A trajectory prediction and safety assessment method based on improved spatiotemporal graph convolution model

Abstract

The present invention provides a track prediction and safety assessment method based on an improved spatiotemporal graph convolution model. A four-dimensional track data set is constructed by cleaning, completing and normalizing the data collected by the Automatic Dependent Surveillance-Broadcast (ADS-B) system. The dynamic identification spatiotemporal graph convolutional neural network (DI-STGCN) is used to predict the track of the airport terminal area, and the accuracy of the model is evaluated by visual analysis and error measurement indicators. At the same time, prior data is obtained through questionnaire surveys, and the posterior safety probability distribution is obtained using a Bayesian neural network. Based on the above research content, a flight safety assessment system is constructed to achieve accurate prediction of flight safety.

Description

Technical Field

[0001] The present invention relates to the field of four-dimensional track prediction, and in particular to a track prediction and safety assessment method based on an improved spatiotemporal graph convolution model. Background Art

[0002] Four-dimensional trajectory prediction is a research focus in the field of civil aviation air traffic management, and has important research significance in the rational arrangement of air traffic. Civil aviation air traffic management includes a variety of prediction methods (such as Kalman filter, particle filter, long short-term memory network, etc.), and the methods for refined air traffic management are still being improved.

[0003] Spatial temporal graph convolutional networks (STGCN) is a deep learning model specifically for modeling spatiotemporal data. By introducing a graph structure and combining the idea of ​​graph convolutional networks (GCN), it can effectively analyze spatiotemporal data, especially in areas such as flight trajectory prediction. STGCN also performs well in tasks such as action recognition and video analysis, and is widely used in sensor data, behavior recognition and other fields, with important development prospects.

[0004] The application of Bayesian method in safety management mainly includes risk assessment, safety decision support, event prediction and safety monitoring improvement. It uses prior information and observation data, and through probabilistic inference, provides reliable safety assessment and management means, providing important support for the safe operation of the civil aviation system. Summary of the invention

[0005] Purpose of the invention: The technical problem to be solved by the present invention is to provide a track prediction and safety assessment method based on an improved spatiotemporal graph convolution model in view of the deficiencies of the prior art, including:

[0006] Step 1: Process and clean the data collected by the automatic dependent surveillance-broadcast system ADS-B to construct a four-dimensional track data set;

[0007] Step 2: Use dynamic identification-spatial temporal graph convolutional networks (DI-STGCN) to predict the flight path of the airport terminal area.

[0008] Step 3: Evaluate the safety of the aircraft trajectory based on the Bayesian method.

[0009] Step 1 includes the following steps:

[0010] Step 1-1, perform timestamp processing on the data, convert the time data into seconds in a day, and then perform Min-Max normalization processing on the time data and the three-dimensional position point coordinates;

[0011] Step 1-2: remove duplicate data and use DBSCAN to cluster the data to obtain a valid data set G C With the noisy dataset G O , respectively retain and eliminate;

[0012] Steps 1-3 use Newton interpolation to complete the data, and perform three Newton interpolation processes on time and latitude, time and longitude, and time and altitude respectively. The formula based on Newton interpolation polynomial is:

[0013]

[0014] where t i is the time, y0 is the dimension value at time t0, C i For the difference, Indicates n * +1st order interpolation polynomial; t is the time corresponding to the track point to be interpolated, j = 0, 1, 2…n * -1, i=1,2,…n * ;

[0015] Steps 1-4: construct the data set and perform Min-Max normalization on the data.

[0016] Steps 1-2 include:

[0017] Extract the i-th track point P from the existing track data i The value x of the corresponding field i , all x i The constructed data set D is represented as D = {x1, x2, ..., x n}, define the local length δ, the neighborhood distance threshold ε and the number of points in the core point neighborhood threshold η according to the needs of forming clusters and eliminating data, where δ and η need to satisfy η≤2δ; the parameter ε is the dimension of the value, η=2ε, and δ is the value to try until a stable data set is formed. The parameter values ​​selected in the invention are ε=3, η=6, and δ is determined by the algorithm according to the specific data. The DBSCAN algorithm uses local traversal, and the clustering efficiency is greater than the method of clustering all data.

