A flight loading optimization method based on fusion of knowledge graph and dynamic scheduling algorithm

Through the combination of knowledge graph and dynamic scheduling algorithm, air logistics distribution is optimized, efficiency and adaptability are improved, and the problem of low loading efficiency in the existing technology is solved.

CN118798763BActive Publication Date: 2024-12-03CIVIL AVIATION LOGISTICS TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411289438.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-14
Publication Date
2024-12-03
Estimated Expiration
2044-09-14

AI Technical Summary

Technical Problem

The existing aviation logistics distribution methods are inefficient, difficult to adapt to the dynamic distribution needs under complex conditions, and cannot effectively take into account transportation efficiency and flight safety.

Method used

Using flight loading optimization method based on knowledge graph and dynamic scheduling algorithm, triple data extraction and inference completion are carried out by obtaining flight data and aviation logistics standards, neo4j graph database is constructed, graph neural network is called to train the knowledge graph model, and Combined with dynamic scheduling algorithms such as ant colony algorithm, we will determine the optimal flight loading plan.

Benefits of technology

It improves loading efficiency and adaptability, can better adapt to transportation modes under different models and multiple restrictions, and solves the shortcomings of traditional loading methods under complex conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118798763B_ABST
    Figure CN118798763B_ABST
Patent Text Reader

Abstract

The present invention provides a flight loading optimization method based on the fusion of knowledge graph and dynamic scheduling algorithm, including: obtaining flight data and current standards of the aviation logistics industry, extracting based on the flight data and current standards of the aviation logistics industry to obtain existing triple data; reasoning and completing the existing triple data based on the reasoning and completion algorithm to obtain reasoning triple data; constructing a neo4j graph database, embedding the existing triple data and the reasoning triple data into the neo4j graph database, calling a graph neural network to train the data constraint relationship of the knowledge graph to obtain a trained knowledge graph model; obtaining external input data and a transportation mode, selecting a corresponding dynamic scheduling algorithm according to the transportation mode, processing the external input data based on the knowledge graph model and the corresponding dynamic scheduling algorithm to obtain an optimal flight loading plan; determining the loading and unloading sequence of the loading in the optimal flight loading plan based on a preset ant colony algorithm.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of aviation logistics technology, and in particular to a flight loading optimization method based on the fusion of knowledge graph and dynamic scheduling algorithm. Background Art

[0002] Air logistics refers to the logistics activities that use air transport to carry out tasks such as cargo transportation, warehousing, and distribution. As one of the modern logistics transportation methods, air logistics is widely used in trade transportation, e-commerce, emergency freight and other fields due to its fast, efficient and global coverage, and plays an important role in the global supply chain. The loading problem is a key link in air logistics - whether it is the passenger baggage and cargo that can be checked in the belly of a passenger plane, or the various types of cargo that need to be carried by a full-cargo aircraft, specific consideration must be given to how to reasonably arrange and allocate the cargo within a limited space in order to balance transportation efficiency and flight safety. This is even more valuable to study at a time when the e-commerce industry is increasingly prosperous, logistics demand is increasing year by year, and the scale of transportation is growing day by day.

[0003] The loading problem involves multiple factors such as the volume, shape, weight, and compatibility of the cargo. It is restricted by physical conditions such as aircraft structure, cabin design, load-bearing restrictions, and transportation methods. It is also related to external factors such as weather conditions, flight schedules, drivers and crew, and even the length of the runway. It is a multi-constraint, multi-condition, and multi-objective optimization problem. It also places high demands on the real-time processing and dynamic adaptation capabilities of loading decisions. However, the loading method is a manual loading mode, and the loading and dispatching are mostly carried out by loaders based on experience and rule models. It has certain applicability in loading tasks with smaller scale and simple scenarios, but its relatively low loading efficiency and non-objective loading mode cannot adapt well to loading problems under complex conditions. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a flight loading optimization method based on the fusion of knowledge graph and dynamic scheduling algorithm, which solves the above-mentioned problems in the prior art.

[0005] According to an embodiment of the present invention, a flight loading optimization method based on the fusion of knowledge graph and dynamic scheduling algorithm includes:

[0006] Obtain flight data and current standards of the aviation logistics industry, extract existing triplet data based on the flight data and current standards of the aviation logistics industry;

[0007] Based on the inference completion algorithm, the existing triple data is inferred and completed to obtain the inference triple data;

[0008] Build a neo4j graph database, embed the existing triple data and inference triple data into the neo4j graph database, call the graph neural network to train the data constraint relationship of the knowledge graph, and obtain the trained knowledge graph model;

[0009] Obtain external input data and transportation mode, select the corresponding dynamic scheduling algorithm according to the transportation mode, process the external input data based on the knowledge graph model and the corresponding dynamic scheduling algorithm, and obtain the optimal flight loading plan;

[0010] The loading and unloading sequence of the optimal flight loading plan is determined based on the preset ant colony algorithm.

