Adaptive grid-based airspace situation prediction method and system

Abstract

The application relates to the technical field of traffic control, in particular to a space situation prediction method and system based on an adaptive grid, which comprises the following steps: collecting information of all aircrafts in a current airspace, including the number, speed, x coordinate, y coordinate, z coordinate, heading, pitch angle and time of the aircrafts; according to the x coordinate, y coordinate and z coordinate of the aircrafts, mapping the collected information of the aircrafts into a grid of a pre-constructed airspace basic grid network, and obtaining an input feature sequence; inputting the input feature sequence into a 3D CNN-GCN-Transformer-based airspace grid situation prediction model, and outputting an airspace situation prediction result. Through the application, accurate prediction of the change of the airspace situation can be realized, thereby providing technical support for air traffic control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of traffic control, in particular to an airspace situation prediction method and system based on adaptive grid. BACKGROUND

[0002] As the key space for aircraft activities, the airspace traffic situation has high dynamicity and strong complexity. To cope with the challenges in safety and efficiency, airspace situation prediction has become a core means to improve the level of airspace management.

[0003] The airspace structure of air traffic is complex and dynamically changing. The traditional airspace description method based on latitude and longitude has been difficult to meet the needs of efficient and accurate management. Therefore, the grid-based airspace representation method has emerged. The airspace is divided into regular and uniform small units, combined with grid coding and spatial relationship, to establish a digital spatial reference framework, which can accurately describe the complex structure and dynamic changes of the airspace, and then realize the fine management and efficient use of the airspace. SUMMARY

[0004] The purpose of the present application is to provide an airspace situation prediction method and system based on adaptive grid, to improve the accuracy of airspace division, and then improve the accuracy of airspace situation analysis.

[0005] To achieve the above-mentioned purpose of the application, the embodiments of the present application provide the following technical solutions:

[0006] An airspace situation prediction method based on adaptive grid, comprising the following steps:

[0007] Collecting information of all aircrafts in the current airspace, including the number, speed, x coordinate, y coordinate, z coordinate, heading, pitch angle, and time of the aircrafts;

[0008] Mapping the collected information of the aircrafts to the grid of the pre-constructed airspace basic grid network according to the x coordinate, y coordinate, and z coordinate of the aircrafts, and obtaining an input feature sequence;

[0009] Inputting the input feature sequence into an airspace grid situation prediction model based on 3D CNN-GCN-Transformer, and outputting an airspace situation prediction result.

[0010] The construction steps of the airspace basic grid network are:

[0011] Firstly, the airspace is uniformly divided into H height layers, and each height layer is divided into MxN basic grids;

[0012] Then, according to the situation complexity of the basic grid, the adaptive splitting and merging of the grid are performed based on the principle of quadtree: for a certain grid, if and If the value of the space-time complexity of the parent grid is greater than the split threshold, then continue to split the parent grid into four child grids, d represents the number of splits, d max represents the maximum number of splits, represents the space-time complexity of the child grid after d times of splitting, is a split threshold; if the space-time complexity of all child grids is less than the split threshold, then the four child grids are merged into one parent grid.

[0013] The construction step of the space domain basic grid network further includes a grid encoding step, comprising:

[0014] In the vertical direction of the flight space domain, taking the sea level as the reference, according to the formula determine the height layer, h is the height layer to which the aircraft belongs, is the actual height of the aircraft, is the floor value; the height encoding is the numerical value of the height layer;

[0015] A quadrilateral is used as the network structure, and (0 0 longitude, 0 0 latitude) is used as the coordinate origin, longitude and latitude are used as the coordinate axis direction, from the coordinate origin, along the latitude axis, use capital letters A-W for encoding, along the longitude axis, use 00-44 for a total of 45 digits for encoding, according to the formula and the network longitude change table, the encoding A m of the mth layer of the horizontal coordinate address composed of x coordinates and y coordinates is obtained. Address = A 1 A 2... A m , A m is composed of the latitude coordinate N m and the longitude coordinate E m .

[0016] The space domain grid situation prediction model based on 3D CNN-GCN-Transformer includes a space domain feature reconstruction module based on residual 3DCNN, a space feature reconstruction module based on GCN, a time-space hybrid attention mechanism module based on a Transformer architecture, and a projection reconstruction module. The space domain feature reconstruction module based on residual 3D CNN is used to extract the space-time situation features of the grid in the local space-time region, the space feature reconstruction module based on GCN is used to construct the space topology based on the air route, the time-space hybrid attention mechanism module based on the Transformer architecture is used to fuse the space-time situation features and the space topology, and the projection reconstruction module is used for feature upsampling and scale mapping to predict the grid situation at the next time step.

[0017] The spatial feature reconstruction module based on residual 3D CNN includes at least two parallel CNN-residual blocks, each of which includes two 3D convolution layers and a first residual connection layer, and the structure of each 3D convolution layer is 3D convolution + ReLU activation function, and the expression of the CNN-residual block is: wherein X represents an input, which includes a sequence of aircraft vector velocity matrices, a sequence of total number of grids information, and a sequence of maximum splitting number of grids, W1 and W2 are parameters of two 3D convolution layers, denotes a three-dimensional convolution operation, ReLU is a nonlinear activation function; for position , the calculation process of the 3D convolution operation is: wherein m=1, 2, 3…M, n=1, 2, 3…N, h=1, 2, 3…H, c in and c out represent input and output channels respectively.

