Machine-Learning-Assisted Signal Interpretation in a Dual-Architecture Polyvinyl Alcohol/Tellurium/Selenium Sensing Platform for Drift-Tolerant Humidity Monitoring and Tactile Classification

Chemical oxidation and viscoelastic deformation can cause substantial signal drift and time-dependent responses in soft wearable sensors, limiting their long-term reliability. Here, we present a machine-learning-assisted signal interpretation framework using a common polyvinyl alcohol (PVA)/tellurium (Te) nanowire/selenium (Se) nanoparticle composite configured into two sensor architectures: an unencapsulated thin-film humidity sensor and a porous piezoresistive pressure sensor. The unencapsulated humidity sensor exhibited severe oxidation-induced ambient aging over 115 days, including a substantial chemical baseline resistance shift and more than 96% loss in raw response. By applying fractional normalization and few-shot XGBoost calibration, functional classification of aged humidity-related signals was partially recovered, achieving 77.3% accuracy on degraded sensor data. In a separate porous pressure-sensing architecture, compression-induced current responses were obtained above an activation threshold of approximately 191 kPa, and LSTM analysis enabled classification of time-series mechanical signals with 98.7% accuracy. These results suggest that algorithmic calibration and time-series learning can extend the functional usability of degraded or mechanically compliant soft sensors, although future work is required to demonstrate fully integrated dual-modal operation and cross-sensitivity suppression.