CN117540626A - A situation prediction method for fixed-wing UAVs based on Bayesian neural network - Google Patents

Abstract

The invention provides a fixed-wing UAV situation prediction method based on Bayesian neural network, which belongs to the technical field of UAV situation prediction; it solves the problem that our UAV cannot attack the enemy UAV in an uncertain environment. The technical problem of uncertainty prediction of future situation. The technical solution is: establish a Bayesian network suitable for time series prediction and collect limited situation information of enemy UAVs; use the situation information of enemy UAVs as input and use the established Bayesian neural network to attack the enemy. Predict the situation of the enemy UAV at the next moment; use the prediction value at a single time as input to predict again to form the situation information of the enemy UAV in the future time period. The beneficial effects of the present invention are: it enables our UAV to use limited situation information to predict the situation of the enemy UAV in the next period of time in the battlefield environment, enables our UAV to seize the initiative on the battlefield, and is conducive to improving the situation of the UAV. Man-machine combat capabilities, thereby reducing the combat loss ratio of our drones.

Description

A situation prediction method for fixed-wing UAVs based on Bayesian neural network

Technical field

The invention relates to the technical field of UAV situation prediction, and in particular to a fixed-wing UAV situation prediction method based on Bayesian neural network.

Background technique

In recent years, fixed-wing UAVs have played a key role in military operations due to their advantages such as long endurance, high-speed flight, and strong payload capacity, and can perform multiple tasks such as reconnaissance, strikes, and aerial confrontation. In this combat environment, predicting and obtaining the situational information of the opponent's UAV is crucial for specifying tactical decisions and implementing strike operations. Predicting the situational information of the enemy's UAV requires not only relying on high-precision sensors such as radar , infrared sensors and cameras, etc., mature and reliable prediction algorithms are also needed to apply data obtained from sensors to complete predictions.

In an air combat environment, in order to reduce the combat damage ratio of our drones, we often need to rely on very reliable algorithms to help the drones predict the future situation information of the enemy drones. However, some conventional algorithms are overconfident in their predictions and are unable to evaluate the uncertainty of the predicted values. These flaws will cause irreparable consequences in military operations. For example, in the article "Task offloading scheme combining deep reinforcement learning and convolutional neural networks for vehicle trajectory prediction in smartcities", the convolutional neural network, as a neural network dedicated to time series, often requires a large amount of data information and is used as a neural network to predict vehicle trajectories. The convolutional neural network of the black box model lacks interpretability and lacks reliability in military operations. The purpose of this invention is to solve the problem that limited data cannot be used to provide reliable predictions of enemy drones in the uncertain environment of the battlefield.

Contents of the invention

The purpose of the present invention is to provide a fixed-wing UAV situation prediction method based on Bayesian neural network. This prediction method can quickly predict the situation information of enemy UAVs and give the uncertainty of the prediction. , helps reduce overfitting of neural networks.

The inventive idea of the present invention is: first, establish a Bayesian network suitable for time series prediction, and then obtain the latest limited situation information of the enemy UAV through our UAV sensor system, and then use the established Bayesian network The Sri Lankan neural network predicts the situation of the enemy UAV at the next moment, and uses the data obtained from the network prediction as input to the network again. After multiple cycles, the situation information of the enemy UAV in the future time period is obtained.

The invention gives the UAV the ability to predict the future situation information of the enemy UAV and gives the uncertainty of the prediction result.

The present invention is achieved through the following measures: a fixed-wing UAV situation prediction method based on Bayesian neural network, including the following steps:

S1. Establish a Bayesian network suitable for time series prediction, and save the trained network parameters and structure to facilitate real-time prediction after later collecting real-time situation information of enemy drones;

S2. Obtain the latest limited situation information of the enemy's UAV through our UAV sensor system, organize the enemy's situation information and pass it to the Bayesian neural network to facilitate rapid prediction by the neural network;

S3. Taking the enemy UAV situation information as input, use the established Bayesian neural network to predict the enemy UAV situation at the next moment;

S4. Use the single-moment prediction value obtained by the network prediction as input to predict again, and splice the prediction results with the original data segments to form the situation data of the enemy UAV in the future time period. Finally, the predicted enemy UAV The situation information of the aircraft is transmitted back to our drone, so that our drone can occupy a favorable position in the future.