[0011] As an embodiment of the present invention, the existing triple data is inferred and completed based on the inference completion algorithm to obtain inference triple data, including:

[0012] Obtain existing triple data and input them into the initial TransR model; wherein the existing triple data includes: relations and entities;

[0013] The relations and entities in the existing triple data are embedded and initialized as random vectors to obtain the entity set and relationship set , based on the relationship Determine all corresponding triplets, obtain the training triplets set, and initialize each relationship Specific relations projected into the projected space ;in, , , , ;

[0014] According to a specific relationship Calculate the triples in the training triple set Head entity and tail entity The value projected into the relational space is calculated as follows:

[0015]

[0016]

[0017] Calculate the score of each triple in the relational space in the training triple set respectively, and the calculation formula is as follows:

[0018]

[0019] in, express norm;

[0020] The parameters of the initial TransR model are optimized through the objective function, and the calculation formula of the objective function is as follows:

[0021]

[0022] in, is a hyperparameter, S For all positive triple sets in the training triple set, is the set of negative triples in the training triple set;

[0023] Use stochastic gradient descent or Adam optimization algorithm to minimize the objective function and update entity embeddings, relation embeddings, and training triple sets;

[0024] Repeat the above steps until the number of iterations is greater than the preset number of iterations or other convergence conditions are met, and a trained TransR model is obtained;

[0025] Select the head entity from the existing triple data and relationship Randomly generate multiple candidate tail entities , and obtain multiple candidate triplets;

[0026] Call the trained TransR model to calculate the energy function value of multiple candidate triples, and select the candidate tail entity whose energy value is less than the preset value The corresponding candidate triples are used as inference triple data.

[0027] As an embodiment of the present invention, a neo4j graph database is constructed, existing triple data and inference triple data are embedded in the neo4j graph database, and a graph neural network is called to train data constraint relationships based on a knowledge graph, including:

[0028] Acquire existing triple data and inferred triple data to construct a graph structure; wherein entities and relationships in the existing triple data and inferred triple data are used as nodes of the graph structure, and triples formed by the existing triple data and inferred triple data are used as edges of the graph structure;

[0029] Build a neo4j graph database to store existing triple data and inference triple data, initialize the GCN graph neural network model and integrate the data constraint relationship of the preset function training knowledge graph;

[0030] The parameters of the knowledge graph model are updated through the stochastic gradient descent algorithm, the gradient is calculated through the back propagation algorithm, and the parameters are updated until the loss of the knowledge graph model tends to stabilize, completing the training of the knowledge graph model.

[0031] As an embodiment of the present invention, the transport mode includes: belly cargo transport of passenger aircraft and cargo aircraft transport;

[0032] The dynamic scheduling algorithm includes: a first preset integrated dynamic scheduling algorithm and a second preset integrated dynamic scheduling algorithm;

[0033] Select the corresponding dynamic scheduling algorithm according to the transportation mode, process the external input data based on the knowledge graph model and the corresponding dynamic scheduling algorithm, and obtain the optimal flight loading plan, including:

[0034] If the transport mode is belly cargo transport, select the first preset integrated dynamic scheduling algorithm, and process the external input data based on the knowledge graph model and the first preset integrated dynamic scheduling algorithm to obtain the optimal flight loading plan; wherein the first preset integrated dynamic scheduling algorithm includes a taboo algorithm and a greedy algorithm;

[0035] If the transportation mode is cargo plane transportation, the second preset integrated dynamic scheduling algorithm is selected, and the external input data is processed based on the knowledge graph model and the second preset integrated dynamic scheduling algorithm to obtain the optimal flight loading plan; wherein the first preset integrated dynamic scheduling algorithm includes genetic algorithm and simulated annealing algorithm.

[0036] As an embodiment of the present invention, external input data is processed based on the knowledge graph model and the first preset integrated dynamic scheduling algorithm to obtain an optimal flight loading plan, including:

[0037] Get external input data and initialize the tabu table , set the maximum number of iterations I and the length of the taboo list , and build a fitness function based on the knowledge graph model; the external input data includes the belly tank capacity set ;

[0038] According to the initialization taboo table Generate several initial solutions , several initial solutions Substitute into the fitness function for calculation and obtain several fitnesses. The fitness calculation formula is as follows

[0039]

[0040]

[0041] in, For fitness, is the objective function, is the penalty coefficient, is a constraint;

[0042] Traverse several initial solutions Corresponding to several fitness values, select the initial solution whose fitness value is greater than the preset fitness value , and obtain the candidate solution set ;

