Multi-UAV Swarm Navigation Method under Communication Delay Constraints

Abstract

The present invention proposes a multi-UAV swarm navigation method under communication delay constraints, which solves the technical problem that it is difficult for a swarm system to achieve stable swarming under the condition of communication delay between UAVs in a UAV swarm. The present invention divides the swarm navigation process into a preparation stage and a flight stage. The preparation stage includes motion modeling of UAVs in the swarm, processing UAV flight state data to obtain training samples, constructing an LSTM model, and training the LSTM model with the training samples. The flight stage includes UAV takeoff initialization, sending interactive flight information, and receiving interactive flight information, finally realizing multi-UAV swarm navigation under communication delay constraints. The present invention has the advantages of small communication and computing overhead and can achieve stable swarming flight of multi-UAVs under communication delay constraints. Applied to UAV swarm navigation, it can avoid the swarming flight jitter caused by time delay during the UAV swarm flight process and reduce the probability of collision during the UAV swarm flight process.

Description

Technical Field

[0001] The present invention belongs to the field of multi-agent collaborative control, mainly relates to the navigation of unmanned aerial vehicle (UAV) swarms, and specifically relates to a method for the navigation of multi-UAV swarms under communication delay constraints, which is mainly applied to the distributed navigation of UAV swarms. Background Art

[0002] At present, with the continuous improvement of the intelligence level of UAVs, the application fields of UAVs are becoming wider and wider. A single UAV can no longer meet the complex mission requirements, and the UAV swarm control technology has become an important research direction in the development of the UAV field. A UAV swarm system refers to an autonomous aerial intelligent system composed of a certain number of UAVs, which realizes the behavioral coordination among them through information interaction and feedback, and excitation and response to jointly complete specific tasks. At present, the UAV swarm control methods are mainly divided into two categories. One is the centralized control method, which relies on the effective communication between the UAVs and the central node, and uses a central controller to coordinate the movement of the UAVs. The other is the distributed control method, in which there is no central controller, and the UAVs complete self-organized collision avoidance through autonomous decision-making. Many existing methods assume that the UAVs can obtain the flight information of the surrounding UAVs in real time through communication. However, in practical applications, there are often problems such as unstable delays and interruptions in the communication between UAVs, resulting in the UAVs being unable to obtain the flight information of the surrounding UAVs in real time, and the surrounding UAVs being unable to obtain the flight information of this UAV in real time, thereby reducing the accuracy of the planned flight path, affecting the stability of the control system, and greatly increasing the probability of UAV collisions.

[0003] Since 2014, research teams have applied the designed swarm algorithms to real outdoor UAV swarms, achieving stable autonomous outdoor flight of up to ten UAVs for the first time. In the algorithms designed by the team, the impact of time delay is considered. Each UAV in the swarm uses the historical flight information of its surrounding neighbor UAVs to plan its own flight path, and adjusts the global parameters to make the system reach a stable swarm state. The problem with this delay solution is that when the number of UAVs in the swarm system is large or the distance between them is small, collisions are likely to occur between UAVs, and parameter adjustment in complex scenarios requires a large amount of computational and communication overhead. In addition, there have been many studies on time delay problems in cybernetics. To solve the delay problem of signal transmission in multi-UAV swarm communication, the general solution is to design a reasonable control protocol, use matrix theory and stability theory to obtain the sufficient conditions for the system to reach lag consensus, and prove that the system can finally reach a consensus state under these conditions. These control theories are based on the assumption that the controller can obtain global information and establish an accurate model for the environment and UAVs. However, in the actual operating environment, due to communication limitations, it is difficult for a single UAV to obtain global information. In addition, the models designed based on control theory need to meet a series of constraints to ensure their usability.

[0004] Existing multi-UAV swarm navigation schemes with communication delay constraints mainly focus on the design of control protocols. Usually, in the control protocol, the controller needs to obtain global information and establish an accurate model for the environment and UAVs, and then derive the constraint conditions that the system needs to meet to finally reach consensus based on the model equations. By proving, it is verified that when the swarm system meets these constraint conditions, the swarm navigation movement can reach a final consensus. The drawback of this technical solution is that it is difficult to ensure the consensus of the swarm in complex environments, and the robustness is low. Even a slight change in the environment may cause the consensus constraint conditions to be not met, resulting in the inability of the system's swarm navigation movement to reach a final consensus. Summary of the Invention

[0005] The object of the present invention is to address the deficiencies of the prior art and propose a multi-UAV swarm navigation method under communication delay that can ensure stable swarm flight of UAVs even when there is a large or small communication delay in the communication between UAVs in the swarm.

[0006] The present invention is a multi-UAV swarm navigation method under communication delay constraints, characterized in that the swarm navigation process is divided into a preparation stage and a flight stage, including the following steps:

[0007] Preparation Stage

[0008] S1. Modeling the movement of drones in the cluster: Neighbor drones in the cluster refer to other drones within a certain communication range of the current drone. The control function is a mapping relationship from flight state to flight speed. When each drone calculates and outputs the flight speed at the next moment through the control function, it only obtains the surrounding drones within its own communication threshold r. c The flight position and flight speed of neighboring drones within the range are used as inputs, and the speed generated by the control function is set to be affected by the autonomous navigation force, the repulsive force generated by neighboring drones, the attractive force generated by neighboring drones, the alignment force generated by neighboring drones, and the repulsive force generated by obstacles; each force corresponds to a control function component, and the repulsive, attractive, and alignment forces generated by neighboring drones and the repulsive force generated by obstacles are divided into their own control function sub-components according to the differences between neighbors and obstacles; the control function sub-components are accumulated to obtain the control function component, and the control function is the sum of all control function components; the output of the drone motion model is the output of the control function, and one flight moment of the drone corresponds to a time step, assuming the step length is Δt. The current drone calculates its own flight speed at the next moment through the control function, and each drone in the cluster will calculate and output the flight speed corresponding to its own flight state through the control function under the flight state at different moments;

