Multi-level document-dependent freight trial calculation data linkage calculation method and system

Abstract

The present application relates to the field of data linkage computing technology, and provides a method and system for the linkage computing of freight trial calculation data with multi-level document dependency. By modeling the spatiotemporal dimensions of time information and geographic identification, the attributes of the document node are expanded from a single business field to a composite vector containing spatiotemporal features, so that the dependency graph can capture the fine-grained dependency relationship of the freight business and improve the spatiotemporal accuracy of data linkage; when the basic information or document information changes, the priority and calculation path of the task queue are updated in real time through deep reinforcement learning to avoid calculation errors caused by change delays; through tensor synthesis operations and parallel sub-task division, the dual verification mechanism of domain knowledge graph and forced dependency edges is ensured to ensure that the entire process from dependency graph construction to calculation result output complies with industry standards, significantly improving the compliance and risk control capabilities of the freight business.

Description

Technical Field

[0001] The present application relates to the field of data linkage calculation technology, and in particular to a method and system for linkage calculation of freight trial calculation data that relies on multiple levels of documents. Background Art

[0002] In the digital operations of online freight platforms, freight cost calculations require the interconnectedness of multiple levels of documents, including carrier orders, shipping orders, and receipts. Typical scenarios include nested multimodal transport segmented billing rules, such as the dynamic superposition of land transport mileage tiered pricing and warehousing handling fees; order status changes triggering chain adjustments to fees, such as the automatic association of shipper order changes with liquidated damages clauses in the carrier contract; and deep coupling of capital flows with logistics documents, such as receipt generation relying on transport track verification and electronic receipt signature, forming a closed-loop data loop for the entire process. However, the problem of how to achieve multi-level document data linkage and interconnected calculations in dynamic scenarios remains unresolved. Summary of the Invention

[0003] In response to the deficiencies of the existing technology, this application provides a method and system for linking calculations of freight trial calculation data that relies on multiple levels of documents.

[0004] In a first aspect, the present application provides a freight trial calculation data linkage calculation system with multi-level document dependency, the system comprising: a dependency determination module, an impact propagation module, a path generation module, and a collaborative calculation module;

[0005] The dependency determination module is used to treat document information as nodes, attach a set of dimension attributes to each node. Each node's dimension attribute set includes time information and geographic identification, and generates a document feature vector. The document feature vector includes document type, amount, and time constraint. Based on each node's dimension attribute set, historical association frequency, and business rules, hierarchical dependencies between nodes are generated. The dependency graph is determined based on the domain knowledge graph and the hierarchical dependencies between nodes.

[0006] The impact propagation module monitors changes to basic and document information in real time, analyzes the impact of these changes on the dependency graph, and generates a task queue with priority tags and computing resources based on document feature vectors. Priority tags include business urgency, impact amount, and time sensitivity.

[0007] The path generation module uses computation time, resource usage, and computation accuracy as constraints, generates a set of candidate paths based on the task queue and dependency graph through deep reinforcement learning, and selects the computation path from the set of candidate paths.

[0008] The collaborative computing module is used to divide the task into multiple parallel subtasks according to the calculation path. Each parallel subtask includes a basic quantity and a floating quantity. The basic quantity and the floating quantity are subjected to tensor synthesis operations to obtain tensor calculation results, and the tensor calculation results are verified for consistency to obtain a verified calculation result set and an error result set. The dependency graph and candidate path set are updated based on the execution effect of the calculation result set and the error result set.

[0009] As an optional implementation, the dependency graph determination logic includes:

[0010] Generate basic dependency edges according to the association weight matrix in the hierarchical dependency relationship between nodes, and label the hierarchical attributes of the basic dependency edges according to the node levels in the hierarchical dependency relationship between nodes to output an initial heterogeneous graph;

[0011] Combined with the domain knowledge graph, business rules are decomposed into a graph structure and marked as mandatory dependency edges to generate a set of rule constraints;

[0012] The nodes in the initial heterogeneous graph are traversed to check whether they meet the rule constraint set, so as to update the initial heterogeneous graph, and the dependency graph is determined based on the updated initial heterogeneous graph.

[0013] As an optional implementation, the hierarchical dependency generation sub-logic includes:

[0014] Extract the dimension attribute set of each node and count the number of linkages between nodes and business rules;

[0015] The historical association frequency is determined based on the number of linkages between nodes, the rule weight is determined based on the business rules, and the spatiotemporal coupling degree is determined by calculating the cosine similarity between nodes. The historical association frequency, rule weight, and spatiotemporal coupling degree are weighted and fused to form an association weight matrix.

[0016] The association weight matrix is ​​divided into multiple node levels through the fuzzy clustering algorithm. The membership of the node level is determined according to the importance of the node. The association weight matrix is ​​updated according to the time difference between nodes. The hierarchical dependency relationship between nodes is generated by combining the node level and the association weight matrix.

[0017] As an optional implementation, the additional sub-logic of the dimension attribute set includes:

[0018] Obtain business system data and IoT data, standardize them and form a raw data set, which includes timestamps, geographic coordinates, document types, amounts, and time constraints.

[0019] Extract the timestamps from the original data set and slice them according to the transportation cycle to form multiple time windows. Within each time window, combine the historical time deviation to output the time fluctuation probability. Combine the time window code and the time fluctuation probability to form time information.

[0020] Extract the geographic coordinates from the original dataset, convert them into H3 hierarchical grids, and combine traffic API data to annotate each grid with a travel time weight and a transportation cost coefficient to form a geographic identifier. The geographic identifier includes the grid code, travel time weight, and transportation cost coefficient.

[0021] The time information and geographic identification are concatenated to form a dimension attribute set, and the dimension attribute set is attached to each document information.

[0022] As an optional implementation, the analysis logic of the influence degree of the dependency graph includes:

[0023] Real-time monitoring of changes to basic information and document information. Analyzing the types of change events through natural language processing methods, including timeliness changes, geographic changes, and rule changes, extracting feature vectors of change events and comprehensively outputting change event features, including the type and scope of impact.

[0024] Dynamically traverse the nodes, basic dependency edges, and mandatory dependency edges of the dependency graph according to the characteristics of the change event to locate the initial impact node;

[0025] The graph diffusion algorithm is used to determine the influence scores of nodes at different times based on the confidence of basic dependency edges and forced dependency edges and the node hierarchy, and the influence score matrix of the dependency graph is output;

[0026] The influence levels are divided according to the influence score matrix of the dependency graph, and the influence levels are superimposed on each grid.

[0027] As an optional implementation, the task queue generation logic includes:

[0028] Assign business urgency based on the impact level, dynamically adjust the business urgency based on the document type in the document feature vector, determine the impact amount by combining the node's impact score at different times and the amount in the document feature vector, and determine the timeliness sensitivity based on the timeliness constraints in the document feature vector to output a priority tag.

[0029] The priority score is determined by weighted summing of business urgency, impacted amount, and time sensitivity, and the weights of business urgency, impacted amount, and time sensitivity are dynamically adjusted based on historical processing records.

[0030] The level of the task queue is determined based on the priority score, computing resources are pre-allocated according to the level of the task queue, and a task queue with priority tags and computing resources is generated.

[0031] As an optional implementation, the screening logic of the calculation path includes:

[0032] Determine the state space and action space of deep reinforcement learning based on task queues with priority tags and computing resources and dependency graphs;

[0033] Computational time, resource usage, and computational accuracy are used as constraints. The reward function of deep reinforcement learning is determined based on the computational time ratio, resource utilization, computational accuracy score, and number of violations, and a set of candidate paths is generated through continuous iteration.

[0034] Based on the grids and time windows in the dependency graph, the candidate path set is spatially and temporally pruned to output an optimized candidate path set. Spatial pruning involves merging tasks within the same grid into grid task groups and executing them in parallel. Temporal pruning involves adjusting tasks with overlapping time windows and low timeliness sensitivity to batch processing.

[0035] Monitor the system status in real time, extract the resource efficiency of each path in the system status, determine the time compliance rate of each path in the optimized candidate path set based on the time constraint, and determine the computational cost of each path. The comprehensive score of each path is obtained to screen out the computational path. The computational path includes the execution order of the task queue, the resource allocation plan, and the configuration of the computational accuracy.

[0036] As an optional implementation, the logic for obtaining the tensor calculation result includes:

[0037] Divide the task into multiple parallel subtasks according to the computational path. Each parallel subtask includes a base quantity and a floating quantity. Through grid encoding and time window encoding, the base quantity and floating quantity are aligned to the same spatiotemporal coordinate system, and a normalized tensor pair is output.

[0038] The document type mapping based on the dimension attribute set of each node and the document feature vector is used as the basic quantity, and the amount and time constraints based on the dimension attribute set of each node and the document feature vector are mapped as floating quantities.

