A method and device for predicting the long-term motion trajectory of an aerospace equipment

Abstract

The present invention discloses a method and apparatus for predicting the long-term motion trajectory of aerospace equipment, comprising: step S1, determining the input and output feature types of a long-term motion trajectory prediction task; step S2, obtaining long-term motion trajectory data for the same type or type of equipment in the same type of task; step S3, training a deep learning model based on the long-term motion trajectory data; and step S4, inputting the pending motion trajectory prediction task into the trained deep learning model for long-term motion trajectory prediction. The technical solution of the present invention improves prediction accuracy, enhances robustness, and ensures a stable and efficient training process.

Description

Technical Field

[0001] The present invention belongs to the technical field of trajectory prediction, and in particular relates to a method and device for predicting the long-term motion trajectory of aerospace equipment based on an Informer model. Background Art

[0002] In a complex and ever-changing battlefield environment, our forces must conduct numerous combat missions, including reconnaissance, maneuvering, establishing positions, and launching long-range weapons, while minimizing enemy reconnaissance and attack. Predicting the trajectory of enemy reconnaissance and strike units is crucial in the planning and execution of these missions, providing crucial information for the planning and execution of covert combat missions and significantly improving their success rate. Compared to other types of combat units, aerospace equipment has relatively fixed motion patterns. Predicting future trajectories based on historical trajectories offers high accuracy and is both feasible and practical. Currently, trajectory prediction methods for aerospace equipment primarily rely on various filtering methods. For short-term trajectory prediction, there are also methods based on geometric calculations and deep learning methods based on recurrent neural networks (RNNs) and long short-term memory networks (LSTMs). However, for long-term trajectory prediction, these existing methods suffer from numerous drawbacks, including low prediction accuracy, poor robustness, and unstable training processes. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method and device for predicting the long-term motion trajectory of aerospace equipment based on the Informer model, which improves the prediction accuracy, has strong robustness, and has a stable and efficient training process.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] A method for predicting long-term motion trajectory of aerospace equipment, comprising:

[0006] Step S1: Determine the input features and output feature types of the long-term motion trajectory prediction task;

[0007] Step S2: Acquire long-term motion trajectory data of the same type or type of equipment in the same type of task;

[0008] Step S3: training a deep learning model based on long-term motion trajectory data;

[0009] Step S4: Input the motion trajectory prediction task to be processed into the trained deep learning model for long-term motion trajectory prediction.

[0010] Preferably, in step S1, the input feature and output feature types are determined by the performance of our reconnaissance equipment and actual combat mission requirements.

[0011] Preferably, in step S3, the deep learning model is a deep learning model based on an Informer model.

[0012] Preferably, in step S3, a gradient descent-based optimizer is used to train a deep learning model based on the Informer model.

[0013] The present invention also provides a device for predicting the long-term motion trajectory of aerospace equipment, comprising:

[0014] Determination module, used to determine the input features and output feature types of the long-term motion trajectory prediction task;

[0015] An acquisition module is used to obtain long-term motion trajectory data of the same type or type of equipment in the same type of tasks;

[0016] A training module for training deep learning models based on long-term motion trajectory data;

[0017] The prediction module is used to input the motion trajectory prediction task to be processed into the trained deep learning model for long-term motion trajectory prediction.

[0018] Preferably, the input feature and output feature types are determined by the performance of our reconnaissance equipment and actual combat mission requirements.

[0019] Preferably, the deep learning model is a deep learning model based on an Informer model.

[0020] Preferably, the training module uses a gradient descent-based optimizer to train a deep learning model based on the Informer model.

[0021] The technical solution of the present invention solves the problem of long-term motion trajectory prediction of aerospace equipment. Given the non-empty historical motion trajectory of aerospace equipment, the motion state of the equipment is predicted, and a motion trajectory prediction of the equipment in the future for a long period of time is given. The motion trajectory information at each moment includes various motion attribute characteristics such as position and speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0023] Figure 1 This is a flow chart of a method for predicting the long-term motion trajectory of aerospace equipment according to an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of visualizing the long-term motion trajectory data of a certain UAV collected in an embodiment of the present invention;

[0025] Figure 3 Schematic diagram of the deep learning model structure based on the Informer model in an embodiment of the present invention;

[0026] Figure 4 This is a flow chart of the deep learning model training process in an embodiment of the present invention;

