Supply chain data real-time integration processing method and system based on multi-source heterogeneous data

Abstract

The invention relates to the technical field of industrial data management, in particular to a supply chain data real-time integration processing method and system based on multi-source heterogeneous data, and the method comprises the steps: obtaining an external compliance data document, calling a document layout analysis model based on a multi-modal converter, extracting a text semantic vector, fusing the text semantic vector with a two-dimensional space coordinate feature to generate a composite index key, and carrying out the real-time integration processing of the supply chain data. Executing clustering analysis to recognize the text blocks and generating structured metadata; carrying out discretization processing on the business transaction data flow, and constructing an associated database with production environment and logistics data as attributes; executing cross-modal entity alignment by using the structured metadata, and establishing a bidirectional pointer link between the unstructured document index and the associated database; receiving a query vector, executing a recursive traversal algorithm based on a bidirectional pointer link, calculating a cascade influence probability as a correlation sorting score, and outputting a retrieval result; through multi-modal feature fusion and cross-modal entity alignment, deep integration and risk linkage retrieval of supply chain heterogeneous data are realized.

Description

Technical Field

[0001] This invention relates to the field of industrial data management technology, specifically to a method and system for real-time integration and processing of supply chain data based on multi-source heterogeneous data. Background Technology

[0002] In existing technologies, supply chain data management faces the challenge of efficiently integrating multi-source heterogeneous data. Supply chains involve a large amount of business transaction records, real-time production logistics sensor data, and unstructured contracts and compliance documents. Traditional data retrieval and management systems typically store structured data and unstructured documents separately, lacking a unified semantic understanding and association mechanism. Processing unstructured documents often neglects visual layout features and spatial arrangement information, leading to inaccurate extraction of key information. Furthermore, in cross-modal data retrieval, existing technologies mainly rely on shallow keyword matching, lacking entity alignment and bidirectional linking capabilities based on business logic. They cannot perform deep traversal along data dependency paths and struggle to quantify the cascading impact of a single node anomaly on the entire downstream supply chain, resulting in search results lacking depth of association and risk warning value.

[0003] To address the challenge of deeply integrating unstructured documents and structured business flow data in supply chain scenarios, a method and system for real-time integration and processing of supply chain data based on multi-source heterogeneous data is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for real-time integration and processing of supply chain data based on multi-source heterogeneous data. It aims to solve the problem of the difficulty in deeply integrating unstructured documents and structured business flow data in the supply chain scenario. By establishing cross-modal bidirectional pointer links and cascading influence probability calculation, it achieves efficient integration and risk association retrieval of multi-source heterogeneous data.

[0005] A method and system for real-time integration and processing of supply chain data based on multi-source heterogeneous data, including: Acquire business transaction data streams, production environment data, logistics trajectory data, and external compliance data documents; For the external compliant data document, a document layout analysis model based on a multimodal transformer is invoked; the semantic vector features of the text and the two-dimensional spatial coordinate features in the document are extracted and fused to generate a composite index key; based on the composite index key, cluster analysis is performed to identify aggregated text blocks in the document, and the information in the aggregated text blocks is mapped to standard attribute fields to generate structured metadata; The business transaction data stream is discretized, a unique identifier is extracted as the index primary key, and an associated database is constructed with production environment data and logistics trajectory data as associated attributes; using the structured metadata, cross-modal entity alignment is performed, and a bidirectional pointer link is established between the unstructured document index and the associated database to form a data dependency path; Receive the query vector, execute a recursive traversal algorithm along the data dependency path based on the bidirectional pointer link, perform node matching and threshold verification in the retrieval path, calculate the cascading influence probability as the relevance ranking score and output the retrieval results.

[0006] Preferably, the invocation of the document layout analysis model based on the multimodal transformer includes: A dual-stream neural network architecture, comprising visual feature extraction and text semantic extraction branches, is employed. The visual feature extraction branch performs convolution operations on document images to capture visual texture features such as table borders, paragraph spacing, and stamp positions. The text semantic extraction branch performs word embedding processing on character sequences within the document to capture contextual linguistic logic features. The visual texture features and linguistic logic features are then input into a cross-attention module to perform feature alignment operations. The correlation weight matrix between visual regions and text content is calculated to obtain a document structure map, thereby achieving the segmentation and localization of unstructured information.

[0007] Preferably, the step of fusing to generate a composite index key specifically includes: The distribution density values ​​of text semantic vector features in the semantic space and the distance of two-dimensional spatial coordinate features from the geometric center in the document layout are calculated respectively. According to the preset field importance rules, the fusion coefficients of semantic features and spatial features are dynamically allocated. Key numerical fields, including contract amount and date, are given higher semantic weights, and position-sensitive fields, including signatures, headers and footers, are given higher spatial weights. The semantic features and spatial features are merged by weighted concatenation and then dimensionality reduction is performed through a fully connected layer to obtain the composite index key.

[0008] Preferably, the step of performing cluster analysis based on the composite index key specifically includes: A density-based adaptive spatial clustering algorithm is employed, using the composite index key as input, to calculate the distance metric between text blocks within the document. Text blocks with a distance less than a preset threshold are grouped into the same semantic cluster, identifying the title region, body text region, and table data region. For the identified table data region, a grid segmentation based on row and column projection is used to separate the content of individual cells. Using a pre-trained semantic matching model, the similarity score between the content of each semantic cluster and the definition description of the standard attribute field is calculated. The attribute corresponding to the highest score is selected as the mapping target to generate standardized key-value pair data.

[0009] Preferably, the step of discretizing the business transaction data stream, extracting a unique identifier as the index primary key, and constructing an association database with production environment data and logistics trajectory data as associated attributes includes: A data alignment mechanism based on a spatiotemporal sliding window is constructed. Taking the order number in the business transaction data stream as the benchmark anchor point, a dynamic time window covering the production cycle and logistics cycle is set. Within the dynamic time window, the production environment data is resampled and interpolated to fill in the time discontinuities caused by sensor failures. The logistics trajectory data is map matched and denoised to correct the drifted positioning coordinates. The processed continuous production status curve and discrete logistics trajectory points are attached to the order number according to the timestamp sequence to construct a hierarchical data storage structure.

[0010] Preferably, the step of establishing a bidirectional pointer link between the unstructured document index and the associated database specifically includes: Based on graph neural networks, entity objects in the associated database and extracted entities in the structured metadata are respectively constructed as nodes in the graph network. The attribute similarity, co-occurrence relationship and business logic dependency between nodes are constructed as edges in the graph network. Node features are propagated through multi-layer graph convolution operations, and the matching probability score between heterogeneous nodes is calculated. When the score exceeds the confidence threshold, a bidirectional pointer link is generated between the two nodes.

[0011] Preferably, the step of receiving the query vector and outputting the search results specifically includes: A hybrid traversal strategy combining depth-first search and breadth-first search with pruning is adopted. Starting from the initial node hit by the received query vector, the algorithm expands to upstream and downstream nodes along bidirectional pointer links. During each jump, the relevance score coefficient between the current path node and the query intent is calculated in real time. Once the coefficient is lower than the preset truncation threshold, the search of the current branch is stopped immediately to perform pruning operation. At the same time, the path nodes are checked to see if they meet the user permission verification rules, and nodes that are not authorized to access are blocked. The algorithm obtains a retrieval subgraph containing directly related data and indirectly potentially influential data. A Bayesian network risk propagation model is established based on the retrieval subgraph. Each node on the path is mapped to a variable node in the network. The conditional probability relationship between nodes is constructed based on historical interruption data. The posterior probability of causing downstream delivery anomalies is calculated as the cascading impact probability based on the query event trigger probability. The probability is weighted and fused with semantic similarity to generate a comprehensive score. The retrieval results containing the priority of business risks are output according to the score.

