effectively solve problems such as the integration of heterogeneous data from multiple enterprises, dynamic data processing, intelligent analysis and decision-making, and improve the collaborative efficiency and management level of the supply chain. Summary of the Invention

[0007] To address the aforementioned issues, this application proposes a supply chain collaborative materials management system and method. Through multi-enterprise heterogeneous data integration, dynamic data intelligent processing, precise risk control and decision-making, continuous optimization, and automatic correction, it achieves high efficiency, accuracy, scientific rigor, and intelligence in supply chain collaborative materials management, effectively improving the collaborative efficiency and management level of the supply chain.

[0008] The objective of this application is achieved through the following technical solution:

[0009] In a first aspect, embodiments of this application provide a supply chain collaborative material management method, the method comprising:

[0010] It receives heterogeneous material data streams from multiple enterprises, constructs an industry knowledge graph containing material categories, supplier relationships, and logistics routes based on field-level semantic mapping of a dynamic ontology library, and generates standardized data tables with spatiotemporal joint indexes.

[0011] Standardized data is injected into a distributed message queue for parallel processing. Dynamic cache partitions are created according to material object tags and have built-in timestamp indexes. A sliding window mechanism is used to scan data and multiple strategies are used to resolve conflicting data within the window.

[0012] Joint analysis is performed on the cleaned data. Vertically, an LSTM-GNN hybrid model is constructed to capture the characteristics of the entire life cycle of materials. Horizontally, an interpretable business rule engine is embedded to define constraints and quantify the contribution of each rule to the anomaly detection.

[0013] Generate enhanced analysis reports; when the similarity between new data and historical cases exceeds a threshold, activate the case transfer learning mechanism to update the model.

[0014] A material status transition matrix is ​​constructed for abnormal data, and correction operations are simulated and automatic corrections are triggered in a digital twin environment.

[0015] Secondly, this application provides a supply chain collaborative material management system for use with the supply chain collaborative material management method described in the embodiments of this application, the system comprising:

[0016] The data standardization module receives heterogeneous material data streams from multiple enterprises, constructs an industry knowledge graph containing material categories, supplier relationships, and logistics routes based on field-level semantic mapping of a dynamic ontology library, and generates standardized data tables with spatiotemporal joint indexes.

[0017] The data storage and resolution module is used to inject standardized data into a distributed message queue for parallel processing, create dynamic cache partitions based on material object tags and have built-in timestamp indexes, scan data using a sliding window mechanism, and perform multi-strategy resolution on conflicting data within the window.

[0018] The joint analysis module is used to perform joint analysis on the cleaned data. Vertically, it constructs an LSTM-GNN hybrid model to capture the characteristics of the entire life cycle of materials, and horizontally embeds an interpretable business rule engine to define constraints and quantify the contribution of each rule to the anomaly detection.

[0019] The report generation module is used to generate enhanced analysis reports. When the similarity between new data and historical cases exceeds a threshold, the case transfer learning mechanism is activated to update the model.

[0020] The anomaly handling module is used to construct a material status transition matrix for abnormal data, simulate correction operations in the digital twin environment, and trigger automatic correction.

[0021] The beneficial effects of this invention include: through field-level semantic mapping in a dynamic ontology library, it can receive heterogeneous material data streams from multiple enterprises, establish mapping relationships between data with the same semantics but different field names, and use an ontology inference engine for consistency checks and conflict resolution, thereby constructing an industry knowledge graph and standardized data tables. This breaks down data barriers between enterprises, achieves efficient integration and sharing of multi-source heterogeneous data, provides a unified and accurate data foundation for subsequent material management, and improves the efficiency of information interaction in all links of the supply chain; standardized data is injected into a distributed message queue for parallel processing, dynamic cache partitions are created according to material object tags and have built-in timestamp indexes, and a sliding window mechanism is used to scan data, performing multi-strategy resolution on conflicting data within the window, enabling… The system rapidly and accurately processes dynamically changing data in the supply chain, effectively resolving data conflicts and ensuring data accuracy and consistency. It constructs a multi-dimensional assessment system for material characteristics, determining the weights of each indicator and calculating the comprehensive characteristic value of materials through principal component analysis. Based on this, it enters a three-level partitioning decision-making logic, designing differentiated index enhancement strategies for partitions with different risk levels to achieve precise control over material risks. Simultaneously, it vertically constructs an LSTM-GNN hybrid model to capture the full lifecycle characteristics of materials, and horizontally embeds an interpretable business rule engine to define constraints and quantify rule contributions. When the deep learning model and rule engine's judgment results conflict, it traces and verifies the source, generating accurate anomaly judgment conclusions to provide scientific and precise decision support for material management. When the similarity between new data and historical cases exceeds a threshold, it activates a case transfer learning mechanism to update the model, enabling the system to continuously adapt to changes in the supply chain environment and continuously optimize management strategies. It constructs a material state transition matrix for abnormal data, simulating correction operations in a digital twin environment. When preset conditions are met, it automatically executes correction operations and synchronously updates the status of all network nodes, achieving automation and intelligence in material management and improving the emergency response capability and management efficiency of the supply chain. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of a supply chain collaborative material management method provided in an embodiment of this application. Detailed Implementation

[0023] The present application will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0024] See Figure 1 This application provides a supply chain collaborative material management method, the method comprising:

[0025] It receives heterogeneous material data streams from multiple enterprises, constructs an industry knowledge graph containing material categories, supplier relationships, and logistics routes based on field-level semantic mapping of a dynamic ontology library, and generates standardized data tables with spatiotemporal joint indexes.

[0026] Standardized data is injected into a distributed message queue for parallel processing. Dynamic cache partitions are created according to material object tags and have built-in timestamp indexes. A sliding window mechanism is used to scan data and multiple strategies are used to resolve conflicting data within the window.

[0027] Joint analysis is performed on the cleaned data. Vertically, an LSTM-GNN hybrid model is constructed to capture the characteristics of the entire life cycle of materials. Horizontally, an interpretable business rule engine is embedded to define constraints and quantify the contribution of each rule to the anomaly detection.