[0043] Determine the candidate solution set Is there an optimal solution that meets the preset screening conditions? , the preset filtering conditions are as follows:

[0044]

[0045] If there is no optimal solution , then re-according to the taboo table Generate several initial solutions Make a judgment; if there is an optimal solution , determine whether the maximum number of iterations is met I or preset termination conditions;

[0046] If not satisfied, re-check according to the taboo table Generate several initial solutions Make a judgment; if satisfied, based on the optimal solution Determine belly capacity set W Is there an element in the unfilled state? If so, call the greedy algorithm to update the optimal solution Until all elements in are fully loaded, the fully loaded optimal solution is obtained , output full load optimal solution is the optimal flight loading solution; if it does not exist, the greedy algorithm is not called and the optimal solution is directly output Optimal flight loading plan.

[0047] As an embodiment of the present invention, determine the candidate solution set Is there an optimal solution that meets the preset screening conditions? When the current taboo table length is greater than the taboo table length ; If it is greater than, the length of the taboo table Carry out capacity expansion.

[0048] As an embodiment of the present invention, a second preset integrated dynamic scheduling algorithm is selected, and external input data is processed based on the knowledge graph model and the second preset integrated dynamic scheduling algorithm to obtain an optimal flight loading plan, including:

[0049] Initialize the population of the genetic algorithm, define the fitness function based on the data constraint relationship of the knowledge graph, and set the initial temperature of the simulated annealing , cooling rate and termination conditions;

[0050] Randomly generate the initial population , each individual in the initial population is a set of solutions;

[0051] Determine whether the initial population meets the evolution termination conditions; if not, continue to evolve the initial population;

[0052] If satisfied, call the fitness function to evaluate the pros and cons of each individual and obtain the evaluation score of each individual;

[0053] Traverse the evaluation scores of each individual, select individuals with evaluation scores greater than the preset score, and obtain the evolution set; if there is no individual with an evaluation score greater than the preset score, generate a new population according to the preset rules, and repeat the above steps until the evolution set is obtained;

[0054] For individual units in the evolutionary set Call the simulated annealing algorithm to perform simulated annealing optimization; Try new solutions in the field , define the energy function and calculate the energy difference between the two , accept this new solution as the current solution with a preset probability, and the energy difference is calculated as follows:

[0055]

[0056] Update the current solution and lower the temperature, repeat the above steps until the maximum number of iterations is reached or the evaluation score of the fitness function no longer increases significantly, output the individual with the highest evaluation score as the final solution, and obtain the optimal flight loading plan.

[0057] Compared with the prior art, the present invention has the following beneficial effects: by integrating knowledge graphs and dynamic scheduling algorithms to adapt to the transportation modes of different types of aircraft and the loading optimization problems under various constraints; it solves the low loading efficiency and subjective loading modes of traditional loading methods that cannot adapt well to the loading problems under complex conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is an overall architecture diagram of an embodiment of the present invention;

[0059] Figure 2 A pattern diagram designed for constructing a knowledge graph model according to another embodiment of the present invention;

[0060] Figure 3 FIG. 4 is a flow chart of another embodiment of the present invention. DETAILED DESCRIPTION

[0061] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0062] like Figures 1 to 3As shown, the embodiment of the present invention proposes a flight loading optimization method based on the fusion of knowledge graph and dynamic scheduling algorithm, including:

[0063] S1: Obtain flight data and current standards of the aviation logistics industry, extract existing triplet data based on the flight data and current standards of the aviation logistics industry;

[0064] S2: Based on the inference completion algorithm, the existing triple data is inferred and completed to obtain the inference triple data;

[0065] S3: Build a neo4j graph database, embed the existing triple data and inference triple data into the neo4j graph database, call the graph neural network to train the data constraint relationship of the knowledge graph, and obtain the trained knowledge graph model;

[0066] S4: Obtain external input data and transportation mode, select the corresponding dynamic scheduling algorithm according to the transportation mode, process the external input data based on the knowledge graph model and the corresponding dynamic scheduling algorithm, and obtain the optimal flight loading plan;

[0067] S5: Determine the loading and unloading sequence of the optimal flight loading plan based on the preset ant colony algorithm.

