Frontiers journals are at the top of citation and impact metrics

Original Research ARTICLE Provisionally accepted The full-text will be published soon. Notify me

Front. Robot. AI | doi: 10.3389/frobt.2019.00027

A deep learning approach to non-linearity in wearable stretch sensors

 Mark Miodownik1*, Ben Oldfrey1, Richard Jackson1 and  Peter Smitham2
  • 1Institute of Making, University College London, United Kingdom
  • 2Institute of Orthopaedics & Musculoskeletal Science, Royal National Orthopaedic Hospital, United Kingdom

There is a growing need for flexible stretch sensors to monitor real time stress and strain in wearable technology. However, developing stretch sensors with linear responses is difficult due to viscoelastic and strain rate dependent effects. Instead of trying to engineer the perfect linear sensor we take a deep learning approach which can cope with non-linearity and yet still deliver reliable results. We present a general method for calibrating highly hysteretic resistive stretch sensors. We show results for textile and elastomeric stretch sensors however we believe the method is directly applicable to any physical choice of sensor material and fabrication, and easily adaptable to other sensing methods, such as those based on capacitance. Our algorithm does not require any a priori knowledge of the physical attributes or geometry of the sensor to be calibrated, which is a key advantage as stretchable sensors are generally applicable to highly complex geometries with integrated electronics requiring bespoke manufacture. The method involves three-stages. The first stage requires a calibration step in which the strain of the sensor material is measured using a webcam while the electrical response is measured via a set of arduino-based electronics. During this data collection stage, the strain is applied manually by pulling the sensor over a range of strains and strain rates corresponding to the realistic in-use strain and strain rates. The correlated data between electrical resistance and measured strain and strain rate are stored. In the second stage the data is passed to a Long Short Term Memory Neural Network (LSTM) which is trained using part of the data set. The ability of the LSTM to predict the strain state given a stream of unseen electrical resistance data is then assessed and the maximum errors established. In the third stage the sensor is removed from the webcam calibration set-up and embedded in the wearable application where the live stream of electrical resistance is the only measure of strain - this corresponds to the proposed use case. Highly accurate stretch topology mapping is achieved for the three commercially available flexible sensor materials tested.

Keywords: Sensors, stretch, flexible, Real-time, Deep Learming, Long Short Term Memory (LSTM)

Received: 19 Jun 2018; Accepted: 02 Apr 2019.

Edited by:

Lucia Beccai, Istituto Italiano di Tecnologia, Italy

Reviewed by:

Yinlai Jiang, University of Electro-Communications, Japan
Massimo Totaro, Istituto Italiano di Tecnologia, Italy  

Copyright: © 2019 Miodownik, Oldfrey, Jackson and Smitham. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

* Correspondence: Prof. Mark Miodownik, University College London, Institute of Making, London, WC1E 7JE, United Kingdom, m.miodownik@ucl.ac.uk