[0028] Generate enhanced analysis reports; when the similarity between new data and historical cases exceeds a threshold, activate the case transfer learning mechanism to update the model.

[0029] A material status transition matrix is ​​constructed for abnormal data, and correction operations are simulated and automatic corrections are triggered in a digital twin environment.

[0030] The working principle and beneficial effects of the above technical solution are as follows:

[0031] In supply chain scenarios, different enterprises exhibit variations in data formats and semantics, resulting in heterogeneous material data flows. The process begins by receiving this data and performing field-level semantic mapping using a dynamic ontology library. This dynamic ontology library contains rich semantic information and mapping rules, capable of identifying fields with identical or similar meanings across different enterprises' data and converting them into standardized semantic representations. Based on this, an industry knowledge graph is constructed. This graph integrates key information such as material categories, supplier relationships, and logistics routes, clearly presenting the connections between various elements of the supply chain in a graphical manner, providing a structured knowledge foundation for subsequent analysis. Simultaneously, standardized data tables with spatiotemporal joint indexes are generated. These indexes organize the data, facilitating rapid location and retrieval of material data within specific time and spatial ranges, achieving orderly data storage and efficient management.

[0032] Standardized data is injected into a distributed message queue, leveraging the parallel computing capabilities of the distributed system to achieve rapid data processing. Dynamic cache partitions are created based on material object tags, with built-in timestamp indexes. This storage structure allows for categorized caching of data according to material type, while the timestamp index facilitates data processing in chronological order. A sliding window mechanism is employed to scan the data, with the sliding window monitoring and analyzing the data within the window in real time at regular intervals or in units of data volume. When conflicting data is detected, a multi-strategy resolution process is implemented. By appropriately selecting and combining strategies, data conflicts are eliminated, ensuring data accuracy and consistency.

[0033] Joint analysis is performed on the cleaned data, constructing a vertical LSTM-GNN hybrid model. Long Short-Term Memory (LSTM) networks excel at processing sequential data and can learn the dynamic changes of materials over time; Graph Neural Networks (GNNs) can uncover the relationship networks between materials. The combination of the two comprehensively captures the characteristics of the entire lifecycle of materials, enabling in-depth analysis of all stages from production and transportation to storage. A horizontally embedded interpretable business rule engine transforms the company's business rules and industry standards into executable constraints. This engine not only validates the material data according to rules but also quantifies the contribution of each rule to anomaly detection, making the anomaly detection process transparent and interpretable, facilitating understanding and decision-making for managers.

[0034] An enhanced analysis report is generated, containing comprehensive analysis results of material data, early warnings of anomalies, and other key information. When the similarity between new data and historical cases exceeds a set threshold, a case transfer learning mechanism is activated. This mechanism transfers successful processing experience and model parameters from historical cases to the processing of new data, rapidly updating the model to adapt to emerging data patterns and business needs, thereby improving the model's generalization ability and predictive accuracy.

[0035] For abnormal data, a material state transition matrix is ​​constructed, which describes the probability and conditions of material transitions between different states. In a digital twin environment, the process of material state changes is simulated to rehearse various corrective operations and evaluate the effectiveness of different corrective strategies. Once the optimal corrective solution is determined, an automatic correction mechanism is triggered to adjust the material state and business processes in the actual supply chain, enabling rapid response and handling of abnormal situations and ensuring the stable operation of the supply chain.

[0036] In one possible implementation, the field-level semantic mapping method of the dynamic ontology library includes:

[0037] Match the predefined semantic models in the dynamic ontology library with the fields of heterogeneous material data streams from multiple enterprises, and establish semantic mapping relationships for data with the same semantics but different field names;

[0038] Natural language processing techniques are used to perform semantic analysis on text fields in the data stream, extract key semantic information, and map it to the corresponding concepts in the dynamic ontology library.

[0039] The ontology inference engine is used to perform consistency checks and conflict resolution on semantic mapping relationships to ensure the accuracy of semantic mapping.

[0040] In one possible implementation, the construction of an industry knowledge graph containing material categories, supplier relationships, and logistics routes includes:

[0041] Extract material attribute information from heterogeneous material data streams from multiple enterprises, classify materials based on attribute information, and construct material category nodes;

[0042] Extract basic information and transaction records of suppliers to construct supplier nodes, and connect supplier nodes with material category nodes through edges to represent the supply relationship between suppliers and materials;

[0043] Obtain information such as logistics routes, times, and means of transport, construct logistics path nodes, and connect material category nodes with logistics path nodes through edges to represent the transportation path of materials;

[0044] The industry knowledge graph is stored and managed using a graph database, and indexes of nodes and edges are built to improve the query and retrieval efficiency of the knowledge graph.

[0045] In one possible implementation, generating a standardized data table with a spatiotemporal joint index using a transfer learning model includes:

[0046] Choose a transfer learning model that has been trained in a relevant field as the base model;

[0047] The heterogeneous material data streams from multiple enterprises are divided according to time and space dimensions, and time and space features are extracted.

[0048] The extracted time features, spatial features, and material data are input into the basic model for fine-tuning and training, so that the model can adapt to the field differences of different enterprises.

[0049] In the generated standardized data table, a spatiotemporal joint index based on timestamps and geographic locations is established to enable quick querying and retrieval of material data within a specific time and spatial range.

[0050] The working principle and effect of the above technical solution are as follows: In supply chain collaboration scenarios, material data flows from different enterprises may have different field names and formats but the same semantic meaning. First, the predefined semantic model in the dynamic ontology library is matched with the fields of the heterogeneous material data flows from multiple enterprises. The semantic model acts like a "translation dictionary," establishing a semantic mapping relationship between the two when encountering data with the same semantic meaning but different field names from different enterprises, thus achieving a unified conversion of data semantics. Second, for text fields in the data flow, natural language processing technology is used for semantic analysis. Through operations such as word segmentation, part-of-speech tagging, and semantic role tagging, key semantic information is extracted and mapped to the corresponding concepts in the dynamic ontology library, integrating unstructured text data into the semantic system. Finally, the established semantic mapping relationship is checked for consistency and conflict resolution using an ontology inference engine. The ontology inference engine judges whether there are contradictions in the mapping relationship based on the logical rules of the ontology. If so, adjustments are made to ensure the accuracy of the semantic mapping, laying a solid foundation for subsequent data processing.

