Unmanned aerial vehicle trajectory prediction method of attention network based on wavelet transform

Abstract

The invention provides an unmanned aerial vehicle trajectory prediction method of an attention network based on wavelet transform, and relates to the field of aircraft trajectory prediction. The method comprises the following steps: acquiring a plurality of historical trajectory points corresponding to the unmanned aerial vehicle in response to a trajectory prediction operation for the unmanned aerial vehicle; constructing an aircraft trajectory prediction model according to wavelet transform and an attention mechanism; and taking the plurality of historical trajectory points as input, and outputting a reconstructed trajectory and a predicted position corresponding to the unmanned aerial vehicle according to the aerial vehicle trajectory prediction model. According to the method and the device, the problems of inaccurate prediction result and poor generalization inevitably caused by predicting the trajectory of the unmanned aerial vehicle only from the time sequence perspective are solved.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle trajectory prediction, and in particular to a method for predicting the trajectory of an unmanned aerial vehicle based on a wavelet transform-based attention network. Background Art

[0002] Short-term prediction of the unmanned aerial vehicle trajectory is to predict the trajectory within a short period of time in the future based on a section of data collected historically. Since the position and attitude of the aerial vehicle are constantly changing during flight, predicting the trajectory of the unmanned aerial vehicle means predicting its position and attitude simultaneously, so it can be classified as a time series prediction problem.

[0003] Although recurrent neural networks such as LSTM are widely used in time series prediction and have achieved certain results. However, due to the characteristics of small size, high speed, and strong mobility of the unmanned aerial vehicle, its trajectory changes violently in a short time. On the other hand, the trajectory and speed of the unmanned aerial vehicle collected by radar inevitably have noise, which will further interfere with the prediction process. Therefore, predicting the trajectory only from the time series perspective will inevitably cause problems such as inaccurate prediction results and poor generalization.

[0004] Therefore, there is an urgent need for a method for predicting the trajectory of an unmanned aerial vehicle based on a wavelet transform-based attention network. Summary of the Invention

[0005] This application provides a method for predicting the trajectory of an unmanned aerial vehicle based on a wavelet transform-based attention network, which solves the problems of inaccurate prediction results and poor generalization that will inevitably occur when predicting the trajectory of an unmanned aerial vehicle only from the time series perspective.

[0006] In the first aspect of the present application, a method for predicting the trajectory of an unmanned aerial vehicle based on a wavelet transform-based attention network is provided. The method includes: in response to a trajectory prediction operation for the unmanned aerial vehicle, obtaining a plurality of historical trajectory points corresponding to the unmanned aerial vehicle; pre-training the aircraft trajectory prediction model through a simulation data set, and adjusting the aircraft trajectory prediction model through a measured data set to construct an aircraft trajectory prediction model according to wavelet transform and attention mechanism; taking the plurality of historical trajectory points as input, and through the aircraft trajectory prediction model, obtaining the original trajectory features corresponding to the unmanned aerial vehicle; through dimension mapping, mapping the original trajectory features to a high-dimensional space, and obtaining the high-dimensional trajectory features corresponding to the unmanned aerial vehicle in the high-dimensional space; by inputting the high-dimensional trajectory features, and according to a preset encoder, outputting the implicit features corresponding to the high-dimensional trajectory features, the preset encoder is composed of a recurrent neural network; by inputting the implicit features, and according to a preset decoder, outputting deep features, the deep features are the trajectory features corresponding to the unmanned aerial vehicle in different scale flight modes; by inputting the deep features, and according to the aircraft trajectory prediction model, outputting the reconstructed trajectory and predicted position corresponding to the unmanned aerial vehicle.

[0007] Optionally, pre-training the aircraft trajectory prediction model through a simulation data set specifically includes: constructing a model training data set through a simulation experiment, and dividing the model training data set into a training set, a validation set and a test set according to a preset ratio; adopting a sliding window cutting operation, and reorganizing the training set, the validation set and the test set respectively according to a preset sliding window width and a preset step length; inputting the training set, the validation set and the test set after the sliding window cutting operation into the aircraft trajectory prediction model, and updating the neural network model parameters according to a loss function.

[0008] Optionally, through dimension mapping, mapping the original trajectory features to a high-dimensional space specifically includes: constructing an embedding layer, and mapping the original trajectory features to a high-dimensional space according to the embedding layer. The embedding layer includes a linear layer and a fully connected layer, and the linear layer and the fully connected layer are used to enhance the extraction ability of trajectory features through the following formula:

[0009]

[0010] where I is the high-dimensional trajectory feature corresponding to the high-dimensional space, σ is the ReLU activation function, is the input trajectory sequence, T is the transpose matrix symbol, N represents the Nth group of samples, M is the number of multiple historical trajectory points, W l1 and W l2 are both weight matrices, which are used to map the original trajectory features to a high-dimensional space.

[0011] Optionally, before inputting the implicit feature and outputting the deep feature according to the preset decoder, the method further includes: constructing a wavelet attention function for updating network parameters to facilitate the preset decoder to obtain the time-frequency representation corresponding to the trajectory attributes of the UAV at different scales. The wavelet attention function is represented by the following formula:

[0012]

[0013] where represents the level of wavelet analysis, L represents the total number of levels of wavelet analysis, represents the wavelet loss function value at the k-th level, WTC k represents the wavelet transform coefficient, d = 6 represents the number of the trajectory attributes, and the number of the trajectory attributes is three coordinate positions and the velocities in three directions, represents the i-th element in the j-th trajectory attribute output by the (k + 1)-th preset decoder.

[0014] Optionally, inputting the implicit feature and outputting the deep feature according to the preset decoder specifically includes: inputting the implicit feature, performing a first operation according to the preset decoder and outputting an enhanced trajectory embedding feature, where the first operation is used to multiply the implicit feature by the attention score; inputting the enhanced trajectory embedding feature, performing a second operation according to the preset decoder and outputting a semantic embedding feature, where the second operation is used to perform a convolution operation on the enhanced trajectory embedding feature by using a shifted wavelet function; inputting the semantic embedding feature, performing a third operation according to the preset decoder and outputting a wavelet embedding feature, where the third operation is used to convert the semantic embedding feature into a wavelet embedding feature through a recurrent neural network; inputting the wavelet embedding feature, performing a fourth operation according to the preset decoder and outputting the deep feature, where the fourth operation is used to perform normalization processing and linear mapping processing.

[0015] Optionally, performing a fourth operation according to the preset decoder and outputting the deep feature specifically includes: outputting the deep feature according to the following formula:

[0016]

[0017] where Q is the deep feature, H w is the wavelet embedding feature, C is the semantic embedding feature, H h is the enhanced trajectory embedding feature, LN() represents normalization processing, FC() represents linear mapping processing, LSTM represents recurrent neural network processing, Conv1d() represents one-dimensional convolution processing, σ represents the ReLU activation function, E represents the attention score, Diag() represents diagonal matrix processing, c iis the state of the i-th intermediate layer, where, is the state of the last hidden layer of the preset encoder, is the state of the first hidden layer of the LSTM, is the hidden layer state, S 2 represents the number of recurrent layers in the recurrent neural network, D is the dimension of the high-dimensional space, R represents the real number field, and H is the implicit feature.

[0018] Optionally, by inputting deep features and according to the aircraft trajectory prediction model, the predicted position corresponding to the unmanned aircraft is output, specifically including: outputting the predicted position corresponding to the unmanned aircraft according to the following formula:

[0019]

[0020] where, is the reconstructed trajectory corresponding to the unmanned aircraft, is the predicted position corresponding to the unmanned aircraft, IDWT() is the operation of performing inverse discrete wavelet transform, operation, Q 0 ,Q 1 ,…,Q L represents deep features at different scales.

[0021] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0022] 1. When the user performs a trajectory prediction operation on the unmanned aircraft, multiple historical trajectory points corresponding to the unmanned aircraft are obtained, an aircraft trajectory prediction model is constructed based on wavelet transform and attention mechanism, and the multiple historical trajectory points are used as inputs. Through the aircraft trajectory prediction model, the original trajectory features corresponding to the unmanned aircraft are obtained. Then, through dimension mapping, the original trajectory features are mapped to a high-dimensional space, and the high-dimensional trajectory features corresponding to the unmanned aircraft in the high-dimensional space are obtained. After that, by inputting the high-dimensional trajectory features and according to the preset encoder, the implicit features corresponding to the high-dimensional trajectory features are output. And by inputting the implicit features and according to the preset decoder, deep features are output. Furthermore, by inputting the deep features and according to the aircraft trajectory prediction model, the reconstructed trajectory and predicted position corresponding to the unmanned aircraft are output. This method effectively improves the accuracy of unmanned aircraft trajectory prediction and solves the problem that predicting the trajectory of an unmanned aircraft only from the time series perspective will inevitably result in inaccurate prediction results and poor generalization.

[0023] 2. A simulation experiment is used to construct a model training dataset, and the model training dataset is divided into a training set, a validation set, and a test set according to a preset ratio. Moreover, a sliding window cutting operation is adopted, and the training set, the validation set, and the test set are reorganized respectively according to a preset sliding window width and a preset step size. Thus, the training set, the validation set, and the test set after the sliding window cutting operation are input into the aircraft trajectory prediction model, effectively avoiding data overfitting and improving the reliability of the simulation experiment construction model in practical applications.

[0024] 3. By inputting implicit features, a first operation is performed according to a preset decoder and enhanced trajectory embedding features are output; by inputting the enhanced trajectory embedding features, a second operation is performed according to the preset decoder and semantic embedding features are output; by inputting the semantic embedding features, a third operation is performed according to the preset decoder and wavelet embedding features are output; by inputting the wavelet embedding features, a fourth operation is performed according to the preset decoder and deep features are output. The fourth operation is used to perform normalization processing and linear mapping processing, thus effectively improving the feature extraction ability and stability of the aircraft trajectory prediction model, enabling the prediction result to more accurately reflect the trajectory change of the unmanned aircraft in a complex environment, and enhancing the expression ability and prediction accuracy of the aircraft trajectory prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 FIG. is a schematic flowchart of a method for predicting the trajectory of an unmanned aircraft based on a wavelet transform-based attention network provided by an embodiment of the present application;

[0026] Figure 2 FIG. is a schematic structural diagram of a wavelet transform-based attention network provided by an embodiment of the present application;

[0027] Figure 3 FIG. is a schematic diagram of a wavelet attention module provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments.

[0029] The terms used in the following embodiments of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification of the present application, the singular forms "a", "an", "the", "above-mentioned", "said", and "this" are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in the present application refers to and includes any or all possible combinations of one or more of the listed items.

[0030] Hereinafter, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as implying or suggesting relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the embodiments of the present application, unless otherwise specified, the meaning of "a plurality" is two or more.

[0031] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0032] Please refer to Figure 1 , which shows a schematic flowchart of a method for predicting the trajectory of an unmanned aerial vehicle based on a wavelet transform-based attention network provided by an embodiment of the present application. The flowchart mainly includes the following steps: S101 to S107.

[0033] Step S101, in response to a trajectory prediction operation for an unmanned aerial vehicle, obtain a plurality of historical trajectory points corresponding to the unmanned aerial vehicle.

[0034] Specifically, please refer to Figure 2 , which shows a schematic diagram of a wavelet transform-based attention network structure provided by an embodiment of the present application. According to the principles of wavelet transform and attention mechanism, for the timing problem in the trajectory prediction operation of an unmanned aerial vehicle, wavelet transform can perform multi-time scale feature analysis on the signal from the frequency domain perspective; while the attention mechanism captures the features in the timing by learning the relationship between each moment and all other moments. Therefore, we adopt a combination of wavelet transform and self-attention mechanism to construct a wavelet transform-based attention network.

[0035] Step S102, pre-train the aircraft trajectory prediction model through a simulation data set, and adjust the aircraft trajectory prediction model through a measured data set to construct an aircraft trajectory prediction model according to wavelet transform and attention mechanism.

[0036] Specifically, a simulation experiment is used to generate a model training data set, and the data set is divided into a training set, a validation set, and a test set according to a preset ratio. Next, a sliding window cutting operation is adopted to reorganize the training set, the validation set, and the test set. Specifically, according to the preset sliding window width and step size, each data set is cut into multiple subsequences to ensure that the temporal characteristics of the data are effectively retained. Finally, based on these data sets after sliding window cutting, an aircraft trajectory prediction model is constructed to ensure that the model can be trained and tested on diverse input data, thereby improving its prediction accuracy and generalization ability.

[0037] In a possible implementation manner, step S102 further includes: pre-training the aircraft trajectory prediction model through a simulation data set, specifically including: constructing a model training data set by means of a simulation experiment, and dividing the model training data set into a training set, a validation set, and a test set according to a preset ratio; adopting a sliding window cutting operation, and reorganizing the training set, the validation set, and the test set respectively according to the preset sliding window width and the preset step size; inputting the training set, the validation set, and the test set after the sliding window cutting operation into the aircraft trajectory prediction model, and updating the neural network model parameters according to the loss function.

[0038] Specifically, model training requires splitting the data, and the purpose of splitting is to maintain the generalization ability of the model, prevent the model from overfitting on the known data, and lack the ability to explain the position data. The training set, the validation set, and the test set are divided according to the ratio of 0.8:0.10:0.10. The training set is used to train the model, and the model parameters are updated by finding the gradient in the training set, and finally the training data is fitted. The validation set is used to check the training status of the model and adjust the model training parameters. The test set is used to finally verify the generalization prediction ability of the model and serve as the test data for testing the performance of the model. Before using the model for multi-step prediction of this time series, the training set, the validation set, and the test set are reorganized respectively in the way of sliding window cutting to meet the data requirements of the model. At the same time, the sliding window cutting can increase the sample size to facilitate the mining of time series patterns. Since the sampling interval of the trajectory is 0.02s, the sliding window width of the training set data is set to 10 and the step size is set to 1. The first 9 points of each sample are the input, and the 10th point is the true label. Each variable is cyclically cut to obtain a set of sliding window data segments:

[0039]

[0040] where X set is the input set, is the true label; the elements in X set and correspond one by one, the length of the elements in set is 1 more than that of X

[0041] Step S103: Use multiple historical trajectory points as inputs, and obtain the original trajectory features corresponding to the unmanned aerial vehicle through the aircraft trajectory prediction model.

[0042] Step S104: Map the original trajectory features to a high-dimensional space through dimensionality mapping, and obtain the high-dimensional trajectory features corresponding to the unmanned aerial vehicle in the high-dimensional space.

[0043] Specifically, each historical trajectory point is first input into a fully connected layer to be converted into a high-dimensional vector feature, that is, the high-dimensional trajectory feature. The purpose of this step is to enhance the features of the input trajectory and convert it to a dimension suitable for subsequent transformations. Each point on the unmanned aerial vehicle trajectory represents a low-dimensional position and velocity vector in three-dimensional space.

[0044] In a possible implementation manner, step S104 further includes: constructing an embedding layer, and mapping the original trajectory features to a high-dimensional space according to the embedding layer. The embedding layer includes a linear layer and a fully connected layer, and the linear layer and the fully connected layer are used to enhance the extraction ability of trajectory features through the following formula.

[0045] Specifically, in order to better extract the implicit features of the flight trajectory for the subsequent network, a low-dimensional vector is mapped to a high-dimensional space through an embedding layer. Therefore, this layer uses a linear layer and a fully connected layer to enhance the extraction ability of trajectory features. The calculation formula is as follows:

[0046]

[0047] where \(I = \mathbb{R}\) M×D is the extracted high-dimensional feature, \(D\) is the vector dimension corresponding to the high-dimensional trajectory feature, \(M\) is the number of historical trajectory points. Among them, \(I\) is the high-dimensional trajectory feature corresponding to the high-dimensional space, \(\sigma\) is the ReLU activation function, is the input trajectory sequence, and the historical trajectory point is a single element in \(X\) in S102 set The superscript \(T\) represents the transpose matrix symbol, \(N\) represents the \(N\)th group of samples, \(M\) is the number of multiple historical trajectory points, \(W\) l1 and \(W\) l2 are both weight matrices, used to map the original trajectory features to a high-dimensional space. and are weight matrices, represents the real number field, here a field of dimension \(d\times M\), and \(d\) is the vector dimension corresponding to the original trajectory feature.

[0048] Step S105: Input the high-dimensional trajectory features, and output the implicit features corresponding to the high-dimensional trajectory features according to a preset encoder. The preset encoder is composed of a recurrent neural network.

[0049] Specifically, the high-dimensional trajectory features are input into an encoder composed of a recurrent neural network, namely a preset encoder, and encoded into implicit feature vectors. The preset encoder is used to extract the implicit features of the trajectory data for subsequent prediction. The input embedding layer can extract the internal features of the trajectory points, such as the dynamic characteristics of a certain point. However, high-level abstract semantic features can provide richer features for trajectory prediction. Therefore, an encoder based on a recurrent neural network is used to construct the temporal dependence of the input trajectory data, extract its dynamic characteristics, and thus provide a more refined internal relationship of the sequence.

[0050] In a possible implementation manner, step S105 further includes: constructing a preset encoder based on a recurrent neural network, and outputting the hidden layer state according to the preset encoder; outputting the hidden layer state as the implicit feature corresponding to the high-dimensional trajectory features.

[0051] Specifically, an LSTM network is used to implement temporal modeling. The features obtained by the input embedding layer are input into the encoder, and the hidden layer state is output:

[0052] H,(h e ,c e )=LSTM(l,h 0 ,c 0 );

[0053] Where are the hidden gate state and forget gate state initialized to 0. is the input of the LSTM layer, and its output consists of two parts: are all the hidden gate states of the last layer of the LSTM, represents the hidden gate state and forget gate state of the last time step, S 1 is the number of hidden layers.

[0054] Step S106, input the implicit features, and output the deep features according to the preset decoder. The deep features are the trajectory features corresponding to the flight modes of the unmanned aerial vehicle at different scales.

[0055] Specifically, by inputting the implicit features, deep features of flight modes at different scales are extracted according to multiple preset decoders, that is, the deep features, and the deep features are the trajectory features corresponding to the flight modes of the unmanned aerial vehicle at different scales.

[0056] In a possible implementation, step S106 further includes: inputting implicit features, performing a first operation according to a preset decoder and outputting enhanced trajectory embedding features, where the first operation is used to multiply the implicit features by attention scores; inputting the enhanced trajectory embedding features, performing a second operation according to the preset decoder and outputting semantic embedding features, where the second operation is used to perform a convolution operation on the enhanced trajectory embedding features using a shifted wavelet function; inputting the semantic embedding features, performing a third operation according to the preset decoder and outputting wavelet embedding features, where the third operation is used to pass through; inputting the wavelet embedding features, performing a fourth operation according to the preset decoder and outputting deep features, where the fourth operation is used to perform normalization processing and linear mapping processing.

[0057] Specifically, a wavelet attention function is constructed, and the wavelet attention function is used to update network parameters so as to facilitate the preset decoder to represent the time-frequency corresponding to the trajectory attributes of the unmanned aerial vehicle at different scales. The wavelet attention function is represented by the following formula:

[0058]

[0059] where l represents the level of wavelet analysis, L represents the total number of levels of wavelet analysis, represents the wavelet loss function value at the k-th level, WTC k represents the wavelet transform coefficient, d = 6 represents the number of the trajectory attributes, and the number of the trajectory attributes is 3 coordinate positions and 3 velocity directions, represents the i-th element in the j-th trajectory attribute output by the (k + 1)-th preset decoder. For the wavelet analysis of the l-th layer, each preset decoder is used to generate a sub-band of a specific time-frequency feature, as shown in the formula:

[0060] Q i = Decoder i (H), i = 0, 1, …, L

[0061] where Q i represents the i-th layer wavelet transform, H is the implicit feature, and Decoder i () represents performing the i-th layer decoding operation on the input H. To enhance the learning ability, a wavelet attention module is added before each preset decoder, which is a self-attention mechanism. The input is the output feature of the previous layer preset decoder, with the shape of [N, L, H], where N is the batch size, L is the sequence length, and H is the hidden layer feature dimension. The output is the input feature and attention weight score of the preset decoder, with the shapes of [N, L′, H] and [N, L′], where L′ is the output sequence length.

[0062] The preset decoder uses the LSTM layer to extract high-dimensional features of the trajectory embedding layer, providing a stable feature for the input trajectory sequence. In the proposed network, the main goal is to predict wavelet coefficients for inverse discrete wavelet transform. However, if only wavelet features themselves are used, such as splitting the time series and inputting each part into the prediction model, it is difficult to provide features with multi-scale time resolution. Therefore, a wavelet attention module is adopted here to learn trajectory features at different scales. Please refer to Figure 3 , which shows a schematic diagram of a wavelet attention module provided by an embodiment of the present application. The wavelet attention module consists of two parts. The first part is used to perform the first operation, that is, multiplying the implicit feature by the attention score. The second part is used to perform the second operation, that is, performing a convolution operation on the enhanced trajectory embedding feature using a shifted wavelet function. To further enhance the feature extraction ability, the trajectory embedding feature obtained by the encoder is multiplied by the attention score to obtain an enhanced feature. Then, the convolution operation extracts features from the obtained enhanced trajectory embedding feature to obtain a semantic embedding feature. The wavelet attention module can be expressed as:

[0063]

[0064] where is the enhanced trajectory embedding feature after linear transformation, and are weight matrices used to linearly transform the enhanced trajectory embedding feature; is also a weight matrix used to explore the importance of each enhanced trajectory embedding feature. σ represents the ReLU activation function, and γ represents the Sigmoid activation function; represents the attention score, which is used to represent the importance of historical trajectory points at each moment. Diag() represents diagonal matrix processing, c i is the i-th intermediate layer state. Among them, is the state of the last hidden layer of the preset encoder, is the state of the first hidden layer of the LSTM, is the hidden layer state, S 2 represents the number of recurrence layers in the recurrent neural network, D is the dimension of the high-dimensional space, represents the real number field, and H is the implicit feature. The original trajectory embedding feature and the trajectory embedding feature multiplied by the weight coefficient are added together to obtain the enhanced trajectory embedding feature. After obtaining the enhanced trajectory embedding feature, a one-dimensional convolution operation is used to convert it into a semantic embedding feature This step is expressed as:

[0065] C = σ(Conv1d(H h ))

[0066] To determine the time dimension C, the length of the wavelet attention module at the L-th layer is reduced to 1 / 2 of the original length L , Conv1d() represents one-dimensional convolution processing, and H h To enhance the trajectory embedding features. At the same time, the discrete wavelet transform uses a shifted wavelet function to perform a convolution operation on the time series. When the filter approaches the edge of the finite signal, the convolution operation needs to obtain values beyond the signal boundary through signal extension. Therefore, the actual length of the wavelet attention module is determined by the wavelet and the signal extension mode. Symmetric extension is usually used to ensure the continuity of the signal boundary. Therefore, the length of the L-th layer is:

[0067]

[0068] where h L is the length of the L-th layer, h L-1 is the length of the (L - 1)-th layer, h i is the length of the wavelet transform coefficient vector at the i-th layer, l is half of the preset filter length, represents the floor operation. After performing the second operation according to the preset decoder and outputting the semantic embedding features, the wavelet embedding features are output through the third operation, that is, the semantic embedding features are transformed into wavelet embedding features through a recurrent neural network:

[0069]

[0070] where H w is the wavelet embedding feature, is the hidden layer state of the last layer of the preset encoder, is the hidden layer state of the first layer of the LSTM, is the state of the hidden layer, which is also initialized with 0, S 2 represents the number of recurrence layers in the recurrent neural network, D is the dimension of the high-dimensional space, represents the real number field. Retrieve the long-term memory of the encoder, initialize the LSTM block with the prior memory of the historical trajectory sequence, and enhance its ability to utilize scale-oriented features. LN() represents the normalization process, and FC() represents the linear mapping process.

[0071] Step S107, by inputting deep features and according to the unmanned aerial vehicle trajectory prediction model, output the reconstructed trajectory and predicted position corresponding to the unmanned aerial vehicle.

[0072] Specifically, by inputting deep features, the reconstructed trajectory and predicted position of the unmanned aerial vehicle are obtained from the outputs of multiple preset decoders according to the inverse discrete wavelet transform. The predicted position is the specific position of the unmanned aerial vehicle at the predicted next moment. The predicted position corresponding to the unmanned aerial vehicle can be output according to the following formula:

[0073]

[0074] Wherein, is the reconstructed trajectory corresponding to the unmanned aerial vehicle, is the predicted position corresponding to the unmanned aerial vehicle, IDWT() is the operation of performing inverse discrete wavelet transform, and Q 0 ,Q 1 ,…,Q L represent deep features at different scales. The inverse discrete wavelet transform operation performs inverse transformation on the wavelet coefficients of each attribute, thereby reconstructing the input historical trajectory sequence and predicting the trajectory points at the next moment. Predefined reconstruction filters and coefficient transformation matrices In each attribute, the reconstruction filter restores a series of attributes in the time domain by iteratively combining low-frequency and high-frequency coefficient pairs. For the jth attribute, the corresponding time series is:

[0075]

[0076] Wherein, Trim() is used to trim the redundant parts in the time series, and lDWT() is the operation of performing inverse discrete wavelet transform, only retaining the first M + 1 elements.

[0077] By adopting the above method in this application, when the user performs a trajectory prediction operation on the unmanned aerial vehicle, multiple historical trajectory points corresponding to the unmanned aerial vehicle are obtained, a flight vehicle trajectory prediction model is constructed according to wavelet transform and the attention mechanism, and the multiple historical trajectory points are used as inputs. Then, through the flight vehicle trajectory prediction model, the original trajectory features corresponding to the unmanned aerial vehicle are obtained. Thus, through dimension mapping, the original trajectory features are mapped to a high-dimensional space, and the high-dimensional trajectory features corresponding to the unmanned aerial vehicle in the high-dimensional space are obtained. After that, by inputting the high-dimensional trajectory features and according to the preset encoder, the implicit features corresponding to the high-dimensional trajectory features are output. And by inputting the implicit features and according to the preset decoder, deep features are output. Furthermore, by inputting the deep features and according to the flight vehicle trajectory prediction model, the reconstructed trajectory and predicted position corresponding to the unmanned aerial vehicle are output. This method effectively improves the accuracy of unmanned aerial vehicle trajectory prediction and solves the problem that predicting the trajectory of the unmanned aerial vehicle only from the time series perspective will inevitably result in inaccurate prediction results and poor generalization.

[0078] The above are only exemplary embodiments disclosed in the present application and should not be used to limit the scope of the present application. That is, any equivalent changes and modifications made in accordance with the teachings disclosed in the present application still fall within the scope covered by the present application. Those skilled in the art will readily think of other embodiments of the present application after considering the specification and the disclosure of the practical truth.

[0079] The present application aims to cover any variations, uses or adaptations of the present application, which follow the general principles of the present application and include common general knowledge or conventional technical means in the technical field not recorded in the present application.

Claims

1. A method for unmanned aerial vehicle trajectory prediction based on wavelet transform attention network, characterized in that: The method comprises: In response to a trajectory prediction operation for an unmanned aerial vehicle, obtaining a plurality of historical trajectory points corresponding to the unmanned aerial vehicle; Pre-training an aircraft trajectory prediction model through a simulation data set, and adjusting the aircraft trajectory prediction model through a measured data set, so as to construct an aircraft trajectory prediction model according to a wavelet transform and an attention mechanism; Taking the plurality of historical trajectory points as input, and obtaining the original trajectory features corresponding to the unmanned aerial vehicle through the aircraft trajectory prediction model; By dimensional mapping, the original trajectory features are mapped to a high-dimensional space, and the high-dimensional trajectory features corresponding to the unmanned aerial vehicle in the high-dimensional space are obtained; By inputting the high-dimensional trajectory feature, and outputting the implicit feature corresponding to the high-dimensional trajectory feature according to a preset encoder, the preset encoder is composed of a recurrent neural network; By inputting the implicit features and outputting deep features according to a preset decoder, the deep features are trajectory features corresponding to the unmanned aerial vehicle in flight modes of different scales; By inputting the deep-level features and outputting the reconstructed trajectory and predicted position corresponding to the unmanned aerial vehicle according to the aircraft trajectory prediction model.

2. The method according to claim 1, characterized in that The pre-training of the aircraft trajectory prediction model by using the simulation data set specifically includes: A simulation experiment is used to construct a model training data set, and the model training data set is divided into a training set, a validation set, and a test set according to a preset ratio; A sliding window cutting operation is adopted, and the training set, the validation set, and the test set are reorganized respectively according to a preset sliding window width and a preset step size; The training set, the validation set and the test set after the sliding window cutting operation are input into the aircraft trajectory prediction model, and the neural network model parameters are updated according to the loss function.

3. The method according to claim 1, characterized in that Mapping the original trajectory features to a high-dimensional space through dimensional mapping specifically includes: An embedding layer is constructed, and the original trajectory features are mapped to the high-dimensional space according to the embedding layer, wherein the embedding layer includes a linear layer and a fully connected layer, and the linear layer and the fully connected layer are used to enhance the extraction capability of trajectory features through the following formula: Wherein, I is the high-dimensional trajectory feature corresponding to the high-dimensional space, σ is the ReLU activation function, is the input trajectory sequence, T is the transposed matrix symbol, N represents the Nth group of samples, M is the number of multiple historical trajectory points, W l1 and W l2 They are all weight matrices, used to map the original trajectory features to the high-dimensional space.

4. The method according to claim 1, characterized in that: Outputting the implicit features corresponding to the high-dimensional trajectory features according to the preset encoder specifically includes: Constructing the preset encoder based on a recurrent neural network, and outputting a hidden layer state according to the preset encoder; The hidden layer state is output as the implicit feature corresponding to the high-dimensional trajectory feature.

5. The method according to claim 1, characterized in that The step of inputting the implicit features and outputting the deep features according to the preset decoder specifically includes: By inputting the implicit feature, performing a first operation according to the preset decoder and outputting an enhanced trajectory embedding feature, wherein the first operation is used to multiply the implicit feature by the attention score; By inputting the enhanced trajectory embedding feature, performing a second operation according to the preset decoder and outputting a semantic embedding feature, wherein the second operation is used to perform a convolution operation on the enhanced trajectory embedding feature using a shifted wavelet function; By inputting the semantic embedding feature, performing a third operation according to the preset decoder and outputting a wavelet embedding feature, the third operation is used to convert the semantic embedding feature into the wavelet embedding feature through a recurrent neural network; By inputting the wavelet embedding feature, performing a fourth operation according to the preset decoder and outputting the deep-level feature, the fourth operation is used to perform normalization processing and linear mapping processing on the wavelet embedding feature.

6. The method according to claim 5, characterized in that The method further includes, before the step of inputting the implicit features and outputting the deep features according to the preset decoder, the method further includes: A wavelet attention function is constructed, and the wavelet attention function is used to update network parameters so that the preset decoder can obtain the time-frequency representation of the trajectory attributes of the unmanned aerial vehicle at different scales. The wavelet attention function is represented by the following formula: in, It represents the level of wavelet analysis, L represents the total number of levels of wavelet analysis, represents the wavelet loss function value at the kth level, h L-k Indicates WTC k Length, WTC k represents the wavelet transform coefficient, d=6 represents the number of the trajectory attributes, the number of the trajectory attributes is 3 coordinate positions and 3 directions of speed, Represents the i-th element in the j-th trajectory attribute output by the k+1-th preset decoder.

7. The method according to claim 6, characterized in that Before adding the wavelet attention module before each of the preset decoders, the method further includes constructing the wavelet attention module, specifically including: The wavelet attention module is constructed by the following formula: in is the enhanced trajectory embedding feature after linear transformation, and is a weight matrix used to linearly transform the enhanced trajectory embedding feature; It is also a weight matrix used to explore the importance of each of the enhanced trajectory embedding features, σ represents the ReLU activation function, and γ represents the Sigmoid activation function; Represents the attention score, which is used to indicate the importance of the historical trajectory point at each moment.

8. The method according to claim 5, characterized in that The performing the fourth operation according to the preset decoder and outputting the deep-level features specifically includes: The deep-level features are output according to the following formula: Among them, Q is the deep feature, H w is the wavelet embedding feature, C is the semantic embedding feature, H h is the enhanced trajectory embedding feature, LN() represents the normalization processing, FC() represents the linear mapping processing, LSTM represents the recurrent neural network processing, Conv1d() represents the one-dimensional convolution processing, σ represents the ReLU activation function, E represents the attention score, Diag() represents the diagonal matrix processing, c i is the i-th intermediate layer state, where is the last hidden layer state of the preset encoder, is the first hidden layer state of LSTM, is the hidden layer state, S2 represents the number of cyclic layers in the recurrent neural network, D is the dimension of the high-dimensional space, represents the real number domain, and H is the implicit feature.

9. The method according to claim 1, characterized in that: The step of inputting the deep-level features and outputting the predicted position corresponding to the unmanned aerial vehicle according to the aircraft trajectory prediction model specifically includes: The predicted position of the unmanned aerial vehicle is output according to the following formula: in, is the reconstructed trajectory corresponding to the unmanned aerial vehicle, is the predicted position corresponding to the unmanned aerial vehicle, IDWT() is an inverse discrete wavelet transform operation, operation, Q0, Q1, ..., Q L Represents the deep features at different scales.

10. The method according to claim 1, characterized in that Each of the preset decoders is used to generate a sub-band with specific time-frequency characteristics, as shown in the formula: Q i =Decoder i (H),i=0,1,…,L Among them, Q i represents the wavelet transform of the i-th layer, H is the implicit feature, and Decoder i () indicates the i-th layer decoding operation on the input H.

Images