[0027] Figure 5 This is the visualization result of long-term motion trajectory prediction of a certain UAV using the long-term motion trajectory prediction method based on the Informer model in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0029] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] Example 1:

[0031] like Figure 1 As shown, an embodiment of the present invention provides a method for predicting a long-term motion trajectory of an aerospace device, wherein the aerospace device includes a drone, and predicts a future motion trajectory of 60 meters in length based on a historical motion trajectory of 30 meters in length, comprising:

[0032] Step S1: Determine the input features and output feature types of the long-term motion trajectory prediction task, wherein the input features and output feature types are determined by the performance of our reconnaissance equipment and the actual combat mission requirements;

[0033] Step S2: Acquire long-term motion trajectory data of the same type or type of equipment in the same type of task;

[0034] Step S3: constructing a deep learning model and training the deep learning model based on the long-term motion trajectory data, wherein the deep learning model is a deep learning model based on the informer model; and using a gradient descent-based optimizer to train the deep learning model based on the informer model;

[0035] Step S4: Input the motion trajectory prediction task to be processed into the trained deep learning model for long-term motion trajectory prediction.

[0036] As an implementation method of an embodiment of the present invention, in step S1, since the combat mission only focuses on the changes in the latitude and longitude coordinates of the UAV, the input features at each moment include position coordinates (2 dimensions) and speed coordinates (2 dimensions), totaling 4 dimensions, and the output features include position coordinates (2 dimensions), totaling 2 dimensions.

[0037] As an implementation method of the embodiment of the present invention, in step S2, long-term motion trajectory data is obtained by the following method, and the visualization diagram is as follows: Figure 2 As shown:

[0038] A track consisting of l+1 alternating straight line segments and l circular arcs (with a central angle no greater than π) is randomly generated within the [0,1]×[0,1] region. The probabilities of l=0, 1, 2, and 3 are set to (0.1, 0.2, 0.5, and 0.2), respectively. Track modes include pass mode and return mode. In pass mode, the unit terminates upon reaching the end of the track, with a total track length of at least 0.6. In return mode, the unit returns to the original path after reaching the end of the track, maintaining a constant speed, with a total track length of at least 0.2. The probabilities are set to (0.7, 0.3), respectively. The speed of the reconnaissance unit follows a uniform distribution over the interval [0.002, 0.005]. The sampling (simulation) interval is 1 second. Under these conditions, 100,000 tracks of length 120 are sampled. Each track contains four features, including position and velocity, at each moment.

[0039] Occlusion is imposed on each trajectory at each moment with probability p to simulate the situation of observation loss in actual combat. For the unoccluded position i in each trajectory, mask mask[i] is set to 1, otherwise mask mask[i] is set to 0, and the four-dimensional features of the corresponding x[i] in the trajectory data are cleared to zero; in order to cover trajectory data under different our reconnaissance capabilities, the occlusion probability of each trajectory is set to (0.2, 0.3, 0.4, 0.1) with p=0.0, 0.15, 0.5, 0.8 respectively. Based on the above conditions, 100,000 occluded trajectories x and the corresponding 100,000 mask sequences mask are obtained. They are randomly divided into a training set of 64,000, a validation set of 16,000, and a test set of 20,000 trajectories for deep learning model training and selection of trajectory prediction.

[0040] The historical trajectory m and the predicted trajectory length n of the prediction task are determined by the training task (the sum of the lengths does not exceed 120). Each trajectory is sampled to obtain a continuous sequence of length m+n, and the starting time of the sequence follows a uniform distribution on [1,121-mn].

[0041] As an implementation method of the embodiment of the present invention, in step S3, the deep learning model structure is as follows Figure 3 As shown in the figure, it is divided into two parts: encoder and decoder. The encoder takes the historical trajectory as input, alternately encodes it through k self-attention layers and k-1 one-dimensional convolutional layers (Conv-1D), and obtains the feature representation (Feature Map) of the historical trajectory; the decoder takes the historical trajectory and a placeholder with the same length as the predicted trajectory (the feature representation and mask are both 0, only indicating the length of the predicted trajectory) as input, alternately passes through l mutual attention layers (Cross-Attention) and l-1 one-dimensional convolutional layers (Conv-1D). At the same time, the Feature Map output by the encoder is also passed as input to each mutual attention layer, decoded by the decoder, and finally passed through a fully connected layer (Fully Connected Layer) to obtain a predicted trajectory with a length equivalent to the placeholder.

