Scene operation situation driving factor extraction and simulation deduction method

Abstract

The invention discloses a scene operation situation driving factor extraction and simulation deduction method, and belongs to the field of airport scene management and control, and the deduction method comprises the following specific steps: I, collecting each group of airport operation data, carrying out the preprocessing, and storing the preprocessed data in a central data warehouse; according to the method, the operation risk possibly caused by data leakage can be reduced, the generalization ability of the whole model is improved, the performance of localized deduction is improved, the dependence on high-performance hardware and the system cost are reduced, deduction errors caused by data quality differences or abnormal data are avoided, and the extraction accuracy of key driving factors is improved; the method can avoid falling into a local optimal solution, provides a more comprehensive optimization strategy for airport scene operation, can efficiently process a deduction problem under a complex constraint condition, helps an airport operation manager to pre-judge the problem in advance, and designs a coping strategy.

Description

Technical Field

[0001] The present invention relates to the field of airport surface control, and in particular to a method for extracting and simulating surface operation situation driving factors. Background Art

[0002] Airport surface operations are an important part of the aviation transportation system, and their operating efficiency directly affects flight punctuality, ground resource utilization, and overall passenger experience. However, the complexity of airport surface operations is increasing day by day, which is mainly reflected in the following aspects: the airport surface involves multiple elements such as flight taxi paths, ground service vehicle scheduling, runway and taxiway resource allocation, and their coordinated operation has highly nonlinear and dynamic characteristics; weather, flight schedule changes, and ground emergencies may all cause drastic changes in the operation situation; the data sources of the surface operation situation are diverse, including flight schedule data, ground monitoring data, meteorological data, etc. How to integrate and mine the key driving factors is the focus of current research. Traditional operation management methods (such as those based on empirical rules or simple statistical analysis) have been unable to cope with the high dynamics and complexity of actual scenarios; therefore, it is particularly important to invent methods for extracting and simulating the driving factors of the surface operation situation.

[0003] The existing methods for extracting and simulating the driving factors of surface operation situations have high operational risks due to possible data leakage, which reduces the generalization ability of the overall model and the performance of localized deduction, and has high deduction errors due to data quality differences or abnormal data. In addition, the existing methods for extracting and simulating the driving factors of surface operation situations are prone to fall into local optimal solutions, and cannot provide a more comprehensive optimization strategy for airport surface operations, and have poor efficiency in dealing with deduction problems under complex constraints. Therefore, we propose a method for extracting and simulating the driving factors of surface operation situations. Summary of the invention

[0004] The purpose of the present invention is to solve the defects in the prior art and to propose a method for extracting and simulating driving factors of scene operation situation.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The method for extracting and simulating driving factors of scene operation situation is as follows:

[0007] Ⅰ. Collect and pre-process the operation data of each group of airports, and store the pre-processed data in the central data warehouse;

[0008] II. Construct an airport surface map of airport surface operation, extract key factors of surface operation situation, and screen key driving factors of airport operation situation and assign weights;

[0009] III. Establish flight taxiing time series, predict taxiing delays and conflicts under different conditions, simulate multiple situation evolution paths in airport operation scenarios, and make dynamic adjustments;

[0010] IV. Model the simulation process as a multi-stage decision problem and simulate path expansion under different conditions while optimizing the trade-off between flight taxi time and ground conflict risk;

[0011] V. Calibrate the probability distribution of the final scene situation prediction path, provide key decision point suggestions, generate a dynamic diagram of the airport scene situation evolution, and provide multi-scenario comparative analysis.

[0012] As a further solution of the present invention, the specific steps of constructing the airport scene map of airport scene operation in step I are as follows:

[0013] S1.1: Use NLP technology to parse the preprocessed flight schedule data and text information in the operation report, and then use the CRF model to identify core entities from the text information. After the entity information is extracted, the twin tower model is used to extract the relationship between each group of entities from the flight data;

[0014] S1.2: The extracted entities are taken as nodes, and the relationship between each group of entities is taken as the edge between two groups of entities. The entities and relationships are stored in a graph database to form an airport scene map. Then the GCN network extracts the structured information between nodes by recursively aggregating the features of neighboring nodes, and extracts the key driving factors affecting the airport operation situation based on the importance of the nodes and the frequency of the relationships.