[0018] In steps 1-4, the following formula is used for normalization:

[0019]

[0020] Where Xmax Indicates the maximum value of the data, x min represents the minimum value of the data, x is the original data, X * is the normalized data.

[0021] Step 2 includes:

[0022] Step 2-1, the DI-STGCN network architecture outputs the feature vector in the graph of M time steps And the corresponding adjacency matrix W∈R n , where N is the number of nodes, Z i is the feature dimension at each time step, R n Represents n-dimensional Euclidean space, where n represents the dimension of the matrix; after two or more spatiotemporal convolution blocks and the final single output layer, the output In order to predict the features of a specific time step after t time steps, The output vector obtained by the neural network predicting the features at a specific time step;

[0023] Step 2-2, the DI-STGCN model allows implicit feature extraction, which can automatically determine the neighbor nodes and time steps that are most critical for prediction;

[0024] In step 2-3, the convolution layer in DI-STGCN is used to capture the evolution of spatiotemporal information, considering the connection relationship between nodes and their neighbors, so as to realize information transfer on the graph structure. The spatiotemporal convolution operation is expressed as:

[0025]

[0026] Among them, H′ is the new feature representation obtained after the spatiotemporal convolution operation, H is the input feature matrix, A is the adjacency matrix, V is the degree matrix, W is the weight matrix, and σ is the activation function;

[0027] In step 2-4, the pooling layer in DI-STGCN is used for feature dimensionality reduction and feature selection, which reduces the spatial and temporal dimensions of the features, reduces the computational complexity, and improves the generalization ability of the model;

[0028] Step 2-5: Perform spatiotemporal modeling. In the temporal domain, several graph convolutional layers are stacked to achieve hierarchical extraction of temporal information. In the spatial domain, the adjacency matrix is ​​used to model the relationship between spatial nodes. Graph convolutional layers are applied alternately in the temporal and spatial order, while considering spatiotemporal information.

[0029] In steps 2-6, DI-STGCN designs parameter combinations for different flight altitude intervals and performs parallel training to capture the characteristics of track data at different altitudes.

[0030] Steps 2-5 include the following steps:

[0031] Step 2-5-1, the time domain convolution block is implemented by inputting U is a two-dimensional tensor, and a one-dimensional convolution is performed along the time dimension, where the number of convolution kernels is 2Z0. Get the result Then use the GLU (Gated Linear Unit) activation function to activate the result, and get the following result:

[0032]

[0033] Among them, P and Q are partial feature vectors output by the time domain convolution block, K t represents the length of the convolution kernel; τ represents the parameter tensor used for the convolution operation, τ *τ Represents the new convolution kernel tensor obtained by operating the convolution kernel τ with itself;

[0034] In a complete space-time graph, input Get the result

[0035] Step 2-5-2, the spatial convolution block is performed on the graph at each time step, by inputting The output Y is:

[0036]

[0037] in:

[0038] T i (U)=2uT i-1 (u)-T i-2 (u),

[0039]

[0040]

[0041] Among them, K represents the number of convolution kernels, θ i represents the weight parameter, T i represents the transposed matrix, represents the normalized Laplacian matrix, I n represents the n*n identity matrix, L represents the Laplace matrix, λ max represents the maximum eigenvalue in the Laplace matrix L, and u is a single element in U;

[0042] The output of the spatial convolution block is an adjacency matrix, and the convolution kernel There are Z0 in total; in a complete space-time graph, input Get the result

[0043] Steps 2-6 include the following steps:

[0044] Step 2-6-1, divide the data into different height intervals, and use different parameter groups for training for different height intervals, including the number of convolution kernels nb_filters, the number of STGCN levels levels and the convolution kernel size kernel_size;

[0045] Step 2-6-2, DI-STGCN trains more than two STGCN models at the same time, and each model is trained using different parameter groups for data features in different height ranges;

[0046] Step 2-6-3, after the training is completed, DI-STGCN integrates the models trained in each height interval into one model to predict and process the data in different height intervals.