[0051] From the heterogeneous material data flow of multiple enterprises, the first step is to extract the attribute information of the materials, such as specifications, models, and uses. Based on these attributes, the materials are classified, and a material category node is constructed for each category, forming the basic elements of the knowledge graph. Next, basic supplier information (such as name, address, and contact information) and transaction records are extracted to construct supplier nodes. Edges connect supplier nodes to material category nodes, and edge attributes can include information such as transaction quantity and transaction time, clearly representing the supply relationship between suppliers and materials. Then, information such as logistics routes, times, and means of transport is obtained to construct logistics path nodes. Similarly, edges connect material category nodes to logistics path nodes, visually presenting the transportation path of materials. Finally, a graph database is used to store and manage the entire industry knowledge graph. Graph databases excel at handling the relationship data between nodes and edges, and indexes are created for nodes and edges. When querying and retrieving the knowledge graph, the target nodes and edges can be quickly located based on the indexes, improving data access efficiency.

[0052] First, a pre-trained transfer learning model in a relevant field is selected as the base model, possessing certain data processing and feature extraction capabilities. The heterogeneous material data streams from multiple enterprises are divided according to time (e.g., order time, transportation time) and space (e.g., production location, transportation route, storage location), extracting time features (e.g., timestamps, time periods) and spatial features (e.g., latitude and longitude, area codes). The extracted time and spatial features, along with the material data, are input into the base model for fine-tuning training. By adjusting model parameters, the model adapts to the differences in fields across different enterprises, learning the commonalities and characteristics of multi-enterprise data, thereby generating a standardized data table. Finally, a spatiotemporal joint index based on timestamps and geographic location is established in the standardized data table, combining time and space dimensions. This enables rapid location and retrieval of material data within a specific time and space range, meeting the needs of real-time supply chain monitoring and analysis.

[0053] In one possible implementation, S2 includes:

[0054] A multi-dimensional evaluation system for material characteristics was constructed. Principal component analysis was used to determine the weights of each primary indicator and to calculate the comprehensive characteristic value of the materials.

[0055] Based on the comparison between the comprehensive feature value and the preset threshold, the three-level partitioning decision logic is entered;

[0056] Based on the timestamp index, differentiated index enhancement strategies are designed for partitions with different risk levels. The differentiated index enhancement strategies include: adding risk event association indexes and supplier change record indexes to high-risk partitions; and retaining only the basic timestamp index for medium and low-risk partitions, but regularly performing index fragmentation defragmentation and performance optimization.

[0057] In one possible implementation, the first indicator includes supplier credit score, historical delivery delay rate, material transaction frequency, real-time inventory turnover rate, material importance, and inventory volatility; the step of determining the weight of each indicator through principal component analysis and calculating the comprehensive characteristic value of the material includes:

[0058] Calculate the covariance matrix of the index, and obtain eigenvalues ​​and eigenvectors through eigenvalue decomposition;

[0059] Principal components are selected based on the cumulative contribution rate of eigenvalues, and the eigenvectors corresponding to the principal components are used as the weights of each indicator.

[0060] The working principle and effects of the above technical solution are as follows:

[0061] When constructing a multidimensional assessment system for material characteristics, the various attributes and influencing factors involved in the supply chain are considered. Multiple dimensions are selected as primary indicators. Principal component analysis (PCA) transforms these indicators linearly into a few independent composite indicators (principal components). These principal components retain as much information as possible from the original indicators. During this process, the variance contribution rate of each principal component is calculated to determine the weight of each primary indicator in the comprehensive assessment. A higher variance contribution rate indicates greater importance of the indicator in describing the material characteristics. Finally, based on the weights and actual values ​​of each indicator, the comprehensive characteristic value of the material is calculated. This value serves as a quantitative indicator for measuring the overall characteristics of the material, providing data support for subsequent decision-making.

[0062] The calculated comprehensive characteristic value of the materials is compared with preset thresholds. These preset thresholds are reference values ​​pre-set based on industry experience, historical enterprise data, and supply chain management objectives. When the comprehensive characteristic value is higher than a certain higher threshold, the material is classified as high-risk; if it falls within the middle range, it is classified as medium-risk; and if it falls below a lower threshold, it is classified as low-risk. Different risk levels correspond to different management strategies and resource allocation methods. After entering the three-level zoning decision-making logic, the system will automatically match the corresponding processing procedures and measures according to the risk level of the materials, achieving differentiated management of materials.

[0063] Building upon the timestamp index, differentiated index enhancement strategies are implemented for partitions with different risk levels. For materials in high-risk partitions, which may face various risks such as supply disruptions, price fluctuations, and quality issues, risk event association indexes and supplier change record indexes are added. The risk event association index records various risk events that occur during the supply chain process, such as transportation delays caused by natural disasters and supplier production accidents. This index allows for quick querying of risk events related to specific materials, facilitating timely risk understanding and response by management. The supplier change record index records information such as supplier changes and contract modifications, helping enterprises grasp the dynamic changes in the upstream supply chain and assess potential risks. For materials in medium- and low-risk partitions, whose risks are relatively low and stable, only the basic timestamp index is retained, while index fragmentation and performance optimization are performed regularly. Index fragmentation occurs during data insertion, deletion, and updates, affecting query efficiency. Regularly defragmenting the index restores its storage structure, improves data query performance, and ensures that computing resources are allocated reasonably while meeting management needs, thereby reducing management costs.

[0064] In one possible implementation, the three-level partitioning decision logic includes:

[0065] If the comprehensive feature value is greater than or equal to the first threshold (high threshold), and the supplier credit score is less than the preset credit score, and the delivery delay rate within the preset time period is greater than the delay rate threshold (e.g., 20%), then the material data will be partitioned independently, and the supplier risk warning process will be initiated to increase the monitoring frequency of the supplier's data.

[0066] If the second threshold is less than the comprehensive feature value and less than the first threshold, and the material transaction frequency is greater than x*historical average (x>1, for example, x=1.5), and the inventory turnover rate is less than the safety threshold, then it is given priority to merge with the same high-frequency materials in the same partition, but the data of this material is isolated and marked to avoid read and write conflicts with other data.

