Airplane wake prediction method based on spatio-temporal fusion deep learning network

Abstract

The application relates to the technical field of intelligent air traffic control and discloses an aircraft wake flow prediction method based on a space-time fusion deep learning network, which comprises the following steps: a laser radar performs vertical section scanning on aircraft wake flow to obtain radial velocity fields at multiple moments and construct a wake flow radial velocity field sequence; a time feature map and a space feature map of the wake flow radial velocity field sequence are extracted through an encoder of the space-time fusion deep learning network; the encoder comprises a time encoder and a space encoder; a meteorological scalar feature is converted into a meteorological boundary condition tensor through a tensor expansion module of the space-time fusion deep learning network, and is fused with the time feature map and the space feature map to obtain a fusion feature; and finally, a predicted wake flow radial velocity field is output by a decoder of the space-time fusion deep learning network. The application realizes high-precision and real-time prediction of the future evolution state of aircraft wake flow.

Description

Technical Field

[0001] This invention relates to the field of intelligent air traffic control technology, and in particular to a method for predicting aircraft wake turbulence based on a spatiotemporal fusion deep learning network. Background Technology

[0002] During flight, the lift generated by the wings creates a pair of strong, counter-rotating vortices behind the aircraft, known as the wake vortex. Figure 2 As shown, the core area of ​​a wake turbulence can have wind speeds of 10-20 m / s and last for several minutes, posing a serious safety threat to subsequent takeoffs and landings. If a following aircraft inadvertently enters the wake turbulence area of ​​the preceding aircraft, it may lead to loss of control, structural damage, or even a crash. To ensure flight safety, the International Civil Aviation Organization (ICAO) has stipulated strict wake separation standards, which directly limit the takeoff and landing capacity of airports. Therefore, accurately predicting the evolution and dissipation of wake turbulence is of great significance for optimizing flight separation and improving airport operational efficiency.

[0003] LiDAR is an active remote sensing technology that detects the distribution and movement of aerosol particles in the atmosphere by emitting laser pulses and receiving the scattered echoes from targets. Doppler lidar can measure the radial velocity of a target and is therefore widely used for detecting aircraft wake ducts. RHI scanning mode is one of the commonly used scanning methods for lidar. Its characteristics include a fixed scanning azimuth angle, a variable elevation angle within a certain range, and scanning on a plane perpendicular to the aircraft's flight path, enabling the acquisition of the two-dimensional velocity field distribution of the wake duct in a vertical profile.

[0004] However, existing methods for monitoring and controlling aircraft wake turbulence mainly rely on empirical rules or fixed time interval models, which are insufficient to reflect the dynamic evolution of wake turbulence under complex weather conditions in real time. They also cannot provide high-precision, real-time predictions of future wake turbulence states, thus affecting runway capacity utilization efficiency and operational safety margins. Furthermore, existing wake turbulence prediction methods based on physical models are computationally complex and lack real-time performance, failing to meet the rapid decision support requirements of air traffic control systems. Summary of the Invention

[0005] To achieve high-precision, real-time prediction of the future evolution of aircraft wake ducts, this invention provides an aircraft wake duct prediction method based on a spatiotemporal fusion deep learning network. It utilizes lidar to collect radial velocity field data of the aircraft wake duct, constructs a spatiotemporal fusion deep learning network architecture to extrapolate and predict the radial velocity field of the wake duct, thereby predicting the next scanning surface of the lidar in advance. Finally, the prediction results are generated and pushed to the control tower for output to various control terminals for notification, thus realizing intelligent control of aircraft wake duct intervals.

[0006] To achieve the above-mentioned objectives, the present invention provides the following technical solution: an aircraft wake prediction method based on a spatiotemporal fusion deep learning network, comprising the following steps: Step 1, a lidar performs a vertical profile scan of the aircraft wake using the RHI scanning mode to obtain radial velocity fields at multiple times; the radial velocity fields are preprocessed to extract radial velocity values ​​at various locations in the radial velocity fields, constructing a wake radial velocity field sequence; Step 2, the temporal and spatial feature maps of the wake radial velocity field sequence are extracted using the encoder of the spatiotemporal fusion deep learning network; the encoder includes a temporal encoder and a spatial encoder; Step 3, the meteorological scalar features are converted into meteorological boundary condition tensors using the tensor expansion module of the spatiotemporal fusion deep learning network, and fused with the temporal and spatial feature maps to obtain fused features; finally, the predicted wake radial velocity field is output by the decoder of the spatiotemporal fusion deep learning network.

[0007] Furthermore, step 1 specifically includes the following steps: After the lidar completes a 28-degree elevation angle scan, it obtains a radial velocity field; after t complete scans, it obtains t radial velocity fields, which are then integrated to obtain a set of t radial velocity fields P={P1,P2,...,P...} t}, where P i Let P represent the radial velocity field obtained after the i-th complete scan, i=1,2,...,t; and let P represent each radial velocity field in the statistical set P. i The radial velocity value is then normalized. ; where X i,(x,y) S represents the normalized radial velocity value at position (x, y) in the i-th radial velocity field, with a value range of [0, 1]. i,(x,y) S represents the radial velocity value at position (x, y) in the i-th radial velocity field; min S is the global minimum value. max The global maximum value is represented by the i-th radial velocity field P. i After extracting the radial velocity values ​​at each position (x, y), they are integrated into the radial velocity field P. i The corresponding wake radial velocity field X i The wake radial velocity fields of t radial velocity fields in set P constitute the wake radial velocity field sequence X={X1,X2,...,X...} t}

