An airport operation risk time series prediction system based on a graph neural network

Abstract

The application provides an airport operation risk time sequence prediction system based on a graph neural network, comprising: a data acquisition module, which acquires airport operation risk and operation situation data, cleans the data, and divides the data into time sequence data according to time periods; a graph structure construction module, which calculates the correlation between risks based on the risk time sequence data, and forms a directed acyclic graph structure with various risks as nodes; a weight generation module, which calculates the correlation of associated nodes with the risk nodes as the center, and generates edge weights according to the strength of the correlation; and a risk prediction module, which processes the graph structure data formed by the risk nodes by using a graph neural network, and realizes the prediction of the risk values of the risk nodes at future moments. The application solves the problem that the coupling relationship of airport operation risks is complex and difficult to accurately predict, fully considers the coupling and correlation relationship of risk factors, and improves the accuracy of risk prediction.

Description

TECHNICAL FIELD

[0001] The application relates to an airport operation risk timing prediction system, in particular to an airport operation risk timing prediction system based on a graph neural network. BACKGROUND

[0002] As one of the three systems of civil aviation transportation, a civil aviation airport is an important infrastructure and component of air traffic. The airport is not only a material basis for ensuring the safe flight of an aircraft, but also a primary condition for the safe take-off and landing of the aircraft. Therefore, it is crucial to ensure the safety of the airport. According to statistics of the International Civil Aviation Organization, more than 100 aviation accidents occur on average every year in the world. Although the level of science and technology has greatly improved the aviation safety of the aircraft in the 21st century, the occurrence of aviation accidents cannot be completely avoided. As a basic condition for passenger travel, the civil aviation airport provides maintenance and inspection services for the aircraft, and provides check-in, security check, luggage consignment and other services for passengers. The operation of the airport is composed of many operation activities, and the safety of the airport includes runway safety, apron safety, aircraft operation safety, special vehicle operation safety and personnel safety and many other aspects. The evaluation and prediction of the airport operation risk help the risk control strategy of the airport managers to provide auxiliary decision basis, thereby improving the management efficiency of the relevant departments and promoting the management mode to be more fine and scientific, which has certain theoretical value and important practical significance for improving the operation efficiency and improving the operation state of the airport.

[0003] The dangerous factors affecting the safe operation of the flight area often exist in different links of the operation process. The role of each link may become the key to affecting the safe operation of the flight area, and all factors related to the safe operation of the flight area may have certain correlation. For example, the snow on the road surface may cause the risk of reducing the road surface friction coefficient, and then cause the risks of route deviation and apron traffic accident, and finally cause the risks of aircraft-aircraft collision and aircraft-vehicle collision. Based on the above analysis, the occurrence of the flight area operation risk is often the result of the interaction of many unsafe factors. As long as a part of negligence may cause adverse consequences, the risks form obvious graph structure data. If the cause of the unsafe event or accident is simply attributed to a single risk of the unsafe event or accident, the safe operation of the flight area cannot be well improved. SUMMARY

[0004] The technical problem to be solved by the application is to provide an airport operation risk timing prediction system based on a graph neural network in view of the shortcomings of the prior art.

[0005] In order to solve the above technical problems, the application discloses an airport operation risk time sequence prediction system based on a graph neural network, which comprises a data acquisition module, a graph structure construction module, a weight generation module and a risk prediction module.

[0006] The data acquisition module is used to acquire the data of the airport operation risk and the operation situation, clean the data, divide the cleaned data into time sequence data according to time periods, and deliver the time sequence data to the graph structure construction module, the weight generation module and the risk prediction module.

[0007] The graph structure construction module is used to analyze the time sequence data, establish the connection relationship between the risks, form a directed acyclic graph structure with various risks as nodes, and deliver the directed acyclic graph structure to the risk prediction module.

[0008] The weight generation module calculates the data correlation between the risk nodes according to the time sequence data, generates the edge weight between the nodes in the directed acyclic graph structure, and delivers the edge weight to the risk prediction module.

[0009] The risk prediction module processes the graph structure data by using a graph neural network prediction method according to the time sequence data of each risk node in the directed acyclic graph structure and the edge weight, and completes the prediction of the risk values of each risk node at a future time.