Further, the step one includes the following steps:

1-1) Collect sufficient enemy UAV situation information data, and organize the collected data information into a three-dimensional array to form a database. The first dimension is the number of enemy UAV situation information collected, and the second dimension is the time step of each piece of situation information data collected, and the third dimension is the number of situation information features input to the neural network;

1-2), use the double loop function and slicing operation to randomly select the training data and test data for training the Bayesian neural network from the data set collected in step 1-1);

1-2-1), use a double loop to select indexes for the first and second dimensions of the collected data;

1-2-2), use the selected index as the starting end of the slice, slice backward, and select a sufficient number of data segments with the same length of each data segment through a loop operation; use the selected index as the starting end of the slice, and slice toward Then perform slicing, and select a sufficient number of data segments with the same data length for each segment through loop operation;

1-2-3). Organize the data segments collected through slicing to form a three-dimensional array as a data set. The first dimension represents the total number of collected data segments, and the second dimension represents the time step of each data segment. The third dimension represents the number of situational information features of the enemy drone;

1-2-4). After shuffling the order of the collected data sets, select 70%-80% of the first dimension of the data set as the training set, and the remaining part as the test set;

1-2-5). For each data segment of the training set and test set in the data set, use the slicing operation to cut out the first half of the second dimension as the data value, and the remaining part as the label. At this time, a total of four sub-data sets are obtained. , respectively, are the input data set used for training, the labeled data set corresponding to the training set, the input data set used for testing, and the labeled data set corresponding to the test set. The third dimension of these four sub-datasets all contains 6 elements, which is the situation characteristic information of the UAV every second, recorded as Among them: 1-3), the number of features_num in the third dimension of the data set obtained in step 1-2-3) is 6, recorded as /> in:

(1)

Among them, n _x is the tangential overload of the UAV, n _z is the normal overload of the UAV, φ is the roll angle of the UAV around the velocity vector, and g is the gravity acceleration. The output feature output_size is 6×1, which means that the predicted data is the above 6 kinematic features of the enemy drone at the 6th second;

1-4), establish a neural network layer with Bayesian characteristics. This neural network layer defines that the weight and bias of each node are random variables sampled from a normal distribution. The weight of this neural network layer is The mean and variance are both two-dimensional matrices, and the size of the first dimension is the number of input features, the size of the second dimension is the number of output features, and the weight and variance of the bias are one-dimensional matrices whose shape is the size of the output features;

1-5). Construct a Bayesian neural network based on the above-defined neural network layer with Bayesian properties. The Bayesian neural network uses the evidence lower bound in variational inference as the loss function and specifies the forward propagation method. Calculate the logarithmic prior distribution, logarithmic posterior distribution and logarithmic likelihood in the neural network layer with Bayesian characteristics, and define the calculation method of the network loss function as the logarithmic posterior distribution minus the logarithmic prior distribution minus log-likelihood;

1-6) Use the above network to train on the training set, and save the trained network model structure and parameters.

Further, in the second step, the situation information data of the enemy UAV is used as input, and the situation information of the enemy UAV at the next moment is predicted through the network.

Further, the step two includes the following steps:,

2-1) Organize the enemy drone situation information collected by our drone in real time into an array that conforms to the input network, recorded as S _b ;

2-2) Input S _b into the Bayesian neural network that has completed training, and the network output will obtain the situation information of the enemy UAV at the next moment.

Further, in the third step, the single-time prediction value is used as input to predict again, constituting the situation information of the enemy UAV in the future time period. The single-time prediction value is used as input to predict again, constituting the situation information of the enemy UAV in the future time period. information.