[0039] Extract the basis quantity from the standardized tensor pair, process the basis quantity through spatiotemporal convolution operation, update the processed basis quantity according to business rules, obtain the constrained basis quantity, and obtain the tensor calculation result after fusing the processed basis quantity, constrained basis quantity and floating quantity.

[0040] As an optional implementation, the dependency graph update logic includes:

[0041] Extract the spatiotemporal dimensions of the tensor calculation results to see if they match the time window and grid in the document feature vector, and verify whether the tensor calculation results comply with the rule constraint set through the domain knowledge graph;

[0042] Compare the tensor calculation results with the average deviation of historical data to verify whether there are any anomalies, and obtain a set of verified calculation results and an error result set;

[0043] Based on the error result set, the basic dependency edges and mandatory dependency edges of the dependency graph are traced back, and the weights of the mandatory dependency edges and basic dependency edges are updated according to the type of error;

[0044] Adjust the attenuation coefficient of the dependency graph according to the execution effect of the calculation result set, and output the updated dependency graph.

[0045] In a second aspect, the present application provides a method for linking and calculating freight trial calculation data with multi-level document dependency, the method comprising: using document information as a node, attaching a set of dimension attributes to each node, and generating a document feature vector;

[0046] Generate hierarchical dependencies between nodes based on the dimension attribute set, historical association frequency, and business rules of each node, and determine the dependency graph based on the domain knowledge graph and the hierarchical dependencies between nodes;

[0047] Monitor changes in basic information and document information in real time, and analyze the impact of changes in basic information and document information on the dependency graph;

[0048] Combine the document feature vector to generate a task queue with priority tags and computing resources;

[0049] Taking computation time, resource usage, and computation accuracy as constraints, deep reinforcement learning is used to generate a set of candidate paths based on the task queue and dependency graph, and the computation path is selected from the set of candidate paths.

[0050] Divide the task into multiple parallel subtasks according to the calculation path. Each parallel subtask includes a basic quantity and a floating quantity. The basic quantity and the floating quantity are combined into a tensor to obtain the tensor calculation result.

[0051] The tensor calculation results are verified for consistency to obtain the verified calculation result set and error result set, and the dependency graph and candidate path set are updated based on the execution effect of the calculation result set and the error result set.

[0052] Compared with the existing technology, the beneficial effects of the present application are: through a closed-loop architecture from dependency graph construction to impact propagation analysis to path optimization generation and then to collaborative calculation and feedback, the technical problems of difficult dynamic modeling of multi-level document dependency relationships, low efficiency of computing resource scheduling and delayed data linkage response in freight trial calculations are solved; through spatiotemporal dimension modeling of time information and geographic identification, the attributes of document nodes are expanded from a single business field to a composite vector containing spatiotemporal features, so that the dependency graph can capture the fine-grained dependency relationships of freight business and improve the spatiotemporal accuracy of data linkage; when basic information or document information changes, the priority and calculation path of the task queue are updated in real time through deep reinforcement learning to avoid calculation errors caused by change lags in traditional static models; through tensor synthesis operations and parallel sub-task division, the computing tasks are dynamically decomposed into different spatiotemporal coordinate systems, combined with the resource pre-allocation strategy of deep reinforcement learning to improve CPU / GPU utilization; the dual verification mechanism of domain knowledge graph and forced dependency edge ensures that the entire process from dependency graph construction to calculation result output complies with industry standards, thereby improving the compliance and risk control capabilities of freight business.

[0053] Document information is abstracted into nodes with time and space dimensions, and hierarchical dependencies are generated through historical association frequencies, business rules, and dimension attribute sets. Combined with the domain knowledge graph, a dependency graph containing basic dependency edges and mandatory dependency edges is generated, breaking through the flat description of document dependencies in traditional database table structures. Rules are decomposed into mandatory dependency edges to avoid execution loopholes caused by implicit rules.

[0054] Monitor change events in real time, parse event types through natural language processing, analyze the impact of changes in basic information and document information on the dependency graph, and generate a priority-tagged task queue based on business urgency, impact amount, and time sensitivity.

[0055] With computational time, resource usage, and computational accuracy as constraints, candidate paths are generated through deep reinforcement learning. Computational paths that include execution order, resource allocation, and accuracy configuration are then screened out through spatiotemporal pruning. Deep reinforcement learning can generate multiple candidate path sets based on task queues and dependency graphs, and then screen out the optimal computational path from the candidate path set.

[0056] The task is divided into parallel subtasks, and the basic quantity and floating quantity are integrated through tensor synthesis operation. After consistency verification, the dependency graph and candidate path set are updated. The parallel subtask mechanism is combined with the alignment of the space-time coordinate system to support the distribution of computing tasks to heterogeneous computing resources such as edge nodes and GPU clusters, and realize the second-level linkage calculation of tens of millions of documents; by tracing the errors of the calculation results and adjusting the attenuation coefficient of the dependency graph, the system can automatically learn the changes in business models. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be derived from these drawings without inventive work. Among them:

[0058] Figure 1 A system flow chart of a multi-level document-dependent freight trial calculation data linkage calculation system provided in an embodiment of the present application;

[0059] Figure 2 A logic diagram for determining a dependency graph of a freight trial calculation data linkage calculation system with multi-level document dependencies provided in an embodiment of the present application;

[0060] Figure 3 A screening logic diagram of the calculation path of the multi-level document-dependent freight trial calculation data linkage calculation system provided in an embodiment of the present application;

[0061] Figure 4 This is a flow chart of a method for linking calculation of freight trial calculation data based on multi-level documents provided in an embodiment of the present application. DETAILED DESCRIPTION

[0062] In order to make the objectives, technical solutions and advantages of the embodiments of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application are clearly and completely described below in conjunction with the drawings in the specification. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0063] Example 1

[0064] like Figure 1 As shown, a system flow chart of a multi-level document-dependent freight trial calculation data linkage calculation system is provided for an embodiment of the present application. The system includes a dependency determination module, an impact propagation module, a path generation module and a collaborative calculation module.

[0065] The dependency determination module is used to take document information as a node, attach a set of dimension attributes to each node, and generate a document feature vector. The document feature vector includes document type, amount and time constraint. The hierarchical dependency relationship between nodes is generated based on the dimensional attribute set, historical association frequency and business rules of each node, and the dependency graph is determined based on the domain knowledge graph and the hierarchical dependency relationship between nodes.

[0066] Furthermore, the additional sub-logic of the dimension attribute set includes:

[0067] Obtain business system data and IoT data, standardize them and form a raw data set, which includes timestamps, geographic coordinates, document types, amounts, and time constraints.

[0068] Extract the timestamps from the original data set and slice them according to the transportation cycle to form multiple time windows. Within each time window, combine the historical time deviation to output the time fluctuation probability. Combine the time window code and the time fluctuation probability to form time information.

[0069] Extract the geographic coordinates from the original dataset, convert them into H3 hierarchical grids, and combine traffic API data to annotate each grid with a travel time weight and a transportation cost coefficient to form a geographic identifier. The geographic identifier includes the grid code, travel time weight, and transportation cost coefficient.

[0070] The time information and geographic identification are concatenated to form a dimension attribute set, and the dimension attribute set is attached to each document information.

[0071] Freight business data comes from a wide range of sources and in various formats. Business system data records the basic information of orders, while IoT data provides real-time transportation status. In order to uniformly process and analyze these data in the future, they need to be standardized to eliminate the impact of data format differences. For business system data, data is pulled according to the API specifications of different business systems, where basic order information, including order number, customer information, and cargo type, is obtained from the ERP system. For IoT data, message queues are used to receive real-time data transmitted by sensors, including vehicle location, driving speed, cargo temperature and humidity, etc. The data is cleaned and converted. For missing values, different filling methods are used according to the data distribution and business logic. If the data is numerical and the data distribution is relatively uniform, the mean is used for filling. If the data has obvious skewness, the median is used for filling. For categorical data, the most frequently occurring value is used for filling.

[0072] At the same time, the data format is converted to ensure that the timestamp format of all data is unified and the geographic coordinates use the same coordinate system. Finally, a standardized original data set is formed. The original data set includes timestamp, geographic coordinates, document type, amount and time constraint. Document type refers to the label in the business system, including purchase order, booking order and customs declaration, as well as cross-border order or domestic order divided by business model, and sea waybill and land waybill divided by transportation mode; the amount includes the original amount and cost structure, among which the original amount refers to the face value of the document, including the order amount and freight amount, and the cost structure refers to the cost of the amount, including transportation fee, warehousing fee and insurance fee; time constraint includes absolute time and flexible time. Absolute time refers to the specified time of the document, that is, delivery within the specified time, while flexible time refers to the allowable delay threshold, that is, the maximum delay time.

