The use of deep learning algorithms to predict mechanical strain from linear acceleration and angular rates of motion recorded from a horse hoof during exercise

Abstract

Mobility solutions offered by living creatures have inspired engineers to capture their locomotion patterns and then develop novel animal-like robots that use legs for locomotion. Exploring relationships amongst mechanical responses and kinematic parameters is essential for both inventing these robots and enhancing computational techniques. Establishment of accurate physical models to quantify mechanical responses of biological systems is challenging because the corresponding variables are multidimensional, dynamic and highly nonlinear. This encourages the advent of data-driven models in mechanical sciences. This paper delves into the use of feedforward and time-series (dynamic) artificial neural networks (ANN) to analyse experimental data recorded from a racing horse exercised up to 60 km/h to then relate hoof mechanical strain to kinematic parameters recorded experimentally. An inertial measurement unit that was comprised of a sensor and data acquisition system package was designed and mounted on the horse's hoof to measure linear accelerations and angular rates of motion. In addition, an instrumented Aluminium horseshoe that was designed and manufactured and contained: (1) inertial sensors including three orthogonal accelerometers and three orthogonal rate gyroscopes; and, (2) a strain gauge located at the middle of the shoe. The horse was warmed up in a steady gallop at around 35 km/h for 1 km then turned around and galloped at increasing speed to 68 km/h back to the finishing line. Nine kinematic parameters, measured during horse exercise, formed the ANNs input variables: hoof linear accelerations along three orthogonal directions (axe,ay,az), hoof angular rates of motion along three orthogonal directions (Gx, Gy, Gz), shoe linear accelerations along three orthogonal directions (axs, ays, azs), and time. Feedforward and time-series ANNs trained using 1,000,000 experimental instances offered excellent reliability for the prediction of mechanical strain from kinematic measurements, i.e. R ≥ 0.97.