[0012] A real-time supply chain data integration and processing system based on multi-source heterogeneous data, the system comprising: Data acquisition module: Acquires business transaction data streams, production environment data, logistics trajectory data, and external compliance data documents; Fusion Clustering Module: For the external compliant data document, it calls the document layout analysis model based on a multimodal transformer; extracts the semantic vector features of the text and the two-dimensional spatial coordinate features in the document, and fuses them to generate a composite index key; performs clustering analysis based on the composite index key, identifies aggregated text blocks in the document, and maps the information in the aggregated text blocks to standard attribute fields to generate structured metadata; Database construction module: Discretizes the business transaction data stream, extracts a unique identifier as the index primary key, and constructs an associated database with production environment data and logistics trajectory data as associated attributes; using the structured metadata, performs cross-modal entity alignment, and establishes a bidirectional pointer link between the unstructured document index and the associated database to form a data dependency path; Query module: Receives the query vector, executes a recursive traversal algorithm along the data dependency path based on the bidirectional pointer link, performs node matching and threshold verification in the retrieval path, calculates the cascading influence probability as the relevance ranking score, and outputs the retrieval results.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. A multimodal transformer with a dual-stream architecture of vision and text is adopted, which integrates visual texture and language logic features and dynamically assigns weights based on the importance of fields, effectively solving the recognition problem caused by the variable layout of supply chain documents. This process significantly improves the extraction accuracy of key fields such as contract amount and seal, and realizes high-precision structured mapping of unstructured documents.

[0014] 2. By constructing bidirectional pointer links, the silos between business flow and document data are broken down. Graph neural networks are used to achieve accurate alignment of cross-modal entities. Combined with the spatiotemporal sliding window mechanism, production logistics data is effectively cleaned and breakpoints and drifts are corrected. This ensures that heterogeneous data forms a tight network structure according to logical dependencies, guaranteeing the integrity and temporal consistency of supply chain data integration.

[0015] 3. By introducing a Bayesian network risk propagation model, it is possible to infer the cascading impact probability of upstream anomalies on downstream delivery along the data dependency path. Combined with pruning strategies and permission verification, it can improve the efficiency of large-scale graph search while outputting search results that include business risk priorities. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of the real-time supply chain data integration and processing method based on multi-source heterogeneous data according to the present invention. Figure 2 This is a schematic diagram of the process for generating composite index keys according to the present invention; Figure 3 This is a schematic diagram illustrating the process of constructing a bidirectional pointer link according to the present invention. Figure 4 This is a system architecture diagram of the supply chain data real-time integration and processing system based on multi-source heterogeneous data according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figures 1 to 4 This invention provides a method and system for real-time integration and processing of supply chain data based on multi-source heterogeneous data, referring to... Figure 1 The flowchart and technical solution are as follows: Real-time integration and processing methods for supply chain data based on multi-source heterogeneous data include: Acquire business transaction data streams, production environment data, logistics trajectory data, and external compliance data documents; For the external compliant data document, a document layout analysis model based on a multimodal transformer is invoked; the semantic vector features of the text and the two-dimensional spatial coordinate features in the document are extracted and fused to generate a composite index key; based on the composite index key, cluster analysis is performed to identify aggregated text blocks in the document, and the information in the aggregated text blocks is mapped to standard attribute fields to generate structured metadata; The business transaction data stream is discretized, a unique identifier is extracted as the index primary key, and an associated database is constructed with production environment data and logistics trajectory data as associated attributes; using the structured metadata, cross-modal entity alignment is performed, and a bidirectional pointer link is established between the unstructured document index and the associated database to form a data dependency path; Receive the query vector, execute a recursive traversal algorithm along the data dependency path based on the bidirectional pointer link, perform node matching and threshold verification in the retrieval path, calculate the cascading influence probability as the relevance ranking score and output the retrieval results.

[0019] Example 1: This embodiment is set in a typical global electronic component supply chain management scenario. The system is deployed to process multi-source heterogeneous data, including chip purchase orders, wafer manufacturing environment monitoring data, cross-border maritime logistics tracks, and various import and export compliance documents, such as certificates of origin, quality inspection reports, and customs declarations.

[0020] First, when processing the external compliant data documents, the system specifically executes the step of calling a document layout analysis model based on a multimodal transformer. This step aims to address the difficulty in automatically extracting and aligning information in unstructured documents (such as scanned documents and PDF images). Specifically, the execution flow of this step is as follows: The system first initializes a two-stream neural network architecture that includes visual feature extraction branches and text semantic extraction branches. Taking a scanned image of a "chip quality inspection report" with a resolution of 2480 pixels by 3508 pixels (A4 paper size, 300 DPI) as an example, the system uses this document image as the input data source.

[0021] The multimodal transformer model adopts the standard Transformer encoder architecture, and the specific parameter configuration is as follows: (1) Visual Feature Extraction Branch: Using a ResNet-50 pre-trained convolutional neural network as the backbone, five convolutional blocks are sequentially executed on the input document image (resolution 2480×3508 pixels). Each convolutional block contains three convolutional layers (kernel size 3×3), with ReLU activation function, and 2×2 max pooling is performed after each convolution. The output feature map of the fifth convolutional block has a dimension of 512×64 (feature map height×width), corresponding to 2048 feature channels. Then, global average pooling is used to compress the feature map to 1×1, and then it is flattened into a 2048-dimensional one-dimensional feature vector, denoted as . .

[0022] (2) Text semantic extraction branch: First, the pre-trained BERT model is called to perform word embedding processing on the text sequence identified in the document. The text vector output by BERT has a dimension of 768. This 768-dimensional vector is then input into the text encoding network (containing 3 fully connected layers with hidden layer dimensions of 768, 512, and 512 respectively). The output dimension of the last layer is 512, denoted as . .

[0023] (3) Feature dimension alignment: due to 2048 dimensions For a 512-dimensional feature, dimension alignment is performed before feature fusion. The L2 norm is performed, and then the result is projected to a 512-dimensional space through a fully connected layer. The projection process uses a weight matrix. A linear transformation is performed on ∈R^{2048×512}, with ReLU as the activation function. The standardized visual feature vector is denoted as... .

[0024] (4) Cross-attention module design: This module adopts a standard multi-head attention mechanism. The number of attention heads is set to 8, and the feature dimension of each attention head is 512 / 8 = 64 dimensions. The cross-attention calculation is as follows: Query vector: by generate, ,in ∈R^{512×512} Key vectors and value vectors: both are derived from... generate, , ,in ∈R^{512×512}.

[0025] Attention weight matrix calculation: ; The outputs of the eight attention heads are concatenated and then fused through a fully connected layer to obtain a 512-dimensional aligned feature representation.

[0026] (5) Model output: The output of the cross attention module is used as the original feature vector of the document structure graph for subsequent clustering and index key generation steps.

[0027] In the visual feature extraction branch, the system performs multi-layer convolution operations on the input document image. Specifically, the system uses convolution kernels to scan the image pixel matrix and extract the low-level visual features of the image. For example, when processing the aforementioned quality inspection report, the convolutional network identifies the table outlines located in the central region of the image. The pixel width of these outlines is mainly distributed between 2 and 5 pixels, forming a five-row, four-column grid structure. Simultaneously, it identifies the paragraph spacing at the bottom of the document, with a vertical pixel distance of approximately 30 pixels. Furthermore, through color channel analysis, the system locates a red stamp area with a diameter of approximately 150 pixels in the lower right corner of the document. The system encodes these captured table outlines, paragraph layouts, and stamp location information into a visual texture feature tensor.

