Data, firmware and code for "Train once, deploy anywhere: physics-based period normalization for portable tiny machine learning on duty-cycle-modulated sensors"

Measurements, models, firmware and scripts behind every number and figure of the paper "Train once, deploy anywhere: physics-based period normalization for portable tiny machine learning on duty-cycle-modulated sensors" (submitted). The experiments use a four-class temperature classifier (LM35 sensor read through a non-inverting duty-cycle modulator). It runs as on-device Random Forests, trained on an ESP32-S3 and deployed without retraining on an STM32H743, with two capture mechanisms (TIM2 with DMA, and interrupts) and with the modulator capacitor changed from 33 to 3.12 nF. Contents: data/: ESP32-S3 training sessions and on-device campaigns; STM32H743 step and staircase logs, with a run catalogue and the rest-period calibration log. models/: originally deployed models, v2 models (period normalization and scale augmentation) and control models (no augmentation). firmware/: ESP32-S3 firmware v1.9.1 and v1.9.3 with the mcpwmcap capture module; STM32H743 port. code/: run_all.py reproduces the tables and figures. Data: CC BY 4.0. Code and firmware: MIT (see LICENSE). Requirements: Python 3, numpy 2.2.6, scikit-learn 1.7.2.