CN105242536A - Unmanned aerial vehicle driving route waypoint calibration method based on BP nerve network - Google Patents

Abstract

The invention discloses a waypoint calibration method for unmanned aerial vehicles based on BP neural network, comprising the following steps: randomly generating 1000 three-dimensional coordinate points; establishing a matrix P of the input layer of BP neural network; listing the output terminals of BP neural network Matrix T; establish a BP neural network; train the BP neural network; fit the waypoint route of the drone. Compared with the prior art, the present invention increases the number of punctuation points that can be marked by the UAV, the punctuation route of the UAV is directional and can improve the UAV autopilot system, and innovatively applies the BP neural network technology, And on this basis, it can quickly and effectively realize the design of the UAV ground station to the waypoint route, which can be applied to the UAV autopilot; as a novel neural network-based autopilot waypoint route scheme, in the future There are great prospects for development in the research of UAV autopilot.

Description

Waypoint Calibration Method of UAV Driving Route Based on BP Neural Network

technical field

The invention relates to the technical field of unmanned aerial vehicle automatic driving, in particular to a waypoint calibration method of an unmanned aerial vehicle driving route.

Background technique

In recent years, the development of drones has been particularly rapid. The scope of use of drones has expanded from military to civilian and scientific research fields. In the military, UAVs can be used for reconnaissance and surveillance, relay communication, electronic countermeasures, battle result assessment, ground (sea) attacks, early warning, etc.; in civilian use, they can be used for atmospheric research, meteorological observation, and experimental verification of new technologies and equipment Wait.

In mid-April 2010, the U.S. Army officially announced the UAV system development roadmap from 2010 to 2035: the Army’s short-term goal is to realize unmanned helicopters, that is, to equip new carrier-based vertical helicopter take-off and landing tactical UAVs , to fill the current urgent war needs. Research on drone autopilots is also intensifying. The unmanned helicopter autopilot MP2128HELI of the Canadian MicroPilot company, the wcPilot1000 micro unmanned helicopter autopilot of the Swiss WcControl company, and the HcliAP unmanned helicopter autopilot developed by the California Institute of Technology are typical representatives.

At present, the research level of domestic UAV autopilot is still in the imitation stage, and the overall research and development is still in its infancy.

The BP neural network belongs to the forward network, which is the core of the neural network and the essence of the entire neural network system, and it is different from the multi-layer perceptron.

Contents of the invention

In view of the above-mentioned prior art and existing problems, the present invention proposes a waypoint calibration method for unmanned aerial vehicle driving routes based on BP neural network, and innovatively applies BP neural network technology in this field to realize The ground station calibrates the flight route of the UAV at the waypoint on the map.

The present invention proposes a kind of unmanned aerial vehicle driving route waypoint calibration method based on BP neural network, comprising the following steps:

Step 1: Randomly generate 1000 three-dimensional coordinate points for simulating 1000 waypoints on the ground station map;

Step 2: Establish the matrix P of the input layer of the BP neural network;

$\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; =</span>\left[\begin{array}{cccccc}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 11</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 12</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 13</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 1999</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 11000</span>}}\\ {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; one</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; two</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; three</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 2999</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 21000</span>}}\\ {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 31</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 32</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 33</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 3999</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 31000</span>}}\end{array}\right]$

Among them, (a ₁₁ , a ₂₁ , a ₃₁ ) represent the first waypoint, (a ₂₁ , a ₂₂ , a ₃₂ ) represent the second waypoint, and so on, (a ₁₉₉₉ , a ₂₉₉₉ , a ₃₉₉₉ ) Indicates the 999th waypoint, (a ₁₁₀₀₀ , a ₂₁₀₀₀ , a ₃₁₀₀₀ ) indicates the 1000th waypoint.