[0010] The airport operation risk time sequence prediction system based on the graph neural network performs the following steps:

[0011] Step S1: The data acquisition module acquires the risk levels of various risks and the related data of the operation situation in the airport operation process, cleans the data, removes the missing data, the error data and the duplicate data, divides the cleaned data into time sequence data of the airport operation risk at fixed time periods, and delivers the time sequence data to the graph structure construction module, the weight generation module and the risk prediction module.

[0012] The related data at least comprises the risk values of the runway intrusion risk, the taxiway conflict risk, the restricted area intrusion risk, the low-visibility risk, the low-altitude wind shear risk, the runway area rain risk, the runway area snow risk, the runway area ice risk, the friction coefficient reduction risk, the route deviation risk, the apron traffic event risk, the aircraft and aircraft collision risk, the aircraft and vehicle scratch risk, the aircraft and personnel scratch risk, the aircraft and building scratch risk and the situation data reflecting the airport operation busy degree, such as the runway utilization rate, the taxiway utilization rate and the vehicle utilization rate.

[0013] Step S2, the graph structure construction module constructs a graph structure data with the cleaned data in step S1 as nodes of the graph, and transmits the generated graph structure data to the risk prediction module;

[0014] Further, the graph structure data is constructed by the following method:

[0015] A Bayesian network structure learning algorithm is used to establish a directed and acyclic connection relationship among the nodes, and the connection relationship among the generated graph structure nodes is adjusted and optimized according to the risk attributes of the nodes, so as to obtain a risk relationship graph structure and complete the construction of the graph structure data.

[0016] Further, the graph structure data can also be constructed by the following method:

[0017] A directed edge is established between the related nodes of the graph, and finally the graph structure data of the risk nodes is constructed by the nodes and the directed edges, which specifically includes:

[0018] The correlation between each target node and all other nodes is calculated, and the nodes with a correlation exceeding a set threshold are selected as connected nodes of the target node, a connection is established between the connected nodes, and the child nodes and the parent nodes are determined according to the risk attributes of the connected nodes, i.e. a directed edge is obtained.

[0019] Further, the correlation between each target node and all other nodes is calculated, and the nodes with a correlation exceeding a set threshold are selected as connected nodes of the target node, which specifically includes:

[0020] The Pearson linear correlation coefficient is calculated to obtain the correlation between node s and all other nodes, and the nodes exceeding a set threshold are selected as connected nodes of node s, which specifically includes:

[0021]

[0022] A sp = 1, {e sp ≥ TH, p ∈ SS}

[0023] wherein e sp represents the Pearson linear correlation coefficient value between node s and node p, represents the risk value of node s at the nth moment in history, represents the average risk value of node s at N moments in history, represents the risk value of other nodes at N moments in history except node s, represents the average risk value of other nodes at N moments in history except node s, A sp = 1 represents that node s is connected with node p, TH represents a set threshold, and SS represents a set of other nodes except node s.

[0024] Step S3, the weight generation module calculates the relevance of the associated nodes of the node and generates the corresponding edge weight according to the strength of the relevance, and transmits the generated edge weight to the risk prediction module;

[0025] The edge weight calculation method comprises:

[0026] For any target node, the relevance between the target node and the nodes connected to it in the graph structure is calculated, and the ratio of the relevance value to the sum of the relevance values of all nodes connected to the target node is taken as the edge weight between the node and the target node; all nodes are traversed to perform the above calculation to obtain the edge weights of all connected nodes, which are as follows:

[0027]

[0028] wherein, w sp represents the edge weight of the target node s and the node p connected to it, l sp represents the Pearson linear correlation coefficient value of the two nodes, and the set S represents the set of all parent nodes associated with the node s.

[0029] The relevance between the target node and the nodes connected to it in the graph structure is calculated, and the specific method comprises:

[0030] The Pearson linear correlation coefficient is used to calculate the relevance l, and the method is as follows:

[0031]

[0032] wherein, l represents the Pearson linear correlation coefficient value of the target node s and the node p connected to it, s i represents the risk value of the node s at the i th moment, represents the average risk value of the target node at all moments, p i represents the risk value of the node p at the i th moment, represents the average risk value of the node p at all moments, and N is the selected time period.