[0067] If the comprehensive feature value is less than or equal to the second threshold, a clustering algorithm (such as K-Means) is used to cluster the materials. Based on the clustering results, the partitions are dynamically merged and a partition association network is established. The data interaction between each partition is monitored in real time. Once abnormal cross-partition data linkage is detected, the partitioning strategy is re-evaluated.

[0068] In the three-level partitioning decision logic, if only some conditions are met (e.g., only a single condition is triggered), then refined processing is performed through a hierarchical response mechanism and a dynamic weighting strategy, including:

[0069] If only the comprehensive characteristic value is greater than or equal to the first threshold, the material will be classified into the "high-concern zone," but it will be allowed to coexist with similar low-risk materials, and the entire lifecycle tracking of the material will be triggered (full-link monitoring from procurement to delivery); and,

[0070] Set a dynamic sampling frequency, for example, increase the quality inspection ratio of each batch to 30% (for example, the baseline is 10%).

[0071] If only the supplier credit score is less than the preset credit score, then add a "supplier risk label" to the existing partition, but maintain the existing partition structure.

[0072] If only the delivery delay rate is greater than the delay rate threshold, the delayed materials will be classified into the "logistics emergency zone," which is independent of the regular logistics data flow.

[0073] The working principle and effects of the above technical solution are as follows:

[0074] For materials meeting the high-risk criteria (comprehensive characteristic value ≥ first threshold, supplier credit score < preset credit score and delivery delay rate within a preset time period > delay rate threshold), they are independently zoned and a supplier risk warning process is initiated, increasing monitoring frequency. This allows for timely detection of potential supplier risks, such as broken cash flow or reduced production capacity, enabling proactive countermeasures such as finding alternative suppliers and adjusting procurement plans to avoid supply disruptions due to supplier issues and reduce supply chain operational risks. When only a single high-risk criterion is met, such as a comprehensive characteristic value ≥ first threshold, materials are classified into a "high-concern zone" and tracked throughout their entire lifecycle. This increases the quality inspection rate, effectively monitoring material quality and supply stability, promptly identifying and resolving potential problems, and comprehensively ensuring supply chain security.

[0075] In medium-risk scenarios (second threshold < comprehensive feature value < first threshold, and material transaction frequency > x * historical average, while inventory turnover rate < safety threshold), priority is given to merging partitions with similar high-frequency materials to reduce the number of data partitions and lower data management complexity. Isolate and mark these materials for storage to avoid data read / write conflicts and ensure data processing efficiency. For low-risk materials (comprehensive feature value ≤ second threshold), a clustering algorithm is used to dynamically merge partitions and establish a partition association network, monitoring data interaction in real time. This approach allows for flexible partition adjustments based on material characteristics, avoiding resource waste and achieving efficient utilization of storage resources. Simultaneously, the dynamically adjusted data management model enables the system to quickly respond to changes in material data, improving overall management efficiency.

[0076] Under the tiered response mechanism and dynamic weighting strategy, the handling of situations where only some conditions are met enables refined data management. Adding a "supplier risk tag" only when the supplier's credit score is lower than a preset credit score, and classifying delayed materials into a "logistics emergency zone" only when the delivery delay rate exceeds a delay rate threshold—this precise data classification and tagging management allows enterprises to quickly locate and handle data on specific problematic materials. Data partitioning, isolation, and labeling prevent interference between data on materials of different risk levels, ensuring data security and stability. It also facilitates enterprises in developing targeted management strategies based on different data characteristics, improving the accuracy and effectiveness of data processing.

[0077] By conducting multi-dimensional analysis and zoning management of material data, enterprises can gain a clearer understanding of the status and risks of materials in the supply chain, providing rich and accurate data support for decisions such as procurement, inventory management, and supplier evaluation.

[0078] In one possible implementation, S2 includes:

[0079] The basic window width is determined based on the material turnover cycle, real-time data flow rate, and the number of supply chain levels.

[0080]

[0081] in, represents the current base window width (the time span of the sliding window at the current moment); where m is the number of material types (i.e., the total number of material types involved in the current window); and i is the material type index. The standard turnover cycle for the i-th type of material; Let be the number of supply chain levels for the i-th type of material; Let be the weight of the i-th material in the current window; Forgetting factor, ∈[0.8,1.2]; QPS is the data flow rate; t is the current time point (in hours);

[0082] Specifically, the weight of the i-th type of material in the current window is determined by its importance score and the proportion of emergency orders for the i-th type of material in the total orders within the current window. ;

[0083] In one possible implementation, the weight of the i-th material in the current window is determined by the following formula:

[0084]

[0085] Based on the second indicator of the current data, the base window width is dynamically adjusted to obtain the adjusted window width; the second indicator includes the data update frequency, the data fluctuation coefficient, and the proportion of abnormal data.

[0086]

[0087] in, This is the adjusted window width. Score the second indicator for the current data. The first coefficient is 0.7 < <1, The second coefficient is 1 < <1.3, The first threshold for the second indicator score of the preset data. The second threshold for the preset data's second indicator score;

[0088] In one possible implementation, the score for the second indicator is determined by the data update frequency, the data fluctuation coefficient, and the proportion of abnormal data.

[0089] In one possible implementation, a second indicator score for the data is obtained by normalizing the data update frequency, data fluctuation coefficient, and proportion of abnormal data, and then taking a weighted average.

[0090] The working principle and effects of the above technical solution are as follows:

[0091] The determination of the basic window width integrates key factors such as material turnover cycle, real-time data flow rate, and the number of supply chain levels. The standard turnover cycle of materials reflects the time required for the materials to complete one cycle from procurement to delivery in the supply chain. Materials with a long turnover cycle mean that they stay in the supply chain for a long time and require a wider window to cover their entire process. The number of supply chain levels reflects the complexity of the material circulation process. The more levels there are, the more complex the enterprises and processes involved, and the larger the window is required for data processing.

