An intelligent digital twin system of wholesale market fusing perception data

Abstract

The application relates to the technical field of digital twinning and risk deduction, and discloses a wholesale market intelligent digital twinning system fusing sensing data, which comprises an edge sensing layer, a central intelligent layer and an application service layer. The edge sensing layer collects multi-source heterogeneous data and generates a virtual state matrix containing visual, internet of things and business modal characteristics. The central intelligent layer constructs a cross-modal spatiotemporal knowledge graph network structure according to the virtual state matrix, extracts modal characteristics of a target entity node, performs tensor fusion calculation, outputs a joint representation vector, extracts neighbor nodes, performs feature aggregation calculation, outputs a context context vector, splices the joint representation vector and the context context vector, maps them into structured prompt word characteristics, inputs a large language model, performs joint probability deduction calculation, and outputs risk deduction result data. The application service layer generates a dynamic response strategy for the target node according to the deduction result data. The application improves the implicit risk early warning accuracy and the safety operation closed loop efficiency.

Description

Technical Field

[0001] This invention relates to the field of digital twin and risk simulation technology, specifically to an intelligent digital twin system for wholesale markets that integrates sensing data. Background Technology

[0002] In existing wholesale market management systems, video surveillance networks, IoT sensing devices, and business transaction systems typically operate independently, with data interaction barriers between these systems. Current risk monitoring methods primarily rely on independent threshold alarms from single data sources, lacking the ability to deeply fuse features from multi-source heterogeneous data such as visual images, IoT sensors, and business records. Because it is impossible to calculate the interaction characteristics of different modalities across multiple dimensions, existing systems struggle to detect hidden risks arising from the interplay of physical space occupancy, environmental monitoring anomalies, and business transaction stagnation. This results in the system failing to issue warnings when a single sensor does not trigger an alarm, leading to low overall accuracy in risk warnings.

[0003] Existing risk assessment models typically analyze each physical entity as an isolated object, failing to effectively transform spatial adjacency, spatial inclusion, and business flow relationships between physical entities. Due to the lack of topological association information of entity nodes and their surrounding environment during the simulation process, existing simulation models cannot extract contextual information by combining the operational characteristics of adjacent entity nodes or devices. This results in risk assessment being detached from the actual physical and business topology environment, leading to poor reliability of risk simulation results.

[0004] Existing systems typically output static alarm prompts after triggering risk warnings, lacking automated intervention and state synchronization mechanisms for complex business scenarios. Current methods cannot automatically extract and instantiate dynamic handling strategies containing specific control parameters based on the spatial coordinates and IoT characteristics of specific entities. After on-site physical intervention, existing systems also cannot automatically use execution feedback data to synchronize and update the historical state of the digital model and re-execute inference calculations to verify whether the risk has truly been eliminated. This lack of underlying execution logic prevents the system from forming a complete control loop from risk discovery and strategy distribution to model state synchronization, resulting in low efficiency in security operation and maintenance response. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent digital twin system for wholesale markets that integrates sensory data. This system solves the problems in existing technologies, such as low accuracy of hidden risk warnings due to isolated multi-source heterogeneous data, risk projection being detached from the physical and business topology environment, and the lack of automated status synchronization and intelligent closed-loop mechanisms for risk handling.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides an intelligent digital twin system for wholesale markets that integrates sensory data, comprising an edge sensing layer, a central intelligence layer, and an application service layer. The edge sensing layer collects multi-source heterogeneous data within the physical space and uses this data to generate a virtual state matrix containing visual modal features, IoT modal features, and business modal features. The central intelligence layer constructs a cross-modal spatiotemporal knowledge graph network structure based on the virtual state matrix. This structure includes a set of entity nodes configured with node category identifiers and a set of topological edges configured with relation type labels. The central intelligence layer extracts the visual modal features, IoT modal features, and business modal features of target entity nodes from the set of entity nodes, performs tensor fusion calculations on these features, and outputs a joint representation vector of the target entity node. The central intelligence layer extracts the neighbor nodes of the target entity node within the cross-modal spatiotemporal knowledge graph network structure and performs feature aggregation calculations, outputting a contextual vector. The central intelligent layer concatenates the joint representation vector and the context vector to generate a comprehensive state representation vector, and maps this comprehensive state representation vector to structured cue word features. The central intelligent layer inputs these structured cue word features into a large language model to perform implicit risk joint probability inference calculations, outputting risk inference result data. The application service layer receives the risk inference result data and generates a dynamic response strategy for the target entity node based on the risk category identifier index in the risk inference result data.

[0007] In constructing a cross-modal spatiotemporal knowledge graph network structure, this invention uses a central intelligent layer to calculate the absolute distance between different entity nodes in the virtual state matrix. When the absolute distance is less than a set adjacent distance threshold, a spatial adjacency relationship is determined between the entity nodes. A spatial containment relationship is determined when the coordinate boundary of one entity node is completely within the coordinate boundary of another entity node. The central intelligent layer compares the business association identifier sequences of different entity nodes. A goods ownership relationship is determined when the business association identifier sequences contain the same goods identification code, and a transaction association relationship is determined when the business association identifier sequences contain the same transaction association code. Through these determination mechanisms, the system aligns the physical space state with the business flow state. Furthermore, the central intelligent layer extracts the timestamp of the discrete time step as a time dimension parameter and appends it to the data record for temporal marking. When the business state parameter changes from "in a transaction" to "transaction completed," the directed edges of the corresponding transaction association relationships are deleted, achieving temporal tracking and dynamic updating of the graph network structure.

[0008] This invention performs feature fusion processing on multi-source heterogeneous data. The central intelligent layer converts visual modality features, IoT modality features, and business modality features into floating-point arrays, and appends a scalar element with a value of 1 to the end of the floating-point arrays, generating augmented spatial visual feature vectors, augmented logistics IoT feature vectors, and augmented transaction business feature vectors. The central intelligent layer performs tensor outer product operations on the aforementioned three augmented feature vectors to generate a multimodal fusion tensor. The tensor outer product operation calculates the product of elements in each dimension, obtaining independent first-order features of the visual modality, IoT modality, and business modality, as well as second- and third-order interaction features of the multimodal combination. The central intelligent layer flattens the multimodal fusion tensor into a one-dimensional multimodal feature vector, and performs a linear projection operation on the multimodal feature vector and weight matrix to obtain a linear mapping vector. The central intelligent layer adds the linear mapping vector to the bias vector and inputs it into a nonlinear activation function for numerical mapping processing, finally outputting a joint representation vector. The aforementioned calculation process extracts cross-modal interaction features.

[0009] This invention utilizes graph neural network aggregation operators to extract contextual vectors. The central intelligent layer uses graph neural network aggregation operators to assign corresponding neighbor weights to the first-order and second-order neighbor nodes of the target entity node based on the relation type label. The central intelligent layer then performs a weighted summation of the visual modal features, IoT modal features, and business modal features of the first-order and second-order neighbor nodes according to the neighbor weights, obtaining the feature aggregation result and outputting it as a contextual vector. The central intelligent layer uses a feature projection layer to linearly map the comprehensive state representation vector, converting the numerical value of the comprehensive state representation vector into a word embedding vector format to generate structured prompt word features, thus achieving a mapping and conversion to the data format of the large language model input layer.

