Flying in air ducts

Abstract

Air ducts are integral to modern buildings but are challenging to access for inspection. Small quadrotor drones offer a potential solution, as they can navigate both horizontal and vertical sections and smoothly fly over debris. However, hovering inside air ducts is problematic due to the airflow generated by the rotors, which recirculates inside the duct and destabilizes the drone, whereas hovering is a key feature for many inspection missions. In this article, we map the aerodynamic forces that affect a hovering drone in a duct using a robotic setup and a force/torque sensor. Based on the collected aerodynamic data, we identify a recommended position for stable flight, which corresponds to the bottom third for a circular duct. We then develop a neural network-based positioning system that leverages low-cost time-of-flight sensors. By combining these aerodynamic insights and the data-driven positioning system, we show that a small quadrotor drone (here, 180 mm) can hover and fly inside small air ducts, starting with a diameter of 350 mm. These results open a new and promising application domain for drones.

Figures (12)

• Figure 1: A-B. Examples of air duct networks. Air ducts are common in most modern buildings because they are necessary for maintaining air quality and temperature. C. Problem considered The objective of this article is to make it possible for small quadrotor drones (here, 180mm) to hover (and fly) air ducts with a diameter from 350 to 560 mm. D. In an air duct, drones are significantly less stable. We tested the same drone with the same controller (based on external motion tracking, Methods), inside and outside a circular air duct with a diameter of 35cm cm and tracked the position for 120 seconds. Because of the aerodynamic recirculations inside a duct, the drone has a hard time hovering at a fixed position.
• Figure 2: A. Experimental setup to measure the aerodynamic forces added by the recirculations. The drone is screwed on a 6-dimensional force/torque sensor, which is fixed to a 7-DOF robotic manipulator. The manipulator makes it possible to measure the forces at 192 different positions, to "map" the aerodynamic forces added by the air duct. B. Forces added by the air duct in a 40cm diameter circular air duct. The arrow shows the direction, the ellipse the variance, and the color the magnitude of the force (N.). The displayed force is the measured force to which the forces outside the duct are subtracted. C. Interpretation of the forces in a 40cm diameter air duct. The blue zone corresponds to the ground effect. In the purple zone, the drone is pushed downward. In the red zones, it is "sucked up" by the walls, that is, these are unstable positions that are likely to lead to collisions. The green zone is the most stable one, as most effects are canceled. Counter-intuitively, the center of the duct is not the most stable position; instead, the airflow is less perturbed at an altitude of about 10 cm (above the ground effect). D. Forces added by the air duct in a 50cm diameter air duct. The pattern is similar to the one in the 40 cm diameter air duct. E. Forces added by the air duct in a 50cm $\times$ 50cm square duct. The arrow shows the direction, the ellipse the variance, and the color the magnitude of the force (N.). The displayed force is the measured force to which the forces outside the duct are subtracted. F. Interpretation of the forces added by the air duct in a 50cm $\times$ 50cm square duct. There is no downward force, but there are clear (1) ground effect (blue), wall effect (red), and ceiling effect (yellow). The center of the duct is stable, but the ceiling and the walls are unstable, as they pull the drone towards the borders.
• Figure 3: Experimental setup to acquire data for the data-driven localization. From left to right, a Lighthouse base station (on a tripod), an LED panel light, a circular air duct in which the drone flies, an Optitrack Trio, and a Lighthouse base station (on a tripod).
• Figure 4: A. 3D view of the drone flying inside a 35cm diameter air duct. The drone is in the middle of the cylinder. The emission cone of the time-of-flight (ToF) sensors is represented in red and is $\phi = 27^\circ$. ToF sensors emitting on the horizontal plane are separated by $\theta = 45^\circ$. $d_i$ and $d_j$ represent 2 distances measured by the ToFs. B-E. Comparison of the position estimations from the geometric and neural network methods against the ground truth inside a 35 cm air duct. The green line represents the Ground Truth measured by the Extended Kalman Filter of the drone helped with the Lighthouse positioning system. The orange and blue lines represent respectively the lateral position outputted by the neural network and the geometrical solution. B. Lateral positions and Absolute Lateral Errors. In this time section of the test set, the geometrical solution is less accurate than the neural network. C. Vertical positions and Absolute Vertical Errors. In this time section of the test set, the geometrical solution is less accurate than the neural network. D. Lateral Errors. This boxplot represents the dispersion of the lateral position error outputted by the neural network and the geometrical solutions with respect to the ground truth in the test set. The neural network outputs a lateral position that is more precise and accurate than the geometrical solution. E. Vertical Errors. This boxplot represents the dispersion of the error between the vertical position measured by the neural network and the ground truth in the test set. The neural network outputs a vertical position that is more precise and accurate than the geometrical solution.
• Figure 5: A. Positions of the drone at altitudes of 115mm and 155mm. Two 2-minute flights are shown in this plot. The red circle represents the 35-cm air duct. The 2 red crosses depict the position the drone must keep for each flight where the drone takes off and hovers at an altitude (Z) of 115mm or 155mm in this air duct. It must stay at $Y = 0$. Two groups of points are visible. The group located below (colors from blue to yellow) are the positions of the drone for the flight at 115mm. The group located above (colors from blue to green) represents the positions of the drone for the flight at 155mm. The neural network used is the one trained on this 35-cm air duct. B. Interquartile Range and Median of Y Positions. This plot depicts the interquartile range (in light green color) and the median (the bright green line) of the lateral positions taken by the drone for different flight altitude targets (represented by the green dots). The target altitude is on the ordinate and the abscissa represents the measured lateral position. It uses the same neural network as the figure A on the left.