[0042] In the embodiment of the present invention, the number of attention layers k=3, the number of mutual attention layers l=2, and the length of the historical trajectory L h =30, predicted trajectory length L p =60 (long sequence prediction) or L p =1 (single point prediction), the feature dimension D in the model input and encoding and decoding process model =256, learning rate lr = 0.0001.

[0043] The self-attention layer and the mutual attention layer encode and decode the historical trajectory as a whole, which is the core component of the model. n} B×n×4 , with x i B×4 Representing the feature representation of the i-th moment in a trajectory, the calculation method of the self-attention layer and the mutual attention layer is as follows:

[0044] Through three linear layers q ,Linear k ,Linear v Transform X separately to obtain the query features of input X Key Features and Value characteristics

[0045] For each position i, calculate the query feature q i With each key feature k jThe inner product of , and do Softmax normalization transformation to get the attention weight a(q i ,k j ), so as to j The weighted summation and the introduction of the residual mechanism are used to obtain the feature representation at position i in the output tensor:

[0046]

[0047] The above formula can be expressed in tensor form as:

[0048] Z=A(Q,K)·V+X

[0049] In this formula, Q, K, and V represent Linear q (X),Linear k (X),Linear v (X), the attention tensor A is calculated as:

[0050] A(Q,K)=Softmax(Q·K)

[0051] Output Attention It is passed through the layer normalization layer LN1, and then transformed through two one-dimensional convolutional layers Conv1 and Conv2 in sequence. The residual mechanism is introduced and passed through the layer normalization layer LN2 to obtain the final output of the self-attention layer:

[0052]

[0053] In the mutual attention layer, the calculation of the Attention output Z must consider both self-attention and mutual attention. The self-attention part is the same as above and is passed to the layer normalization layer LN; the mutual attention part replaces the key feature K and value feature V in the layer with Linear k ,Linear v The transformation of the historical trajectory feature representation is performed, and then the calculation is performed according to the self-attention method. Finally, the residual mechanism is introduced for the above two items:

[0054]

[0055] In this formula Represents Linear k (Feature) and Linear v (Feature).

[0056] In the embodiment of the present invention, the input and output dimensions of each linear layer Linear are 256, the input and output dimensions of each one-dimensional convolutional layer Conv are 256, the convolution kernel size is 1, and the activation function Activate is ELU.

[0057] The feature dimension of the original data of the trajectory sequence is 4 (position and speed), which needs to be encoded first and transformed into a high-dimensional feature sequence (D model The encoding process not only expands the dimension but also needs to characterize the relative position of features at each moment in the sequence. Otherwise, the model will have difficulty coping with variable-length and missing inputs. This is also one of the key technologies for handling variable-length and missing sequence prediction tasks.

[0058] The original data encoding includes two parts: feature encoding and position encoding. Feature encoding transforms the 4-dimensional original data into D through a one-dimensional convolution layer Conv. model dimensional feature representation, the input and output dimensions of this layer are 4 and D respectively model :

[0059] Emb token (X) = Conv(X)

[0060] The position encoding uses the encoding method in Transformer, which is a timestamp related only to the moment position and dimension position:

[0061]

[0062] Among them, pos represents the moment position in the sequence, 2j and 2j+1 represent the dimension position of the feature representation, L x is a large constant. For a fixed offset step number o, PE pos+o Can be made by PE pos and PE pos+1 (or PE pos-1 ) is linearly represented with constant coefficients to ensure the relativity of position encoding to cope with variable-length and missing inputs.

[0063] The complete raw data encoding is represented as:

[0064] Emb(X)=Emb token (X)+PE

[0065] In the embodiment of the present invention, D model =256, position coding constant L x =5000.

[0066] The model loss function is defined as the weighted mean square error (WMSE) of the predicted trajectory data, where the weight is higher for the position closer to the front of the trajectory. and the true trajectory y=(y1,y2,…,y m ), the loss function is calculated as follows:

[0067]

[0068] The hyperparameter γ∈(0,1] is the attenuation rate. The larger the γ is, the more attention is paid to the prediction accuracy of the trajectory end. In this project, it is set to 0.95.