[0047] Step 3 includes the following steps:

[0048] Step 3-1, obtaining experts’ opinions on airport trajectory prediction and safety through questionnaire survey;

[0049] Step 3-2, calculate the prior probability of the Bayesian method based on the questionnaire data;

[0050] Step 3-3, through dynamic identification, for different track data sets in the flight phase, the predicted data obtained by DI-STGCN is used as the observation data, combined with the prior probability, and the posterior safety probability of the corresponding track is obtained through the Bayesian neural network;

[0051] Step 3-4, based on the a posteriori safety probability and combined with the entropy weight method to determine the weights, a flight safety assessment system is constructed for the flight stage.

[0052] Step 3-4 includes the following steps:

[0053] Step 3-4-1, combined with the posterior safety probability distribution, based on the probability, divide the safety intervals of longitude, latitude, and altitude into five levels according to the probability of occurrence, and assign values ​​1 to 5. The higher the score, the higher the safety.

[0054] Step 3-4-2, combined with the real track data, the importance weights of longitude, latitude and altitude in the flight process are determined by entropy weight method, and the corresponding weights w1, w2 and w3 are obtained respectively;

[0055] Step 3-4-3, calculate the safety score of each flight's track point according to the following formula:

[0056] B=w1X1+w2X2+w3X3

[0057] Among them, X1 is the longitude matrix, X2 is the latitude matrix, X3 is the altitude matrix, and B is the score matrix.

[0058] The present invention also provides a storage medium storing a computer program or instruction. When the computer program or instruction is executed, the method for trajectory prediction and safety assessment based on the improved spatiotemporal graph convolution model is implemented.

[0059] The present invention studies a method for trajectory prediction and safety assessment based on an improved spatiotemporal graph convolution model that combines deep learning and probabilistic inference. It can not only predict the trajectory more accurately, but also make up for the lack of dynamic analysis in the aviation field, and has certain practicality and innovation. By performing graph structure modeling and neural network training on spatiotemporal data, this technology can better capture spatiotemporal information and improve the accuracy and reliability of flight predictions. In terms of achieving the goal of "four-type airports", this technology can improve airport safety, intelligence and efficiency, and has a certain driving effect and reference value on the sustainable development of the aviation transportation system.

[0060] Beneficial effects: The present invention adopts a trajectory prediction and safety assessment method based on an improved spatiotemporal graph convolution model. Compared with traditional methods, this innovative technology not only integrates the spatiotemporal graph convolution algorithm and the calculation of the Bayesian network, but also has the ability to perform online and offline predictions according to actual needs. It is flexible and reliable and is an innovative technology. This method can objectively reflect the various factors that affect the trajectory, and perform dynamic predictions based on the flight phase and some operating data to provide a reference for the takeoff and landing of flights. This method that combines deep learning and probabilistic inference not only improves the accuracy and efficiency of aviation spatiotemporal data analysis and trajectory prediction, but also provides strong support for air traffic management. The comprehensive advantages of this technology will help promote the development of safety management and operational efficiency in the aviation field in a more sustainable direction. Flight trajectory data and risk prediction values ​​can be predicted for flight conditions at different stages, which will help airports better arrange air traffic in the approach area, and to a certain extent, it has a reference significance for the takeoff and landing of flights. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0062] Figure 1 It is a comprehensive algorithm flow.

[0063] Figure 2 This is the algorithm flow of the spatiotemporal graph convolutional neural network (STGCN).

[0064] Figure 3 This is the Bayesian neural network algorithm process.

[0065] Figure 4 This is a comparison curve between the three-dimensional track prediction value and the actual value during the descent cruise phase.

[0066] Figure 5 is the posterior safety probability density histogram of flight longitude.

[0067] Figure 6 It is a comparative histogram of flight safety after Bayesian safety assessment.

