Drone Formation Area Search Using Bio-inspired Neural Networks

As a new type of operational force, drones have developed rapidly, and their application scope has expanded from traditional reconnaissance, surveillance, hydro-meteorological information support, and保障 fields to various industries. They play an increasingly important role in tracking, positioning, remote control, telemetry, and digital transmission. For the area coverage search problem of drones, common algorithms include heuristic-based coverage algorithms, pheromone-based multi-drone coverage algorithms, and swarm intelligence algorithms such as Ant Colony Optimization (ACO) and Particle Swarm Optimization (PSO). While these algorithms can improve coverage to some extent, they involve significant computational complexity and, due to the large number of agents, do not fully realize the distributed characteristics of分工合作 within the群体. Therefore, it is essential to define interaction rules and methods among drones to enable the drone formation to quickly and effectively create a comprehensive map.

To address these requirements, I introduce a search model based on the Bio-inspired Neural Network (BINN) to represent environmental state information. During area coverage, the model leverages its excitation and inhibition functions to allow drones to rapidly interact with the environment and the drone formation in unknown areas, thereby creating complete map information. Model Predictive Control (MPC), through formation control, guides the drone formation based on map information to enhance area search capabilities in unknown environments.

In this article, I present a method for drone formation area search that integrates BINN and MPC. The approach aims to improve搜索 efficiency and environmental adaptability by enabling drones to perform cooperative搜索 in unknown scenarios. I will detail the design of the drone formation control method, the environmental model based on BINN, and provide simulation results to validate the effectiveness of the proposed method. Throughout, I emphasize the importance of drone formation in achieving distributed and efficient搜索 tasks.

Design of Drone Formation Control Method

The control of drone formation is critical for coordinated area search. I begin by defining the动力学模型 of individual drones and then apply MPC for formation control.

Kinematic Model of Drones

I assume that each drone is equipped with an autopilot for airspeed, heading, and track channels, comprising 3 translational degrees of freedom and 2 rotational degrees of freedom. This allows me to construct an equivalent second-order system动力学模型 for the drone, as shown in Equation (1).

$$
\begin{aligned}
\dot{x}_U &= V_U \cos \chi_U \cos \gamma_U \\
\dot{y}_U &= V_U \sin \chi_U \cos \gamma_U \\
\dot{z}_U &= V_U \sin \gamma_U \\
\dot{V}_U &= -\frac{1}{\tau_v} V_U + \frac{1}{\tau_v} V_U^c \\
\dot{\chi}_U &= -\frac{1}{\tau_{\chi}} \chi_U + \frac{1}{\tau_{\chi}} \chi_U^c \\
\dot{\gamma}_U &= -\frac{1}{\tau_{\gamma}} \gamma_U + \frac{1}{\tau_{\gamma}} \gamma_U^c
\end{aligned}
$$

Here, \( x_U \) represents the velocity along the x-axis for drone \( U \), \( V_U \) is the horizontal speed, \( \chi_U \) is the heading angle, \( \gamma_U \) is the rate of altitude change, \( y_U \) and \( z_U \) are velocities along the y and z axes, respectively. The derivatives \( \dot{V}_U \), \( \dot{\chi}_U \), and \( \dot{\gamma}_U \) denote changes in horizontal speed, heading angle, and altitude rate, with \( \tau_v \), \( \tau_{\chi} \), and \( \tau_{\gamma} \) as time constants. The control inputs are \( V_U^c \), \( \chi_U^c \), and \( \gamma_U^c \).

I also consider practical constraints for the drone, as given in Equation (2).

$$
\begin{aligned}
V_{\text{min}} &\leq V_U \leq V_{\text{max}} \\
|\dot{\chi}_U| &\leq \frac{n_{\text{max}} g}{V_U} \\
\gamma_{\text{min}} &\leq \gamma_U \leq \gamma_{\text{max}}
\end{aligned}
$$

In this formulation, \( V_{\text{min}} \) and \( V_{\text{max}} \) are the minimum and maximum horizontal speeds, \( n_{\text{max}} \) is the maximum lateral load factor, \( g \) is gravitational acceleration, and \( \gamma_{\text{min}} \) and \( \gamma_{\text{max}} \) are the minimum and maximum altitude change rates, with \( \gamma_{\text{min}} < 0 \).