[0018] The spatial feature construction module based on GCN includes multiple parallel GCN-residual blocks, each of which includes a graph convolution layer and a second residual connection layer, and the graph convolution layer first performs graph convolution and ReLU activation operation on the input to upgrade the input from 2 layers to 4 layers, and then performs graph convolution and ReLU activation operation once again to downgrade the 4 layers to 2 layers, and let denotes the node feature matrix after the layer graph convolution, wherein g is the number of nodes, is the feature dimension, and the original input feature is wherein is the input sequence after adding self-connection to the original input sequence A; is the degree matrix of , used for symmetric normalization; is a weight matrix to be learned; ReLU is a ReLU activation function, L is the total number of graph convolution layers, is a normalization process.

[0019] The spatio-temporal hybrid attention mechanism module based on the Transformer architecture is stacked by multiple Transformer encoding layers with the same structure, each of which includes a multi-head self-attention mechanism sublayer and a positional feedforward network sublayer connected in series, and both the multi-head self-attention mechanism sublayer and the positional feedforward network sublayer are connected with a residual connection and layer normalization; X t-1 denotes the output of the t-1 time step, and the calculation process of each Transformer encoding layer in the t time step is:

[0020] Self-attention mechanism sublayer: ; position feed-forward network sublayer: ; wherein, , , , wherein r represents the rth attention head, R is the total number of attention heads, d k is the number of columns; X is the feature matrix output after processing by the residual 3D CNN based spatial feature reconstruction module, are trainable projection parameter matrices, , , is a learnable weight matrix; b1 and b2 are bias terms; , wherein x is the output feature of the self-attention mechanism sublayer and the position feed-forward network sublayer, , are the mean and variance of the input in the feature dimension, respectively; γ, β are learnable scaling and shifting parameters, is a numerical stability constant.

[0021] The projection reconstruction module first maps the output feature of the Transformer encoding layer at the last time step, denoted as , to a new feature , X N is the output feature of the Transformer encoding layer, with a dimension of , b represents the batch size, represents the spatial dimension, and q represents the feature dimension; then dimension reorganization and expansion are performed, , , wherein is a reorganization transformation that reorganizes the feature from to , and satisfies , W P and b P are trainable projection parameters; finally, by upsampling Zʺ, the dimension is mapped to a larger resolution spatial coordinate, , is a spatial upsampling transformation.

[0022] An adaptive grid-based spatial situation prediction system, comprising:

[0023] An information acquisition unit for acquiring information of all aircraft in the current airspace, including the number, speed, x coordinate, y coordinate, z coordinate, heading, pitch angle, and time of the aircraft;

[0024] a grid mapping unit, configured to map the collected aircraft information into a grid of a pre-constructed airspace base grid network according to x, y and z coordinates of the aircraft, and obtain an input feature sequence;

[0025] a situation prediction unit, configured to input the input feature sequence into a 3D CNN-GCN-Transformer-based airspace grid situation prediction model, and output an airspace situation prediction result.

[0026] The 3D CNN-GCN-Transformer-based airspace grid situation prediction model comprises a residual 3D CNN-based airspace feature reconstruction module, a GCN-based spatial feature reconstruction module, a Transformer architecture-based spatio-temporal hybrid attention mechanism module and a projection reconstruction module. The residual 3D CNN-based airspace feature reconstruction module is configured to extract an airspace situation feature of a grid in a local spatio-temporal region. The GCN-based spatial feature reconstruction module is configured to construct a spatial topology based on an air route. The Transformer architecture-based spatio-temporal hybrid attention mechanism module is configured to fuse the airspace situation feature and the spatial topology. The projection reconstruction module is configured to perform up-sampling and scale mapping of the feature, and predict a grid situation at a next time step.

[0027] For the spatio-temporal characteristics of the airspace situation, the fusion-type deep learning model framework can effectively integrate the multi-modal spatio-temporal features in the airspace, break through the limitations of single modal information, and improve the overall representation ability of the airspace situation. The prediction model constructs an adjacency graph of the grid, clearly defines the topological relations of the airspace operation such as air route corridors, key nodes and intersection regions, and helps to identify potential bottlenecks and congestion trends. The spatio-temporal attention mechanism in the Transformer module is introduced to effectively capture the dynamic changes and spatial interaction features of the airspace situation over time, and enhance the perception ability of the prediction model to complex spatio-temporal patterns. Meanwhile, the projection reconstruction module is used to realize the state mapping from coarse granularity to high resolution, further improve the accurate prediction effect of local hot spot areas and key airspace, and meet the fine and dynamic management requirements in complex airspace environment.

[0028] Compared with the prior art, the application extracts the situation characteristics of the spatial grid in the local space-time region by using a three-dimensional convolutional neural network, ensures the perception ability of the situation distribution and short-time dynamic change, introduces a graph convolutional network to construct a space topology based on a route, expresses the spatial dependence relationship between different grids formed by the fixed route, then a Transformer module captures the long-term dependence and global interaction characteristics in the evolution process of the spatial grid situation through a multi-head space-time hybrid attention mechanism, deeply fuses the extracted local and topological characteristics, and finally realizes the fine prediction of the situation (the number of splitting times) of the grid at the next time step through a projection mapping and reconstruction module, thereby accurately realizing the spatial situation prediction and providing technical support for air traffic control.