[0073] The unified data format and structure make subsequent data processing and analysis more efficient and accurate, avoid errors and omissions caused by inconsistent data formats, improve the system's utilization of data, and provide an accurate and standardized data basis for the subsequent extraction of time information and geographic identification. If data standardization is not thorough, it will lead to problems such as inaccurate time slicing and geographic coordinate conversion errors, thereby affecting the construction of the entire dimension attribute set.

[0074] The timeliness of freight business is very critical. Different time points and time periods have a significant impact on transportation costs and timeliness. By slicing the timestamps and analyzing the probability of timeliness fluctuations in combination with historical timeliness deviations, the time characteristics of the freight business can be described more accurately. First, according to different transportation scenarios and business needs, the appropriate transportation cycle is determined. For domestic short-distance transportation, the transportation cycle is days, while for international long-distance transportation, the transportation cycle is weeks or months. Then, the timestamps are time-sliced ​​according to the transportation cycle to form multiple time windows. Preferably, one day is used as the transportation cycle, and one day is divided into 24-hour time windows. In each time window, historical transportation data is collected and timeliness deviations are analyzed. In the case of a bad situation, a machine learning algorithm is used to predict the time deviation, output the time fluctuation probability, and assign a unique code to each time window. The code of the time window and the time fluctuation probability are combined to form time information; thus, the time pattern of the freight business can be grasped more accurately, and a more accurate time basis can be provided for the subsequent hierarchical dependency generation and task scheduling. That is, in the time window with high time fluctuation probability, more transportation resources can be arranged in advance to cope with possible delays. Time information will be an important part of the dimension attribute set, affecting the calculation of the spatiotemporal coupling between nodes. In the hierarchical dependency generation, nodes with similar time have a higher coupling degree, which affects the association weight between nodes.

[0075] The geographical factors of freight business have a significant impact on transportation costs, timeliness and route selection. Converting geographic coordinates into H3 hierarchical grids and combining traffic API data to annotate the travel time weight and capacity cost coefficient for each grid can describe geographic spatial information in more detail. Use the H3 library to convert geographic coordinates into the encoding of the H3 hierarchical grid. The H3 grid is a hexagonal grid system with good spatial consistency and scalability. According to different business needs, choose the appropriate grid level. For freight business within the city, choose a finer grid level, while for long-distance transportation business, choose a coarser grid level. Call the AutoNavi Map API to obtain traffic information for each grid, including road congestion. and average driving speed, etc. Based on this information, the travel time weight and capacity cost coefficient are marked for each grid, where the travel time weight reflects the difficulty of travel within the grid, and the capacity cost coefficient reflects the transportation cost within the grid; this can more accurately evaluate the transportation cost and timeliness of different geographical areas, providing a more scientific basis for path planning and resource allocation. When selecting a transportation route, grids with lower travel time weights are given priority. Geographical identification will be another important component of the dimensional attribute set, and together with time information, it will affect the calculation of the spatiotemporal coupling between nodes. In the generation of hierarchical dependency relationships, geographically adjacent nodes have higher coupling, which affects the association weight between nodes.

[0076] Time information and geographic identifiers are spliced ​​together to form a dimension attribute set. Attaching the dimension attribute set to each document information can provide more comprehensive node attribute information for subsequent hierarchical dependency generation and dependency graph construction; time information and geographic identifiers are spliced ​​together to form a composite data structure containing time information and geographic identifiers, that is, the time information and geographic identifiers are stored together in JSON format, and the dimension attribute set is attached to the node corresponding to each document information through SQL statements in database operations. In graph databases, including Neo4j, the dimension attribute set is stored as a node attribute; each document node has richer attribute information, so that subsequent hierarchical dependency generation and dependency graph construction can consider more factors, improve the accuracy and reliability of the dependency graph, and provide comprehensive node attribute information for hierarchical dependency generation. The node's dimension attribute set will serve as an important basis for calculating historical association frequency, rule weight, and spatiotemporal coupling.

[0077] Furthermore, the generation logic of hierarchical dependencies includes:

[0078] Extract the dimension attribute set of each node and count the number of linkages between nodes and business rules;

[0079] The historical association frequency is determined based on the number of linkages between nodes, the rule weight is determined based on the business rules, and the spatiotemporal coupling degree is determined by calculating the cosine similarity between nodes. The historical association frequency, rule weight, and spatiotemporal coupling degree are weighted and fused to form an association weight matrix.

[0080] The association weight matrix is ​​divided into multiple node levels through the fuzzy clustering algorithm. The membership of the node level is determined according to the importance of the node. The association weight matrix is ​​updated according to the time difference between nodes. The hierarchical dependency relationship between nodes is generated by combining the node level and the association weight matrix.

[0081] The generation of hierarchical dependency relationships requires comprehensive consideration of the node's dimensional attribute set, historical association frequency, and business rules. By extracting the dimensional attribute set of each node and counting the number of linkages between nodes and business rules, data support can be provided for the subsequent association weight calculation. The dimensional attribute set of each node, including time information and geographic identification, is extracted from the graph database. The required node attributes are obtained using database query statements. The number of linkages between nodes is counted by analyzing the document association relationships in historical transportation data. That is, association rule mining is used to find document combinations that often appear together and count their number of occurrences. At the same time, business rules are sorted out to clarify the dependencies and constraints between different types of documents. This provides an accurate data basis for the subsequent association weight calculation, allowing the association weight to more realistically reflect the actual dependency between nodes. The node's dimensional attribute set, linkage number, and business rules will serve as input for calculating historical association frequency, rule weight, and spatiotemporal coupling, directly affecting the generation of the association weight matrix.

[0082] The historical association frequency, rule weight and spatiotemporal coupling reflect the strength of the association between nodes from different perspectives. By weighted fusion of them to form an association weight matrix, multiple factors can be comprehensively considered to more comprehensively describe the hierarchical dependency between nodes. The historical association frequency is calculated according to the number of linkages between nodes. The more linkages there are, the higher the historical association frequency is. The historical association frequency is expressed by a simple statistical method, that is, the ratio of the number of linkages to the total number of linkages. The rule weight is determined based on the business rules. A higher rule weight is assigned to the mandatory dependent business rules, while a lower rule weight is assigned to the optional dependent business rules. The cosine phase between nodes is calculated to obtain the historical association frequency. Similarity is used to determine the spatiotemporal coupling. The time information and geographic identification of the nodes are converted into vector form, and then the cosine similarity between them is calculated. The higher the cosine similarity, the higher the spatiotemporal coupling. Finally, the historical association frequency, rule weight and spatiotemporal coupling are weighted and fused to form an association weight matrix. The distribution of weights will be adjusted according to business experience and actual conditions. The association weight matrix can comprehensively consider multiple factors and more accurately describe the hierarchical dependency relationship between nodes, avoiding the limitations of a single factor in judging the dependency relationship. The association weight matrix will be used as the input of the fuzzy clustering algorithm to divide the node hierarchy. Different association weights will lead to different node hierarchy division results.

[0083] Dividing the association weight matrix into multiple node levels through the fuzzy clustering algorithm can more clearly display the hierarchical structure between nodes, determine the node level membership according to the importance of the node, and update the association weight matrix according to the time difference between nodes, so that the hierarchical dependency is more in line with the actual business situation; use the fuzzy clustering algorithm to perform cluster analysis on the association weight matrix, set the appropriate number of clusters and clustering parameters, and divide the nodes into different levels; determine the node level membership according to the importance of the node, and the importance of the node needs to be evaluated based on factors such as the business value and correlation of the node, among which the core business node has a higher membership and the edge business node has a lower membership; update the association weight matrix according to the time difference between nodes. The larger the time difference, the lower the association weight will be. Through the attenuation function To update the association weights, represents the updated association weight, represents the association weight before updating, represents a constant, Represents the attenuation coefficient, attenuation coefficient Determined according to time constraints, cross-border logistics has a long process cycle. The value is small and the decay speed is slow. Domestic express delivery has a short process cycle. The larger the value, the faster the decay rate. Represents the time difference between nodes; combines the node hierarchy and the associated weight matrix to generate the hierarchical dependency between nodes. In the graph database, the hierarchical dependency between nodes is represented by adding edges and edge attributes; a clear node hierarchy structure and accurate hierarchical dependency can provide more effective guidance for subsequent dependency graph construction and task scheduling. The hierarchical dependency between nodes will serve as an important basis for determining the dependency graph, affecting the generation of basic dependency edges and the annotation of hierarchical attributes.