[0010] This invention utilizes a large language model to perform joint probability inference calculations for implicit risks. The central intelligent layer injects structured cue word features into the input layer of the large language model, invokes the multi-head self-attention mechanism module and feedforward neural network module within the large language model to perform forward propagation calculations, and outputs hidden layer feature vectors. The central intelligent layer inputs the hidden layer feature vectors into a normalized exponential function operator to calculate the probability distribution values ​​of the hidden layer feature vectors at each risk category identifier index in a preset set of implicit risk categories. Based on the risk category identifier index, the central intelligent layer matches a set threshold in the risk trigger threshold matrix; when the inferred probability value is greater than the corresponding set threshold, it determines that the target entity node has implicit risk.

[0011] This invention generates dynamic response strategies based on a preset response strategy database. The application service layer extracts a basic handling logic sequence containing variable placeholders from the preset response strategy database based on the risk category identifier index in the risk simulation result data. The application service layer extracts the spatial coordinate boundaries, business modal features, and IoT modal features of first-order neighbor nodes identified as device nodes from the target entity node, and maps these features to the variable placeholders corresponding to the basic handling logic sequence, generating a dynamic response strategy that includes security inspection work orders and control signaling.

[0012] This invention further performs state synchronization and closed-loop verification operations on the operation and maintenance side and the value-added side. On the operation and maintenance side, the central intelligent layer receives field execution feedback data and equipment execution feedback data to generate a state synchronization trigger signal, extracts continuously aligned data frames, and obtains the perceived data update value. The central intelligent layer uses the perceived data update value to replace the corresponding historical value, updates the cross-modal spatiotemporal knowledge graph network structure, and re-calls the large language model to perform joint probability inference calculation. When the recalculated inference probability value is less than or equal to the matched set threshold, a risk elimination confirmation signal is generated. On the value-added side, the value-added service layer extracts the issuance timestamp and completion timestamp of the security inspection work order from the field execution feedback data to calculate the risk response timeliness value and counts the risk trigger frequency value. The value-added service layer uses the credit evaluation quantification model to perform linear weighted calculation based on preset weight coefficients, outputs a dynamic credit rating score, and matches the supply chain finance adjustment strategy containing credit limit adjustment parameters and the insurance rate adjustment strategy containing the basic premium multiplier in the value-added strategy database.

[0013] This invention provides an intelligent digital twin system for wholesale markets that integrates sensor data. It has the following beneficial effects: 1. This invention extracts visual modal features, IoT modal features, and business modal features of entity nodes through a central intelligent layer, and performs tensor fusion calculation to output a joint representation vector. This mechanism utilizes tensor outer product operations to calculate the product of different modal features across various dimensions, extracting cross-modal interaction features of physical space occupancy, environmental monitoring, and business transaction status. This feature fusion method breaks down data silos between video surveillance, IoT sensing, and business systems in traditional wholesale markets, enabling the system to accurately deduce hidden risks that a single physical sensor cannot trigger an alarm based on the joint distribution of multi-source heterogeneous data, thus improving the overall accuracy of risk warning.

[0014] 2. Based on the construction of a cross-modal spatiotemporal knowledge graph network structure, this invention utilizes graph neural network aggregation operators to extract the contextual vectors of target entity nodes and their neighboring nodes, and concatenates these vectors with joint representation vectors to map them into structured prompt word features, which are then input into a large language model. This processing structure transforms the spatial adjacency, inclusion, and business flow relationships between physical entities into computable topological weight information. The large language model, combined with the aforementioned input vectors containing topologically related environmental features, performs probabilistic inference calculations, effectively avoiding isolated risk assessments detached from the actual physical environment and improving the reliability of risk inference results.

[0015] 3. This invention, through the application service layer, extracts a basic handling logic sequence containing variable placeholders from a preset response strategy database based on risk simulation results. It then replaces the feature mappings of the target entity node and its first-order neighbor nodes into the variable placeholders to generate a dynamic response strategy. Simultaneously, the central intelligent layer updates the cross-modal spatiotemporal knowledge graph network structure based on on-site execution feedback data and re-executes the simulation calculation to verify the risk elimination status. This design achieves a complete intelligent closed loop from underlying risk discovery, automatic generation of security inspection work orders and equipment control signaling, to the synchronization of the digital twin model status after physical intervention, effectively improving the efficiency of security operation and maintenance response in complex business scenarios. Attached Figure Description

[0016] Figure 1 This is a system architecture diagram of the present invention. Detailed Implementation

[0017] The technical solutions in 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 the appendix Figure 1 This invention provides an intelligent digital twin system for wholesale markets that integrates sensing data, comprising: an edge sensing layer, a central intelligence layer, an application service layer, and a value-added service layer.

[0019] The edge perception layer is deployed within the physical space of the wholesale market to collect multi-source heterogeneous data from physical entities. This layer includes image acquisition devices and IoT sensor devices. Image acquisition devices, including edge computing devices and cameras, are used to collect customer flow data, vehicle flow data, and data on violations. IoT sensor devices, including RFID devices, are used to collect transaction flow data, logistics trajectory data, and inventory status data. The edge perception layer performs unified format conversion and time window alignment preprocessing on the collected multi-source heterogeneous data before uploading it to the central intelligent layer.

[0020] The central intelligent layer communicates with the edge perception layer to process and jointly infer the received multi-source heterogeneous data. The central intelligent layer includes a digital twin construction module, a cross-modal knowledge graph module, and a large language model inference module. The digital twin construction module maps the received multi-source heterogeneous data to physical market entities, extracts corresponding spatial and business attributes based on the entity's unique identifier, and constructs a corresponding virtual digital entity model. The cross-modal knowledge graph module, based on the virtual digital entity model, extracts feature vectors of physical entities in spatial, logistical, and business dimensions, and uses a tensor fusion network to calculate the joint representation vector of the physical entity. The large language model inference module receives the joint representation vector, combines it with the graph structure information to calculate the probability of occurrence of different types of risk events, and generates corresponding handling strategies.

[0021] The application service layer communicates with the central intelligence layer to execute closed-loop operations related to physical environment maintenance. The application service layer is configured with large-screen display terminals, computer maintenance terminals, and mobile terminals. It receives the handling strategies generated by the central intelligence layer, dispatches corresponding task orders to the large-screen display terminals, computer maintenance terminals, or mobile terminals based on risk categories, and simultaneously records the processing progress information of the task orders.

[0022] The value-added service layer communicates with the central intelligence layer to execute value-added operations for business management. The value-added service layer has a data platform that receives entity joint representation vectors and risk probability data output from the central intelligence layer to construct operational profiles for each physical entity. Based on these operational profiles, the value-added service layer sends supply chain finance connection information and transaction matching information to physical entity terminals that meet preset trigger conditions.

