Cold chain risk intelligent management and control system based on ANN-GERT hybrid modeling

Abstract

The application discloses a cold chain risk intelligent management and control system based on ANN-GERT hybrid modeling, and relates to the field of cold chain logistics data processing.The system comprises a ternary extraction module, which maps effective records into flow edges, compares flow program bits to determine rollback, extracts return triggers, receiving and re-entry sections, and assembles the sections into a return ternary chain set; a parameter generation module, which splices the ternary chain into an input unit, maps output quality and cost and other multi-dimensional risk parameters through a pre-trained neural network; a super-edge construction module, which positions network boundaries, instantiates return super-edges, constructs a transfer function in combination with parameters, and outputs an enhanced risk network; and an attribution management and control module, which analyzes the transfer function in combination with a role vector to generate node contribution values, extracts core nodes, and generates instructions to trigger hardware linkage locking.The system solves the problem that one-way tracking cannot analyze abnormal backtracking nonlinear losses, and realizes accurate attribution and targeted physical management and control.

Description

Technical Field

[0001] This invention relates to the field of cold chain logistics data processing, specifically to a cold chain risk intelligent management and control system based on ANN-GERT hybrid modeling. Background Technology

[0002] Food cold chain logistics often involves the collaborative operation of multiple legal entities, including warehouse management systems and transportation scheduling systems. Throughout the complete fulfillment lifecycle, controlled food products undergo multiple stages, including receiving and inspection, refrigerated storage, and sorting and loading. The underlying hardware and software environments deployed at each business end are heterogeneous, and the system log chains accumulated during cross-entity handover and fulfillment for the same batch of goods often exhibit multiple sources and discontinuous flow. In actual operations, when abnormal situations such as exceeding environmental limits or failing handover inspection occur, the conventional unidirectional, monotonically increasing business sequence is disrupted. The target logistics batch will be pulled back to the pre-emptive quality control state, including re-inspection, anomaly isolation, and repackaging. After the pre-emptive anomaly handling is completed, the related business nodes are reconnected to the normal forward fulfillment chain through system truncation operations.

[0003] Existing control technologies are limited by unidirectional tracking logic, focusing on environmental threshold alarms and forward trajectory recording. When a business flow reversal occurs where the endpoint flow program bit is less than the starting flow program bit, the forward tracking link breaks directly into an isolated blind spot, and the control system cannot continuously track complex trajectories containing abnormal feedback loops. With the generation of abnormal return paths, the accumulation of potential food spoilage penalty losses and sunk costs from repeated operations breaks the linear transmission law, exhibiting a highly complex nonlinear dynamic superposition state. Due to the lack of deep graph theory analysis and topology folding reconstruction capabilities for multi-source discontinuous logs in the underlying computing architecture, existing platforms cannot map multi-stage related business risks in a multi-dimensional hidden space, nor can they construct frequency domain transfer operators to quantify the superposition effects of superimposed edge losses. The lack of graph computation simplification capabilities leads to systemic risk attribution distortion and miscalculation of single node contribution. The control backend cannot accurately pinpoint the core links with the most profound impact on the overall risk spread within a massive cluster of concurrent nodes. Consequently, it cannot automatically assemble and verify data payloads and issue instructions, ultimately preventing the targeted triggering of interlocking mechanisms such as unauthorized access control locks in automated warehouses or abnormal freezing of electronic access control systems in cold storage. For discrete data chains containing anomaly backtracking loops, how to extract multi-entity structured correlation paths and construct an algorithmic foundation with nonlinear topological transmission analysis capabilities to achieve accurate attribution of tracing nodes and cut off the systemic spread of quality hazards has become a core technical problem that urgently needs to be solved. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a cold chain risk intelligent management and control system based on ANN-GERT hybrid modeling. This system solves the problem that the unidirectional tracking logic cannot analyze nonlinear transmission losses in abnormal backtracking scenarios, resulting in distorted risk attribution and the inability to target and trigger linkage interlocking mechanisms.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a cold chain risk intelligent management and control system based on ANN-GERT hybrid modeling, comprising: The ternary extraction module is used to map the valid business records stripped from the multi-entity cold chain handover data set to standard cold chain business status nodes to generate actual flow edges. When comparing the endpoint flow program bits to determine the backtracking, it extracts the set of return trigger segments, return acceptance segments, and re-entry segments, and assembles them into a global return ternary chain set. The parameter generation module is used to concatenate the global back-back ternary chain set into a back-back memory input unit, and output the parameters for continued performance success, quality risk bearing parameters, cost bearing parameters and role contribution allocation vector after being pre-trained by the artificial neural network model. The superedge construction module is used to locate the topological boundary of the global back-back ternary chain set in the basic forward cold chain process network to instantiate the back-back superedge. It constructs the back-back superedge risk cost transfer function by combining the quality risk bearing parameters, cost bearing parameters and continued performance success parameters, and outputs the back-back enhanced risk network. The attribution control module is used to define the roles of nodes in the enhanced risk network, analyze the return super-edge risk cost transfer function to extract the return contribution benchmark value, combine the role contribution allocation vector to generate the contribution value of a single return node, and obtain the total return node contribution value through iterative accumulation to extract the core node, generate intelligent control instructions to trigger the hardware linkage interlocking mechanism.