[0029] Other technical advantages of the application are described in the following embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiments will be briefly introduced as follows. It should be understood that the following drawings only show some of the embodiments of the application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0031] Figure 1 The flowchart of the spatial situation prediction method based on adaptive grid in the embodiments.

[0032] Figure 2 The schematic diagram of the situation influence of the aircraft on the grid.

[0033] Figure 3 The schematic diagram of the splitting and merging of the grid.

[0034] Figure 4 The schematic diagram of the grid encoding rule.

[0035] Figure 5 The architecture diagram of the spatial grid situation prediction model based on 3D CNN-GCN-Transformer.

[0036] Figure 6 The structure diagram of the CNN-residual block.

[0037] Figure 7 The working flowchart of the GCN-residual block.

[0038] Figure 8 The working flowchart of the space-time hybrid attention mechanism module based on the Transformer architecture.

[0039] Figure 9 A structural block diagram of the adaptive grid-based airspace situation prediction system. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0041] Referring to Figure 1 The adaptive grid-based airspace situation prediction method provided in the embodiments includes the following steps:

[0042] S10, collecting information of all aircrafts in the current airspace, including aircraft number, speed, x coordinate, y coordinate, z coordinate, heading, and pitch angle.

[0043] The aircraft number is used to represent the identity of the aircraft, and the speed, x coordinate, y coordinate, z coordinate, heading, and pitch angle are used to represent the situation of the aircraft. The collected information can also include time, which is used to represent the time characteristics.

[0044] S20, mapping the collected information of the aircraft into the grid of a pre-constructed airspace basic grid network according to the x coordinate, y coordinate, and z coordinate of the aircraft, and obtaining an input feature sequence.

[0045] S30, inputting the input feature sequence into an airspace grid situation prediction model based on a 3D CNN-GCN-Transformer model, and outputting an airspace situation prediction result.

[0046] Constructing an airspace basic grid network is to gridize the airspace for representation. The gridization representation of the airspace mainly includes the description of three core spatial elements of "point, line, and surface". Through the gridization representation of these elements, a comprehensive description of the airspace structure can be achieved.

[0047] The "point" structure is the basic unit in the airspace, which is usually used to describe static positions, such as take-off and landing points, aircraft position points, etc. Due to the large number and wide distribution of points, grid coding is used to label each point, which can be accurately positioned through grid coordinates. The grid representation is a triple array structure , where Level is the height encoding of the grid, Address is the address encoding, and Property is the property encoding. The height encoding reflects the height information of the grid in space, the address encoding represents its latitude and longitude information, and the property encoding indicates the airspace type where the grid is located. According to the "National Airspace Foundation Classification Method", airspace types are divided into 7 categories (A, B, C, D, E, G, W), among which categories A-E are controlled airspace and categories G and W are uncontrolled airspace. The airspace grid model has both spatial and attribute connotations, which can effectively support airspace situation recognition.

[0048] The "line" structure is used to describe dynamic paths or routes in airspace, such as aircraft tracks, approach routes, etc. Since the line structure has continuity, it is necessary to record each grid cell passed by the path during gridization to ensure the continuity and accuracy of the path. The gridization representation of a line can include two methods: 1) Grid edge description method, which describes the boundary of the path by describing the boundaries of the grid (such as the meridian and latitude lines of latitude and longitude); 2) Two-point line description method, which describes the path by connecting the geographical coordinates of two points. These methods can ensure the continuity of the line structure while converting it into manageable spatial units through the grid system.

[0049] The "face" structure is used to describe regional functional areas in airspace, such as no-fly zones, airspace control zones, etc. These areas usually have clear boundaries and areas, which can be accurately described through gridization methods. The gridization representation methods of the face structure can include grid area description method and multi-point line description method. The grid area description method represents a face area by combining multiple grid cells, with each grid marked by a grid code. The multi-point line description method connects multiple geographical coordinate points to form the boundary of the face area, and then determines the grid within the area through filling method.

[0050] For "line" and "face" structures, the gridization process needs to complete the continuity description through spatial filling. As an example, the odd-even scanning method is used to complete the area filling in this embodiment, the basic idea of which is to scan the polygon along its boundary through scanning lines, and determine whether the grid on each scanning line belongs to the interior of the polygon according to the parity of the intersection points between the scanning line and the polygon. The specific processing flow is as follows:

[0051] Step 1, row scanning: divide the entire area into several horizontal scanning lines according to the horizontal direction, and process row by row.

[0052] Step 2, calculate intersection points: for each scanning line, calculate the intersection points with the polygon boundary. If the scanning line intersects with the boundary of the polygon, these intersection points are the intersection points.

[0053] Step 3, Parity Determination: Determine whether a grid on the scan line belongs to the interior of a polygon based on the number of intersection points. If the number of intersection points is odd, the grid between the intersection points belongs to the interior of the polygon and is marked as a filled area; if the number of intersection points is even, the scan line does not intersect the polygon, and the grid does not belong to the filled area.

[0054] Step 4, Update scan lines: Scan line by line until the entire area has been scanned.

[0055] Construct a basic spatial grid network by uniformly dividing the spatial domain into H height layers, which can be denoted as h={1,2…,H}. Each height layer is further divided into M×N basic grids, which can be denoted as s={1,2…,(M×N)}.

