Hierarchical Crack-Engineered Strain Sensors for Machine-Learning-Enabled Multimodal Recognition and Edge Computing in Ultra-Low-Power Wearables

Next-generation wearable electronics require multimodal sensing with high sensitivity, a wide linear strain range, and low power consumption, yet existing strain sensing systems face inherent trade-offs among these metrics. Here, we introduce a hierarchically engineered Thickness Gradient and Surface Topology (TGST) strain sensor with a crack-controlled architecture, achieving a gauge factor of 273.33 and a linear response up to 150% strain. Leveraging these capabilities, we developed an ML-driven Ensemble Sequential Decoupling Model (ESDM) that enables a single sensor to separate multiple overlapping stimuli, including pulse, gesture, sound, and pressure, reducing reliance on multiple dedicated sensors and improving power efficiency. We further integrate a distributed TGST sensor array into an edge computing module enabled by an Ensemble Convolutional Neural Network Reconstruction Model (ECNNRM), enabling high-accuracy real-time motion tracking with 85% energy savings. This ultra-low-power framework advances real-time health monitoring, fall detection, and human-machine interaction, offering a scalable pathway toward ML-enabled telehealth applications.