[0023] Based on the aforementioned system architecture, the edge perception layer performs multi-source heterogeneous data collection operations in the physical space, providing basic data input for the central intelligent layer.

[0024] The edge perception layer comprises image acquisition devices and IoT sensor devices. The image acquisition devices capture consecutive visual image frames of the target area. They execute target detection algorithms to extract features from these consecutive visual image frames, obtaining spatial visual features. These spatial visual features include the spatial coordinates and area occupied by the target entity. The image acquisition devices arrange these spatial visual features in chronological order, forming a visual perception data sequence.

[0025] IoT sensor devices include RFID readers and environmental sensors. RFID readers read data from RFID tags to obtain cargo identification codes and location information. Environmental sensors collect temperature and humidity values ​​of the target area. The IoT sensor devices arrange the location information, temperature values, and humidity values ​​in chronological order to form an IoT sensing data sequence.

[0026] The central intelligence layer connects to the business system database via a network interface. It extracts structured business records from the database. These records include order creation time, inventory changes, and transaction status. The central intelligence layer then arranges these records in chronological order to form a sequence of business transaction data.

[0027] The central intelligent layer receives visual perception data sequences, IoT perception data sequences, and business transaction data sequences. To address the different sampling frequencies of image acquisition devices, IoT sensor devices, and business system databases, the central intelligent layer sets a unified reference time alignment window.

[0028] The central intelligence layer divides the visual perception data sequence, IoT perception data sequence, and business transaction data sequence into time axes, using the start and end points of the reference time alignment window as boundaries. The central intelligence layer extracts spatial visual features, location status information, temperature values, humidity values, and structured business records that fall within the same reference time alignment window.

[0029] For continuous data with multiple sampled values ​​within the baseline time alignment window, the central intelligent layer performs aggregation operations by calculating the average value to obtain a unique representative value. For discrete state data with multiple sampled values ​​within the baseline time alignment window, the central intelligent layer extracts the last sampled value as the unique representative value. For data with missing sampled values ​​within the baseline time alignment window, the central intelligent layer performs interpolation operations by calling the corresponding value from the previous baseline time alignment window to complete the value. The central intelligent layer combines the spatial visual features, location status information, temperature values, humidity values, and structured business records processed by aggregation and interpolation operations into a unified aligned data frame for the corresponding discrete time step. The central intelligent layer stores the continuous aligned data frames in the data platform and provides them for use by the digital twin construction module.

[0030] After the central intelligent layer stores the continuous alignment data frames to the data platform, the central intelligent layer calls the aforementioned alignment data frames to further perform the mapping of physical entities to virtual space and the continuous state synchronization operation.

[0031] The central intelligence layer establishes a set of physical market entities within the digital twin building module. This set includes fixed stalls, public passageways, cold storage equipment, and mobile vehicles. The central intelligence layer assigns a globally unique identifier to each physical entity in the set. Based on these globally unique identifiers, the central intelligence layer constructs a set of virtual digital entities in virtual space that correspond one-to-one with the structure of the physical market entity set.

[0032] The central intelligence layer configures an attribute vector for each virtual digital entity in the set of virtual digital entities. The attribute vector includes spatial size parameters, business status parameters, and environmental monitoring parameters. The central intelligence layer reads aligned data frames stored in the data platform. It parses the aligned data frames, extracting spatial visual features, location status information, temperature values, humidity values, and structured business records. The central intelligence layer maps spatial visual features and location status information to spatial size parameters, temperature and humidity values ​​to environmental monitoring parameters, and structured business records to business status parameters. The central intelligence layer then writes the mapped parameters into the attribute vector of the corresponding virtual digital entity using a globally unique identifier.

[0033] The central intelligent layer updates the state of the virtual digital entity set continuously in discrete time steps using state update instructions. The central intelligent layer defines the virtual state matrix as a matrix composed of the attribute vectors of all virtual digital entities. The central intelligent layer retrieves the virtual state matrix from the previous discrete time step and extracts the aligned data frame for the current discrete time step.

[0034] The central intelligent layer parses the aligned data frame of the current discrete time step, extracting the globally unique identifier and the corresponding perceived data update value carried in the aligned data frame. The central intelligent layer matches the corresponding virtual digital entity in the virtual state matrix and replaces the historical value of the corresponding attribute vector in the corresponding virtual digital entity with the perceived data update value. After completing the value replacement of the attribute vector, the central intelligent layer combines them to obtain the virtual state matrix of the current discrete time step. The aforementioned parsing, matching, and replacement process realizes the real-time synchronous update of the physical market entity state and the virtual digital entity state.

[0035] The central intelligence layer transmits the virtual state matrix of the current discrete time step to the application service layer. The large-screen display terminal of the application service layer receives the virtual state matrix of the current discrete time step and extracts the spatial dimension parameters and business state parameters recorded in the virtual state matrix. The large-screen display terminal converts the spatial dimension parameters into three-dimensional coordinates and the business state parameters into primitive state labels. Based on the three-dimensional coordinates and primitive state labels, it renders the three-dimensional model parameters to achieve information visualization. Simultaneously, the central intelligence layer transmits the virtual state matrix of the current discrete time step to the cross-modal knowledge graph module, serving as the basic input for subsequent data fusion and risk inference in the cross-modal knowledge graph module.

[0036] After the cross-modal knowledge graph module receives the virtual state matrix for the current discrete time step, the central intelligent layer further parses the virtual state matrix to construct a structured cross-modal spatiotemporal knowledge graph. The central intelligent layer sets the entity node set, topological edge set, relation type set, and time dimension parameter in the cross-modal spatiotemporal knowledge graph.

[0037] The central intelligence layer establishes a set of entity nodes based on the aforementioned set of virtual digital entities. Each entity node in the entity node set corresponds to a virtual digital entity in the set of virtual digital entities. The central intelligence layer assigns a node category identifier to each entity node. Node category identifiers include merchant nodes, region nodes, goods nodes, and equipment nodes. The central intelligence layer stores the spatial size parameters, environmental monitoring parameters, and business status parameters recorded in the virtual state matrix as the visual modal features, IoT modal features, and business modal features of the corresponding entity nodes, respectively.

[0038] The central intelligent layer parses the spatial coordinate set and business association identifier sequence in the virtual state matrix to establish a topological edge set. The central intelligent layer calculates the absolute distance between different entity nodes in the spatial coordinate set. When the absolute distance is less than a set adjacency threshold, the central intelligent layer determines that there is a spatial adjacency relationship between entity nodes; when the coordinate boundary of one entity node is completely inside the coordinate boundary of another entity node, it determines that there is a spatial containment relationship between entity nodes. The central intelligent layer compares the business association identifier sequences of different entity nodes; when the business association identifier sequences of different entity nodes have the same identifier code, it determines that there is a goods ownership relationship or a transaction association relationship between entity nodes. The topological edge set contains directed edges connecting different entity nodes. The central intelligent layer assigns a relationship type label from the relationship type set to each directed edge. The relationship type labels include spatial containment, spatial adjacency, goods ownership, and transaction association relationships. The central intelligent layer connects region nodes and device nodes through spatial containment and spatial adjacency relationships. The central intelligent layer connects merchant nodes and goods nodes through goods ownership and transaction association relationships.