[0056] By representing airspace using "points, lines, and surfaces," airspace is divided into uniform discrete units, i.e., grids. However, when dealing with large-scale airspace, the uniform grid division method has certain limitations. Choosing smaller grid units to ensure control accuracy will lead to an exponential increase in computational load; using larger grid units to reduce computational burden may fail to effectively identify key areas in the airspace, thus affecting management effectiveness. Furthermore, factors such as route distribution, the clustering of key points, and the density of flight activity all influence the controller's level of attention to different areas. Areas with higher airspace complexity, such as route junctions and approach areas, require more detailed monitoring and management; while open areas can employ coarser-grained control strategies. This heterogeneous nature of airspace requires the grid to adaptively adjust its granularity and distribution.

[0057] Therefore, based on basic rasterization, this paper proposes an adaptive raster modeling method that splits and merges based on airspace situation characteristics. The adaptive raster splitting mechanism, driven by aircraft complexity, ensures an optimal balance between refined management and computational efficiency. In areas of higher complexity, the raster splitting depth increases, resulting in a non-uniform distribution of airspace partitions. This characteristic maps differences in airspace situation to raster distribution features, effectively characterizing the spatial evolution trend of airspace complexity. By analyzing and predicting the raster distribution, the spatial distribution characteristics of airspace situation complexity can be intuitively displayed, clearly indicating areas requiring priority attention to controllers.

[0058] Participation is possible Figure 2 A schematic diagram illustrating the situational impact of an aircraft on a grid, showing the two vertices p of the boundary edge closest to the aircraft. l and p r Midpoint p m and the grid center point p cthe average of the potential field of the grid as a measure of the influence of the aircraft on the basic grid complexity. Meanwhile, based on the superposition of the influence of all aircraft in the airspace, the overall situation index of the airspace grid is constructed, and the continuous time is divided into discrete time steps At each time step, the situation index is updated:

[0059] wherein, represents the situation influence of all aircraft in the airspace on the grid at time t, i.e., the airspace situation complexity of the grid at time t, 、 、 and are the left vertex, right vertex, middle point and center point of the nearest boundary edge of the aircraft i and the grid at time t, respectively; I is the total number of aircrafts that have influence on the grid situation within the time step.

[0060] By setting a split threshold , it is determined whether the grid needs to be further split according to the airspace situation complexity of the grid. If , further splitting is needed, thereby improving the identification accuracy of the region; d represents the number of splitting, d max represents the maximum number of splitting, represents the airspace situation complexity of the sub-grid after d times of splitting. In this embodiment, the splitting rule adopts a quadtree method. The quadtree is a technique for recursively partitioning a data space, and the basic idea is to divide the space (parent grid) into four sub-regions (sub-grids) each time, and each sub-grid can continue to be recursively split until the maximum number of splitting d max is reached. The maximum number of splitting d max is related to the type of airspace. Different airspace types have different requirements for safety separation and management accuracy of aircraft, so the maximum number of splitting of the grid needs to be dynamically adjusted to ensure that the granularity meets the specific needs.

[0061] Similarly, when the airspace situation complexity of all sub-grids is less than the splitting threshold, the four sub-grids are merged into one parent grid using the quadtree method. In the merging process, the original sub-grids are deleted, and the parent grid becomes the new leaf grid. As Figure 3 shown, the evolution of grid complexity will affect the dynamic adjustment of the grid, and thus lead to differentiated merging and splitting.

[0062] After dividing the grid, it needs to be encoded. The encoding of the grid includes height encoding and horizontal address encoding. The height encoding is in the vertical direction of the flight airspace, taking the sea level as the reference, and dividing once every certain height distance, for example, 300 m, thereby obtaining a number of height spaces. The specific calculation is as follows:​ where h is the height layer the aircraft belongs to, is the actual height of the aircraft, in meters, is the floor function.

[0063] The horizontal address coding can be based on the GeoSOT grid, which recursively divides the airspace from degrees (the coarsest division unit) to minutes, seconds, and so on, so that each layer can represent a smaller spatial area with higher precision, and each area has a unique address code. A quadrilateral is taken as the basic grid network structure of the airspace, and (0 0 , 0 0 ) is taken as the coordinate origin, and longitude and latitude are taken as the coordinate axis direction. The first layer of the network uses a 4 ° x 4 ° network structure, and from the coordinate origin, the grid is represented by capital letters A-W along the latitude axis and 00-44 (a total of 45 numbers) along the longitude axis. Based on the quadtree splitting method, the parent grid is split into 4 child grids, so the second layer network structure is 2 ° x 2 ° , and the grids are represented in order clockwise as A-D, as shown in Figure 4 . The parent grid and the child grid belong to the relative concept, for example, grid 1 is split into 4 grids 2, then grid 1 is called the parent grid, and grid 2 is called the child grid; if one of the grids 2 is further split into 4 grids 3, then grid 2 is the parent grid, and grid 3 is the child grid.

[0064] Based on this logic, the airspace is divided into finer and finer precision levels, and the specific airspace basic grid network level information is shown in Table 1.

[0065] Table 1: Network level information table

[0066]

[0067] 500 meters is already a very small grid, so only 11 levels are recorded in Table 1. As an example, for a position point coordinate (a ° bʹcʺN, d ° eʹfʺE), the horizontal address coding based on the GeoSOT grid is: , where the latitude coordinate N m and the longitude coordinate E m of the mth layer of the coordinate together constitute the coding A m of the layer. The horizontal address coding of the coordinate is Address = A 1 A 2... A m , S mThe grid precision size corresponding to the mth layer.