[0015] S1.3: Generate node embedding vectors using graph embedding technology, combine with principal component analysis to reduce dimensionality, and then calculate the contribution of each driving factor to situation prediction to quantify the importance of key driving factors.

[0016] As a further solution of the present invention, the core entities described in S1.1 specifically include groups of entities such as flight number, aircraft type, taxiway number, ground service vehicle and time information; the relationships between the entities specifically include various types of relationships such as the path relationship between the flight and the taxiway and the service relationship between the flight and the ground vehicle.

[0017] As a further solution of the present invention, the specific steps of screening the key driving factors of the airport operation situation and assigning weights in step III are as follows:

[0018] S2.1: Collect the data storage systems of different airports and use the data storage systems of each airport as nodes. The central server builds an initial set of global models according to the preset network architecture and parameters. Each node builds a local model. The architecture and parameters of the built global model are then sent to each node and used as the architecture and parameters of each local model.

[0019] S2.2: The factors extracted from the airport scene map are used as the local features of each node. Then, each airport node trains the corresponding local model through the local features. Then, each local model outputs the prediction results through forward propagation, and then the architecture and parameters of the local model are updated through the local loss function and back propagation algorithm.

[0020] S2.3: The updated local model architecture and parameter information are uploaded to the central server. The central server performs weighted aggregation on each set of received architecture and parameter information to obtain a new global model architecture and parameters, and replaces the original global model architecture and parameters. The KL divergence is then used to calculate the similarity between the local model output probability distribution and the pseudo target, and the cross entropy loss function is used to calculate the gap between the local model output and the true label. The KL divergence and the cross entropy loss function are then combined through weighted combination to obtain the total loss function of the local model.

[0021] S2.4: The local model of each node receives the updated global model architecture, parameters and soft labels, and replaces the architecture and parameters of the original local model. Then, the true value of the local feature data is used as the true label. The local model of each node is jointly trained based on the local true label and pseudo label, with the goal of minimizing the local loss function.

[0022] S2.5: In each training iteration of the local model, the gradient of the total loss function is calculated, and the parameters of the student model are adjusted using the gradient descent method according to the calculated gradient value. When the total loss function converges to the preset range, the training of the local model is stopped, and the diagnostic accuracy of the local model and the global model are compared through the verification data set, and the reasoning speed and resource usage of the student model are verified. If the local model meets the preset requirements, it is used as the final lightweight model and deployed to each airport node;

[0023] S2.6: Each airport node obtains the importance score of each factor through the local model, integrates the importance scores through the federated learning mechanism, assigns weights to the key driving factors according to the integrated importance scores, and then normalizes the weights. Based on the final weights, the ranking of key driving factors is generated to provide a quantitative explanation of the important driving factors.

[0024] As a further solution of the present invention, the specific steps of simulating multiple situation evolution paths in the airport operation scenario and dynamically adjusting in step III are as follows:

[0025] S3.1: Extract the initial situation status of the current scenario from the airport operation data, including the current flight taxi path and taxi time, ground service vehicle scheduling, runway usage and queue length, and define the objective function based on minimizing taxi time and optimizing resource utilization;

[0026] S3.2: Set the initial temperature, cooling rate, and stop temperature, generate a neighborhood state from the current state by randomly adjusting the taxiway occupancy or runway priority allocation and modifying the scheduling strategy of the ground service vehicle, and calculate the objective function values ​​of the current state and the new state;

[0027] S3.3: If the objective function value of the current state is less than the objective function value of the new state, the new state is accepted and the current state is replaced. If the objective function value of the current state is greater than the objective function value of the new state, the acceptance probability of the new state is calculated and the new state is accepted with the acceptance probability.

[0028] S3.4: If the new state is accepted, the new state is set as the current state, otherwise it remains unchanged. After each round of new state replacement is completed, the current temperature is updated according to the preset cooling rate until the current temperature reaches the stop temperature, the state update is stopped, and the current state and the corresponding key driving factor configuration are output as the optimal path for state evolution.

