CNN-AM-BiLSTM neural network model-based track prediction method

Abstract

The invention discloses a track prediction method based on a CNN-AM-BiLSTM neural network model, and the method comprises the steps: collecting the track data of an air target through radar equipment, and constructing a data set; reducing the track data to a range of [0, 1] by using a normalization algorithm, and dividing into a training set, a test set and a verification set; building a CNN-AM-BiLSTM neural network model, wherein the CNN-AM-BiLSTM neural network model comprises a convolution module, an attention mechanism module, a sequence extraction module and an output module combination; taking the data set as the input of a neural network model, starting to train the neural network model, continuously adjusting hyper-parameters in the training process, and obtaining an optimal hyper-parameter group according to the change of a loss function; and inputting air target track data detected by radar equipment into the neural network model, and outputting future track data of the air target after model checking calculation. According to the method, the flight paths of various aerial targets can be predicted, and the prediction precision is improved.

Description

Technical Field

[0001] The present invention belongs to the technical fields of trajectory prediction and deep learning. More specifically, it relates to a trajectory prediction method based on a CNN-AM-BiLSTM neural network model. Background Art

[0002] With the development of information technology and artificial intelligence technology in modern warfare, more and more high-tech applications are used on aerial targets such as unmanned aerial vehicles and fighter jets, making aerial targets more efficient and intelligent. This enables aerial targets to respond more quickly when facing ground air defense weapons and may also make some complex movements after being locked by ground air defense weapons to get rid of the lock. This series of operations will bring great interference to the aiming and shooting of our air defense weapons. There are various types of air raid weapons in modern warfare with different performances. Relying solely on a certain weapon cannot fully achieve effective air defense. It is necessary to comprehensively use various air defense weapons to build a complete multi-layer air defense system with long, medium, and short ranges, and high, medium, and low altitudes, and conduct interception layer by layer to achieve the purpose of effective air defense. Artillery is the main weapon for long-range strikes and fire suppression now, but it is very passive in the face of aerial targets. Therefore, future artillery should also have a certain air defense ability. Air defense weapons mainly rely on radar to detect aerial targets at a long distance. Using radar to predict the future trajectory of targets is of great significance for improving the shooting accuracy of artillery air defense.

[0003] At present, the artillery system does not have air defense capabilities, and traditional air defense weapons such as self-propelled anti-aircraft guns mostly use traditional algorithms to predict the trajectories of aerial targets. This method cannot make full use of historical trajectory data, and the algorithm has insufficient generality and low prediction accuracy. Summary of the Invention

[0004] A trajectory prediction method based on a CNN-AM-BiLSTM neural network model is provided to achieve the prediction of the trajectories of various aerial targets and improve the prediction accuracy.

[0005] The technical solution to achieve the purpose of the present invention is as follows:

[0006] A trajectory prediction method based on a CNN-AM-BiLSTM neural network model includes:

[0007] Step 1: Collect the trajectory data of aerial targets through radar equipment and construct a data set;

[0008] Step 2: Use the min-max normalization algorithm to scale the trajectory data to the range of [0, 1], and divide it into a training set, a test set, and a validation set;

[0009] Step 3: Build a CNN-AM-BiLSTM neural network model, including a convolutional module, an attention mechanism module, a sequence extraction module, and an output module combination;

[0010] The feature extraction module adopted is stacked by three convolutional modules. Each convolutional module consists of a one-dimensional CNN layer, a max-pooling layer, and a batch normalization BN layer, and the rectified linear unit ReLU is used as the activation function; a dropout layer is added to the last convolutional module to reduce overfitting;

[0011] The attention mechanism module is used to calculate attention to obtain the final attention value;

[0012] The sequence extraction module consists of two BiLSTM layers and a dropout layer; BiLSTM contains two LSTM layers, one in the front, i.e., the forward LSTM, and the other in the back, i.e., the backward LSTM; the forward LSTM can obtain the past information of the input data, and the backward LSTM can obtain the future data information of the input data, and then the outputs of the two hidden layers are combined;

[0013] The output module contains a Flatten layer and two fully connected layers; the Flatten layer transforms the shape of the data, converting the multi-dimensional input data into a one-dimensional vector for subsequent processing by the fully connected layer; the fully connected layer performs a non-linear transformation on the eigenvalue output by the BiLSTM layer, and finally generates a prediction result;

[0014] Step 4: Use the data set in Step 2 as the input of the CNN-AM-BiLSTM neural network model obtained in Step 3, start training the CNN-AM-BiLSTM neural network model, continuously adjust the hyperparameters during the training process, and obtain the optimal hyperparameter group according to the change of the loss function;

[0015] Step 5: Input the air target track data detected by the radar device into the CNN-AM-BiLSTM neural network model, and output the future track data of the air target after model verification.