Further, the step three includes the following steps:

3-1) Organize the enemy UAV situation information at a single moment into a one-dimensional array Sh _h that is the same as the second-dimensional feature element of the two-dimensional array S _b ;

3-2) Splice _the one-dimensional array S _T0 into _the two _- dimensional array S , forming a new two-dimensional array S _T1 ;

3-3) Input the obtained two-dimensional array S _T1 into the Bayesian neural network through a loop to obtain the situation information of the enemy UAV at the next moment, and loop the operation in 2) above to obtain the two-dimensional array S _T2 ;

3-4) Loop through the above 3-1)-3-3) to obtain the situation information data S _T of the enemy UAV in the future period. At a single moment when all the situation data is predicted by the Bayesian neural network, the enemy has no Composition of human-machine situation data;

3-5) Transmit the enemy's situation information data in the future period back to our drone, and our drone will use this situation information to seize an advantageous situation in the air battle.

Compared with the prior art, the beneficial effects of the present invention are:

(1) This invention introduces uncertainty modeling. Bayesian neural network can add uncertainty to the prediction results, which is not available in traditional neural networks. In UAV situation prediction, environmental changes, Factors such as sensor noise can lead to uncertainty, while Bayesian neural networks can provide more accurate prediction results and better quantify the credibility of predictions.

(2) The present invention performs well under limited data conditions, but in UAV situation prediction, insufficient data is a common challenge. Traditional neural networks are prone to overfitting in this case, while Bayesian neural networks reduce the risk of overfitting by introducing prior probabilities, thereby performing more robustly on small data sets.

(3) The Bayesian neural network in the present invention provides a high degree of interpretability and is particularly suitable for explaining enemy UAV situation prediction. Unlike traditional neural networks, Bayesian neural networks are easier to explain their prediction logic and help decision makers understand the principles behind the model's prediction results.

(4) The Bayesian neural network in the present invention can show stronger generalization ability, can handle uncertainty more effectively, reduce the risk of over-fitting, can better adapt to unseen data, and can Handle environmental and data instability and excel in complex tasks.

(5) The Bayesian Bayesian neural network in the present invention can still provide effective predictions when only limited data of enemy drones are collected, and can provide relatively reliable information on the enemy drones in a certain period of time in the future. machine situation information.

Description of drawings

The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

Figure 1 is an overall flow chart of the fixed-wing UAV situation method in an uncertain environment based on Bayesian neural network provided by the present invention.

Figure 2 is a flow chart of post-processing after collecting enemy fixed-wing UAV situation data provided by the present invention.

Figure 3 is a network training flow chart of the method for predicting the situation of a fixed-wing UAV in an uncertain environment based on the Bayesian neural network provided by the present invention.

Figure 4 is a histogram of the MSE comparison between the predicted trajectory and the actual trajectory in each step of the five-step prediction method of the fixed-wing UAV situation prediction in an uncertain environment based on Bayesian neural network provided by the present invention; wherein, Figure 4 ( a) is the first step, Figure 4(b) is the second step, Figure 4(c) is the third step, Figure 4(d) is the fourth step, and Figure 4(e) is the fifth step.

FIG. 5 is a flow chart of the fixed-wing UAV situation prediction method in an uncertain environment based on Bayesian neural network provided by the present invention when using the data obtained from network prediction to predict again.

Figure 6 is a comparison chart between the predicted trajectory and the actual trajectory obtained by the fixed-wing UAV situation prediction method in an uncertain environment based on Bayesian neural network provided by the present invention.

Figure 7 is a comparison chart between the predicted trajectory and the actual trajectory obtained by replacing the Bayesian neural network with a long short-term memory neural network using the fixed-wing UAV situation prediction method in an uncertain environment based on the Bayesian neural network provided by the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described here are only used to explain the present invention and are not intended to limit the present invention.

Example 1

This embodiment provides a fixed-wing UAV situation prediction method based on Bayesian neural network, which includes the following steps:

Step 1) Establish a Bayesian network suitable for time series prediction, and save the trained network parameters and structure;

Step 2) Obtain the latest situation information of the enemy's UAV through our UAV sensor system, organize the enemy's situation information and pass it to the Bayesian neural network;

Step 3) Use the established Bayesian neural network to predict the situation of the enemy UAV at the next moment, and finally transmit the predicted situation information of the enemy UAV back to our UAV.