[0068] As an example, the spatial representation form of "point, line, surface" based on GeoSOT grid encoding is shown in Table 2 as follows:

[0069] Table 2: Different forms of spatial representation form table

[0070]

[0071] After constructing the spatial basic grid network, according to the x coordinate, y coordinate and z coordinate of the aircraft, the position point of the aircraft can be mapped to the corresponding grid, and the speed, heading, pitch angle and aircraft number are also mapped to the corresponding grid as the attribute of the grid. All aircraft information can form a multi-item input feature sequence, which is specifically an aircraft vector speed matrix sequence , a flight line adjacency matrix sequence , a grid total number information sequence , a maximum number of grid splits sequence , the x coordinate, y coordinate and z coordinate determine the position point of the aircraft, the heading and pitch angle determine the flight line of the aircraft, the aircraft vector speed matrix sequence is composed of the speed of all aircrafts, the flight line adjacency matrix sequence is determined by the flight line and position point of all aircrafts, and the grid total number information sequence and the maximum number of grid splits sequence are set by the user.

[0072] The spatial grid situation prediction model based on 3D CNN-GCN-Transformer takes the aircraft vector speed matrix sequence, the flight line adjacency matrix sequence, the grid total number information sequence and the maximum number of grid splits sequence as the input feature sequence. For any height layer, the influence of the aircraft on the spatial situation is based on the microscopic scale, and combined with the quadtree splitting rule, each grid is recursively divided and complexity evaluated.

[0073] According to the grid encoding rule, define: , and let represent the collection of all grid encodings with splitting number d composed of , for example is AA, AB, AC, AD, BA, BB, BC, BD, CA, CB, CC, CD, DA, DB, DC, DD. In the quadtree structure, any grid with splitting number d can be represented by , is the grid encoding in the basic grid network of the height layer h, represents the encoding obtained by the grid along the quadtree splitting path.

[0074] By judging the safety situation scalar value of the grid at time t whether greater than a split threshold , to determine whether the grid needs to be further subdivided, while setting the subsequent complexity of no need to be subdivided to 0. When d < d max and , the grid continues to be subdivided into 4 sub-grids, and the coding rule is .

[0075] Reference can be made to Figure 5 , which shows the architecture of the spatial grid situation prediction model based on 3D CNN-GCN-Transformer. The spatial grid situation prediction model uses a three-dimensional convolutional neural network to extract the spatial situation features of the spatial grid in the local space-time region, ensuring the perception ability of the situation distribution and short-time dynamic change. Secondly, the graph convolution network is introduced to construct the spatial topology based on the route, expressing the spatial dependence relationship between different grids formed by the fixed route. Then, the Transformer module captures the long-term dependence and global interaction characteristics in the evolution process of the spatial situation through the multi-head space-time hybrid attention mechanism, and deeply fuses the extracted local spatial situation and topological features. Finally, through projection mapping and reconstruction, the feature upsampling and scale mapping are realized, and the fine prediction of the next time step grid situation (split number) is completed.

[0076] Specifically, the spatial grid situation prediction model includes a spatial feature reconstruction module based on residual 3D CNN, a spatial feature reconstruction module based on GCN, a space-time hybrid attention mechanism module based on the Transformer architecture, and a projection reconstruction module.

[0077] The spatial feature reconstruction module based on residual 3D CNN includes a plurality of (at least two) parallel CNN-residual blocks, each CNN-residual block including two 3D convolution layers and a first residual connection layer. The structure of each layer of 3D convolution layer is 3D convolution + ReLU activation function. The structure of a CNN-residual block is as shown in Figure 6 . The expression of each CNN-residual block is: , wherein X represents the input, which includes the aircraft vector velocity matrix sequence, the grid total number information sequence, and the maximum split number sequence of the grid, W1 and W2 are the parameters of the two layers of 3D convolution, , and the 3D convolution operation is represented as , and ReLU is the nonlinear activation function. For the position , the 3D convolution operation calculation process is: , wherein traverses the spatial dimensions of the 3D convolution kernel, m = 1, 2, 3…M, n = 1, 2, 3…N, h = 1, 2, 3…H, c in , and c outThese represent the input and output channels, respectively.

[0078] Airspace situation features include aircraft speed, grid split count, and total number of grid cells, forming multidimensional heterogeneous data. Simply stacking these features directly will fail to effectively extract key local features, limiting the model's ability to represent complex situations. This paper introduces 3D convolution and a residual structure (ResBlock) to fully explore the deep correlations between multidimensional airspace features and improve feature extraction accuracy. Simultaneously, the residual structure effectively alleviates the gradient vanishing problem that easily occurs during the training of deep networks, laying the foundation for stable training and performance improvement of the subsequent airspace grid situation model. A multi-CNN-residual block cascade approach is adopted, using layer-by-layer nonlinear transformations to gradually extract hierarchical feature representations of the airspace situation, providing a foundation for subsequent spatial correlation modeling and spatiotemporal dependency construction.