[0033] Step S4, the risk prediction module takes the time series data input by the data acquisition module as the input feature value of each node in the graph structure data, takes the edge weight input by the weight generation module as the input feature value of each directed edge in the graph structure data, and uses a graph neural network to predict the risk value of each risk node at a future moment, which specifically comprises:

[0034] The graph neural network learns a d-dimensional vector x n of each node by a graph embedding method, wherein the x n contains the information of the current node and the connected nodes, and the calculation method is as follows:

[0035] x n = f w (l n , l co[n] , x ne[n] , l ne[n] )

[0036] Wherein, l n Indicates the input feature of the node, that is, the node risk value at the historical moment, l co[n] Indicates the feature of the edge connected with the node, that is, the edge weight described in step S3, x ne[n] Indicates the embedding vector of the node connected with the node, l ne[n] Indicates the feature of the node connected with the node, that is, the historical risk value of the node connected with the node, f w Indicates the conversion function of the graph neural network;The output value of each node is finally calculated by the following method:

[0037] o n = g w (x n , l n )

[0038] Wherein, o n Is the output value of the node, that is, the predicted risk value of the node at the next moment or the next few moments, g w Indicates the output function of the graph neural network.

[0039] Step S5, verify the accuracy of the prediction of the graph neural network, if the accuracy requirement is not reached, repeat the execution of step S1 to step S4, until the prediction result of the graph neural network meets the accuracy requirement.

[0040] Beneficial effects:

[0041] In the application, the airport operation risk time sequence prediction based on the graph neural network is established, the directed acyclic graph structure is established with each risk as a node according to the correlation of each risk, and the graph neural network is used for risk time sequence prediction. The coupling relationship between risks is fully considered, which greatly improves the accuracy of risk prediction.

[0042] First, the airport operation risk correlation relationship based on data driving is established, and the correlation relationship between each risk is no longer formulated by an expert system or past research. The problem that the expert system is difficult to obtain and the expert system has subjectivity and deviation from the actual situation is solved.

[0043] Second, the coupling correlation between risks is fully considered, the influence of other risks on the risk is considered while considering the single risk time sequence change, the weight relationship of risk correlation is established, so that the coupling relationship between each risk can be more accurately established.

[0044] Third, the graph neural network is used to process the graph structure data to realize more accurate airport operation risk prediction. The graph neural network has good modeling capability for graph structure data, can learn the relationship or logical reasoning mode hidden in the graph structure data from the relationship between the nodes and edges, and thus realizes accurate prediction of the coupling risk. BRIEF DESCRIPTION OF DRAWINGS

[0045] The above and / or other aspects of the present application will become apparent and more readily appreciated from the following description, taken in conjunction with the accompanying drawings, in which:

[0046] Figure 1 is a schematic diagram of the system framework of the present application.

[0047] Figure 2 is a flowchart of the establishment of a directed acyclic graph based on the K2 algorithm of the present application.

[0048] Figure 3 is a schematic diagram of the risk node relationship of the present application.

[0049] Figure 4 is a flowchart of the establishment of a directed acyclic graph based on the correlation of the present application. DETAILED DESCRIPTION

[0050] The present application aims to solve the problem that the coupling relationship of airport operation risk is complex and difficult to accurately predict. A graph neural network-based airport operation risk time series prediction system is constructed. The coupling relationship of risk factors is fully considered. The airport operation risk graph structure is established based on risk factors as nodes. The graph neural network is used for risk prediction based on historical risk data to improve the accuracy of risk prediction.

[0051] The system comprises a data acquisition module, a graph structure construction module, a weight generation module and a risk prediction module.

[0052] The data acquisition module is used to acquire the relevant data of airport operation risk and operation situation, clean the data, remove missing data, error data and duplicate data, and divide the collected historical data into time series data according to time periods. The time series data is transmitted as input to the graph structure construction module, the weight generation module and the risk prediction module.

[0053] The graph structure construction module is used to analyze the collected various risk time series data, establish the connection relationship between risks, form a directed acyclic graph structure with various risks as nodes, and transmit the graph structure as input to the risk prediction module.

[0054] The weight generation module calculates data correlation between risk nodes according to time sequence data of each risk node, generates edge weights between nodes in the graph structure, and the edge weights are input to the risk prediction module;

[0055] The risk prediction module predicts risk values of each risk node at a future time by using a graph neural network prediction technology according to historical time sequence data of each risk node in the graph structure and the edge weights in the graph structure.