[0039] The central intelligent layer uses a time dimension parameter to perform temporal labeling on entity nodes and directed edges. It extracts the timestamp of the current discrete time step as a time dimension parameter and appends it to the data records of entity nodes and directed edges. The central intelligent layer receives updates to the virtual state matrix for consecutive discrete time steps. Based on the progression of discrete time steps, the central intelligent layer updates the visual modal features, IoT modal features, and business modal features of entity nodes. Based on changes in business state parameters, the central intelligent layer dynamically adds or deletes directed edges between entity nodes. The central intelligent layer combines the temporally labeled and dynamically updated set of entity nodes and topological edges to generate a cross-modal spatiotemporal knowledge graph network structure corresponding to the discrete time step, which is then provided to subsequent processing modules for tensor fusion computation.

[0040] After the cross-modal knowledge graph module receives the virtual state matrix for the current discrete time step, the central intelligent layer further parses the virtual state matrix to construct a structured cross-modal spatiotemporal knowledge graph. The central intelligent layer sets the entity node set, topological edge set, relation type set, and time dimension parameter in the cross-modal spatiotemporal knowledge graph.

[0041] The central intelligence layer establishes a set of entity nodes based on the aforementioned set of virtual digital entities. Each entity node in the entity node set corresponds to a virtual digital entity in the set of virtual digital entities. The central intelligence layer assigns a node category identifier to each entity node. Node category identifiers include merchant nodes, region nodes, goods nodes, and equipment nodes. The central intelligence layer stores the spatial size parameters, environmental monitoring parameters, and business status parameters recorded in the virtual state matrix as the visual modal features, IoT modal features, and business modal features of the corresponding entity nodes, respectively.

[0042] The central intelligent layer parses the set of spatial coordinates and business association identifier sequences in the virtual state matrix to establish a set of topological edges. The central intelligent layer calculates the absolute distances between different entity nodes in the spatial coordinate set. When the absolute distance is less than a set adjacent distance threshold, the central intelligent layer determines that there is a spatial adjacency relationship between the entity nodes. The central intelligent layer parses spatial size parameters to obtain the coordinate boundaries of the corresponding entity nodes. When the coordinate boundary of one entity node is completely inside the coordinate boundary of another entity node, the central intelligent layer determines that there is a spatial containment relationship between the entity nodes.

[0043] The central intelligent layer compares the business association identifier sequences of different entity nodes. The business association identifier sequence includes a goods identifier code and a transaction association code. When the central intelligent layer finds the same goods identifier code in the business association identifier sequences of different entity nodes, it determines that there is a goods ownership relationship between the entity nodes; when the same transaction association code exists in the business association identifier sequences of different entity nodes, it determines that there is a transaction association relationship between the entity nodes. The topological edge set contains directed edges connecting different entity nodes. The central intelligent layer assigns a relationship type label from the relationship type set to each directed edge. The relationship type labels include spatial inclusion relationship, spatial adjacency relationship, goods ownership relationship, and transaction association relationship. The central intelligent layer connects region nodes and device nodes through spatial inclusion relationships and spatial adjacency relationships. The central intelligent layer connects merchant nodes and goods nodes through goods ownership relationships and transaction association relationships.

[0044] The central intelligence layer uses a time dimension parameter to perform temporal labeling on entity nodes and directed edges. It extracts the timestamp of the current discrete time step as a time dimension parameter and appends it to the data records of entity nodes and directed edges. The central intelligence layer receives virtual state matrix updates for consecutive discrete time steps. Based on the progression of discrete time steps, the central intelligence layer updates the visual modal features, IoT modal features, and business modal features of entity nodes.

[0045] The central intelligent layer dynamically adds or deletes directed edges between entity nodes based on changes in business state parameters. When a business state parameter changes from "in a transaction" to "transaction completed," the central intelligent layer deletes the corresponding directed edges related to the transaction. The central intelligent layer combines the temporally labeled and dynamically updated set of entity nodes and topological edges to generate a cross-modal spatiotemporal knowledge graph network structure corresponding to the discrete time step, which is then provided to subsequent processing modules for tensor fusion computation.

[0046] After generating the aforementioned cross-modal spatiotemporal knowledge graph network structure corresponding to discrete time steps, the central intelligent layer performs heterogeneous modality feature vector extraction based on the cross-modal spatiotemporal knowledge graph network structure. The central intelligent layer then identifies the target entity node for risk inference within the cross-modal spatiotemporal knowledge graph network structure. The central intelligent layer extracts the visual modality features, IoT modality features, and business modality features stored in the target entity node.

[0047] The central intelligent layer analyzes visual modal features. It extracts the spatial area and stacking height values ​​from these features. The central intelligent layer concatenates these values ​​according to a predefined dimensional order, generating a floating-point array with a first fixed dimension. This first fixed-dimensional floating-point array is defined as the spatial visual feature vector. .

[0048] The central intelligent layer analyzes IoT modal features. It extracts location status information and temperature values ​​from these features. The central intelligent layer extracts the timestamps corresponding to the location status information and calculates the RFID dwell time of the target entity node based on the timestamps. The central intelligent layer calls upon the system's built-in preset temperature standard value, calculates the difference between the temperature value and the preset temperature standard value, and obtains the temperature deviation value. The central intelligent layer concatenates the RFID dwell time value and the temperature deviation value according to a preset dimensional order to generate a floating-point array with a second fixed dimension. The central intelligent layer defines this second fixed-dimensional floating-point array as the logistics IoT feature vector. .

[0049] The central intelligence layer analyzes business modal features. It extracts order creation time, inventory change values, and transaction flow status from these features. Based on the order creation time and inventory change values, the central intelligence layer calculates the current sales rate of the target entity node. It then calls the system's built-in preset sales rate benchmark value, calculates the difference between the current sales rate and the preset benchmark value, and obtains the sales rate deviation value. The central intelligence layer also calculates the order flow frequency value per unit time based on the transaction flow status. Finally, it concatenates the sales rate deviation value and the order flow frequency value according to a preset dimensional order, generating a floating-point array with a third fixed dimension. This third fixed-dimensional floating-point array is defined as the transaction business feature vector. .

[0050] The central intelligent layer for spatial visual feature vectors Logistics IoT Feature Vector and transaction business feature vector Perform scalar augmentation operations. The central intelligent layer performs spatial visual feature vector... Append a scalar element with a value of 1 to the end of the vector to generate an augmented space visual feature vector. The central intelligent layer in the feature vector of logistics IoT Append a scalar element with a value of 1 to the end of the vector to generate an augmented logistics IoT feature vector. The central intelligent layer in the transaction business feature vector Append a scalar element with a value of 1 to the end of the vector to generate an augmented transaction feature vector. .

