CN115328203B - Large-scale unmanned aerial vehicle cluster formation simulation acceleration method and system based on data driving - Google Patents
Large-scale unmanned aerial vehicle cluster formation simulation acceleration method and system based on data driving Download PDFInfo
- Publication number
- CN115328203B CN115328203B CN202211054695.2A CN202211054695A CN115328203B CN 115328203 B CN115328203 B CN 115328203B CN 202211054695 A CN202211054695 A CN 202211054695A CN 115328203 B CN115328203 B CN 115328203B
- Authority
- CN
- China
- Prior art keywords
- unmanned aerial
- aerial vehicle
- vehicle cluster
- cluster formation
- dimensional
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05D—SYSTEMS FOR CONTROLLING OR REGULATING NON-ELECTRIC VARIABLES
- G05D1/00—Control of position, course or altitude of land, water, air, or space vehicles, e.g. automatic pilot
- G05D1/10—Simultaneous control of position or course in three dimensions
- G05D1/101—Simultaneous control of position or course in three dimensions specially adapted for aircraft
- G05D1/104—Simultaneous control of position or course in three dimensions specially adapted for aircraft involving a plurality of aircrafts, e.g. formation flying
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Abstract
A large-scale unmanned aerial vehicle cluster formation simulation acceleration method and system based on data driving relates to the technical field of digital simulation and aims to solve the problem that a traditional linear agent model cannot effectively accelerate the large-scale unmanned aerial vehicle cluster formation simulation process while the simulation precision is guaranteed. According to the invention, the original high-dimensional unmanned aerial vehicle cluster state data is subjected to dimension reduction, and the deep neural network is used for fitting the linear proxy model error, so that the complexity of large-scale unmanned aerial vehicle cluster formation simulation modeling and the required computing resources during simulation computing are effectively reduced, and a certain simulation precision is ensured. According to the invention, the simulation speed of the large-scale unmanned aerial vehicle cluster formation behavior dynamic process is obviously improved by establishing the low-dimensional characteristic subspace of the original large-scale unmanned aerial vehicle cluster formation state, and the composite proxy model formed by combining the linear initial model and the deep neural network can accelerate the simulation speed of the large-scale unmanned aerial vehicle cluster formation behavior and keep higher simulation precision.
Description
Technical Field
The invention relates to the technical field of digital simulation, in particular to a large-scale unmanned aerial vehicle cluster formation simulation acceleration method and system based on data driving.
Background
Unmanned aerial vehicle has extensive use in military field and civilian field owing to have characteristics such as low cost, unmanned aerial vehicle casualties, equipment are simple, convenient operation and nimble reliable, in order to make unmanned aerial vehicle can exert better ability, carries out more complicated task, needs adopt the mode of many unmanned aerial vehicle formations flight to carry out the task in coordination. Because a large-scale unmanned aerial vehicle formation system is complex and corresponding physical experiments are difficult to perform, a digital simulation mode becomes a main mode for testing and developing large-scale unmanned aerial vehicle formation behaviors, and the simulation provides a good means for feasibility verification of an algorithm. However, when the number of the unmanned aerial vehicles in the unmanned aerial vehicle formation is large in scale, the simulation time consumption problem is relatively prominent, and the simulation time consumption problem becomes an important factor restricting the simulation development. Therefore, a method for accelerating the simulation speed of large-scale unmanned aerial vehicle cluster formation is necessary to be researched.
The currently common main mode for solving the problems of high model complexity and low analog simulation calculation speed in large-scale simulation is to simulate the simulation process of a high-precision physical model by adopting a proxy model. The calculation result of the proxy model is very close to that of the original model, but the calculation amount of the solution is small. The proxy model is built using a data-driven, bottom-up approach. It is generally assumed that the internal exact process of the original model run is unknown (and sometimes may be known), while the input-output behavior of the model is known. The proxy model is built by computing the responses (outputs) of the original model at carefully selected limited points (inputs). This process is also referred to as behavioral modeling or black box model. The method of using the proxy model to replace the physical experiment and the simulation experiment is very common in engineering design, and the proxy model can also be used in many other experiments or scientific fields with large solution calculation amount. The main challenges faced by this approach are: how to build a proxy model as accurate as possible using as few high-precision model solutions as possible.
Although the proxy model method can greatly reduce the amount of calculation in large-scale simulation compared with the original model fitted with the proxy model method, the defects of the proxy model method are also prominent. The proxy model method can only simulate a model or system approximating linearity or linearity, and for a model or system with a strong degree of nonlinearity, the proxy model method needs to increase an approximation item to improve local precision or needs to adopt a multi-segment respective simulation approximation mode. The mode of improving the simulation capability of the model agent model obviously increases the calculated amount, and the precision can not be ensured. For a large-scale unmanned aerial vehicle cluster formation simulation extremely strong nonlinear system, the proxy model method has the problems that the calculated amount cannot be effectively improved and the model approximation accuracy is poor, so that the proxy model method is difficult to apply to the scenes.
Disclosure of Invention
In view of the above problems, the invention provides a large-scale unmanned aerial vehicle cluster formation simulation acceleration method and system based on data driving, so as to solve the problem that the conventional linear proxy model cannot effectively accelerate the large-scale unmanned aerial vehicle cluster formation simulation process while ensuring the simulation accuracy.
According to an aspect of the invention, a data-driven large-scale unmanned aerial vehicle cluster formation simulation acceleration method is provided, and the method comprises the following steps:
acquiring unmanned aerial vehicle cluster formation tracks comprising multi-time unmanned aerial vehicle cluster formation state data;
inputting the unmanned aerial vehicle cluster formation track into a linear agent model simulating the unmanned aerial vehicle cluster formation process for training to obtain a trained linear agent model;
training an error fitting model based on a neural network to correct the prediction error of the linear agent model;
inputting initial state data of the unmanned aerial vehicle cluster formation to be predicted into the trained linear agent model, correcting prediction errors of the linear agent model by using the trained error fitting model based on the neural network, and obtaining a state prediction result of the unmanned aerial vehicle cluster formation to be predicted.