[0056] The system performs the following steps:

[0057] In step S1, the data acquisition module acquires risk levels and operation situation data of various risks in the airport operation process, including runway intrusion risk, taxiway conflict risk, restricted area intrusion risk, low visibility risk, low-altitude wind shear risk, runway area rain risk, runway area snow risk, runway area ice risk, friction coefficient reduction risk, route deviation risk, apron traffic event risk, aircraft and aircraft collision risk, aircraft and vehicle scratch risk, aircraft and personnel scratch risk, aircraft and building scratch risk, and situation data reflecting the busy degree of airport operation such as runway utilization rate, taxiway utilization rate, and vehicle utilization rate. The data acquisition module cleans the data to remove missing data, error data, and duplicate data, and divides the collected data into fixed time periods to obtain time sequence data of airport operation risks, which are input to the graph structure construction module, the weight generation module, and the risk prediction module;

[0058] In step S2, the graph structure construction module constructs graph structure data by taking operation risks and operation situations as nodes, traverses all nodes according to the input time sequence data of each node, calculates the correlation between each two nodes, establishes a directed edge for the correlated nodes, and finally constructs the graph structure data of the risk nodes by nodes and edges. The generated graph structure data is input to the risk prediction module;

[0059] In step S3, the weight generation module calculates the correlation of associated nodes with the node as the center, and generates corresponding edge weights according to the strength of the correlation. The stronger the correlation, the higher the weight. The generated edge weights are input to the risk prediction module;

[0060] In step S4, the risk prediction module takes the historical risk data input by the data acquisition module as the input feature value of each node in the graph structure, takes the edge weight input by the weight generation module as the input feature value of each directed edge in the graph structure, and predicts the risk values of each risk node at the next or several time points by using a graph neural network.

[0061] Step S5: verifying the prediction accuracy of the graph neural network, if the accuracy requirement is not met, more operation risk data is acquired in step S1, the graph neural network is continuously trained until the prediction result meets the accuracy requirement.

[0062] Embodiments:

[0063] The application provides an airport operation risk time series prediction system based on a graph neural network, which comprises a data acquisition module, a graph structure construction module, a weight generation module and a risk prediction module, as shown in Figure 1 .

[0064] The data acquisition module is used for acquiring airport operation risk and operation situation data, cleaning the data, removing missing data, error data and duplicate data, and dividing the acquired historical data into time series data according to time periods and inputting the time series data into the graph structure construction module, the weight generation module and the risk prediction module;

[0065] The graph structure construction module establishes the connection relationship between risks based on the correlation of various risk time series data, forms a directed acyclic graph structure with various risks as nodes and inputs the graph structure into the risk prediction module;

[0066] The weight generation module calculates the data correlation between risk nodes according to the time series data of each risk node, generates the edge weight between each node in the network and inputs the edge weight into the risk prediction module;

[0067] The risk prediction module uses the graph neural network prediction technology to realize the prediction of the risk values of each risk node at the future moment according to the historical time series data of each risk node in the graph structure and the edge weight in the graph structure.

[0068] The application provides two embodiments, which are respectively aimed at different graph structure construction modes to realize optimal prediction results.

[0069] Embodiment 1:

[0070] This embodiment is aimed at airport operation risk, and uses the graph neural network to realize risk prediction, which comprises the following steps:

[0071] Step S1: The data collection module collects the risk levels and operation situation related data of various risks in the airport operation process, including: runway incursion risk, taxiway conflict risk, restricted area incursion risk, low visibility risk, low-level wind shear risk, runway area rain risk, runway area snow risk, runway area ice risk, friction coefficient reduction risk, route deviation risk, apron traffic event risk, aircraft and aircraft collision risk, aircraft and vehicle collision risk, aircraft and personnel collision risk, aircraft and building collision risk, and other risk values, as well as runway utilization rate, taxiway utilization rate, vehicle utilization rate and other situation data reflecting the busy degree of airport operation. The data collection module cleans the data, removes missing data, error data and duplicate data, and divides the collected data into fixed time periods to obtain time series data of airport operation risk, which is input to the graph structure construction module, the weight generation module and the risk prediction module;