[0016] Compared with the prior art, the significant advantages of the present invention are:

[0017] (1) The present invention uses regular expressions to process track data, greatly reducing the data processing speed and improving the data utilization rate.

[0018] (2) The present invention uses a stack of three convolutional neural networks, which increases the ability to extract spatial features of data compared with traditional single-layer convolutional neural networks. Moreover, after each convolutional layer, the batch normalization (BN) algorithm is adopted as an effective regularization strategy, which can reduce the shift of internal covariates, improve the training performance of the network, and enhance the generalization ability of the network. A dropout layer is added in the last convolutional module to reduce the overfitting problem of subsequent modules.

[0019] (3) The present invention uses an attention mechanism to strengthen the attention to the spatial features of data and highlight the factors that have a greater impact on the spatial features, providing a better weight reference for the subsequent sequence learning module to learn the temporal correlation of data.

[0020] (4) The present invention uses a stack of two bidirectional long short-term memory networks, which strengthens the learning of temporal patterns and the ability to capture the long-term temporal dependence of time series data compared with traditional single-layer bidirectional long short-term memory networks.

[0021] (5) The present invention uses a stack of three neural network models, which can combine the advantages of the three models, increase the attention and extraction ability of spatial features of data, better learn the time dependence relationship, improve the accuracy of trajectory prediction compared with traditional methods, and enhance the generalization ability of the model. Description of the Drawings

[0022] Figure 1 is a flowchart of a trajectory prediction method based on a CNN-AM-BiLSTM neural network model.

[0023] Figure 2 is a flowchart of data preprocessing in the present invention.

[0024] Figure 3 is a parameter diagram of the CNN-AM-BiLSTM neural network model in the present invention.

[0025] Figure 4 is a schematic diagram of the structure of the CNN-AM-BiLSTM neural network model in the present invention.

[0026] Figure 5 is a flowchart of training the CNN-AM-BiLSTM neural network in the present invention.

[0027] Figure 6 is a flowchart of trajectory prediction based on the CNN-AM-BiLSTM neural network in the present invention. Detailed Embodiments

[0028] The specific embodiments of the present invention will be described below in conjunction with the accompanying drawings, so that those skilled in the art can better understand the present invention. It should be particularly noted that in the following description, when the detailed description of known functions and designs may obscure the main content of the present invention, these descriptions will be omitted here.

[0029] Reference Figures 1 to 6 , a trajectory prediction method based on a CNN-AM-BiLSTM neural network model will be elaborated on for the specific implementation scheme of the present invention, including the following steps:

[0030] Step 1, collect the trajectory data of air targets through radar equipment to construct a dataset. Usually, a radar system can detect multiple data such as the speed, azimuth angle, elevation angle, and slant range of air targets. Therefore, three data items, namely the azimuth angle, elevation angle, and slant range of the target, are selected to construct an air target trajectory dataset;

[0031] Step 2, use the min-max normalization algorithm to scale the trajectory data to the range of [0, 1]. The formula of the min-max normalization algorithm is as follows:

[0032]

[0033] where X new is the normalized data, X is the original data, X min is the minimum value in the dataset, and X max is the maximum value in the dataset. After normalizing the data, use the data partitioning algorithm to divide the trajectory dataset into three parts, namely the training set accounting for 75%, the test set accounting for 15%, and the validation set accounting for 10%. These operations facilitate subsequent training of the neural network.