[0029] As a further solution of the present invention, the specific steps of calibrating the probability distribution of the final scene situation prediction path and providing key decision point suggestions in step V are as follows:

[0030] S4.1: The taxiway occupancy, flight queue length, ground service resource allocation and subsequent states generated by simulation are taken as nodes, and a decision search tree is constructed based on each set of information. Each node in the tree represents a specific state of airport surface operation, and the root node of the tree is the current initial state.

[0031] S4.2: Assign a probability to each path to indicate the likelihood of its occurrence, sample the deduced paths, calibrate the probabilities based on historical operating data, and smooth the path probability distribution based on the probability calibration formula;

[0032] S4.3: Traverse the decision search tree, give priority to transfer paths with higher probabilities, and retain multiple groups of low-probability paths to explore potential suboptimal solutions. Then, the final situation status results of different paths are merged through weighted average to generate a comprehensive situation assessment. Based on the comprehensive situation status and combined with key driving factors, explainable decision recommendations are generated, and the core reasons for the situation changes are explained based on the SHAP value.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. The present invention sends the architecture and parameters of the global model built by the central server to each node, and uses them as the architecture and parameters of each local model, and uses the factors extracted from the airport scene map as the local features of each node. Then, each airport node trains and updates the architecture and parameters of the corresponding local model through local features, and uploads the updated local model architecture and parameter information to the central server. The central server performs weighted aggregation on each set of received architecture and parameter information to obtain a new global model architecture and parameters, and replaces the original global model architecture and parameters. Then, the KL divergence and cross entropy loss function are combined through weighted combination to obtain the total loss function of the local model. The local model of each node receives the updated global model architecture, parameters and soft labels, and replaces the original local model architecture and parameters. Then, the true value of the local feature data is used as the true label. The local model of each node is jointly trained according to the local true label and the pseudo label, respectively, with the goal of minimizing the local loss function. In each training iteration of the local model, The gradient of the total loss function is calculated, and the parameters of the student model are adjusted using the gradient descent method according to the calculated gradient value. When the total loss function converges to the preset range, the training of the local model is stopped. The diagnostic accuracy of the local model and the global model are compared through the verification data set, and the reasoning speed and resource occupancy of the student model are verified. If the local model meets the preset requirements, it will be used as the final lightweight model and deployed to each airport node. Each airport node obtains the importance score of each factor through the local model, and integrates the importance scores through the federated learning mechanism. According to the integrated importance scores, weights are assigned to the key driving factors, and then the weights are normalized. Based on the final weights, the ranking of key driving factors is generated, and a quantitative explanation of the important driving factors is provided, which can reduce the operational risks that may be caused by data leakage, improve the generalization ability of the overall model, improve the performance of localized deduction, reduce dependence on high-performance hardware and system costs, avoid deduction errors caused by data quality differences or abnormal data, and improve the extraction accuracy of key driving factors.

[0035] 2. The scene operation situation driving factor extraction and simulation deduction method extracts the initial situation state of the current scene from the airport operation data, including the current flight taxi path and taxi time, ground service vehicle scheduling, runway usage and queue length, defines the objective function based on minimizing taxi time and optimizing resource utilization, sets the initial temperature, cooling rate and stop temperature, generates a neighborhood state from the current state by randomly adjusting the taxiway occupancy or runway priority allocation and modifying the scheduling strategy of ground service vehicles, and calculates the objective function values ​​of the current state and the new state. If the objective function value of the current state is less than the objective function value of the new state, the new state is accepted and replaced. Previous state, if the objective function value of the current state is greater than the objective function value of the new state, then calculate the acceptance probability of the new state, and accept the new state with the acceptance probability. If the new state is accepted, set the new state as the current situation state, otherwise it remains unchanged. After each round of new state replacement is completed, update the current temperature according to the preset cooling rate until the current temperature reaches the stop temperature, stop state update, and output the current situation state and the corresponding key driving factor configuration. As the optimal path for situation evolution, it can avoid falling into the local optimal solution, provide a more comprehensive optimization strategy for airport surface operations, and efficiently handle deduction problems under complex constraints, helping airport operation managers to predict problems in advance and design response strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0037] Figure 1 This is a flowchart of the scene operation situation driving factor extraction and simulation deduction method proposed in the present invention. DETAILED DESCRIPTION

[0038] Example 1

[0039] Reference Figure 1 , the scene operation situation driving factor extraction and simulation deduction method, the specific steps of the deduction method are as follows:

[0040] Collect and pre-process the operation data of each group of airports, and store the pre-processed data in the central data warehouse.

