Control allocation algorithms for quadrotor drones that convert controller commands into rotor thrusts have trouble managing minor but computationally complex perturbations such as ‘blade flapping’. A neural network approach, however, is set to overcome the challenge.

Control allocation algorithms for quadrotor unmanned aerial vehicles (UAVs) avoid incorporating relatively minor aerodynamic effects such as ‘blade flapping’ and induced drag in order to simplify the computation. But this simplification can sometimes mean that commands sent by a controller are not always properly generated by the device. A new control allocation method —the method of ensuring a command from a user and is implemented by the device with a tight fidelity to the user’s intent— making use of a neural network could make such errors a thing of the past.

Paper Information

S. P. Madruga, A. H. B. M. Tavares, S. O. D. Luiz, T. P. Nascimento, and A. M. N. Lima, “Aerodynamic effects compensation on multi-rotor UAVs based on a neural network control allocation approach,” IEEE/CAA J. Autom. Sinica, vol. 9, no. 2, pp. 295–312, Feb. 2022.

https://www.ieee-jas.net/en/article/doi/10.1109/JAS.2021.1004266

The use of UAVs, or drones, has enjoyed substantial growth over the last decade across a range of applications, from package delivery to military deployment. Alongside this growth, researchers have increasingly focussed on improved control techniques for these vehicles, in particular for quadrotors—a type of helicopter with four rotor blades (rotors).

Control allocation algorithms take the control input or command from a user and ensure that the command is implemented by the quadrotor UAV’s actuators and effectors. The former is the component of the device that is responsible for controlling the mechanism that performs a movement, while the latter is the part of the device that interacts with the environment to carry that movement out. In other words, a control allocation algorithm—sometimes referred to as a motor mixing matrix, control mixer or blender—converts controller commands into individual rotors’ thrusts.

While this may seem a straightforward task, quadrotor UAVs are considered classic unstable dynamical systems. In physics, dynamical systems are those where the state of the system’s ensemble of constituent elements varies over time, and so are governed by differential equations. To predict how the system will behave in the future, these equations must be solved, and a whole field of mathematics, called stability theory, is devoted to the “stability” of such solutions. Stability in this case refers to whether or how the trajectories of the dynamical systems are maintained in the face of tiny changes or perturbations to initial conditions.

There are several aerodynamic and gyroscopic effects that produce such minor perturbations in quadrotors, in particular ‘blade flapping’ and induced drag, that are very difficult to incorporate into a simplified control allocation model. And quadrotors are particularly susceptible to such aerodynamics effects due to their small size.

“Indeed, despite how these aerodynamic effects complicate the natural stability of quadrotors, in most control allocation efforts, such effects are simply ignored,” said Tiago do Nascimento of the Department of Computer Systems at Brazil’s Universidade Federal da Paraiba and lead author of the paper.

As a result of neglecting such effects, the torque and thrust commands sent by a controller are not always properly generated by the actuators because the motors cannot achieve the necessary rotation speed.

“In effect, an error is deliberately inserted in the control allocation by the disregard of aerodynamic effects because to do otherwise would be too computationally challenging,” Prof. Do Nascimento added.

So the researchers devised a new control allocation method making use of a “function fitting” neural network that would be able to consider these aerodynamic effects. Neural networks are a form of machine learning in which a computer learns to carry out a piece of work by being trained on examples. Function fitting is the process of training a neural network on a set of inputs with the aim of producing a set of target outputs.

This method allowed the researchers to improve controller performance without the need for directly using aerodynamic effects equations in the control algorithm.

Upon experimenting with quadrotor UAVs, the researchers found that their novel control allocation approach did indeed improve the performance of quadrotor controllers.

They compared the mechanical torques commanded by the flight controller to those actually generated by the actuators in the UAV. They found that that the neural network was able to closely match the controller’s input with the actual result, while the classic allocation matrix still produced a gap between command and output.

The researchers also found the neural network approach improved the quality of the thrust and torque signals, delivering much less “noisy” behavior.