[0072] Step S2: The graph structure construction module constructs the graph structure data with operation risk and operation situation as nodes. According to the time series data input by each node, all nodes are traversed, the correlation between each two nodes is calculated, and directed edges are established for related nodes. Finally, the graph structure data of the risk node is constructed by nodes and edges. The generated graph structure data is transmitted as input to the risk prediction module;

[0073] Specifically, the Bayesian network structure learning algorithm is used, and the K2 algorithm is taken as an example to establish a directed acyclic connection relationship for each node to form a set {x1, x2, … x n} in order. The first i-1 nodes are selected as candidate parent nodes of node x i , and the Bayesian information criterion score of the Bayesian network structure formed by node x i and each candidate parent node is calculated. The score of the highest candidate parent node after joining is compared with the score without joining the candidate parent node. If the score increases, the node is taken as the parent node of node i, and the edge from the parent node to node i is added. If the score does not increase, the node is discarded from the candidate parent node set. When the score after adding the parent node no longer increases or reaches the in-degree of node i, the process stops. Otherwise, the candidate parent node of node x i is selected repeatedly, and the specific process is shown in Figure 2 . According to the actual situation, the connection relationship between the generated graph structure nodes is adjusted and optimized (the graph structure generated by the Bayesian network may have a risk association relationship that does not conform to the actual airport operation situation, and the graph structure needs to be adjusted), and the finally obtained risk relationship graph structure is shown in Figure 3 .

[0074] Step S3: The weight generation module calculates the relevance of the associated nodes of the node and generates the corresponding edge weight according to the strength of the relevance. The stronger the relevance, the higher the weight. The generated edge weight is transmitted as input to the risk prediction module.

[0075] Specifically, the Pearson linear correlation coefficient (PLCC) is used to calculate the correlation between the data of the connected nodes in the graph structure, as follows.

[0076]

[0077] wherein l represents the PLCC value of the two nodes, s i represents the risk value of the node at the i-th moment, represents the average risk value of the node at all moments, p i represents the risk value of the associated node at the i-th moment, represents the average risk value of the associated node at all moments, and N is the selected time period. The edge weight of each connected node is determined on the basis of the PLCC, as follows.

[0078]

[0079] wherein w ij represents the edge weight of the two connected nodes i, j, l ij represents the PLCC value of the two nodes, and S represents the set of all parent nodes associated with node j.

[0080] Step S4: The risk prediction module takes the historical risk data input by the data acquisition module as the input feature value of each node in the graph structure, takes the edge weight input by the weight generation module as the input feature value of each directed edge in the graph structure, and uses a graph neural network to predict the risk value of each risk node at the next or several moments. The graph neural network is a method for processing graph structure data combined with neural networks and deep learning. One of its typical applications is the node classification problem. In this type of problem, node i has some features, and the label corresponding to node i is t iThen given a part of the graph G with node label information, the graph neural network is trained with the labeled nodes as training set to predict the label of unlabeled nodes (reference: Defferrard, M.; Bresson, X.; and Vandergheynst, P. 2016. Convolutional neural networks on graphs with fast localized spectral filtering. In Advances in neural information processing systems, 3844-3852.).

[0081] Specifically, the graph neural network learns a d-dimensional vector x n for each node through graph embedding, which contains the information of the current node and the connected nodes in x n .

[0082] x n = f w (l n ,l co[n] ,x ne[n] ,l ne[n] )(3)

[0083] wherein, l n represents the input feature of the node, i.e. the risk value of the node at the previous 10 time points, l co[n] represents the feature of the edge connected to the node, i.e. the edge weight obtained by the weight generation module, x ne[n] represents the embedding vector of the node connected to the node, and l ne[n] represents the feature of the node connected to the node, i.e. the historical risk value of the node connected to the node. The output value of each node is calculated as follows.

[0084] o n = g w (x n ,l n )(4)

[0085] wherein, o n is the output value of the node, i.e. the predicted risk value of the node at the next time or several times;

[0086] Step S5: verifying the prediction accuracy of the graph neural network, if the accuracy requirement is not met, more running risk data is obtained in step S1, and the graph neural network is continuously trained until the prediction result meets the accuracy requirement.