[0008] Furthermore, the spatial encoder includes t channels with identical structures. Each channel includes a multi-scale pyramid convolutional unit and a convolution-gated recurrent unit. The output of the convolution-gated recurrent unit in the previous channel is connected to the input of the convolution-gated recurrent unit in the next channel, and the output of the convolution-gated recurrent unit in the last channel is connected to the input of the spatiotemporal feature fusion module. The wake radial velocity field X at each moment in the wake radial velocity field sequence X... i The corresponding channel is input for processing, and the radial velocity field X of the i-th wake is obtained. i The operation after multi-scale pyramid convolution units is represented as follows: Among them, F i,S1 Represents the feature map output by the multi-scale pyramid convolutional unit in the i-th channel; Conv k×k This refers to a convolution operation with a k×k kernel in a multi-scale pyramid convolutional unit, where k=3,5,7; Concat is a concatenation operation; feature map F i,S1 The hidden state update process, where the input is into a convolutional gated recurrent unit with the same channel i, is represented as follows: ; ; ; ; where H i H represents the hidden state at time i; i-1 Let z be the hidden state at time i-1; i W is the update gate at time i; z U z b z These are the input weight matrix of the update gate, the hidden state weight matrix, and the bias, respectively. r is the sigmoid activation function; i W is the reset gate at time i; r U r b r These are the input weight matrix of the reset gate, the hidden state weight matrix, and the bias, respectively. Let W be the candidate hidden state at time i; tanh is the tanh activation function; W h U h b h These are the input weight matrix of the candidate hidden state, the hidden state weight matrix, and the bias, respectively. This represents element-wise multiplication; the feature map F output by the convolutional gated recurrent unit of the last channel t. t,S2 That is, the spatial feature map F output by the spatial encoder. S .

[0009] Furthermore, the time encoder includes t channels with identical structures. Each channel includes multiple stacked convolutional layers and momentum-driven deformable convolutional gated loop units. The output of the momentum-driven deformable convolutional gated loop unit in the previous channel is connected to the input of the momentum-driven deformable convolutional gated loop unit in the next channel. The output of the momentum-driven deformable convolutional gated loop unit in the last channel is connected to the input of the spatiotemporal feature fusion module. The wake radial velocity field X at each moment in the wake radial velocity field sequence X is... i The corresponding channel is input for processing, and the radial velocity field X of the i-th wake is obtained. i After multiple stacked convolutional layers, the feature map F is obtained. i,T1 Feature map F i,T1 The momentum-driven deformable convolution gated recurrent unit, which is input into the same channel i, stores the hidden state H from the previous time step. i-1 Feature map F at the current time i,T1 Channel concatenation is performed to jointly predict the sampling offset of the convolution kernel. The calculation formula is: Among them, Conv offset The offset prediction convolution operation is represented; Concat is the concatenation operation; feature map F i,T1 The deformable convolution output after adaptive adjustment by the momentum-driven deformable convolution gated recurrent unit is expressed as: Among them, F i,T2 The feature map is adjusted by the momentum-driven deformable convolutional gated recurrent unit in the i-th channel; w k p is the convolution weight; p0 is the output pixel position; p k This refers to the sampling location of the convolution kernel; The feature map F represents the sampling offset guided by the hidden state of the previous time step; K represents the set of sampling positions of the convolution kernel, k=1,2,...,K; i,T2 Compared to the hidden state H in the previous moment i-1 In the input recurrent gating mechanism, the momentum-driven deformable convolutional gated recurrent unit outputs the feature map F. i,T3 The input is fed into the momentum-driven deformable convolutional gated recurrent unit of the (i+1)th channel, and the feature map F output by the momentum-driven deformable convolutional gated recurrent unit of the last channel t is... t,T3 That is, the time feature map F output by the time encoder. T .

[0010] Furthermore, in step 3, the fusion feature is: Among them, F fused For fusion features; Conv 1×1 Indicates convolutional layer operation; M envFor the meteorological boundary condition tensor; F T For time feature maps; F S This is a spatial feature map.

[0011] Furthermore, the total loss function of the spatiotemporal fusion deep learning network includes a vortex-aware weighted loss function based on the wake velocity gradient and a multi-objective physical constraint loss function based on environmental perception; the vortex-aware weighted loss function L based on the wake velocity gradient... W-MSE for: ; Among them, V pred (i,j) is the predicted value at position (i,j) in the radial velocity field of the wake; V true (i,j) is the true value at position (i,j) in the radial velocity field of the wake; W(i,j) is the wake velocity gradient function; H and W are the spatial resolutions of the radial velocity field in the elevation direction and the radial distance direction, respectively, i=1,2,...,H, j=1,2,...,W, and N=H×W; α is the gradient penalty constant; For V true The gradient of (i,j); the multi-objective physical constraint loss function L based on environment perception. Struct for: ;in, For V pred The gradient of (i,j).

[0012] Furthermore, the total loss function is: Where Loss is the total loss function; For L W-MES The weights; For L Struct The weights; ;in, For weighted mapping networks; E t is the real-time meteorological feature vector; Sotfmax is the Sotfmax activation function.