[0068] Figure 7 Histogram of Bayesian safety assessment scores for different flight phases. DETAILED DESCRIPTION

[0069] The present invention provides a track prediction and safety assessment method based on an improved spatiotemporal graph convolution model, such as Figure 1 As shown, the specific steps include:

[0070] (1) Process and clean Automatic Dependent Surveillance-Broadcast (ADS-B) data to construct a four-dimensional track data set;

[0071] (2) Use the spatiotemporal graph convolutional neural network (DI-STGCN) to predict the flight path in the airport terminal area;

[0072] (3) Evaluate the safety of aircraft trajectory based on Bayesian method, such as Figure 3 shown.

[0073] The track data set constructed in step (1) includes the IATA flight number, the ICAO flight number, the departure airport, the landing airport, the time, and the longitude, latitude, altitude, and speed at the corresponding time. The specific initial track data set example is shown in Table 1:

[0074] Table 1

[0075]

[0076] Step (1) The specific steps are as follows:

[0077] 1.1) Perform timestamp processing on the data, convert the time data into seconds in a day, and then perform Min-Max normalization processing on the time data and the three-dimensional position point coordinates;

[0078] 1.2) Remove duplicate data, use DBSCAN for clustering, and extract track points P from the existing track data i The value x of the corresponding field i , all x i Construct a data set D = {x1, x2, ..., x n}, define the local length δ, neighborhood distance threshold ε and number of points in the core point neighborhood threshold η according to the needs of forming clusters and removing data, where δ and η need to satisfy η≤2δ. Parameter ε is the dimension of the value, η=2ε, and δ is the value to try.

[0079] To form a stable data set, the parameter values ​​selected in the present invention are ε=3, η=6, and δ is determined by the algorithm according to the specific data. The effective data set G is obtained C With the noisy dataset G O , respectively retain and remove; this DBSCAN algorithm uses local traversal, and the clustering efficiency is greater than the method of clustering all data. ;

[0080] 1.3) Newton interpolation method is used to complete the data. Three Newton interpolation processes are performed on time and latitude, time and longitude, and time and altitude respectively. The formula based on Newton interpolation polynomial is as follows:

[0081]

[0082] where t i is the time, y0 is the dimension value at time t0, C i For the difference, Indicates n * +1st order interpolation polynomial; t is the time corresponding to the track point to be interpolated, j = 0, 1, 2…n * -1, i=1,2,…n * ;

[0083] 1.4) Construct the data set and perform Min-Max normalization on the data. The specific formula is as follows:

[0084]

[0085] where x max Indicates the maximum value of the data, x min represents the minimum value of the data, x is the original data, X * is the normalized data;

[0086] The data set after the above data processing and cleaning is shown in Table 2:

[0087] Table 2

[0088]

[0089] Among them, Figure 2 As shown, the specific principle of step (2) is as follows:

[0090] 2.1) DI-STGCN network architecture outputs the feature vector in the graph of M time steps And the corresponding adjacency matrix W∈R n , where N is the number of nodes, Z i is the feature dimension at each time step, R n Represents n-dimensional Euclidean space, where n represents the dimension of the matrix. After multiple spatiotemporal convolution blocks and the final single output layer, the output is In order to predict the features of a specific time step after t time steps, The output vector obtained by the neural network predicting the features at a specific time step;

[0091] 2.2) The core of this DI-STGCN model allows implicit feature extraction, which can automatically determine the most critical neighbor nodes and time steps for prediction;

[0092] 2.3) The convolution layer in DI-STGCN is mainly used to capture the evolution of spatiotemporal information, considering the connection relationship between nodes and their neighbors, so as to realize information transfer on the graph structure. Specifically, the spatiotemporal convolution operation is expressed as:

[0093]

[0094] Among them, H is the input feature matrix, H′ is the new feature representation obtained after the spatiotemporal convolution operation, A is the adjacency matrix, V is the degree matrix, W is the weight matrix, and σ is the activation function;

[0095] 2.4) The pooling layer in DI-STGCN is mainly used for feature dimensionality reduction and feature selection, which reduces the spatial and temporal dimensions of the features, reduces the computational complexity, and improves the generalization ability of the model;