[0006] Compared with existing technologies, it has the following advantages: This proposed intelligent cold chain risk management system, based on ANN-GERT hybrid modeling, effectively solves the problem of traditional management technologies being unable to accurately analyze chain breaks and transmission losses when facing business flow regressions. For the reverse backtracking loop composed of multi-source discontinuous logs, this system compares endpoint flow program bits to structurally assemble scattered flow breakpoints into a global backtracking ternary chain set. Then, a pre-trained artificial neural network model is introduced to perform temporal average pooling and multi-branch hidden layer mapping on the multi-dimensional state tensor, outputting risk and cost parameters with realistic dimensions. The system instantiates equivalent backtracking hyperedges in the basic forward cold chain process network and constructs a risk cost transfer function by combining moment generating function auxiliary variables and exponential mapping operators. This processing mechanism overcomes the limitation of existing technologies that can only perform unidirectional tracing based on directed acyclic graphs, successfully transforming the dynamic superposition process of food spoilage penalty losses and repetitive sunk costs into a precisely calculable frequency domain transfer operator, establishing a nonlinear quantitative calculation benchmark for complex cross-entity handover trajectories in the feedback topology space.

[0007] Building upon the completion of complex topology quantitative modeling, this system further eliminates the technical defects of systemic risk attribution distortion and node contribution assessment misalignment in large-scale concurrent networks. The management backend extracts the return contribution benchmark value by performing first-order derivative operations on the risk cost transfer function, and constructs the node evidence carrying ratio by combining the total number of valid business records at the underlying level. It then extracts the single-return node contribution value using multi-dimensional parameter multiplication operations. Combined with heap sorting algorithms and anti-collision comparison mechanisms, the platform can accurately pinpoint the core links with the most profound impact on overall risk diffusion from massive concurrent data. Based on defined node role labels and business action type fields, the scheduling engine automatically assembles and verifies data payloads and issues intelligent management instructions, targeting and triggering hardware-linked locking actions such as automated warehouse interception locks or cold storage access control abnormal freezing. This entire mechanism establishes a technical closed loop from underlying multi-source log tracing and analysis to targeted intervention by terminal hardware, cutting off the systemic spread path of food quality hazards in the supply chain network at its source. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the system framework of the present invention.

[0009] Figure 2 This is a schematic diagram of the system execution flow of the present invention. Detailed Implementation

[0010] 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.

[0011] Please see Figures 1 to 2 This application provides a cold chain risk intelligent management and control system based on ANN-GERT hybrid modeling, including a ternary extraction module, a parameter generation module, a hyperedge construction module, and an attribution management and control module; The ternary extraction module is used to acquire the multi-entity cold chain handover data set for the target cold chain batch. In the distributed IoT architecture, business terminals such as the warehouse management system and transportation scheduling system, which operate across legal entities, push logs to the backend server in real time through a distributed message queue. The system extracts the target cold chain batch identifier from each log, performs data aggregation and cleaning on the multi-entity cold chain handover data set, and extracts the valid business records belonging to that batch. For duplicate records caused by network retransmission within the same time window for the same business entity, the system uses evidence record identifiers for hash verification and deduplication. Valid business records must have underlying architectural fields such as batch identifier, business action type, and business entity source. If key traceability primary keys such as batch identifier or action type are missing, the system marks the associated record as an isolated breakpoint and isolates it, prohibiting it from participating in topology construction; if non-critical environmental features such as node temperature or equipment status are missing, the system extracts the median value of historical similar samples to fill the corresponding fields to ensure the continuity of the data link.

[0012] After data cleaning, the system calls the standard cold chain business status node table pre-deployed in the database to map valid business records to corresponding standard cold chain business status nodes. The standard cold chain business status node table serves as a global reference dictionary, containing built-in codes for entity operation nodes such as receiving and inspection, refrigerated storage, and sorting and loading. When a valid business record contains a complete node code, the system directly performs precise addressing mapping. If the underlying record lacks a node code due to system heterogeneity, the system automatically triggers an action type auxiliary mapping mechanism. This mechanism extracts the business action type field from the valid business record and iterates through and matches preset action correlation weights within the global dictionary. These action correlation weights are based on a preset business rule dictionary configuration, which persistently stores the strong binding mapping probability values ​​between each standard node code and the valid business action type. The system selects the node with the highest correlation weight as the target mapping point, eliminating data mapping gaps in multi-source heterogeneous environments.