[0079] The spatial feature construction module based on GCN includes multiple parallel GCN-residual blocks. Each GCN-residual block includes a graph convolutional layer and a second residual connection layer. The graph convolutional layer first performs graph convolution and ReLU activation on the input (the output of the spatial feature reconstruction module based on residual 3D CNN and the flight path adjacency matrix sequence) to increase the input dimension from 2 layers to 4 layers, and then performs graph convolution and ReLU activation to reduce the 4 layers to 2 layers. The working process is as follows: Figure 7 As shown. Let Indicates the first The node feature matrix after layer graph convolution, where g is the number of nodes. For feature dimensions. If If the original input features are (the output of the spatial feature reconstruction module based on residual 3D CNN and the sequence of flight path adjacency matrices), then... in, The input sequence after adding a self-loop is the original input sequence A (the flight path adjacency matrix sequence and the aircraft vector velocity matrix sequence, grid total number information sequence, and grid maximum split number sequence output after processing by the residual 3D CNN-based airspace feature reconstruction module). for The degree matrix is ​​used for symmetric normalization; It is the weight matrix that needs to be learned; The ReLU activation function is used to improve the model's nonlinear expressive power, and L is the total number of graph convolutional layers. This is for normalization purposes.

[0080] The movement trajectory formed by the aircraft flying along the fixed route constructs a complex airspace network topology, and its adjacency matrix can better reflect the spatial correlation characteristics in the airspace. By using GCN to model this graph structure, it can effectively fuse and transfer spatial features, thereby mining the correlation features with global dependency.

[0081] As shown in Figure 8 , the spatio-temporal hybrid attention mechanism module based on the Transformer architecture is stacked by multiple Transformer encoding layers with the same structure. Each Transformer encoding layer includes a multi-head self-attention mechanism (Multi-Head Self-Attention) sublayer and a position-wise feed-forward network (Position-wise Feed-Forward Network) sublayer connected in series, and each sublayer is connected with a residual connection and layer normalization. If X t-1 represents the output of the t-1 time step, then the calculation process of each Transformer encoding layer in the t time step is as follows:

[0082] Self-attention mechanism sublayer:

[0083] Position-wise feed-forward network sublayer:

[0084] wherein, , , , wherein r represents the rth attention head, R is the total number of attention heads, d k is the column number of to maintain numerical stability; X is the feature matrix output after processing by the residual 3D CNN-based airspace feature reconstruction module, are trainable projection parameter matrices, , , are learnable weight matrices; b1 and b2 are bias terms. wherein x is the output feature of each sublayer (self-attention mechanism sublayer and position-wise feed-forward network sublayer), , are the mean and variance of the input in the feature dimension, respectively; γ and β are learnable scaling and translation parameters, is a numerical stability constant.

[0085] The evolution process of the spatial situation not only embodies the dynamic change in the time dimension, but also involves the interaction between spatial positions. By introducing the Transformer architecture to construct the spatio-temporal hybrid attention mechanism, the overall modeling in the time and space dimensions can be realized, so that direct interaction between any spatio-temporal positions can be achieved without relying on multi-layer information transmission, effectively capturing the correlation between long time series and distant positions. At the same time, the attention mechanism has self-adaptability and can dynamically focus on the spatio-temporal positions most relevant to the current situation, breaking through the constraints of traditional models in time and space processing, avoiding information loss caused by separate modeling of time and space features, and improving the modeling accuracy of the evolution process of complex spatial situations.

[0086] The projection reconstruction module is to effectively map the fused multi-dimensional features to the target high-dimensional space. The projection reconstruction module first maps the output features of the Transformer encoding layer at the last time step, denoted as , to , where N is the output feature of the Transformer encoding layer, with a dimension of , b represents the batch size, represents the spatial dimension, and q represents the feature dimension; then, dimension reorganization and expansion are performed, , where is a reorganization transformation that reorganizes the features from to , and satisfies , W P and b P are trainable projection parameters; finally, by upsampling Zʺ, the dimension is mapped to a larger resolution spatial coordinate, , is a spatial upsampling transformation.

[0087] The projection reconstruction module realizes the mapping from low-resolution features, such as 8x8, to high-resolution predictions, 8x16x8x16. Each original base grid is subdivided into 16x16 sub-grids, achieving finer grid splitting prediction and significantly improving the spatial resolution and detail expression ability of the spatial situation prediction.

[0088] To verify the effectiveness of the spatial situation prediction model proposed in this paper, the ADS-B data of all aircraft (including take-off, landing, and crossing flights) in the Chengdu terminal airspace from December to the 12th to the 17th of the following month in a certain year is selected as experimental data. Data preprocessing includes the following steps: 1) Establish a Cartesian coordinate system with the airport as the origin, where the X, Y, and Z axes are converted from longitude, latitude, and flight altitude, respectively; 2) The data only contain the flight speed of the XY axis, so the aircraft climb trend is determined by the height change, and the climb angle is set to 5° elevation and the descent angle is 3° pitch.

[0089] In addition, to ensure the integrity and accuracy of the data, the Z-score anomaly detection method is used to eliminate abnormal values that deviate from the normal flight pattern, and each timestamp is rounded to ensure that all aircraft data are aligned at the same time step. The processed data set information includes: aircraft number (NUM), speed (SPEED), x-coordinate (x / km), y-coordinate (y / km), z-coordinate (z / km), heading (ANGLE), pitch angle (PITCH), time (TIME), and refined time (Time-refine). The refined time refers to the rounded time, such as 3 minutes and 24 seconds modified to 3 minutes and 20 seconds.