[0084] Specifically, if Figure 2 As shown, the determination logic of the dependency graph includes:

[0085] Generate basic dependency edges according to the association weight matrix in the hierarchical dependency relationship between nodes, and label the hierarchical attributes of the basic dependency edges according to the node levels in the hierarchical dependency relationship between nodes to output an initial heterogeneous graph;

[0086] Combined with the domain knowledge graph, business rules are decomposed into a graph structure and marked as mandatory dependency edges to generate a set of rule constraints;

[0087] The nodes in the initial heterogeneous graph are traversed to check whether they meet the rule constraint set, so as to update the initial heterogeneous graph, and the dependency graph is determined based on the updated initial heterogeneous graph.

[0088] Based on the association weight matrix in the hierarchical dependency relationship between nodes, basic dependency edges are generated, and the hierarchical attributes of the basic dependency edges are labeled according to the node hierarchy. This can display the abstract hierarchical dependency relationship in a graphical manner, which is convenient for subsequent analysis and processing. The association weight matrix is ​​traversed, and for node pairs whose association weight is greater than the set weight threshold, basic dependency edges are added to the graph database. The weight of the basic dependency edge is set according to the value in the association weight matrix, and the hierarchical attributes of the basic dependency edge are labeled according to the node hierarchy. For the dependency edge between the core layer node and the middle layer node, the labeled hierarchical attributes are core and middle, while for the dependency edge between the middle layer node and the edge layer node, the labeled hierarchical attributes are middle and edge. Then, the initial heterogeneous graph is output, which contains nodes and basic dependency edges, as well as the hierarchical attributes of the edges. The initial heterogeneous graph intuitively displays the hierarchical dependency relationship between nodes, provides a basis for the subsequent addition of rule constraints and update of the dependency graph, and affects the generation of the rule constraint set and the update of the initial heterogeneous graph.

[0089] Combined with the domain knowledge graph, the business rules are decomposed into a graph structure and marked as mandatory dependency edges to generate a rule constraint set, which can ensure that the dependency graph complies with business rules and industry specifications; the domain knowledge graph is analyzed to extract business rules related to freight business, including that dangerous goods transportation must first undergo a safety assessment, and that customs declarations must be submitted before the arrival of goods. The business rules are decomposed into a graph structure and implemented using the query and operation statements of the graph database. The rules that trigger safety assessments for dangerous goods transportation are converted into edges in the graph structure and marked as mandatory dependency edges to generate a rule constraint set. The rule constraint set contains all mandatory dependency edges and corresponding rule descriptions; the rule constraint set can ensure that the dependency graph complies with business rules and industry specifications, avoiding business risks caused by violations of rules. The rule constraint set will be used to check whether the nodes in the initial heterogeneous graph meet the rule requirements, thereby updating the basic dependency edges in the initial heterogeneous graph.

[0090] Traversing the nodes in the initial heterogeneous graph and checking whether they comply with the rule constraint set to update the initial heterogeneous graph can ensure the accuracy and compliance of the dependency graph; traversing the nodes and basic dependency edges in the initial heterogeneous graph and checking them according to the rule constraint set. If a node or edge does not meet the rule requirements, corresponding processing is performed. When a dangerous goods transportation node has no associated safety assessment node, a corresponding mandatory dependency edge is added. The dependency graph is determined based on the updated initial heterogeneous graph. In the graph database, the attributes of the nodes and edges, as well as the structure of the graph, are updated to finally obtain a dependency graph that complies with business rules and actual conditions. The accuracy and compliance of the dependency graph are guaranteed, which can provide a reliable foundation for subsequent impact propagation analysis and path generation. The dependency graph will serve as an important input for the impact propagation module and the path generation module, and will be used to analyze the impact of changes in basic information and document information on the dependency graph, as well as to generate candidate path sets and screen calculation paths.

[0091] The impact propagation module is used to monitor changes in basic information and document information in real time, analyze the impact of changes in basic information and document information on the dependency graph, and generate a task queue with priority tags and computing resources based on the document feature vector. The priority tags include business urgency, impact amount, and time sensitivity.

[0092] Specifically, the analysis logic of the influence degree of the dependency graph includes:

[0093] Real-time monitoring of changes to basic information and document information. Analyzing the types of change events through natural language processing methods, including timeliness changes, geographic changes, and rule changes, extracting feature vectors of change events and comprehensively outputting change event features, including the type and scope of impact.

[0094] Dynamically traverse the nodes, basic dependency edges, and mandatory dependency edges of the dependency graph according to the characteristics of the change event to locate the initial impact node;

[0095] The graph diffusion algorithm is used to determine the influence scores of nodes at different times based on the confidence of basic dependency edges and forced dependency edges and the node hierarchy, and the influence score matrix of the dependency graph is output;

[0096] The influence levels are divided according to the influence score matrix of the dependency graph, and the influence levels are superimposed on each grid.

[0097] In freight business, basic information and document information may change at any time. Timely capture of these changes and accurate analysis of their types and characteristics are the basis for subsequent evaluation of their impact on the dependency graph. Only by clarifying the specific circumstances of the change event can the dependency graph be processed in a targeted manner. In terms of real-time monitoring, data monitoring programs can be set at the interface between the business system and the data source. For the business system, the trigger mechanism of the database is used to trigger the data monitoring program immediately when the basic information or document information is updated. For IoT data, changes in sensor data are received in real time through the message queue. The type of change event is parsed through natural language processing methods, and the text description of the change information is input into the model. The model outputs the type of change event, including time change, geographical change, and rule change. The feature vector of the change event is extracted. , including the impact duration and affected time window in timeliness changes, the H3 grid code and affected radius in geographic changes, and the affected business rules in rule changes, that is, the process adjustment of export tax rebates, to form a feature vector. Then, these feature vectors are combined to output change event features containing the type of change event and the scope of impact. This enables rapid and accurate identification of change events, providing a clear direction for subsequent positioning of impact nodes and assessment of the degree of impact, enabling the system to respond to business changes in a timely manner and avoid calculation errors or business risks caused by untimely processing of changes. Accurate change event features provide a basis for locating the initial impact node. Different types of change events and impact scopes will guide the system to find the affected starting node in the dependency graph, thereby starting subsequent impact diffusion analysis.

[0098] Dependency graphs contain numerous nodes, and change events typically do not affect all nodes. By dynamically traversing the nodes, base dependency edges, and mandatory dependency edges of the dependency graph based on the characteristics of the change event, the initial nodes directly affected by the change can be quickly located, avoiding unnecessary traversals of the entire dependency graph and improving analysis efficiency. The starting conditions for traversal are determined based on the characteristics of the change event. If the change is time-sensitive, traversal begins at the nodes related to timeliness; if the change is geographical, traversal begins at the nodes related to that geographic location. Leveraging the query capabilities of the graph database, adjacent nodes are sequentially checked to see if they meet the impact range of the change event based on the relationships between base dependency edges and mandatory dependency edges. By setting a certain traversal depth or condition, infinite traversals are avoided. When a node that meets the conditions is found, it is marked as the initial impact node. This narrows the scope of impact analysis, improves the relevance and efficiency of the analysis, avoids a comprehensive search of the entire dependency graph, and saves computing resources and time. After determining the initial impact nodes, the subsequent graph diffusion algorithm can start from these nodes and accurately calculate the impact scores of nodes at different times, thereby more accurately assessing the impact of the change on the dependency graph.

[0099] To comprehensively assess the impact of a change on the dependency graph, it is necessary to quantify the impact of each node at different times. Using a graph diffusion algorithm, combined with the confidence of basic and mandatory dependency edges and the node hierarchy, we can more accurately simulate the propagation of impact in the graph. We select an appropriate graph diffusion algorithm to iteratively calculate the impact score for each node based on the edge weights between nodes and the node hierarchy. The edge weights between nodes represent the confidence of basic and mandatory dependency edges. During the calculation, we consider time and assign different weights to different times. Initially, the impact is concentrated near the initially affected node. Over time, the impact gradually spreads. After multiple iterative calculations, we output a matrix containing the impact scores of each node at different times. This quantitatively presents the impact of the change on each node in the dependency graph and its temporal trends, providing concrete data support for the subsequent categorization of impact levels. The impact score matrix is ​​the basis for categorizing impact levels. Based on the scores in the impact score matrix, we can categorize impact levels, providing a more intuitive understanding of the overall impact of the change on the dependency graph.

[0100] A simple impact score matrix is ​​not intuitive enough and difficult to apply directly to business decisions. Dividing the impact into different levels and superimposing them on the grid can more clearly show the impact distribution of change events in geographic space, making it easier for business personnel to allocate resources and make decisions. Based on the impact score matrix, the threshold for dividing the impact level is determined, and the impact score is divided into three levels: high, medium, and low. For each grid, the average impact score of the nodes in the grid is calculated, and the impact level of the grid is determined based on the divided level threshold. The impact level is superimposed on the geographic grid and can be visualized through the geographic information system. It intuitively shows the impact distribution of change events in geographic space, helping business personnel quickly understand which areas are more affected and which areas are less affected, providing a visual basis for resource allocation and decision-making. The impact level will be used to allocate business urgency in task queue generation. Different impact levels correspond to different business urgency, thereby affecting the priority sorting of tasks.