[0069] Furthermore, the optimizer in step S3 selects Adam. Figure 4 As shown in the figure, the training process adopts the standard batch training process of supervised learning: in each new training round, a fixed-size batch of training data is selected each time until the entire training set is traversed; for each batch of training data selected, the loss function is first calculated, and then the model parameters are updated based on the gradient descent method using the Adam optimizer; after each training round, the model is verified, and if the current model performs best on the verification set, the current model is saved; after all training rounds are completed, the currently saved optimal model is tested on the test set to evaluate the final performance of the model.

[0070] An example of a UAV long-term motion trajectory prediction problem is as follows: Figure 5 As shown in the figure, since the switching timing between straight flight and turning flight of the UAV is uncertain, multiple predicted trajectories within a certain range can become reasonable subsequent motion trajectories. Therefore, it is normal for there to be a certain deviation between the predicted result and the actual motion trajectory.

[0071] Example 2:

[0072] An embodiment of the present invention further provides a device for predicting the long-term motion trajectory of aerospace equipment, comprising:

[0073] Determination module, used to determine the input features and output feature types of the long-term motion trajectory prediction task;

[0074] An acquisition module is used to obtain long-term motion trajectory data of the same type or type of equipment in the same type of tasks;

[0075] A training module for training deep learning models based on long-term motion trajectory data;

[0076] The prediction module is used to input the motion trajectory prediction task to be processed into the trained deep learning model for long-term motion trajectory prediction.

[0077] As an implementation method of an embodiment of the present invention, the input feature and output feature types are determined by the performance of our reconnaissance equipment and actual combat mission requirements.

[0078] As an implementation method of an embodiment of the present invention, the deep learning model is a deep learning model based on an Informer model.

[0079] As an implementation method of an embodiment of the present invention, the training module uses a gradient descent-based optimizer to train a deep learning model based on the Informer model.

[0080] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for predicting the long-term motion trajectory of aerospace equipment, characterized in that: include: Step S1: Determine the input features and output feature types of the long-term motion trajectory prediction task; Step S2: Acquire long-term motion trajectory data of the same type or type of equipment in the same type of task; Step S3: training a deep learning model based on long-term motion trajectory data; Step S4: input the motion trajectory prediction task to be processed into the trained deep learning model for long-term motion trajectory prediction; In step S1, the input feature and output feature types are determined by the reconnaissance equipment performance and actual combat mission requirements; In step S3, the deep learning model is a deep learning model based on the Informer model; In step S3, a gradient descent-based optimizer is used to train a deep learning model based on the Informer model; In step S2, a track consisting of l+1 straight line segments and l circular arcs alternately connected is randomly generated in the area [0,1]×[0,1], with the probabilities of l=0, 1, 2, and 3 set to (0.1, 0.2, 0.5, and 0.2), respectively. The track mode includes two modes: pass mode and return mode. In pass mode, the unit terminates after reaching the end of the track, and the total track length is not less than 0.

6. In return mode, the unit returns to the original path after reaching the end of the track, keeping the speed unchanged, and the total track length is not less than 0.