[0013] Subsequently, based on the upstream and downstream handover documents and business status flow relationships of the target cold chain batch, the system constructs a directed graph topology represented by an adjacency matrix in memory, generating the actual flow edge set for that batch. The actual flow edge set contains multiple directed flow records indicating the actual logistics movement trajectory. The system traverses each actual flow edge in the actual flow edge set and queries the flow program bits corresponding to the starting and ending nodes of the associated actual flow edge. The system's underlying computing unit extracts the ending and starting flow program bits of the actual flow edge (here, the ending and starting flow program bits are collectively referred to as endpoint flow program bits) and executes a numerical comparison instruction. The flow program bits are monotonically increasing values ​​pre-hardcoded in the node dictionary, used to characterize the business progress of a node in the standard forward fulfillment process. When the comparison result determines that the ending flow program bit is strictly less than the starting flow program bit, the system determines that a business flow reversal has occurred, immediately cuts off the current forward tracking link, and instantiates the actual flow edge that triggered the reverse state transition as a return trigger segment. The process of extracting the return trigger segment completely eliminates the dependence on local timestamps, avoiding misjudgments of the flow direction caused by asynchronous clocks on various business servers.

[0014] After locking the return trigger segment in memory, the system initiates sequence tracing along the outward direction from the start of the return trigger segment. The system continuously reads the consecutive flow records that immediately follow and whose node type attributes belong to the preceding quality control state. The preceding quality control state includes an enumerated set of types consisting of re-inspection, anomaly isolation, and disk reassembly. The system stores the captured consecutive flow records into a dynamic array and combines them to generate a set of return connection segments. During tracing, the system verifies the endpoint node type of the consecutive flow records in real time. Once it detects that the type identifier of the target node exceeds the anomaly control area and the pointer to the flow program bit is greater than the preceding position of the return trigger segment, the system triggers a forced boundary truncation command. Outgoing flow edges that cross the quality control category are extracted by the system and confirmed as re-entry segments.

[0015] Finally, the system identifies and extracts the set of return trigger segments, return acceptance segments, and re-entry segments, and then structurally assembles them according to the topological flow association sequence, instantiating them into a return ternary chain in the memory stack. The return ternary chain is structurally represented as a structured business topology unit containing three dimensions: abnormal triggering, abnormal acceptance, and fulfillment re-entry. If multiple independent flow program bit reversal events occur within the global lifecycle of the target cold chain batch, the system iteratively generates multiple independent return ternary chains based on traversal matching rules, encapsulates them into a global return ternary chain set for that batch, and uses this set as a high-dimensional standard input tensor to fill subsequent artificial neural networks and risk simulation models.

[0016] The parameter generation module receives the global return ternary chain set output by the ternary extraction module and generates corresponding return risk parameters based on each return ternary chain in the global return ternary chain set. The system traverses each return ternary chain in the global return ternary chain set and calls the associated underlying business records. The system extracts return trigger features from the business records corresponding to the return trigger segment, return acceptance features from the business records corresponding to the return acceptance segment set, and re-entry features from the business records corresponding to the re-entry segment. The return trigger features cover the node code before return triggering, the node code after return triggering, and the return trigger reason field; the return acceptance features cover the return acceptance action type, re-inspection result, and acceptance segment cost record field; the re-entry features cover the action type after re-entry and the goods status field after re-entry. The above-extracted return trigger features, return acceptance features, and re-entry features are collectively referred to as underlying business features.

[0017] After extracting the three types of features mentioned above, the system performs feature cleaning and tensor alignment operations at the data preprocessing layer. For categorical fields such as node codes and action types in the extracted features, the system calls a preset entity embedding layer to map them into computable dense vectors. For numerical fields such as cost records and product status indicators, the system performs maximum and minimum value normalization processing according to the data distribution range of the same product category. Considering that within a specific historical sampling time window, the baseline operating cost of a certain type of standardized cold chain product may be in a constant billing state, or the continuous monitoring indicator readings of controlled products in an extremely steady-state environment may be absolutely consistent, which objectively leads to the maximum and minimum values ​​of the historical data of the extracted field being exactly equal, making the feature range in the normalization formula zero. To avoid the underlying floating-point operation overflow caused by the above-mentioned extreme steady-state data distribution, the system introduces a very small constant term as a smoothing factor in the denominator of the normalization division formula, or directly maps the numerical field to a preset default normalized scalar when the data distribution range is zero.

[0018] If a categorical field contains null values, the system uniformly maps them to a preset unknown category baseline vector. If numerical environmental indicators such as product status are missing, the system queries the database, extracts the median value of historical handover samples under the same product category, and fills it into the corresponding field. If cost records are missing, the system queries the pre-configured standard operating cost table within the cold chain management system, extracts the baseline operating cost value corresponding to the same action type, and fills it into the corresponding feature position. For multiple continuous flow records contained in the return acceptance segment set, after extracting the dense vector and normalized value of a single record, the system uses a time-series average pooling operation to aggregate the variable-length record features into a single fixed-length return acceptance feature tensor (i.e., the fixed-length acceptance tensor).