[0087] Example 2:

[0088] This embodiment addresses airport operational risks by employing a graph neural network for risk prediction, including the following steps:

[0089] Step S1: The data acquisition module collects relevant data on the risk levels and operational status of various risks during airport operations, including: runway incursion risk, taxiway conflict risk, restricted area incursion risk, low visibility risk, low-altitude wind shear risk, pavement rain risk, pavement snow risk, pavement icing risk, reduced friction coefficient risk, route deviation risk, apron traffic incident risk, aircraft-to-aircraft collision risk, aircraft-to-vehicle scraping risk, aircraft-to-personnel scraping risk, and aircraft-to-building scraping risk, as well as status data reflecting the airport's operational busyness, such as runway utilization rate, taxiway utilization rate, and vehicle utilization rate. The data acquisition module cleans the data, removing missing, erroneous, and duplicate data, and divides the collected data into fixed time periods to obtain time-series data of airport operational risks. This data is then input into the graph structure construction module, weight generation module, and risk prediction module.

[0090] Step S2: The graph structure construction module uses operational risks and operational status as nodes to construct graph structure data. It iterates through all nodes based on the time-series data input to each node, calculates the correlation between every two nodes, and establishes directed edges for related nodes. Finally, the graph structure data of the risk nodes is constructed from the nodes and edges. The generated graph structure data is then passed as input to the risk prediction module.

[0091] Specifically, sorting all nodes yields the set {x1, x2, ..., x}. n Starting from node i, calculate the Pearson linear correlation coefficient (PLCC) between node i and nodes i+1 to n, and select nodes that exceed the set threshold as connected nodes to node i, as shown below.

[0092]

[0093] Among them, e ij This represents the PLCC value of node i relative to other nodes. This represents the risk value of node i over N historical time points. This represents the average risk value of node i over N historical time points. This represents the risk value of nodes other than node i at N historical time points. Let A represent the average risk value of nodes other than node i over N historical time points. ij= 1 indicates that node i is connected with node j, TH represents a set threshold value, and S represents a set of nodes except node i+1 to node n. In turn, until i = n stops, the connection relationship of all nodes is obtained. After determining the connection relationship, the child node and the parent node are determined according to the risk attribute of the connected two nodes, and the process is as shown in Figure 4 , and the finally obtained risk relationship diagram structure is as shown in Figure 3 .

[0094] Step S3: The weight generation module calculates the relevance of the associated nodes of the node as the center, and generates the corresponding edge weight according to the strength of the relevance. The stronger the relevance, the higher the weight, and the generated edge weight is transmitted as input to the risk prediction module.

[0095] Specifically, the Pearson linear correlation coefficient (PLCC) is used to calculate the relevance between the data of the connected nodes in the graph structure, as follows.

[0096]

[0097] wherein, l represents the PLCC value of two nodes, s i represents the risk value of node i moment, represents the average risk value of all moments of the node, p i represents the risk value of the associated node i moment, represents the average risk value of all moments of the associated node, and N is the selected time period. On the basis of PLCC, the edge weight of each connected node is determined, as follows.

[0098]

[0099] wherein, w ij represents the edge weight of two connected nodes i, j, l ij represents the PLCC value of two nodes, and S represents a set of all parent nodes associated with node j.

[0100] Step S4: The risk prediction module takes the historical risk data input by the data acquisition module as the input feature value of each node in the graph structure, takes the edge weight input by the weight generation module as the input feature value of each directed edge in the graph structure, and uses the graph neural network to predict the risk value of each risk node at the next or several moments.

[0101] Specifically, the graph neural network learns the d-dimensional vector x n of each node through graph embedding, and the information of the current node and the connected nodes is contained in x n , and the calculation formula is as follows.

[0102] xn =f w (l n ,l co[n] ,x ne[n] ,l ne[n] ) (3)

[0103] wherein, l n represents the input feature of the node, i.e., the node risk value at the historical 10 time points, l co[n] represents the feature of the edge connected to the node, i.e., the edge weight obtained by the weight generation module, x ne[n] represents the embedding vector of the node connected to the node, l ne[n] represents the feature of the node connected to the node, i.e., the historical risk value of the node connected to the node. The output value of each node is finally calculated as follows.

[0104] o n =g w (x n ,l n ) (4)

[0105] wherein, o n is the output value of the node, i.e., the predicted risk value of the node at the next time or several time points;

[0106] Step S5: verifying the prediction accuracy of the graph neural network, if the accuracy requirement is not met, more running risk data is acquired in step S1, the graph neural network is continuously trained until the prediction result meets the accuracy requirement.