[0041] Construct an airport surface map of airport surface operations, extract key factors of surface operation status, and at the same time screen the key driving factors of airport operation status and assign weights.

[0042] Specifically, NLP technology is used to parse the preprocessed flight schedule data and text information in the operation report, and then the CRF model is used to identify core entities from the text information. After the entity information is extracted, the dual-tower model is used to extract the relationship between each group of entities from the flight data. The extracted entities are used as nodes, and the relationship between each group of entities is used as the edge between the two groups of entities. The entities and relationships are stored in a graph database to form an airport scene map. After that, the GCN network recursively aggregates the features of neighboring nodes to extract the structured information between nodes, and extracts the key driving factors that affect the airport operation situation based on the importance of the nodes and the frequency of the relationships. The node embedding vector is generated using graph embedding technology, combined with principal component analysis for dimensionality reduction, and then the contribution value of each driving factor to the situation prediction is calculated to quantify the importance of the key driving factors.

[0043] It should be further explained that the core entities specifically include flight number, aircraft type, taxiway number, ground service vehicle and time information entities; the relationship between the entities specifically includes the path relationship between the flight and the taxiway and the service relationship between the flight and the ground vehicle.

[0044] Specifically, data storage systems of different airports are collected, and the data storage systems of each airport are used as nodes. The central server builds a set of initial global models according to the preset network architecture and parameters, and each node builds a local model. The architecture and parameters of the built global model are then sent to each node and used as the architecture and parameters of each local model. The factors extracted from the airport scene map are used as the local features of each node. Then, each airport node trains the corresponding local model through local features, and then each local model outputs the prediction result through forward propagation. Then, the architecture and parameters of the local model are updated through the local loss function and the back propagation algorithm, and the updated local model architecture and parameter information are uploaded to the central server. The central server performs weighted aggregation on the received sets of architecture and parameter information to obtain a new global model architecture and parameters, and replaces the original global model architecture and parameters. Then, the KL divergence is used to calculate the similarity between the local model output probability distribution and the pseudo target, and the cross entropy loss function is used to calculate the gap between the local model output and the true label. Then, the KL divergence and the cross entropy loss function are combined through weighted combination to obtain the The total loss function of the local model. The local model of each node receives the updated global model architecture, parameters and soft labels, and replaces the architecture and parameters of the original local model. Then, the true value of the local feature data is used as the true label. The local model of each node is jointly trained based on the local true label and pseudo label, with the goal of minimizing the local loss function. In each training iteration of the local model, the gradient of the total loss function is calculated, and the parameters of the student model are adjusted using the gradient descent method based on the calculated gradient value. When the total loss function converges to the preset range, the training of the local model is stopped. The diagnostic accuracy of the local model and the global model are compared through the verification data set, and the reasoning speed and resource occupancy of the student model are verified. If the local model meets the preset requirements, it will be used as the final lightweight model and deployed to each airport node. Each airport node obtains the importance score of each factor through the local model, integrates the importance scores through the federated learning mechanism, assigns weights to the key driving factors according to the integrated importance scores, and then normalizes the weights. Based on the final weights, the ranking of key driving factors is generated to provide a quantitative explanation of the important driving factors.

[0045] Example 2

[0046] Reference Figure 1 , the scene operation situation driving factor extraction and simulation deduction method, the specific steps of the deduction method are as follows:

[0047] Establish flight taxi time series, predict taxi delays and conflicts under different conditions, simulate multiple situation evolution paths in airport operation scenarios, and make dynamic adjustments.

[0048] Specifically, the initial situation state of the current scenario is extracted from the airport operation data, including the taxi path and taxi time of the current flight, the scheduling of ground service vehicles, the use of runways and the length of the queue. The objective function is defined based on minimizing the taxi time and optimizing the resource utilization, and the initial temperature, cooling rate and stop temperature are set. By randomly adjusting the taxiway occupancy or runway priority allocation and modifying the scheduling strategy of the ground service vehicle, a neighborhood state is generated from the current state, and the objective function values ​​of the current state and the new state are calculated. If the objective function value of the current state is less than the objective function value of the new state, the new state is accepted and replaced with the current state. If the objective function value of the current state is greater than the objective function value of the new state, the acceptance probability of the new state is calculated, and the new state is accepted with the acceptance probability. If the new state is accepted, the new state is set as the current situation state, otherwise it remains unchanged. After each round of new state replacement is completed, the current temperature is updated according to the preset cooling rate until the current temperature reaches the stop temperature, the state update is stopped, and the current situation state and the corresponding key driving factor configuration are output as the optimal path for situation evolution.

