An unmanned aerial vehicle cluster communication network node deployment method based on inter-sensory integration

Abstract

The application discloses a kind of unmanned aerial vehicle cluster communication network node deployment methods based on the integration of mutual perception, belong to unmanned aerial vehicle cluster communication field;It will use Interacting Multiple Model (IMM) to improve Extended Kalman Filter (EKF), and introduce weight parameter in the prediction of different measurement error positions.Finally, the sequence of NARX prediction guides EKF input, and the position of unmanned aerial vehicle is adjusted according to the direction of ground crowd movement to realize unmanned aerial vehicle tracking.The application can realize the rapid deployment of unmanned aerial vehicle cluster communication network topology under the condition of communication burst interruption, and the unmanned aerial vehicle cluster obtains the position perception information of demand distribution by using past communication signals, and the NARX network of multi-step prediction enhances data correlation and improves the prediction performance of EKF.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) swarm communication, and more particularly to a method for deploying UAV swarm communication network nodes based on integrated sensing and communication. Background Technology

[0002] Sensor-integrated technology refers to new applications and services that enhance perception in new usage scenarios. It provides multi-dimensional, wide-area perception services, collecting spatial information about unknown objects and the surrounding environment of connected devices. Its implementation takes various forms, but based on its principles, it can be categorized into three types: communication-assisted perception, perception-assisted communication, and integrated communication-perception design. Combining the deployment of UAV swarm communication networks with sensor-integrated technology is expected to improve the deployment capabilities of swarm communication networks.

[0003] Traditional positioning and deployment methods include geometric positioning and deployment based on heuristic positioning algorithms. Geometric positioning and deployment methods use physical quantities such as angle of arrival and time difference of arrival in polar or Euler coordinate systems to solve for position information and complete the deployment task. Heuristic positioning and deployment methods treat the UAV as a point mass and deploy it within space. However, both of these methods may encounter local optima or have difficulty converging.

[0004] In engineering practice, using integrated sensing and communication design requires updating communication or sensing equipment and redesigning the integrated communication and sensing waveform, which is quite difficult to implement. However, through experiments and theoretical analysis, it can be seen that the integrated sensing and communication technology has the characteristics of low latency and high precision, which can better meet the high dynamic real-time deployment requirements of UAV swarms. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention implements a method for deploying nodes in a drone swarm communication network based on integrated sensing and communication. When the number of communication nodes carried by a drone increases or decreases, the drone swarm can autonomously predict the distribution of ground demand based on historical data. Based on this prediction, the drone swarm can achieve uneven coverage and more quickly reconstruct the communication network after changes in communication nodes, enabling real-time, efficient deployment of drone communication nodes according to ground demand.

[0006] Based on the above objectives, the technical solution provided by the present invention is as follows:

[0007] A method for deploying nodes in a drone swarm communication network based on integrated sensing and communication includes the following steps:

[0008] Step 1: The drone swarm is evenly deployed to the communication target area. N drones are evenly distributed in the airspace above the area with insufficient communication capacity, where N is the total number of drones in the drone swarm.

[0009] Step 2: Obtain the location information of user nodes that need to connect to communication base stations within the target area within 24 hours through a drone swarm;

[0010] Step 3: Store the collected dataset in a Data table, where the user node location information includes longitude, latitude, and time;

[0011] Step 4: Calculate the model of communication demand distribution over time based on the location information of user nodes. Train the Data dataset using a regression model with a nonlinear external input and a self-attention mechanism (NARX). The training result is Questtimenet. The trained model can predict the ground demand distribution over time.

[0012] Step 5: Determine whether the current drone has insufficient communication capacity by using the communication capacity threshold set by the drone; if the communication capacity is sufficient, proceed to steps 3 to 5; if the communication capacity is insufficient, proceed to step 6.

[0013] Step 6: Based on the failure time of the UAV swarm communication network and Questtimenet, predict the location distribution of ground communication needs;

[0014] Step 7: Based on the location of ground demand, use EKF to fuse the UAV location at the next moment with the ground demand distribution location data to predict the UAV location at the next moment. Based on the uneven distribution of demand, calculate the non-uniform distribution of the UAV swarm communication topology.