[0028] Meanwhile, in the text semantic extraction branch, the system first identifies character sequences within the document using an Optical Character Recognition (OCR) engine and converts the identified text into computer-readable strings. Next, these character sequences undergo word segmentation and word embedding processing. For example, for characters such as "temperature resistance test," "pass," and "GB / T2423" appearing in the report, the system maps them to numerical vectors with 512 dimensions. This process not only encodes the meaning of the words themselves but also captures the linguistic logic features of the context through positional encoding; for instance, it identifies the word "pass" as a conclusive description immediately following the test item "temperature resistance test."

[0029] Furthermore, the system simultaneously inputs the extracted visual texture features and linguistic logic features into the cross-attention module to perform feature alignment. During this process, the cross-attention module calculates the correlation weight matrix between visual regions, such as the coordinate range of a cell in a table, and text content, such as "125 degrees Celsius".

[0030] Specifically, the system calculates the matching degree between each text vector and each visual feature region. For example, for the text content "qualified", the system calculates the association weight between its visual features and the red stamp area at the bottom of the document. If the calculated weight value reaches 0.85, which is the set threshold, it is determined that there is a strong correlation between the two. For the text content "batch number: 20230801", the system calculates the association weight between its visual features and the grid area in the first row and second column of the table to determine its spatial correspondence.

[0031] Furthermore, based on the calculated association weight matrix, the system generates a document structure graph. This document structure graph is represented by a tree-like or graph-like data structure, clearly defining the division and location of unstructured information. For example, the document structure graph defines the image coordinate range [200, 500] to [800, 600] as the "core parameter area," and binds the visual table structure within this area to text nodes containing semantics such as "voltage" and "current," thereby achieving accurate parsing of unstructured document content and providing an aligned feature base for subsequent extraction of aggregated text blocks and generation of structured metadata.

[0032] Employing a multimodal transformer with a dual-stream architecture encompassing both visual and textual elements, along with a cross-attention mechanism, it can simultaneously capture both visual textures and linguistic logic features of the layout, effectively solving the challenge of accurately locating and parsing unstructured information such as tables and seals in complex supply chain documents.

[0033] Furthermore, after obtaining the semantic vector features and two-dimensional spatial coordinate features of the text, the system performs a step of fusing them to generate a composite index key. This step aims to dynamically adjust the feature weights based on the attribute characteristics of different fields, generating index data with strong representational capabilities. (Refer to...) Figure 2 This is a schematic diagram of the process for generating composite index keys according to the present invention; Specifically, the execution flow of this step is as follows: The system first calculates the distribution density of text semantic vector features in the semantic space. Taking an electronic component procurement contract as an example, the system extracts the semantic vector of the text block containing "total contract amount," such as a 768-dimensional BERT vector, and calculates the average Euclidean distance between this vector and all text vectors in the document, quantifying its semantic uniqueness. Simultaneously, the system calculates the two-dimensional spatial coordinate features of this text block, including the distances of the top-left and bottom-right corners relative to the geometric center of the document layout. For example, for an A4-sized document with the center point coordinates set to (1050, 1485), the system calculates the straight-line pixel distance from the center point of the text block to that center point.

[0034] Subsequently, the system dynamically allocates the fusion coefficients of semantic and spatial features based on preset field importance rules. For key numerical fields such as "contract amount," "delivery date," and "material code," the system determines that their semantic information has higher distinguishability than their positional information, and therefore assigns higher weight coefficients to semantic features, for example, setting the semantic weight to 0.8 and the spatial weight to 0.2. For position-sensitive fields such as "company seal," "seale seal," and "header and footer," the system determines that their fixed positions on the page are more identifiable than their specific text content, and therefore assigns higher weight coefficients to spatial features, for example, setting the spatial weight to 0.7 and the semantic weight to 0.3.

[0035] Further optimization involves introducing an adaptive weighting mechanism based on feature information entropy to accurately calculate the fusion ratio of semantic vectors and spatial coordinates when generating composite index keys. Specifically, the system calculates Shannon entropy on the extracted text semantic vector feature matrix to quantify its information dispersion. Simultaneously, it calculates the normalized variance of spatial coordinate features in the layout distribution. When the semantic entropy value of a field (such as "remarks") is greater than a preset threshold and the spatial variance is large, it is determined that its content dispersion is high, and the system automatically increases the projection dimension of the semantic features at the input of the fully connected layer. Conversely, for fields with fixed formats (such as "order number"), if their spatial location variance approaches zero, the system suppresses the weight of the semantic dimension through the Sigmoid activation function. Through this entropy feedback loop based on the statistical characteristics of data distribution, the parameter matrix of the feature fusion layer is dynamically adjusted. Through adaptive weighting based on information entropy, it can ensure that even when the document layout is distorted or the text content is blurred, a highly robust unique index key can still be generated, significantly improving the clustering accuracy and retrieval recall of heterogeneous data in the vector space.

[0036] After determining the weights, the system merges the semantic feature vector and the spatial feature vector through weighted concatenation to form a high-dimensional feature vector. This high-dimensional feature vector is then input into a fully connected layer network for dimensionality reduction, compressing the dimension to a preset length of 128 dimensions, ultimately yielding a composite index key for subsequent processing.

[0037] The specific process of calculating the distribution density value and dynamically allocating the fusion coefficient is as follows: For each independent text block identified in the document, the system extracts semantic vectors through a pre-trained language model and calculates its local density using the K-nearest neighbor density estimation method. Specifically, the semantic density features are obtained by calculating the Euclidean distance between the current text block and its neighboring vectors, taking the reciprocal of the average distance, and standardizing the result. At the same time, the system determines the geometric center point of the document layout, calculates the straight-line distance between the text block coordinates and the center point, and uses half the length of the document diagonal as a normalization factor for division to obtain the normalized spatial distance features.

[0038] Based on this, the system dynamically allocates weights according to field importance rules: for key numerical fields such as contract amounts, higher semantic weights and lower spatial weights are assigned; for location-sensitive fields such as official seals, higher spatial weights and lower semantic weights are assigned. The system identifies text block types and matches corresponding coefficients using a text classifier. The weighted semantic density, spatial distance, and original semantic vectors are concatenated into a fused feature vector, which is then input into a fully connected neural network model. Nonlinear activation functions are used for feature transformation and dimensionality reduction layer by layer. The final output is a low-dimensional vector restricted to a specific interval, which serves as a composite index key for subsequent clustering analysis and entity alignment.

[0039] By dynamically allocating the fusion weights of semantic and spatial features, targeted feature enhancements are performed on different types of fields such as contract amount, date, and signature, significantly improving the composite index key's ability to represent and distinguish key business information.

[0040] Furthermore, based on the generated composite index key, the system performs a clustering analysis step to identify the document structure and extract structured data. Specifically, the execution flow of this step is as follows: The system employs a density-based adaptive spatial clustering algorithm, using the composite index key of all text blocks within the document as input data points. The system calculates the distance metric between any two text blocks, which combines semantic similarity and spatial proximity. A distance threshold of 0.15 is set. When the distance metric between two text blocks is less than this threshold, the system merges them into the same semantic cluster. Through multiple iterations, the system clusters scattered text blocks into different functional areas. For example, the company name and logo at the top are grouped into the title area, the large sections of legal clauses in the middle are grouped into the body text area, and densely packed areas containing fields such as "product name," "specification," and "unit price" are identified as table data areas.

[0041] For the identified table data region, the system performs a grid segmentation operation based on row and column projection. The system performs pixel projection statistics in the horizontal and vertical directions on the binarized image of the table region. Based on the distribution characteristics of projection peaks and troughs, it determines the coordinates of the table row dividing lines and column dividing lines, such as Y-axis pixel coordinates of 500, 550, and 600. According to these dividing lines, the system physically segments the table region into independent cell contents and extracts the text information in each cell.