[0101] Specifically, the task queue generation logic includes:

[0102] Assign business urgency based on the impact level, dynamically adjust the business urgency based on the document type in the document feature vector, determine the impact amount by combining the node's impact score at different times and the amount in the document feature vector, and determine the timeliness sensitivity based on the timeliness constraints in the document feature vector to output a priority tag.

[0103] The priority score is determined by weighted summing of business urgency, impacted amount, and time sensitivity, and the weights of business urgency, impacted amount, and time sensitivity are dynamically adjusted based on historical processing records.

[0104] The level of the task queue is determined based on the priority score, computing resources are pre-allocated according to the level of the task queue, and a task queue with priority tags and computing resources is generated.

[0105] When dealing with the impact of change events on the dependency graph, it is necessary to prioritize related tasks to ensure that resources can be allocated to important tasks first. By comprehensively considering business urgency, impact amount and time sensitivity, the importance of tasks can be more comprehensively evaluated. Business urgency is assigned based on the level of impact. The higher the level of impact, the higher the business urgency. A high impact level corresponds to a high business urgency. Business urgency is dynamically adjusted according to the document type in the document feature vector. Different types of documents will have different importance. Urgent orders have higher business urgency. Combined with the impact score of the node at different times and the document characteristics, the business urgency is adjusted dynamically according to the document type in the document feature vector. Different types of documents will have different importance. Urgent orders have higher business urgency. The amount in the feature vector determines the impact amount. The higher the impact score and the larger the amount, the higher the impact amount. The timeliness sensitivity is determined based on the timeliness constraint in the document feature vector. The stricter the timeliness requirement, the higher the timeliness sensitivity. The business urgency, impact amount, and timeliness sensitivity are combined to output a priority tag. This comprehensively considers multiple factors, more accurately assesses the importance of tasks, provides a reasonable basis for task priority sorting, and avoids the problem of inaccurate priority judgment caused by a single factor. The priority tag is the basis for calculating the priority score. The priority of the task will be determined based on these tags in the future, thereby affecting the hierarchical division of the task queue.

[0106] In order to prioritize tasks more accurately, it is necessary to convert priority tags into specific priority scores. At the same time, dynamically adjust weights based on historical processing records to make the priority scores more in line with actual business conditions. The priority score is determined by weighted summation of business urgency, impact amount, and time sensitivity, where priority score = business urgency × weight 1 + impact amount × weight 2 + time sensitivity × weight 3. Analyze historical processing records and adjust weights based on task processing results, where task processing results include whether the task was completed on time and whether errors occurred. If a factor has a greater impact on the successful execution of the task in historical processing, its weight is increased. This results in a specific priority score, which is convenient for task sorting. Dynamically adjusting weights makes priority sorting more in line with actual business needs and improves the efficiency and accuracy of task processing. The priority score will be used to determine the level of the task queue. The higher the priority score, the higher the level of the task in the queue and the higher the priority it receives.

[0107] The hierarchy of the task queue is determined according to the priority score of the task, which can classify tasks according to their importance and facilitate the rational allocation of resources. Pre-allocation of computing resources can ensure that tasks have sufficient resource support during execution and improve the efficiency of task processing. According to the range of priority scores, the task queue is divided into different hierarchies, with tasks with high priority scores as the first-level queue, tasks with medium priority scores as the second-level queue, and tasks with low priority scores as the third-level queue. Computing resources are pre-allocated according to the hierarchy of the task queue. For tasks in the high-priority queue, more computing resources are allocated, and for tasks in the low-priority queue, fewer computing resources are allocated. Tasks are generated into task queues with priority tags and computing resources according to the hierarchy and pre-allocated computing resources. This achieves the rational classification of tasks and the effective allocation of resources, improves the overall processing efficiency of the system, ensures that important tasks can be processed first, and avoids resource waste. The generated task queue will be used as the input of the path generation module to generate a set of candidate paths. The hierarchy of the task queue and the allocation of computing resources will affect the strategy and results of path generation.

[0108] The path generation module is used to use computing time, resource usage, and computing accuracy as constraints, and generate a set of candidate paths based on the task queue and dependency graph through deep reinforcement learning, and filter out the computing path from the candidate path set.

[0109] Specifically, if Figure 3 As shown, the filtering logic for calculating the path includes:

[0110] Determine the state space and action space of deep reinforcement learning based on task queues with priority tags and computing resources and dependency graphs;

[0111] Computational time, resource usage, and computational accuracy are used as constraints. The reward function of deep reinforcement learning is determined based on the computational time ratio, resource utilization, computational accuracy score, and number of violations, and a set of candidate paths is generated through continuous iteration.

[0112] Based on the grids and time windows in the dependency graph, the candidate path set is spatially and temporally pruned to output an optimized candidate path set. Spatial pruning involves merging tasks within the same grid into grid task groups and executing them in parallel. Temporal pruning involves adjusting tasks with overlapping time windows and low timeliness sensitivity to batch processing.

[0113] Monitor the system status in real time, extract the resource efficiency of each path in the system status, determine the time compliance rate of each path in the optimized candidate path set based on the time constraint, and determine the computational cost of each path. The comprehensive score of each path is obtained to screen out the computational path. The computational path includes the execution order of the task queue, the resource allocation plan, and the configuration of the computational accuracy.

[0114] The state space includes task characteristics, spatiotemporal characteristics, and system status. Task characteristics include business urgency, impact amount, time sensitivity, and resource type. Spatiotemporal characteristics include grid encoding and time window encoding. System status includes currently available computing resources and historical path time consumption. The action space includes the execution order selection of the task queue and the adjustment of computing accuracy.

[0115] To screen out the optimal computing path, deep reinforcement learning needs to fully understand the current task and system status in order to make appropriate decisions. The state space contains information such as task characteristics, spatiotemporal characteristics, and system status, which allows deep reinforcement learning to grasp the urgency of the task, spatiotemporal distribution, and system resource conditions. The action space defines the operations that deep reinforcement learning can take, including the execution order selection of the task queue and the adjustment of the calculation accuracy, providing the model with a means to optimize the path; the task characteristics include a four-dimensional feature vector composed of business urgency, impact amount, time sensitivity, and resource type, among which business urgency includes three levels: high, medium, and low. Time sensitivity is dynamically calculated by the ratio of the remaining time constraint to the task processing cycle. The resource type is associated with the computing resource pool corresponding to the encoding of the grid, including CPU intensive and GPU intensive. Preferably, the grid of the hub node is preferentially matched with the GPU cluster.

[0116] The grid encoding in the spatiotemporal features generates multi-resolution embedding vectors using the H3 spatial indexing algorithm, capturing the spatial associations of geographic units at different levels, including cities, regions, and hubs. The encoding of the time window is obtained through the steps in the dependency determination module described above. The system state includes the currently available computing resources and historical path durations. The currently available computing resources can be obtained by the number of CPU cores and memory size, while the historical path duration records the time spent on previous path calculations. The action space includes the execution order selection of the task queue and the adjustment of computational precision. The execution order selection of the task queue refers to the permutation and combination of tasks in the task queue to generate all possible execution orders as one of the actions that deep reinforcement learning can choose. The computational precision is adjusted by pre-setting several different computational precision levels, including high precision, medium precision, and low precision, and deep reinforcement learning can choose among these levels. By accurately constructing the state space and action space, deep reinforcement learning can make decisions based on comprehensive information. The state space allows deep reinforcement learning to understand the actual situation of the current task and system, while the action space provides the model with diverse optimization paths, thereby increasing the possibility of screening out more optimal computational paths.

[0117] In order to make deep reinforcement learning optimize the path in the desired direction, it is necessary to define a reward function. By calculating the time consumption ratio, resource utilization, calculation accuracy score and number of violations, the quality of the path selected by deep reinforcement learning can be quantified. In the continuous iteration process, deep reinforcement learning will try to select the path that can obtain higher rewards based on the feedback of the reward function, thereby generating a set of candidate paths that meet the constraints; record the actual calculation time of each candidate path and compare it with the preset benchmark time to obtain the calculation time ratio. The benchmark time is calculated based on historical experience or theory; monitor the use of computing resources by each candidate path during execution and calculate resource utilization, including calculating CP U usage rate and memory occupancy rate, etc.; evaluate the calculation accuracy score based on the error rate between the calculation result and the true value; check whether each candidate path violates the rules in the dependency graph, such as whether it violates the hierarchical dependency between tasks, etc., and count the number of violations; perform a weighted combination of the calculation time ratio, resource utilization, calculation accuracy score and number of violations to obtain the value of the reward function; deep reinforcement learning continuously explores and selects in the state space and action space, updates its own strategy based on the feedback of the reward function, and after multiple iterations, generates a series of different candidate paths to form a candidate path set; the candidate path set generated through continuous iteration contains a variety of calculation paths, which increases the chance of screening out the optimal path.