[0034] Step 3, build a CNN-AM-BiLSTM neural network model mainly composed of a convolutional module, an attention mechanism module, a sequence extraction module, and an output module. The structure and parameters of the trajectory prediction neural network model are as Figure 3 and 4 shown, specifically including the following steps:

[0035] (1) The feature extraction module adopted is stacked by three convolutional modules. Each convolutional module consists of a one-dimensional CNN layer, a max pooling layer, and a batch normalization (BN) layer, and the rectified linear unit (ReLU) is used as the activation function. Its formula is as follows. In the formula, a i represents the i-th input data, and max represents the maximum value function.

[0036] a i = max(0, a i )

[0037] Higher-level features can be better extracted from the input by stacking three convolutional modules, which helps to represent the input data more clearly. The ReLU function is used as the activation function to avoid the problem of gradient vanishing or explosion, and at the same time improve the model convergence speed. At the end of each convolutional module, the batch normalization algorithm is used as an effective regularization strategy. The calculation method of the BN layer is as follows:

[0038]

[0039] where N defines the size of the mini-batch sample data volume, x i and y i are the input and output of the i-th observation in the mini-batch sample. μ represents the mean of the mini-batch sample, γ is the standard deviation of the mini-batch sample, and σ 2 represents the variance of the mini-batch sample. ε is a constant close to zero to ensure stability, γ is the scaling parameter, β is the bias parameter, represents the number of features. In addition to having a regularization effect, it can also reduce the internal covariate shift, improve the training performance of the network, and enhance the generalization ability of the network. A dropout layer is added in the last convolutional module to reduce the overfitting problem of the subsequent modules. In the first convolutional module, the shape of the convolutional kernel is 16*3*1, and the shape of the max-pooling layer is 16 / 2 / 1. In the second convolutional module, the shape of the convolutional kernel is 32*3*1, and the shape of the max-pooling layer is 32 / 2 / 1. In the third convolutional module, the shape of the convolutional kernel is 64*1*1, and the shape of the max-pooling layer is 64 / 2 / 1.

[0040] (2) After the convolutional module is the attention mechanism (AM) module. Placing the attention mechanism module after the convolutional module strengthens the attention to the spatial features of the data and highlights the factors that have a greater impact on the spatial features. Calculating attention involves three stages. In the first stage, the similarity or correlation between the query and each Key is calculated as follows.

[0041] s t =tanh(W h h t +b h )

[0042] where s t is the attention score of the t-th input vector, W h , b h are the weight and bias of the AM respectively, h t represents the t-th input vector, and t represents the serial number of the input vector. In the second stage, the scores obtained in the first stage are normalized, and the softmax function is used to convert the attention scores into the values given in the formula:

[0043]

[0044] where z r represents the r-th input vector, and z t represents the t-th component in the input vector. exp(z r ) represents the returned a t is the new attention score after normalizing the attention score of the t-th input vector. In the third stage, a weighted sum is performed on a t according to the weight coefficient to obtain the final attention value s, as shown in the formula:

[0045]

[0046] (3) After the attention mechanism module is the sequence extraction module, which consists of two layers of BiLSTM and a dropout layer. The number of BiLSTM neurons in the first layer is 64, and the number of BiLSTM neurons in the second layer is 32. BiLSTM has the ability to capture past and future data in the data sequence, which is very important for track prediction because it utilizes the sequential information in both directions. BiLSTM adds the ability to extract past information on the basis of LSTM. It contains two layers of LSTM, one in the front, namely the forward LSTM, and the other in the back, namely the backward LSTM. The forward LSTM can obtain the past information of the input data, and the backward LSTM can obtain the future data information of the input data, and then combine the outputs of the two hidden layers. LSTM has three gate structures: (1) The forget gate f, which determines how much information from the previous cell state needs to be forgotten and how much needs to be retained in the current memory cell state. (2) The input gate i, which determines how much of the current network input information needs to be saved in the current memory cell state. (3) The output gate o, which controls how much information from the current memory cell state needs to be output to the external state.

[0047] The first step of the LSTM operation is used to determine which information to discard or retain from the cell state. This decision is selectively forgotten by the forget gate through the sigmoid function. It receives h t-1 and x t , and for each number in the cell state C t-1 , the output value ranges from 0 to 1. The closer to 0, the more information should be forgotten, and the closer to 1, the more information should be retained. Among them, f t represents the output of the forget gate, σ represents the sigmoid function, x r represents the r-th input vector of the sigmoid function, h t-1 is the output value of the previous moment, x t is the current input value, and C t-1Represents the input state unit at the previous moment, W f and b f are the weights and biases of the forget gate, W f [h t-1 ,x t is the abbreviation of W f h t-1 +W f x t The same applies to the following use of [].