Drone Formation Control Based on MPC

To control the drone formation, I use MPC to manage position and velocity errors. The position error \( e_c \) for the cluster state is defined in Equation (3).

$$
e_c = c_p – c_d
$$

where \( c_p \) is the current position and \( c_d \) is the desired position.

The velocity error \( e_v \) is given by Equation (4).

$$
e_v = \dot{c}_p – \dot{c}_d
$$

Here, \( \dot{c}_p \) is the current velocity and \( \dot{c}_d \) is the desired velocity.

Combining these, I derive a discrete state-space equation for the errors, as shown in Equation (5).

$$
\begin{bmatrix} e_U(k+1) \\ \dot{e}_U(k+1) \end{bmatrix} = \begin{bmatrix} 1 & \Delta t \\ 0 & 1 \end{bmatrix} \begin{bmatrix} e_U(k) \\ \dot{e}_U(k) \end{bmatrix} + \begin{bmatrix} 0 \\ \Delta t \end{bmatrix} u(k)
$$

In this equation, \( e_U(k) \) is the position error at time \( k \), \( \dot{e}_U(k) \) is the velocity error at time \( k \), \( \Delta t \) is the sampling time, and \( u(k) \) is the input vector.

I express this in standard state-space form, as per Equation (6).

$$
\begin{aligned}
x(k+1) &= A x(k) + B u(k) \\
y(k+1) &= C x(k+1)
\end{aligned}
$$

where \( x(k) \) is the system state variable, \( A \) is the state coefficient matrix, \( B \) is the input coefficient matrix, \( y(k) \) is the system output variable, and \( C \) is the coefficient matrix relating output to state. Since the error is used as the state variable, the output \( y = x \).

For MPC, I solve a Quadratic Programming (QP) problem. The optimization function \( q(z) \) is defined in Equation (7).

$$
q(z) = z^T H z + f^T z
$$

Here, \( q(z) \) is the function to optimize, \( z \) is the optimization variable, \( H \) is a positive definite symmetric matrix that weights state tracking errors, and \( f \) is a matrix related to deviations from the reference trajectory.

The cost function \( J \) minimized in the QP problem is given by Equation (8).

$$
J = J_q + J_r + J_f
$$

with components:

$$
\begin{aligned}
J_q &= \sum_{i=0}^{N-1} x(k+i|k)^T Q x(k+i|k) \\
J_r &= \sum_{i=0}^{N-1} u(k+i|k)^T R u(k+i|k) \\
J_f &= x(k+N|k)^T F x(k+N|k)
\end{aligned}
$$

where \( N \) is the prediction horizon, \( i \) is the step index, \( Q \) is the weighting matrix for errors, \( R \) is the weighting matrix for inputs, and \( F \) is the terminal error matrix. By solving this QP problem to minimize \( J \), I achieve effective control of the drone formation.

The use of MPC in drone formation control allows for real-time adjustments based on预测 models, ensuring that the formation maintains desired trajectories while adapting to environmental changes. This is crucial for area search missions where drones must navigate未知 terrain cooperatively.

Environmental Model Design Based on Bio-inspired Neural Network

As a prerequisite for cooperative target搜索 in drone formation, I model the environment using a grid-based approach inspired by collective behaviors in insects. The core idea is that environmental stimuli prompt individual actions, which leave information that influences other individuals. Over time, the model adjusts activity states in nodes, effectively构建 dynamic environments during搜索 without complex global optimization.

In the decision-making phase, I assume each drone is an intelligent agent with a 360° field of view, maximum turn limit angle \( \theta_{\text{max}} \), minimum turn limit angle \( \theta_{\text{min}} \), and communication within a perception range \( L_{\text{max}} \). I discretize the search area into a two-dimensional grid and construct a 2D BINN model based on this grid. In the model, each grid point represents a biological neuron, as illustrated below.