[0013] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention proposes an aircraft wake prediction method based on a spatiotemporal fusion deep learning network, which can accurately predict the future evolution of the radial velocity field of the wake, achieve real-time prediction capability, meet the requirements of wake prediction timeliness in actual airport operations, effectively extract the local spatial features of wake data, make full use of the spatial structure information and local correlation of the velocity field, accurately capture the long-distance spatiotemporal dependence in the wake evolution process, improve the modeling ability of complex nonlinear evolution laws, and construct a spatiotemporal fusion deep learning network deep architecture, which can also obtain good prediction performance of wake evolution flow under limited data conditions. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of the method of the present invention.

[0016] Figure 2 This is a schematic diagram of the RHI scan model according to an embodiment of the present invention.

[0017] Figure 3 This is a schematic diagram of the spatiotemporal fusion deep learning network in an embodiment of the present invention.

[0018] Figure 4 This is a schematic diagram illustrating the extension of meteorological scalar features in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance, or suggesting any such actual relationship or order between these entities or operations.

[0021] Terminology Explanation: 1. Aircraft Wake Vortex: A pair of strong, counter-rotating vortices generated behind an aircraft during flight, posing a serious safety threat to aircraft taking off or landing subsequently.

[0022] 2. LiDAR (Light Detection and Ranging): A radar system that uses laser beams to detect the position, velocity, and other characteristics of a target.

[0023] 3. RHI Scan Mode (Range Height Indicator): Distance-height display scanning mode, with a fixed azimuth angle and a variable elevation angle, scanning on a vertical plane.

[0024] 4. Radial velocity field: The velocity component distribution along the lidar detection direction.

[0025] This invention is achieved through the following technical solutions, such as... Figure 1 As shown, an aircraft wake prediction method based on a spatiotemporal fusion deep learning network is proposed to achieve high-precision spatiotemporal fusion prediction of the radial velocity field of the aircraft wake. The method includes the following steps: Step 1, the lidar uses the RHI scanning mode to perform vertical profile scanning of the aircraft wake to obtain the radial velocity field at multiple times; the radial velocity field is preprocessed to extract the radial velocity values ​​at each position in the radial velocity field and construct the wake radial velocity field sequence.

[0026] During the data acquisition phase, a Doppler lidar system was used to detect the aircraft wake, and the lidar used the RHI scanning mode to perform vertical profile scanning. Figure 2 A schematic diagram of the RHI scanning model is shown. The scanning azimuth is fixed and set perpendicular to the runway extension direction, making the scanning plane approximately orthogonal to the aircraft's flight path. A lidar is installed at a predetermined position to the side of the airport runway. The laser beam continuously scans within a set elevation angle range in a two-dimensional plane perpendicular to the aircraft's flight path. After the aircraft passes the scanning plane, the wake vortex enters the laser scanning area. By measuring the Doppler frequency shift of the scattered echo generated by the interaction between the laser pulse and aerosol particles, the radial velocity information of the wake in the laser line-of-sight direction is obtained. Thus, after one complete elevation angle scan, a two-dimensional radial velocity field is obtained, which is used to characterize the spatial distribution of the wake at a specific moment. After t complete scans, t radial velocity fields can be obtained, which are integrated to form the set P = {P1, P2, ..., P...}. t}, where P i Let represent the radial velocity field obtained after the i-th complete scan, i=1,2,...,t, where t can be understood as the current time. This scheme predicts the wake radial velocity field at time t+1 using the radial velocity fields of the previous t time steps. The radial velocity field can be viewed as a single-channel two-dimensional matrix.

[0027] As a more specific implementation, the laser radar scanning azimuth angle is fixed at... Perpendicular to the runway direction, the scanning rate is [number] times per second. Elevation range set to The elevation resolution is The detection radius ranges from 68 to 502 meters, with a radial resolution of 10 meters. In other words, the lidar scans along the elevation direction at a rate of [number] seconds. After 14 seconds, a complete scan from bottom to top can be performed. Thus, a radial velocity field P is obtained. i In the next 14 seconds, a complete scan was performed from top to bottom. Thus, the next radial velocity field P is obtained. i+1 Each radial velocity field is stored in the form of a two-dimensional matrix with a size of H×W (H and W are the spatial resolutions of the radial velocity field in the elevation and radial distance directions, respectively). Each matrix element corresponds to the radial velocity value at the corresponding position (x, y) in the radial velocity field. The extracted radial velocity values ​​can be used as inputs for subsequent model training and prediction.

[0028] In the data preprocessing stage, to accelerate model training and improve numerical stability, each radial velocity field collected in set P is normalized to extract the radial velocity values. First, the radial velocity field P in set P is statistically analyzed. i The radial velocity values ​​are used to determine the global minimum and global maximum values. Then, the radial velocity values ​​are normalized. ; where X i,(x,y) S represents the normalized radial velocity value at position (x, y) in the i-th radial velocity field, with a value range of [0, 1]. i,(x,y) S represents the radial velocity value at position (x, y) in the i-th radial velocity field; min S is the global minimum value. max This is the global maximum value.

[0029] The i-th radial velocity field P i After extracting the radial velocity values ​​at each position (x, y), they are integrated into the radial velocity field P. i The corresponding wake radial velocity field X i The wake radial velocity fields of t radial velocity fields in set P constitute the wake radial velocity field sequence X={X1,X2,...,X...} t}, i=1,2,...,t.

