You Are the Wind Beneath My Sensors

In the world of unmanned aerial vehicles (UAVs), quadcopters have taken center stage. These vehicles are very nimble and stable in flight, and have proven themselves to be extremely useful in a wide range of applications time and again. But there is just one little problem — battery life. Spinning rotors are not very energy-efficient, to say the least, so onboard batteries are drained quickly. And that, in turn, severely limits the flight time and range of quadcopters.

Given that insects, birds, and other flying creatures, can take to the skies in a far more energy-efficient manner than today’s quadcopters, researchers have increasingly been turning to the natural world for inspiration in building better UAVs. Unfortunately, machines that mimic these natural creatures are far more challenging to design and control.

It is known that flying creatures have senses — some of which we hardly understand — that make it possible for them to continuously adapt to changing conditions and make lightning-fast corrections to maintain stable flight. Unlocking the secrets of these senses may be the key to more efficient flight.

A pair of engineers at the Institute of Science Tokyo have honed in on one of these special senses that has been discovered in hummingbirds and other winged animals. This sensory data is collected by mechanical receptors positioned on the creatures’ wings that recognize strain forces. It is believed that this provides information about changes in the wind, or possibly other environmental data.

To better understand this natural system and how it might be mimicked, the team developed an artificial copy of it for the purpose of classifying wind direction. Their work utilizes biomimetic flexible wings each equipped with a set of seven commercial strain gauges that pass data into a convolutional neural network (CNN) for processing of the wing deformation data.

The wings, modeled after hummingbird and hawk moth anatomy, were built with a 3D-printed frame that replicates natural flight feather shafts and were covered with a lightweight polyimide film. The wings were attached to a custom flapping mechanism driven by a DC motor, producing flapping motions at variable rates. This setup was tested in a wind tunnel under controlled conditions to simulate hovering flight.

The CNN model first segmented the time-series strain data into either full flapping cycles or shorter segments. It then extracted features from the strain data as it passed through the layers of the model, ultimately classifying the wind direction using a softmax output layer. The model was trained using supervised learning with labeled datasets for various wind directions and a no-wind condition.

Experimental results demonstrated that the system could classify wind direction with a high level of accuracy. With data from all seven strain gauges over a full flapping cycle, the accuracy reached 99.5 percent. Even with shorter data segments, the accuracy remained high at 85.2 percent. Using a single strain gauge, classification accuracy ranged from 95.2 percent to 98.8 percent over a full flapping cycle, though it dropped significantly with shorter data lengths.

The researchers’ work demonstrates that alternative methods of flight might be made more practical in the future through the use of lightweight biomimetic sensing mechanisms. It is hoped that additional enhancements, such as the inclusion of a recurrent neural network for data processing, will make the system even more effective in the days ahead.