[0096] 2.5) Perform spatiotemporal modeling. In the temporal domain, several graph convolutional layers are stacked to achieve hierarchical extraction of temporal information. In the spatial domain, the adjacency matrix is ​​used to model the relationship between spatial nodes. Graph convolutional layers are applied alternately in the temporal and spatial order, while considering spatiotemporal information. The specific principles are as follows:

[0097] 2.5.1) The time domain convolution block is implemented by inputting U is a two-dimensional tensor, and a one-dimensional convolution is performed along the time dimension, where the number of convolution kernels is 2Z0. Get the result Then use the GLU (Gated Linear Unit) activation function to activate the result, and get the following result:

[0098]

[0099] Among them, P and Q are partial feature vectors output by the time domain convolution block, K t Represents the length of the convolution kernel. τ represents the parameter tensor used for the convolution operation, τ*τ Represents the new convolution kernel tensor obtained by operating the convolution kernel τ with itself.

[0100] In a complete space-time graph, input Results available

[0101] 2.5.2) The spatial convolution block is performed on the graph at each time step, by inputting The output is as follows:

[0102]

[0103] T i (U)=2uT i-1 (u)-T i-2 (u),

[0104]

[0105]

[0106] Among them, K represents the number of convolution kernels, θ i represents the weight parameter, T i represents the transposed matrix, represents the normalized Laplacian matrix, O n represents the n*n identity matrix, L represents the Laplace matrix, λ max represents the largest eigenvalue in the Laplace matrix L, and u is a single element in U.

[0107] Its output is the adjacency matrix, the convolution kernel There are Z0 in total; in a complete space-time graph, input Get the result

[0108] 2.6) DI-STGCN designs parameter combinations for different intervals of flight altitude and performs parallel training to capture the characteristics of aviation data at different altitudes. The specific principles are as follows:

[0109] 2.6.1) Divide the data into different height intervals, as shown in Table 3:

[0110] Table 3

[0111]

[0112] Different parameter groups are used for training at different height intervals, including the number of convolution kernels (nb_filters), the number of STGCN layers (levels) and the size of the convolution kernel (kernel_size). The parameter combinations are as follows:

[0113] nb_filters=32, levels=4, kernel_size=3, if h∈H1

[0114] nb_filters=64, levels=5, kernel_size=3, if h∈H2

[0115] nb_filters=128, levels=6, kernel_size=3, if h∈H3

[0116] nb_filters=256, levels=7, kernel_size=3, if h∈H4

[0117] 2.6.2) DI-STGCN can train multiple STGCN models at the same time, and each model uses different parameter groups to train data features in different height ranges;

[0118] 2.6.3) After training, DI-STGCN integrates the models trained in each height interval into one model to predict and process data in different height intervals;

[0119] 2.7) Visualize the training results and compare the predicted values ​​with the true values. Figure 4 shown.

[0120] Among them, the specific steps of step (3) are as follows:

[0121] 3.1) Obtain experts’ opinions on airport trajectory prediction and safety through questionnaire survey;

[0122] 3.2) Calculate the prior probability of the Bayesian method based on the questionnaire data;

[0123] 3.3) Through dynamic identification, the prediction data obtained by DI-STGCN is used as the observation data for the different track data sets in the flight phase. Combined with the prior probability, the posterior safety probability of the corresponding track is obtained through the Bayesian neural network. Taking longitude as an example, the frequency distribution with respect to longitude is as follows: Figure 5 The frequencies corresponding to the intervals are shown in Table 4:

[0124] Table 4

[0125]

[0126]

[0127] 3.4) Based on the a posteriori safety probability and the frequency, a scoring system is constructed for the track data of a certain dimension in proportion; combined with the real track data, the importance weights of longitude, latitude and altitude in the flight process are determined by the entropy weight method. The weights determined by the example are shown in Table 5:

[0128] Table 5

[0129]

[0130] The entropy weight method is used to determine the weights, and the scoring system is combined with the weights of each dimension to construct a flight safety assessment system;

[0131] 3.5) If Figure 6 As shown, for different flights, the corresponding safety score is calculated according to the following formula.