Step 1), establish a Bayesian network suitable for time series prediction, and save the trained network parameters and structure. The specific steps are as follows:

1-1), as shown in Figure 2, the processing flow after collecting enemy situational data is to collect sufficient, mutually independent and non-repeating situational information data of enemy drones. The data shape is (tracks_num, time_long, features_num), where tracks_num represents the number of enemy UAV situation information collected, time_long represents the time step of each piece of enemy UAV situation information data collected, and features_num represents the situation of the enemy UAV entered into the network. Number of information features.

1-2), use the double loop function and slicing operation to randomly select the training data and test data for training the Bayesian neural network from the data set collected in step 1-1);

1-2-1), use a double loop function to select the index idex for the first dimension tracks_num and the second dimension time_long of the data collected in step 1-1), and the total amount selected is batch_size∈(1000, tracks_num×time_long);

1-2-2), take the index idex selected in step 1-2-1) as the starting end of the slice, select 6 consecutive time steps backward as a piece of data, and select a total of batch_size data segments, each segment is long is 6, and each data segment is marked as α;

1-2-3), the data selected in steps 1-2-1) and 1-2-2) is a three-dimensional array of shape (batch_size, time_step, features_num), where batch_size=3000 is for training and testing. For the total batch, time_step=6 is the time step of a single segment of UAV situation information, and features_num represents the number of situation information features of the enemy UAV;

1-2-4). After shuffling the order of the collected data sets, select 80% of the batch_size as the training set trian_data, with a shape of (2400, 6, features_num), and the remaining 20% as the test set test_data, with a shape of (800, 6,features_num);

1-2-5). For each data segment α in the training set train_data and test set test_data, the first 5 seconds are segmented as the data value, and the 6th second is used as the label. At this time, a total of 4 data sets are obtained: :

The input data used for training input_train has a shape of (2400,5,features_num);

The label data set output_train corresponding to the training data set has a shape of (2400,1,features_num);

The input data used for testing input_test has a shape of (800,5,features_num);

The label data set output_test corresponding to the test data set has a shape of (800,1,features_num). The collection and division of the data set are shown in Figure 2.

1-3). The number of features_num in the third dimension of the data set obtained in step 1-2-3) is 6, which is recorded as in:

(1)

n _x is the tangential overload of the UAV, n _z is the normal overload of the UAV, φ is the roll angle of the UAV around the velocity vector, and g is the gravity acceleration. The output feature output_size is 6×1, which means that the predicted data is the above 6 kinematic features of the enemy drone at the 6th second;

1-4), establish a neural network layer Linear_BBB with Bayesian characteristics. This neural network layer defines that the weight w and bias b of each node are sampled from a normal distribution with a mean of 0 and a variance of 1. Random variables, the mean w_mu of the weight and the variance w_rho of the weight in the neural network layer are matrices with a shape of (input_size, putput_size), and the weight b_mu of the bias and the variance of the bias b_rho are matrices with a shape of (output_size);

1-5) Construct the network MLP_BBB and create the Linear_BBB instance defined above in the network. This Bayesian neural network uses the evidence lower bound ELBO (Evidence Lower Bound) in variational inference as the loss function and specifies the forward propagation method. Calculate the log prior distribution log_prior, log posterior distribution log_post and log likelihood log_like in the Linear_BBB layer, and define the calculation formula of the network loss function sample_elambo as:

loss=log_post-log_prior-log_like (2)

In the loss function of the network, log_post-log_prior is the complexity cost, and log_like is the error cost;

1-6) Use the collected enemy situation information data to train the built Bayesian neural network. Set the training batch epoch_num before training. Follow the process in Figure 2 to complete the network training and save the trained network. Parameters and structure. When training the network and testing the network, the change curve of the loss function is shown in Figure 4. The change curve of the complexity cost and error cost that constitute the loss function is shown in Figure 5.