Each neuron has an associated activity value, and within \( L_{\text{max}} \), neurons stimulate or inhibit others. The BINN model is described by Equation (9).

$$
\frac{dx_{j_1}}{dt} = -A x_{j_1} + (B – x_{j_1}) \left([I_{j_1}^e]^+ + \sum_{i=1}^{n} w_{j_1,j_2} x_{j_2}\right) – (D + x_{j_1}) [I_{j_1}^o]^-
$$

where \( x_{j_1} \) is the activity value of neuron \( j_1 \), \( -A \) is the decay rate of \( x_{j_1} \), \( x_{j_2} \) is the activity value of adjacent neuron \( j_2 \), \( B \) and \( D \) are the upper and lower bounds of \( x_{j_2} \) (i.e., \( x_{j_2} \in [-D, B] \)), \( I_{j_1}^e \) is the excitatory input with \( [I_{j_1}^e]^+ = \max(I_{j_1}^e, 0) \), \( n \) is the number of neurons connected to neuron \( i \), \( w_{j_1,j_2} \) is the weight coefficient between neurons, and \( I_{j_1}^o \) is the inhibitory input with \( [I_{j_1}^o]^- = \max(-I_{j_1}^o, 0) \).

The lateral connection weight between neurons \( j_1 \) and \( j_2 \) in the same layer is computed using Equation (10).

$$
w_{j_1,j_2} =
\begin{cases}
a & \text{if } \| \mathbf{p}_{j_1} – \mathbf{p}_{j_2} \| \leq R_n \\
0 & \text{if } \| \mathbf{p}_{j_1} – \mathbf{p}_{j_2} \| > R_n
\end{cases}
$$

Here, \( \mathbf{p}_{j_1} \) and \( \mathbf{p}_{j_2} \) are position vectors of neurons \( j_1 \) and \( j_2 \), \( R_n \) is the distance threshold between two neurons, and \( a \) is a constant.

In the grid map, each grid point can be in one of three states: unsearched, searched, or obstacle. These states directly affect the neuron activity value and are represented by a state flag, as in Equation (11).

$$
I_i =
\begin{cases}
1 & \text{if } S(G_i) = 1 \text{ (unsearched)} \\
0 & \text{if } S(G_i) = 0 \text{ (searched)} \\
-1 & \text{if } S(G_i) = -1 \text{ (obstacle)}
\end{cases}
$$

where \( I_i \) is the state function for the \( i \)-th grid point, and \( S(G_i) \) denotes the state of grid point \( G_i \).

To model the probability distribution of targets, I define the target existence probability for grid \( (a,b) \) as \( p_{a,b} \in [0,1] \). At time \( k \), the Target Probability Map (TPM) maintained by drone \( U \) is given by Equation (12).

$$
\text{TPM}_U(k) = \{ p_{a,b}(k) \}
$$

where \( \text{TPM}_U(k) \) changes with time, and \( p_{a,b}(k) \) represents the probability in the mission area.

After receiving a mission, drones take time to reach the area, during which the target probability distribution动态 changes. Based on prior information, I classify target position distributions into two types.

If the target position is unknown, it is assumed to be uniformly distributed across the mission area. The probability density function is given by Equation (13).

$$
f(x,y) =
\begin{cases}
\frac{1}{N_x N_y} & \text{if } (x,y) \in D \\
0 & \text{if } (x,y) \notin D
\end{cases}
$$

Here, \( f(x,y) \) is the probability density of a target at location \( (x,y) \) in the mission area, \( N_x \) and \( N_y \) are the total number of grid cells in the x and y directions, and \( D \) is the mission region.

If the initial target position is known but its velocity is unknown, let the prior information indicate an initial position \( (x^*, y^*) \). The actual target position \( (x_0, y_0) \) is a random variable, with \( x_0 \) and \( y_0 \) independent and following two-dimensional normal distributions \( N(x^*, \sigma_0^2) \) and \( N(y^*, \sigma_0^2) \), respectively. The probability density function is given by Equation (14).

$$
f(x_0, y_0) = \frac{1}{2\pi \sigma_0^2} \exp\left( -\frac{(x_0 – x^*)^2 + (y_0 – y^*)^2}{2\sigma_0^2} \right)
$$

where \( \sigma_0 \) is the standard deviation of the normal distribution.

This BINN-based environmental model enables the drone formation to adaptively perceive and map unknown areas, facilitating efficient搜索 through distributed coordination.

