Wing stiffness and neural encoding properties determine optimal sensor locations on a flapping wing.

A: Top: Strain over time is simulated at each possible sensor location in an Euler-Lagrange model of a flapping wing [23, 31]. The wing may be subject to rotation in different axes. Strain is shown for an example sensor location (blue dot). Middle: Strain at each location is encoded by a neural-inspired sensor. Strain is convolved with a linear filter and passed through a static nonlinearity to generate a probability of firing a spike over time, P(fire). Spikes are probabilistically generated, so that exact spike time varies from wingbeat to wingbeat. Bottom: The full data set, a matrix of spike times at all sensor locations for each wingbeat, is used to find a subset of sensors from which information about rotation can be read out. Linear discriminant analysis is used to find the vector w, such that when the data are projected onto w the means of the flapping only and flapping with rotation data are maximally separated. B: Top: Changing wing stiffness produces different strain (shown for identical location) over the course of each wingbeat. Bottom: Changing the threshold of the neural encoding nonlinearity alters of the probability of firing over the course of the wingbeat, with higher thresholds resulting in lower probability. Hawkmoth image in A courtesy of Armin J Hinterwirth.