Step 2): Perceive the situation information data S _b of the enemy UAV in real time through the sensor system of our UAV. This data should be the latest 5 seconds of situation information of the enemy UAV, that is, the shape of S _b is ( 5,6), where the first dimension of 5 represents the situation information of the enemy UAV in the last five seconds, and the second dimension of 6 represents the 6 kinematic characteristics of the enemy UAV per second. The perceived The data is passed to the Bayesian neural network that has been trained in steps 1-6).

Step 3),

3-1) Organize the enemy drone situation information at a single moment into a two-dimensional array Sh _h that has the same shape as the last two dimensions of the three-dimensional array S _b ;

3-2) Splice the two-dimensional array S _T0 according to the first dimension to _the three _- dimensional array _S Three-dimensional array S _T1 ;

3-3) Input the obtained three-dimensional array S _T1 into the Bayesian neural network through a loop to obtain the situation information of the enemy UAV at the next moment, and loop the operation in 2) above to obtain the three-dimensional array S _T2 ;

3-4) Loop through the above 3-1)-3-3) to obtain the situation information data S _T of the enemy UAV in the future period. At a single moment when all the situation data is predicted by the Bayesian neural network, the enemy has no The human-machine situation data composition and program cycle flow chart are shown in Figure 5;

3-5) Transmit the enemy's situation information data in the future period back to our drone, and our drone will use this situation information to seize an advantageous situation in the air battle.

Example 2

Referring to Figure 7, the technical solution provided by this embodiment is a fixed-wing situation prediction method based on long short-term memory neural network. The specific steps are as follows:

Step 1) Establish a long-short-term memory neural network suitable for time series prediction, and save the network parameters and structure.

Step 2) Obtain the latest situation information of the enemy's UAV through our UAV sensor system, organize the enemy's situation information and pass it to the Bayesian neural network;

Step 3) Use the established Bayesian neural network to predict the situation of the enemy UAV at the next moment, and finally transmit the predicted situation information of the enemy UAV back to our UAV.

Step 1), establish a long short-term memory neural network suitable for time series prediction, and save the trained network parameters and structure. The specific steps are as follows:

Steps 1-1) to 1-3) are the same as in Embodiment 1 above.

1-4) Establish a 3-layer long short-term memory neural network, use the mean square error MSE as the loss function, and stipulate that in each training, the loss function is calculated, and the parameters and structure of the trained long short-term memory neural network are save.

Step 2) and step 3) are the same as the above-mentioned Embodiment 1. The trajectory map of the UAV predicted using the long short-term memory neural network and the actual trajectory map are compared as shown in Figure 7.

Comparing Figure 6 with Figure 7 , the prediction result obtained by the Bayesian neural network used in this embodiment in Figure 6 is more accurate and closer to the actual trajectory of the enemy drone. It also shows that the Bayesian neural network proposed in this embodiment has stronger generalization ability when there is less training data.

The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.

Claims (6)

1. A situation prediction method of a fixed wing unmanned aerial vehicle based on a Bayesian neural network is characterized by comprising the following steps:

1) Establishing a database according to the situation information of the enemy fixed wing unmanned aerial vehicle, and constructing a Bayesian neural network for time sequence prediction;

2) The situation information of the enemy unmanned aerial vehicle is taken as input, and the situation information of the enemy unmanned aerial vehicle at the next moment is predicted through a network;

3) And predicting again by taking the single-moment predicted value obtained by network output as input to form situation information of future time periods of the enemy unmanned aerial vehicle, and transmitting back to the unmanned aerial vehicle.

2. The method of claim 1, wherein creating a database based on enemy fixed wing unmanned aerial vehicle situation information comprises the steps of:

1) Collecting sufficient situation information data of the enemy unmanned aerial vehicle, arranging the collected data information into a three-dimensional array to form a database, wherein the first dimension is the quantity of the collected situation information data of the enemy unmanned aerial vehicle, the second dimension is the time step of each piece of the collected situation information data, and the third dimension is the characteristic quantity of the situation information of the input neural network;

2) The collected data is randomly selected and partitioned into training and testing sets using a double loop and slice operation.