[0015] Furthermore, ground demand can vary by region and month. The NARX regression model with nonlinear external input can still update the dataset (Data1.1) after the Questtimenet results are output.

[0016] Furthermore, based on the Questtimenet model, the distribution of ground communication demand when communication nodes fail is calculated, or the density of ground communication access points is calculated. Based on the density, the distribution of ground communication demand is calculated, and based on the demand, the distribution of the drone swarm is calculated.

[0017] Furthermore, in step 4, the training results Questtimenet are stored in each UAV to predict the distribution of ground demand based on time; and the distribution density of ground demand nodes is calculated based on the predicted ground node locations.

[0018] Furthermore, in step 6, the drone swarm inputs the command time point into the Quest timenet network. Quest timenet calculates the distribution density of ground communication devices, determines the number of drones to be deployed based on the density information, and divides the ground into square areas. The square areas are then graded according to the drone density, with each density level corresponding to a group of drones. Based on the distribution density of ground communication nodes, the drones calculate the position of each node when the drone swarm is unevenly distributed, and by default, the drone positions are randomly selected within the square areas.

[0019] Furthermore, in step 4, the regression model training dataset of the self-attention mechanism NARX nonlinear external input is trained with Questtimenet, the neural network has a latency of 5 steps and 10 layers, the positions of the moving nodes on the ground are recorded, the node distribution density is calculated, and the distribution density is used to simulate the ground demand.

[0020] Compared with the prior art, the advantages of the present invention are:

[0021] 1. In the deployment of drone swarm communication networks, energy consumption is a major limiting factor for the swarm's operating time. However, to achieve a network deployment that matches ground requirements, more ground information is needed, and drones need to carry more sensors, which will ultimately lead to greater energy consumption and shorten the operating time of the swarm communication network. However, the integrated communication and sensing technology can utilize the location information in the existing communication data packets to help improve the drone's location perception capabilities and avoid further energy consumption.

[0022] 2. This invention can address situations such as insufficient communication traffic supply and communication gaps in more unexpected scenarios. In further application scenarios, the following situations may occur:

[0023] The sudden damage to ground communication base stations has created a large-scale communication gap, necessitating the rapid redeployment of the drone swarm communication network.

[0024] Sudden crowd gatherings can lead to an exponential increase in network communication demand at any given moment, necessitating the replenishment of bandwidth and traffic for drone swarm communication networks within a short period.

[0025] The drone swarm stopped flying and crashed due to malfunctions, insufficient power, or other reasons, causing the network to malfunction.

[0026] Thirdly, the calculation of the UAV swarm communication network topology mainly includes two parts: node position prediction and multi-node connection. This invention modifies the state prediction function of the Extended Kalman Filter (EKF) through an autoregressive network with a nonlinear external input of the attention mechanism NARX, thereby improving the ability of UAV swarms to select the optimal location during deployment.

[0027] Fourthly, the self-attention mechanism in NARX enhances the capture of long-term dependencies: Self-attention allows the model to focus on the correlations between different time steps when processing sequential data, which helps capture long-term dependencies in the sequence. Introducing self-attention into the NARX model allows it to better understand the connections between historical data and more accurately predict future values, and it also enables parallel computation: the computation of self-attention can be performed in parallel, making the model more computationally efficient during training and inference. Since NARX models are typically applied to tasks with long input sequences, such as time series analysis, the parallel computational capability of self-attention is highly beneficial for accelerating model training and prediction. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating a method for deploying nodes in a drone swarm communication network based on integrated sensing and communication, according to an embodiment of the present invention.

[0029] Figure 2 Conceptual diagram for deploying drone networks;

[0030] Figure 3 Flowchart of the improved EKF algorithm (EKF algorithm flowchart);

[0031] Figure 4 This is a schematic diagram of the NARX-IEKF-ISAC principle. Detailed Implementation

[0032] In the following description, well-known functions or structures will not be described in detail to avoid unnecessary confusion with the content of this invention.