[0051] The central intelligent layer will augment spatial visual feature vectors Enhanced Logistics IoT Feature Vectors and augmented transaction feature vector It is stored in a memory matrix and provided to the tensor fusion network for outer product fusion calculation.

[0052] In augmenting spatial visual feature vectors Enhanced Logistics IoT Feature Vectors and augmented transaction feature vector After being stored in the memory matrix, the central intelligent layer calls the tensor fusion network to perform cross-modal feature fusion calculation.

[0053] The central intelligent layer extracts augmented spatial visual feature vectors from the memory matrix. Enhanced Logistics IoT Feature Vectors and augmented transaction feature vector The central intelligent layer performs tensor outer product operations on the aforementioned three augmented feature vectors.

[0054] The central intelligence layer calculates the product of elements in each dimension of the three augmented feature vectors using tensor outer product operations. The output of the tensor outer product operation includes independent first-order features of the visual modality, IoT modality, and business modality; second-order interaction features of any two modalities; and third-order interaction features of all three modalities. The central intelligence layer defines the output of the tensor outer product operation as a multimodal fusion tensor.

[0055] The central intelligent layer configures a learnable weight matrix in the tensor fusion network. With bias vector The central intelligent layer flattens the multimodal fusion tensor into a one-dimensional multimodal feature vector. The central intelligent layer then combines the multimodal feature vector with the weight matrix. Perform a linear projection operation to obtain a linearly mapped vector. The central intelligent layer then compares the linearly mapped vector with the bias vector. Perform matrix addition to obtain the affine transformation vector.

[0056] The central intelligent layer is configured with a nonlinear activation function in the tensor fusion network. The central intelligent layer inputs the affine transformation vector into a nonlinear activation function. Numerical mapping is performed within this layer. The central intelligent layer uses a nonlinear activation function. The output is defined as the joint representation vector of the target entity nodes. .

[0057] Joint representation vector executed by the central intelligent layer The calculation formula is as follows: In the above formula, This represents the tensor outer product operator. This represents the tensor flattening operator.

[0058] The central intelligent layer will compute the obtained joint representation vector. This is written into the data structure of the target entity nodes in the cross-modal spatiotemporal knowledge graph. The central intelligent layer extracts the joint representation vector. The target entity node data structure is provided to the large language model inference module as the basic input features for subsequent joint probability inference of implicit risks.

[0059] After the central intelligent layer writes the computed joint representation vector into the data structure of the target entity node in the cross-modal spatiotemporal knowledge graph network structure, the central intelligent layer initiates the implicit risk inference process based on the large language model. The central intelligent layer extracts the target entity node data structure containing the joint representation vector in the large language model inference module and simultaneously reads the cross-modal spatiotemporal knowledge graph network structure at the corresponding discrete time step.

[0060] The central intelligent layer performs graph traversal operations within the cross-modal spatiotemporal knowledge graph network structure. Centered on the target entity node, the central intelligent layer extracts first-order and second-order neighbor nodes within the cross-modal spatiotemporal knowledge graph network structure based on a set hop count threshold. The central intelligent layer extracts the visual modal features, IoT modal features, business modal features, and corresponding relationship type labels of the first-order and second-order neighbor nodes.

[0061] The central intelligent layer invokes the graph neural network aggregation operator. It inputs the visual modal features, IoT modal features, and business modal features of first-order and second-order neighbor nodes, along with their corresponding relationship type labels, into the graph neural network aggregation operator. The central intelligent layer uses the graph neural network aggregation operator to assign corresponding neighbor weights to the first-order and second-order neighbor nodes based on the relationship type labels. Based on these neighbor weights, the central intelligent layer performs a weighted summation of the visual modal features, IoT modal features, and business modal features of the first-order and second-order neighbor nodes to obtain the feature aggregation result. The central intelligent layer outputs the feature aggregation result as a contextual vector for the target entity node. This contextual vector contains neighborhood association information of the target entity node in both physical space and business flow dimensions.

[0062] The central intelligent layer extracts the joint representation vector and context vector of the target entity node. The central intelligent layer then performs a concatenation operation on the joint representation vector and context vector according to a preset feature dimension. The output of the concatenation operation is defined as the comprehensive state representation vector.

[0063] The central intelligent layer is configured with a feature projection layer within the natural language processing framework. The central intelligent layer inputs the comprehensive state representation vector to the feature projection layer for linear mapping. The feature projection layer maps the numerical values ​​of the comprehensive state representation vector to a word embedding vector format readable by the large language model input layer. The central intelligent layer defines the mapped word embedding vector format as structured cue word features.

[0064] The central intelligent layer injects structured cue word features as a preceding continuous input vector into the input layer of the locally deployed large language model. These structured cue word features serve as contextual preconditions for the large language model to perform subsequent implicit risk joint probability inference within the input layer.

[0065] After the central intelligence layer injects structured prompt word features into the input layer of the locally deployed large language model, it uses the large language model to perform joint probability inference calculations of implicit risks. The central intelligence layer pre-defines a set of implicit risk categories within the large language model. This set includes multiple mutually exclusive business flow risk types and physical security risk types. The central intelligence layer assigns a corresponding risk category identifier index to each risk type in the set of implicit risk categories. .

[0066] The central intelligent layer extracts the pre-trained parameters and fine-tuned network parameters configured in the large language model. These parameters are obtained by training the large language model using historical risk event records. The central intelligent layer takes the structured cue word features as the input sequence to the large language model and calls the multi-head self-attention mechanism module and feedforward neural network module within the large language model to perform forward propagation calculations on the structured cue word features. Through forward propagation calculations, the large language model extracts the association weights from the structured cue word features and outputs the hidden layer feature vector corresponding to the target entity node.

[0067] The central intelligent layer configures a normalized exponential function operator in the output layer of the large language model. The central intelligent layer inputs the hidden layer feature vectors calculated from the forward propagation into the normalized exponential function operator for probability space mapping. The normalized exponential function operator calculates the index of each risk category identifier in a predefined set of implicit risk categories for the hidden layer feature vectors. The probability distribution values ​​are defined as follows. The central intelligent layer defines the output of the normalized exponential function operator as the joint probability distribution data of the target entity nodes.

[0068] The central intelligent layer extracts the risk category identifier index from the joint probability distribution data of risks. The corresponding projected probability value. The central intelligent layer retrieves the pre-stored risk trigger threshold matrix from the system database. The risk trigger threshold matrix records the risk category identifier index. The mapping relationship between the risk category and the set threshold. The central intelligent layer indexes the risk category. The central intelligent layer matches the corresponding set threshold in the risk trigger threshold matrix. It then compares each projected probability value with the matched set threshold one by one. When a projected probability value exceeds the corresponding set threshold, the central intelligent layer determines that the target entity node has a corresponding risk category identifier index. The hidden risks it points to.

[0069] The central intelligent layer extracts and determines the risk category identifier index of any latent risks. The central intelligence layer packages the aforementioned information, including the inferred probability value and the globally unique identifier of the target entity node in the cross-modal spatiotemporal knowledge graph, to generate structured risk inference result data. This data is then transmitted to the application service layer and the value-added service layer, providing them with data as the basis for generating response strategies.