[0042] Finally, the pre-trained semantic matching model is a deep neural network built on a Transformer architecture (such as BERT). After pre-training on a large-scale general corpus and fine-tuning through comparative learning of synonym pairs in the supply chain domain, it possesses the ability to map texts with different expressions to the same semantic vector space. Semantic similarity is calculated by transforming the cell text to be matched and the standard attribute fields into high-dimensional dense feature vectors, and then calculating the cosine similarity between the two feature vectors (i.e., the ratio of the vector dot product to the product of their magnitudes). This value quantifies the degree of semantic association between the two and is used to determine the confidence level of the mapping relationship. The system uses the semantic matching model to match each semantic cluster and cell content with predefined standard attribute fields in the database. For example, for the extracted cell text "Unit Price Including Tax", the system calculates its semantic similarity score with the standard field "Unit_Price_Tax_Incl". If the calculated similarity score is 0.96, which is the highest among all candidate fields, the system selects "Unit_Price_Tax_Incl" as the mapping target, establishes a binding relationship between the value "12.50" in the cell and the standard field, generates standardized key-value pair data, and completes the transformation from unstructured document to structured metadata.

[0043] The specific implementation of the density-based adaptive spatial clustering algorithm is as follows: First, the system defines a distance metric to measure the similarity between two text blocks. For each text block's corresponding composite index key, the system uses a weighted Euclidean distance algorithm for calculation. To accurately distinguish the different impacts of semantic and spatial features on the clustering results, the system divides the composite index key into semantic and spatial dimension intervals, assigning different weight coefficients to each. Specifically, the system assigns a higher total weight (0.6) to the semantic dimension interval and a lower total weight (e.g., 0.4) to the spatial dimension interval. The specific weight of each dimension is the total weight of that interval divided by the number of dimensions contained in that interval, thus ensuring that the distance calculation simultaneously reflects the semantic similarity of the text content and the proximity of the page layout.

[0044] Secondly, the system calculates an adaptive threshold for determining cluster boundaries, which automatically adjusts based on document complexity. The system first calculates the distances between all text block pairs within the document, forming a distance distribution set, and then calculates the statistical mean and standard deviation of this set. To accommodate different document layout characteristics, the system sets the adaptive threshold to the mean minus a certain percentage of the standard deviation (e.g., the mean minus 0.5 times the standard deviation). Through this dynamic calculation method, for documents with dense and complex text distribution, the system automatically generates a smaller threshold to achieve refined clustering; while for sparsely and uniformly distributed documents, a larger threshold is generated to avoid over-segmentation.

[0045] Finally, the system performs density-based clustering using the aforementioned distance metric and adaptive threshold. The specific steps include: traversing each unvisited index key and calculating the number of neighboring index keys within the adaptive threshold distance range; if the number within that neighborhood is not less than a preset minimum number of neighboring points (e.g., 3), the current index key is marked as a core point, and a new cluster is created; subsequently, starting from this core point, the system assigns all points within its neighborhood to the same cluster, and recursively expands this cluster, including points adjacent to boundary points that satisfy the density condition; for index keys that ultimately cannot be assigned to any cluster, the system marks them as noise points and ignores them in subsequent processing. This process ultimately outputs a set of clusters of several text blocks, each cluster corresponding to an independent semantic region in the document.

[0046] By utilizing density-based adaptive spatial clustering and table grid cutting techniques, it is possible to intelligently identify and separate text and table regions in a document, and achieve standardized mapping through semantic matching, thus ensuring efficient extraction and structured transformation of unstructured document content.

[0047] Furthermore, the system performs discretization processing on the business transaction data stream and establishes a multi-source data alignment and storage mechanism based on spatiotemporal dimensions. Specifically, the execution flow of this step is as follows: The system first receives real-time business transaction data streams from the Enterprise Resource Planning (ERP) system. This data stream typically exists as a continuous JSON message queue. The system discretizes this data stream, breaking it down into independent transaction event records and extracting unique identifiers as index keys. For example, the system parses a transaction record containing "Order Number: ORD-20231115-CN", "Material Code: MCU-9900", and "Order Time: November 15, 2023, 09:00:00", and determines "ORD-20231115-CN" as the baseline anchor for all subsequent data associations.

[0048] Next, the system constructs a data alignment mechanism based on a spatiotemporal sliding window. Using the order number "ORD-20231115-CN" as a benchmark, the system queries the production scheduling plan and logistics delivery notification associated with the order, setting a dynamic time window covering both the production and logistics cycles. Assuming the production cycle begins on November 10th and the logistics is expected to end on November 20th, the system sets the time span of this dynamic time window to 00:00 on November 10th to 23:59 on November 20th, and uses this time period as the query range to retrieve raw data from the underlying industrial IoT database and logistics positioning platform.

[0049] Within the dynamic time window, the system performs resampling and interpolation processing on the retrieved production environment data. Taking temperature sensor data from the wafer manufacturing workshop as an example, the sensor is preset to report data every 30 seconds. The system detected that during the period from 14:00:00 to 14:05:00 on November 12th, data was missing due to network fluctuations, i.e., a time breakpoint existed. Based on the temperature values ​​collected at adjacent times, which were 24.5 degrees Celsius and 24.8 degrees Celsius respectively, the system automatically generated filler data for this breakpoint period using a linear interpolation algorithm, such as 24.6 degrees Celsius and 24.7 degrees Celsius, to complete the production status curve and ensure data continuity.

[0050] Simultaneously, within the dynamic time window, the system performs map matching and noise reduction processing on the retrieved logistics trajectory data. Taking the GPS data of the truck carrying the goods in this order as an example, the system obtains a set of raw latitude and longitude coordinates. The system detects that one of the coordinate points deviates from the actual road grid, indicating it is located in farmland next to the highway, which is considered drift noise. The system calls the basic geographic information database, executes the Hidden Markov Model algorithm, and forcibly attaches the drifted coordinates to the nearest G4 highway road network trajectory, providing corrected coordinates to ensure the authenticity of the logistics trajectory.

[0051] Furthermore, when constructing the physical storage structure of the relational database, a spatiotemporal multidimensional data linearization storage strategy based on Hilbert curves is adopted to replace the traditional linear timestamp index. Specifically, the system maps the timestamp dimension of production environment data, the two-dimensional spatial dimension of latitude and longitude of logistics trajectories, and the hash value of order IDs into a normalized multidimensional hypercube. Using the Hilbert space-filling curve algorithm, the discrete points within this multidimensional hypercube are projected into a one-dimensional integer sequence, and this integer sequence is used as the physical clustering index of the underlying database. When data is written to disk, the system sorts and stores data pages according to this index value, so that data records that are close in spatiotemporal neighborhood (such as logistics points and environmental data in the same time period and geographically close) are also stored contiguously on the physical disk sectors. By maintaining the spatial locality of data through the Hilbert curve, the system significantly reduces disk seek time and page replacement overhead when retrieving cross-modal data, and greatly improves the efficiency of range query and concurrent reading of massive supply chain spatiotemporal data.

[0052] Finally, the system appends the processed continuous production status curves, such as temperature and humidity readings every minute, and the corrected discrete logistics trajectory points, such as latitude and longitude coordinates every ten minutes, to the order number "ORD-20231115-CN" according to a unified timestamp sequence. The system adopts a hierarchical data storage structure, such as a key-value wide table or a document-oriented database structure, with the order number as the root node, "production stage" and "logistics stage" as child nodes, and the cleaned time-series data as leaf nodes for physical storage, thus completing the construction of the relational database.