[0009] S2. Process the flight status data of the drone to obtain training samples: a Data space is allocated in the internal storage space of each drone control chip to store the flight status data of the drone; each drone sends its own flight position and flight speed information to neighboring drones and receives the flight position and flight speed information of neighboring drones; collects the flight status data of drones flying in clusters in a real cluster environment or a simulated environment and puts it into the Data space; reads all the flight status data of the drones and divides them into an input feature set and an output feature set, each piece of data saved in the form of a data record in the input feature set is the flight status data of the drone at multiple moments before the current moment; each piece of data saved in the form of a data record in the output feature set is the flight position and flight speed data of the drone at multiple moments after the current moment; one piece of data in the input feature set corresponds to one piece of data in the output feature set, and each piece of data saved in the form of a data record in the input feature set and each piece of data saved in the form of a data record in the output feature set are matched one by one to form a pair of training samples, and all data records in the input feature set and the output feature set are summarized to obtain a training sample set of the drone flight status;

[0010] S3. Construct an LSTM model and train the LSTM model with training samples: The input data of the long short-term memory (LSTM) model is the flight state data of the UAV at the current moment and multiple previous moments, and the output prediction data is the flight position and flight speed of the UAV at multiple future moments. The constructed LSTM model includes at least two hidden layers and an attention layer, and the Adam optimizer is used. The input data is input into the first hidden layer of the LSTM model, and the data is processed layer by layer and output to the attention layer. After being processed by the attention layer, the prediction data is output. Before training, initialize the weights of the LSTM model, import the training sample set of the UAV flight state to train the LSTM model parameters, and obtain the trained LSTM model.

[0011] Flight phase

[0012] S4. UAV takeoff initialization: Apply the trained LSTM model to all UAVs, initialize the storage space in the internal chip of the UAV, including the Data space. At this moment, the time stamp t = 0, initialize the flight position and flight speed for each UAV, and set the UAV flight time step to Δt.

[0013] S5. Send interactive flight information: The UAV sends interactive flight information through a wireless communication device. In the initial flight phase, the UAV sends the real flight position and flight speed at the current moment as interactive flight information to neighboring UAVs. In the predicted flight phase, the UAV sends the real flight position and flight speed at the current moment and the predicted flight positions and flight speeds at multiple future moments as interactive flight information to neighboring UAVs. The predicted flight positions and predicted flight speeds are calculated and output according to the trained LSTM model.

[0014] S6. Receive interactive flight information: The current UAV obtains the interactive flight information sent by neighboring UAVs through communication. The current UAV calculates its own flight speed at the next moment through a control function according to the interactive flight information. When calculating, it is necessary to read the flight position and flight speed of neighboring UAVs at the current moment. However, the interactive flight information sent by some neighboring UAVs cannot be received by the current UAV on time due to communication delays. If the interactive flight information of a certain neighboring UAV at the current moment is not received by the current UAV, the current UAV reads the latest predicted flight position and flight speed regarding the current moment in the received interactive flight information of this neighboring UAV, and uses them to calculate the sub-components of the control function generated by all the influences corresponding to this neighboring UAV. If the interactive flight information sent by a neighboring UAV is received, the current UAV directly reads the interactive flight information of this neighboring UAV at the current moment, and uses the flight position and flight speed of this neighboring UAV at the current moment to calculate the sub-components of the control function generated by all the influences corresponding to this neighboring UAV. When calculating the flight speed of the current UAV at the next moment through the control function, the flight position and flight speed of the current UAV itself, the flight positions and flight speeds of all neighboring UAVs within the communication range, and the position information of obstacles are used as the input of the control function to calculate the flight speed of the current UAV at the next moment, so as to achieve multi-UAV cluster navigation under communication delays.

[0015] The present invention solves the technical problem that it is difficult for UAVs to stably cluster under the condition of more or less uncertain communication delays in the communication between UAVs in a UAV swarm.

[0016] The present invention has the following advantages compared with the prior art:

[0017] Small communication and computing overhead: In the present invention, when a UAV plans its next movement path, it only needs to obtain the flight information of a certain number of UAVs closest to itself within a certain communication range, without obtaining the information of all UAVs globally to establish an accurate model, which can greatly reduce the communication overhead and computing overhead. The UAV is less affected by uncertainties when planning a path in a complex environment, and it has stronger robustness compared with traditional control methods.

[0018] Achieve stable multi-UAV cluster flight under more or less communication delay constraints: The present invention introduces an LSTM model to predict the flight speed and flight position of a UAV after a certain time delay and send them to neighboring UAVs. When a neighboring UAV cannot obtain the real-time flight position and flight speed information of this UAV due to communication delays, it can read the predicted information at the current moment from the historical prediction information of this UAV. This prediction model uses the historical flight state data of multiple moments to predict the flight position and flight speed of the UAV in multiple future moments, and can achieve a high prediction accuracy after training, and can cope with more or less communication delays in the communication network, and can reduce the collision risk brought by communication delays to the system.

[0019] Compared with traditional control algorithms, the present invention can better solve the problem of multi-UAV swarm navigation under communication delay constraints, reduce the communication and computing overhead in the swarm, improve the prediction accuracy of the LSTM model through historical experience data, and perform interpolation operations based on the model prediction values under the condition of uncertain communication delay in the system, thereby improving the reliability of the UAV swarm system under communication delay constraints. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a schematic flowchart of the present invention;

[0021] Figure 2 is a schematic diagram of the swarm environment where UAV i is located under the condition of limited communication range;

[0022] Figure 3 is a schematic diagram of the LSTM model framework of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0023] Embodiment 1

[0024] In view of the fact that the distributed control method has better autonomy, the present invention uses the distributed control method. Each UAV flying autonomously in the system needs to obtain the flight information of surrounding neighbor UAVs through wireless communication, including flight position, flight speed and other information, and then uses this information to plan its flight path.

[0025] Each UAV flying autonomously in the UAV swarm system needs to obtain the flight information of surrounding neighbor UAVs through wireless communication, and then uses this information to plan its flight path. Many existing swarm methods assume that UAVs can obtain the flight information of surrounding UAVs in real time through communication, but in actual applications, there are often problems such as unstable delays and interruptions in the communication between UAVs, resulting in the UAV being unable to obtain the flight information of surrounding UAVs in real time, thus leading to collisions.