[0070] After the risk simulation results are transmitted to the application service layer and the value-added service layer from the central intelligence layer, this embodiment further clarifies the execution process of the joint probability simulation of implicit risks using a specific business scenario within the physical market.

[0071] The central intelligent layer extracts an entity node with the node category identifier of a merchant node as the target entity node within the cross-modal spatiotemporal knowledge graph network structure. The central intelligent layer extracts the visual modality features, IoT modality features, and business modality features corresponding to the target entity node. In the data record of the current discrete time step, the visual modality features corresponding to the target entity node include a stacking height value exceeding a preset benchmark value; the IoT modality features include extended RFID dwell time and increased temperature deviation values; and the business modality features include a negative deviation in sales rate. The central intelligent layer converts the aforementioned visual modality features, IoT modality features, and business modality features into augmented spatial visual feature vectors, augmented logistics IoT feature vectors, and augmented transaction business feature vectors, respectively.

[0072] The central intelligent layer invokes the tensor fusion network to perform tensor outer product operations on the aforementioned three augmented feature vectors, followed by flattening and linear projection operations. The central intelligent layer outputs a joint representation vector of the target entity node. This joint representation vector combines cross-feature information from physical space occupancy, environmental temperature changes, and business transaction stagnation within the data structure.

[0073] The central intelligent layer traverses the cross-modal spatiotemporal knowledge graph network structure to extract first-order neighbor nodes that have spatial adjacency with the target entity node. These first-order neighbor nodes are device nodes categorized as cold storage equipment. The central intelligent layer then uses a graph neural network aggregation operator to perform a weighted summation of the visual modal features, IoT modal features, and business modal features of the device nodes, generating a contextual vector. This contextual vector contains operational features related to local temperature fluctuations caused by exhaust heat dissipation from adjacent cold storage equipment.

[0074] The central intelligent layer concatenates the joint representation vector and the context vector, and generates structured cue word features after linear mapping by the feature projection layer. The central intelligent layer injects these structured cue word features into the input layer of the large language model. The large language model performs forward propagation computation, outputting the probability distribution values ​​of the hidden layer feature vectors within a predefined set of latent risk categories. This set of latent risk categories includes the risk of smoldering fires in goods. The central intelligent layer extracts the inferred probability values ​​corresponding to the risk of smoldering fires in goods from the set of latent risk categories.

[0075] When the inferred probability value exceeds the corresponding set threshold in the risk triggering threshold matrix, the central intelligent layer determines that the target entity node poses a risk of smoldering fire. The aforementioned calculation process, through tensor fusion of cross-modal data and large language model inference, calculates and outputs the implicit risks that have not reached the single device alarm threshold from multi-source heterogeneous data such as cargo accumulation in physical space, temperature fluctuations of IoT devices, and sales stagnation in business systems, generating corresponding risk inference result data.

[0076] After the structured risk simulation results are transmitted from the central intelligence layer to the application service layer and the value-added service layer, the application service layer performs dynamic strategy generation operations based on the risk simulation results data.

[0077] The application service layer receives the risk simulation results data. It then parses the data, extracting the globally unique identifier of the target entity node, the risk category identifier index, and the corresponding simulation probability value.

[0078] The application service layer configures a preset response strategy database in the system storage module. This database stores the mapping between risk category identifier indexes and basic handling logic sequences. Based on the extracted risk category identifier indexes, the application service layer queries and retrieves the corresponding basic handling logic sequences from the preset response strategy database. These basic handling logic sequences contain preset variable placeholders.

[0079] The application service layer invokes the cross-modal spatiotemporal knowledge graph data structure of the target entity node at the current discrete time step through the central intelligence layer. The application service layer extracts the spatial coordinate boundaries and business modal features of the target entity node. Based on the spatial adjacency relationships in the topological edge set, the application service layer extracts first-order neighbor nodes associated with the target entity node, categorized as device nodes. The application service layer extracts the IoT modal features of the first-order neighbor nodes. The application service layer maps and replaces the spatial coordinate boundaries, business modal features, and IoT modal features of the first-order neighbor nodes into the variable placeholders corresponding to the basic processing logic sequence, generating a dynamic response strategy for the target entity node.

[0080] The dynamic response strategy includes on-site intervention commands and equipment adjustment parameters. The application service layer parses the on-site intervention commands and generates a security inspection work order containing the spatial coordinate boundaries of the target entity node. Based on the spatial coordinate boundaries of the target entity node, the application service layer matches the corresponding grid management area in the system spatial database. The application service layer then distributes the security inspection work order to the mobile terminal bound to the grid management area.

[0081] The application service layer synchronously parses the device adjustment parameters and generates control signaling for the aforementioned first-order neighbor nodes. The application service layer then sends the control signaling to the corresponding physical device control nodes to execute the corresponding physical state adjustment operations. This process realizes a closed-loop control mechanism from implicit risk deduction to physical space intervention.

[0082] After the application service layer sends the security inspection work order to the mobile terminal bound to the grid management area and sends the control signal to the corresponding physical device control node, the central intelligent layer, in collaboration with the application service layer, further executes the intelligent closed-loop verification and status synchronization operation on the operation and maintenance side.

[0083] The mobile terminal receives a security inspection work order. Maintenance personnel execute on-site handling operations based on the on-site intervention instructions and spatial coordinate boundaries in the security inspection work order. The mobile terminal collects on-site images and handling confirmation text after the on-site handling operation is completed. The mobile terminal transmits the on-site images and handling confirmation text as on-site execution feedback data to the application service layer. Simultaneously, the physical device control node receives control signals and executes physical state adjustment operations. The physical device control node extracts the device operating status parameters after execution. The physical device control node transmits the device operating status parameters as device execution feedback data to the application service layer. The application service layer sends the on-site execution feedback data and device execution feedback data back to the central intelligence layer.

[0084] The central intelligent layer receives field execution feedback data and device execution feedback data, and generates state synchronization triggering signals. The edge perception layer continuously performs multi-source heterogeneous data acquisition operations within the physical space. Based on the state synchronization triggering signals, the central intelligent layer extracts continuous aligned data frames after the completion of field handling operations. The central intelligent layer parses the continuous aligned data frames and extracts the updated perception data values. The central intelligent layer matches the target entity node with its corresponding first-order neighbor node in the virtual state matrix. The central intelligent layer uses the updated perception data values ​​to replace the historical values ​​of the corresponding attribute vectors of the target entity node and its first-order neighbor node.

[0085] Based on the numerically replaced virtual state matrix, the central intelligent layer updates the visual modal features, IoT modal features, and business modal features of the target entity node and its first-order neighbor nodes in the cross-modal spatiotemporal knowledge graph network structure. Based on the updated cross-modal spatiotemporal knowledge graph network structure, the central intelligent layer re-extracts the joint representation vector and context vector of the target entity node. The central intelligent layer concatenates the re-extracted joint representation vector and context vector into a comprehensive state representation vector, and then maps it to structured cue word features using a feature projection layer.