Further, the state data comprises a three-dimensional position and a three-dimensional speed state of the unmanned aerial vehicle.
Further, after the unmanned aerial vehicle cluster formation track comprising multi-time unmanned aerial vehicle cluster formation state data is obtained, the unmanned aerial vehicle cluster formation state data is subjected to dimensionality reduction processing, low-dimensional unmanned aerial vehicle cluster characteristic state data is obtained, and then the low-dimensional unmanned aerial vehicle cluster characteristic state data is input into a linear agent model simulating an unmanned aerial vehicle cluster formation process to be trained, so that a trained linear agent model is obtained.
Furthermore, in the fourth step, after the dimensionality reduction processing is carried out on the initial state data of the unmanned aerial vehicle cluster formation to be predicted, the initial state data is input into the trained linear agent model, the prediction error of the linear agent model is corrected by using the trained error fitting model based on the neural network, and the state prediction result of the unmanned aerial vehicle cluster formation to be predicted is obtained; and recovering the characteristic state data of the low-dimensional unmanned aerial vehicle cluster in the state prediction result into high-dimensional unmanned aerial vehicle cluster state data.
And further, performing dimensionality reduction on the unmanned aerial vehicle cluster formation state data by adopting a principal component analysis method to obtain low-dimensional unmanned aerial vehicle cluster characteristic state data.
Further, the third step specifically comprises:
step three, initializing the predicted low-dimensional characteristic states of the unmanned aerial vehicle cluster at the moments of t =0 and t =1 to be
Wherein z is 0 ,z 1 Respectively representing the actual low-dimensional characteristic states of the unmanned aerial vehicle cluster at the t =0 moment and the t =1 moment in the unmanned aerial vehicle cluster formation track;
step three, at the moment that t is larger than 1, utilizing the trained linear agent model to carry out clustering characteristic state on the low-dimensional unmanned aerial vehicle
And (3) predicting:
in the formula, alpha and beta represent model parameters of the linear proxy model; as a component multiplication of a vector;
respectively representing the low-dimensional unmanned aerial vehicle cluster characteristic states predicted by using a linear agent model at t-1 and t-2 moments;
thirdly, correcting the prediction error of the linear agent model by using the neural network phi to obtain the corrected low-dimensional unmanned aerial vehicle cluster characteristic state
In the formula, w represents a vector of the formation destination position of the unmanned aerial vehicle cluster relative to the initial position of the cluster center; θ represents a network parameter of the neural network Φ;
step three, calculating a loss function by using the average absolute error, and updating a network parameter theta;
and step three, iteratively executing the step three to the step three four until the maximum iteration times is reached, and stopping executing to obtain a trained error fitting model.
According to another aspect of the invention, a data-driven large-scale unmanned aerial vehicle cluster formation simulation acceleration system is provided, and the system comprises:
a state trajectory acquisition module configured to acquire unmanned aerial vehicle cluster formation trajectories including multi-time unmanned aerial vehicle cluster formation state data; the state data comprises a three-dimensional position and a three-dimensional speed state of the unmanned aerial vehicle;
the linear agent model training module is configured to input the unmanned aerial vehicle cluster formation track into a linear agent model simulating an unmanned aerial vehicle cluster formation process for training to obtain a trained linear agent model;
an error correction model training module configured to train a neural network-based error fitting model to correct a prediction error of the linear proxy model;
and the state prediction module is configured to input initial state data of the unmanned aerial vehicle cluster formation to be predicted into the trained linear agent model, correct prediction errors of the linear agent model by utilizing the trained error fitting model based on the neural network, and obtain a state prediction result of the unmanned aerial vehicle cluster formation to be predicted.
Further, the system further comprises a dimension reduction processing module configured to: after an unmanned aerial vehicle cluster formation track comprising multi-time unmanned aerial vehicle cluster formation state data is obtained, performing dimensionality reduction processing on the unmanned aerial vehicle cluster formation state data to obtain low-dimensional unmanned aerial vehicle cluster characteristic state data; and then inputting the low-dimensional unmanned aerial vehicle cluster characteristic state data into the linear agent model training module for training.
Further, the state prediction module carries out dimension reduction processing on initial state data of the unmanned aerial vehicle cluster formation to be predicted by the dimension reduction processing module; and after the state prediction result of the unmanned aerial vehicle cluster formation to be predicted is obtained, recovering the low-dimensional unmanned aerial vehicle cluster characteristic state data in the state prediction result into high-dimensional unmanned aerial vehicle cluster state data.
Further, the specific process of training the error fitting model based on the neural network in the error correction model training module includes:
step three, initializing the predicted low-dimensional characteristic states of the unmanned aerial vehicle cluster at the moments of t =0 and t =1 to be
Wherein z is 0 ,z 1 Respectively representing the actual low-dimensional characteristic states of the unmanned aerial vehicle cluster at the moments of t =0 and t =1 in the unmanned aerial vehicle cluster formation track;
step three, at the moment that t is larger than 1, utilizing the trained linear agent model to carry out clustering characteristic state on the low-dimensional unmanned aerial vehicle
And (3) prediction is carried out:
in the formula, alpha and beta represent model parameters of the linear proxy model; as a component multiplication of a vector;
respectively representing the low-dimensional unmanned aerial vehicle cluster characteristic states predicted by using a linear agent model at t-1 and t-2 moments;
thirdly, correcting the prediction error of the linear agent model by using the neural network phi to obtain the corrected low-dimensional unmanned aerial vehicle cluster characteristic state
In the formula, w represents a vector of the formation destination position of the unmanned aerial vehicle cluster relative to the initial position of the cluster center; θ represents a network parameter of the neural network Φ;
step three, calculating a loss function by using the average absolute error, and updating a network parameter theta;
and step three, iteratively executing the step two to the step three four until the maximum iteration times is reached, and stopping executing to obtain a trained error fitting model.