[0026] Existing UAV swarm solutions considering communication delay constraints are mainly based on the design of control protocols. Usually, in the control protocol, the controller needs to obtain global information and establish an accurate model for the environment and UAVs, and then derive the constraint conditions that the system needs to meet to achieve final consistency according to the model. By proving and verifying that when the swarm system meets these constraint conditions, the swarm navigation movement can achieve final consistency. The deficiencies of this technical solution are that it is difficult to ensure consistency in complex environments and the robustness is low. Even if the environment changes slightly, it may cause the consistency constraint conditions not to be met, resulting in the swarm navigation movement of the system not achieving final consistency. To overcome these defects, the present invention proposes a multi-UAV swarm navigation method under communication delay constraints.

[0027] The present invention is a multi-UAV cluster navigation method under communication delay constraints. Refer to Figure 1 , Figure 1 , which is a schematic flow diagram of the present invention. The present invention divides the cluster navigation process into a preparation stage and a flight stage, including the following steps:

[0028] Preparation stage

[0029] S1. Model the motion of UAVs in the cluster: In the cluster, neighboring UAVs refer to other UAVs within a certain communication range of the current UAV. The control function is a mapping relationship from the flight state to the flight speed. When each UAV calculates and outputs the flight speed at the next moment through the control function, it only obtains the flight positions and flight speeds of neighboring UAVs within its own communication threshold range as inputs. The communication threshold is set to r c , that is, |x i - x j | < r c (x i and x j respectively represent the positions of the current UAV i and the neighboring UAV j). It is assumed that the speed generated by the control function is affected by the autonomous navigation force, the repulsive force generated by neighboring UAVs, the attractive force generated by neighboring UAVs, the alignment force generated by neighboring UAVs, and the repulsive force generated by obstacles; each force corresponds to a control function component. The repulsive, attractive, and alignment forces generated by neighboring UAVs and the repulsive force generated by obstacles are divided into their respective control function sub-components according to different neighbors and obstacles; the control function sub-components are accumulated to obtain the control function component, and the control function is the sum of all control function components; the output of the UAV motion model is the output of the control function. One flight moment of the UAV corresponds to one time step. Let the time step be Δt. The current UAV calculates its own flight speed at the next moment through the control function. Each UAV in the cluster will calculate and output the flight speed corresponding to its own flight state through the control function at different flight states at different times.

[0030] Each control function component in the UAV motion model contains a certain number of parameters. According to the cluster target, a certain number of evaluation indicators are set, such as the speed consistency evaluation indicator and the collision risk evaluation indicator. Optimize the parameters so that the cluster can achieve better evaluation results, and then apply the corresponding parameters to the control function to act on all UAVs.

[0031] S2. Process the flight status data of the drone to obtain training samples: a Data space is allocated in the internal storage space of each drone control chip to store the flight status data of the drone. The drones in the cluster communicate with each other through wireless communication equipment. Each drone sends its own flight position and flight speed information to the neighboring drone and receives the flight position and flight speed information of the neighboring drone. The flight status data of the drones flying in the cluster in the real cluster environment or simulation environment is collected and put into the Data space. Read all the flight status data of the drones and divide them into an input feature set and an output feature set. Each data saved in the form of a data record in the input feature set is the flight status data of the drone at the current moment and multiple moments before; each data saved in the form of a data record in the output feature set is the flight position and flight speed data of the drone at multiple moments after the current moment; one piece of data in the input feature set corresponds to one piece of data in the output feature set, and each piece of data saved in the form of a data record in the input feature set corresponds to each piece of data saved in the form of a data record in the output feature set, forming a pair of training samples, and summarizing all data records in the input feature set and the output feature set to obtain a training sample set of the drone flight status.

[0032] S3. Build an LSTM model and train the LSTM model with training samples: The input data of the long short-term memory network LSTM model is the flight status data of the drone at the current moment and multiple moments before, and the output prediction data is the flight position and flight speed of the drone at multiple moments in the future; the constructed LSTM model framework can be found in Figure 3 The model contains at least two hidden layers and one attention layer, and uses the Adam optimizer; the input data is input into the first hidden layer of the LSTM model, and the data is processed layer by layer and output to the attention layer, and the predicted data is output after the attention layer operation processing; the weight of the LSTM model is initialized before training, and the training sample set of the UAV flight status is imported to train the LSTM model parameters, and the trained LSTM model is obtained.

[0033] Flight phase

[0034] S4, UAV takeoff initialization: Save the LSTM model parameters trained in step S3 inside all UAV control chips. The UAV can input historical flight status data to obtain the predicted flight position and flight speed information through the LSTM model. Initialize the storage space in the UAV internal chip, including the Data space, the initial timestamp t = 0, initialize the flight position and flight speed for each UAV, and set the UAV flight time step to Δt.

[0035] S5. Sending Interactive Flight Information: The UAV sends interactive flight information through a wireless communication device. In the initial flight stage, the UAV sends the real flight position and flight speed at the current moment as the interactive flight information to neighboring UAVs. In the predicted flight stage, the UAV sends the real flight position and flight speed at the current moment, as well as the predicted flight positions and flight speeds at multiple future moments, as the interactive flight information to neighboring UAVs. The predicted flight positions and flight speeds are calculated and output according to the trained LSTM model.

[0036] S6. Receiving Interactive Flight Information: The current UAV obtains the interactive flight information sent by neighboring UAVs in step S5 through communication. The current UAV calculates its flight speed at the next moment through a control function based on the interactive flight information, and it needs to read the flight position and flight speed of neighboring UAVs at the current moment during the calculation. However, the interactive flight information sent by some neighboring UAVs may not be received by the current UAV on time due to communication delays. If the interactive flight information of a certain neighboring UAV at the current moment is not received by the current UAV, the current UAV reads the latest predicted flight position and flight speed regarding the current moment in the interactive flight information of this neighboring UAV that has been received, and uses them to calculate the sub-components of the control function generated by all the influences corresponding to this neighboring UAV. Since the interactive flight information of neighboring UAVs predicts the flight positions and flight speeds at multiple future moments, when there are uncertain communication delays in UAV communication, the current UAV can read the latest obtained predicted flight positions and flight speeds of this neighboring UAV at multiple future moments. If the interactive flight information sent by a neighboring UAV is received, the current UAV directly reads the interactive flight information of this neighboring UAV at the current moment to obtain the flight position and flight speed of this neighboring UAV at the current moment, and uses them to calculate the sub-components of the control function generated by all the influences corresponding to this neighboring UAV. When calculating the flight speed of the current UAV at the next moment through the control function, the flight position and flight speed of the current UAV itself, the flight positions and flight speeds of all neighboring UAVs within the communication range, and the position information of obstacles are used as the input of the control function to calculate the flight speed of the current UAV at the next moment, that is, the sub-components of the control function corresponding to the repulsive, attractive, and alignment forces generated by all neighboring UAVs within the communication range of the current UAV, the sub-component of the control function corresponding to the self-driving force of the current UAV, and the sub-component of the control function corresponding to the obstacle force are accumulated to obtain the speed of the current UAV at the next moment, so as to avoid collisions with neighboring UAVs or obstacles in the cluster and achieve multi-UAV cluster navigation under communication delays.