[0132] B=w1X1+w2X2+w3X3

[0133] Among them, X1 is the longitude matrix, X2 is the latitude matrix, X3 is the altitude matrix, and B is the score matrix.

[0134] 3.6) If Figure 7 As shown, for different flight phases, the corresponding safety scores are calculated.

[0135] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium can store a computer program, and when the computer program is executed by the data processing unit, the invention content of a method for trajectory prediction and safety assessment based on an improved spatiotemporal graph convolution model provided by the present invention and some or all of the steps in each embodiment can be run. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0136] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of computer programs and their corresponding general hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention are essentially or partly contributed to the prior art can be embodied in the form of a computer program, i.e., a software product, which can be stored in a storage medium and includes several instructions for enabling a device including a data processing unit (which can be a personal computer, a server, a single-chip microcomputer, a MUU or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0137] The present invention provides a method for track prediction and safety assessment based on an improved spatiotemporal graph convolution model. There are many methods and approaches to implement the technical solution. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention. All components not specified in this embodiment can be implemented using existing technologies.

Claims

1. A track prediction and safety assessment method based on an improved spatiotemporal graph convolution model, characterized in that: The following steps are involved: Step 1: Process and clean the data collected by the automatic dependent surveillance-broadcast system ADS-B to construct a four-dimensional track data set; Step 2: Use the dynamic identification spatiotemporal graph convolutional neural network DI-STGCN to predict the flight path of the airport terminal area; Step 2 includes: Step 2-1, the DI-STGCN network architecture outputs the feature vector in the graph of M time steps And the corresponding adjacency matrix E∈R n , where N is the number of nodes, Z i is the feature dimension at each time step, R n Represents n-dimensional Euclidean space, where n represents the dimension of the matrix; after two or more spatiotemporal convolution blocks and the final single output layer, the output In order to predict the features of a specific time step after t time steps, The output vector obtained by the neural network predicting the features at a specific time step; Step 2-2, the DI-STGCN model allows implicit feature extraction, which can automatically determine the neighbor nodes and time steps that are most critical for prediction; In step 2-3, the convolution layer in DI-STGCN is used to capture the evolution of spatiotemporal information, considering the connection relationship between nodes and their neighbors, so as to realize information transfer on the graph structure. The spatiotemporal convolution operation is expressed as: Among them, H′ is the new feature representation obtained after the spatiotemporal convolution operation, H is the input feature matrix, A is the adjacency matrix, V is the degree matrix, W is the weight matrix, and σ is the activation function; In step 2-4, the pooling layer in DI-STGCN is used for feature dimensionality reduction and feature selection, reducing the spatial and temporal dimensions of the features; Step 2-5: Perform spatiotemporal modeling. In the temporal domain, several graph convolutional layers are stacked to achieve hierarchical extraction of temporal information. In the spatial domain, the adjacency matrix is ​​used to model the relationship between spatial nodes. Graph convolutional layers are applied alternately in the temporal and spatial order, while considering spatiotemporal information. Step 2-6, DI-STGCN designs parameter combinations for different intervals of flight altitude and performs parallel training to capture the characteristics of track data at different altitudes; Step 3: Evaluate the safety of the aircraft trajectory based on the Bayesian method.

2. The method according to claim 1, characterized in that Step 1 includes the following steps: Step 1-1, perform timestamp processing on the data, convert the time data into seconds in a day, and then perform Min-Max normalization processing on the time data and the three-dimensional position point coordinates; Step 1-2: remove duplicate data and use DBSCAN to cluster the data to obtain a valid data set G C With the noisy dataset G O , respectively retain and eliminate; Steps 1-3 use Newton interpolation to complete the data, and perform three Newton interpolation processes on time and latitude, time and longitude, and time and altitude respectively. The formula based on Newton interpolation polynomial is: where t i is the time, y0 is the dimension value at time t0, C i For the difference, Indicates n * +1st order interpolation polynomial; t is the time corresponding to the track point to be interpolated, j = 0, 1, 2...n * -1, i = 1, 2, ... n * ; Steps 1-4: construct the data set and perform Min-Max normalization on the data.