[0068] The working principle of the above technical solution: in actual use, the triple data is extracted by obtaining flight data and the current standards of the aviation logistics industry, wherein the aviation data is the aircraft model information of each airline and the parameter information covering its center of gravity envelope range, load-bearing limit, cabin distribution and various other physical constraints; the current standards of the aviation logistics industry are to organize the various parameter information of various transport goods (containers) according to the "Dangerous Goods Regulations" formulated by the International Air Transport Association and other logistics industry standards; the existing triple data is extracted based on the above information, for example, (airline, management, aircraft); (airline, formulation, flight plan); (airline, regulations, passenger weight conversion standard); (aircraft, design, model structure); (aircraft, design, model attributes); (aircraft, allowed, center of gravity envelope range); (aircraft, allowed, load range); (aircraft, attachment, cabin situation); (cabin, design, distribution area); (cabin, allowed, unit load-bearing); (cabin, design, physical constraint); (cabin, allowed, cargo type); (cargo, belongs to, cargo type); (cargo, has, value attribute) ; (cargo, design, shape, volume, weight attributes); (cargo, whether, boxed transportation); (cargo, association, other cargo); (climate, impact, aircraft balance state); (aircraft plan, constraints, direct / connecting / transfer); (flight plan, assignment, crew information), etc.; then call the reasoning completion algorithm to complete the existing triple data to obtain the reasoning triple data; then build a neo4j graph database, embed the existing triple data and reasoning triple data into the neo4j graph database, call the graph neural network to train the data constraint relationship of the knowledge graph, and obtain the trained knowledge graph model; then input external data and transportation mode, select the corresponding dynamic scheduling algorithm according to the transportation mode, process the external input data based on the knowledge graph model and the corresponding dynamic scheduling algorithm, and obtain the optimal flight loading plan; among them, the external input data is generally input by the loaders used, including local climate conditions, flight plan, crew information and aircraft cargo hold capacity, etc.; after obtaining the optimal flight loading plan, by providing an additional ant colony algorithm, the loaders are allowed to plan the loading and unloading sequence of baggage and cargo according to the loading plan.

[0069] The beneficial effects of the above technical solution: Through the above technical solution, by integrating knowledge graphs and dynamic scheduling algorithms to adapt to the transportation modes of different types of aircraft and the loading optimization problems under various constraints; the low loading efficiency and subjective loading modes of traditional loading methods that cannot adapt well to loading problems under complex conditions are solved.

[0070] In one embodiment, the existing triple data is inferred and completed based on the inference completion algorithm to obtain inference triple data, including:

[0071] S21: Obtain existing triple data and input it into the initial TransR model; wherein the existing triple data includes: relations and entities;

[0072] S22: Initialize the relationships and entities in the existing triple data into random vectors to obtain the entity set and relationship set , based on the relationship Determine all corresponding triplets, obtain the training triplets set, and initialize each relationship Specific relations projected into the projected space ;in, , , , ;

[0073] S23: Based on a specific relationship Calculate the triples in the training triple set Head entity and tail entity The value projected into the relational space is calculated as follows:

[0074]

[0075]

[0076] Calculate the score of each triple in the relational space in the training triple set respectively, and the calculation formula is as follows:

[0077]

[0078] in, express norm;

[0079] S24: Optimize the parameters of the initial TransR model through the objective function. The calculation formula of the objective function is as follows:

[0080]

[0081] in, is a hyperparameter, S For all positive triple sets in the training triple set, is the set of negative triples in the training triple set;

[0082] S25: Use stochastic gradient descent or Adam optimization algorithm to minimize the objective function and update entity embeddings, relation embeddings, and training triple sets;

[0083] S26: Repeat the above steps until the number of iterations is greater than the preset number of iterations or other convergence conditions are met, and a trained TransR model is obtained;

[0084] S27: Select the head entity in the existing triple data and relationship Randomly generate multiple candidate tail entities , and obtain multiple candidate triplets;

[0085] S28: Call the trained TransR model to calculate the energy function value of multiple candidate triples, and select the candidate tail entity whose energy value is less than the preset value The corresponding candidate triples are used as inference triple data;

[0086] The working principle of the above technical solution is: the existing triple data is inferred and completed by the inference completion algorithm. The inference completion algorithm generally adopts either the TransR model or the TransE model. When the TransR model is adopted, the existing triple data is first obtained and input into the initial TransR model; then the relations and entities in the existing triple data are respectively embedded and initialized as random vectors to obtain the entity set and relationship set , based on the relationship Determine all corresponding triplets, obtain the training triplets set, and initialize each relationship Specific relations projected into the projected space ; Then according to the specific relationship Calculate the triples in the training triple set Head entity and tail entity The value projected to the relational space; then, the score of each triple in the training triplet in the relational space is calculated according to the calculated value of the relational space; then, the parameters of the initial TransR model are optimized based on the score in the relational space through the objective function, and the calculation formula of the objective function is as follows:

[0087]