[0037] Traditional control schemes usually obtain global information and establish an accurate model for the environment and the UAVs. Then, based on the model, equations are constructed to derive the constraint conditions that the system needs to meet to achieve time-delay consistency finally. The problem with this scheme is that the construction of the model requires obtaining global information and has low robustness. Even a slight change in the environment may cause the system consistency constraints not to be met, making the system unstable. The multi-UAV cluster navigation scheme under communication delay constraints proposed by the present invention first models the motion of the UAVs. The UAVs only obtain the flight information of neighbor UAVs within a certain communication range, design the flight scenario and reasonable evaluation indicators, train the parameter values in the control function, apply the parameter values with better evaluation effects to the control function, and then obtain the flight state data of the UAVs in the real and simulated environments, divide it into input feature data and output feature data, and use it as the training samples of the LSTM model. Optimize the model parameters, use this model as a prediction model, and use the historical flight state data of the UAVs at multiple moments to predict the flight positions and flight speed information at multiple future moments. In this way, when the real-time flight position and flight speed sent by the UAV are not received by neighbor UAVs due to time delay, the neighbor UAVs can read the predicted information sent by this UAV at the previous moment.

[0038] The scheme of the present invention can complete the cluster motion of the system without obtaining global information, and improve the prediction accuracy of the LSTM model through historical experience data. Under the condition of uncertain communication delay, perform interpolation operations through the model prediction values, thereby improving the reliability of the UAV cluster system under communication delay constraints; by predicting the flight positions and flight speeds of the UAVs at multiple future moments to cope with the possible uncertain communication delay in the network, it can better ensure the stability of the cluster.

[0039] Embodiment 2

[0040] The multi-UAV cluster navigation method under communication delay constraints is the same as that in Embodiment 1. The modeling of the motion of the UAVs in the cluster described in step S1 includes the following steps:

[0041] S11. Set the control function: The control function is the sum of all flight speeds generated by the current UAV's autonomous navigation force, the repulsive force generated by neighbor UAVs, the attractive force generated by neighbor UAVs, the alignment force generated by neighbor UAVs, and the repulsive force generated by obstacles on the speed of the current UAV. The calculation formula of the motion model control function for the speed of the next moment of UAV i is:

[0042]

[0043] Among them, represents the speed of the next moment calculated by UAV i; v i represents the speed of UAV i at the current moment; v preferRepresents the expected speed value during the flight of the UAV. Represents the speed control sub-component generated by the attracting force of neighboring UAV j on UAV i. Represents the speed control sub-component generated by the aligning force of neighboring UAV j on UAV i. Represents the speed control sub-component generated by the repulsive force of neighboring UAV j on UAV i. Represents the speed control sub-component generated by obstacle c on UAV i during the flight of UAV i. Each speed control sub-component is affected by the flight positions and speeds of neighboring UAVs. Taking as an example, the positions of neighboring UAVs need to be obtained when calculating this sub-component. The closer the two UAVs are, the greater the repulsive effect of UAV j on UAV i. The specific magnitude of the influence is determined by the repulsive force parameter.

[0044] Each control function component contains a certain number of parameters, such as the repulsive coefficient corresponding to the repulsive force. In this embodiment, a certain number of evaluation indicators are set according to the cluster target, such as the speed consistency evaluation indicator and the collision risk evaluation indicator. The parameters are optimized to enable the cluster to achieve better evaluation results, and then the corresponding parameters are applied to the control function and act on all UAVs.

[0045] S12. The UAV obtains its flight state and calculates its flight speed at the next moment: The flight state of the current UAV includes the flight position and speed of the current UAV itself, as well as the flight positions and speeds of all neighboring UAVs within its communication range and the obstacle position information. One flight moment of the UAV corresponds to a time step, and the time step is set as Δt. The flight state is used as the input of the control function to calculate the flight speed of the current UAV at the next moment, so as to avoid collisions with other UAVs and ensure the realization of the UAV cluster.

[0046] The control function here is the speed update rule during the movement of the UAV. In the present invention, the UAV only needs to obtain the flight position and speed information of neighboring UAVs within a certain communication range r c to complete the speed update, without the need to obtain global information.

[0047] Embodiment 3

[0048] The multi-UAV cluster navigation method under communication delay constraints is the same as that in Embodiments 1-2. For the construction of the LSTM model described in step S3 of the present invention, the schematic diagram of the LSTM model framework is as Figure 3 , and the LSTM model is trained with training samples, including the following steps:

[0049] S31. Construct the LSTM model: The LSTM model is a deep learning prediction model. The framework of the LSTM model constructed in the present invention is referred to Figure 3, the model predicts the flight position and flight speed of the UAV. The input data of the LSTM model are the flight state data of the UAV at the current moment and multiple previous moments. The flight state data include the current flight position and flight speed of the UAV itself, as well as the flight positions and flight speeds of all neighboring UAVs within its communication range and the obstacle position information; the output prediction data are the predicted flight positions and flight speeds of the UAV at multiple future moments. The constructed LSTM model uses the Adam optimizer, and the value range of the Adam optimization learning rate is [0.0001, 0.2]. Two hidden layers are set, and the number of neurons in the single-layer LSTM hidden layer ranges from [20, 60]. The input data is transmitted to the first hidden layer, and the hidden layer performs mathematical calculations on the input data and passes it to the next hidden layer. After the last hidden layer is processed, a time series vector is output. This time series vector undergoes the operation and processing of the attention layer to obtain the predicted flight positions and flight speeds of the UAV at multiple future moments as the output.