[0019] Based on this, the underlying computational unit executes the feature concatenation formula: in, Indicates the return memory input unit; This represents a fixed-length feature tensor composed of return trigger features (i.e., a fixed-length tensor corresponding to the trigger features). This represents a fixed-length feature tensor (i.e., a fixed-length acceptance tensor) formed by dimensionality reduction of the back-receiving features through temporal average pooling. This represents a fixed-length feature tensor composed of re-entry features (i.e., the fixed-length tensor corresponding to the re-entry features). The back-to-memory input unit is a high-dimensional feature tensor composed of the aforementioned three fixed-length feature tensors, horizontally concatenated, used to characterize the numerical state of the target cold chain batch undergoing the back-to-memory topology loop.

[0020] Subsequently, the back-to-memory input unit is loaded into a pre-trained artificial neural network model deployed on the computing cluster for hidden layer mapping. The pre-trained artificial neural network model employs a three-branch network architecture, including independently operating back-to-the-end branch, back-to-the-end branch, and re-entry branch. The system processes the back-to-the-end feature tensor through the back-to-the-end branch, outputting a trigger encoding vector; processes the back-to-the-end feature tensor through the back-to-the-end branch, outputting an end-to-the-end encoding vector; and processes the re-entry feature tensor through the re-entry branch, outputting a re-entry encoding vector. The system previously constrained the encoding vectors output by each independent branch to an equal-length dimension to ensure the mathematical validity of subsequent element-wise multiplication operations. To capture the business-related coupling relationships between different roles, the system performs hidden layer vector interaction at the feature fusion layer.

[0021] The vector interaction and cascading formulas are as follows: in, Represents the return fusion vector; Indicates the trigger encoding vector; Indicates the continuation of the encoded vector; This indicates re-entry into the encoding vector; This represents the element-wise multiplication operation of vectors (i.e., the Hadamard product). The back-fusion vector combines independent encoded vectors with product interaction terms to characterize the deep feature space relationships of risk propagation and superposition at different stages of the flow.

[0022] The system will return the fused vector. The multi-task output layer of the pre-trained artificial neural network model is fed into a fully connected layer and processed by differentiated activation functions configured for different output branches to output specific backflow risk parameters. The multi-task output layer has four independent output branches: the continuation of successful performance parameter output branch uses a Sigmund activation function to constrain the numerical range; the quality risk carrying parameter output branch uses a linear rectified activation function or a smoothed linear rectified activation function to implement non-negative truncation; the role contribution allocation vector output branch uses a normalized exponential activation function to achieve vector element normalization and non-negative distribution; and the cost carrying parameter output branch uses a linear identity mapping. The system uses pre-stored cost extremum parameters to perform inverse normalization mapping on the relative cost values ​​output by the network, restoring the specific cost unit carried by the relative cost value.

[0023] The network prediction calculation formula can be expressed as an end-to-end mapping relationship as follows: in, Indicates that the network weight parameters are Artificial neural network models; The parameter representing the continued successful performance is a dimensionless probability factor whose value ranges from zero to one after being constrained by the Sigmund activation function. The quality risk bearing parameter is a dimensionless scalar whose value is not less than zero after being constrained by a linear rectified activation function or a smoothed linear rectified activation function. This represents the cost-bearing parameter, whose value is the actual consumption value of a specific cost unit after being output by linear identity mapping and subjected to inverse normalization mapping, carrying the configuration of the management system. This represents the role contribution allocation vector, which is a probability distribution vector that covers the weight allocation ratio of the triggering role, the receiving role, and the re-entry role after being constrained by the normalized exponential activation function, and the sum of the values ​​of each element in the vector is always one.

[0024] The network weight parameters of the pre-trained artificial neural network model are obtained through backpropagation training using historical cold chain batch data. During the offline training phase, the system extracts historical training samples from the data warehouse and constructs multi-dimensional training labels. (1) Construction of performance tag: Construct a continued performance success tag based on the end performance results after the historical batches have completed the return processing; (2) Risk label construction: Based on the quality grade table, the historical handover and acceptance results are normalized and mapped to construct quality risk labels; (3) Cost label construction: Extract the extreme value parameters of cost from historical training samples, perform maximum and minimum value normalization on the accumulated actual disposal costs of historical batches, and construct return cost labels; (4) Contribution label construction: The system introduces a Laplace smoothing term to statistically analyze the smoothing ratio of the processing costs of the return trigger segment, return acceptance segment and re-entry segment in the total abnormal cost, and constructs role contribution labels to prevent the computer from being divided into zero due to zero cost of the trigger segment.