[0048]

[0049] f t =σ(W f [h t-1 ,x t +b f )

[0050] The second step is to generate new information that needs to be updated. This step is divided into two sub - stages. The first sub - stage is that the input gate determines which values need to be updated through the sigmoid function. The second sub - stage is that tanh is used to generate new candidate values - which can be used to add to the cell state. In the next step, the above two sub - stages need to be combined to achieve the update of the cell state. Finally, the result of the sigmoid output is multiplied by the input information processed by tanh to determine which are the key information to be retained, where i t is the input of the input gate, represents the current input cell state, i.e., the intermediate variable, W i and b i are the weights and biases of the input gate, W c and b c are the weights and biases of the current state.

[0051] i t =σ(W i [h t-1 ,x t +b i )

[0052]

[0053] Now update the old cell state C t-1 to C t , by multiplying the old state by f t to forget the information that was previously decided to be forgotten, and then adding This is the new candidate value, and then scale the proportion of the updated value determined for each state. Adding the results obtained from the first step and the second step is the process of discarding the unnecessary information and adding new information, and the information transmitted to the next state can be obtained, where Ct is the cell state at the current moment, and ⊙ represents element-wise multiplication.

[0054]

[0055] The last step is the decision-making process of the model output. In this process, the hidden state of the next time step will be determined, which contains the previous input information. This process first obtains the output through the sigmoid function, and then uses the tanh function to scale the new cell state of the previous time step to ensure that the value is between -1 and 1. After that, the scaled new cell state is multiplied pairwise with the output obtained from the sigmoid function to determine the information carried by the hidden state. Finally, the new information will be used as the current output and further passed to the next time step, where O t represents the output of the output gate, and h t represents the hidden state at the current moment, and W o and b o are the weights and biases of the output gate.

[0056] O t = σ(W o [h t-1 , x t + b o )

[0057] h t = o t tanh(C t )

[0058] (4) After passing through the sequence extraction module, the data arrives at the output module, which consists of a Flatten layer and two fully connected layers. The Flatten layer transforms the shape of the data, converting the multi-dimensional input data into a one-dimensional vector for subsequent processing by the fully connected layers. The fully connected layers perform a series of non-linear transformations on the eigenvalue output by the BiLSTM layer and finally generate the prediction result. After passing through the prediction module, the azimuth angle, elevation angle, and slant range of the aerial target are output.

[0059] Step 4. Adjusting the hyperparameters of the CNN-AM-BiLSTM neural network model during the training process is crucial for the trained model. Finding suitable hyperparameters mainly involves adjusting and optimizing most hyperparameters through a large number of experiments. For example, several optimization algorithms such as Stochastic Gradient Descent (SGD), RMSprop algorithm, and Adam were compared. According to the comparison results, Adam can improve the accuracy of the built model and was selected as the optimization algorithm. Mean Squared Error (MSE) was used as the loss function, which can be backpropagated to update the weights and biases. The initial learning rate of the model was adjusted to 0.002 and gradually linearly decreased to 0.0005 by the last training. This helps to maintain a relatively stable pace during the learning process. The set number of training was 120, the batch size was 32, and each experiment was repeated more than ten times to limit the influence of random factors on the network performance.

[0060] The mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), accuracy (ACC), and R 2 were used as five evaluation criteria to evaluate the performance of the built model and make a fair comparison with other models. The equations of MAE, RMSE, MAPE, ACC, and R 2 are shown as follows:

[0061]

[0062] In the above formulas, n samples is the number of predicted data, the predicted value is the actual value is y i , and the average value of the actual values is

[0063] The proposed method was compared with KF, CNN, LSTM, BiLSTM, and CNN-LSTM to verify the effectiveness, superiority, and generalization ability of the method. Table 1 shows the results of all the proposed methods in predicting the azimuth angle, elevation angle, and slant range. It can be seen from the results that the method has the best performance in predicting the three coordinates, and the MAE, RMSE, and MAPE are all the lowest, and the R 2 score is the highest. This indicates that the method is correct in predicting the air target track and has a competitive advantage compared with other methods. Among them, the Kalman filter model has the worst performance in track prediction, with the highest MAE, RMSE, and MAPE indicators, and the lowest R 2 score. This is mainly because the Kalman filter needs to change the parameters of the state transition matrix according to the general shape of the track, and this method has poor generalization ability and cannot be applied to complex and diverse target tracks.