[0118] There will be some redundant or unreasonable paths in the generated candidate path set, which will increase the complexity and computational complexity of subsequent screening. Through spatial pruning and temporal pruning, these invalid paths can be removed, the screening range can be reduced, the screening efficiency can be improved, and the execution mode of the path can be optimized to improve resource utilization efficiency. Among them, spatial pruning refers to merging tasks within the same grid into grid task groups in the dependency graph. For example, multiple transportation tasks within the same geographical area are merged into one grid task group, and the tasks within the grid task group are executed in parallel to make full use of computing resources. Parallel computing is achieved through multi-threading or distributed computing technologies. Temporal pruning is a method of merging tasks within the same grid into a grid task group in the dependency graph. The purpose of pruning is to find tasks with overlapping time windows and low timeliness sensitivity, including some non-urgent document processing tasks. If their processing time windows overlap, batch processing is considered, and these tasks are adjusted to batch processing to reduce the overhead of task switching and improve processing efficiency. Spatial pruning and temporal pruning effectively reduce the number of candidate paths and reduce the complexity of screening. At the same time, through task merging and batch processing, resource utilization efficiency is improved, computing time and resource consumption are reduced, and the output optimized candidate path set is more streamlined and efficient, providing more valuable candidate paths for subsequent path screening and reducing the workload and time cost of screening.

[0119] Although the set of optimized candidate paths has been streamlined to a certain extent after spatial pruning and temporal pruning, it still contains multiple paths and needs to be further screened to find the optimal computational path. By comprehensively considering factors such as resource efficiency, time compliance rate and computational cost, the path that best meets actual needs can be selected. The system status is monitored in real time, and the efficiency of computing resources used by each path during execution is recorded to determine resource efficiency, that is, the ratio of the amount of tasks completed per unit time to the amount of resources consumed. According to the time constraint, each path is checked to see if it can complete the task on time, and the time compliance rate is calculated, that is, the ratio of the number of tasks completed on time to the total number of tasks is counted. The computational cost needs to consider the cost of computing resources consumed by each path during execution. Cost, including electricity and equipment depreciation, as well as additional costs due to violations; resource efficiency, time compliance rate and computing cost are weighted and combined to obtain a comprehensive score for each path. According to the comprehensive score, the path with the highest score is selected as the calculation path. The calculation path includes the execution order of the task queue, resource allocation plan and configuration of calculation accuracy; by comprehensively considering multiple factors, the selected calculation path can achieve a better balance in resource utilization, timeliness and cost, thereby improving the efficiency and benefits of the entire freight trial calculation data linkage calculation. The selected calculation path provides a clear execution plan for the collaborative calculation module, which can divide the task into multiple parallel subtasks according to the path for subsequent calculation and processing.

[0120] The collaborative computing module is used to divide the task into multiple parallel subtasks according to the calculation path. Each parallel subtask includes a basic quantity and a floating quantity. The basic quantity and the floating quantity are subjected to tensor synthesis operations to obtain tensor calculation results, and the tensor calculation results are verified for consistency to obtain a verified calculation result set and an error result set. The dependency graph and candidate path set are updated based on the execution effect of the calculation result set and the error result set.

[0121] Specifically, the logic for obtaining tensor calculation results includes:

[0122] Divide the task into multiple parallel subtasks according to the computational path. Each parallel subtask includes a base quantity and a floating quantity. Through grid encoding and time window encoding, the base quantity and floating quantity are aligned to the same spatiotemporal coordinate system, and a normalized tensor pair is output.

[0123] The document type mapping based on the dimension attribute set of each node and the document feature vector is used as the basic quantity, and the amount and time constraints based on the dimension attribute set of each node and the document feature vector are mapped as floating quantities.

[0124] Extract the basis quantity from the standardized tensor pair, process the basis quantity through spatiotemporal convolution operation, update the processed basis quantity according to business rules, obtain the constrained basis quantity, and obtain the tensor calculation result after fusing the processed basis quantity, constrained basis quantity and floating quantity.

[0125] Tasks often have certain complexity and correlation. Dividing the tasks in the calculation path into multiple parallel subtasks can make full use of multi-core processors or distributed computing resources to improve computing efficiency. At the same time, since the basic quantities and floating quantities of different tasks will have different spatiotemporal dimensions, aligning them to the same spatiotemporal coordinate system can facilitate subsequent unified calculations and analysis, and avoid calculation errors or result deviations caused by inconsistent dimensions; according to the dependencies and resource allocation plans of the tasks in the calculation path, the tasks are split into multiple independent parallel subtasks. For document calculation tasks involving multiple transportation links, they are divided according to factors such as transportation stage and geographical area; using the grid code and time window code as index, each task is divided into The base quantities and floating quantities in each parallel subtask are mapped to the same spatiotemporal coordinate system. The base quantities of transportation costs in different geographical grids and the floating quantities caused by market fluctuations in different time windows are matched and aligned through unified grid coding and time window coding. The aligned base quantities and floating quantities are standardized so that they have the same data format and range to facilitate subsequent tensor synthesis operations. Through parallel subtask division and spatiotemporal alignment, the computing efficiency can be significantly improved, and the advantages of multi-core processors or distributed computing resources can be fully utilized. At the same time, the output of the standardized tensor pair provides a unified data basis for subsequent tensor synthesis operations, reducing the complexity of data processing and improving the accuracy and reliability of the calculation results.

[0126] The dimension attribute set and document feature vector of each node contain rich business information. Mapping this information to basic quantity and floating quantity can better reflect the characteristics and changing patterns of freight trial calculation data. Basic quantity usually represents relatively stable business factors, among which the basic transportation cost determined by the document type, and floating quantity reflects the business factors that change with time and market factors, such as fluctuations in amount and changes in time constraints. Through this mapping, freight trial calculation data can be modeled and calculated more accurately; according to the different document types, a mapping relationship between document type and basic quantity is established, among which for different types of transport documents, including ocean bills of lading and air waybills, their corresponding Based on the basic quantities such as basic transportation costs and handling fees; combining the time information and geographical identification in the node's dimensional attribute set, as well as the amount and time constraints in the document feature vector, a floating amount mapping model is established. According to the market conditions and transportation needs of different geographical areas, the floating ratio of the amount is calculated. According to the historical time deviation and current time constraints, the floating cost related to time is determined; by mapping the node's dimensional attribute set and the document feature vector to the basic quantity and floating amount, the dynamic changes of the freight trial calculation data can be more accurately reflected, and the accuracy and reliability of the calculation results can be improved. At the same time, this mapping method makes the calculation process more in line with business logic and facilitates the interpretation and analysis of the calculation results.

[0127] After obtaining the mapping of standardized tensor pairs to basic quantities and floating quantities, the basic quantities need to be further processed to meet the requirements of business rules. The spatiotemporal convolution operation can capture the characteristics and change patterns of the basic quantities in the spatiotemporal dimensions. The processed basic quantities are updated in combination with business rules to ensure that the calculation results meet actual business needs. Finally, the processed basic quantities, constrained basic quantities and floating quantities are fused to obtain the final tensor calculation results, which can comprehensively consider various business factors and improve the accuracy and reliability of the calculation results.

[0128] The basic quantities in the standardized tensor pairs are processed by the spatiotemporal convolutional neural network, which can simultaneously capture the characteristics and changing laws of the basic quantities in the time and space dimensions. For the basic quantities of transportation costs in different geographical grids in different time windows, their spatiotemporal correlations and trends can be extracted through spatiotemporal convolution operations. According to business rules, including transportation policies and market price fluctuations, the basic quantities processed by the spatiotemporal convolution operation are updated to obtain the constrained basic quantities. If a region introduces a new transportation policy that increases transportation costs, then the corresponding basic quantities of the region need to be adjusted accordingly. The processed basic quantities, constrained basic quantities and floating quantities are fused, and the tensor synthesis algorithm is used to obtain the final tensor calculation result, that is, through weighted summation. The three are integrated in this way, and the weights need to be determined based on business experience or historical data. Through spatiotemporal convolution operations and updates to business rules, the characteristics and changing patterns of the basic quantities in the spatiotemporal dimensions can be captured more accurately, ensuring that the calculation results meet actual business needs. The final tensor calculation results comprehensively consider multiple factors such as the basic quantity, constrained basic quantity, and floating quantity, which improves the accuracy and reliability of the calculation results and provides strong support for subsequent consistency verification and decision-making. The obtained tensor calculation results are the objects of subsequent consistency verification, and their accuracy and reliability directly affect the verification results and subsequent dependency graph updates. If the tensor calculation results are accurate and reliable, the pass rate of the consistency verification will be higher, and it can more effectively provide data support for the update of the dependency graph.