[0025] Next, the system constructs a multi-task joint loss function. For the parameters of continued successful performance and role contribution allocation vectors, the system calculates the cross-entropy loss between the network's predicted output value and the corresponding real business training label. For the parameters of quality risk bearing capacity and cost bearing capacity, the system calculates the mean squared error loss between the network's predicted output value and the corresponding real business training label. The system minimizes the joint loss using a gradient descent optimization algorithm, updates the network weight parameters, and outputs a converged artificial neural network model for online use.

[0026] The hyperedge construction module receives the return risk parameters output by the parameter generation module and obtains the basic forward cold chain process network for the target cold chain batch. The basic forward cold chain process network is a directed acyclic graph model pre-stored in a graph database, consisting of a set of business state nodes and a set of actual flow edges. The system locates the topological boundaries of each return ternary chain within the global return ternary chain set in the basic forward cold chain process network. The system extracts the starting node of the return trigger segment in the return ternary chain and identifies it as the starting node of the return hyperedge; it also extracts the ending node of the re-entry segment in the return ternary chain and identifies it as the ending node of the return hyperedge.

[0027] After determining the start and end nodes, the system performs a topology folding operation in the in-memory graph structure. The system extracts the node and edge attribute tables associated with the in-memory graph structure and sets a global traversal visibility field in these tables. The visibility fields of the return trigger segment, the return receiving segment set, and the underlying subgraph structure object corresponding to the re-entry segment are configured to skip traversal. Configuring the skip traversal state ensures that no redundant path calculations occur during the calculation of the global network equivalent transfer value, while preserving the complete topological relationship of the underlying subgraph structure within a local scope for subsequent evidence tracing. The processor instantiates a topological connection edge between the start and end nodes, confirms this topological connection edge as an equivalent return hyperedge, and persists it to the graph database.

[0028] The formula for constructing the back hyperedge is as follows: in, Indicates returning the superedge; Indicates the starting node of the returned superedge; This represents the endpoint node of the return hyperedge. A return hyperedge is a computationally calculable edge in graph theory that spans the exception handling loop and directly connects the previous performance interruption point with the subsequent performance recovery point. The system directly attaches the unique identifier set of the original return ternary chain in key-value pairs to the return hyperedge attribute dictionary in the graph database, constructing a mapping association for underlying data penetration and preventing the loss of underlying real flow records and handover documents due to topology folding operations.

[0029] Furthermore, the system calls the quality risk bearing parameters and cost bearing parameters output by the parameter generation module to calculate the hyperedge risk adjustment cost of the returned hyperedge.

[0030] The formula for calculating the adjusted cost of supermarginal risk is as follows: in, Indicates the cost of adjusting for super-marginal risk; This indicates the cost-bearing parameter, whose value carries the specific cost unit configured in the management system. The quality risk bearing parameter represents the actual sunk cost. The constant term represents the actual sunk cost. The quality risk bearing parameter, as a punitive amplification factor, is superimposed on the sunk cost to quantify the additional loss risk to the overall batch value due to potential food spoilage. The super-marginal risk adjustment cost is a comprehensive scalar that integrates direct economic consumption and potential quality penalties. Through basic arithmetic operations involving addition and multiplication, the dimensionless quality risk weights are legally converted into a cost-dimensional calculation benchmark value.

[0031] The system's underlying computing unit further incorporates the parameters for continued successful performance to construct a return-over-edge risk cost transfer function.

[0032] The formula for constructing the transfer function is as follows: in, This represents the transfer function for the risk cost of returning to the superedge; The parameter representing the success of continued performance is a dimensionless probability factor characterizing the probability of survival upon return. The superedge risk adjustment cost is represented by s; s represents the moment generating function auxiliary variable, whose dimension is the reciprocal of the cost unit; exp represents the exponential mapping operator with the natural constant as its base. The return superedge risk cost transfer function is a frequency domain mapping operator based on stochastic network graph theory, used to quantify the overall transmission impact of complex return paths containing probabilistic branches and resource consumption on the global process network. The superedge risk adjustment cost has a cost unit, while the moment generating function auxiliary variable has a cost reciprocal unit; their product is a dimensionless pure number, ensuring the mathematical validity of the exponential mapping operator when performing exponentiation.

[0033] Finally, the system writes the generated backflow superedge and its attached backflow superedge risk cost transfer function into the basic forward cold chain process network, and outputs the backflow enhanced risk network.

[0034] The formula for network cascading reconfiguration is as follows: in, This indicates a return to an enhanced risk network; This represents the set of business status nodes experienced by the target cold chain batch; This represents the actual set of edges used in the target cold chain batch. This represents a backflow hyperedge with a transfer function attached. Instantiating the backflow hyperedge transforms the original directed acyclic graph model into a stochastic network model containing feedback loops. The system subsequently uses an equivalent network simplification algorithm to handle the newly added feedback loops, making the backflow-enhanced risk network the global operational foundation for risk attribution of execution nodes and system control and scheduling.