The beneficial technical effects of the invention are as follows:
according to the method, a data-driven method is adopted to construct a simulation agent model of large-scale unmanned aerial vehicle cluster formation, and the complexity of simulation modeling of the large-scale unmanned aerial vehicle cluster formation and the calculation resources required during simulation calculation can be effectively reduced by performing dimension reduction on original high-dimensional unmanned aerial vehicle cluster state data and using a method of fitting a deep neural network with linear agent model errors, and a certain simulation precision is ensured; compared with the traditional method, the method has the following advantages: 1) The simulation speed of the dynamic process of the large-scale unmanned aerial vehicle cluster formation behavior can be remarkably improved by establishing a low-dimensional feature subspace of the original large-scale unmanned aerial vehicle cluster formation state; 2) The composite proxy model combined by the linear initial model and the deep neural network can accelerate the behavior simulation speed of the large-scale unmanned aerial vehicle cluster formation while keeping high simulation accuracy.
Drawings
The present invention may be better understood by reference to the following description taken in conjunction with the accompanying drawings, which are incorporated in and form a part of this specification, and which are used to further illustrate preferred embodiments of the present invention and to explain the principles and advantages of the present invention.
Fig. 1 is a flowchart of a large-scale unmanned aerial vehicle cluster formation simulation acceleration method based on data driving according to an embodiment of the present invention;
FIG. 2 is a schematic structural diagram of an error fitting model based on a long-term and short-term memory network according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a loss value variation curve in the process of training the error fitting model based on the long-short term memory network in the embodiment of the invention;
FIG. 4 is a schematic diagram of a dynamic process of formation behavior of a large-scale unmanned aerial vehicle cluster in a two-dimensional scene in the embodiment of the present invention;
FIG. 5 is a schematic diagram of a dynamic process of formation behavior of a large-scale unmanned aerial vehicle cluster in a three-dimensional scene in the embodiment of the present invention;
fig. 6 is a schematic diagram for comparing the actual model of the formation behavior of the large-scale drone cluster with the running time of the agent model in the embodiment of the present invention;
fig. 7 is a schematic structural diagram of a large-scale unmanned aerial vehicle cluster formation simulation acceleration system based on data driving according to an embodiment of the present invention.
Detailed Description
In order that those skilled in the art will better understand the disclosure, exemplary embodiments or examples of the disclosure are described below with reference to the accompanying drawings. It is obvious that the described embodiments or examples are only some, but not all embodiments or examples of the invention. All other embodiments or examples obtained by a person of ordinary skill in the art based on the embodiments or examples of the present invention without any creative effort shall fall within the protection scope of the present invention.
The embodiment of the invention provides a large-scale unmanned aerial vehicle cluster formation simulation acceleration method based on data driving, and as shown in figure 1, the method comprises the following steps:
acquiring unmanned aerial vehicle cluster formation tracks comprising multi-time unmanned aerial vehicle cluster formation state data; the state data comprises a three-dimensional position and a three-dimensional speed state of the unmanned aerial vehicle;
inputting the unmanned aerial vehicle cluster formation track into a linear agent model simulating the unmanned aerial vehicle cluster formation process for training to obtain a trained linear agent model;
training an error fitting model based on a neural network to correct the prediction error of the linear agent model;
inputting initial state data of the unmanned aerial vehicle cluster formation to be predicted into the trained linear agent model, correcting prediction errors of the linear agent model by using the trained error fitting model based on the neural network, and obtaining a state prediction result of the unmanned aerial vehicle cluster formation to be predicted.
In this embodiment, preferably, after the formation trajectory of the unmanned aerial vehicle cluster including the multi-time unmanned aerial vehicle cluster formation state data is obtained, the unmanned aerial vehicle cluster formation state data is subjected to dimensionality reduction processing to obtain low-dimensional unmanned aerial vehicle cluster characteristic state data, and then the low-dimensional unmanned aerial vehicle cluster characteristic state data is input into a linear agent model simulating the unmanned aerial vehicle cluster formation process to be trained, so that a trained linear agent model is obtained.
In this embodiment, preferably, in the fourth step, after performing dimensionality reduction on initial state data of the to-be-predicted unmanned aerial vehicle cluster formation, inputting the initial state data into a trained linear agent model, and correcting a prediction error of the linear agent model by using a trained error fitting model based on a neural network to obtain a state prediction result of the to-be-predicted unmanned aerial vehicle cluster formation; and recovering the characteristic state data of the low-dimensional unmanned aerial vehicle cluster in the state prediction result into high-dimensional unmanned aerial vehicle cluster state data.
In this embodiment, preferably, a principal component analysis method is adopted to perform dimensionality reduction on the unmanned aerial vehicle cluster formation state data, and low-dimensional unmanned aerial vehicle cluster feature state data is obtained.