[0050] S32. Train the LSTM model with training samples: Initialize the weights of the LSTM network model, set the parameter weights of the LSTM network model and the input data to a normal distribution, randomly generate the initial values of the parameters with a normal distribution, import the training sample set to train the LSTM model parameters, and obtain the trained LSTM model.

[0051] Long short-term memory (LSTM) is a powerful deep learning prediction model that can be effectively applied to time series problems. Currently, this method has been applied to many spatio-temporal prediction applications. LSTM is a special recurrent neural network, which is more advanced than ordinary recurrent neural networks. In addition to standard neurons, LSTM also designs special memory units to save the information in the data sequence. Therefore, it can effectively extract the sequence context information and has obvious advantages in dealing with complex time series problems. In the present invention, the LSTM model is introduced, and by inputting the historical flight state data of the UAV at multiple moments, the flight position and flight speed information of the UAV at multiple future moments can be predicted.

[0052] Embodiment 4

[0053] The multi-UAV cluster navigation method under communication delay constraints is the same as that in Embodiments 1-3. The sending and interacting flight information described in step S5 of the present invention includes the following steps:

[0054] S51. Sending information in the initial flight stage: The UAV sends interactive flight information through a wireless communication device. The UAV uses the historical flight state data of consecutive h moments for prediction. When the current moment k < h, the UAV sends its current real flight position and flight speed as interactive flight information to neighboring UAVs.

[0055] S52. Prediction Flight Phase Information Sending: When the current time k≥h, the UAV outputs the predicted flight positions and flight speeds of the UAV at multiple future times according to the trained LSTM model, and takes the real flight position and flight speed at the current time and the predicted flight positions and flight speeds at multiple future times as interactive flight information and sends it to neighboring UAVs.

[0056] The interactive flight information sent by the UAV not only includes its own flight position and flight speed at the current time, but also includes the predicted values of its own flight positions and flight speeds at multiple future times. In this way, when a neighboring UAV does not obtain the flight position and flight speed information of this UAV at the current time, it can read the predicted values of the flight position and flight speed of this UAV at the current time from the historical received information. This method can cope with the possible large or small communication delays in the communication network.

[0057] Embodiment 5

[0058] The multi-UAV cluster navigation method under communication delay constraints is the same as that in Embodiments 1-4. The received information in step S6 of the present invention includes the following steps:

[0059] S61. The UAV receives the interactive flight information sent by neighboring UAVs: The current UAV communicates to obtain the interactive flight information sent by neighboring UAVs in step S5. The current UAV calculates its own flight speed at the next moment through a control function according to the interactive flight information. When calculating, it is necessary to read the flight position and flight speed of the neighboring UAV at the current time.

[0060] S62. The interactive flight information of the neighboring UAV at the current time has been received: The current UAV reads the real flight position and flight speed of the neighboring UAV at the current time for calculating all control function sub-components generated by the corresponding influence of the neighboring UAV.

[0061] S63. The real flight position and flight speed of the neighboring UAV at the current time have not been received: If the interactive flight information of a certain neighboring UAV at the current time has not been received by the current UAV, the current UAV reads the latest predicted flight position and flight speed about the current time in the received interactive flight information of this neighboring UAV for calculating all control function sub-components generated by the corresponding influence of the neighboring UAV. Since the interactive flight information of the neighboring UAV predicts the flight positions and flight speeds at multiple future times, when there is an uncertain communication delay in the communication between UAVs, the current UAV can read the latest obtained predicted flight positions and flight speeds of this neighboring UAV at multiple future times.

[0062] S64. Obtain the flight state by splicing: Splice and fuse the current flight position and flight speed of the UAV itself, the flight positions and flight speeds of all neighboring UAVs within the communication range, and the position information of obstacles to obtain the flight state of the UAV. Use this flight state as the input of the control function to calculate the flight speed of the current UAV at the next moment, that is, the sub-components of the control function corresponding to the repulsive, attractive, and alignment forces generated by all neighboring UAVs within the communication range of the current UAV, the sub-component of the control function corresponding to the self-driving force of the current UAV, and the sub-component of the control function corresponding to the obstacle force are accumulated to obtain the flight speed of the current UAV at the next moment.

[0063] S65. Determine whether the current flight is over: If the UAV receives a flight termination instruction or the current flight time of the cluster has reached the maximum flight time, it indicates that the current cluster flight is over, and step S66 is executed. Otherwise, jump back to step S5 to send interactive flight information.

[0064] S66. End the current cluster movement.

[0065] The multi-UAV cluster navigation method under communication delay constraints of the present invention belongs to the field of multi-agent cooperative control and is mainly applied in UAV distributed cluster algorithms. Due to the communication delay in the communication between UAVs in practical applications, the UAVs cannot obtain the current flight information of surrounding neighboring UAVs in real time at the current moment to accurately plan a collision-free path, resulting in low system stability. The present invention uses an LSTM-based neural network to predict the flight information of UAVs after a certain time delay, and then shares the predicted information with surrounding neighboring UAVs. The neighboring UAVs can plan their own movement paths through the predicted information of this UAV. This method can effectively avoid the cluster flight jitter caused by time delay during the UAV cluster flight and significantly reduce the probability of collision during the UAV cluster flight.

[0066] The following gives a more detailed example to further illustrate the present invention.

[0067] Embodiment 6

[0068] The multi-UAV cluster navigation method under communication delay constraints is the same as that in Embodiments 1-5, and the overall process is shown in Figure 1 . In this implementation case, the UAV cluster problem is simplified. It is assumed that the flight altitudes of the UAVs are the same, and the position information of the UAVs is represented by p = (α, β), where α represents the component of the UAV position on the x-axis, and β represents the component of the UAV position on the y-axis. The flight speed of the UAV is represented by v = (v x , v y ), where v x represents the component of the UAV speed on the x-axis, and v yRepresents the component of the UAV's speed on the y-axis. The UAV refreshes its flight position and flight speed every time step Δt (one time step corresponds to one moment). The sampling timestamp of the UAV at time k is denoted as t k , and the sampling timestamp of the previous moment at time k is t k-1 , and there is Δt = t k-1 - t k .