[0129] Specifically, the update logic of the dependency graph includes:

[0130] Extract the spatiotemporal dimensions of the tensor calculation results to see if they match the time window and grid in the document feature vector, and verify whether the tensor calculation results comply with the rule constraint set through the domain knowledge graph;

[0131] Compare the tensor calculation results with the average deviation of historical data to verify whether there are any anomalies, and obtain a set of verified calculation results and an error result set;

[0132] Based on the error result set, the basic dependency edges and mandatory dependency edges of the dependency graph are traced back, and the weights of the mandatory dependency edges and basic dependency edges are updated according to the type of error;

[0133] Adjust the attenuation coefficient of the dependency graph according to the execution effect of the calculation result set, and output the updated dependency graph.

[0134] The spatiotemporal dimensions of the tensor calculation results need to be matched with the time windows and grids in the document feature vector to ensure the spatiotemporal consistency of the calculation results with the actual business scenarios. At the same time, the tensor calculation results are checked by rule verification through the domain knowledge graph to ensure that the calculation results comply with the business rules and constraints, and to avoid calculation results that do not conform to the actual business. The spatiotemporal dimensions of the tensor calculation results are compared with the time windows and grids in the document feature vector to check whether there are matching errors. For the calculation results of a certain geographic grid within a specific time window, check whether it is consistent with the document feature vector corresponding to the grid and time window. Use the business rules and constraints in the domain knowledge graph to verify the tensor calculation results, including checking whether the calculation results conform to the business rules and constraints. Whether the reasonable range of transportation costs and time constraints meet the requirements; through spatiotemporal dimension matching and rule verification, the spatiotemporal consistency of tensor calculation results with actual business scenarios and compliance with business rules can be ensured, thereby improving the accuracy and reliability of calculation results. At the same time, this verification mechanism can promptly discover errors and anomalies that may exist in the calculation process, and provide a basis for subsequent error processing and dependency graph updates. The results of spatiotemporal dimension matching and rule verification will determine whether the tensor calculation results can pass the consistency verification, thereby affecting the generation of verified calculation result sets and error result sets. If the matching and verification pass, the calculation result will enter the verified calculation result set, otherwise, it will enter the error result set, providing data support for subsequent tracing and updates.

[0135] Compare the tensor calculation results with historical data to verify whether there are any anomalies, so as to timely discover problems and risks in the calculation process. By dividing the verified calculation result set and the error result set, the calculation results can be classified and managed, which facilitates subsequent tracing and updating. Calculate the average deviation between the tensor calculation results and the historical data, set a reasonable deviation threshold. If the deviation between the calculation result and the average value of the historical data is greater than the deviation threshold, it is considered that the calculation result is abnormal. According to the results of the abnormal verification, the tensor calculation results are divided into the verified calculation result set and the error result set. The verified calculation results are included in the subsequent analysis and decision-making. For the calculation results with abnormalities, It is included in the error result set for further tracing and processing; through exception verification and result set division, problems and risks in the calculation process can be discovered in a timely manner, the calculation results can be classified and managed, and the efficiency and accuracy of data processing can be improved. At the same time, the generation of error result sets can provide data support for subsequent tracing and updating, which helps to continuously optimize the dependency graph and calculation model. The division of the verified calculation result set and the error result set is the basis for subsequent tracing and updating of the dependency graph. The error result set will be used to trace the basic dependency edges and forced dependency edges of the dependency graph, and update its weight according to the type of error. The verified calculation result set will be used to evaluate the execution effect of the calculation results and adjust the attenuation coefficient of the dependency graph.

[0136] Based on the error result set, the basic dependency edges and mandatory dependency edges of the dependency graph are traced back, and their weights are updated according to the type of error. This can reflect the deviation between the calculation results and the actual business situation, adjust the structure and parameters of the dependency graph in a timely manner, and improve the accuracy and reliability of the dependency graph; based on the abnormal calculation results in the error result set, the basic dependency edges and mandatory dependency edges in the dependency graph are traced back. For a document with an abnormal calculation result, the dependency relationship between it and other documents is checked to determine the relevant basic dependency edges and mandatory dependency edges; based on the type and degree of the error, the weights of the traced basic dependency edges and mandatory dependency edges are updated. If the calculation error is caused by an inaccurate association relationship of a certain basic dependency edge, , then the weight of the basic dependency edge is appropriately adjusted. If the calculation error is caused by the failure to meet the constraint conditions of the forced dependency edge, the weight of the forced dependency edge is adjusted accordingly; by updating the dependency edge weight, the inaccuracy and unreasonableness in the dependency graph can be corrected in a timely manner, and the accuracy and reliability of the dependency graph can be improved. At the same time, the updated dependency graph can better reflect the actual business situation and provide more accurate guidance for subsequent calculations and analysis. The update of the dependency edge weight will directly affect the structure and parameters of the dependency graph, and then affect the subsequent calculation path generation and task scheduling. The updated dependency graph can provide more accurate dependencies and constraints, which is helpful to generate more optimized calculation paths and task queues.

[0137] Adjusting the attenuation coefficient of the dependency graph according to the execution effect of the calculation result set can reflect the effectiveness and adaptability of the dependency graph in practical applications. A reasonable attenuation coefficient can balance the stability and dynamics of the dependency graph, ensuring that the dependency graph can promptly reflect changes in the business environment. Based on the verified calculation result set, evaluate the execution effect of the calculation results, including the accuracy and timeliness of the calculation results. For example, by comparing the differences between the calculation results and the actual business data, evaluate the accuracy of the calculation results, and by checking the processing time of the calculation results, evaluate the timeliness of the calculation results. Adjust the attenuation coefficient of the dependency graph according to the evaluation results of the execution effect. If the execution effect of the calculation result is good, it means that the structure and parameters of the dependency graph are relatively reasonable. Appropriately reduce the attenuation coefficient to maintain the stability of the dependency graph. If the execution effect of the calculation result is poor, it means that the dependency graph needs to be adjusted. Adjust the decay coefficient appropriately to speed up the update of the dependency graph, and output the updated dependency graph according to the adjusted decay coefficient and the updated dependency edge weight; by adjusting the decay coefficient of the dependency graph, the dependency graph can better adapt to changes in the business environment, improve the effectiveness and adaptability of the dependency graph, and the updated dependency graph can provide more accurate dependencies and constraints for subsequent calculation path generation, task scheduling, and collaborative calculations, thereby improving the computing efficiency and accuracy of the entire system. The updated dependency graph will serve as the basis for subsequent calculation path generation and task scheduling, providing more accurate dependencies and constraints for the path generation module, helping to generate more optimized calculation paths and task queues, and also providing a more accurate dependency graph for the impact propagation module, enabling it to more accurately analyze the impact of changes in basic information and document information on the dependency graph.

[0138] Example 2

[0139] like Figure 4 As shown, a flowchart of a method for linking calculation of freight trial calculation data based on multi-level document dependency is provided for an embodiment of the present application. The method includes:

[0140] Take the document information as a node, attach a set of dimension attributes to each node, and generate a document feature vector;

[0141] Generate hierarchical dependencies between nodes based on the dimension attribute set, historical association frequency, and business rules of each node, and determine the dependency graph based on the domain knowledge graph and the hierarchical dependencies between nodes;

[0142] Monitor changes in basic information and document information in real time, and analyze the impact of changes in basic information and document information on the dependency graph;

[0143] Combine the document feature vector to generate a task queue with priority tags and computing resources;

[0144] Taking computation time, resource usage, and computation accuracy as constraints, deep reinforcement learning is used to generate a set of candidate paths based on the task queue and dependency graph, and the computation path is selected from the set of candidate paths.

[0145] Divide the task into multiple parallel subtasks according to the calculation path. Each parallel subtask includes a basic quantity and a floating quantity. The basic quantity and the floating quantity are combined into a tensor to obtain the tensor calculation result.

[0146] The tensor calculation results are verified for consistency to obtain the verified calculation result set and error result set, and the dependency graph and candidate path set are updated based on the execution effect of the calculation result set and the error result set.

[0147] Since the principle of solving the problem by the method in the embodiment of the present application is similar to that of the system described above in the embodiment of the present application, the implementation of the method refers to the implementation of the system, and the repeated parts will not be repeated.