In this embodiment, preferably, the specific steps of step three include:
step three, initializing the predicted low-dimensional characteristic states of the unmanned aerial vehicle cluster at the moments of t =0 and t =1 to be
Wherein z is 0 ,z 1 Respectively representing the actual low-dimensional characteristic states of the unmanned aerial vehicle cluster at the t =0 moment and the t =1 moment in the unmanned aerial vehicle cluster formation track;
step three, at the moment that t is larger than 1, utilizing the trained linear agent model to carry out clustering characteristic state on the low-dimensional unmanned aerial vehicle
And (3) prediction is carried out:
in the formula, alpha and beta represent model parameters of the linear proxy model; as a component multiplication of a vector;
respectively representing the low-dimensional unmanned aerial vehicle cluster characteristic states predicted by using a linear agent model at t-1 and t-2 moments;
thirdly, correcting the prediction error of the linear agent model by using the neural network phi to obtain the corrected low-dimensional unmanned aerial vehicle cluster characteristic state
In the formula, w represents a vector of the formation destination position of the unmanned aerial vehicle cluster relative to the initial position of the cluster center; theta represents a network parameter of the neural network phi;
step four, calculating a loss function by using the average absolute error, and updating a network parameter theta;
and step three, iteratively executing the step three to the step three four until the maximum iteration times is reached, and stopping executing to obtain a trained error fitting model.
The invention further provides a large-scale unmanned aerial vehicle cluster formation simulation acceleration method based on data driving. Firstly, acquiring a large amount of high-precision unmanned aerial vehicle cluster formation data by using a large-scale unmanned aerial vehicle cluster formation simulation simulator for training a subsequent model; secondly, performing dimensionality reduction processing on the high-dimensional large-scale unmanned aerial vehicle cluster formation state data by adopting a principal component analysis method to obtain low-dimensional feature data capable of effectively representing the unmanned aerial vehicle cluster state; then, establishing an initial agent model for simulating the formation process of the unmanned aerial vehicle cluster based on the linear model, wherein the initial agent model is used for roughly simulating the dynamic process of the formation behavior of the unmanned aerial vehicle cluster; then, establishing an error fitting network for repairing the simulation error of the initial linear agent model based on the long-term and short-term memory network, wherein the error fitting network is used for correcting the simulation error of the initial linear agent model and improving the simulation precision of the dynamic process of the formation behavior of the unmanned aerial vehicle cluster; and finally, simulating the dynamic process of the large-scale unmanned aerial vehicle cluster formation behavior by using the established linear initial model and the error fitting network based on the long-short term memory network, and accelerating the simulation speed. The steps of the method of the embodiment of the present invention are explained in detail below.
The method comprises the following steps: acquiring a large amount of high-precision unmanned aerial vehicle cluster formation data by using a large-scale unmanned aerial vehicle cluster formation simulation simulator for training a subsequent model;
according to the embodiment of the invention, the data acquired by using the large-scale unmanned aerial vehicle cluster formation simulator can be represented as follows:
wherein τ is i (i is more than or equal to 1 and less than or equal to N) represents the acquired ith unmanned aerial vehicle cluster formation track, and is represented as tau i ＝[x 0 ,x 1 ,…x t …,x Ti ]Wherein
Indicating the formation status of the cluster of drones at time t, i.e.
c represents the number of drones in the cluster, and 6 represents the state dimension of each drone. Each unmanned plane every momentIs expressed as a three-dimensional position and three-dimensional velocity state of the drone, i.e.
T i Represents the ith trace τ i Of the length of (c).
The large-scale unmanned aerial vehicle cluster formation simulation simulator comprises a dynamic model of an unmanned aerial vehicle, a navigation guidance control model, a formation algorithm and other modules and is used for generating large-scale unmanned aerial vehicle cluster formation data in a specified environment.
Step two: performing dimensionality reduction processing by adopting a principal component analysis method aiming at the high-dimensional large-scale unmanned aerial vehicle cluster formation state data generated in the step one to obtain low-dimensional characteristic data capable of effectively representing the unmanned aerial vehicle cluster state;
according to the embodiment of the invention, firstly, a data set generated by a large-scale unmanned aerial vehicle cluster formation simulator is used
The cluster formation state data at all the time in (1) are combined into a matrix to obtain a combined state matrix:
the joint state matrix X represents all states of the drone cluster system generated by the large scale drone cluster formation simulator.
And then, compressing the combined state matrix X by using a principal component analysis method to obtain low-dimensional characteristic data capable of effectively representing the unmanned aerial vehicle cluster state. The data compression process is as follows:
Z＝U(X-x μ )
in the formula (I), the compound is shown in the specification,
for compression matrix, u is the data dimension of the cluster state after compression, x μ Is the mean of all states X, Z is the low-dimensional cluster feature state data after compression。
Step three: establishing an initial linear agent model for simulating the formation process of the unmanned aerial vehicle cluster on the basis of the low-dimensional feature data Z after dimension reduction in the second step, wherein the initial linear agent model is used for roughly simulating the dynamic process of the formation behavior of the unmanned aerial vehicle cluster;
according to the embodiment of the invention, considering that large-scale unmanned aerial vehicle cluster formation has certain inertia in the motion process, the established initial linear agent model for simulating the unmanned aerial vehicle cluster formation process is as follows:
in the formula (I), the compound is shown in the specification,
parameters, z, respectively, representing the initial linear proxy model t-1 ,z t-2 Respectively representing the low-dimensional characteristic data of the unmanned aerial vehicle cluster at the time t-1 and the time t-2,
the low-dimensional feature of the unmanned aerial vehicle cluster at time t, which indicates the initial linear proxy model prediction, indicates a component multiplication of the vector. The above parameters can be obtained by solving a linear least squares equation.
