The ever-growing demand of soft strain sensors in ubiquitous electronics poses urgent need of device durability innovations to reduce resource waste. As most soft sensors use non-degradable materials, it is important to ensure durable sensor performance with reliable sensing signals over its lifecycle to reduce device waste, yet facing pervasive signal predicaments of nonlinearity, hysteresis, cycling attenuation, and batch inconsistency. Herein, instead of empirical experiments via material or structural engineering, we propose an efficient computational calibration framework to comprehensively address these signal predicaments by utilizing hierarchical physics-guided machine learning (ML) models. Using an eco-friendly carbon waste-based strain sensor as a case study, physics-guided ML models are developed and show both high computation efficiency and learning accuracy in terms of real-time sensing signal calibration of the developed sensor, which automatically calibrates its unsatisfied sensing performance to equivalently linear, non-hysteresis, long-term stable, and batch consistent one without trial-and-error experiments. As final demonstrations, the ML-driven calibration models are capable of expanding sensor’s reliable working lifetime for >3000 times than its own counterpart for multiple robotic tasks, facilitating their long-term usages during practical applications. The high demand of soft strain sensors requires solutions to reduce resource waste. Here, the authors propose a machine learning framework that corrects unstable signals in soft strain sensors, making them more accurate, reliable and longer-lasting, and extending their usable lifetime by more than 3,000 times.
Li et al. (Thu,) studied this question.