[0053] The specific implementation steps for setting the dynamic time window and processing the data in the spatiotemporal sliding window are as follows: The system first estimates the entire supply chain cycle by accessing the historical database based on order attributes. For production cycle estimation, the system queries the average production time (i.e., the average number of days from order confirmation to production completion) for a specific material code in historical order records, using this as the baseline production cycle for that order. For logistics cycle estimation, the system queries the historical average transportation time for the corresponding transportation mode (such as air, sea, or land transport) based on the geographical distance between the shipping and receiving locations, and sets this as the logistics cycle. Finally, the system determines the coverage of the dynamic time window: setting the start point of the time window as a preset buffer period before the order confirmation time, and the end point as the sum of the production completion time and the logistics cycle, thus constructing a time span covering the entire process from production preparation to logistics delivery.

[0054] Within a defined time window, raw sensor data sequences are acquired from an industrial IoT platform. First, missing data detection is performed, iterating through the time series. If the time interval between any two adjacent data points exceeds a preset normal sampling threshold, data is considered missing for that time period. For the detected missing time periods, the system uses linear interpolation to fill in the gaps. This involves establishing a linear proportional relationship between the preceding and following valid data points within the missing segment, calculating the value at any given moment within the missing segment, and thus filling the gap. Subsequently, to achieve temporal alignment of multi-source data, the system resamples the filled production environment data based on the sampling frequency of the logistics data. Through interpolation calculations, the sampling interval of the production data is adjusted to match that of the logistics data, generating a standard dataset with a unified timestamp sequence.

[0055] The system obtains the original latitude and longitude coordinate sequence from satellite positioning services and performs denoising and matching operations sequentially. First, denoising preprocessing is performed, calculating the distance and speed of movement between adjacent positioning points. If the calculated speed exceeds the physically reasonable limit or an excessive displacement occurs within a very short time interval, the point is marked as an abnormal noise point and removed. Next, the system uses a Hidden Markov Model combined with a geographic information database for map matching: the denoised original satellite positioning coordinate sequence is used as the observation sequence, and the actual road grid segments are used as the hidden state sequence. An observation probability model (usually following a normal distribution) is established based on the distance between the positioning point and the road segment, and a state transition probability model is established based on the topology of the road network. The system calls an optimal path search algorithm (such as the Viterbi algorithm) to find the optimal hidden state sequence with the highest probability among all possible paths, thereby determining the road segment that best matches the actual driving trajectory. Finally, the system calculates the projection point of the original positioning point on the optimal road segment and replaces the original coordinates with this corrected projection coordinate, completing trajectory correction.

[0056] A hierarchical key-value pair wide table structure is constructed to store the processed data in a relational database. This structure uses a unique order number as the global root node primary key, and divides the root node into two first-level child nodes: "Production Stage" and "Logistics Stage". The "Production Stage" child node stores environmental time-series data (such as temperature, humidity, and pressure) after resampling and interpolation processing, while the "Logistics Stage" child node stores trajectory coordinate data corrected by map matching. In terms of data format, a common lightweight data exchange format or a distributed columnar storage format is adopted to ensure that all production elements and logistics trajectory data strictly correspond to a unified timestamp sequence, facilitating subsequent retrieval and correlation analysis.

[0057] A data alignment and repair mechanism based on a spatiotemporal sliding window was constructed, which effectively solved the problems of missing production environment data and drifting logistics trajectory. It accurately mounts discrete multidimensional data to the order dimension, ensuring the temporal integrity and accuracy of the underlying related database.

[0058] Furthermore, the system establishes bidirectional pointer links between the unstructured document index and the associated database. This step uncovers the complex relationships hidden among multi-source data by constructing a graph neural network model. (See also...) Figure 3 This is a schematic diagram of the process for constructing a bidirectional pointer link according to the present invention; specifically, the execution flow of this step is as follows: The system first constructs a heterogeneous graph network oriented towards the supply chain domain. The system instantiates entity objects from the associated database, such as order records, material entries, and supplier profiles, as "structured nodes" in the graph network. For example, the order number "ORD-20231115-CN" stored in the database and its associated supplier "Company X" are constructed as a structured node A. Simultaneously, the system extracts structured metadata from external compliant documents, such as identified invoice headers and material descriptions on packing lists, as "document nodes" in the graph network. For example, the text block "Buyer: Company X" extracted from a PDF invoice and its coordinate position in the document are constructed as a document node B.

[0059] Next, the system constructs edges between nodes according to preset rules, including calculating attribute similarity, analyzing co-occurrence relationships, and defining business logic dependencies.

[0060] Regarding attribute similarity, the system calculates the edit distance between the supplier name "Field 1" in node A and the text "Field 2" in node B. If the calculated string similarity is 0.92, a "name similarity edge" with a weight of 0.92 is established between the two nodes.

[0061] Regarding co-occurrence relationships, if the database record's entry time is "2023-11-20" and the document's parsed signing date is "2023-11-18", and the time difference between the two is within the preset three-day tolerance range, the system establishes a "time co-occurrence edge".

[0062] In terms of business logic dependencies, the system establishes directed "logical dependency edges" between order nodes and invoice nodes based on supply chain rules, such as "orders come first, invoices come later".

[0063] Furthermore, in the process of establishing bidirectional pointer links, a heterogeneous graph attention network mechanism based on meta-path guidance is adopted to filter noisy connections in the graph structure. The system predefines multiple meta-path patterns with clear data dependencies, such as "order node - logistics node - location node" or "order node - document node - amount field". When performing graph convolution operations, the system does not perform uniform aggregation across the entire graph, but only samples neighborhood features along these predefined meta-paths. A self-attention mechanism is introduced on each meta-path to calculate the contribution coefficient of the current path to the feature representation of the target node. Only when the aggregation contribution of a meta-path exceeds the confidence threshold is a physical pointer link generated between the first and last nodes of that path, and redundant edges with low contribution are removed. Through the attention mechanism guided by meta-paths, it is ensured that the generated bidirectional pointer links strictly follow effective data dependency logic, which greatly improves the accuracy and convergence speed of retrieval traversal while ensuring the sparsity of the index graph.

[0064] Subsequently, the system performs feature propagation operations based on multi-layer graph convolutional neural networks (GCNs). The system inputs the constructed graph structure into a pre-trained graph neural network model, which contains three convolutional layers. The specific structure and parameter configuration are as follows: The first convolutional layer receives an initial node feature vector of dimension 768 as input, sets the output dimension to 512, and uses a linear rectified function as the activation function to extract preliminary local correlation features. The second convolutional layer receives the output of the first layer, sets the output dimension to 256, and also uses a linear rectified function as the activation function to further compress the feature space and perform nonlinear transformation. The third convolutional layer receives the output of the second layer, sets the output dimension to 128, and uses a sigmoid activation function to strictly limit the output value to the range of 0-1, ultimately generating a low-dimensional feature representation that incorporates neighborhood information.

[0065] In the first convolutional layer, node A aggregates the feature information of its neighboring nodes, and the same applies to node B. After three layers of graph convolutional operations, the node features are fused with the topological and semantic information of their local neighborhood. The system calculates the matching probability score between heterogeneous nodes. In this example, after feature propagation, the model determines that although nodes A and B have slight differences in name, due to strong temporal co-occurrence and logical dependency, the calculated matching probability score is 0.95.

[0066] Finally, the system performs threshold verification and link generation. The system sets the confidence threshold to 0.85. Since the calculated matching probability score (0.95) exceeds this threshold, the system determines that node A and node B are different manifestations of the same entity. The system then generates a bidirectional pointer link between the two nodes: A forward pointer points from the associated database to an unstructured document, stored in the format of "database primary key ID -> document storage path + page index + bounding box coordinates". For example, it can directly jump from an order record to the specific amount area of ​​the scanned invoice. The reverse pointer points from the unstructured document to the associated database, with the storage format being "document metadata ID -> database table name + record primary key ID". For example, clicking on the material row in the invoice preview image will index the corresponding production environment temperature curve data in the database.