Step four: establishing an error fitting network for repairing the simulation error of the initial linear agent model established in the third step based on the long-short term memory network, wherein the error fitting network is used for correcting the simulation error of the initial linear model and improving the simulation precision of the dynamic process of the formation behavior of the unmanned aerial vehicle cluster;
according to the embodiment of the invention, a neural network model phi is established based on a long-term and short-term memory network so as to approximate the fitting error of an initial linear proxy model, namely:
in the formula, delta z t Representing the output value of the neural network model phiAnd correcting the simulation error of the initial linear model, wherein w is a vector of the formation destination position of the unmanned aerial vehicle cluster relative to the initial position of the cluster center, and is represented by w = [ x ] g -x 0 ,y g -y 0 ,z g -z 0 ]Wherein [ x ] g ,y g ,z g ]Destination location for unmanned aerial vehicle cluster formation, [ x [ ] 0 ,y 0 ,z 0 ]And forming an initial position of a center for the unmanned aerial vehicle cluster. Fig. 2 shows the established error fitting network based on the long-short term memory network.
It should be noted that, although the neural network used in the embodiment of the present invention is a long-short term memory network, the implementation process of the present invention does not depend on which kind of neural network is used, and other neural network models may be used to perform error correction on the linear proxy model in addition to the long-short term memory network.
Training parameters of a neural network model phi based on the compressed low-dimensional cluster feature state data Z, as shown in the following training flow:
step five: simulating the dynamic process of the large-scale unmanned aerial vehicle cluster formation behavior by using the initial linear agent model established in the third step and the trained error fitting network in the fourth step;
according to the embodiment of the invention, firstly, based on the initial low-dimensional unmanned aerial vehicle cluster formation state z 0 ,z 1 Predicting t (t) by using initial linear proxy model established in step three>2) The characteristic state of the cluster formation of the low-dimensional unmanned aerial vehicles at the moment is as follows:
the prediction error of the initial linear model is then repaired using the neural network model Φ, as follows:
and finally, restoring the low-dimensional unmanned aerial vehicle cluster characteristic data to the high-dimensional unmanned aerial vehicle cluster state data as follows:
the following experiments were further used to verify the technical effects of the present invention.
The correctness and the rationality of the invention are verified by adopting a digital simulation mode. Firstly, a large-scale unmanned aerial vehicle cluster formation consisting of 50 rotor unmanned aerial vehicles is constructed in a Python environment, and the original unmanned aerial vehicle cluster formation data for simulation agent model training are randomly generated under 100 different sets of initial states. In addition, a neural network model is built by adopting a Pytrch architecture. The simulation test software environment is Windows 10+ Python3.7, and the hardware environment is I9-9820X CPU + GTX1080TiGPU +64.0GB RAM.
The experiment first trains the error fitting neural network based on the long-short term memory network established in fig. 2, and the loss value curve in the training process is shown in fig. 3. The maximum training round number Max _ episde is 500, and the number of steps s considered when calculating the neural network loss function is 30. As can be seen from fig. 2, the training loss value decreases sharply as the number of training rounds increases at the beginning of training, stabilizes below 2000 after the number of training rounds reaches 100 rounds, and then remains substantially unchanged as the number of training rounds increases, indicating that the training process of the network converges. The training process above illustrates that the error fitting neural network based on the long-short term memory network established by the invention can learn stable network parameters through training data.
Secondly, the effectiveness of the method is verified through simulation of a large-scale unmanned aerial vehicle cluster formation behavior dynamic process formed by 50 rotor unmanned aerial vehicles at a time. In the process of using a principal component analysis method to reduce the dimension of the original unmanned aerial vehicle cluster data, 300-dimensional original state data is compressed to 80-dimensional low-dimension feature data. The original unmanned aerial vehicle cluster formation track, the low-dimensional unmanned aerial vehicle cluster formation track after the compression dimensionality reduction, the unmanned aerial vehicle cluster formation track predicted by the initial linear agent model and the unmanned aerial vehicle cluster formation track after the neural network correction are shown in fig. 4 and 5. According to the diagram, the characteristic state of the low-dimensional unmanned aerial vehicle cluster formation after dimensionality reduction by the principal component analysis method can well represent the original unmanned aerial vehicle cluster formation state, and the trajectory change process of the original unmanned aerial vehicle cluster formation can be basically and completely reproduced. The difference between the unmanned aerial vehicle cluster formation track predicted by the initial linear agent model and the original unmanned aerial vehicle cluster formation track is large, and the difference is larger and larger along with the movement of the unmanned aerial vehicle cluster formation, which shows that the formation behavior of the unmanned aerial vehicle cluster cannot be predicted accurately for a long time only by the linear agent model. On the basis, the unmanned aerial vehicle cluster formation track corrected by the neural network is basically overlapped with the original unmanned aerial vehicle cluster formation track, and the dynamic change process of the unmanned aerial vehicle cluster formation behavior can be well simulated.
The experiment results show that the method of principal component analysis dimensionality reduction, initial linear proxy model and neural network error fitting model can well fit the dynamic process of large-scale unmanned aerial vehicle cluster formation behaviors. The degree of simulation acceleration thereof is test verified below.
In the test process, 10 tests are randomly performed on the generation process of the original unmanned aerial vehicle cluster formation data and the generation process of the data for simulating the unmanned aerial vehicle cluster formation behavior by using the simulation agent model provided by the invention respectively, and the running time of each test is recorded, and fig. 6 shows the actual running time of the unmanned aerial vehicle cluster formation model and the running time of the simulation agent model provided by the invention. As can be seen from the figure, the actual model operating time is approximately 16.26s, and the proxy model operating time is approximately 0.40s, and therefore, the operating speed of the proxy model is approximately 40.62 times the operating speed of the actual model.