[0033] Drone swarms assist in emergencies by extending communication range, acting as relay stations for real-time communication. However, frequent communication link interruptions during network construction due to drone and user movement pose challenges to real-time reconstruction. Furthermore, the complex communication environment of drone swarms leads to differences in observation errors among individual drones. This parameter significantly impacts the accuracy of drone position prediction during network deployment and affects link reconstruction. To address the communication link reconstruction problem in dynamic drone swarm deployments, this patent proposes an integrated NARX-IEKF-ISAC algorithm. This algorithm introduces the NARX neural network into an improved EKF localization algorithm and uses communication data for training to improve drone localization accuracy. First, this paper uses the feedforward neural network NARX to train the crowd position data in the communication dataset. Second, the Interacting Multiple Model (IMM) is used to improve the Extended Kalman Filter (EKF), introducing weight parameters in the prediction of positions with different measurement errors. Finally, the sequence predicted by NARX guides the EKF input, and the direction of ground crowd movement adjusts the drone position to achieve drone tracking. This method enables rapid deployment of a drone swarm communication network topology in the event of sudden communication interruptions. The drone swarm acquires demand distribution location perception information by using past communication signals, and the NARX network enhances data correlation through multi-step prediction, thereby improving EKF prediction performance. This patent empowers drone swarm communication with communication-assisted perception integrated sensing technology, enabling swarm tasks with on-demand network deployment and improving swarm collaboration performance.

[0034] Reference Figures 1 to 4 A method for deploying nodes in a drone swarm communication network based on integrated sensing, comprising:

[0035] S1: The drone swarm is evenly deployed in the target communication area. N drones are evenly distributed in the airspace above the area with insufficient communication capacity, where N is the total number of drones in the drone swarm. The drones are the drones in the drone swarm.

[0036] S2: Obtain the location information of user nodes that need to connect to the communication base station within the target area within 24 hours, and store the collected dataset in a Data table, which includes longitude, latitude, and time, where the time format can be restricted to "2023 / 03 / 03 00:00:00".

[0037] S3: Calculate a model of demand distribution over time based on the location information of nodes requiring communication. Train the Data dataset using a regression model with a nonlinear external input and a self-attention mechanism (NARX). The training result is Questtimenet. The trained model can predict the ground demand distribution over time.

[0038] S4: Determine if the number of communication nodes has changed. Due to faults, energy depletion, or other reasons, the structure of the drone swarm communication network needs to be recalculated.

[0039] S5: Based on the failure time of the UAV swarm communication network and Questtimenet, predict the location distribution of ground communication needs.

[0040] S6: Based on the location of ground demand, use EKF to fuse the drone's location at the next moment with the ground demand distribution data to predict the drone's location at the next moment. Based on the uneven distribution of demand, calculate the non-uniform distribution of the drone swarm communication topology.

[0041] S1 includes:

[0042] The drone swarm receives coverage area information. This invention emphasizes decentralized processing within the drone swarm, disregards issues such as decision-making centers, and assumes that the number of drones is sufficient to achieve full coverage of the communication area.

[0043] Based on the coverage area information received from the ground, the drones are evenly deployed over the area.

[0044] S2 includes:

[0045] The location information of user nodes that need to connect to communication base stations within the target area is obtained through mobile communication data packets within 24 hours.

[0046] The location of terrestrial mobile communication devices is stored in an information table to prepare for improving the accuracy of location awareness.

[0047] The information table:

[0048] The information table is based on a 24-hour cycle and can be accumulated. To prevent data table overflow and excessive data storage pressure on the drone cluster, the maximum recording duration for this method is 356 days. The data storage fields have been given above. The time field needs to be in a fixed format for the convenience of S3 training, as clearly stated above.

[0049] S3 includes:

[0050] The self-attention mechanism NARX can train the stored data using an externally input nonlinear autoregressive network; the training results, Questtimenet, are stored in each UAV, and these results can predict the distribution of ground demand over time; based on the predicted ground node locations, the distribution density of ground demand nodes can be calculated.

[0051] S4 includes:

[0052] Determine whether the drone swarm communication network needs to be recalculated; if so: the communication network needs to be deployed or re-deployed due to natural disasters, improper human operation, communication demand overload, insufficient drone energy, or other reasons.