Claims

1. A multi-level document-dependent freight calculation data linkage calculation system, characterized by: include: Dependency determination module, influence propagation module, path generation module and collaborative calculation module; The dependency determination module is used to treat document information as nodes, attach a set of dimension attributes to each node. Each node's dimension attribute set includes time information and geographic identification, and generates a document feature vector. The document feature vector includes document type, amount, and time constraint. Based on each node's dimension attribute set, historical association frequency, and business rules, hierarchical dependencies between nodes are generated. The dependency graph is determined based on the domain knowledge graph and the hierarchical dependencies between nodes. The impact propagation module monitors changes to basic and document information in real time, analyzes the impact of these changes on the dependency graph, and generates a task queue with priority tags and computing resources based on document feature vectors. Priority tags include business urgency, impact amount, and time sensitivity. The path generation module uses computation time, resource usage, and computation accuracy as constraints, generates a set of candidate paths based on the task queue and dependency graph through deep reinforcement learning, and selects the computation path from the set of candidate paths. The collaborative computing module is used to divide the task into multiple parallel subtasks according to the calculation path. Each parallel subtask includes a basic quantity and a floating quantity. The basic quantity and the floating quantity are subjected to tensor synthesis operations to obtain tensor calculation results, and the tensor calculation results are verified for consistency to obtain a verified calculation result set and an error result set. The dependency graph and candidate path set are updated based on the execution effect of the calculation result set and the error result set.

2. The multi-level document-dependent freight trial calculation data linkage calculation system according to claim 1, characterized in that: The determination logic of the dependency graph includes: Generate basic dependency edges according to the association weight matrix in the hierarchical dependency relationship between nodes, and label the hierarchical attributes of the basic dependency edges according to the node levels in the hierarchical dependency relationship between nodes to output an initial heterogeneous graph; Combined with the domain knowledge graph, business rules are decomposed into a graph structure and marked as mandatory dependency edges to generate a set of rule constraints; The nodes in the initial heterogeneous graph are traversed to check whether they meet the rule constraint set, so as to update the initial heterogeneous graph, and the dependency graph is determined based on the updated initial heterogeneous graph.

3. The multi-level document-dependent freight trial calculation data linkage calculation system according to claim 2, characterized in that: The generation sub-logic of the hierarchical dependency relationship includes: Extract the dimension attribute set of each node and count the number of linkages between nodes and business rules; The historical association frequency is determined based on the number of linkages between nodes, the rule weight is determined based on the business rules, and the spatiotemporal coupling degree is determined by calculating the cosine similarity between nodes. The historical association frequency, rule weight, and spatiotemporal coupling degree are weighted and fused to form an association weight matrix. The association weight matrix is ​​divided into multiple node levels through the fuzzy clustering algorithm. The membership of the node level is determined according to the importance of the node. The association weight matrix is ​​updated according to the time difference between nodes. The hierarchical dependency relationship between nodes is generated by combining the node level and the association weight matrix.

4. The multi-level document-dependent freight trial calculation data linkage calculation system according to claim 3, characterized in that: The additional sub-logic of the dimension attribute set includes: Obtain business system data and IoT data, standardize them and form a raw data set, which includes timestamps, geographic coordinates, document types, amounts, and time constraints. Extract the timestamps from the original data set and slice them according to the transportation cycle to form multiple time windows. Within each time window, combine the historical time deviation to output the time fluctuation probability. Combine the time window code and the time fluctuation probability to form time information. Extract the geographic coordinates from the original dataset, convert them into H3 hierarchical grids, and combine traffic API data to annotate each grid with a travel time weight and a transportation cost coefficient to form a geographic identifier. The geographic identifier includes the grid code, travel time weight, and transportation cost coefficient. The time information and geographic identification are concatenated to form a dimension attribute set, and the dimension attribute set is attached to each document information.

5. The multi-level document-dependent freight trial calculation data linkage calculation system according to claim 4, characterized in that: The analysis logic of the impact degree of the dependency graph includes: Real-time monitoring of changes to basic information and document information. Analyzing the types of change events through natural language processing methods, including timeliness changes, geographic changes, and rule changes, extracting feature vectors of change events and comprehensively outputting change event features, including the type and scope of impact. Dynamically traverse the nodes, basic dependency edges, and mandatory dependency edges of the dependency graph according to the characteristics of the change event to locate the initial impact node; The graph diffusion algorithm is used to determine the influence scores of nodes at different times based on the confidence of basic dependency edges and forced dependency edges and the node hierarchy, and the influence score matrix of the dependency graph is output; The influence levels are divided according to the influence score matrix of the dependency graph, and the influence levels are superimposed on each grid.

6. The multi-level document-dependent freight trial calculation data linkage calculation system according to claim 5, characterized in that: The generation logic of the task queue includes: Assign business urgency based on the impact level, dynamically adjust the business urgency based on the document type in the document feature vector, determine the impact amount by combining the node's impact score at different times and the amount in the document feature vector, and determine the timeliness sensitivity based on the timeliness constraints in the document feature vector to output a priority tag. The priority score is determined by weighted summing of business urgency, impacted amount, and time sensitivity, and the weights of business urgency, impacted amount, and time sensitivity are dynamically adjusted based on historical processing records. The level of the task queue is determined based on the priority score, computing resources are pre-allocated according to the level of the task queue, and a task queue with priority tags and computing resources is generated.

7. The multi-level document-dependent freight trial calculation data linkage calculation system according to claim 6, characterized in that: The screening logic of the calculation path includes: Determine the state space and action space of deep reinforcement learning based on task queues with priority tags and computing resources and dependency graphs; Computational time, resource usage, and computational accuracy are used as constraints. The reward function of deep reinforcement learning is determined based on the computational time ratio, resource utilization, computational accuracy score, and number of violations, and a set of candidate paths is generated through continuous iteration. Based on the grids and time windows in the dependency graph, the candidate path set is spatially and temporally pruned to output an optimized candidate path set. Spatial pruning involves merging tasks within the same grid into grid task groups and executing them in parallel. Temporal pruning involves adjusting tasks with overlapping time windows and low timeliness sensitivity to batch processing. Monitor the system status in real time, extract the resource efficiency of each path in the system status, determine the time compliance rate of each path in the optimized candidate path set based on the time constraint, and determine the computational cost of each path. The comprehensive score of each path is obtained to screen out the computational path. The computational path includes the execution order of the task queue, the resource allocation plan, and the configuration of the computational accuracy.

8. The multi-level document-dependent freight trial calculation data linkage calculation system according to claim 7, characterized in that: The logic for obtaining the tensor calculation result includes: Divide the task into multiple parallel subtasks according to the computational path. Each parallel subtask includes a base quantity and a floating quantity. Through grid encoding and time window encoding, the base quantity and floating quantity are aligned to the same spatiotemporal coordinate system, and a normalized tensor pair is output. The document type mapping based on the dimension attribute set of each node and the document feature vector is used as the basic quantity, and the amount and time constraints based on the dimension attribute set of each node and the document feature vector are mapped as floating quantities. Extract the basis quantity from the standardized tensor pair, process the basis quantity through spatiotemporal convolution operation, update the processed basis quantity according to business rules, obtain the constrained basis quantity, and obtain the tensor calculation result after fusing the processed basis quantity, constrained basis quantity and floating quantity.

9. The multi-level document-dependent freight trial calculation data linkage calculation system according to claim 8, characterized in that: The update logic of the dependency graph includes: Extract the spatiotemporal dimensions of the tensor calculation results to see if they match the time window and grid in the document feature vector, and verify whether the tensor calculation results comply with the rule constraint set through the domain knowledge graph; Compare the tensor calculation results with the average deviation of historical data to verify whether there are any anomalies, and obtain the verified calculation result set and error result set; Based on the error result set, the basic dependency edges and mandatory dependency edges of the dependency graph are traced back, and the weights of the mandatory dependency edges and basic dependency edges are updated according to the type of error; Adjust the attenuation coefficient of the dependency graph according to the execution effect of the calculation result set, and output the updated dependency graph.

10. A method for linking freight calculation data based on multi-level document dependencies, characterized in that: include: Take the document information as a node, attach a set of dimension attributes to each node, and generate a document feature vector; Generate hierarchical dependencies between nodes based on the dimension attribute set, historical association frequency, and business rules of each node, and determine the dependency graph based on the domain knowledge graph and the hierarchical dependencies between nodes; Monitor changes in basic information and document information in real time, and analyze the impact of changes in basic information and document information on the dependency graph; Combine the document feature vector to generate a task queue with priority tags and computing resources; Taking computation time, resource usage, and computation accuracy as constraints, deep reinforcement learning is used to generate a set of candidate paths based on the task queue and dependency graph, and the computation path is selected from the set of candidate paths. Divide the task into multiple parallel subtasks according to the calculation path. Each parallel subtask includes a basic quantity and a floating quantity. The basic quantity and the floating quantity are combined into a tensor to obtain the tensor calculation result. The tensor calculation results are verified for consistency to obtain the verified calculation result set and error result set, and the dependency graph and candidate path set are updated based on the execution effect of the calculation result set and the error result set.

Images