According to the test results, the simulation acceleration method for the large-scale unmanned aerial vehicle cluster formation based on data driving can reduce the original unmanned aerial vehicle cluster formation state data through a principal component analysis method, roughly estimate the dynamic process of unmanned aerial vehicle cluster formation behaviors by using an initial linear agent model, and remarkably improve the simulation speed of the dynamic process of the unmanned aerial vehicle cluster formation behaviors under the condition of keeping the simulation precision of the unmanned aerial vehicle cluster formation behaviors by using a composite method for repairing prediction errors by using a neural network based on a long-short term memory network. According to the method, accurate simulation and simulation acceleration of the large-scale unmanned aerial vehicle cluster formation behavior dynamic process can be achieved, and a new technical thought is provided for an implementation mode of accurate simulation and simulation acceleration of large-scale simulation entity cooperative behaviors.
Another embodiment of the present invention provides a large scale unmanned aerial vehicle cluster formation simulation acceleration system based on data driving, as shown in fig. 7, the system includes:
a state trajectory acquisition module 10 configured to acquire an unmanned aerial vehicle cluster formation trajectory including multi-time unmanned aerial vehicle cluster formation state data; the state data comprises a three-dimensional position and a three-dimensional speed state of the unmanned aerial vehicle;
the linear agent model training module 30 is configured to input the unmanned aerial vehicle cluster formation track into a linear agent model simulating an unmanned aerial vehicle cluster formation process for training, and obtain a trained linear agent model;
an error correction model training module 40 configured to train a neural network based error fitting model to correct the prediction error of the linear proxy model;
and the state prediction module 50 is configured to input initial state data of the formation of the unmanned aerial vehicle cluster to be predicted into the trained linear agent model, correct prediction errors of the linear agent model by using the trained error fitting model based on the neural network, and acquire a state prediction result of the formation of the unmanned aerial vehicle cluster to be predicted.
In this embodiment, preferably, the system further includes a dimension reduction processing module 20, and the dimension reduction processing module 20 is configured to: after an unmanned aerial vehicle cluster formation track comprising multi-time unmanned aerial vehicle cluster formation state data is obtained, performing dimensionality reduction processing on the unmanned aerial vehicle cluster formation state data to obtain low-dimensional unmanned aerial vehicle cluster characteristic state data; and then inputting the feature state data of the low-dimensional unmanned aerial vehicle cluster into the linear agent model training module 30 for training.
In this embodiment, preferably, in the state prediction module 50, the dimension reduction processing module 20 is used to perform dimension reduction processing on the initial state data of the to-be-predicted unmanned aerial vehicle cluster formation; and after the state prediction result of the unmanned aerial vehicle cluster formation to be predicted is obtained, recovering the low-dimensional unmanned aerial vehicle cluster characteristic state data in the state prediction result into high-dimensional unmanned aerial vehicle cluster state data.
In this embodiment, preferably, the specific process of training the error fitting model based on the neural network in the error correction model training module 40 includes:
step three, initializing the predicted low-dimensional characteristic states of the unmanned aerial vehicle cluster at the moments of t =0 and t =1 to be
Wherein z is 0 ,z 1 Respectively representing the actual low-dimensional characteristic states of the unmanned aerial vehicle cluster at the t =0 moment and the t =1 moment in the unmanned aerial vehicle cluster formation track;
step two, at the moment t is larger than 1, the trained linear agent model is utilized to carry out clustering on the low-dimensional unmanned aerial vehicles
And (3) prediction is carried out:
in the formula, alpha and beta represent model parameters of the linear proxy model; as a component multiplication of a vector;
respectively representing the low-dimensional unmanned aerial vehicle cluster characteristic states predicted by using a linear agent model at t-1 and t-2 moments;
thirdly, correcting the prediction error of the linear agent model by using the neural network phi to obtain the corrected low-dimensional unmanned aerial vehicle cluster characteristic state
In the formula, w represents a vector of the formation destination position of the unmanned aerial vehicle cluster relative to the initial position of the cluster center; theta represents a network parameter of the neural network phi;
step four, calculating a loss function by using the average absolute error, and updating a network parameter theta;
and step three, iteratively executing the step two to the step three four until the maximum iteration times is reached, and stopping executing to obtain a trained error fitting model.
The functions of the large-scale unmanned aerial vehicle cluster formation simulation acceleration system based on data driving in the embodiment of the invention can be described by the large-scale unmanned aerial vehicle cluster formation simulation acceleration method based on data driving, so that the detailed part of the system embodiment can be referred to the method embodiment, and the detailed description is omitted here.
While the invention has been described with respect to a limited number of embodiments, those skilled in the art, having benefit of this description, will appreciate that other embodiments can be devised which do not depart from the scope of the invention as described herein. The present invention has been disclosed in an illustrative rather than a restrictive sense, and the scope of the present invention is defined by the appended claims.
Claims (8)
1. A large-scale unmanned aerial vehicle cluster formation simulation acceleration method based on data driving is characterized by comprising the following steps:
acquiring an unmanned aerial vehicle cluster formation track comprising multi-time unmanned aerial vehicle cluster formation state data;
inputting the unmanned aerial vehicle cluster formation track into a linear agent model simulating the unmanned aerial vehicle cluster formation process for training to obtain a trained linear agent model;
step three, training an error fitting model based on a neural network to correct the prediction error of the linear agent model; the method comprises the following specific steps:
step three, initializing the predicted low-dimensional characteristic states of the unmanned aerial vehicle cluster at the moments of t =0 and t =1 to be
Wherein z is 0 ,z 1 Respectively representing the actual low-dimensional characteristic states of the unmanned aerial vehicle cluster at the moments of t =0 and t =1 in the unmanned aerial vehicle cluster formation track;
step two, at the moment t is larger than 1, the trained linear agent model is utilized to carry out clustering on the low-dimensional unmanned aerial vehicles
And (3) prediction is carried out:
in the formula, alpha and beta represent model parameters of the linear proxy model; e represents a component multiplication of the vector;
respectively representing the low-dimensional unmanned aerial vehicle cluster characteristic states predicted by using a linear agent model at t-1 and t-2 moments;
thirdly, correcting the prediction error of the linear agent model by using the neural network phi to obtain the corrected low-dimensional unmanned aerial vehicle cluster characteristic state
In the formula, w represents a vector of the formation destination position of the unmanned aerial vehicle cluster relative to the initial position of the cluster center; θ represents a network parameter of the neural network Φ;
step three, calculating a loss function by using the average absolute error, and updating a network parameter theta;
step three, iteratively executing the step two to the step three four until the maximum iteration times is reached, and stopping executing to obtain a trained error fitting model;
inputting initial state data of the unmanned aerial vehicle cluster formation to be predicted into the trained linear agent model, correcting prediction errors of the linear agent model by using the trained error fitting model based on the neural network, and obtaining a state prediction result of the unmanned aerial vehicle cluster formation to be predicted.