[0030] Step 2: Extract the temporal and spatial feature maps of the wake radial velocity field sequence through the encoder of the spatiotemporal fusion deep learning network.

[0031] Aircraft wake evolution prediction is based on the prediction of the radial velocity field of future wakes using historical wake observation sequences. Aircraft wakes have obvious multi-scale vortex structure characteristics, with a strong vortex core structure in the central region of the wake, while the outer region shows a gradually decaying velocity distribution. Figure 3The diagram shows the structure of the spatiotemporal fusion deep learning network, which includes an encoder, a spatiotemporal feature fusion module, a channel confusion attention module, and a decoder connected in sequence. The encoder is responsible for learning the spatiotemporal features in the wake radial velocity field sequence X and gradually reducing the spatial size of the feature map. The decoder is responsible for reconstructing the feature map output by the decoder and gradually enlarging the spatial size of the feature map to obtain the final prediction result.

[0032] The encoder includes a spatial encoder and a time encoder. The spatial encoder spatially encodes the wake velocity field sequence X, and the time encoder temporally encodes the wake radial velocity field sequence X. Please continue reading. Figure 3 Assuming the current time is t, then the wake radial velocity field sequence X={X1,...,X2} collected from time 1 to time t will be used. t The time encoder and spatial encoder are input respectively, and the final decoder outputs the predicted wake radial velocity field at time t+1.

[0033] Traditional convolutional networks, due to their limited receptive field, struggle to simultaneously capture wake structures at different scales. Therefore, this scheme introduces a multi-scale pyramid convolutional structure when spatially encoding the wake radial velocity field sequence X using a spatial encoder. Wake structure features are extracted through different convolutional kernel scales. The spatial encoder comprises t identical channels, each including a multi-scale pyramid convolutional unit and a convolutionally gated recurrent unit (ConvGRU). The output of the convGRU in the previous channel is connected to the input of the convGRU in the next channel, and the output of the convGRU in the last channel is connected to the input of the spatiotemporal feature fusion module. The wake radial velocity field X at each moment in the wake radial velocity field sequence X is... i The corresponding channel is processed. Taking the processing of the i-th channel as an example, the radial velocity field X of the i-th wake is... i The operation after multi-scale pyramidal convolution units can be represented as: Among them, F i,S1 Represents the feature map output by the multi-scale pyramid convolutional unit in the i-th channel; Conv k×k The multi-scale pyramid convolutional unit performs a convolution operation with a kernel of k×k, where k=3,5,7; Concat is a concatenation operation. Through the feature maps extracted by the multi-scale pyramid convolutional unit, the model can simultaneously capture both fine-grained wake vortex core structures and large-scale wake diffusion structures.

[0034] Subsequently, feature map F i,S1 The hidden state is fed into a convolutionally gated recurrent unit (ConvGRU) with the same channel i to further model the evolution of spatial features at consecutive time steps. The hidden state update process of the convolutionally gated recurrent unit can be represented as: ; ; ; ; where H i H represents the hidden state at time i; i-1 Let z be the hidden state at time i-1; i W is the update gate at time i; z U z b z These are the input weight matrix of the update gate, the hidden state weight matrix, and the bias, respectively. r is the sigmoid activation function; i W is the reset gate at time i; r U r b r These are the input weight matrix of the reset gate, the hidden state weight matrix, and the bias, respectively. Let W be the candidate hidden state at time i; tanh is the tanh activation function; W h U h b h These are the input weight matrix of the candidate hidden state, the hidden state weight matrix, and the bias, respectively. This indicates element-wise multiplication.

[0035] Through the aforementioned recurrent gating mechanism, the model can accumulate wake structure information time-by-time, thereby obtaining the high-level semantics of the wake spatial structure. The feature map F output by the convolutional gating recurrent unit of the i-th channel is... i,S2 The feature map F output by the convolutional gated recurrent unit of the (i+1)th channel and the convolutional gated recurrent unit of the last channel t is... t,S2 That is, the spatial feature map F output by the spatial encoder. S .

[0036] The time encoder is mainly used to capture the displacement and deformation characteristics of the wake vortex core in a continuous time series. Unlike the spatial encoder, which focuses on the structural characteristics of the wake vortex core, the wake evolution process has obvious spatial displacement characteristics, similar to the optical flow changes in fluid motion. Traditional convolution kernels are fixed in spatial position, while the wake structure often exhibits irregular deformation during evolution. In order to further match the fluid momentum and physical inertia characteristics of aircraft wakes, this scheme improves the limitation of traditional deformable convolution relying only on the current moment to extract the offset, and proposes a momentum-driven deformable convolution gated recurrent unit (MD-ConvGRU). The time encoder also includes t channels with identical structures. Each channel includes multiple stacked convolutional layers (Conv) and momentum-driven deformable convolutional gated recurrent units (MD-ConvGRUs). The output of the MD-ConvGRU in the previous channel is connected to the input of the MD-ConvGRU in the next channel, and the output of the MD-ConvGRU in the last channel is connected to the input of the spatiotemporal feature fusion module. The wake radial velocity field X at each moment in the wake radial velocity field sequence X. i The corresponding channel is processed. Taking the processing of the i-th channel as an example, the radial velocity field X of the i-th wake is... i After multiple stacked convolutional layers, the feature map F is obtained. i,T1 .