[0092] The weight of a resource within the current window is determined by its importance score and the proportion of urgent orders. Resources with higher importance scores and a larger proportion of urgent orders have a greater impact on the stability and efficiency of the supply chain, and are given higher weight in the window width calculation to allow for more time and space for data processing. A forgetting factor is used to control the influence of historical data. The trigonometric function term in the formula introduces a time-periodic factor, taking into account the differences in data processing needs at different times of the day, such as busier business during the day and relatively calmer business at night, thus enabling the window width to dynamically change over time.

[0093] This formula measures the strategic position of materials in the supply chain by considering their value, scarcity, and criticality to production. The proportion of urgent orders reflects the urgency of material demand, and materials with a high proportion of urgent orders require priority processing. This formula quantifies and normalizes the importance and urgency of materials, determining the weight of each material in the window width calculation. This makes the window width calculation more aligned with actual business needs, ensuring that critical materials and urgent orders receive timely and sufficient data processing.

[0094] The base window width is dynamically adjusted based on the second metric of the current data. The second metric encompasses data update frequency, data volatility coefficient, and the proportion of outlier data. A weighted average is calculated by normalizing each of these three factors separately before obtaining the second metric score.

[0095] Data update frequency reflects the real-time nature of the data. Frequently updated data requires a narrower window for timely processing to avoid data backlog. Data volatility coefficient measures the severity of data changes. Data with large fluctuations requires a window with stronger adaptability. The proportion of outlier data reflects the quality of the data and potential risks. When there is a lot of outlier data, a wider window is needed for detailed analysis and processing.

[0096] Based on the comparison between the second indicator score and the preset threshold, the window width is adjusted to ensure sufficient data processing. This dynamic adjustment mechanism allows the sliding window width to change flexibly according to the actual state of the data, achieving efficient and accurate processing of material data.

[0097] In one possible implementation, S3 includes:

[0098] Joint analysis of the cleaned data includes:

[0099] a) Vertical intelligent analysis: Construct a deep learning model that integrates temporal and topological features, wherein:

[0100] The LSTM network is used to process the time-series characteristics of materials, such as inventory, order frequency, and quality inspection pass rate, and to extract the time-series pattern of its entire life cycle.

[0101] The multi-level relationships between suppliers, logistics, and warehousing are modeled using graph neural networks (CNNs); node features include historical transaction credibility scores and real-time logistics delays; edge weights reflect the degree of collaboration between nodes.

[0102] Among them, historical transaction credibility score Obtained using the following formula:

[0103]

[0104] The design incorporates a spatiotemporal cross-attention mechanism, dynamically weighting key time nodes and core topology nodes to achieve multi-scale fusion of temporal characteristics and network structure.

[0105] b) Horizontal rule analysis: Embedding an interpretable business rule engine, where:

[0106] Dynamic business constraint sets are generated based on historical data mining and knowledge graph derivation.

[0107] Quantify the contribution of each rule to anomaly detection and use a dynamic weighting strategy to prioritize the rules;

[0108] c) Joint decision-making mechanism: When the decision results of the deep learning model and the rule engine conflict, the following operations are performed:

[0109] Explain the source of the model's output features and identify key influencing factors;

[0110] Perform reverse verification on the triggering rules to evaluate their applicability in the current scenario;

[0111] The final anomaly determination conclusion is generated based on the source tracing and verification results.

[0112] In one possible implementation, the spatiotemporal attention mechanism is implemented by:

[0113] The temporal feature vector output by the LSTM is converted into a query vector Q;

[0114] The topological node features output by the GNN are converted into key-value matrices K and V;

[0115] Dynamic weighted fusion of temporal features and topological features is achieved through attention weight calculation.

[0116] In one possible implementation, the joint decision-making mechanism further includes:

[0117] Establish an interactive training mechanism between the model and rules, including encoding rule trigger states as feature vectors and inputting them into the deep learning model; and feeding back the feature importance of the model output to the rule engine to optimize the constraints.

[0118] Build a conflict case knowledge base for model fine-tuning and rule base version iteration.

[0119] The working principle and effects of the above technical solution are as follows:

[0120] In longitudinal intelligent analysis, a deep learning model that integrates temporal and topological features is constructed. LSTM network and graph neural network (GNN) are used to process different features of materials respectively, and multi-scale fusion is achieved through spatiotemporal cross-attention mechanism.

[0121] LSTM networks excel at processing data with time-series characteristics. For time-series features such as inventory levels, order frequency, and quality inspection pass rates, LSTM networks, through memory units and gating mechanisms, can effectively capture long-term dependencies and patterns in the time series throughout the entire lifecycle of materials, extracting key time-series feature information. GNNs, on the other hand, are used to model multi-level relationships between suppliers, logistics, and warehousing. Suppliers, logistics nodes, and warehousing nodes are treated as nodes in a graph. Node features include historical transaction credibility scores (calculated by the number of successful transactions / total number of transactions × log(1 + years of cooperation), reflecting the reliability of supplier transactions and the stability of cooperation) and real-time logistics delays. Edge weights reflect the closeness of cooperation between nodes, thereby constructing a topological model of the supply chain and uncovering potential relationships between nodes.

[0122] The spatiotemporal cross-attention mechanism converts the temporal feature vector output by LSTM into a query vector Q, and the topological node features output by GNN into key-value matrices K and V. By calculating attention weights, key time nodes and core topological nodes are dynamically weighted, realizing multi-scale fusion of temporal features and network structure, enabling the model to simultaneously utilize the temporal change patterns of materials and the topological structure information of the supply chain for analysis.

[0123] The horizontal rule analysis incorporates an interpretable business rule engine, which generates a dynamic set of business constraints based on historical data mining and knowledge graph derivation. By analyzing historical material data and combining industry knowledge and enterprise business rules, logical relationships and constraints are extracted from the knowledge graph to form a rule set applicable to the current business scenario. Simultaneously, the contribution of each rule to anomaly detection is quantified, and a dynamic weighting strategy is used to prioritize rules. The weights of rules are adjusted according to their importance and effectiveness in different scenarios, ensuring that key rules play a leading role in anomaly detection, thereby improving the accuracy and efficiency of rule analysis.