[0035] The attribution control module receives the backtracking enhanced risk network and superedge risk adjustment cost output by the hyperedge construction module, and simultaneously acquires the global backtracking ternary chain set generated by the ternary extraction module and the role contribution allocation vector output by the parameter generation module. For each backtracking ternary chain in the global backtracking ternary chain set, the system uses the topological relationships and hierarchical depth preserved in the backtracking enhanced risk network to lock the subordinate role of each node in the abnormal topology loop. The system traverses the set of business state nodes in the backtracking enhanced risk network and executes node role definition instructions. If a business state node is included in the backtracking trigger segment, the system assigns the business state node a backtracking trigger role; if a business state node is included in the backtracking acceptance segment set, the system assigns the business state node a backtracking acceptance role; if a business state node is included in the re-entry segment, the system assigns the business state node a re-entry role. The node role definition instructions rely on hard matching of the instantiated topology boundaries in memory and do not rely on external manual experience settings.

[0036] Based on the aforementioned node role labels, the underlying computing unit extracts the equivalent-processed backbound hyperedge features in the backbound enhanced risk network. The system performs a first-order derivative operation on the backbound hyperedge risk cost transfer function attached to the backbound hyperedge at the moment generating function auxiliary variable's value being zero, extracting the expected value that integrates the continued fulfillment success parameter and the hyperedge risk adjustment cost, and uses this as the backbound contribution benchmark value. For each business status node belonging to the same role, the system counts the total number of valid business records associated with each business status node in the underlying database. The system executes the node evidence carrying ratio calculation instruction, extracts the number of valid business records contained in the current business status node, and counts the total number of valid business records of all business status nodes belonging to the same role as the current business status node. The system calculates the ratio of the current number of valid business records to the aforementioned total number of valid business records, and confirms the generated ratio as the node evidence carrying ratio. Since a business status node participating in topology folding must have at least one underlying log support, the aforementioned total number of valid business records is strictly greater than zero, eliminating the risk of division-by-zero errors in the underlying computing unit. Based on the node evidence carrying ratio, the system further calls the corresponding role weight parameters in the role contribution allocation vector. The system's underlying computing unit performs a multiplication operation on the role weight parameters, the node evidence carrying ratio, and the return contribution benchmark value to generate a single return node contribution value for the current business state node in the abnormal topology loop. The single return node contribution value integrates the global neural network prediction weights and the local true record density into a comprehensive scalar, used to quantify the actual cost loss and potential quality loss that a single business state node should bear in the abnormal topology loop.

[0037] When handling complex business scenarios with multiple abnormal loops, the target cold chain batch is associated with multiple independent return ternary chains. The system initializes the memory accumulator and traverses the global return ternary chain set for each business status node. The system performs a loop accumulation operation on the single return node contribution value generated by the business status node in each return ternary chain to obtain the total return node contribution value of the business status node. Subsequently, the server memory pool calls the heap sort algorithm to sort the total return node contribution values ​​of all business status nodes in descending order, extracts the top-ranked business status node and marks it as the core node that urgently needs key monitoring. If there are multiple business status nodes with strictly equal total return node contribution values, the system compares the total number of valid business records associated with the conflicting nodes in the global lifecycle, and extracts the business status node with the highest total number of valid business records as the core node that urgently needs key monitoring; if the total number of valid business records is still equal, the system calls the underlying log entry sequence number corresponding to the conflicting node to perform a first-in-first-out comparison, and extracts the business status node with the smallest underlying log entry sequence number as the core node that urgently needs key monitoring, eliminating node sorting collisions in the scheduling engine. Core nodes represent the core business processes that have the most profound impact on final cost spillover and quality degradation after crossing multiple abnormal network branches.

[0038] The system retrieves the node role tags and business action type fields bound to the core nodes requiring key monitoring in the memory stack and triggers the action scheduling engine. If the core node is bound to a return trigger role, the action scheduling engine generates a handover evidence verification instruction; if the core node is bound to a return accepting role, the action scheduling engine calls a preset instruction scheduling strategy library, performs key-value pair matching based on the core node's business action type field, and extracts the corresponding re-inspection or anomaly isolation re-inspection instruction; if the core node is bound to a re-entry role, the action scheduling engine generates subsequent performance verification instructions. The action scheduling engine structurally assembles the relevant parameter fields and fills the data payload of the intelligent control instruction with the target cold chain batch identifier, the core nodes requiring key monitoring, the node role tags bound to the core nodes, the total return node contribution value corresponding to the core nodes, and the generated specific verification action instructions. The system sends the intelligent control instruction to the cold chain management platform or execution terminal through a standard interface protocol. The intelligent control command forcibly switches the target cold chain batch from the regular transportation scheduling queue to the high-priority anomaly handling queue at the system level, and simultaneously triggers the hardware linkage interlocking mechanism of the physical execution terminal (such as activating the unauthorized interception lock of the automated warehouse or the abnormal freezing state of the cold storage electronic access control). By accurately issuing targeted verification tasks, the closed-loop elimination of quality risks accumulated in the complex return network of the cold chain supply chain is achieved.