3. The method according to claim 2, characterized in that Steps 1-2 include: Extract the i-th track point P from the existing track data i The value x of the corresponding field i , all x i The constructed data set D is represented as D = {x1, x2, ..., x n }, define the local length δ, the neighborhood distance threshold ε, and the number of points in the core point neighborhood threshold η according to the needs of forming clusters and removing data, where δ and η need to satisfy η≤2δ; the parameter ε is the dimension of the value, η=2ε.

4. The method according to claim 3, characterized in that In steps 1-4, the following formula is used for normalization: where x max Indicates the maximum value of the data, x min represents the minimum value of the data, x is the original data, X * is the normalized data.

5. The method according to claim 4, characterized in that Steps 2-5 include the following steps: Step 2-5-1, the time domain convolution block is implemented by inputting U is a two-dimensional tensor, and a one-dimensional convolution is performed along the time dimension, where the number of convolution kernels is 2Z0. Get the result Then use the GLU activation function to activate the result and get the following result: Among them, P and Q are partial feature vectors output by the time domain convolution block, K t represents the length of the convolution kernel; τ represents the parameter tensor used for the convolution operation, τ *τ Represents the new convolution kernel tensor obtained by operating the convolution kernel τ with itself; In a complete space-time graph, input Get the result Step 2-5-2, the spatial convolution block is performed on the graph at each time step, by inputting The output Y is: in: T i (U)=2uT i-1 (u)-T i-2 (u), Among them, K represents the number of convolution kernels, θ i represents the weight parameter, T i represents the transposed matrix, represents the normalized Laplacian matrix, I n represents the n*n identity matrix, L represents the Laplace matrix, λ max represents the maximum eigenvalue in the Laplace matrix L, and u is a single element in U; The output of the spatial convolution block is an adjacency matrix, and the convolution kernel There are Z0 in total; in a complete space-time graph, input Get the result 6. The method according to claim 5, characterized in that Steps 2-6 include the following steps: Step 2-6-1, divide the data into different height intervals, and use different parameter groups for training for different height intervals, including the number of convolution kernels nb_filters, the number of STGCN levels levels and the convolution kernel size kernel_size; Step 2-6-2, DI-STGCN trains more than two STGCN models at the same time, and each model is trained using different parameter groups for data features in different height ranges; Step 2-6-3, after training is completed, DI-STGCN integrates the models trained in each height interval into one model to predict and process data in different height intervals.

7. The method according to claim 6, characterized in that Step 3 includes the following steps: Step 3-1, obtaining experts’ opinions on airport trajectory prediction and safety through questionnaire survey; Step 3-2, calculate the prior probability of the Bayesian method based on the questionnaire data; Step 3-3, through dynamic identification, for different track data sets in the flight phase, the predicted data obtained by DI-STGCN is used as the observation data, combined with the prior probability, and the posterior safety probability of the corresponding track is obtained through the Bayesian neural network; Step 3-4, based on the a posteriori safety probability and combined with the entropy weight method to determine the weights, a flight safety assessment system is constructed for the flight stage.

8. The method according to claim 7, characterized in that Step 3-4 includes the following steps: Step 3-4-1, combined with the posterior safety probability distribution, based on the probability, divide the safety intervals of longitude, latitude, and altitude into five levels according to the probability of occurrence, and assign values ​​1 to 5. The higher the score, the higher the safety. Step 3-4-2, combined with the real track data, the importance weights of longitude, latitude and altitude in the flight process are determined by entropy weight method, and the corresponding weights w1, w2 and w3 are obtained respectively; Step 3-4-3, calculate the safety score of each flight's track point according to the following formula: B=w1X1+w2X2+w3X3 Among them, X1 is the longitude matrix, X2 is the latitude matrix, X3 is the altitude matrix, and B is the score matrix.

9. A storage medium, characterized in that: A computer program or instruction is stored, and when the computer program or instruction is executed, the method according to any one of claims 1 to 8 is implemented.

Images