Developing bio-inspired methodologies for encoding angular position from strain

Abstract

As mechanical systems rely more on closed-loop control, the sensors which supply feedback information are essential. Additionally, in systems where sensor function is critical, sensor redundancy is important to retain functionality if one or more sensors fail. Redundancy can be achieved through multiple high-fidelity sensors which measure the same type of information, such as gyroscopes or accelerometers. However, multiple high-fidelity sensors can increase cost significantly. This thesis explores the potential to replace or augment the functionality of angular position sensors using strain measurements. Strain gauges are already used in system health monitoring systems. By utilizing these already implemented sensors to measure angular position, we can remove the additional cost of redundant angular position sensors. However, for complex systems, the mapping between strain and angular position is unclear. By incorporating reduced order, physics-based models into machine learning techniques, we can efficiently transform high-order strain data into angular position. To demonstrate the potential of using alternative sensing methods, we developed a reduced order model of a parametrically excited flexible pendulum. Inspiration for this simplified system comes from insect halteres, which are small sensory organs evolved from insect hind wings which provide rapid information about body rotation. The parametrically excited flexible pendulum allows a single axis of rotation and single direction of flexibility to be paired, and their relationship studied. By varying parameters within the model such as pendulum length and modulus as well as parametric excitation amplitude and frequency, the Gaussian process regression learning can be optimized to reduce training time and increase untrained prediction accuracy. Inputs of strain and parametric excitation position along with their respective first and second derivatives are then analyzed to determine which inputs are interrelated and therefore un-necessary, thus reducing the input required. This provides the essential first steps towards using machine learning to implement multiple sensor, deformation based, multi axial angular position sensing in complex systems.