[0039] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A cold chain risk intelligent management and control system based on ANN-GERT hybrid modeling, characterized in that, include: The ternary extraction module is used to map the valid business records stripped from the multi-entity cold chain handover data set to standard cold chain business status nodes to generate actual flow edges. When comparing the endpoint flow program bits to determine the backtracking, it extracts the set of return trigger segments, return acceptance segments, and re-entry segments, and assembles them into a global return ternary chain set. The process by which the ternary extraction module extracts the set of return trigger segments, return acceptance segments, and re-entry segments is as follows: Compare the endpoint flow program bit with the starting flow program bit of the actual flow edge. If the endpoint flow program bit is less than the starting flow program bit, cut off the forward tracking link and instantiate the corresponding actual flow edge as a return trigger segment. Initiate sequence tracing along the out-degree direction of the starting point of the return trigger segment, intercept continuous flow records whose node type attributes belong to the previous quality control state, and generate a set of return connection segments; Real-time verification of the end node type of continuous flow records. When the target node's type identifier exceeds the previous quality control state and the pointer to the flow program bit is greater than the previous return trigger bit, a forced boundary truncation command is triggered to extract the outgoing flow edge and confirm it as a re-entry segment. The parameter generation module is used to concatenate the global back-back ternary chain set into a back-back memory input unit, and output the parameters for continued performance success, quality risk bearing parameters, cost bearing parameters and role contribution allocation vector after being pre-trained by the artificial neural network model. The superedge construction module is used to locate the topological boundary of the global back-back ternary chain set in the basic forward cold chain process network to instantiate the back-back superedge. It constructs the back-back superedge risk cost transfer function by combining the quality risk bearing parameters, cost bearing parameters and continued performance success parameters, and outputs the back-back enhanced risk network. The process of instantiating and returning the hyperedge by the hyperedge construction module is as follows: Extract the starting node of the return trigger segment of the corresponding ternary chain in the global return ternary chain set, and confirm the starting node of the return trigger segment as the starting point of the return superedge; Extract the re-entry segment termination node of the corresponding ternary chain and confirm the re-entry segment termination node as the return superedge endpoint; Extract the node and actual flow edge attribute table of the basic forward cold chain process network, set the global traversal visibility field in the node and actual flow edge attribute table, and configure the global traversal visibility field of the underlying subgraph structure object corresponding to the ternary chain mapping to skip traversal state. Instantiate a topological connection edge between the start and end points of the return hyperedge, confirm the topological connection edge as a return hyperedge, and establish a bottom-level data penetration mapping by attaching the unique identifier set of the underlying subgraph structure object to the attribute dictionary of the return hyperedge. The attribution control module is used to define the roles of nodes in the enhanced risk network, analyze the return super-edge risk cost transfer function to extract the return contribution benchmark value, combine the role contribution allocation vector to generate the contribution value of a single return node, and obtain the total return node contribution value through iterative accumulation to extract the core node, generate intelligent control instructions to trigger the hardware linkage interlocking mechanism.

2. The intelligent cold chain risk management system based on ANN-GERT hybrid modeling as described in claim 1, characterized in that, The ternary extraction module maps to standard cold chain business status nodes, including: Data aggregation and cleaning are performed on the multi-entity cold chain handover data set to extract valid business records. When the valid business record is missing node encoding, the business action type field contained in the valid business record is extracted. Based on the preset business rule dictionary, the action correlation weight is matched and the node with the highest action correlation weight is selected as the target mapping point. A directed graph topology represented by an adjacency matrix is ​​constructed in memory to generate the actual flow edge set.

3. The intelligent cold chain risk management system based on ANN-GERT hybrid modeling as described in claim 1, characterized in that, The parameter generation module maps the global return ternary chain set to a return fusion vector, including: Extract the underlying business features corresponding to the return trigger segment, the return acceptance segment set, and the re-entry segment respectively, and transform the underlying business features into corresponding dense vectors and normalized values. After obtaining the dense vector and normalized value of a single flow record, the temporal average pooling operation is called to aggregate multiple records in the back-to-back segment set into a fixed-length back-to-back tensor, and then horizontally concatenates it with the fixed-length tensors corresponding to the trigger feature and re-entry feature to form a back-to-back memory input unit. The back-to-memory input unit is fed into the pre-trained artificial neural network model, and the trigger encoding vector, the accept encoding vector, and the re-entry encoding vector of equal length are mapped and output through independent back-to-trigger branches, back-to-accept branches, and re-entry branches, respectively. In the feature fusion layer, the trigger and receiving encoding vectors, and the receiving and re-entry encoding vectors are multiplied element-wise, and then concatenated and fused with each independent encoding vector to generate a return fusion vector.