[0037] Subsequently, feature map F i,T1 The momentum-driven deformable convolution gated recurrent unit (MD-ConvGRU) input to the same channel i will store the hidden state H from the previous time step (or the previous channel). i-1 Feature map F at the current time i,T1 Channel concatenation is performed to jointly predict the sampling offset of the convolution kernel. The hidden state H in the previous moment i-1 It includes the historical trajectory and evolution trend of the wake. The calculation formula is: Among them, Conv offset This represents the offset prediction convolution operation, used to process the feature map F at the current time step. i,T1 Compared to the previous hidden state H i-1 The splicing features are used for convolution calculation.

[0038] Assume the input is a feature map F i,T1 The deformable convolution output after adaptive adjustment by the momentum-driven deformable convolution gated recurrent unit can be expressed as: Among them, F i,T2 The feature map is adjusted by the momentum-driven deformable convolutional gated recurrent unit in the i-th channel; w kp is the convolution weight; p0 is the output pixel position; p k This refers to the sampling location of the convolution kernel; K represents the sampling offset guided by the hidden state of the previous time step; K represents the set of sampling positions of the convolution kernel, that is, the set of indices of all discrete sampling points within the receptive field of the convolution kernel, k=1,2,...,k.

[0039] Through this spatiotemporal coupling structure, the motion trend of the wake vortex core in the previous moment is directly used as a physical prior, guiding the model in the current moment to actively search for the possible drift location of the vortex core in space. Subsequently, the feature map F... i,T2 Compared to the hidden state H in the previous moment i-1 In the input recurrent gating mechanism, the momentum-driven deformable convolutional gated recurrent unit outputs the feature map F. i,T3 The input is fed into the momentum-driven deformable convolutional gated recurrent unit of the (i+1)th channel, and the feature map F output by the momentum-driven deformable convolutional gated recurrent unit of the last channel t is... t,T3 That is, the time feature map F output by the time encoder. T .

[0040] Step 3: The meteorological scalar features are converted into meteorological boundary condition tensors by the tensor expansion module of the spatiotemporal fusion deep learning network, and then fused with the temporal feature map and the spatial feature map to obtain the fused features; finally, the decoder of the spatiotemporal fusion deep learning network outputs the predicted wake radial velocity field.

[0041] In real-world operating scenarios, meteorological environmental conditions such as temperature, humidity, and wind speed have a significant physical modulating effect on the dissipation speed and descent trajectory of aircraft wakes. Considering that meteorological parameters are one-dimensional global scalars, and the wake's temporal characteristic diagram F... T and spatial feature map F S Belonging to two-dimensional spatial tensors, this scheme solves the problem of fusing multi-source heterogeneous data through feature tensor expansion and multimodal fusion mechanisms.

[0042] The acquired one-dimensional meteorological scalar features, such as temperature, humidity, and wind speed, are input into the tensor expansion module, such as... Figure 4 As shown, the tensor expansion module expands and replicates each one-dimensional scalar in the spatial dimension, generating a two-dimensional meteorological boundary condition tensor M with a resolution completely consistent with the temporal and spatial feature maps of the wake. env .

[0043] Then, the time feature map F T Spatial feature map F S and meteorological boundary condition tensor M env The common input spatiotemporal feature fusion module concatenates the data along the channel dimension and performs cross-channel information interaction and dimensionality reduction fusion through convolutional layers, outputting the fused feature F.fused : Among them, Conv 1×1 This indicates the operation of the convolutional layer.

[0044] To further enhance the ability of spatiotemporal fusion deep learning networks to extract key features of the wake, the fusion feature F fused The input is fed into the channel confusion attention module, where the channel confusion attention mechanism is used to process the fused feature F. fused Grouping and channel obfuscation operations are performed to promote cross-channel information fusion between different feature groups. Subsequently, through an efficient channel attention unit (ECA), a one-dimensional convolutional local cross-channel interaction strategy is used to adaptively calculate the attention weights for each channel. The calculation process is as follows: ; ; ; ;in, Represents the input features of the i-th channel group; Conv i (·) represents the convolution operation of the i-th channel group; The convolution output feature of the i-th channel group is represented by F; i = 1, 2, ..., G, where G is the total number of channel groups; ChannelShuffle represents the channel scrambling operation; F shuffle This represents the feature map after channel obfuscation, used to reassemble and interact information between different channel groups, enhancing cross-group feature fusion capabilities; GAP(·) represents the global average pooling operation; Conv1D k (·) represents a one-dimensional convolution operation with a kernel size of k, used to perform local cross-channel information interaction in the channel dimension, thereby modeling the correlation between adjacent channels; W represents the sigmoid activation function; c Indicates channel attention weights; F represents element-wise multiplication along the channel dimension; g This represents the output of the channel obfuscation attention module. The channel obfuscation attention module effectively suppresses invalid airflow noise in the background wind field and significantly enhances the ability to represent high-dimensional features of the wake vortex core and strong wind shear regions.

[0045] Ultimately, feature F g The input decoder gradually recovers the spatial resolution through upsampling and convolution operations, and predicts and generates the radial velocity field of the wake at future time steps: Where Y represents the predicted wake radial velocity field; Decoder represents the decoder. Through decoder operations, the model can recover the high-resolution wake structure while preserving the continuity of the spatial structure. Finally, Y is denormalized to output the wake radial velocity field V at the next time step. pred .

