Explainable Machine Learning for Human Activity Recognition Using Auxetic cTPU Knee-Worn Sensors

Abstract

This paper presents a wearable soft strain sensor based on a commercially available conductive thermoplastic polyurethane (cTPU) 3D-printed as an auxetic soft metamaterial for human activity recognition. The growing demand for flexible and wearable electronics, driven by advances in artificial intelligence, highlights the importance of such sensors in healthcare, medical rehabilitation, soft robotics, and human–machine interfaces. The auxetic cTPU sensor was mechanically and electrically characterized through empirical measurements and validated against numerical simulations. A single sensor mounted on a knee brace was used to collect gait signals across four activities: running, walking, standing, and sitting. Two classification approaches were investigated. A Long Short-Term Memory (LSTM) network was trained directly on the raw time-series signal, with the best configuration achieving 96% accuracy using Relative Standard Deviation Normalization with 50 hidden units. Traditional machine learning models, namely Random Forest and XGBoost, were trained on 30 extracted time-domain and frequency-domain features per motion cycle, achieving 100% and 97.33% accuracy, respectively, under five-fold cross-validation. To enhance model transparency, explainability analysis using SHAP identified power spectral density and the first harmonic frequency as the most consistently influential features across both models, with dynamic activities driven by frequency characteristics and stationary activities distinguished by signal mean amplitude. The results demonstrate that auxetic cTPU soft strain sensors combined with machine learning and explainable artificial intelligence provide an accurate and interpretable solution for wearable human activity recognition, highlighting their potential for applications in robotics, healthcare, and human–robot interfaces.