[0090] Through statistical analysis of experimental data, the coordinate distribution range of aircraft ADS-B data is: X: [-128km, 128km], Y: [-160km, 96km], Z: [0km, 12km]. To improve computational efficiency and facilitate expression, the basic grid level M=5 is selected, i.e. the longitude and latitude resolution of the basic grid unit is set to 16ʹ×16ʹ (32km×32km), and 0.3km is used as the basic height layer unit, thereby constructing a cubic grid unit. Based on the average speed of the aircraft, which is about 800km / h, to ensure that the grid division can identify high dynamic flight environment, the finest granularity is set to level M=9, i.e. the grid resolution is (2km×2km). This scale determines the maximum splitting number as 4, which enables the grid to adaptively adjust at different scales, thereby balancing the calculation accuracy and computational efficiency.

[0091] Several other prediction models are also designed for comparison experiments, and the experimental results are shown in Table 3:

[0092] Table 3: Comparison of experimental results of different prediction models

[0093]

[0094] It is concluded through analysis that the prediction accuracy of the CNN-GCN-spatiotemporal-Transformer model using the AMSE loss function is significantly improved compared to the model using the MSE loss function in high fragmentation levels (i.e. complex situation areas). Considering the prediction accuracy and the ability to identify high complexity areas, the CNN-GCN-spatiotemporal-Transformer model with β = 4 achieves a better balance between accuracy and robustness, and exhibits stronger overall performance.

[0095] For details, please refer to Figure 9 Based on the same inventive concept, the embodiment simultaneously provides a self-adaptive grid-based airspace situation prediction system, which comprises:

[0096] An information collection unit is configured to collect information of all aircrafts in the current airspace, including the number, speed, x coordinate, y coordinate, z coordinate, heading, pitch angle and time of the aircrafts.

[0097] A grid mapping unit is configured to map the collected information of the aircrafts into the grid of a pre-constructed airspace basic grid network according to the x coordinate, y coordinate and z coordinate of the aircrafts, and obtain an input feature sequence.

[0098] A situation prediction unit is configured to input the input feature sequence into a 3D CNN-GCN-Transformer-based airspace grid situation prediction model, and output an airspace situation prediction result.

[0099] For the working process and structure of each component unit, please refer to the related description in the foregoing method embodiment. To save space, no further description is given here.

[0100] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for spatial situation prediction based on adaptive grid, characterized in that, The method comprises the following steps: Collecting information of all aircrafts in the current airspace, including the number, speed, x coordinate, y coordinate, z coordinate, heading, pitch angle and time of the aircrafts; Mapping the collected information of the aircrafts into the grid of a pre-constructed airspace basic grid network according to the x coordinate, y coordinate and z coordinate of the aircrafts, and obtaining an input feature sequence; Inputting the input feature sequence into an airspace grid situation prediction model based on 3D CNN-GCN-Transformer, and outputting an airspace situation prediction result; The airspace grid situation prediction model based on 3D CNN-GCN-Transformer comprises an airspace feature reconstruction module based on residual 3D CNN, a spatial feature reconstruction module based on GCN, a spatio-temporal hybrid attention mechanism module based on a Transformer architecture, and a projection reconstruction module, the airspace feature reconstruction module based on residual 3D CNN is used for extracting airspace situation features of the grid in a local spatio-temporal region, the spatial feature reconstruction module based on GCN is used for constructing a spatial topology based on a route, the spatio-temporal hybrid attention mechanism module based on the Transformer architecture is used for fusing the airspace situation features and the spatial topology, and the projection reconstruction module is used for up-sampling and scale mapping of features to predict a grid situation at a next time step; The airspace feature reconstruction module based on residual 3D CNN comprises at least two parallel CNN-residual blocks, each CNN-residual block comprises two 3D convolution layers and a first residual connection layer, and the structure of each 3D convolution layer is 3D convolution + ReLU activation function; The spatial feature construction module based on GCN comprises a plurality of parallel GCN-residual blocks, each GCN-residual block comprises a graph convolution layer and a second residual connection layer; The spatio-temporal hybrid attention mechanism module based on the Transformer architecture is stacked by a plurality of Transformer encoding layers with the same structure, each Transformer encoding layer comprises a multi-head self-attention mechanism sublayer and a position feedforward network sublayer connected in series, and the multi-head self-attention mechanism sublayer and the position feedforward network sublayer are both connected with a residual connection and layer normalization.

2. The adaptive grid based airspace situation prediction method of claim 1, wherein, The construction steps of the airspace basic grid network are as follows: First, the airspace is uniformly divided into H height layers, and each height layer is divided into MxN basic grids; Then according to the situation complexity of the basic grid, the grid adaptive splitting and merging is performed based on the principle of quadtree: for a certain grid, if and then continue to split four sub-grids, d represents the number of splitting, d max represents the maximum number of splitting, represents the spatial situation complexity of the sub-grid after d times of splitting, is the splitting threshold; if the spatial situation complexity of all sub-grids is less than the splitting threshold, then the four sub-grids are merged into a parent grid.