[0067] By building bidirectional pointer links between unstructured documents and related databases based on graph neural networks, and deeply mining the attribute similarity and business dependencies between nodes, high-confidence alignment and interconnection of cross-modal heterogeneous entities are achieved.

[0068] After establishing a heterogeneous data graph with bidirectional pointer links, the system enters the user interaction and intelligent retrieval phase. At this point, the system not only needs to respond to users' explicit query requests but also needs to uncover potential business risks behind the data.

[0069] Furthermore, the system receives query vectors and outputs retrieval results, and then performs retrieval and Bayesian risk inference based on a hybrid traversal strategy.

[0070] The construction and reasoning process of the Bayesian network risk propagation model includes: A probabilistic dependency network structure is constructed, mapping entity nodes in the retrieval subgraph to random variables with discrete state value ranges. Connections representing business dependencies are transformed into directed edges with probabilistic influence. Structural loops are eliminated by introducing auxiliary nodes or adjusting directions to ensure that the topology meets the requirement of directed acyclicity. Conditional probability parameters are calculated by retrieving historical abnormal event databases to statistically analyze the frequency of occurrence of each state of child nodes under specific combinations of parent node states. A smoothing algorithm is used to correct estimation biases caused by sparse samples, and a conditional probability table is generated and stored.

[0071] Furthermore, probabilistic reasoning is performed to parse user query commands to determine evidence variables. Combining this with a conditional probability table, a probabilistic reasoning algorithm is used to calculate the posterior probability of the target node under the current evidence conditions, thereby quantifying the cascading impact of risk propagation downstream along directed edges. Multi-path integration is implemented: if multiple propagation paths exist from the source node to the target node, a weighted average is calculated based on the number of key nodes involved in each path, or the maximum or minimum path probability is selected according to risk warning requirements to obtain the final comprehensive risk propagation probability.

[0072] Specifically, the execution flow of this step is as follows: The system first receives a natural language query command input by the user, such as "Query for order delivery risks caused by abnormal wafer production temperatures in November 2023". The system converts the text into a high-dimensional query vector and searches it in the graph network index, hitting one or more initial nodes. In this example, the initial nodes are "Wafer Manufacturing Workshop Environmental Monitoring Records (2023-11)" and the associated "Order Set".

[0073] Next, the system employs a hybrid traversal strategy combining depth-first search (DFS) with pruning and breadth-first search (BFS). Starting from the initial node, the system expands upstream and downstream along pre-established bidirectional pointer links.

[0074] In the first-level traversal (BFS mode), the system accesses directly related nodes, such as the "production temperature curve data" for a specific batch and the corresponding "purchase order ORD-20231115". At this point, the system calculates the relevance score coefficient between the current path node and the query vector. Since they are directly related, the coefficient is set to 1.0.

[0075] In the second-level traversal (DFS mode), the system traces downstream from the order node to the "customer's shipping address" node. At this point, the system detects that this node is purely geographic location information, which has a weak semantic relevance to the query intent of "temperature anomaly risk," and the relevance score coefficient drops to 0.15.

[0076] The system performs a threshold check, setting the truncation threshold to 0.2. Since 0.15 is below this threshold, the system immediately stops searching for that branch, performs a pruning operation, and no longer continues to traverse other historical orders of that customer, thereby saving computing resources.

[0077] Simultaneously, during the traversal, the system checks whether each path node meets the user permission verification rules. Assuming the current user is a "Production Engineer," when the system traverses to the associated "Contract Business Terms" node (which contains price-sensitive information), it identifies that the user does not have permission to access "Finance" classified data. The system then blocks this node and does not include it in the final search subgraph, ultimately obtaining a search subgraph that includes "Temperature Anomalies," "Affected Wafer Batches," "Related Order Status," and "Logistics Shipping Notifications."

[0078] Subsequently, the system establishes a Bayesian network risk propagation model based on the retrieved subgraph. The system maps each node on the subgraph path to a variable node in the Bayesian network. Based on historical interruption data (which does not assume an absolute single-factor causal relationship between nodes, but rather uses statistical inference based on multi-dimensional historical fault correlation data, specifically retrieving all abnormal event records related to the target node from a historical database spanning more than 5 years, including sensor data anomalies, business interruption events, and non-conforming marks in quality inspection reports), the system constructs conditional probability relationships between nodes.

[0079] The probability of cascading impact refers to the conditional probability that, given an anomaly at an upstream node, the anomaly propagates downstream along the supply chain dependency path, ultimately causing an anomaly at a critical target node (such as a delivery delay).

[0080] The probability is calculated based on the chain rule of conditional probability and the Markov assumption. For a dependency path {X1→X2→...→Xn}, the cascade probability is calculated as follows: P(Xn=abnormal|X1=abnormal)=P(Xn=abnormal|X{n-1}=abnormal)×P(X{n-1}=abnormal|X{n-2}=abnormal)×...×P(X2=abnormal|X1=abnormal); This formula reveals the cascading mechanism of risk propagation along the supply chain dependent path. It deconstructs the overall cross-level probability of an anomaly from the source node X1 to the end node Xn into a product of the "local transmission probabilities" between each pair of adjacent nodes in the path. This shows that the propagation of anomalies is carried out in a "relay-like" manner. The magnitude of the overall risk depends on the ability of each intermediate link in the path to accept and transmit anomalies. Furthermore, since the probability value is usually less than 1, the product operation also reflects the objective law that the risk gradually decreases as the link lengthens.

[0081] For example, the system sets the conditional probability to 0.85 for the "temperature exceeds standard" node and the "yield test fails" node; and sets the conditional probability to 0.9 for the "yield test fails" node and the "final delivery delayed" node.

[0082] The system uses a query event (temperature anomaly) as an evidence variable input to trigger probability inference calculations. The system calculates the posterior probability of causing downstream delivery anomalies, which is 0.85 multiplied by 0.9 equals 0.765. This value of 0.765 represents the cascading effect probability, quantifying the likelihood that the current production anomaly will negatively impact the final delivery.

[0083] Finally, the system weights and fuses the cascading effect probability with semantic similarity. The system sets the semantic matching score, i.e., the matching degree between the query term and the document content, with a weight of 40%, and the cascading effect probability, i.e., the degree of business risk, with a weight of 60%.

[0084] Suppose a search result has a mediocre text match (semantic score of 0.6) but an extremely high risk probability of 0.765, and its overall score is improved to the top of the list; while another result has a perfect text match but belongs to resolved historical archive data (risk probability of 0.01), and its overall score is reduced.

[0085] By employing a hybrid traversal algorithm with pruning strategy combined with a Bayesian network risk model, the retrieval efficiency and security are improved, while the probability of the cascading impact of upstream node anomalies on downstream delivery is quantitatively assessed, thus realizing intelligent recommendation based on business risk priority.

[0086] Reference Figure 4 The system architecture diagram is shown below, and the integrated configuration of the system at the hardware and software levels is as follows: The data acquisition module is deployed at the system's edge access layer and is configured with multiple protocol adapters. For business transaction data streams, this module establishes a long connection with the enterprise's ERP system (such as SAP or Oracle) through an encrypted API interface, listening to and capturing JSON-formatted order messages in real time. For production environment data, this module connects to the industrial IoT platform through an MQTT protocol gateway, collecting sensor time-series data at a concurrent rate of one thousand messages per second. For logistics trajectory data, this module synchronizes vehicle latitude and longitude information through a Beidou / GPS satellite positioning interface. For external compliant data documents, this module is configured with RPA (Robotic Process Automation) scripts to periodically download PDF or image-formatted document files from the FTP servers of customs or quality inspection agencies and write the raw data to the system's raw data buffer.