Step 3: Establish the output terminal matrix T of the BP neural network;

$\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; =</span>\left[\begin{array}{cccccc}{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 11</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 12</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 13</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 1999</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 11000</span>}}\\ {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; one</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; two</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; three</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 2999</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 21000</span>}}\\ {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 31</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 32</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 33</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 3999</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 31000</span>}}\end{array}\right]$

Among them, (b ₁₁ , b ₂₁ , b ₃₁ ) represents the first waypoint passed by the system output, (b ₁₂ , b ₂₂ , b ₃₂ ) represents the second waypoint passed by the system output, and so on, (b ₁₉₉₉ , b ₂₉₉₉ , b ₃₉₉₉ ) represent the 999th waypoint passed by the system output, (b ₁₁₀₀₀ , b ₂₁₀₀₀ , b ₃₁₀₀₀ ) represent the 1000th waypoint passed by the system output. At the same time, the matrix P is equal to the matrix T, that is, P=T.

Step 4: Establish a BP neural network, net=netff(P, T, 3), wherein, P and T represent the input and output matrices of the BP neural network respectively, and 3 represents that the designed BP neural network has a hidden layer, and the implicit The number of layer neurons is 3;

Step 5: Train the BP neural network, [net,tr]=train(net,P,T), use the input and output matrices to train the BP neural network net, and obtain a new neural network net, tr is used to record training The number of steps epoch and performance perf;

Step 6: For the matrix output of the BP neural network, use the plot3 function to connect the waypoints represented by the matrix to generate the waypoint route of the UAV.

Compared with the prior art, the present invention increases the number of punctuation points that can be marked by the UAV, the punctuation route of the UAV is directional and can improve the UAV autopilot system, and innovatively applies the BP neural network technology, And on this basis, quickly and effectively realize the design of the waypoint route of the UAV ground station, which can be applied to the autopilot of the UAV;

As a novel neural network-based autopilot waypoint route scheme, it has great development prospects in the future research on UAV autopilot.

Description of drawings

Figure 1 is a schematic diagram of the calibration route under 100 waypoints;

Fig. 2 is a schematic diagram of the error performance curve between the actual flight route and the design route under 100 waypoints;

Among them, Train, Validation, and Test represent the performance of training results, the performance of inspection results, and the performance of verification results, respectively. Performance indicates system performance, and Epochs indicates the number of steps in system training. It can be seen from the figure that the errors of the three curves all drop to 10 ^-5 (percentage value) at 1000 steps.

Figure 3 is a schematic diagram of the calibration route under 1000 waypoints;

Fig. 4 is a schematic diagram of the error performance curve between the actual flight route and the design route under 1000 waypoints;

Among them, Train, Validation, Test, and Best represent the performance of training results, the performance of inspection results, the performance of verification results, and the performance of the system in the best state, respectively. Performance indicates system performance, and Epochs indicates the number of steps in system training. BestValidationPerformanceis0.00046723atepoch1000 indicates that the best validation performance is 0.00046723 when training to 1000 steps. At the same time, the system performance is represented by MSE (MeanSquaredError) with the minimum mean square error. It can be seen that the curves ①②③ in the figure overlap, indicating that the performance of training, testing, verification, and the best results are similar.

Fig. 5 is the overall flow chart of the present invention.

detailed description

The technical solution of the present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments.

There are less than 100 route waypoint calibrations for the automatic driving of ordinary UAVs. The present invention can increase the number of waypoints to 1000 based on BP neural network (ErrorBackPropagation, namely the error backpropagation algorithm neural network). Although it takes a certain amount of time to use the BP neural network to train the flight path, the present invention utilizes the innovative point based on the BP neural network, and the calibration accuracy can reach the order of 10 ^-4 (the ratio of the error value to the actual value). At the same time, when using the BP neural network, the direction of the waypoint can be calibrated to judge whether the flying direction of the UAV is correct. When the unmanned aerial vehicle flies between two waypoints, the error between the actual flight route and the designed route can be given by using the calibration result of the present invention, and if the error is too large, the system will give an alarm. In this way, it is beneficial to further improve the UAV autopilot.

Simulation of the present invention is carried out under MATLAB:

First utilize the randi function in MATLAB to generate random three-dimensional waypoint data (respectively adopting a group of 100 waypoints and 1000 waypoints of the present invention to compare); limit the data range, simulate the three-dimensional point in the air, and make the waypoint reach 1000 pieces.

Secondly, establish a BP neural network (using a three-layer neural network, which includes an input layer, an output layer, and a hidden layer), the hidden layer uses the tansig function, the output layer uses the purelin function, and the weight/threshold learning function of the BP network It is learngdm, and its performance function is MSE (MeanSquareError), that is, the mean square error function is used as the index of performance analysis;

Finally, curve fitting is performed on the above waypoints to form the flight path of the UAV.

The BP neural network uses 60% of the input data for training, 20% for testing, and 20% for verification. It adopts an early termination strategy to prevent over-fitting and improve accuracy. As shown in Figure 2 and Figure 4, the schematic diagram of the error performance curve between the actual flight route and the design route under 00 waypoints and the error performance curve diagram between the actual flight route and the design route under 1000 waypoints are respectively given , it can be seen that the errors are all below 10 ^-4 .

When using MATLAB simulation, use ">" to indicate the direction of the route and the direction of each waypoint. In this way, the UAV will not fly in the opposite direction along the route when it is flying forward. In the process of using the BP neural network to realize the UAV punctuation route, the direction between any two punctuation points can be given, that is, which punctuation point the UAV passes through first, and which punctuation point it passes through.

When the UAV is flying, its real-time flight data can be observed at the ground station. By calculating the distance between the current position point on the actual flight route and the two waypoints, that is, the BP neural network realizes the punctuation route of the UAV. When When the vertical distance between the actual flight position of the UAV and the route fitted by the neural network training reaches the set value, the system can give an alarm and judge the error of the UAV's current flight status, so that the UAV can drive automatically The instrument makes quick adjustments to reduce errors. This is just an error evaluation method, and there may be other equivalent methods.

For example, the UAV is flying in the sky in three-dimensional space, corresponding to any point of the UAV as a waypoint, and setting its three-dimensional coordinates (x, y, z), latitude, longitude and height from the reference plane. This reference plane can be set according to the actual situation. The latitude range is (-90°, 90°), the negative value indicates the south latitude, the positive value indicates the north latitude; the longitude range is (-180°, 180°), the negative value indicates the west longitude, and the positive value indicates the east longitude; the height range is ( 0,10000), the unit is meter, when generating height data, the maximum value is 10000 meters. At present, most drones on the market support 50 or 100 waypoints. The present invention trains the BP neural network on the basis of 1000 waypoints to achieve the purpose of drawing lines.

For the above 1000 waypoints, a BP neural network is established, the hidden layer uses the tansig function, the output layer uses the purelin function, the training function of the BP network is trainlm, the weight/threshold learning function of the BP network is learngdm, and its performance function is MSE (MeanSquareError), the mean square error as a function of error performance.

On the basis of the trained neural network, input 1000 waypoints, and make the network output 1000 waypoints correspondingly.

For these 1000 waypoints, curve fitting is performed to form the flight path of the drone.

The schematic diagram of the calibration route under 1000 waypoints as shown in Figure 3, the calibration method includes the following steps:

Step 1: Randomly generate 1000 three-dimensional coordinate points for simulating 1000 waypoints on the ground station map;

Step (2), set up the matrix P of BP neural network input layer;

$\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; =</span>\left[\begin{array}{cccccc}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 11</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 12</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 13</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 1999</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 11000</span>}}\\ {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; one</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; two</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; three</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 2999</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 21000</span>}}\\ {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 31</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 32</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 33</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 3999</span>}}& {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 31000</span>}}\end{array}\right]$

Among them, (a ₁₁ , a ₂₁ , a ₃₁ ) represent the first waypoint, (a ₂₁ , a ₂₂ , a ₃₂ ) represent the second waypoint, and so on, (a ₁₉₉₉ , a ₂₉₉₉ , a ₃₉₉₉ ) Indicates the 999th waypoint, (a ₁₁₀₀₀ , a ₂₁₀₀₀ , a ₃₁₀₀₀ ) indicates the 1000th waypoint.

Step (3), list BP neural network output terminal matrix T;

$\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; =</span>\left[\begin{array}{cccccc}{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 11</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 12</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 13</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 1999</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 11000</span>}}\\ {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; one</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; two</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; three</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 2999</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 21000</span>}}\\ {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 31</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 32</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 33</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 3999</span>}}& {\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 31000</span>}}\end{array}\right]$

Among them, (b ₁₁ , b ₂₁ , b ₃₁ ) represents the first waypoint passed by the system output, (b ₁₂ , b ₂₂ , b ₃₂ ) represents the second waypoint passed by the system output, and so on, (b ₁₉₉₉ , b ₂₉₉₉ , b ₃₉₉₉ ) represent the 999th waypoint passed by the system output, (b ₁₁₀₀₀ , b ₂₁₀₀₀ , b ₃₁₀₀₀ ) represent the 1000th waypoint passed by the system output. At the same time, the matrix P is equal to the matrix T, that is, P=T.