Simulation and Analysis

To validate the proposed method, I conduct simulations with multiple drone formations performing cooperative搜索. First, I construct a two-dimensional grid environment model and, based on initial target prior information, build static and dynamic target probability distribution models. Then, drones obtain external environmental stimuli by detecting obstacles, searched and unsearched areas, and adjust their thresholds through cluster density, obstacle avoidance, and搜索 state information. Next, drones determine decisions that maximize收益 functions based on combined information. Finally, I update and feedback decision model strategies according to the results of cooperative搜索 for static and dynamic targets until all搜索 tasks are completed. The parameters used in the simulation are listed in Table 1, and the drone formation parameters are in Table 2.

Table 1: Initial Positions and Obstacle Parameters for Drone协同搜索
Name Description Coordinates (x, y, h/r) [m]
Drone Formation 1 Drones: 1(2.5,2.0,3.0); 2(2,3,4); 3(2,2,5); 4(2.2,2.3,4.0) As listed
Drone Formation 2 Drones: 1(95,3,4); 2(95.7,4.0,3.5); 3(95.5,4.5,4.5); 4(95.7,3.8,4.0) As listed
Drone Formation 3 Drones: 1(94.0,96.5,4.2); 2(93,96,4); 3(95.5,95.0,4.5); 4(94.2,95.8,4.2) As listed
Drone Formation 4 Drones: 1(3.5,98.0,3.6); 2(3.0,97.0,4.7); 3(2.0,97.6,5.4); 4(2.8,97.5,4.6) As listed
Search Area Obstacles Circular obstacles: 1(45,25,6); 2(70,60,10); 3(30,10,5); 4(80,40,5); 5(40,80,6); 6(42,60,7); Rectangular obstacle: (65,20, width:6, length:14, height:8) As listed, with h for height and r for radius
Table 2: Parameters for Drone Formation Control
Parameter Description Value/Unit
τ_v Time constant for horizontal speed 0.01
τ_χ Time constant for heading angle 0.01
τ_λ Time constant for altitude change 0.01
V_max Maximum speed 20 m/s
V_min Minimum speed 5 m/s
n_max Maximum lateral load factor 1
n_min Minimum lateral load factor -1
λ_max Maximum flight path angle limit π/18 rad
λ_min Minimum flight path angle limit -π/18 rad
t Sampling time 0.1 s
l Constraint interval 150

In the simulation, I use four drone formations, each consisting of four drones, to search for four static targets and three dynamic targets in the area. Static target搜索 is based on the uniform distribution model, while dynamic target搜索 uses a two-dimensional normal distribution probability density function modeled as a Wiener random process. The trajectories of the drone formation during area search for static and dynamic targets are depicted in the results, and the average coverage rate is analyzed.

From the simulation results, I observe that in a scenario including both static and dynamic targets, with a run time of 2500 seconds, the proposed method successfully locates all targets. In contrast, methods without the BINN model only find five targets. Additionally, the flight trajectories indicate that non-BINN methods have lower搜索 coverage, demonstrating the superiority of the proposed approach. The average coverage rate over time is shown in comparative plots, where the BINN-based method consistently achieves higher coverage, especially as time progresses.

These findings highlight the effectiveness of integrating BINN with MPC for drone formation area search. The BINN model enhances environmental perception and adaptability, allowing the drone formation to efficiently navigate and cover未知 areas while maintaining coordinated control through MPC.

Conclusion

In this article, I have proposed a drone formation area search method based on a Bio-inspired Neural Network model. Under conditions of unknown targets and environmental information, drones use the BINN model for environmental perception and MPC for formation control to perform area搜索. Simulations in environments with both static and dynamic targets show that the BINN mechanism significantly improves搜索 efficiency and environmental adaptability compared to non-BINN approaches.

The integration of BINN and MPC enables the drone formation to achieve distributed cooperation, rapid interaction, and effective mapping of unknown regions. This method holds promise for various applications, including search and rescue, surveillance, and environmental monitoring, where coordinated drone formations are essential. Future work may explore extensions to three-dimensional environments, real-time implementation challenges, and integration with other swarm intelligence algorithms for enhanced performance.

Overall, the drone formation approach presented here demonstrates the potential of bio-inspired models in advancing autonomous systems for complex搜索 tasks. By leveraging neural network dynamics and predictive control, drones can operate collaboratively in dynamic and uncertain settings, paving the way for more intelligent and adaptive robotic systems.

Scroll to Top