[0086] The central intelligent layer injects structured prompt word features into the input layer of the large language model, and then re-invokes the large language model to perform joint probability inference calculations for implicit risks. The central intelligent layer obtains the recalculated inference probability values. The central intelligent layer compares the recalculated inference probability values ​​with the set thresholds matched in the risk trigger threshold matrix.

[0087] When the recalculated probability value is less than or equal to a set matching threshold, the central intelligent layer determines that the latent risk corresponding to the target entity node has been eliminated. The central intelligent layer generates a risk elimination confirmation signal and sends it to the application service layer. Upon receiving the risk elimination confirmation signal, the application service layer updates the business status parameters of the security inspection work order to the completed status. The central intelligent layer extracts the structured prompt word features generated before this risk simulation calculation as model input sample data, and extracts on-site execution feedback data and equipment execution feedback data as real label data to verify whether the risk actually exists. The central intelligent layer combines the model input sample data and the real label data into historical risk event records for archiving and storage, providing them to the large language model as training samples for subsequent updates and fine-tuning of network parameters. The aforementioned verification, feature reconstruction, and parameter archiving processes realize an intelligent closed loop from physical space state intervention to digital space model correction.

[0088] After the central intelligent layer and the application service layer complete the intelligent closed-loop verification and status synchronization operations on the operation and maintenance side, the value-added service layer executes the intelligent closed-loop operations on the value-added side based on the archived historical risk event records.

[0089] The value-added service layer extracts the globally unique identifier of the target entity node. Based on the globally unique identifier, the value-added service layer retrieves the historical risk event record corresponding to the target entity node from the system storage module. The historical risk event record contains the projected probability value of each occurrence of implicit risks of the target entity node, the risk category identifier index, and the on-site execution feedback data returned by the operation and maintenance side.

[0090] The value-added service layer analyzes the on-site execution feedback data. It extracts the issuance and completion timestamps of security inspection work orders recorded in the on-site execution feedback data. The value-added service layer calculates the time difference between the issuance and completion timestamps to obtain the risk response timeliness value. Based on the risk category identifier index in historical risk event records, the value-added service layer calculates the risk trigger frequency value of the target entity node within a preset time period.

[0091] The value-added service layer configures a credit rating quantification model within the system. This model incorporates a first preset weighting coefficient and a second preset weighting coefficient. The value-added service layer inputs the risk response timeliness value and the risk trigger frequency value as features into the credit rating quantification model. The model multiplies the risk response timeliness value by the first preset weighting coefficient and the risk trigger frequency value by the second preset weighting coefficient. Finally, the model sums these two products and outputs the dynamic credit rating score for the target entity node.

[0092] The value-added services layer matches corresponding supply chain finance adjustment strategies and insurance premium adjustment strategies from the value-added strategy database based on the dynamic credit rating score. The supply chain finance adjustment strategy includes credit limit adjustment parameters. The insurance premium adjustment strategy includes the basic premium multiplier.

[0093] The value-added service layer transmits supply chain finance adjustment strategies and insurance premium rate adjustment strategies to the servers of externally authorized financial institutions via application programming interfaces (APIs). The financial institution servers update the financial service agreements of the target entity nodes according to these strategies. The value-added service layer receives the agreement update confirmation data from the financial institution servers. It then extracts the update data as updated business status parameters and transmits these values ​​to the central intelligence layer. The central intelligence layer uses these updated business status parameters to replace the historical values ​​of the corresponding business modality features of the target entity nodes in the cross-modal spatiotemporal knowledge graph. This process achieves an intelligent closed loop from monitoring entity operational risks to quantifying and adjusting commercial value.

[0094] Based on the aforementioned intelligent closed loop from monitoring operational risks to quantifying and adjusting commercial value, this embodiment provides an electronic device. The electronic device includes a processor, a memory, a communication interface, and a communication bus.

[0095] The processor, memory, and communication interface are electrically connected and exchange data with each other through a communication bus. The communication bus provides a data transmission channel between various hardware components within the electronic device.

[0096] The communication interface is used to establish a data transmission link between electronic devices and external hardware nodes. Electronic devices receive multi-source heterogeneous data and updated sensing data values ​​transmitted from the edge sensing layer through the communication interface, and also transmit security inspection work orders and control signals to external mobile terminals and physical device control nodes through the communication interface.

[0097] The memory is used to store computer-readable instructions and business data generated during system operation. System business data includes the aforementioned virtual state matrix, cross-modal spatiotemporal knowledge graph network structure, large language model pre-training parameters, fine-tuned network parameters, and historical risk event records. The memory includes random access memory and non-volatile storage media. Non-volatile storage media include read-only memory, flash memory, or hard disk.

[0098] The processor reads and executes computer-readable instructions stored in memory. When executing these instructions, the processor drives the electronic device to run the logic code for the aforementioned cross-modal spatiotemporal knowledge graph construction, heterogeneous modal feature vector extraction, tensor fusion network derivation calculation, large language model joint probability inference calculation, dynamic strategy generation, intelligent closed-loop verification and state synchronization operations on the operation and maintenance side, and intelligent closed-loop operations on the value-added side. By executing this logic code, the processor achieves implicit risk inference and physical space intervention for the target entity nodes.

[0099] This embodiment provides a computer-readable storage medium. The computer-readable storage medium stores computer-executable instructions. These computer-executable instructions are read and executed by a processor in the aforementioned electronic device.

[0100] When executing computer execution instructions, the processor implements the various computational logic and control operations performed by the central intelligence layer, application service layer, and value-added service layer in the aforementioned embodiments. The processor executes computer execution instructions to implement the various execution steps of the implicit risk deduction and physical space intervention method provided in the aforementioned embodiments.

[0101] The processor executes computer instructions to sequentially achieve cross-modal spatiotemporal knowledge graph construction, heterogeneous modal feature vector extraction, tensor fusion network derivation calculation, large language model joint probability inference calculation, dynamic strategy generation, intelligent closed-loop verification and state synchronization operation on the operation and maintenance side, and intelligent closed-loop operation on the value-added side.

[0102] Computer-readable storage media are non-transitory computer-readable storage media. Specific physical forms of non-transitory computer-readable storage media include read-only memory, random access memory, erasable programmable read-only memory, flash memory, disk storage, and optical disk storage media. Electronic devices retrieve corresponding computer execution instructions from non-transitory computer-readable storage media for computation and processing via an internal communication bus.

