A Deep Learning Approach to Non-linearity in Wearable Stretch Sensors

Area of Science:

• Materials Science
• Electrical Engineering
• Machine Learning

Background:

• Wearable technology requires flexible stretch sensors for real-time monitoring.
• Non-linear responses in sensors due to viscoelasticity and strain rate effects pose calibration challenges.

Purpose of the Study:

• To present a general, deep learning-based method for calibrating highly hysteretic resistive stretch sensors.
• To demonstrate the method's applicability to various sensor materials and adaptability to other sensing modalities.

Main Methods:

• A three-stage calibration process involving webcam strain measurement and electrical response recording.
• Utilizing a Long Short-Term Memory Neural Network (LSTM) to learn the complex relationship between sensor resistance and strain.
• Validating the LSTM's predictive accuracy on unseen electrical resistance data.

Main Results:

• The deep learning algorithm successfully calibrated hysteretic stretch sensors without prior knowledge of sensor physics or geometry.
• Highly accurate stretch topology mapping was achieved for three commercial flexible sensor materials.
• The method is adaptable to complex geometries and various sensor types, including capacitive sensors.

Conclusions:

• Deep learning offers a robust solution for calibrating non-linear stretch sensors, overcoming traditional limitations.
• The proposed method provides a versatile and accurate approach for strain monitoring in wearable applications.
• This technique enables reliable performance from inherently non-linear sensor materials.