[0088] in, is a hyperparameter, S For all the positive triples in the training triple set, The negative triple set in the training triple set; use the stochastic gradient descent or Adam optimization algorithm to minimize the objective function, and update the entity embedding, relationship embedding and training triple set; repeat the above steps until the number of iterations is greater than the preset number of iterations or the convergence condition, and the trained TransR model is obtained; for the head entity in the existing triple data and relationship Randomly generate multiple candidate tail entities , and obtain multiple candidate triplets; call the trained TransR model to calculate the energy function value of multiple candidate triplets, and select the candidate tail entity whose energy value is less than the preset value The corresponding candidate triples are used as inference triple data; when the TransE model is used for processing, except for the processing step S24, the rest is the same as the TransR model processing, and the objective function in the processing step S24 is replaced by the loss function, and the loss function formula is as follows:

[0089]

[0090] in, is the set of positively sampled triplets, is a set of negative sampling triplets, is the marginal hyperparameter, is a distance function (e.g. or norm), ;

[0091] Among them, the training triple set is based on the relationship Positive sampling and negative sampling are performed respectively, that is, the training triplet set includes a positive triplet set and a negative triplet set;

[0092] Beneficial effects of the above technical solution: Through the above technical solution, by standardizing the extraction of current standards in the aviation logistics industry, the knowledge graph can actively learn existing knowledge and infer and supplement possible implicit rules by itself. The formed knowledge base can better serve as the "population selection" or "taboo rule" constraints of subsequent dynamic scheduling algorithms, thereby optimizing the algorithm processing flow.

[0093] In one embodiment, a neo4j graph database is constructed, existing triple data and inference triple data are embedded in the neo4j graph database, and a graph neural network is called to train data constraint relationships based on a knowledge graph, including:

[0094] S31: Acquire existing triple data and inferred triple data to construct a graph structure; wherein entities and relationships in the existing triple data and inferred triple data are used as nodes of the graph structure, and triples formed by the existing triple data and inferred triple data are used as edges of the graph structure;

[0095] S32: Build a neo4j graph database to store existing triple data and inference triple data, initialize the GCN graph neural network model and integrate the data constraint relationship of the preset function training knowledge graph;

[0096] S33: Update the parameters of the knowledge graph model through the stochastic gradient descent algorithm, calculate the gradient through the back propagation algorithm, and update the parameters until the loss of the knowledge graph model tends to be stable, thus completing the training of the knowledge graph model;

[0097] The working principle and beneficial effects of the above technical solution: Collect and organize the existing triple data and inference triple data of the knowledge graph;

[0098] Construct a graph structure with the existing triple data and inference triple data of the knowledge graph; wherein entities and relationships are used as nodes of the graph, and triples are used as edges of the graph; create a new neo4j graph database to store the above triple data; initialize the GCN graph neural network model and integrate the preset function to train the triple data of the knowledge graph; wherein the preset function includes:

[0099] Message passing aggregation function, the functional relationship is as follows:

[0100]

[0101] in, Representation Node In the The aggregation result of the layer, Representation Node The set of neighbor nodes of represents the normalization constant;

[0102] Neighbor feature aggregation function, the functional relationship is as follows:

[0103]

[0104] in, Representation Node In the The feature representation of the layer, Indicates The weight matrix of the layer, represents the activation function (such as );

[0105] Update function, the functional relationship is as follows:

[0106]

[0107] in, Representation Node In the The feature representation of the layer, Indicates The weight matrix of the layer, Representation Node The set of neighbor nodes of represents the normalization constant;

[0108] The above functions work together on each layer of GCN, and gradually and iteratively learn the representation of nodes in the graph through the process of information transmission and feature update; the message passing aggregation function is responsible for extracting information from neighbor nodes, the neighbor feature aggregation function is responsible for fusing the information of neighbor nodes with the features of the central node, and the update function is responsible for updating the feature representation of the central node; finally, the parameters of the model are updated using algorithms such as stochastic gradient descent, the gradient is calculated through the back propagation algorithm, and the parameters are updated until the loss of the knowledge graph model stabilizes.

[0109] In one embodiment, the transport modes include: belly cargo transport on passenger aircraft and cargo aircraft transport;

[0110] The dynamic scheduling algorithm includes: a first preset integrated dynamic scheduling algorithm and a second preset integrated dynamic scheduling algorithm;

[0111] Select the corresponding dynamic scheduling algorithm according to the transportation mode, process the external input data based on the knowledge graph model and the corresponding dynamic scheduling algorithm, and obtain the optimal flight loading plan, including:

[0112] S41: If the transport mode is belly cargo transport, select a first preset integrated dynamic scheduling algorithm, and process the external input data based on the knowledge graph model and the first preset integrated dynamic scheduling algorithm to obtain an optimal flight loading plan; wherein the first preset integrated dynamic scheduling algorithm includes a taboo algorithm and a greedy algorithm;