[0049] The simulation process is modeled as a multi-stage decision-making problem, and path expansion under different conditions is simulated while optimizing the trade-off between flight taxi time and ground conflict risk.

[0050] Calibrate the probability distribution of the final scene situation prediction path, provide key decision point suggestions, generate a dynamic diagram of the airport scene situation evolution, and provide multi-scenario comparative analysis.

[0051] Specifically, the taxiway occupancy, flight queue length, ground service resource allocation and subsequent states generated by deduction simulation are taken as nodes, and a decision search tree is constructed based on each group of information. Each node in the tree represents a specific situation state of airport surface operation. The root node of the tree is the current initial situation state. A probability is assigned to each path to indicate the possibility of its occurrence. The deduced path is sampled, the probability is calibrated according to historical operation data, and the path probability distribution is smoothed based on the probability calibration formula. The decision search tree is traversed, and transfer paths with higher probabilities are given priority. Multiple groups of low-probability paths are retained to explore potential suboptimal solutions. The final situation state results of different paths are then merged through weighted averaging to generate a comprehensive situation assessment. Based on the comprehensive situation state and combined with key driving factors, explainable decision recommendations are generated, and the core reasons for the situation changes are explained based on the SHAP value.

Claims

1. A method for extracting and simulating driving factors of scene operation situation, characterized in that: The specific steps of the deduction method are as follows: Ⅰ. Collect and pre-process the operation data of each group of airports, and store the pre-processed data in the central data warehouse; II. Construct an airport surface map of airport surface operation, extract key factors of surface operation situation, and screen key driving factors of airport operation situation and assign weights; III. Establish flight taxiing time series, predict taxiing delays and conflicts under different conditions, simulate multiple situation evolution paths in airport operation scenarios, and make dynamic adjustments; IV. Model the simulation process as a multi-stage decision problem and simulate path expansion under different conditions while optimizing the trade-off between flight taxi time and ground conflict risk; V. Calibrate the probability distribution of the final scene situation prediction path, provide key decision point suggestions, generate a dynamic diagram of the airport scene situation evolution, and provide multi-scenario comparative analysis.

2. The method for extracting and simulating driving factors of scene operation situation according to claim 1 is characterized in that: The specific steps of constructing the airport surface map of airport surface operation described in step I are as follows: S1.1: Use NLP technology to parse the preprocessed flight schedule data and text information in the operation report, and then use the CRF model to identify core entities from the text information. After the entity information is extracted, the twin tower model is used to extract the relationship between each group of entities from the flight data; S1.2: The extracted entities are taken as nodes, and the relationship between each group of entities is taken as the edge between two groups of entities. The entities and relationships are stored in a graph database to form an airport scene map. Then the GCN network extracts the structured information between nodes by recursively aggregating the features of neighboring nodes, and extracts the key driving factors affecting the airport operation situation based on the importance of the nodes and the frequency of the relationships. S1.3: Generate node embedding vectors using graph embedding technology, combine with principal component analysis to reduce dimensionality, and then calculate the contribution of each driving factor to situation prediction to quantify the importance of key driving factors.

3. The method for extracting and simulating driving factors of scene operation situation according to claim 2 is characterized in that: The core entities described in S1.1 specifically include flight number, aircraft type, taxiway number, ground service vehicle and time information entities; the relationships between the entities specifically include the path relationship between the flight and the taxiway and the service relationship between the flight and the ground vehicle.