[0124] When the judgment results of the deep learning model and the rule engine conflict, a joint decision-making mechanism is activated. First, the interpretability of the model's output features is traced, using model interpretation technology to identify the key factors that influence the judgment result and understand the basis for the model's decision. Second, the triggering rule is reverse-verified, combining actual data and background information of the current business scenario to evaluate the applicability of the rule in this scenario. Finally, based on the traceability and verification results, a final anomaly judgment conclusion is generated, comprehensively considering model analysis and rule judgment to arrive at a more accurate and reliable decision.

[0125] Furthermore, the joint decision-making mechanism establishes an interactive training mechanism between the model and rules. Rule trigger states are encoded as feature vectors and input into the deep learning model, enabling the model to learn the business logic inherent in the rules. The importance of the features output by the model is fed back to the rule engine to optimize constraints, achieving dynamic updates and optimization of the rules. Simultaneously, a conflict case knowledge base is built to record and analyze each conflict case for model fine-tuning and rule base version iteration, continuously improving the accuracy and adaptability of the model and rules.

[0126] In one possible implementation, S4 includes:

[0127] When new material data arrives, its spatiotemporal feature vector is extracted; the spatiotemporal feature vector includes time phase encoding, geographic raster encoding, and material status encoding.

[0128] When the similarity exceeds the threshold, the bottom feature extraction layer of the model is frozen, and only the top decision network is fine-tuned.

[0129] In one possible implementation, the similarity calculation employs a hybrid metric method, including:

[0130] Improved cosine similarity calculation for structured data :

[0131]

[0132] in, , For the data timestamps of two structured vectors; , Let j be the numerical value of the j-th dimension of two structured vectors;

[0133] Extract knowledge graph embedding vectors from unstructured data and calculate Euclidean distance;

[0134] The final similarity score is obtained by dynamically weighting and fusing structured and unstructured similarities.

[0135] The working principle and effects of the above technical solution are as follows:

[0136] When new material data arrives, the system extracts multi-dimensional features to generate spatiotemporal feature vectors. Temporal phase encoding digitizes time information, such as dividing a day into multiple time periods and encoding transaction and transportation times into specific numerical vectors to reflect the time-series characteristics of the material data. Geographic raster encoding grids the geographical location of the materials, with each grid corresponding to a code, thus transforming the geographical information of the materials into a computer-processable form, facilitating the analysis of the spatial distribution and flow patterns of the materials. Material status encoding identifies the current state of the materials, such as inventory status, transportation status, and quality inspection status, integrating this status information into the feature vectors. Through these three encoding methods, the spatiotemporal and status information of new material data is transformed into computable and analyzable feature vectors, providing a foundation for subsequent processing.

[0137] Similarity calculation employs a hybrid metric approach, processing structured and unstructured data separately and then fusing the results. For structured data, an improved cosine similarity formula is used, incorporating a timestamp factor into the traditional cosine similarity calculation to reflect the timeliness of the data; data more recent than the current time receives a higher weight in the similarity calculation.

[0138] For unstructured data, knowledge graph embedding vectors are first extracted to map the unstructured data into a low-dimensional vector space. Then, Euclidean distance is calculated; the smaller the Euclidean distance, the higher the similarity between the unstructured data and the feature space. Finally, through dynamic weighting, the two similarity calculation results are fused according to the importance of structured and unstructured data in different scenarios to obtain the final similarity value, which is used to determine the similarity between new material data and historical data. Accurate similarity calculation and targeted model adjustment enable the system to better identify the relationship and differences between new material data and historical data.

[0139] When the similarity between new material data and historical data exceeds a threshold, in order to ensure that the model can utilize existing knowledge while quickly adapting to the characteristics of the new data, the system adopts a strategy of freezing the model's bottom feature extraction layer and only fine-tuning the top-level decision network. The model's bottom feature extraction layer typically learns the general basic features of the data. These features have a certain degree of universality in different data scenarios, and freezing the bottom layer can preserve the general knowledge already learned by the model. The top-level decision network is responsible for making the final decision judgment based on the features extracted from the bottom layer. Fine-tuning the top-level network for new data enables the model to quickly adjust its decision logic based on the existing foundation, adapt to the characteristics of the new material data, and achieve efficient processing and accurate analysis of the new data.

[0140] By extracting spatiotemporal feature vectors, material data is transformed into a unified, computable form, facilitating rapid similarity calculations and model processing. The hybrid metric method employs optimal calculation methods for different data types and reduces unnecessary computational complexity through dynamic weighted fusion of results.

[0141] Freezing the model's bottom-level feature extraction layer avoids retraining a large number of parameters at the bottom level, saving computational resources and time. In practical applications, training deep learning models often consumes significant computational resources and time, especially for complex underlying network structures. This approach significantly reduces model training costs and improves resource utilization efficiency while maintaining model performance, enabling enterprises to effectively manage and analyze new material data at a lower cost. Fine-tuning only the top-level decision network allows the model to quickly adapt to the characteristics of new material data while retaining existing knowledge, enhancing its adaptability to different scenarios and data changes, improving its generalization ability, and ensuring high accuracy and reliability when processing various new material data.

[0142] In one possible implementation, S5 includes:

[0143] A material state transition probability model is generated based on the real-time supply chain network topology. The node state includes three dimensions: inventory level, quality inspection status, and the amount of logistics in transit. The transition edge weights are obtained using the following formula:

[0144]

[0145] In one possible implementation, the following simulations are executed in parallel within a virtual environment:

[0146] Monte Carlo simulation: Generating key parameters with random perturbations, such as 500+ risk scenarios;

[0147] Discrete event simulation: typical abnormal events such as injection equipment failure and demand surge;

[0148] Reinforcement learning exploration: Iterative optimization and correction of strategies through Q-learning;

[0149] Intelligent correction trigger: When both of the following conditions are met simultaneously: simulated profit gain ≥ k3 × operating cost, where k3 is a constant greater than 1.

[0150] When consensus is reached among collaborative nodes with critical path redundancy ≥ a security threshold and a preset percentage, a correction operation is automatically executed and the status of all network nodes is synchronously updated; the preset percentage is, for example, 90%.