[0113] S42: If the transport mode is cargo aircraft transport, select a second preset integrated dynamic scheduling algorithm, and process the external input data based on the knowledge graph model and the second preset integrated dynamic scheduling algorithm to obtain an optimal flight loading plan; wherein the first preset integrated dynamic scheduling algorithm includes a genetic algorithm and a simulated annealing algorithm;

[0114] The working principle and beneficial effects of the above technical solution are as follows: By targeting different transportation modes, the knowledge graph model is integrated with different dynamic scheduling algorithms to determine the optimal loading plan for flights under different transportation models; for example, for passenger aircraft, the belly of the aircraft is mainly used to transport passengers' checked baggage and other cargo; in this case, passenger checked baggage should be the primary transportation consideration; this goal is subject to the flight cut-off time, which is generally 30 to 45 minutes, and the taboo search algorithm is selected to try to obtain the optimal solution under the constraints of the knowledge graph to meet the passenger checked baggage scheduling needs. For the remaining belly space, the greedy algorithm is additionally called in Load as much cargo as possible under the constraints of the knowledge graph; for all-cargo aircraft, there are no additional constraints when transporting cargo, and the allowed loading calculation time is longer; optimize the genetic algorithm to make it more suitable for the field of aviation logistics, and consider using the optimal solution space obtained after multiple rounds of evolution as the input of the simulated annealing algorithm for secondary tuning to ensure the superiority of the loading scheme; finally, according to the actual needs of passenger aircraft belly transportation, all-cargo aircraft transportation, and loading and unloading planning, try to combine the taboo search algorithm with the greedy algorithm, the improved genetic algorithm with the simulated annealing algorithm, and the ant colony algorithm to construct an algorithm model that can adapt to different situational modes.

[0115] In one embodiment, external input data is processed based on the knowledge graph model and the first preset integrated dynamic scheduling algorithm to obtain an optimal flight loading plan, including:

[0116] S411: Obtain external input data and initialize taboo table , set the maximum number of iterations I and the length of the taboo list , and build a fitness function based on the knowledge graph model; the external input data includes the belly tank capacity set ;

[0117] S412: Initialize the taboo table Generate several initial solutions , several initial solutions Substitute into the fitness function for calculation and obtain several fitnesses. The fitness calculation formula is as follows

[0118]

[0119]

[0120] in, For fitness, is the objective function, is the penalty coefficient, is a constraint;

[0121] S413: Traverse several initial solutions Corresponding to several fitness values, select the initial solution whose fitness value is greater than the preset fitness value , and obtain the candidate solution set ;

[0122] S414: Determine the candidate solution set Is there an optimal solution that meets the preset screening conditions? , the preset filtering conditions are as follows:

[0123]

[0124] If there is no optimal solution , then re-according to the taboo table Generate several initial solutions Make a judgment; if there is an optimal solution , determine whether the maximum number of iterations is met I or preset termination conditions;

[0125] If not satisfied, re-check according to the taboo table Generate several initial solutions Make a judgment; if satisfied, based on the optimal solution Determine belly capacity set W Is there an element in the unfilled state? If so, call the greedy algorithm to update the optimal solution Until all elements in are fully loaded, the fully loaded optimal solution is obtained , output full load optimal solution is the optimal flight loading solution; if it does not exist, the greedy algorithm is not called and the optimal solution is directly output Optimal flight loading plan;

[0126] At the same time, determine the candidate solution set Is there an optimal solution that meets the preset screening conditions? When the current taboo table length is greater than the taboo table length ; If it is greater than, the length of the taboo table Carry out capacity expansion.

[0127] In one embodiment, a second preset integrated dynamic scheduling algorithm is selected, and external input data is processed based on the knowledge graph model and the second preset integrated dynamic scheduling algorithm to obtain an optimal flight loading plan, including:

[0128] S421: Initialize the population of the genetic algorithm, define the fitness function based on the data constraint relationship of the knowledge graph, and set the initial temperature of the simulated annealing , cooling rate And the termination conditions; among them, when the genetic algorithm population is initialized, various basic data such as the evolution termination conditions of the genetic algorithm will be initialized;

[0129] S422: Randomly generate initial population , each individual in the initial population is a set of solutions;

[0130] S423: Determine whether the initial population meets the evolution termination condition; if not, continue to evolve the initial population;

[0131] S424: If satisfied, calling the fitness function to evaluate the pros and cons of each individual and obtaining the evaluation score of each individual;

[0132] S425: traverse the evaluation score of each individual, select individuals with evaluation scores greater than the preset score, and obtain an evolution set; if there is no individual with an evaluation score greater than the preset score, generate a new population according to the preset rules, and repeat the above steps S422-S424 until the evolution set is obtained;

[0133] S426: Evolution of individual units in the set Call the simulated annealing algorithm to perform simulated annealing optimization; Try new solutions in the field , define the energy function and calculate the energy difference between the two , accept this new solution as the current solution with a preset probability, and the energy difference is calculated as follows:

[0134]

[0135] S427: Update the current solution and lower the temperature, repeat the above step S426 until the maximum number of iterations is reached or the evaluation score of the fitness function no longer increases significantly, output the individual with the highest evaluation score as the final solution, and obtain the optimal flight loading plan.