[0053] or;

[0054] If not, the UAV will remain in standby mode to maintain its control configuration and only collect location information from ground mobile communication devices.

[0055] S5 includes:

[0056] The drone swarm inputs the command time point into the Questtimenet network. Questtimenet calculates the distribution density of ground communication equipment, determines the number of drones to be deployed based on the density information, and divides the ground into square areas. The square areas are then classified according to the drone density, with each density level corresponding to a group of drones.

[0057] The drone calculates the location of each node when the drone swarm is unevenly distributed based on the distribution density of ground communication nodes, and by default, the drone location can be randomly selected within a square area.

[0058] The invention emphasizes that it equates the concept of ground demand size with the size of the number of ground mobile demand nodes, rather than directly measuring communication traffic.

[0059] S6 includes:

[0060] Broadcast the position of each drone i calculated in S5 to the drone cluster;

[0061] The communication nodes use the shortest path method to connect multiple nodes to form a communication network, enabling the uneven distribution of UAVs according to ground communication needs.

[0062] The following is a more specific example:

[0063] Table 1 Communication Location Data Table

[0064] 2009 / 2 / 10 9:33 39.966145 116.314737 2011 / 9 / 21 12:31 39.97469 116.3162 2008 / 8 / 3 8:32 39.7947949 116.4857149 2008 / 7 / 23 0:49 39.934474 116.430084 2009 / 3 / 7 7:29 40.350871 116.020288 2010 / 3 / 19 5:43 43.886274 125.311187 2012 / 2 / 16 1:48 39.981325 116.3281533

[0065] In the current deployment of UAV swarm communication networks, improving the correlation between node locations and ground requirements is crucial for enhancing network deployment capabilities. A UAV swarm communication network refers to a group of collaboratively working UAVs achieving information transmission and task coordination through communication connections. To ensure efficient network operation and successful mission completion, the following aspects need to be strengthened:

[0066] The first part discusses the critical importance of accurate drone positioning for network deployment. Drones need precise positioning capabilities that meet application requirements to accurately send and receive information. While advanced positioning technologies such as GPS or INS can provide high-precision location data, they cannot reflect the connection between ground communication needs and network deployment. Therefore, enhanced awareness of ground needs is necessary. Drones can better perceive their own location and understand the ground conditions, thus participating more effectively in communication networks.

[0067] Secondly, a reliable communication link needs to be established between ground requirements and the location of the drones. Ground requirements refer to the instructions and information that the drone swarm needs to execute tasks, such as flight path planning and target designation. To ensure that the drones accurately receive and understand ground requirements, a stable and reliable communication link must be established. The communication coverage area of ​​drones is beam-shaped, and the accuracy and effectiveness of the communication coverage area need to be improved according to the distribution of ground requirements.

[0068] Furthermore, communication between drone locations and ground requirements should be flexible and adaptive. Drone swarms typically operate in dynamic environments, with locations and mission requirements potentially changing at any time. Therefore, the communication network needs to be capable of automated adjustment and adaptation to meet the needs of different scenarios. Employing an intelligent network management and scheduling system can monitor drone locations and ground requirements in real time, enabling dynamic node positioning and network deployment, thereby improving the performance and efficiency of the network topology.

[0069] In conclusion, strengthening the relationship between UAV location and ground requirements is crucial for improving the performance of UAV swarm communication network deployment. By ensuring accurate UAV location, establishing reliable communication links, and implementing flexible and adaptive communication mechanisms, network deployment performance can be effectively improved, enabling UAV swarm communication networks to better meet various mission requirements.

[0070] The ground communication demand model is obtained by training on NARX. The NARX model is embedded into the extended Kalman filter to realize the communication positioning data to assist in improving the perception positioning data, thereby improving the deployment capability of UAV swarm network. In order to further optimize the prediction capability of communication demand, a self-attention mechanism is introduced into the neural network. This attention mechanism can enhance the correlation between input data and provide a stronger guarantee for the multi-step prediction performance of NARX neural network.

[0071] By combining the self-attention mechanism NARX training model with the improved extended Kalman filter algorithm, UAVs can use location perception information from 5G and other mobile communications to assist in the deployment of communication networks, realizing communication-assisted perception in the integration of sensing and communication, and understanding the working principle of perception-assisted communication.