[0069] S1. Model the motion of UAVs in the cluster: Neighbor UAVs in the cluster refer to other UAVs within the certain sensing range of the current UAV. For the schematic diagram of the UAV cluster environment under the condition of limited communication range, see Figure 2 shown. In the present invention, the flight state of the UAV includes the flight position and flight speed of the current UAV itself, the flight positions and flight speeds of all neighbor UAVs within the communication range, and the position information of obstacles. The flight state is used as the input of the control function to calculate the flight speed of the UAV at the next moment. The control function is a mapping relationship formula from the flight state to the flight speed. In the present invention, it is set that the speed generated by the control function is affected by the autonomous navigation force, the repulsive force generated by neighbor UAVs, the attractive force generated by neighbor UAVs, the alignment force generated by neighbor UAVs, and the repulsive force generated by obstacles; each force corresponds to a component of the control function. The calculation formula of the control function for the speed of UAV i at the next moment in the motion model is:

[0070]

[0071] In this embodiment, in order to simplify the motion process of the UAV cluster, the influence of obstacles on the cluster motion is not considered. Set the flight speed v prefer generated by the UAV under the action of the autonomous navigation force to be 4 m / s. When not affected by other forces, the UAV will fly at a speed of 4 m / s. The direction of the autonomous navigation force is the direction of the current speed v i of the UAV. Affected by the relative distance between neighbor UAV j and the current UAV i, the position of the neighbor UAV needs to be obtained when calculating this component. The closer the distance between the two UAVs, the greater the repulsive effect of UAV j on UAV i. Affected by the speed consistency between neighbor UAV j and the current UAV i, the speed of the neighbor UAV needs to be obtained when calculating this component. The lower the speed consistency between the two UAVs, the greater the alignment effect of UAV j on UAV i. Inputting the positions and speeds of the current UAV i and all its neighbor UAVs into the control function for calculation can obtain the speed of the current UAV i at the next moment At the next moment, the UAV flies at a new speed.

[0072] S2. Process the UAV flight state data to obtain training samples: Divide a Data space in the internal storage space of each UAV control chip to store the UAV flight state data; UAVs in the cluster communicate through wireless communication devices, including sending and receiving the interaction of flight information. Each UAV sends its own flight position and flight speed information to neighboring UAVs and receives the flight position and flight speed information of neighboring UAVs; Collect the UAV flight state data of the cluster flight in the real cluster environment or simulation environment and put it into the Data space.

[0073] The flight position sequence of UAV i at time c and previous times is:

[0074]

[0075] where represents the flight position of UAV i at time k.

[0076] The flight speed sequence of UAV i at time c and previous times is:

[0077]

[0078] where represents the flight speed of UAV i at time k.

[0079] At time c, each UAV reads its own h historical position and speed information, and reads the flight positions and flight speeds of the s UAVs closest in distance within its communication perception range for h consecutive times, and predicts the flight position and flight speed of the current UAV at the next f times.

[0080] The prediction information sequence is:

[0081]

[0082] where represents the flight position and flight speed of UAV i predicted at time c for itself at time c + f.

[0083] Read the flight state data of all UAVs and divide it into an input feature set and an output feature set. Each piece of data saved in the form of a data record in the input feature set is the flight state data of the UAV at the previous h times before the current time, and each piece of data saved in the form of a data record in the output feature set is the flight position and flight speed data of the UAV at the next h times after the current time. One piece of data in the input feature set corresponds to one piece of data in the output feature set. Correspond one by one the data saved in the form of data records in the input feature set and the data saved in the form of data records in the output feature set to form a pair of training samples, and summarize all the data records in the input feature set and the output feature set to obtain the training sample set of the UAV flight state.

[0084] S3. In the setting of model parameters, set the value range of the number of neurons in the hidden layer to [20, 60], the value range of the Adam optimization learning rate to [0.0001, 0.2], and the value range of the Dropout learning rate to [0.2, 0.5]. Input the input data into the first hidden layer of the LSTM model, and the data is processed layer by layer and output to the attention layer. After being processed by the attention layer, the predicted data is output. Subsequently, initialize the weights of the network model, divide the training samples in S2, 80% of the training samples are used for LSTM model training, and 20% of the data is used for test evaluation. The fitness evaluation is the root-mean-square error RMSE (Root-Mean-Square Error) of the model on the test set

[0085]

[0086] Among them, y(i) represents the i-th true value, y′(i) represents the i-th predicted value, and n is the length of the test set label

[0087] Save the LSTM model after training

[0088] S4. UAV takeoff initialization: Save the LSTM model trained in S3 inside all UAV control chips, and initialize the storage space in the UAV internal chip, including the Data space. Each UAV saves historical flight state data and the predicted flight positions and flight speeds at the next f time moments sent by neighboring UAVs in the Data space. The new prediction information will overwrite the old prediction information, and each record of flight position and flight speed has the UAV ID and the corresponding time stamp

[0089] Initialize the time stamp t = 0, initialize the flight position and flight speed for each UAV, and set the UAV flight time step to Δt

[0090] S5. Send interactive flight information: The UAV sends interactive flight information through a wireless communication device. When the current time k < h, the UAV sends the current true flight position and flight speed as interactive flight information to neighboring UAVs. When the current time k ≥ h, the UAV outputs the predicted flight positions and flight speeds of the UAV at the next f time moments according to the LSTM model trained in S3, and sends the current true flight position and flight speed and the predicted flight positions and flight speeds at the next f time moments as interactive flight information to neighboring UAVs

[0091] S6. The drone receives interactive flight information: The drone i obtains the interactive flight information sent by neighboring drones in step S5 through communication at time k. The current drone calculates the speed at the next moment through a control function. When calculating, it is necessary to read the flight position of neighboring drones at the current moment. And the flight speed However, the interactive flight information sent by some neighboring drones may not be received by the current drone on time due to communication delays. If the interactive flight information of a certain neighboring drone at the current time k is not received by the current drone, the current drone reads the latest predicted flight position and flight speed regarding the current time k in the interactive flight information of this neighboring drone that has been received. This predicted data was sent by the neighboring drone before time k; if received, directly read the real flight position of the neighboring drone at the current moment. And the flight speed Combine the flight position and flight speed of the neighboring drone with the flight position And the flight speed of the current drone itself to obtain the flight state, and use the flight state as the input of the control function to calculate the flight speed of the current drone at the next moment.