[0087] The fusion clustering module is deployed as a standalone AI inference service on a high-performance computing server equipped with a GPU accelerator card (such as NVIDIA A100). This module reads external compliant data documents from the raw data buffer via an internal fiber optic network. The system integrates an OCR optical character recognition engine and a pre-trained deep learning model based on the Transformer architecture. During runtime, the module utilizes the parallel computing capabilities of the GPU to perform convolution operations and attention mechanisms on the document images, automatically extracting semantic vectors and spatial coordinates. It generates composite index keys through feature concatenation operations in memory and calls a clustering algorithm library to divide the document content into blocks, ultimately transforming the unstructured document into a structured metadata stream, which is then sent to a message broker (such as Kafka).

[0088] The database building module is deployed at the data storage layer, integrating a multimodal database cluster, including a time-series database (such as InfluxDB) for storing time-series data and a graph database (such as Neo4j) for storing relationships. This module subscribes to data streams in the message middleware, performs discretization logic on business transaction data, and creates entity nodes in the graph database. Simultaneously, the alignment engine within this module mounts production and logistics data to the corresponding primary key nodes based on timestamps and geographic coordinates. Furthermore, this module executes a cross-modal entity alignment procedure, scanning the structured metadata and node features in the relational database. When the matching degree meets the requirements, a bidirectional pointer link is written at the underlying storage engine of the graph database, physically establishing a direct addressing path between the document storage address and the database record ID.

[0089] The query module is deployed at the application service layer, providing a RESTful query interface. Upon receiving a query vector from the front-end application, this module invokes the graph computing engine, using a memory graph traversal algorithm to retrieve data along bidirectional pointer links. Internally, this module integrates a Bayesian network inference engine, which reads historical fault database data in real-time during the retrieval process and calculates the probability of cascading effects. Finally, the module weights and sorts the risk probability and semantic matching degree, encapsulates it into a JSON-formatted result set containing risk warnings, and returns it to the client, thus achieving a closed-loop system from underlying data collection to top-level intelligent applications.

[0090] By solidifying the above data processing logic into a modular system architecture, the system achieves automated collection and real-time parallel processing of massive amounts of multi-source heterogeneous data, significantly improving the system's response speed and operational stability in high-concurrency scenarios, and facilitating rapid deployment and integration in actual supply chain operations.

[0091] This embodiment constructs a supply chain data processing system based on a distributed microservice architecture, integrating hardware and software configurations to achieve data closure. At the edge layer, the system utilizes multi-protocol adapters to collect ERP transactions, IoT production data, GPS logistics data, and compliance documents. At the computing layer, GPU servers are deployed to run Transformer models to extract multimodal features from documents and generate structured metadata. At the storage layer, time-series and graph databases are integrated, business data is mounted through spatiotemporal alignment, and bidirectional pointer links are established between unstructured indexes and relational databases. At the application layer, a RESTful interface is provided, combining graph computing and Bayesian networks to perform risk-oriented recursive retrieval, ultimately outputting structured results containing cascading impact probabilities.

[0092] Example 2 This embodiment focuses on the scenario of cross-border biopharmaceutical, such as cold chain logistics for vaccines or blood products, and describes in detail the specific implementation steps of "establishing two-way pointer links" and "calculating the probability of cascading effects". In this scenario, the data sources involve multinational multilingual documents, high-frequency temperature control sensor data, and strict compliance inspection documents.

[0093] When processing a batch of imported biological agents, the system faces the problem of a lack of unified association keys between standardized enterprise resource planning records in the relational database and unstructured third-party logistics documents (such as scanned air waybills and handwritten handover forms); the system performs the following steps to establish a deep connection: First, a heterogeneous graph network of nodes is constructed. The system instantiates purchase orders from the enterprise's internal system into structured entity nodes, whose attributes include "Material Name: Recombinant Protein Vaccine", "Temperature Control Requirement: -70 degrees Celsius ± 2 degrees", and "Supplier Name: Biopharmaceutical Co., Ltd." Simultaneously, the system parses a scanned portable document of unknown origin, extracts the text block "Reference Number: 8892 / BP / EXP" located in the upper right corner of the document and the red stamp area at the bottom of the document, and constructs them into document entity nodes.

[0094] Furthermore, the system initializes the graph neural network, establishes potential connection edges between structured entity nodes and document entity nodes, and performs the following calculations: First, semantic feature calculation is performed. The system calculates the semantic embedding vector cosine similarity between the supplier name in the structured node attributes and the header text parsed from the document node (such as the English expression "Biopharmaceutical Logistics Department"). The calculation results show that the two have extremely high semantic similarity. Secondly, by performing rule feature calculations, the system identifies that the order number sequence in the structured node is a subset of the document node reference number string, and assigns the maximum weight to the rule matching. Finally, the spatiotemporal constraint calculation was performed. The structured record showed the expected delivery time as "10:00 AM on November 20th", while the signing timestamp extracted from the document node was "2:30 PM on November 20th" and the location coordinates were resolved to Airport C, which is the location of the supplier. The time difference between the two is within a reasonable threshold of four hours, and the spatial logic is consistent.

[0095] After three layers of graph convolution feature aggregation, the system calculates the matching probability score of the two nodes belonging to the same business entity based on the above features. Since the score exceeds the preset high confidence threshold, the system establishes a bidirectional pointer at the bottom layer of the graph database: the forward pointer points directly from the purchase order record to the coordinate area of ​​the "Special Goods Operation Instructions" paragraph on the second page of the document file; the reverse pointer directly indexes from the document metadata to the time-series database form containing the temperature data of the entire life cycle of this batch of vaccines. Furthermore, based on Bayesian networks, cascaded risk retrieval: when quality inspectors input a natural language query command such as "investigate the cause of the inactivation of a specific batch of vaccines," the system executes the following mixed traversal and inference steps: Step 1: Combine traversal and pruning The system uses this specific batch number as the initial anchor point and spreads upstream and downstream along the bidirectional pointer established above; For the logistics path branch, the system traverses to the "airport cold storage" node and retrieves the IoT ambient temperature data for that time period; the system detects that there is a data gap of up to 30 minutes within a specific time period, indicating that the sensor is offline; for the document path branch, the system jumps along the bidirectional pointer to the associated "airport ground handover form" document node; the system performs semantic analysis on the handwritten notes area of ​​the document and identifies key phrases containing the meaning of "ice addition delay"; Meanwhile, for the financial path branch, the system calculates its semantic relevance coefficient with "active failure". Since the result is lower than the truncation threshold, the system immediately stops searching for that branch to save resources.

[0096] Step 2: Bayesian Risk Inference The system constructs a local Bayesian network, mapping "missing sensor data," "abnormal document records," "exceeding actual temperature limits," and "vaccine activity failure" to variable nodes in the network. Prior probabilities are set based on historical data, and the retrieved evidence is input into the model: the state of "missing sensor data" is true, and the state of "documents containing ice-adding delay semantics" is also true. Inference is made based on a conditional probability table, the logic being that even if the sensor has no data records due to a malfunction in the case of "ice-adding delay," the probability of actual temperature drift is still extremely high. Based on this, the system deduces downstream nodes and calculates the cascading impact probability of "vaccine activity failure" to a relatively high value of 0.88.

