Characterization of flexible and stretchable sensors using neural networks

Flexible and stretchable sensors made of highly deformable materials have been an active research area with many promising applications. These sensors have the advantages of being highly compliant and elastic, which improves the poor ventilation of the traditional rigid sensors. However, the major drawbacks of flexible and stretchable sensors are the non-linearity and hysteresis in their response, as well as other performance criteria such as precision and repeatability, which may further deteriorate with usage and therefore require pre-emptive calibration from time to time. Most of the existing works often concentrate on new designs and materials and often undermine these issues. To achieve greater precision sensing, this paper proposes an approach with four feature extractors, namely, long short-term memory, gated recurrent units, temporal convolutional networks (TCNs), and a fully convolutional network, to characterize the properties of such sensors. The proposed approach can serve as a calibration method as well as an end-to-end measurement method depending on the settings of the input and the output. We adopted various public datasets to validate the performance of the proposed approach. The experimental results show that the model with a TCN as the feature extractor can give highly promising results with a median error of 0.66% on a kirigami-like sensor and less than 3% on microfluidic-based pressure sensors. With a high performance on representative datasets, the proposed approach is believed to be extendable to other stretchable sensors and actuators for performance analyses, which could greatly increase the adaptation of such sensors in many engineering applications.