[0107] The method provided by the application can be applied to the existing airport flight area surface operation monitoring system, generate the risk value of the observation node at the future time, form a risk data file in json format, and give early warning information to the observation node exceeding the risk threshold on the display screen of the existing airport flight area surface operation monitoring system. In airport operation, airport operation personnel can discover nodes that may produce risks in time according to the risk data of each observation node generated by the system, and take corresponding safety management measures to eliminate hidden dangers and avoid risks, thereby improving airport operation safety.

[0108] In a specific implementation, the application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium can store a computer program, and the computer program can run part or all of the steps in the invention content and each embodiment of the airport operation risk time sequence prediction system based on the graph neural network when executed by the data processing unit. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0109] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present application can be realized by means of a computer program and the corresponding general hardware platform. Based on such understanding, the technical solutions in the embodiments of the present application can be embodied in the form of a computer program, i.e. a software product, which can be stored in a storage medium, including a plurality of instructions for causing a device (which can be a personal computer, a server, a single-chip microcomputer, a MUU or a network device, etc.) comprising a data processing unit to execute the method described in various embodiments or some parts of the embodiments of the present application.

[0110] The present application provides a kind of based on the thinking and method of airport operation risk time series prediction system of graph neural network, the method and approach for specifically realizing this technical solution are many, above-mentioned only is preferred embodiment of the present application, it should be pointed out, for the ordinary skilled in the art, on the premise of not departing from the principle of the present application, can also make a number of improvements and refinements, these improvements and refinements also should be regarded as the protection scope of the present application. The components not explicitly described in the embodiments can be realized by existing technology.

Claims

1. An airport operation risk time series prediction system based on a graph neural network, characterized in that, The system comprises a data collection module, a graph structure construction module, a weight generation module, and a risk prediction module. The data collection module is configured to collect data of airport operation risks and operation situations, clean the data, and divide the cleaned data into time series data according to time periods, and then transmit the time series data to the graph structure construction module, the weight generation module, and the risk prediction module. The graph structure construction module is configured to analyze the time series data, establish a connection relationship between risks, form a directed acyclic graph structure with various risks as nodes, and transmit the directed acyclic graph structure to the risk prediction module. The weight generation module is configured to calculate data correlation between risk nodes according to the time series data, generate edge weights between nodes in the directed acyclic graph structure, and transmit the edge weights to the risk prediction module. The risk prediction module is configured to process graph structure data by using a graph neural network prediction method according to time series data of each risk node in the directed acyclic graph structure and the edge weights, and complete prediction of risk values of each risk node at a future time.

2. The airport operation risk time series prediction system based on a graph neural network according to claim 1, characterized in that, The system performs the following steps: In step S1, the data collection module collects risk levels and related data of operation situations of various risks in the airport operation process, cleans the data to remove missing data, error data, and duplicate data, divides the cleaned data into time series data of airport operation risks according to fixed time periods, and transmits the time series data to the graph structure construction module, the weight generation module, and the risk prediction module. In step S2, the graph structure construction module constructs graph structure data by taking the cleaned data in step S1 as nodes of the graph, and transmits the generated graph structure data to the risk prediction module. In step S3, the weight generation module calculates the correlation of associated nodes with a node as the center, generates corresponding edge weights according to the strength of the correlation, and transmits the generated edge weights to the risk prediction module. In step S4, the risk prediction module takes the time series data input by the data collection module as input feature values of each node in the graph structure data, takes the edge weights input by the weight generation module as input feature values of each directed edge in the graph structure data, and uses a graph neural network to predict risk values of each risk node at a future time. In step S5, the accuracy of the prediction of the graph neural network is verified, and if the accuracy requirement is not met, steps S1 to S4 are repeatedly executed until the prediction result of the graph neural network meets the accuracy requirement.

3. The airport operation risk time series prediction system based on a graph neural network according to claim 2, characterized in that, The related data in step S1 at least includes risk values of runway incursion risk, taxiway conflict risk, restricted area incursion risk, low-visibility risk, low-altitude wind shear risk, runway area rain risk, runway area snow risk, runway area ice risk, friction coefficient reduction risk, route deviation risk, apron traffic event risk, aircraft-aircraft collision risk, aircraft-vehicle collision risk, aircraft-person collision risk, and aircraft-building collision risk, and situation data of runway utilization rate, taxiway utilization rate, and vehicle utilization rate.