2. The data-driven large-scale unmanned aerial vehicle cluster formation simulation acceleration method according to claim 1, wherein the state data comprises three-dimensional position and three-dimensional speed states of the unmanned aerial vehicles.
3. The data-driven large-scale unmanned aerial vehicle cluster formation simulation acceleration method according to claim 2, characterized in that after an unmanned aerial vehicle cluster formation track comprising multi-time unmanned aerial vehicle cluster formation state data is obtained, the unmanned aerial vehicle cluster formation state data is subjected to dimensionality reduction processing to obtain low-dimensional unmanned aerial vehicle cluster feature state data, and then the low-dimensional unmanned aerial vehicle cluster feature state data is input into a linear agent model simulating an unmanned aerial vehicle cluster formation process to be trained to obtain a trained linear agent model.
4. The data-driven large-scale unmanned aerial vehicle cluster formation simulation acceleration method based on the claim 3 is characterized in that in the fourth step, after the dimensionality reduction processing is carried out on the initial state data of the unmanned aerial vehicle cluster formation to be predicted, the initial state data are input into a trained linear agent model, the prediction error of the linear agent model is corrected by using the trained error fitting model based on the neural network, and the state prediction result of the unmanned aerial vehicle cluster formation to be predicted is obtained; and recovering the characteristic state data of the low-dimensional unmanned aerial vehicle cluster in the state prediction result into high-dimensional unmanned aerial vehicle cluster state data.
5. The data-driven large-scale unmanned aerial vehicle cluster formation simulation acceleration method according to claim 4, characterized in that a principal component analysis method is adopted to perform dimensionality reduction processing on unmanned aerial vehicle cluster formation state data to obtain low-dimensional unmanned aerial vehicle cluster feature state data.
6. A large-scale unmanned aerial vehicle cluster formation simulation acceleration system based on data driving is characterized by comprising:
a state trajectory acquisition module configured to acquire unmanned aerial vehicle cluster formation trajectories including multi-time unmanned aerial vehicle cluster formation state data; the state data comprises a three-dimensional position and a three-dimensional speed state of the unmanned aerial vehicle;
the linear agent model training module is configured to input the unmanned aerial vehicle cluster formation track into a linear agent model simulating an unmanned aerial vehicle cluster formation process for training to obtain a trained linear agent model;
an error correction model training module configured to train a neural network-based error fitting model to correct a prediction error of the linear proxy model; the specific process of training the error fitting model based on the neural network comprises the following steps:
step three, initializing the predicted low-dimensional characteristic states of the unmanned aerial vehicle cluster at the moments of t =0 and t =1 to be
Wherein z is 0 ,z 1 Respectively representing the actual low-dimensional characteristic states of the unmanned aerial vehicle cluster at the moments of t =0 and t =1 in the unmanned aerial vehicle cluster formation track;
step two, at the moment t is larger than 1, the trained linear agent model is utilized to carry out the non-existence of the low dimensionHuman machine cluster feature state
And (3) prediction is carried out:
in the formula, alpha and beta represent model parameters of the linear proxy model; e represents a component multiplication of the vector;
respectively representing the low-dimensional unmanned aerial vehicle cluster characteristic states predicted by using a linear agent model at t-1 and t-2 moments;
thirdly, correcting the prediction error of the linear agent model by using the neural network phi to obtain the corrected low-dimensional unmanned aerial vehicle cluster characteristic state
In the formula, w represents a vector of the formation destination position of the unmanned aerial vehicle cluster relative to the initial position of the cluster center; θ represents a network parameter of the neural network Φ;
step three, calculating a loss function by using the average absolute error, and updating a network parameter theta;
step three, iteratively executing the step two to the step three four until the maximum iteration times is reached, and stopping executing to obtain a trained error fitting model;
and the state prediction module is configured to input initial state data of the unmanned aerial vehicle cluster formation to be predicted into the trained linear agent model, correct prediction errors of the linear agent model by utilizing the trained error fitting model based on the neural network, and obtain a state prediction result of the unmanned aerial vehicle cluster formation to be predicted.
7. The data-driven large-scale unmanned aerial vehicle cluster formation simulation acceleration system according to claim 6, wherein the system further comprises a dimension reduction processing module configured to: after an unmanned aerial vehicle cluster formation track comprising multi-time unmanned aerial vehicle cluster formation state data is obtained, performing dimensionality reduction processing on the unmanned aerial vehicle cluster formation state data to obtain low-dimensional unmanned aerial vehicle cluster characteristic state data; and then inputting the low-dimensional unmanned aerial vehicle cluster characteristic state data into the linear agent model training module for training.