3. The prediction method according to claim 2, wherein the randomly selecting and dividing the collected data into the training set and the test set using the double loop and slice operation comprises the steps of: 1) Selecting an index for the first dimension and the second dimension of the collected data using a double loop; 2) Taking the selected index as the start end of slicing, slicing backwards, and selecting enough data segments with the same data length by cyclic operation; 3) The data segments collected through slicing are arranged to form a three-dimensional array as a data set, wherein the first dimension represents the total number of the collected data segments, the second dimension represents the time step of each segment of data segment, and the third dimension represents the situation information feature quantity of the enemy unmanned aerial vehicle; 4) After the collected data sets are disordered, 75 to 90 percent of the first dimension of the data sets are selected as training sets, and the rest of the data sets are selected as test sets; 5) For each data segment of the training set and the test set in the data set, slicing the first half of the second dimension into data values and the rest into labels by using slicing operation, and obtaining four sub-data sets, namely an input data set for training, a label data set corresponding to the training set, an input data set for testing and a label data set corresponding to the test set, wherein the third dimension of the four sub-data sets contains 6 elements, is situation characteristic information of the unmanned aerial vehicle per second and is recorded asWherein (1)>The position increment and the +.A unit time of the unmanned plane in the x, y and z directions are respectively as follows>For the speed increment of unmanned plane per unit time, +.>The yaw angle increment and the pitch angle increment of the unmanned aerial vehicle in unit time are respectively adopted.

4. The prediction method according to claim 2, wherein the construction of the bayesian neural network for the time series comprises the steps of:

1) Establishing a neural network layer with Bayesian characteristics, wherein the neural network layer defines that the weight and bias of each node are random variables obtained by sampling in normal distribution, the mean value and variance of the weight in the neural network layer are two-dimensional matrixes, the first dimension is the number of input features, the second dimension is the number of output features, and the weight and variance of the bias are one-dimensional matrixes with the shape of the number of output features;

2) Constructing a Bayesian neural network according to the defined neural network layer with Bayesian characteristics, wherein the Bayesian neural network uses the evidence lower bound in variation inference as a loss function, and provides that the log prior distribution, the log posterior distribution and the log likelihood in the neural network layer with Bayesian characteristics are calculated in a forward propagation method, and the calculation mode of the network loss function is defined as the log posterior distribution minus the log prior distribution minus the log likelihood;

3) Training on the training set by using the network, and storing the trained network model structure and parameters.

5. The prediction method according to claim 1, wherein the situation information of the enemy unmanned aerial vehicle is input, and the situation information of the enemy at the next moment is predicted through a network, and the method comprises the following steps:

1) The enemy unmanned aerial vehicle situation information collected by the unmanned aerial vehicle in real time is arranged into an array conforming to an input network and recorded as S _b ；

2) Will S _b And inputting the information into the trained Bayesian neural network, and outputting the information by the network to obtain situation information of the enemy unmanned aerial vehicle at the next moment.

6. The prediction method according to claim 1, wherein the predicting again using the predicted value of the single moment as an input constitutes situation information of a future period of time of the enemy unmanned aerial vehicle, and comprises the steps of:

1) The situation information of the enemy unmanned aerial vehicle at a single moment is arranged into a two-dimensional array S _b One-dimensional array S with same shape in rear two dimensions _h ；

2) Will be one-dimensional array S _T0 Splicing to a two-dimensional array S in order of a first dimension _b Two-dimensional array S formed by the extreme end _X Two-dimensional array S cut out according to the order of the first dimension _X Form a new two-dimensional array S _T1 ；

3) The two-dimensional array S is obtained by circulation _T1 Inputting a Bayesian neural network to obtain situation information of the enemy unmanned aerial vehicle at the next moment, and circularly performing the operation in the step 2) to obtain a two-dimensional array S _T2 ；

4) Circulating the steps 1) to 3) to obtain situation information data S of future time periods of the enemy unmanned aerial vehicle _T The situation data are all composed of situation data of the enemy unmanned aerial vehicle at a single moment predicted by a Bayesian neural network;

5) And transmitting situation information data of the enemy future period to the unmanned aerial vehicle, and utilizing the situation information, the unmanned aerial vehicle seizes the favorable situation in the air combat.