4. The intelligent cold chain risk management system based on ANN-GERT hybrid modeling as described in claim 3, characterized in that, The parameter generation module sends the return fusion vector to the multi-task output layer to generate return risk parameters, including: The returned fusion vector is fed into the multi-task output layer of the pre-trained artificial neural network model, and feature mapping operations are performed through a fully connected layer and a differential activation function. In the multi-task output layer, the branch for continuing to fulfill the successful parameters calls the Sigmund activation function to generate the parameters for continuing to fulfill the successful parameters. The quality risk bearing parameter output branch calls the linear rectification activation function or the smooth linear rectification activation function to generate the quality risk bearing parameter; The output branch of the role contribution allocation vector calls the normalized exponential activation function to generate the role contribution allocation vector. The cost-bearing parameter output branch calls the linear identity mapping to output the relative cost value, and combines the pre-stored cost extremum parameter to perform an inverse normalization mapping on the relative cost value to generate the cost-bearing parameter.

5. The intelligent cold chain risk management system based on ANN-GERT hybrid modeling according to claim 4, characterized in that, The pre-trained artificial neural network model undergoes backpropagation training based on historical cold chain batch data during the offline training phase, including: Historical training samples are extracted from the data warehouse, and a continued fulfillment success label is constructed based on the end-of-life fulfillment results after the historical batches have completed the return processing. Quality risk labels are constructed by normalizing and mapping historical handover and acceptance results based on the quality grading table. Extract the cost extreme value parameters of historical training samples, and perform maximum and minimum value normalization on the accumulated actual disposal costs of historical batches to construct return cost labels; Introducing a Laplace smoothing term to statistically analyze the smoothing percentage of the handling costs of the return trigger segment, return acceptance segment, and re-entry segment in the total anomaly cost, and constructing role contribution labels; Construct a multi-task joint loss function, calculate the cross-entropy loss between the network predicted output value and the corresponding real business training label for the parameters of continued successful performance and role contribution allocation vector, and calculate the mean squared error loss between the network predicted output value and the corresponding real business training label for the parameters of quality risk bearing and cost bearing. The network weight parameters are updated by minimizing the joint loss function of multiple tasks using the gradient descent optimization algorithm.

6. The intelligent cold chain risk management system based on ANN-GERT hybrid modeling as described in claim 1, characterized in that, The hyperedge construction module combines quality risk tolerance parameters, cost tolerance parameters, and continued performance success parameters to construct a return hyperedge risk cost transfer function, including: A constant term is introduced to represent the benchmark sunk cost. The quality risk bearing parameter is superimposed on the benchmark sunk cost as a punitive amplification factor and multiplied with the cost bearing parameter to calculate the super-marginal risk adjustment cost. By introducing a moment generating function auxiliary variable, multiplying the moment generating function auxiliary variable with the hyperedge risk adjustment cost to generate a product term, performing a transformation on the product term using an exponential mapping operator with the natural constant as the base, and then multiplying it with the continued performance success parameter, a return hyperedge risk cost transfer function based on random network frequency domain mapping is constructed.

7. The intelligent cold chain risk management system based on ANN-GERT hybrid modeling as described in claim 6, characterized in that, The attribution control module parses the return superedge risk cost transfer function and combines it with the role contribution allocation vector to generate the contribution value of a single return node, including: The first derivative operation is performed on the return superedge risk cost transfer function at the point where the auxiliary variable of the moment generating function takes zero, and the expected value is extracted as the return contribution benchmark value. For any business status node in the enhanced risk network, extract the number of valid business records contained in the current business status node, count the total number of valid business records of all business status nodes belonging to the same role as the current business status node, and calculate the ratio of the number of valid business records to the total number of valid business records to confirm the node evidence carrying ratio. Call the corresponding role weight parameter in the role contribution allocation vector, perform a multiplication operation on the role weight parameter, the node evidence carrying ratio and the return contribution benchmark value to generate the single return node contribution value of the current business status node in the abnormal topology loop.

8. The intelligent cold chain risk management system based on ANN-GERT hybrid modeling according to claim 7, characterized in that, The attribution control module extracts core nodes and generates intelligent control instructions, including: Traverse the global return ternary chain set and perform a loop accumulation operation on the single return node contribution value generated by the business status node in each return ternary chain to obtain the total return node contribution value; Use the heap sort algorithm to sort the total return node contribution values ​​of the business status nodes in descending order to extract the core nodes; When the total return node contribution values ​​are equal, compare the total number of valid business records and extract the business status node with the highest total number of valid business records as the core node. When the total number of valid business records is still equal, the underlying log entry sequence number is called to perform a first-in-first-out comparison, and the business status node with the smallest underlying log entry sequence number is extracted as the core node. Obtain the node role label and business action type field bound to the core node, trigger the action scheduling engine to perform key-value pair matching, extract the corresponding specific verification action instruction, fill the specific verification action instruction into the data payload of the intelligent management and control instruction, and issue it for execution.

Images