8. The system according to claim 7, wherein the state prediction module performs dimension reduction on initial state data of the formation of the unmanned aerial vehicle cluster to be predicted by using the dimension reduction processing module; and after the state prediction result of the unmanned aerial vehicle cluster formation to be predicted is obtained, recovering the characteristic state data of the low-dimensional unmanned aerial vehicle cluster in the state prediction result into high-dimensional unmanned aerial vehicle cluster state data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211054695.2A CN115328203B (en) | 2022-08-31 | 2022-08-31 | Large-scale unmanned aerial vehicle cluster formation simulation acceleration method and system based on data driving |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211054695.2A CN115328203B (en) | 2022-08-31 | 2022-08-31 | Large-scale unmanned aerial vehicle cluster formation simulation acceleration method and system based on data driving |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115328203A CN115328203A (en) | 2022-11-11 |
| CN115328203B true CN115328203B (en) | 2023-03-14 |
Family
ID=83928739
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211054695.2A Active CN115328203B (en) | 2022-08-31 | 2022-08-31 | Large-scale unmanned aerial vehicle cluster formation simulation acceleration method and system based on data driving |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115328203B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110007688A (en) * | 2019-04-25 | 2019-07-12 | 西安电子科技大学 | A kind of cluster distributed formation method of unmanned plane based on intensified learning |
| CN112819303A (en) * | 2021-01-22 | 2021-05-18 | 中国人民解放军国防科技大学 | PCE agent model-based aircraft tracking efficiency evaluation method and system |
| CN113886953A (en) * | 2021-09-27 | 2022-01-04 | 中国人民解放军军事科学院国防科技创新研究院 | Unmanned aerial vehicle intelligent simulation training method and device based on distributed reinforcement learning |
| CN113887134A (en) * | 2021-09-27 | 2022-01-04 | 南京航空航天大学 | Large-scale lightweight unmanned cluster test simulation platform and working method |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111694365B (en) * | 2020-07-01 | 2021-04-20 | 武汉理工大学 | Unmanned ship formation path tracking method based on deep reinforcement learning |
-
2022
- 2022-08-31 CN CN202211054695.2A patent/CN115328203B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110007688A (en) * | 2019-04-25 | 2019-07-12 | 西安电子科技大学 | A kind of cluster distributed formation method of unmanned plane based on intensified learning |
| CN112819303A (en) * | 2021-01-22 | 2021-05-18 | 中国人民解放军国防科技大学 | PCE agent model-based aircraft tracking efficiency evaluation method and system |
| CN113886953A (en) * | 2021-09-27 | 2022-01-04 | 中国人民解放军军事科学院国防科技创新研究院 | Unmanned aerial vehicle intelligent simulation training method and device based on distributed reinforcement learning |
| CN113887134A (en) * | 2021-09-27 | 2022-01-04 | 南京航空航天大学 | Large-scale lightweight unmanned cluster test simulation platform and working method |
Non-Patent Citations (1)
| Title |
|---|
| 闫茂德 ; 周小强 ; 张逊逊 ; 朱旭 ; .关于异构无人器地空联合编队控制仿真研究.2016,第33卷(第10期),361-365. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115328203A (en) | 2022-11-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Chee et al. | Knode-mpc: A knowledge-based data-driven predictive control framework for aerial robots | |
| KR20190031318A (en) | Domain Separation Neural Networks | |
| CN112100856A (en) | Automatic driving joint simulation method based on multiple platforms | |
| CN107633105B (en) | Improved hybrid frog-leaping algorithm-based quad-rotor unmanned aerial vehicle parameter identification method | |
| WO2022156182A1 (en) | Methods and apparatuses for constructing vehicle dynamics model and for predicting vehicle state information | |
| CN115239593A (en) | Image restoration method, image restoration device, electronic device, and storage medium | |
| CN114290339B (en) | Robot realistic migration method based on reinforcement learning and residual modeling | |
| Khatiri et al. | Simulation-based test case generation for unmanned aerial vehicles in the neighborhood of real flights | |
| Blanco-Claraco et al. | Multibody dynamic systems as Bayesian networks: Applications to robust state estimation of mechanisms | |
| CN115328203B (en) | Large-scale unmanned aerial vehicle cluster formation simulation acceleration method and system based on data driving | |
| CN111553954B (en) | Online luminosity calibration method based on direct method monocular SLAM | |
| CN113641099A (en) | Impedance control imitation learning training method for surpassing expert demonstration | |
| Su et al. | A ROS based open source simulation environment for robotics beginners | |
| CN115958606A (en) | Mechanical arm obstacle avoidance track planning method and device and electronic equipment | |
| CN115098941A (en) | Unmanned aerial vehicle digital twin control method and platform for agile deployment of intelligent algorithm | |
| CN113977580A (en) | Mechanical arm simulation learning method based on dynamic motion primitives and adaptive control | |
| Delcaro et al. | Automatic dimensionality reduction of Twin-in-the-Loop Observers | |
| Cook et al. | Towards real-time robot simulation on uneven terrain using neural networks | |
| CN112068434B (en) | Iterative learning control method for rehabilitation mechanical device based on extended state observer | |
| CN113985732B (en) | Adaptive neural network control method and device for aircraft system | |
| Omainska et al. | Rigid Motion Gaussian Processes With SE (3) Kernel and Application to Visual Pursuit Control | |
| CN117407319B (en) | Target generation method and device for radar system software test and electronic equipment | |
| CN117863152A (en) | Robot assembly learning method based on teaching rewarding state machine and residual error reinforcement learning | |
| KR102484249B1 (en) | Method and apparatus for improving the accuracy of 6-dof flight simulation using flight test data | |
| CN114035594B (en) | Method and device for controlling aircraft subject to class clearance lag and output constraints |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |