ESP32-Based Intelligent Wearable System for Human Activity Recognition and Behavioral Analytics

This preprint documents the end-to-end engineering design, hardware fabrication, fault-isolation debugging, embedded firmware architecture, and host-side machine learning analytics of an integrated wearable system for real-time posture telemetry and human activity recognition. The wearable hardware integrates a 30-pin ESP32-WROOM-32 dual-core microcontroller and an MPU6050 6-axis MEMS inertial measurement unit onto a custom single-sided CNC-milled copper printed circuit board (PCB) housed within a 3D-printed wearable enclosure. Through a systematic three-stage empirical hardware troubleshooting protocol, an initial collector-emitter thermal breakdown fault in an NPN transistor driving an inductive vibration motor was resolved via a solid-state architectural pivot to a noise-immune 3.3V digital telemetry topology. On the embedded side, the firmware executes a 3.0-second startup auto-calibration baseline zeroing routine, computes 3D Euler orientation angles via four-quadrant $\text{atan2}$ kinematics, and enforces a 3.0-second non-blocking millis() sustained tilt debounce finite state machine ($\pm 25.0^\circ$ threshold). Real-time kinematic vectors, Signal Vector Magnitude ($SVM$), and gyroscope rotational energy are broadcast simultaneously at 5 Hz over wired USB UART and wireless Bluetooth Classic Serial Port Profile (SPP). On the host workstation, an interactive Python dashboard ingests the telemetry stream to render a 2D Pitch/Roll bullseye target, multi-axis biomechanics radar profiles, gamified wellness progression, and a scikit-learn Random Forest classifier trained across 1,080 sliding-window feature vectors achieving 96.8% validation accuracy.