[0072] The principle behind the integration of sensing and sensing technologies:

[0073] The Extended Kalman Filter (EKF) integrates sensing and communication location information to achieve unified communication and sensing. For the problem of UAV swarm flight position prediction, the EKF algorithm uses observation column vectors and state vectors to predict the position, where S(k) contains three-dimensional coordinates v. x (k), v y (k), v z (k), velocity v in the orthogonal direction x (k), v y (k), v z (k), orthogonal acceleration a x (k), a y (k), a z The L(k) and M(K) movement state vectors contain the movement distance of L(k) and observation information with θ(k) as the direction.

[0074] S(k)=[x(k),y(k),z(k),v x (k), v y (k), v z (k), a x (k), a y (k), a z (k)] T (1)

[0075] M(k) = [L(k), θ(k)] T (2)

[0076] The calculation model also considers the error models (3)(4) of the state vector S(K) and observation vector M(K), where the error values ​​of S(k) and M(k) follow a Gaussian distribution. and The covariance matrix of the error values

[0077]

[0078]

[0079] The position of the UAV at any given time can be determined based on the state vector S(k) and the observation matrix M(k).

[0080]

[0081] Where (x0(k), y0(k), z0(k)) is the initial position of the UAV;

[0082] The Extended Kalman Filter (EKF) mainly consists of the following four steps:

[0083] First, the state vector and position vector each predict the information at time k based on the information at time k-1.

[0084]

[0085] Where f(·) represents the ground crowd movement trajectory function, and h(·) represents the UAV inertial measurement prediction function.

[0086] The second step is to calculate the state transition probability matrix using formula (17).

[0087]

[0088] The state matrix and the measurement matrix are correlated in the motion model. This correlation is called the filter gain and can be expressed as (8).

[0089]

[0090] The third step is to update the predicted values ​​of the state matrix and the measurement matrix based on the filter gain as follows (9).

[0091]

[0092] The algorithm flow of EKF is as described above. However, when the UAV swarm is networked, the position needs to be updated in real time according to the needs of the ground, and the communication needs of the ground users need to be used as the state transition function to improve the deployment capability of the UAV swarm in highly dynamic scenarios. Therefore, the NARX time series nonlinear prediction model is used to train the state transition model based on the relevant time series dataset.

[0093] State transition model - self-attention mechanism NARX nonlinear autoregressive model with external input trained on the GEOLIA Etrajectory 3.0 dataset to simulate the changing state of ground node communication requirements over time:

[0094] y(t)=f(y(t-1), y(t-2),..., y(t-ny), u(t-1), u(t-2),..., u(t-nu)) (10)

[0095] Modeling the input sequence X and the external input sequence U using a self-attention mechanism:

[0096] Z X =SelfAttention(X) (11)

[0097] Z U =SelfAttention(U) (12)

[0098] Applying a self-attention mechanism to the input sequence X and the external input sequence U respectively yields their attention representations Z.X and Z U Self-attention mechanisms can enhance the model's ability to learn important correlations between the input sequence and the external input sequence.

[0099] Represent attention as Z X and Z U Perform a connection and use it as the extended input sequence Z:

[0100] Z = Concatenate(Z) X Z U (13)

[0101] Z X and Z U By concatenating these sequences, we obtain the extended input sequence Z. This allows the model to comprehensively consider the correlation between the input sequence X and the external input sequence U, thus enabling better predictions for multiple future time steps.

[0102] Multi-step prediction using the expanded sequence Z:

[0103] Y h at = NARX M ultiStep p rediction(Z) (14)

[0104] The extended sequence Z is used in the NARX model for multi-step prediction tasks. The NARX model can be a multi-layered nonlinear model to better capture the complex relationships of the input sequence. According to formula (10), the predicted value is predicted through feedback from several past states, which improves the accuracy of the predicted state. y(t) represents the value of the target variable at time t, and u(t) represents the value of the external input variable at time t. f(·) is a nonlinear function that represents the relationship between the target variable and the external input variable.