Step (4), set up BP neural network, net=netff (P, T, 3), wherein, P, T represent BP neural network input and output matrix respectively, 3 represents that the BP neural network of design has a hidden layer, and The number of hidden layer neurons is 3;

Step (5), train this BP neural network, [net, tr]=train (net, P, T), use input, output matrix pair

BP neural network net is used for training, and a new neural network net is obtained at the same time, and tr is used to record the number of training steps epoch and performance perf;

Step (6), for the matrix output of the BP neural network, use the plot3 function to map the route points represented by the matrix

Make a connection to generate the waypoint route of the drone.

As shown in Figure 4, it is the error performance under 1000 waypoints, and the method includes the following steps:

Step (7): On the basis of step (1), step (2), step (3), step (4), step (5), and step (6), use the MSE method to estimate the systematic error performance, that is, use plotperf() function.

The schematic diagram of the calibration route shown in Figure 1, its calibration method and error performance calculation method are the same as above, 100 three-dimensional waypoints are randomly generated for simulating 100 waypoints on the map of the ground station; the 100 waypoints shown in Figure 2 The error performance between the actual flight path and the designed path under the point.

Claims (1)

1., based on a unmanned plane drive route destination scaling method for BP neural network, it is characterized in that, the method comprises the following steps:

Step (1), random generation 1000 three-dimensional coordinate points, for simulating 1000 destinations on land station's map;

Step (2), sets up the matrix P of BP neural network input layer;

$P=\left[\begin{array}{cccccc}{a}_{11}& a_{12}& a_{13}& ...& a_{1999}& a_{11000}\\ a_{21}& a_{22}& a_{23}& ...& a_{2999}& a_{21000}\\ a_{31}& a_{32}& a_{33}& ...& a_{3999}& a_{31000}\end{array}\right]$

Wherein, (a ₁₁, a ₂₁, a ₃₁) represent the 1st destination, (a ₂₁, a ₂₂, a ₃₂) represent the 2nd destination, by that analogy, (a ₁₉₉₉, a ₂₉₉₉, a ₃₉₉₉) represent the 999th destination, (a ₁₁₀₀₀, a ₂₁₀₀₀, a ₃₁₀₀₀) represent the 1000th destination.

Step (3), lists BP neural network output terminal matrix T;

$T=\left[\begin{array}{cccccc}{b}_{11}& b_{12}& b_{13}& ...& b_{1999}& b_{11000}\\ b_{21}& b_{22}& b_{23}& ...& b_{2999}& b_{21000}\\ b_{31}& b_{32}& b_{33}& ...& b_{3999}& b_{31000}\end{array}\right]$

Wherein, (b ₁₁, b ₂₁, b ₃₁) represent that system exports the 1st destination of process, (b ₁₂, b ₂₂, b ₃₂) represent that system exports the 2nd destination of process, by that analogy, (b ₁₉₉₉, b ₂₉₉₉, b ₃₉₉₉) represent that system exports the 999th destination of process, (b ₁₁₀₀₀, b ₂₁₀₀₀, b ₃₁₀₀₀) represent that system exports the 1000th destination of process.Meanwhile, matrix P is equal with matrix T, i.e. P=T.

Step (4), sets up BP neural network, net=netff (P, T, 3), and wherein, P, T represent BP neural network input and output matrix respectively, and 3 represent that the BP neural network of design has a hidden layer, and hidden layer neuron number is 3;

Step (5), trains this BP neural network, [net, tr]=train (net, P, T), use input, output matrix are trained BP neural network net, and obtain new neural network net, tr is for recording step number epoch and the performance perf of training simultaneously;

Step (6), for the Output matrix of BP neural network, uses plot3 function to carry out line to the way point representated by matrix, generates the destination route of unmanned plane.