[0064] Table 1 Comparison of error indicators of six model evaluation methods

[0065]

[0066] The prediction accuracies of the six models in azimuth, elevation angle, and slant range are shown in Table 2. The prediction accuracy of the CNN-AM-BiLSTM model is the highest among the six models. Among them, the prediction accuracy of the KF model in azimuth reaches 66.49%, in elevation angle reaches 68.43%, and in slant range reaches 58.63%. The prediction accuracy of the CNN model in azimuth reaches 80.73%, in elevation angle reaches 89.09%, and in slant range reaches 86.97%. The prediction accuracy of the LSTM model in azimuth reaches 72.39%, in elevation angle reaches 83.2%, and in slant range reaches 89.26%. The prediction accuracy of the BiLSTM model in azimuth reaches 78.77%, in elevation angle reaches 88.22%, and in slant range reaches 87.36%. The prediction accuracy of the CNN-LSTM model in azimuth reaches 91.32%, in elevation angle reaches 94.07%, and in slant range reaches 97.78%. The prediction accuracy of the CNN-LSTM model is generally higher than that of single models. After adding the attention mechanism to the CNN-LSTM model, the prediction accuracy is further improved. The prediction accuracy of the CNN-AM-BiLSTM model in azimuth reaches 98.63%, in elevation angle reaches 99.51%, and in slant range reaches 99.91%. Compared with traditional models, the CNN-AM-BiLSTM model effectively utilizes a large amount of historical data and can also make relatively accurate predictions when facing complex and diverse flight tracks. Compared with single deep learning models such as CNN, LSTM, and BiLSTM, the CNN-AM-BiLSTM model combines the advantages of single models, can extract spatial features of flight track data, increases the learning weight of important spatial features, and can also learn the time characteristics of flight track data.

[0067] Table 2 Comparison of Prediction Accuracies of Six Models

[0068]

[0069]

[0070] Step 5: Input the air target track data detected by the radar device into the CNN-AM-BiLSTM neural network model. After model verification, output the future track data of the air target. Input the azimuth angle, elevation angle, and slant range of the air target detected by the radar into the CNN-AM-BiLSTM neural network model. After verification, the CNN-AM-BiLSTM neural network model will output the predicted azimuth angle, elevation angle, and slant range of the future air target, ultimately realizing the prediction of the future track of the air target.

[0071] The established neural network model is mainly used to predict the tracks of various air targets, thereby providing effective reference data for the fire control computer and accelerating the calculation of the firing lead by the fire control computer. Moreover, the neural network track prediction model based on CNN-AM-BiLSTM refers to a large amount of historical data and can predict the tracks of different air targets, having strong versatility. The prediction accuracy has also been improved to a certain extent compared with the traditional air defense weapon track prediction algorithm.

Claims

1. A track prediction method based on a CNN-AM-BiLSTM neural network model, characterized in that Including: Step 1: Collect the track data of air targets through a radar device and construct a dataset. Step 2: Use the min-max normalization algorithm to scale the track data to the range of [0, 1], and divide it into a training set, a test set, and a validation set. Step 3: Build a CNN-AM-BiLSTM neural network model, including a convolutional module, an attention mechanism module, a sequence extraction module, and an output module combination. The feature extraction module adopted is stacked by three convolutional modules. Each convolutional module consists of a one-dimensional CNN layer, a max-pooling layer, and a batch normalization BN layer, and the rectified linear unit ReLU is used as the activation function; a dropout layer is added to the last convolutional module to reduce overfitting. The attention mechanism module is used to calculate attention and obtain the final attention value. The sequence extraction module consists of two layers of BiLSTM and a dropout layer; BiLSTM contains two layers of LSTM, one in the front, i.e., the forward LSTM, and the other in the back, i.e., the backward LSTM; the forward LSTM can obtain the past information of the input data, and the backward LSTM can obtain the future data information of the input data, and then combine the outputs of the two hidden layers. The output module contains a Flatten layer and two fully connected layers; the Flatten layer transforms the shape of the data, converting the multi-dimensional input data into a one-dimensional vector for subsequent processing by the fully connected layer. The fully connected layer performs a non-linear transformation on the eigenvalue output by the BiLSTM layer and finally generates a prediction result. Step 4: Use the dataset in Step 2 as the input of the CNN-AM-BiLSTM neural network model obtained in Step 3, start training the CNN-AM-BiLSTM neural network model, continuously adjust the hyperparameters during the training process, and obtain the optimal hyperparameter group according to the change of the loss function. Step 5: Input the track data of the air target detected by the radar device into the CNN-AM-BiLSTM neural network model, and output the future track data of the air target after model verification.