[0151] The automatic correction process includes:

[0152] The blockchain evidence storage mechanism writes key parameters for correcting decisions into the consortium blockchain, including: correction trigger timestamp, simulation result hash value, and collaborative node signature set.

[0153] The working principle and effects of the above technical solution are as follows:

[0154] The above steps construct a material status transition probability model based on the real-time supply chain network topology. The node status encompasses three dimensions: inventory level, quality inspection status, and logistics in-transit quantity, comprehensively reflecting the actual situation of materials in the supply chain. The transition edge weight is determined by the formula "Transition edge weight = Historical successful transition count / Total transition attempts × 1 / (1 + Current path load rate)". The ratio of historical successful transition count to total transition attempts reflects the historical reliability of the transition path, while the current path load rate reflects the busyness of the current path. The higher the load rate, the lower the edge weight, thereby dynamically adjusting the probability of transition.

[0155] Three simulations are performed in parallel within a virtual environment:

[0156] Monte Carlo simulations depict various possible supply chain risk scenarios, helping to identify potential problems and challenges.

[0157] Discrete event simulations demonstrate how the supply chain responds and changes when faced with unforeseen circumstances, in order to assess the system's robustness.

[0158] Reinforcement learning exploration: Iterative optimization and correction strategy using Q-learning algorithm.

[0159] By continuously trying different correction strategies and adjusting them based on the reward values ​​from environmental feedback, the optimal correction solution can be gradually found.

[0160] The correction operation will be executed automatically when all three of the following conditions are met:

[0161] The simulated profit gain is greater than or equal to k3 × operating cost (where k3 is a constant greater than 1), ensuring that the correction operation is economically feasible, i.e., the profit is greater than the cost.

[0162] Critical path redundancy ≥ safety threshold ensures sufficient backup capacity for the critical path of the supply chain to cope with uncertainties during the correction process;

[0163] A consensus reached by a preset percentage (e.g., 90%) of collaborative nodes reflects the synergy and consistency among nodes in the supply chain, and adjustments are only made when the majority of nodes agree.

[0164] Once automatic correction is triggered, a blockchain notarization mechanism will be used. Key parameters of the correction decision, such as the correction trigger timestamp, the hash value of the simulation result, and the set of signatures of the collaborating nodes, will be written into the consortium blockchain. The comparison between the simulated benefit gain and the operating cost in the intelligent correction triggering conditions avoids unnecessary waste of resources, improves the operational efficiency and economic benefits of the supply chain, and the correction operation can only be triggered when a predetermined proportion of collaborating nodes reach a consensus, which promotes communication and collaboration among the nodes in the supply chain.

[0165] This application embodiment also provides a supply chain collaborative material management system, used in the supply chain collaborative material management method described in this application embodiment, the system comprising:

[0166] The data standardization module receives heterogeneous material data streams from multiple enterprises, constructs an industry knowledge graph containing material categories, supplier relationships, and logistics routes based on field-level semantic mapping of a dynamic ontology library, and generates standardized data tables with spatiotemporal joint indexes.

[0167] The data storage and resolution module is used to inject standardized data into a distributed message queue for parallel processing, create dynamic cache partitions based on material object tags and have built-in timestamp indexes, scan data using a sliding window mechanism, and perform multi-strategy resolution on conflicting data within the window.

[0168] The joint analysis module is used to perform joint analysis on the cleaned data. Vertically, it constructs an LSTM-GNN hybrid model to capture the characteristics of the entire life cycle of materials, and horizontally embeds an interpretable business rule engine to define constraints and quantify the contribution of each rule to the anomaly detection.

[0169] The report generation module is used to generate enhanced analysis reports. When the similarity between new data and historical cases exceeds a threshold, the case transfer learning mechanism is activated to update the model.

[0170] The anomaly handling module is used to construct a material status transition matrix for abnormal data, simulate correction operations in the digital twin environment, and trigger automatic correction.

[0171] The working principle and effect of the above technical solution are the same as those in the method embodiments of this application, and will not be repeated here.

[0172] This application describes the invention from the perspectives of purpose, performance, progress, and novelty, and it meets the functional enhancement and use requirements emphasized by the Patent Law. The above description and drawings are merely preferred embodiments of this application and are not intended to limit this application. Therefore, all structures, devices, features, etc., that are similar to or identical to those of this application, i.e., all equivalent substitutions or modifications made in accordance with the scope of this patent application, shall fall within the scope of protection of this patent application.

Claims

1. A supply chain collaborative material management method, characterized by, The method comprises: S1: receiving multi-enterprise heterogeneous material data streams, constructing an industry knowledge graph containing material categories, supplier relationships and logistics paths based on field-level semantic mapping of a dynamic ontology library, and generating a standardized data table with a joint time and space index; S2: injecting the standardized data into a distributed message queue for parallel processing, creating dynamic cache partitions according to material object tags and embedding a timestamp index, and using a sliding window mechanism to scan the data and execute multi-strategy resolution on conflicting data in the window; S3: implementing joint analysis on the cleaned data, vertically constructing an LSTM-GNN hybrid model to capture material life cycle characteristics, horizontally embedding an interpretable business rule engine to define constraint conditions, and quantifying the contribution of each rule to abnormal judgment; S4: generating an enhanced analysis report, and activating a case transfer learning mechanism to update the model when the similarity between new data and historical cases exceeds a threshold; S5: constructing a material state transition matrix for abnormal data, simulating and correcting operations in a digital twin environment, and triggering automatic correction; The S3 comprises: Building a deep learning model that integrates time series and topological features, including: using an LSTM network to extract time series patterns of the material life cycle; modeling the multi-level relationship between suppliers, logistics and warehouses through a graph neural network; and designing a spatio-temporal attention mechanism to dynamically integrate time series features and network topology; Embedding an interpretable business rule engine, including: generating a dynamic set of business constraints based on historical data mining and knowledge graph derivation to quantify the contribution of each rule to abnormal judgment, and using a dynamic weight strategy to prioritize rules; When the judgment results of the deep learning model and the rule engine conflict, the following operations are performed: Trace the key influencing factors of the model output features for explainability; Reverse verify the triggered rules to assess their applicability in the current scenario; Generate a final abnormal judgment conclusion based on the traceability and verification results.