3. The adaptive grid based airspace situation prediction method of claim 2, wherein, The construction steps of the airspace basic grid network further comprise a grid encoding step, comprising: In the vertical direction of the flight airspace, with the sea level as the reference, the formula is used to determine the height layer, h is the height layer to which the aircraft belongs, is the actual height of the aircraft, is the floor function; the height code is the numerical value of the height layer; The quadrangle is used as the network structure, and (0 0 longitude, 0 0 latitude) is used as the coordinate origin and the longitude and latitude as the coordinate axis direction. Starting from the coordinate origin, the latitude axis is encoded with capital letters A-W, and the longitude axis is encoded with 00-44, a total of 45 digits. According to the formula and the network longitude change table, the encoding A m of the mth layer of the horizontal coordinate address composed of x coordinate and y coordinate is obtained. The horizontal address encoding of the horizontal coordinate address is Address=A 1 A 2... A m , A m is composed of the latitude coordinate N m and the longitude coordinate E m . 4.The adaptive grid-based space domain situation prediction method according to claim 1, characterized in that, The expression of the CNN-residual block is: where X represents the input, which includes the aircraft vector velocity matrix sequence, the grid total number information sequence, and the maximum splitting number sequence of the grid, W1 and W2 are parameters of two layers of 3D convolution respectively, is represented as a three-dimensional convolution operation, is a ReLU nonlinear activation; for position , the calculation process of the 3D convolution operation is: where m = 1, 2, 3…M, n = 1, 2, 3…N, h = 1, 2, 3…H, c in and c out represent the input and output channels respectively.

5. The adaptive grid based airspace situation prediction method of claim 1, wherein, The graph convolution layer first performs graph convolution and ReLU activation operation on the input to upgrade the input from 2 layers to 4 layers, and then performs graph convolution and ReLU activation operation once again to downgrade the 4 layers to 2 layers, that is, represents the first node feature matrix after layer graph convolution, where g is the number of nodes, is the feature dimension, and the original input feature , wherein, is the input sequence after adding self-connection on the original input sequence A; is the degree matrix for symmetric normalization; is the weight matrix to be learned; ReLU activation function, L is the total number of graph convolution layers, is the normalization processing.

6. The adaptive grid based airspace situation prediction method of claim 1, wherein, X t-1 denotes the output at time step t−1, then the computation of each Transformer encoding layer at time step t is given by: Self-attention mechanism sublayer: ; Position feedforward network sublayer: ; wherein, , , , wherein r represents the rth attention head, R is the total number of attention heads, d k is the number of columns; X is the feature matrix output after processing by the residual 3D CNN-based spatial domain feature reconstruction module, respectively trainable projection parameter matrices, , , is a learnable weight matrix; b1 and b2 are bias terms; , wherein x is the output feature of the self-attention mechanism sublayer and the position feedforward network sublayer, , respectively the mean and variance of the input in the feature dimension; γ, β are respectively learnable scaling and translation parameters, is a numerical stability constant.

7. The adaptive grid based airspace situation prediction method of claim 1, wherein, The projection reconstruction module first maps the output feature of the Transformer encoding layer at the last time step, denoted as , to a feature space, denoted as , where N is the output feature of the Transformer encoding layer, and has a dimension of , b represents the batch size, represents the spatial dimension, and q represents the feature dimension. Dimension reorganization and expansion are then performed, , wherein is a reorganization transform that reorganizes the features from to and satisfies W P and b P are trainable projection parameters; finally, the dimensions are mapped to spatial coordinates of larger resolution by upsampling Zʺ, , is a spatial upsampling transform.

8. An adaptive grid based airspace situation prediction system, characterized by, Comprise: An information collection unit configured to collect information of all aircrafts in the current airspace, including the number, speed, x coordinate, y coordinate, z coordinate, heading, pitch angle and time of the aircrafts; A grid mapping unit configured to map the collected information of the aircrafts into the grid of a pre-constructed airspace basic grid network according to the x coordinate, y coordinate and z coordinate of the aircrafts, and obtain an input feature sequence; A situation prediction unit configured to input the input feature sequence into an airspace grid situation prediction model based on 3D CNN-GCN-Transformer, and output an airspace situation prediction result. The 3D CNN-GCN-Transformer-based space domain grid situation prediction model comprises a residual 3D CNN-based space domain feature reconstruction module, a GCN-based space feature reconstruction module, a Transformer architecture-based space-time hybrid attention mechanism module, and a projection reconstruction module. The residual 3D CNN-based space domain feature reconstruction module is used to extract space domain situation features of a grid in a local space-time region. The GCN-based space feature reconstruction module is used to construct a space topology based on a flight path. The Transformer architecture-based space-time hybrid attention mechanism module is used to fuse the space domain situation features and the space topology. The projection reconstruction module is used to perform feature up-sampling and scale mapping to predict a grid situation at a next time step. The residual 3D CNN-based space domain feature reconstruction module comprises at least two parallel CNN-residual blocks. Each CNN-residual block comprises two 3D convolution layers and a first residual connection layer. The GCN-based space feature reconstruction module comprises a plurality of parallel GCN-residual blocks. Each GCN-residual block comprises a graph convolution layer and a second residual connection layer. The Transformer architecture-based space-time hybrid attention mechanism module is stacked by a plurality of Transformer encoding layers with the same structure. Each Transformer encoding layer comprises a multi-head self-attention mechanism sub-layer and a position feed-forward network sub-layer connected in series. Both the multi-head self-attention mechanism sub-layer and the position feed-forward network sub-layer are connected with a residual connection and layer normalization.

Images