[0046] When training the spatiotemporal fusion deep learning network, the preprocessed dataset (i.e., the wake radial velocity field sequence X) is divided into three subsets: training set, validation set, and test set. The training set comprises 60% of the dataset and is used to train the model parameters and update the network weights through backpropagation. The validation set comprises 20% of the dataset and is used for model selection and hyperparameter tuning, while also monitoring overfitting during training. The test set comprises 20% of the dataset. The model training employs the Adam adaptive optimizer, which combines the advantages of momentum and RMSprop methods. It can adaptively adjust the learning rate of each parameter, exhibits high robustness to hyperparameters, and has a fast convergence speed. The initial learning rate of the Adam optimizer is set to 0.0001, the first-order moment decay coefficient is 0.9, the second-order moment decay coefficient is 0.999, and the numerical stability constant is 1e-8.

[0047] Existing deep learning prediction methods typically use Standard Mean Squared Error (MSE), but MSE treats all regions equally, which can easily lead to predictions of high-risk areas such as strong vortex cores in the wake being smoothed out by the background wind field. To address this, this invention improves the loss function of the spatiotemporal fusion deep learning network by proposing a vortex-sensing weighted loss function based on wake velocity gradients and a multi-objective physical constraint loss function based on environment perception. The vortex-sensing weighted loss function based on wake velocity gradients converges faster and more stably, but its accuracy is slightly lower; while the multi-objective physical constraint loss function based on environment perception has higher accuracy, but is more difficult to train. Therefore, this scheme combines the two, balancing training stability and convergence speed. By utilizing the smooth gradient and strong guidance of the vortex-sensing term, the problem of training difficulties and easy divergence in the physical constraint loss function is solved, while making the overall training process more stable, converging faster, and less oscillating, and reducing sensitivity to hyperparameters and initialization. Furthermore, it can significantly improve the final prediction accuracy and physical rationality, retain the high accuracy and strong physical consistency brought about by multi-objective physical constraints (such as the velocity field and vortex core structure being more in line with fluid laws), and avoid the problems of insufficient accuracy and large physical deviation when using vortex sensing loss alone.

[0048] A vortex-sensing weighted loss function based on the wake velocity gradient is used to measure the difference between the predicted velocity field and the true velocity field. The calculation formula is as follows: ; ; among which, L W-MSE V is the vortex-sensing weighted loss function based on the wake velocity gradient; pred (i,j) is the predicted value at position (i,j) in the radial velocity field of the wake; V true(i,j) represents the true value at position (i,j) in the radial velocity field of the wake; W(i,j) is the wake velocity gradient function; H and W are the spatial resolutions of the radial velocity field in the elevation direction and radial distance direction, respectively, i=1,2,...,H, j=1,2,...,W, and N=H×W; α is the gradient penalty constant, used to enhance the regression accuracy of the model in strong vortex regions; For V true The gradient of (i,j). During validation, after each epoch of training, the model performance is evaluated on the validation set. The validation set is input into the model to calculate the loss on the validation set. The validation loss is recorded to monitor overfitting. Finally, the parameters of the best-performing model on the validation set are saved as the final model.

[0049] The network is trained and optimized more complexly end-to-end based on a multi-objective physical constraint loss function that is based on environmental awareness. The calculation formula is as follows: ; among which, L Struct This is a loss function for multi-objective physical constraints based on environment perception. For V pred The gradient of (i,j).

[0050] As one possible implementation, the total loss function of a spatiotemporal fusion deep learning network is: Where Loss is the total loss function; For L W-MES The weights; For L Struct The weight of L. W-MSE Using the spatial gradient of the real velocity field Construct a weighted matrix to highlight the vortex core region while preventing the prediction ambiguity problem common in deep learning; L Struct This requires that the predicted value and the true value remain consistent in the velocity gradient space.

[0051] Unlike traditional fixed weight allocation, the total loss function introduces a lightweight weight mapping network. (e.g., a gated network consisting of two fully connected layers), whose input is the real-time meteorological feature vector E. t Through weight mapping network The adaptive weights for the two loss functions are dynamically generated. This mechanism ensures that the model can automatically adjust its emphasis on numerical accuracy, physical laws, and morphological structure based on current temperature, humidity, and wind speed conditions. The calculation formula is as follows: ;in, For weighted mapping networks; Softmanx is the activation function.

[0052] As another possible implementation method, the total loss function of the spatiotemporal fusion deep learning network is: Where Loss is the total loss function; For L W-MES The weights; For L Struct The weight.

[0053] To ensure the model focuses solely on the wake vortex core structure and suppresses noise interference, a non-smooth sparse learning function is constructed to sparsely assign weights to the two loss functions, thereby improving the model's generalization ability and making it more interpretable and consistent with physical logic. The non-smooth sparse learning function is as follows: ;in, For non-smooth coefficient learning functions; For regularization parameters; For regularization terms; These are the weights that need optimization; This is the total loss function.

[0054] Search for the optimal using momentum iteration technique Make the function To achieve the minimum value, a non-smooth sparse learning function is implemented by introducing a non-smooth regularization term to achieve weight sparsity, thereby automatically selecting features, reducing complexity, preventing overfitting, and improving interpretability and computational efficiency.

[0055] The first moment of momentum is: The second-order moment momentum is: ;constraint: and ,at the same time , Where l represents the number of momentum iterations, and the optimal momentum is found through momentum iteration techniques. When the iteration converges, for example, when it has iterated to L times, the searched... Make the non-smooth sparse learning function If the minimum value is reached, then l = 0, 1, 2, ..., L. The first-order moment decay coefficient under the l-th iteration; These are the second-order moment iteration coefficients in the l-th iteration; The weights are those for the l-th iteration; , This represents the Hadamard product. In the initial state (i.e., l=0),... , m0=0, v0=0 .

[0056] With respect to the second moment momentum v l+1 A two-stage correction is performed. The first stage uses the standard GELU activation function to apply the correction to v. l+1 Make corrections to obtain The second stage uses the Softplus activation function to... Make corrections to obtain .

[0057] The standard GELU activation function is: Where x is the parameter that needs to be updated.

[0058] Performing a second-order Taylor expansion of the standard GELU activation function at x = 0 and discarding the remainder, we obtain: ; will v l+1 Substituting x, we obtain the first stage correction. : .

[0059] The Softplus activation function is: Performing a second-order Taylor expansion of the Softplus activation function at x=0 and discarding the remainder, we obtain: ;in, As a smoothing parameter, when v0=0, it avoids T1(x) and T2(x) from being equal to 0 or approaching 0.

[0060] Will Substituting x, we obtain the second-stage correction. : ; after completing the (l+1)th iteration, the second moment momentum After correction, the weights in the (l+1)th iteration are obtained. : ;in, These are free parameters.

[0061] Step 4: Use the trained spatiotemporal fusion deep learning network to obtain the real-time radial velocity field and predict the wake radial velocity field at future moments.

[0062] In the application phase, a trained spatiotemporal fusion deep learning network is used for real-time prediction. Assuming the current time is t, the actual radial velocity field from time 1 to t is acquired via LiDAR scanning. After preprocessing, this data is output to the spatiotemporal fusion deep learning network. The model output is then inversely normalized to obtain the predicted wake radial velocity field at time t+1. At time t+1, the actual radial velocity field at time t+1 and the actual wake radial velocity field are input into the spatiotemporal fusion deep learning network. Based on the actual data from time 1 to t+1, the predicted data for time t+2 is predicted. This process is repeated to continuously improve the accuracy of the model's predictions.

[0063] The forecast results can assist air traffic controllers in dynamic flight management. Based on the predicted wake location, intensity, and diffusion trend, controllers can adjust aircraft takeoff and landing sequences and intervals in conjunction with airport runway usage, weather conditions, and air traffic flow. For example, if the forecast indicates that a wake generated by a certain aircraft will persist in a specific area, controllers can schedule subsequent aircraft to delay takeoffs and landings or adjust their flight paths to avoid the wake's influence area, thereby effectively reducing the risk of wake interference and ensuring the safety and efficiency of flight operations. Simultaneously, the actual data at the predicted time is synchronized to the dataset, providing data support for continuous model improvement.

[0064] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting aircraft wake turbulence based on a spatiotemporal fusion deep learning network, characterized in that, Includes the following steps: Step 1: The lidar uses the RHI scanning mode to perform a vertical profile scan of the aircraft wake to obtain the radial velocity field at multiple moments; the radial velocity field is preprocessed to extract the radial velocity values ​​at each position in the radial velocity field and construct the wake radial velocity field sequence. Step 2: Extract the temporal and spatial feature maps of the wake radial velocity field sequence using the encoder of the spatiotemporal fusion deep learning network; the encoder includes a temporal encoder and a spatial encoder; Step 3: The meteorological scalar features are converted into meteorological boundary condition tensors by the tensor expansion module of the spatiotemporal fusion deep learning network, and then fused with the temporal feature map and the spatial feature map to obtain the fused features; finally, the decoder of the spatiotemporal fusion deep learning network outputs the predicted wake radial velocity field.

2. The aircraft wake prediction method based on spatiotemporal fusion deep learning network according to claim 1, characterized in that, Step 1 specifically includes the following steps: After the lidar completes a 28-degree elevation angle scan, it obtains a radial velocity field. After t complete scans, t radial velocity fields are obtained, which are then integrated to form a set of t radial velocity fields P={P1,P2,...,P...} t }, where P i This represents the radial velocity field obtained after the i-th complete scan, where i = 1, 2, ..., t; Each radial velocity field P in the statistical set P i The radial velocity value is then normalized. ； Among them, X i,(x,y) S represents the normalized radial velocity value at (x,y) in the i-th radial velocity position field, with a value range of [0,1]; i,(x,y) S represents the radial velocity value at position (x, y) in the i-th radial velocity field; min S is the global minimum value. max It is the global maximum value; The i-th radial velocity field P i After extracting the radial velocity values ​​at each position (x, y), they are integrated into the radial velocity field P. i The corresponding wake radial velocity field X i The wake radial velocity fields of t radial velocity fields in set P constitute the wake radial velocity field sequence X={X1,X2,...,X...} t } 3. The aircraft wake prediction method based on spatiotemporal fusion deep learning network according to claim 1, characterized in that, The spatial encoder includes t channels with identical structures. Each channel includes a multi-scale pyramid convolutional unit and a convolutional gated loop unit. The output of the convolutional gated loop unit in the previous channel is connected to the input of the convolutional gated loop unit in the next channel. The output of the convolutional gated loop unit in the last channel is connected to the input of the spatiotemporal feature fusion module. The radial velocity field of the wake at each moment in the wake radial velocity field sequence X i The corresponding channel is input for processing, and the radial velocity field X of the i-th wake is obtained. i The operation after multi-scale pyramid convolution units is represented as follows: ； Among them, F i,S1 Represents the feature map output by the multi-scale pyramid convolutional unit in the i-th channel; Conv k×k This represents a convolution operation with a k×k kernel in a multi-scale pyramid convolution unit, where k = 3, 5, 7; Concat is the concatenation operation; i = 1, 2, ..., t; Feature map F i,S1 The hidden state update process, where the input is into a convolutional gated recurrent unit with the same channel i, is represented as follows: ； ； ； ； Among them, H i H represents the hidden state at time i; i-1 Let z be the hidden state at time i-1; i W is the update gate at time i; z U z b z These are the input weight matrix of the update gate, the hidden state weight matrix, and the bias, respectively. r is the sigmoid activation function; i W is the reset gate at time i; r U r b r These are the input weight matrix of the reset gate, the hidden state weight matrix, and the bias, respectively. Let W be the candidate hidden state at time i; tanh is the tanh activation function; W h U h b h These are the input weight matrix of the candidate hidden state, the hidden state weight matrix, and the bias, respectively. This indicates element-wise multiplication; The feature map F output by the convolutionally gated recurrent unit of the last channel t t,S2 That is, the spatial feature map F output by the spatial encoder. S .

4. The aircraft wake prediction method based on spatiotemporal fusion deep learning network according to claim 1, characterized in that, The time encoder includes t channels with identical structures. Each channel includes multiple stacked convolutional layers and momentum-driven deformable convolutional gated loop units. The output of the momentum-driven deformable convolutional gated loop unit in the previous channel is connected to the input of the momentum-driven deformable convolutional gated loop unit in the next channel. The output of the momentum-driven deformable convolutional gated loop unit in the last channel is connected to the input of the spatiotemporal feature fusion module. The radial velocity field of the wake at each moment in the wake radial velocity field sequence X i The corresponding channel is input for processing, and the radial velocity field X of the i-th wake is obtained. i After multiple stacked convolutional layers, the feature map F is obtained. i,T1 ; Feature map F i,T1 The momentum-driven deformable convolution gated recurrent unit, which is input into the same channel i, stores the hidden state H from the previous time step. i-1 Feature map F at the current time i,T1 Channel concatenation is performed to jointly predict the sampling offset of the convolution kernel. The calculation formula is: ； Among them, Conv offset This indicates an offset prediction convolution operation; Concat is a concatenation operation. Feature map F i,T1 The deformable convolution output after adaptive adjustment by the momentum-driven deformable convolution gated recurrent unit is expressed as: ； Among them, F i,T2 The feature map is adjusted by the momentum-driven deformable convolutional gated recurrent unit in the i-th channel; w k p is the convolution weight; p0 is the output pixel position; p k This refers to the sampling location of the convolution kernel; This represents the sampling offset guided by the hidden state of the previous time step; K represents the set of sampling positions of the convolution kernel, k=1,2,...,K; Feature map F i,T2 Compared to the hidden state H in the previous moment i-1 In the input recurrent gating mechanism, the momentum-driven deformable convolutional gated recurrent unit outputs the feature map F. i,T3 The input is fed into the momentum-driven deformable convolutional gated recurrent unit of the (i+1)th channel, and the feature map F output by the momentum-driven deformable convolutional gated recurrent unit of the last channel t is... t,T3 That is, the time feature map F output by the time encoder. T .

5. The aircraft wake prediction method based on spatiotemporal fusion deep learning network according to claim 1, characterized in that, In step 3, the fusion feature is: ； Among them, F fused For fusion features; Conv 1×1 Indicates convolutional layer operation; M env For the meteorological boundary condition tensor; F T For time feature maps; F S This is a spatial feature map.

6. The aircraft wake prediction method based on spatiotemporal fusion deep learning network according to claim 1, characterized in that, The total loss function of the spatiotemporal fusion deep learning network includes a vortex-aware weighted loss function based on wake velocity gradient and a multi-objective physical constraint loss function based on environment perception. Eddy sensing weighted loss function L based on wake velocity gradient W-MSE for: ； ； Among them, V pred (i,j) is the predicted value at position (i,j) in the radial velocity field of the wake; V true (i,j) is the true value at position (i,j) in the radial velocity field of the wake; W(i,j) is the wake velocity gradient function; H and W are the spatial resolutions of the radial velocity field in the elevation direction and the radial distance direction, respectively, i=1,2,...,H, j=1,2,...,W, and N=H×W; α is the gradient penalty constant; For V true The gradient of (i,j); Multi-objective physical constraint loss function L based on environmental perception Struct for: ； in, For V pred The gradient of (i,j).

7. The aircraft wake prediction method based on spatiotemporal fusion deep learning network according to claim 6, characterized in that, The total loss function is: ； Where Loss is the total loss function; For L W-MES The weights; For L Struct The weights; ； in, For weighted mapping networks; E t is the real-time meteorological feature vector; Sotfmax is the Sotfmax activation function.

Images