2. The supply chain collaborative material management method according to claim 1, characterized by, The industry knowledge graph containing material categories, supplier relationships and logistics paths comprises: Extracting attribute information of materials from multi-enterprise heterogeneous material data streams, classifying materials according to attribute information, and constructing material category nodes; Constructing supplier nodes through supplier information, and connecting supplier nodes and material category nodes through edges; Constructing logistics path nodes through logistics transportation information, and connecting material category nodes and logistics path nodes through edges; Using a graph database to store and manage the industry knowledge graph, and establishing indexes for nodes and edges.

3. The supply chain collaboration material management method of claim 1, wherein, The S2 comprises: Building a multi-dimensional evaluation system for material features, determining the weight of each first indicator through principal component analysis, and calculating the comprehensive feature value of the material; the first indicators include supplier credit score, historical delivery delay rate, material transaction frequency, real-time inventory turnover rate, material importance and inventory volatility; Based on the comparison of the comprehensive feature value and the preset threshold, enter the three-level partition decision logic; On the basis of the timestamp index, design differentiated index enhancement strategies for different risk level partitions.

4. The supply chain collaboration material management method according to claim 3, characterized by, The three-level partition decision logic comprises: If the comprehensive feature value is greater than or equal to the first threshold value, the supplier credit score is less than the preset credit score, and the delivery delay rate in the preset time period is greater than the delay rate threshold value, the material data is independently partitioned, and a supplier risk early warning process is started, and the monitoring frequency of the supplier data is increased; If the second threshold value is less than the comprehensive feature value and the material transaction frequency is greater than x*the historical average value, and the inventory turnover rate is less than the safety threshold value, the partition is preferentially merged with similar high-frequency materials, but the material data is marked for isolated storage; If the comprehensive feature value is less than or equal to the second threshold value, a clustering algorithm is used to cluster the materials, dynamic partitioning is performed according to the clustering results, a partition correlation network is established, and real-time monitoring of data interaction in each partition is performed. If the cross-partition data abnormally links, the partition strategy re-evaluation is triggered.

5. The supply chain collaboration material management method of claim 1, wherein, The S2 comprises: The base window width is determined according to the material turnover period, the real-time data flow rate and the number of supply chain levels; ； wherein, is the current base window width; wherein, m is the number of material categories; i is the material category index, is the standard turnover period of the ith material; is the number of supply chain levels of the ith material; is the weight of the ith material in the current window; is the forgetting factor ∈ [0.8, 1.2]; QPS is the real-time data flow rate; t is the current time point; The second index of the current data is used to dynamically adjust the base window width to obtain an adjusted window width; the second index includes data update frequency, data fluctuation coefficient and abnormal data proportion.

6. The supply chain collaboration material management method of claim 1, wherein, The S2 comprises: Confidence decay fusion processing is performed on the window edge overlapping data; Attention mechanism model is used to identify numerical, time and logical conflicts; The base strategy and the enhanced strategy are executed in layers; the base strategy includes timestamp priority coverage and data source weighted average, and the enhanced strategy includes path backtracking verification and multi-version data fusion; Q-learning model is used to dynamically select the strategy combination to generate a conflict resolution trace map.

7. The supply chain collaboration material management method of claim 1, wherein, The S4 comprises: When new material data arrives, its spatiotemporal feature vector is extracted; the spatiotemporal feature vector includes time phase coding, geographic grid coding and material state coding; When the similarity exceeds the threshold value, the bottom feature extraction layer of the model is frozen, and only the top decision network is fine-tuned.

8. The supply chain collaboration material management method of claim 1, wherein, The S5 comprises: A material state transition probability model is generated based on a real-time supply chain network topology; the node state includes three-dimensional features of inventory level, quality inspection state and logistics in-transit quantity; and the transfer edge weight is obtained by the following formula: ； When the simulated revenue gain is greater than or equal to k3*operation cost, the key path redundancy is greater than or equal to the safety threshold value, and a preset proportion of the above collaborative nodes reach a consensus, a correction operation is automatically performed and the node state of the whole network is updated synchronously; k3 is a constant greater than 1.

9. A supply chain collaborative material management system for implementing the supply chain collaborative material management method according to claim 1, characterized by, The system comprises: A data standardization module is configured to receive multiple enterprise heterogeneous material data streams, construct an industry knowledge graph including material categories, supplier relationships and logistics paths based on field-level semantic mapping of a dynamic ontology library, and generate a standardized data table with a spatiotemporal joint index; A data storage resolution module is configured to inject the standardized data into a distributed message queue for parallel processing, create dynamic cache partitions according to material object labels and embed timestamp indexes, scan data using a sliding window mechanism, and perform multi-strategy resolution on conflicting data in the window. The joint analysis module is configured to perform joint analysis on the data after cleaning, longitudinally construct an LSTM-GNN hybrid model to capture features of the whole life cycle of the materials, laterally embed an interpretable business rule engine to define constraint conditions, and quantify the contribution of each rule to the abnormality determination; The report generation module is configured to generate an enhanced analysis report and activate a case migration learning mechanism to update the model when new data has a similarity to historical cases that exceeds a threshold; The abnormality processing module is configured to construct a material state transition matrix for abnormal data, simulate correction operations in a digital twin environment, and trigger automatic correction. The joint analysis module includes: A deep learning model that fuses time series and topological features, including: using an LSTM network to extract time series patterns of the whole life cycle of the materials; modeling the multi-level correlation between suppliers, logistics, and warehouses through a graph neural network; and designing a spatio-temporal attention mechanism to dynamically fuse time series features and network topological structures; An interpretable business rule engine, including: generating a dynamic business constraint set based on historical data mining and knowledge graph derivation to quantify the contribution of each rule to the abnormality determination, and using a dynamic weight strategy to prioritize the rules; When the determination results of the deep learning model and the rule engine conflict, the following operations are performed: Performing interpretable traceability on the model output features to locate key influencing factors; Performing reverse verification on the triggered rules to evaluate the applicability of the rules in the current scenario; Generating a final abnormality determination conclusion based on the traceability and verification results.

Images