2. The probabilities are set to (0.7, 0.3), respectively. The speed of the reconnaissance unit follows a uniform distribution on [0.002, 0.005], and the sampling interval is 1s. Based on the above conditions, 100,000 tracks of length 120 are obtained. Each track contains four dimensions of features, namely position and speed, at each moment. For each moment of each trajectory, occlusion is applied with probability p to simulate the situation of observation loss in actual combat. For the unoccluded position i in each trajectory, mask mask[i] is set to 1, otherwise mask mask[i] is set to 0, and the four-dimensional features of the corresponding x[i] in the trajectory data are cleared. In order to cover trajectory data under different reconnaissance capabilities, the occlusion probability of each trajectory is set to (0.2, 0.3, 0.4, 0.1) with p=0.0, 0.15, 0.5, and 0.8 respectively. Based on the above conditions, 100,000 occluded trajectories x and the corresponding 100,000 mask sequences mask are obtained. They are randomly divided into a training set of 64,000, a validation set of 16,000, and a test set of 20,000 trajectories for deep learning model training and selection of trajectory prediction. The historical trajectory m and the predicted trajectory length n of the prediction task are determined by the training task. Each trajectory is sampled to obtain a continuous sequence of length m+n, and the starting time of the sequence follows a uniform distribution on [1,121-mn]. The deep learning model structure consists of an encoder and a decoder. The encoder takes the historical trajectory as input and encodes it alternately through k self-attention layers and k-1 one-dimensional convolutional layers to obtain the feature representation of the historical trajectory. The decoder takes the historical trajectory and a placeholder of the same length as the predicted trajectory as input and alternately passes it through l mutual attention layers and l-1 one-dimensional convolutional layers. At the same time, the feature map output by the encoder is also passed as input to each mutual attention layer and decoded by the decoder. Finally, a fully connected layer is passed to obtain a predicted trajectory of the same length as the placeholder. Taken from the number of attention layers k=3, the number of mutual attention layers l=2, and the length of the historical trajectory L h =30, predicted trajectory length L p =60 or L p =1, the feature dimension D during model input and encoding and decoding model =256, learning rate lr=0.0001; The self-attention layer and the mutual attention layer encode and decode the historical trajectory as a whole; for the input tensor X={x1,…x n } B×n×4 , with x i B×4 Representing the feature representation of the i-th moment in a trajectory, the calculation method of the self-attention layer and the mutual attention layer is as follows: Through three linear layers q ,Linear k ,Linear v Transform X separately to obtain the query features of input X Key Features Sum value characteristics For each position i, calculate the query feature q i With each key feature k j The inner product of , and do Softmax normalization transformation to get the attention weight a(q i ,k j ), so as to j The weighted summation and the introduction of the residual mechanism are used to obtain the feature representation at position i in the output tensor: z i =∑ j a(q i ,k j )·v j +x i The above formula can be expressed in tensor form as: Z=A(Q,K)·V+X In this formula, Q, K, and V represent Linear q (X),Linear k (X),Linear v (X), the attention tensor A is calculated as: A(Q,K)=Softmax(Q·K) Output Attention It is passed through the layer normalization layer LN1, and then transformed through two one-dimensional convolutional layers Conv1 and Conv2 in sequence. The residual mechanism is introduced and passed through the layer normalization layer LN2 to obtain the final output of the self-attention layer: In the mutual attention layer, the calculation of the Attention output Z must consider both self-attention and mutual attention; the self-attention is fed into the layer normalization layer LN; the mutual attention part replaces the key feature K and value feature V in the layer with Linear k ,Linear v The transformation of the historical trajectory feature representation is performed, and then the calculation is performed according to the self-attention method. Finally, the residual mechanism is introduced for the above two items: in, Represents Linear k (Feature) and Linear v (Feature); The input and output dimensions of each linear layer are 256, the input and output dimensions of each one-dimensional convolutional layer Conv are 256, the convolution kernel size is 1, and the activation function Activate is ELU; The feature dimension of the original data of the trajectory sequence is 4, which is first encoded and transformed into a high-dimensional feature sequence D model dimension; The original data encoding includes two parts: feature encoding and position encoding; feature encoding transforms the 4-dimensional original data into D through a one-dimensional convolution layer Conv model dimensional feature representation, the input and output dimensions of this layer are 4 and D respectively model : Emb token (X)=Conv(X) The position encoding uses the encoding method in Transformer, which is a timestamp related only to the moment position and dimension position: Among them, pos represents the moment position in the sequence, 2j and 2j+1 represent the dimension position of the feature representation, L x is a constant; for a fixed offset step number o, PE pos+o Can be made by PE pos and PE pos+1 or PE pos-1 Expressed linearly with constant coefficients; The complete raw data encoding is represented as: Emb(X)=Emb token (X)+PE。 2. A device for predicting the long-term motion trajectory of aerospace equipment that implements the method for predicting the long-term motion trajectory of aerospace equipment according to claim 1, characterized in that: include: Determination module, used to determine the input features and output feature types of the long-term motion trajectory prediction task; An acquisition module is used to obtain long-term motion trajectory data of the same type or type of equipment in the same type of tasks; A training module for training deep learning models based on long-term motion trajectory data; The prediction module is used to input the pending motion trajectory prediction task into the trained deep learning model for long-term motion trajectory prediction; The input characteristics and output characteristics are determined by the performance of the reconnaissance equipment and the actual combat mission requirements; The deep learning model is a deep learning model based on the Informer model; The training module uses a gradient descent-based optimizer to train a deep learning model based on the Informer model.

Images