[0092] When constructing and optimizing the parameters of the LSTM model in the present invention, the flight state data of the drone at multiple historical moments is used as the input, which can improve the prediction accuracy; the interactive flight information sent by the drone not only includes its own flight position and flight speed at the current moment, but also includes the predicted values of its own flight positions and flight speeds at multiple future moments. In this way, when a neighboring drone does not obtain the flight position and flight speed of this drone at the current moment, it can read the predicted value of this drone regarding the current moment from the historical received information, and can cope with the large or small communication delays that may exist in the communication network.

[0093] In summary, the multi-UAV swarm navigation method under communication delay constraints of the present invention belongs to the field of multi-agent cooperative control, and solves the technical problem that it is difficult for a swarm system to achieve stable clustering under the condition that there is communication delay among UAVs in a UAV swarm. It is mainly applied in UAV distributed swarm algorithms. Due to the existence of communication delay in the communication between UAVs in practical applications, the UAVs cannot obtain the flight information of surrounding neighbor UAVs in real time at the current moment to accurately plan a collision-free path, resulting in low system stability. The present invention uses an LSTM-based neural network to predict the flight information of UAVs after a certain time delay, and then shares the predicted information with surrounding neighbor UAVs, and the neighbor UAVs can plan their own motion paths through the predicted information of this UAV. The present invention divides the swarm navigation process into a preparation stage and a flight stage. The preparation stage includes modeling the motion of UAVs in the swarm, processing the flight state data of UAVs to obtain training samples, constructing an LSTM model and training the LSTM model with the training samples. The flight stage includes UAV takeoff initialization, sending and receiving interactive flight information, and finally realizing multi-UAV swarm navigation under communication delay constraints. The present invention has the advantages of small communication and computing overhead and can achieve stable swarm flight of multiple UAVs under communication delay constraints. When applied to UAV swarm navigation, it can avoid the swarm flight jitter caused by time delay during the UAV swarm flight process and reduce the probability of collision during the UAV swarm flight process.

Claims

1. A multi-UAV swarm navigation method under communication delay constraints, characterized in that, The cluster navigation process is divided into a preparation stage and a flight stage, including the following steps: Preparation stage: S1. Model the motion of UAVs in the cluster: In the cluster, neighbor UAVs refer to other UAVs within a certain communication range of the current UAV. The control function is a mapping relationship from the flight state to the flight speed. When each UAV calculates and outputs the flight speed at the next moment through the control function, it only obtains the flight positions and flight speeds of the neighbor UAVs within its communication threshold r c range as inputs. It is assumed that the speed generated by the control function is affected by the autonomous navigation force, the repulsive force generated by neighbor UAVs, the attractive force generated by neighbor UAVs, the alignment force generated by neighbor UAVs, and the repulsive force generated by obstacles; each force corresponds to a control function component. The repulsive, attractive, and alignment forces generated by neighbor UAVs and the repulsive force generated by obstacles are divided into their respective control function sub-components according to different neighbors and obstacles; the control function sub-components are accumulated to obtain the control function component, and the control function is the sum of all control function components; the output of the UAV motion model is the output of the control function. One flight moment of the UAV corresponds to a time step. Let the time step be Δt. The current UAV calculates its own flight speed at the next moment through the control function. Each UAV in the cluster will calculate and output the flight speed corresponding to its own flight state through the control function under different flight states at different moments; S2. Process the UAV flight state data to obtain training samples: Divide a Data space in the internal storage space of each UAV control chip to store the UAV flight state data; each UAV sends its own flight position and flight speed information to neighboring UAVs and receives the flight position and flight speed information of neighboring UAVs; collect the UAV flight state data of the cluster flight in the real cluster environment or simulation environment and put it into the Data space; read all the UAV flight state data and divide it into an input feature set and an output feature set. Each piece of data saved in the form of a data record in the input feature set is the UAV flight state data at the current moment and multiple moments before; Each piece of data saved in the form of a data record in the output feature set is the flight position and flight speed data of the UAV at multiple moments after the current moment; One piece of data in the input feature set corresponds to one piece of data in the output feature set. One-to-one correspondence is established between each piece of data saved in the form of a data record in the input feature set and each piece of data saved in the form of a data record in the output feature set to form a pair of training samples. Summarize all the data records in the input feature set and the output feature set to obtain the training sample set of the UAV flight state; S3. Construct an LSTM model and train the LSTM model with training samples: The input data of the long short-term memory network LSTM model is the UAV flight state data at the current moment and multiple moments before, and the output prediction data is the flight position and flight speed of the UAV at multiple future moments; the constructed LSTM model contains at least two hidden layers and one attention layer, and uses the Adam optimizer; input the input data into the first hidden layer of the LSTM model, and the data is processed layer by layer and output to the attention layer, and the prediction data is output after being processed by the attention layer operation; Initialize the weights of the LSTM model before training, import the training samples of the UAV flight state training sample set to train the LSTM model parameters, and obtain the trained LSTM model; Flight stage S4. UAV takeoff initialization: Apply the trained LSTM model to all UAVs, initialize the storage space in the UAV internal chip, including the Data space. At this time, the timestamp t = 0, initialize the flight position and flight speed for each UAV, and set the UAV flight time step to Δt; S5. Send interactive flight information: The UAV sends interactive flight information through a wireless communication device. In the initial flight stage, the UAV sends the current real flight position and flight speed as the interactive flight information to neighboring UAVs. In the predicted flight stage, the UAV sends the current real flight position and flight speed and the predicted flight position and flight speed at multiple future moments as the interactive flight information to neighboring UAVs. The predicted flight position and flight speed are calculated and output according to the trained LSTM model; S6. Receive interactive flight information: The current UAV obtains the interactive flight information sent by neighboring UAVs through communication. The current UAV calculates its flight speed at the next moment through a control function according to the interactive flight information. When calculating, it is necessary to read the flight position and flight speed of neighboring UAVs at the current moment. However, the interactive flight information sent by some neighboring UAVs may not be received by the current UAV on time due to communication delays. If the interactive flight information of a certain neighboring UAV at the current moment is not received by the current UAV, the current UAV reads the latest predicted flight position and flight speed regarding the current moment in the received interactive flight information of this neighboring UAV, and uses them to calculate the sub-components of the control function generated by all the influences corresponding to this neighboring UAV. If the interactive flight information sent by a neighboring UAV is received, the current UAV directly reads the interactive flight information of this neighboring UAV at the current moment, and uses the flight position and flight speed of this neighboring UAV at the current moment to calculate the sub-components of the control function generated by all the influences corresponding to this neighboring UAV. When calculating the flight speed of the current UAV at the next moment through the control function, the flight position and flight speed of the current UAV itself, the flight positions and flight speeds of all neighboring UAVs within the communication range, and the position information of obstacles are used as the input of the control function to calculate the flight speed of the current UAV at the next moment, so as to achieve multi-UAV cluster navigation under communication delays.

