Lightweight Framework for Time-Varying Error Compensation in Soft Strain Sensors

Soft strain sensors offer compliant, lightweight, and low-cost solutions for wearable sensing and human–machine interfaces. However, their long-term performance is often degraded by baseline drift and nonlinear hysteresis caused by the viscoelastic behavior and aging of soft materials. Existing compensation approaches commonly rely on recurrent or adaptive learning models that require substantial computational resources and long historical data. Here, we present a lightweight framework for real-time compensation of time-varying errors in soft strain sensors. The framework combines an exponential moving average (EMA) algorithm for continuous drift tracking and baseline stabilization with a compact feedforward neural network (FNN) for residual nonlinearity and hysteresis compensation. Using a hydraulic filament sensor subjected to randomized cyclic stretching over two hours, the proposed method achieved a displacement prediction root-mean-square error (RMSE) of 1.18 mm (2.95% of the operating range) and a coefficient of determination (R²) of 0.99. The framework also reduced hysteresis errors from 5.76% to 1.84%. Compared with recurrent and attention-based neural network architectures, the proposed approach achieved superior accuracy while maintaining low computational complexity, making it suitable for real-time deployment. As a proof of concept, the framework was integrated into a wearable knee-motion monitoring system, achieving a knee-angle estimation RMSE of 3.63°. These results demonstrate an effective and scalable strategy for improving the long-term reliability of soft strain sensors.