[0105] In NARX, the self-attention mechanism aims to increase the temporal dependencies in the input sequence, thereby achieving better temporal predictability. Based on the dot product of the query and label values ​​calculated by the attention mechanism, a score is calculated and normalized to highlight the largest value in the matrix while suppressing other components much lower than the maximum value. This process can be implemented using the softmax function. The corresponding mathematical formula for this process is:

[0106]

[0107] d k This represents the dimension of the word vector. We divide by... This is to prevent QK T If the value is too large, it will cause overflow during softmax calculation.

[0108] The training result, Quest time net, will replace the S(k) matrix in the conventional EKF. The introduction of this prediction model can combine ground communication needs when the deployment location of UAV nodes is determined. In the results of the actual implementation, it can be shown that dense deployment can be achieved in the stage of dense communication demand, and the location of communication demand coefficient can be sparse, thereby improving the quality of service (Qos) of ground users.

Claims

1. A method for deploying nodes in a drone swarm communication network based on integrated sensing, characterized in that, Specifically, the following steps are included: Step 1: The drone swarm is evenly deployed to the communication target area. N drones are evenly distributed in the airspace above the area with insufficient communication capacity, where N is the total number of drones in the drone swarm. The drones are the drones in the drone swarm. Step 2: Obtain the location information of user nodes that need to connect to communication base stations within the target area within 24 hours through a drone swarm; Step 3: Store the collected dataset in the Data table. The user node location information includes longitude, latitude, and time. Step 4: Calculate the model of communication demand distribution over time based on the location information of user nodes. Use the self-attention mechanism NARX nonlinear external input autoregressive model to train the Data dataset and obtain the training result Quest time net. The trained model can predict the ground demand distribution over time. Step 5: Determine whether the current drone has insufficient communication capacity by using the communication capacity threshold set by the drone; if the communication capacity is sufficient, proceed to steps 3 to 5; if the communication capacity is insufficient, proceed to step 6. Step 6: Based on the drone swarm communication network failure time and Quest time net, predict the location distribution of ground communication needs; Step 7: Based on the location of ground demand, use EKF to fuse the drone location at the next moment with the ground demand distribution location data to predict the drone location at the next moment. Based on the uneven distribution of demand, calculate the non-uniform distribution of the drone swarm communication topology.

2. The method for deploying nodes in a UAV swarm communication network based on sensor integration according to claim 1, characterized in that, Ground demand can vary by region and month. The NARX autoregressive model with nonlinear external input can still update the dataset Data1.1 and the data model Questtime net after the Quest time net results are output.

3. The method for deploying nodes in a UAV swarm communication network based on integrated sensing as described in claim 1, characterized in that, Based on the Quest Time Net model, the distribution of ground communication demand when a communication node fails is calculated, or the density of ground communication access points is calculated. Based on the density, the distribution of ground communication demand is calculated, and based on the demand, the distribution of the drone swarm is calculated.

4. The method for deploying nodes in a UAV swarm communication network based on integrated sensing as described in claim 1, characterized in that, In step 4, the training result Quest time net is stored in each UAV to predict the distribution of ground demand based on time; and the distribution density of ground demand nodes is calculated based on the predicted ground node locations.

5. The method for deploying nodes in a drone swarm communication network based on integrated sensing as described in claim 1, characterized in that, In step 6, the drone swarm inputs the command time point into the Quest Timenet network. Quest Timenet calculates the distribution density of ground communication devices, determines the number of drones to be deployed based on the density information, and divides the ground into square areas. The square areas are graded according to the drone density, and each density level corresponds to a group of drones. The drones calculate the position of each node when the drone swarm is unevenly distributed based on the distribution density of ground communication nodes, and randomly select drone positions within the square areas by default.

6. The method for deploying nodes in a drone swarm communication network based on sensor integration according to claim 1, characterized in that, In step 4, the self-attention mechanism NARX nonlinear external input autoregressive model training dataset is used to train the Quest time net, the neural network has a latency of 5 steps and 10 layers, the positions of the moving nodes on the ground are recorded, the node distribution density is calculated, and the distribution density is used to simulate the ground demand.

Images