[0136] In one embodiment, by embedding the knowledge graph model into the cloud device, the dynamic scheduling algorithm and the external input interface for inputting external input data are sunk to the edge setting. Through the edge-cloud collaborative processing mode, the processing advantage of "cloud-side overall control and edge-side adaptation" can be fully utilized, reducing the computing load of both cloud and edge ends while clarifying the responsibilities of each end, and greatly improving the efficiency of load planning with a small amount of network load; among them, edge-cloud collaboration is a computing architecture that combines the advantages of edge computing and cloud computing; it achieves resource optimization and efficient computing by processing some computing tasks at the edge node close to the data source, and handing over complex tasks or tasks that require a lot of computing resources to the cloud for processing. Edge collaborative operations can reduce the processing pressure of the cloud to a certain extent, and the direct intervention of the edge can also improve the response speed of the task and provide a redundant backup of the edge node. However, when working collaboratively, additional scheduling overhead and network resource overhead are inevitable.

[0137] Certain data transmission delay: There is a certain delay when data is transmitted between edge devices and the cloud, which can be described as:

[0138]

[0139] in, represents data transmission delay, D represents data volume, and B represents bandwidth;

[0140] Total delay of edge-cloud collaborative link. Compared with cloud computing or edge computing, the delay overhead required for edge-cloud collaborative operation is shown in the following formula:

[0141]

[0142] in, represents the total delay, Indicates the task Data transmission delay;

[0143] Resource utilization optimization. The resource utilization optimization brought about by the additional overhead can be described as:

[0144]

[0145] in, Indicates The resource utilization of edge nodes, Indicates the total number of edge nodes.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A flight loading optimization method based on the fusion of knowledge graph and dynamic scheduling algorithm, characterized in that: include: Obtain flight data and current standards of the aviation logistics industry, extract existing triplet data based on the flight data and current standards of the aviation logistics industry; Based on the inference completion algorithm, the existing triple data is inferred and completed to obtain the inference triple data; Build a neo4j graph database, embed the existing triple data and inference triple data into the neo4j graph database, call the graph neural network to train the data constraint relationship of the knowledge graph, and obtain the trained knowledge graph model; Obtain external input data and transportation mode, select the corresponding dynamic scheduling algorithm according to the transportation mode, process the external input data based on the knowledge graph model and the corresponding dynamic scheduling algorithm, and obtain the optimal flight loading plan; Determine the loading and unloading sequence of the optimal flight loading plan based on the preset ant colony algorithm; The modes of transport include: belly cargo transport and cargo aircraft transport; The dynamic scheduling algorithm includes: a first preset integrated dynamic scheduling algorithm and a second preset integrated dynamic scheduling algorithm; Select the corresponding dynamic scheduling algorithm according to the transportation mode, process the external input data based on the knowledge graph model and the corresponding dynamic scheduling algorithm, and obtain the optimal flight loading plan, including: If the transport mode is belly cargo transport, select the first preset integrated dynamic scheduling algorithm, and process the external input data based on the knowledge graph model and the first preset integrated dynamic scheduling algorithm to obtain the optimal flight loading plan; wherein the first preset integrated dynamic scheduling algorithm includes a taboo algorithm and a greedy algorithm; If the transport mode is cargo aircraft transport, the second preset integrated dynamic scheduling algorithm is selected, and the external input data is processed based on the knowledge graph model and the second preset integrated dynamic scheduling algorithm to obtain the optimal flight loading plan; wherein the first preset integrated dynamic scheduling algorithm includes a genetic algorithm and a simulated annealing algorithm; Based on the knowledge graph model and the first preset integrated dynamic scheduling algorithm, the external input data is processed to obtain the optimal flight loading plan, including: Get external input data and initialize the tabu table , set the maximum number of iterations I and the length of the taboo list , and build a fitness function based on the knowledge graph model; the external input data includes the belly tank capacity set ; According to the initialization taboo table Generate several initial solutions , several initial solutions Substitute into the fitness function for calculation and obtain several fitnesses. The fitness calculation formula is as follows in, For fitness, is the objective function, is the penalty coefficient, is a constraint; Traverse several initial solutions Corresponding to several fitness values, select the initial solution whose fitness value is greater than the preset fitness value , and obtain the candidate solution set ; Determine the candidate solution set Is there an optimal solution that meets the preset screening conditions? , the preset filtering conditions are as follows: If there is no optimal solution , then re-according to the taboo table Generate several initial solutions Make a judgment; if there is an optimal solution , determine whether the maximum number of iterations is met I or preset termination conditions; If not satisfied, re-check according to the taboo table Generate several initial solutions Make a judgment; if satisfied, based on the optimal solution Determine belly capacity set W Is there an element in the unfilled state? If so, call the greedy algorithm to update the optimal solution Until all elements in are fully loaded, the fully loaded optimal solution is obtained , output full load optimal solution is the optimal flight loading solution; if it does not exist, the greedy algorithm is not called and the optimal solution is directly output Optimal flight loading plan; Determine the candidate solution set Is there an optimal solution that meets the preset screening conditions? When the current taboo table length is greater than the taboo table length ; If it is greater than, the length of the taboo table Carry out capacity expansion; The second preset integrated dynamic scheduling algorithm is selected, and the external input data is processed based on the knowledge graph model and the second preset integrated dynamic scheduling algorithm to obtain the optimal flight loading plan, including: Initialize the population of the genetic algorithm, define the fitness function based on the data constraint relationship of the knowledge graph, and set the initial temperature of the simulated annealing , cooling rate and termination conditions; Randomly generate the initial population , each individual in the initial population is a set of solutions; Determine whether the initial population meets the evolution termination conditions; if not, continue to evolve the initial population; If satisfied, call the fitness function to evaluate the pros and cons of each individual and obtain the evaluation score of each individual; Traverse the evaluation scores of each individual, select individuals with evaluation scores greater than the preset score, and obtain the evolution set; if there is no individual with an evaluation score greater than the preset score, generate a new population according to the preset rules, and repeat the above steps until the evolution set is obtained; For individual units in the evolutionary set Call the simulated annealing algorithm to perform simulated annealing optimization; Try new solutions in the field , define the energy function and calculate the energy difference between the two , accept this new solution as the current solution with a preset probability, and the energy difference is calculated as follows: Update the current solution and lower the temperature, repeat the above steps until the maximum number of iterations is reached or the evaluation score of the fitness function no longer increases significantly, output the individual with the highest evaluation score as the final solution, and obtain the optimal flight loading plan.