[0097] Step 3: Output the results The system outputs search results, presenting the handwritten record of "Ice Addition Delay" in the "Airport Ground Handover Form" as the highest-risk evidence, and highlighting the sensor data breakpoints corresponding to that time period. The system prompts the user: Although the IoT data did not trigger an alarm (due to data loss), based on document semantic reasoning, this step is the root cause of vaccine failure, and its comprehensive risk score ranks first in the search list.

[0098] This embodiment is applied to the cross-border pharmaceutical cold chain scenario. By fusing semantic and spatiotemporal features through graph convolutional networks, a bidirectional pointer link is established between standardized records and handwritten documents. A risk model is constructed using Bayesian networks. In the case of missing sensor data, the probability of cascading effects leading to vaccine failure is inferred by combining document semantics, thereby achieving accurate positioning and intelligent early warning of hidden business risks.

[0099] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method and system for real-time integration and processing of supply chain data based on multi-source heterogeneous data, characterized in that: include: Acquire business transaction data streams, production environment data, logistics trajectory data, and external compliance data documents; For the external compliant data document, a document layout analysis model based on a multimodal transformer is invoked; the semantic vector features of the text and the two-dimensional spatial coordinate features in the document are extracted and fused to generate a composite index key; based on the composite index key, cluster analysis is performed to identify aggregated text blocks in the document, and the information in the aggregated text blocks is mapped to standard attribute fields to generate structured metadata; The business transaction data stream is discretized, a unique identifier is extracted as the index primary key, and an associated database is constructed with production environment data and logistics trajectory data as associated attributes; using the structured metadata, cross-modal entity alignment is performed, and a bidirectional pointer link is established between the unstructured document index and the associated database to form a data dependency path; Receive the query vector, execute a recursive traversal algorithm along the data dependency path based on the bidirectional pointer link, perform node matching and threshold verification in the retrieval path, calculate the cascading influence probability as the relevance ranking score and output the retrieval results.

2. The method and system for real-time integration and processing of supply chain data based on multi-source heterogeneous data as described in claim 1, characterized in that, The invocation of the document layout analysis model based on the multimodal transformer includes: A dual-stream neural network architecture, comprising visual feature extraction and text semantic extraction branches, is employed. The visual feature extraction branch performs convolution operations on document images to capture visual texture features such as table borders, paragraph spacing, and stamp positions. The text semantic extraction branch performs word embedding processing on character sequences within the document to capture contextual linguistic logic features. The visual texture features and linguistic logic features are then input into a cross-attention module to perform feature alignment operations. The correlation weight matrix between visual regions and text content is calculated to obtain a document structure map, thereby achieving the segmentation and localization of unstructured information.

3. The method and system for real-time integration and processing of supply chain data based on multi-source heterogeneous data as described in claim 1, characterized in that, The steps for generating composite index keys through fusion specifically include: The distribution density values ​​of text semantic vector features in the semantic space and the distance of two-dimensional spatial coordinate features from the geometric center in the document layout are calculated respectively. According to the preset field importance rules, the fusion coefficients of semantic features and spatial features are dynamically allocated. Key numerical fields, including contract amount and date, are given higher semantic weights, and position-sensitive fields, including signatures, headers and footers, are given higher spatial weights. The semantic features and spatial features are merged by weighted concatenation and then dimensionality reduction is performed through a fully connected layer to obtain the composite index key.

4. The method and system for real-time integration and processing of supply chain data based on multi-source heterogeneous data as described in claim 1, characterized in that, The steps of performing cluster analysis based on the composite index key specifically include: A density-based adaptive spatial clustering algorithm is employed, using the composite index key as input, to calculate the distance metric between text blocks within the document. Text blocks with a distance less than a preset threshold are grouped into the same semantic cluster, identifying the title region, body text region, and table data region. For the identified table data region, a grid segmentation based on row and column projection is used to separate the content of individual cells. Using a pre-trained semantic matching model, the similarity score between the content of each semantic cluster and the definition description of the standard attribute field is calculated. The attribute corresponding to the highest score is selected as the mapping target to generate standardized key-value pair data.

5. The method and system for real-time integration and processing of supply chain data based on multi-source heterogeneous data according to claim 1, characterized in that, The process of discretizing the business transaction data stream, extracting a unique identifier as the index primary key, and constructing an association database with production environment data and logistics trajectory data as associated attributes includes: A data alignment mechanism based on a spatiotemporal sliding window is constructed. Taking the order number in the business transaction data stream as the benchmark anchor point, a dynamic time window covering the production cycle and logistics cycle is set. Within the dynamic time window, the production environment data is resampled and interpolated to fill in the time discontinuities caused by sensor failures. The logistics trajectory data is map matched and denoised to correct the drifted positioning coordinates. The processed continuous production status curve and discrete logistics trajectory points are attached to the order number according to the timestamp sequence, and a hierarchical data storage structure is constructed to obtain an associated database.

6. The method and system for real-time integration and processing of supply chain data based on multi-source heterogeneous data according to claim 1, characterized in that, The steps for establishing a bidirectional pointer link between the unstructured document index and the associated database specifically include: The unstructured document index is a document feature index built based on the composite index key; Based on graph neural networks, entity objects in the associated database and extracted entities in the structured metadata are respectively constructed as nodes in the graph network. The attribute similarity, co-occurrence relationship and business logic dependency between nodes are constructed as edges in the graph network. Node features are propagated through multi-layer graph convolution operations, and the matching probability score between heterogeneous nodes is calculated. When the score exceeds the confidence threshold, a bidirectional pointer link is generated between the two nodes.

7. The method and system for real-time integration and processing of supply chain data based on multi-source heterogeneous data according to claim 1, characterized in that, The steps of receiving the query vector and outputting the search results specifically include: A hybrid traversal strategy combining depth-first search and breadth-first search with pruning is adopted. Starting from the initial node hit by the received query vector, the algorithm expands to upstream and downstream nodes along bidirectional pointer links. During each jump, the relevance score coefficient between the current path node and the query intent is calculated in real time. Once the coefficient is lower than the preset truncation threshold, the search of the current branch is stopped immediately to perform pruning operation. At the same time, the path nodes are checked to see if they meet the user permission verification rules, and nodes that are not authorized to access are blocked. The algorithm obtains a retrieval subgraph containing directly related data and indirectly potentially influential data. A Bayesian network risk propagation model is established based on the retrieval subgraph. Each node on the path is mapped to a variable node in the network. The conditional probability relationship between nodes is constructed based on historical interruption data. The posterior probability of causing downstream delivery anomalies is calculated as the cascading impact probability based on the query event trigger probability. The probability is weighted and fused with semantic similarity to generate a comprehensive score. The retrieval results containing the priority of business risks are output according to the score.

8. A real-time supply chain data integration and processing system based on multi-source heterogeneous data, characterized in that, The system includes: Data acquisition module: Acquires business transaction data streams, production environment data, logistics trajectory data, and external compliance data documents; Fusion Clustering Module: For the external compliant data document, it calls the document layout analysis model based on a multimodal transformer; extracts the semantic vector features of the text and the two-dimensional spatial coordinate features in the document, and fuses them to generate a composite index key; performs clustering analysis based on the composite index key, identifies aggregated text blocks in the document, and maps the information in the aggregated text blocks to standard attribute fields to generate structured metadata; Database construction module: Discretizes the business transaction data stream, extracts a unique identifier as the index primary key, and constructs an associated database with production environment data and logistics trajectory data as associated attributes; using the structured metadata, performs cross-modal entity alignment, and establishes a bidirectional pointer link between the unstructured document index and the associated database to form a data dependency path; Query module: Receives the query vector, executes a recursive traversal algorithm along the data dependency path based on the bidirectional pointer link, performs node matching and threshold verification in the retrieval path, calculates the cascading influence probability as the relevance ranking score, and outputs the retrieval results.

Images