4. The method for extracting and simulating driving factors of scene operation situation according to claim 2 is characterized in that: The specific steps for screening the key drivers of airport operation status and assigning weights in step III are as follows: S2.1: Collect the data storage systems of different airports and use the data storage systems of each airport as nodes. The central server builds an initial set of global models according to the preset network architecture and parameters. Each node builds a local model. The architecture and parameters of the built global model are then sent to each node and used as the architecture and parameters of each local model. S2.2: The factors extracted from the airport scene map are used as the local features of each node. Then, each airport node trains the corresponding local model through the local features. Then, each local model outputs the prediction results through forward propagation, and then the architecture and parameters of the local model are updated through the local loss function and back propagation algorithm. S2.3: The updated local model architecture and parameter information are uploaded to the central server. The central server performs weighted aggregation on each set of received architecture and parameter information to obtain a new global model architecture and parameters, and replaces the original global model architecture and parameters. The KL divergence is then used to calculate the similarity between the local model output probability distribution and the pseudo target, and the cross entropy loss function is used to calculate the gap between the local model output and the true label. The KL divergence and the cross entropy loss function are then combined through weighted combination to obtain the total loss function of the local model. S2.4: The local model of each node receives the updated global model architecture, parameters and soft labels, and replaces the architecture and parameters of the original local model. Then, the true value of the local feature data is used as the true label. The local model of each node is jointly trained based on the local true label and pseudo label, with the goal of minimizing the local loss function. S2.5: In each training iteration of the local model, the gradient of the total loss function is calculated, and the parameters of the student model are adjusted using the gradient descent method according to the calculated gradient value. When the total loss function converges to the preset range, the training of the local model is stopped, and the diagnostic accuracy of the local model and the global model are compared through the verification data set, and the reasoning speed and resource usage of the student model are verified. If the local model meets the preset requirements, it is used as the final lightweight model and deployed to each airport node; S2.6: Each airport node obtains the importance score of each factor through the local model, integrates the importance scores through the federated learning mechanism, assigns weights to the key driving factors according to the integrated importance scores, and then normalizes the weights. Based on the final weights, the ranking of key driving factors is generated to provide a quantitative explanation of the important driving factors.

5. The method for extracting and simulating driving factors of scene operation situation according to claim 4 is characterized in that: The specific steps of simulating multiple situation evolution paths and dynamically adjusting them in the airport operation scenario described in step III are as follows: S3.1: Extract the initial situation status of the current scenario from the airport operation data, including the current flight taxi path and taxi time, ground service vehicle scheduling, runway usage and queue length, and define the objective function based on minimizing taxi time and optimizing resource utilization; S3.2: Set the initial temperature, cooling rate, and stop temperature, generate a neighborhood state from the current state by randomly adjusting the taxiway occupancy or runway priority allocation and modifying the scheduling strategy of the ground service vehicle, and calculate the objective function values ​​of the current state and the new state; S3.3: If the objective function value of the current state is less than the objective function value of the new state, the new state is accepted and the current state is replaced. If the objective function value of the current state is greater than the objective function value of the new state, the acceptance probability of the new state is calculated and the new state is accepted with the acceptance probability. S3.4: If the new state is accepted, the new state is set as the current state, otherwise it remains unchanged. After each round of new state replacement is completed, the current temperature is updated according to the preset cooling rate until the current temperature reaches the stop temperature, the state update is stopped, and the current state and the corresponding key driving factor configuration are output as the optimal path for state evolution.

6. The method for extracting and simulating driving factors of scene operation situation according to claim 5 is characterized in that: The specific steps of calibrating the probability distribution of the final scene situation prediction path and providing key decision point suggestions described in step V are as follows: S4.1: The taxiway occupancy, flight queue length, ground service resource allocation and subsequent states generated by simulation are taken as nodes, and a decision search tree is constructed based on each set of information. Each node in the tree represents a specific state of airport surface operation, and the root node of the tree is the current initial state. S4.2: Assign a probability to each path to indicate the likelihood of its occurrence, sample the deduced paths, calibrate the probabilities based on historical operating data, and smooth the path probability distribution based on the probability calibration formula; S4.3: Traverse the decision search tree, give priority to transfer paths with higher probabilities, and retain multiple groups of low-probability paths to explore potential suboptimal solutions. Then, the final situation status results of different paths are merged through weighted average to generate a comprehensive situation assessment. Based on the comprehensive situation status and combined with key driving factors, explainable decision recommendations are generated, and the core reasons for the situation changes are explained based on the SHAP value.

Images