4. The airport operation risk time series prediction system based on a graph neural network according to claim 3, characterized in that, The method for constructing the graph structure data in step S2 comprises: The Bayesian network structure learning algorithm is used to establish a directed and non-cyclic connection relationship for each node, and then the connection relationship between the nodes of the generated graph structure is adjusted and optimized according to the risk attributes of the nodes, so as to obtain a risk relationship graph structure and complete the construction of the graph structure data.

5. The airport operation risk time series prediction system based on a graph neural network according to claim 3, characterized in that, The method for constructing the graph structure data in step S2 comprises the following steps: A directed edge is established between the related nodes in the nodes of the graph, and finally the graph structure data of the risk nodes is constructed by the nodes and the directed edges, and specifically comprises the following steps: The correlation between each target node and all other nodes is calculated, the nodes with a correlation exceeding a set threshold are selected as the connected nodes of the target node, a connection is established between the connected nodes, and after the connection is established, the child nodes and the parent nodes are determined according to the risk attributes of the connected nodes, i.e. the directed edge is obtained.

6. The airport operation risk time series prediction system based on a graph neural network according to claim 4 or 5, characterized in that, The method for calculating the edge weight in step S3 comprises the following steps: For any target node, the correlation between the target node and the connected nodes in the graph structure is calculated, and the ratio of the correlation value to the sum of the correlation values of all connected nodes of the target node is taken as the edge weight between the node and the target node; all nodes are traversed to perform the above calculation to obtain the edge weights of all connected nodes.

7. The airport operation risk time series prediction system based on a graph neural network according to claim 6, characterized in that, The method for calculating the correlation between the target node and the connected nodes in the graph structure in step S3 comprises the following steps: The Pearson linear correlation coefficient is used to calculate the correlation l, and the method is as follows: wherein, l represents the Pearson linear correlation coefficient value of the target node s and the node p connected therewith, s i represents the risk value of the node s at the i th moment, represents the average risk value of the target node at all moments, p i represents the risk value of the node p at the i th moment, represents the average risk value of the node p at all moments, and N is the selected time period.

8. The airport operation risk time series prediction system based on a graph neural network according to claim 7, characterized in that, The method for calculating the edge weight in step S3 comprises the following steps: where w sp denotes the edge weight between target node s and its connected node p, l sp denotes the Pearson linear correlation coefficient value between two nodes, and set S denotes the set of all parent nodes associated with node s.

9. The airport operation risk time series prediction system based on a graph neural network according to claim 8, characterized in that, The method for predicting the risk values of the risk nodes at a future time point by using the graph neural network in step S4 comprises the following steps: The graph neural network learns a d-dimensional vector x for each node by a graph embedding method n , x n contains information of the current node and the connected nodes, and the calculation method is as follows: x n = f w (l n , l co[n] , x ne[n] , l ne[n] ) wherein, l n represents the input feature of the node, i.e., the historical risk value of the node at the time, l co[n] represents the feature of the edge connected to the node, i.e., the edge weight described in step S3, x ne[n] represents the embedding vector of the node connected to the node, l ne[n] represents the feature of the node connected to the node, i.e., the historical risk value of the node connected to the node, f w represents the conversion function of the graph neural network; finally, the output value of each node is calculated by the following method: o n = g w (x n , l n ) wherein o n is the output value of the node, i.e. the predicted risk value of the node at the next time or times, g w denotes the output function of the graph neural network.

10. The airport operation risk time series prediction system based on a graph neural network according to claim 5, wherein, The method for calculating the correlation between each target node and all other nodes in step S2 comprises the following steps: The Pearson linear correlation coefficient is used to calculate the correlation between the node s and all other nodes, and the nodes exceeding the set threshold are selected as the connected nodes of the node s, and the specific method is as follows: A sp = 1, {e sp ≥ TH, p e SS} where e sp represents the Pearson linear correlation coefficient value between node s and node p, represents the risk value of node s at the nth time in history, represents the average risk value of node s at N times in history, represents the risk value of other nodes except node s at N times in history, represents the average risk value of other nodes except node s at N times in history, A sp = 1 represents that node s is connected with node p, TH represents a set threshold value, and SS represents a set of other nodes except node s.

Images