2. The multi-UAV cluster navigation method under communication delay constraints according to claim 1, wherein The motion modeling of UAVs in the cluster described in step S1 includes the following steps: S11. Set the control function: The control function is the sum of all flight speeds generated by the influence of the current UAV's autonomous navigation force, the repulsive force generated by neighboring UAVs, the attractive force generated by neighboring UAVs, the alignment force generated by neighboring UAVs, and the repulsive force generated by obstacles on the current UAV's speed. The calculation formula of the control function regarding the speed of UAV i at the next moment in the motion model is: Among them, represents the speed of the drone i at the next moment calculated; v i represents the speed of the drone i at the current moment; v prefer represents the expected speed value during the flight of the drone, represents the speed control sub-component of the neighbor drone j on the drone i due to the attracting force; represents the speed control sub-component of the neighbor drone j on the drone i due to the aligning force; represents the speed control sub-component of the neighbor drone j on the drone i due to the repulsive force; represents the speed control sub-component of the obstacle c on the drone i during the flight of the drone; S12. The UAV obtains its flight state and calculates its flight speed at the next moment: The flight state of the current UAV includes the flight position and flight speed of the current UAV itself, as well as the flight positions and flight speeds of all neighboring UAVs within its communication range and the position information of obstacles. One flight moment of the UAV corresponds to a time step, and the time step is set as Δt. The flight state is used as the input of the control function to calculate the flight speed of the current UAV at the next moment, so as to avoid collisions between the current UAV and other UAVs and ensure the realization of cluster navigation.

3. The multi-UAV swarm navigation method under communication delay constraints according to claim 1, wherein The steps of constructing the LSTM model and training the LSTM model with training samples described in step S3 include the following steps: S31. Construct an LSTM model: The LSTM model is a deep learning prediction model used to predict the flight position and flight speed of the drone. The input data of the LSTM model is the flight state data of the drone at the current moment and multiple previous moments. The flight state data includes the current flight position and flight speed of the drone itself, as well as the flight positions and flight speeds of all neighboring drones within its communication range and the obstacle position information. The output prediction data is the predicted flight positions and flight speeds of the drone at multiple future moments. The constructed LSTM model uses the Adam optimizer, contains at least two hidden layers, sets the Adam learning rate and the number of neurons in each hidden layer of the LSTM, transmits the input data to the first hidden layer, and the hidden layer performs mathematical calculations on the input data and passes it to the next hidden layer. After the last hidden layer finishes processing, it outputs a time series vector, and this time series vector undergoes operation processing by the attention layer to obtain the predicted flight positions and flight speeds of the drone at multiple future moments for output. S32. Train the LSTM model with the training sample set: Initialize the weights of the LSTM network model, set the parameter weights of the LSTM network model and the input data to a normal distribution, randomly generate the initial values of the parameters using the normal distribution, import the training sample set to train the LSTM model parameters, and obtain the trained LSTM model.

4. The multi-UAV swarm navigation method under communication delay constraints according to claim 1, characterized in that, The steps of sending and interacting flight information described in step S5 include the following steps: S51. Sending information in the initial flight stage: The drone sends interactive flight information through a wireless communication device. The drone uses the historical flight state data of consecutive h moments for prediction. When the current moment k < h, the drone sends its current real flight position and flight speed as interactive flight information to neighboring drones. S52. Sending information in the predicted flight stage: When the current moment k ≥ h, the drone outputs the predicted flight positions and flight speeds of the drone at multiple future moments according to the trained LSTM model, and sends the current real flight position and flight speed and the predicted flight positions and flight speeds at multiple future moments as interactive flight information to neighboring drones.

5. The multi-UAV cluster navigation method under communication delay constraints according to claim 1, characterized in that, The steps of receiving information described in step S6 include the following steps: S61. The drone receives the interactive flight information sent by neighboring drones: The current drone communicates to obtain the interactive flight information sent by neighboring drones. The current drone calculates its flight speed at the next moment through a control function according to the interactive flight information, and the flight position and flight speed of the neighboring drone at the current moment need to be read during the calculation. S62. The interactive flight information of the neighboring drone at the current moment has been received: The current drone reads the current real flight position and flight speed of the neighboring drone for calculating all the control function sub-components generated by the corresponding influence of the neighboring drone. S63. The true flight position and flight speed of the neighboring UAV at the current moment are not received: If the current UAV does not receive the interactive flight information of a certain neighboring UAV at the current moment, the current UAV reads the latest predicted flight position and flight speed regarding the current moment in the received interactive flight information of the neighboring UAV, and uses them to calculate the sub-components of the control function generated by all the influences corresponding to the neighboring UAV; S64. Obtain the flight state by splicing: The flight state of the UAV is obtained by splicing and fusing the flight position and flight speed of the current UAV itself, the flight positions and flight speeds of all neighboring UAVs within the communication range, and the position information of the obstacles. The flight state is used as the input of the control function to calculate the flight speed of the current UAV at the next moment; S65. Determine whether the current flight is over: If the UAV receives a flight termination instruction or the flight time of the current cluster reaches the maximum flight time, it indicates that the current cluster flight is over, and step S66 is executed. Otherwise, it jumps back to step S5 to send the interactive flight information; S66. End the current cluster movement.

Images