2. The track prediction method based on the CNN-AM-BiLSTM neural network model according to claim 1, wherein The batch normalization algorithm that makes up the batch normalization BN layer is: where N is the size of the batch sample data, x i and y i are the input and output of the i-th observation in the batch sample, μ represents the mean of the batch sample, γ is the standard deviation of the batch sample, σ 2 represents the variance of the batch sample, ε is a constant, γ is a scaling parameter, β is a bias parameter, represents the number of features.

3. The track prediction method based on the CNN-AM-BiLSTM neural network model according to claim 1, characterized in that In the first convolutional module, the shape of the convolutional kernel is 16*3*1, and the shape of the max-pooling layer is 16 / 2 / 1; in the second convolutional module, the shape of the convolutional kernel is 32*3*1, and the shape of the max-pooling layer is 32 / 2 / 1; in the third convolutional module, the shape of the convolutional kernel is 64*1*1, and the shape of the max-pooling layer is 64 / 2 / 1.

4. The track prediction method based on the CNN-AM-BiLSTM neural network model according to claim 1, wherein Attention calculation includes three stages: The first stage: The similarity or correlation between the query and each Key. s t = tanh(W h h t + b h ) where s t is the attention score of the t-th input vector, and W h , b h are the weight and bias of AM respectively, h t represents the t-th input vector, and t represents the serial number of the input vector; In the second stage, normalize the scores obtained in the first stage, and use the softmax function to convert the attention scores into the values given in the formula. where z r represents the r-th input vector, and z t represents the t-th component in the input vector. exp(z r ) represents returning e zr , and a t is the new attention score after normalizing the attention score of the t-th input vector; In the third stage, a weighted sum is performed on a according to the weight coefficient t : Thus, the final attention value s is obtained.

5. The track prediction method based on the CNN-AM-BiLSTM neural network model according to claim 4, characterized in that The first step of the LSTM operation is to determine what information to discard or retain from the cell state. This decision is made by the forget gate through sigmoid for selective forgetting, receiving h t-1 and x t , and for each number in the cell state C t-1 , the output value ranges between 0 and 1. The closer to 0, the more it should be forgotten, and the closer to 1, the more it should be retained; The second step is to generate new information that needs to be updated. The first stage is that the input gate determines which values need to be updated through the sigmoid function; The second stage is that tanh is used to generate new candidate values The above two stages are combined to update the unit state; finally, the result output by sigmoid is multiplied by the input information processed by tanh to determine which key information needs to be retained; The last step is the decision-making process of the model output. In this process, the hidden state of the next time step will be determined, which contains the previous input information. First, the output is obtained through the sigmoid function, and then the new cell state of the previous time step is scaled using the tanh function to ensure that the values are between -1 and 1. After that, the scaled new cell state and the output obtained from the sigmoid function are multiplied pairwise to determine the information carried by the hidden state. Finally, the new information will be used as the current output and further passed to the next time step.

6. The method for predicting a flight path based on the CNN-AM-BiLSTM neural network model according to claim 1, wherein The number of BiLSTM neurons in the first layer is 64, and the number of BiLSTM neurons in the second layer is 32.

7. The track prediction method based on the CNN-AM-BiLSTM neural network model according to claim 1, wherein The mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), accuracy (ACC), and R 2 are used as five evaluation criteria to evaluate the performance of the established model.

Images