[0103] 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 smart digital twin system for wholesale markets that integrates sensor data, characterized in that, It includes the edge perception layer, the central intelligence layer, and the application service layer; The edge perception layer collects multi-source heterogeneous data in the physical space and uses the multi-source heterogeneous data to generate a virtual state matrix containing visual modal features, IoT modal features and business modal features; The central intelligent layer constructs a cross-modal spatiotemporal knowledge graph network structure based on the virtual state matrix. The cross-modal spatiotemporal knowledge graph network structure includes a set of entity nodes configured with node category identifiers and a set of topological edges configured with relation type labels. The central intelligent layer extracts the visual modal features, IoT modal features, and business modal features of the target entity nodes in the entity node set, and performs tensor fusion calculation on the visual modal features, IoT modal features, and business modal features to output the joint representation vector of the target entity nodes; The central intelligent layer extracts the neighbor nodes of the target entity node in the cross-modal spatiotemporal knowledge graph network structure and performs feature aggregation calculation, outputting a contextual vector; The central intelligent layer concatenates the joint representation vector and the context vector to generate a comprehensive state representation vector, and maps the comprehensive state representation vector to structured cue word features; The central intelligent layer inputs the structured prompt word features into the large language model to perform implicit risk joint probability inference calculation and outputs risk inference result data; The application service layer receives the risk simulation results data and generates a dynamic response strategy for the target entity node based on the risk category identifier index in the risk simulation results data.

2. The intelligent digital twin system for wholesale markets that integrates sensing data according to claim 1, characterized in that, The central intelligent layer calculates the absolute distance between different entity nodes in the virtual state matrix. When the absolute distance is less than a set adjacent distance threshold, it determines that there is a spatial adjacency relationship between the entity nodes. When the coordinate boundary of one entity node is completely inside the coordinate boundary of another entity node, it determines that there is a spatial inclusion relationship between the entity nodes. The central intelligent layer compares the business association identifier sequences of different entity nodes. When the business association identifier sequences contain the same goods identifier code, it determines that there is a goods ownership relationship between the entity nodes. When the business association identifier sequences contain the same transaction association code, it determines that there is a transaction association relationship between the entity nodes.

3. The intelligent digital twin system for wholesale markets that integrates sensing data according to claim 1, characterized in that, The central intelligent layer converts visual modal features, IoT modal features, and business modal features into floating-point arrays and appends scalar elements with a value of 1 to the end of the floating-point arrays to generate augmented spatial visual feature vectors, augmented logistics IoT feature vectors, and augmented transaction business feature vectors. The central intelligent layer performs tensor outer product operations on the augmented spatial visual feature vectors, augmented logistics IoT feature vectors, and augmented transaction business feature vectors to generate a multimodal fusion tensor. The central intelligent layer flattens the multimodal fusion tensor into a one-dimensional multimodal feature vector and performs linear projection operations on the multimodal feature vectors and weight matrices to obtain a linear mapping vector. The central intelligent layer adds the linear mapping vector to the bias vector and inputs it into a nonlinear activation function for numerical mapping processing, outputting a joint representation vector.

4. The intelligent digital twin system for wholesale markets that integrates sensing data according to claim 1, characterized in that, The central intelligent layer uses graph neural network aggregation operators to assign corresponding neighbor weights to the first-order and second-order neighbor nodes of the target entity node based on the relation type label. The central intelligent layer performs weighted summation calculation on the visual modal features, IoT modal features, and business modal features of the first-order and second-order neighbor nodes according to the neighbor weights, obtains the feature aggregation result, and outputs it as a contextual vector. The central intelligent layer uses a feature projection layer to perform linear mapping on the comprehensive state representation vector, and maps the numerical value of the comprehensive state representation vector into a word embedding vector format to generate structured prompt word features.

5. The intelligent digital twin system for wholesale markets that integrates sensing data according to claim 1, characterized in that, The central intelligent layer injects structured prompt word features into the input layer of the large language model, calls the multi-head self-attention mechanism module and feedforward neural network module inside the large language model to perform forward propagation calculation, and outputs hidden layer feature vectors. The central intelligent layer inputs the hidden layer feature vectors into the normalized exponential function operator to calculate the probability distribution values ​​of each risk category identifier index in the preset set of hidden risk categories. The central intelligent layer matches the set threshold in the risk trigger threshold matrix according to the risk category identifier index. When the inferred probability value is greater than the corresponding set threshold, it determines that the target entity node has hidden risk and generates risk inference result data.

6. The intelligent digital twin system for wholesale markets that integrates sensing data according to claim 1, characterized in that, The application service layer extracts the basic handling logic sequence containing variable placeholders from the preset response strategy database based on the risk category identifier index in the risk simulation result data; the application service layer extracts the spatial coordinate boundary, business modal characteristics, and IoT modal characteristics of the first-order neighbor nodes whose node category is device node from the spatial coordinate boundary, business modal characteristics, and IoT modal characteristics of the first-order neighbor nodes; the application service layer maps and replaces the spatial coordinate boundary, business modal characteristics, and IoT modal characteristics of the first-order neighbor nodes into the variable placeholders corresponding to the basic handling logic sequence, generating a dynamic response strategy containing security inspection work orders and control signaling.

7. The intelligent digital twin system for wholesale markets that integrates sensing data according to claim 1, characterized in that, The central intelligent layer receives field execution feedback data and equipment execution feedback data to generate a status synchronization trigger signal; the central intelligent layer extracts continuous aligned data frames and obtains the perceived data update value based on the status synchronization trigger signal; the central intelligent layer uses the perceived data update value to replace the historical values ​​of the target entity node and its corresponding first-order neighbor node, updates the cross-modal spatiotemporal knowledge graph network structure, and re-calls the large language model to perform joint probability inference calculation of implicit risks; the central intelligent layer generates a risk elimination confirmation signal when the recalculated inference probability value is less than or equal to the set matching threshold.

8. The intelligent digital twin system for wholesale markets that integrates sensing data according to claim 1, characterized in that, It also includes a value-added services layer; The value-added service layer extracts the issuance and completion timestamps of security inspection work orders from the on-site execution feedback data to calculate the risk response timeliness value, and counts the risk trigger frequency value of the target entity node within the preset time period; The value-added service layer uses a credit rating quantification model to perform a linear weighted calculation based on the first preset weight coefficient and the second preset weight coefficient on the risk response time value and the risk trigger frequency value, and outputs a dynamic credit rating score. The value-added service layer matches supply chain finance adjustment strategies, which include credit limit adjustment parameters, and insurance premium rate adjustment strategies, which include basic premium multipliers, in the value-added strategy database based on dynamic credit rating scores.

9. The intelligent digital twin system for wholesale markets that integrates sensing data according to claim 1, characterized in that, The central intelligent layer extracts the timestamp of the current discrete time step as a time dimension parameter; the central intelligent layer appends the time dimension parameter to the data records of the entity node set and the topology edge set for temporal marking; when the business status parameter changes from "in a transaction" to "transaction completed", the central intelligent layer deletes the directed edges of the corresponding transaction association, and combines the temporally marked entity node set and the dynamically updated topology edge set to generate the cross-modal spatiotemporal knowledge graph network structure of the corresponding discrete time step.

10. The intelligent digital twin system for wholesale markets that integrates sensing data according to claim 1, characterized in that, Visual modal features include space occupancy area and stacking height; IoT modal features include RFID dwell time calculated based on timestamps and temperature deviation calculated based on preset temperature standards; business modal features include sales rate deviation calculated based on order creation time and inventory changes, as well as order turnover frequency per unit time.

Images