2. The flight loading optimization method based on the fusion of knowledge graph and dynamic scheduling algorithm according to claim 1 is characterized in that: Based on the inference completion algorithm, the existing triple data is inferred and completed to obtain the inference triple data, including: Obtain existing triple data and input them into the initial TransR model; wherein the existing triple data includes: relations and entities; The relations and entities in the existing triple data are embedded and initialized as random vectors to obtain the entity set and relationship set , based on the relationship Determine all corresponding triplets, obtain the training triplets set, and initialize each relationship Specific relations projected into the projected space ;in, , , , ; According to a specific relationship Calculate the triples in the training triple set Head entity and tail entity The value projected into the relational space is calculated as follows: Calculate the score of each triple in the relational space in the training triple set respectively, and the calculation formula is as follows: in, express norm; The parameters of the initial TransR model are optimized through the objective function, and the calculation formula of the objective function is as follows: in, is a hyperparameter, S For all positive triple sets in the training triple set, is the set of negative triples in the training triple set; Use stochastic gradient descent or Adam optimization algorithm to minimize the objective function and update entity embeddings, relation embeddings, and training triple sets; Repeat the above steps until the number of iterations is greater than the preset number of iterations or other convergence conditions are met, and a trained TransR model is obtained; Select the head entity from the existing triple data and relationship Randomly generate multiple candidate tail entities , and obtain multiple candidate triplets; Call the trained TransR model to calculate the energy function value of multiple candidate triples, and select the candidate tail entity whose energy value is less than the preset value The corresponding candidate triples are used as inference triple data.

3. The flight loading optimization method based on the fusion of knowledge graph and dynamic scheduling algorithm according to claim 1 is characterized in that: Build a neo4j graph database, embed the existing triple data and inference triple data into the neo4j graph database, and call the graph neural network to train the data constraint relationship based on the knowledge graph, including: Acquire existing triple data and inferred triple data to construct a graph structure; wherein entities and relationships in the existing triple data and inferred triple data are used as nodes of the graph structure, and triples formed by the existing triple data and inferred triple data are used as edges of the graph structure; Build a neo4j graph database to store existing triple data and inference triple data, initialize the GCN graph neural network model and integrate the data constraint relationship of the preset function training knowledge graph; The parameters of the knowledge graph model are updated through the stochastic gradient descent algorithm, the gradient is calculated through the back propagation algorithm, and the parameters are updated until the loss of the knowledge graph model tends to stabilize, completing the training of the knowledge graph model.