Inter-floor Noise Detection with TinyML (On-Sensor AI)

[ English | 한국어 ]

• DBpia: https://www.dbpia.co.kr/journal/articleDetail?nodeId=NODE11743327
• PDF: paper/Addressing Inter-floor Noise Issues in Apartment Buildings using On-Sensor AI Embedded with TinyML on Ultra-Low-Power Systems.pdf

This repository contains the implementation assets behind the paper: “Addressing Inter-floor Noise Issues in Apartment Buildings using On-Sensor AI Embedded with TinyML on Ultra-Low-Power Systems” (March 2024).

The project targets real-time inter-floor noise discrimination using Arduino Nano 33 BLE boards and an on-device CNN (TensorFlow Lite for Microcontrollers), avoiding server-side inference.

The original chronological research log has been reorganized into a portfolio-first structure focused on reproducibility, traceability, and paper-to-code mapping.

Paper

• PDF: paper/Addressing Inter-floor Noise Issues in Apartment Buildings using On-Sensor AI Embedded with TinyML on Ultra-Low-Power Systems.pdf
• Summary: docs/paper_summary.md
• Paper-to-code map: docs/paper_to_code_map.md

Quickstart

1. Create an environment and install dependencies:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

2. Run repository sanity checks:

make smoke

3. Rebuild dataset windows (optional, if regenerating processed data):

make preprocess

4. Train/evaluate/export model:

make train

5. Convert .tflite to C array for firmware:

make tflite2cc

Reproducing Key Results

• Full reproduction guide: docs/reproduce.md
• Experimental summary artifacts: experiments/results/

Repository Structure

• paper/: included paper PDF and citation metadata
• docs/: methodology summary, mapping, reproduction, audit notes
• src/: reproducible Python scripts (collection helpers, preprocessing, training)
• embedded/: on-device TinyML firmware (final floor_noise_v5) + BLE examples
• data/: raw captures and processed train/eval windows
• notebooks/: paper-related notebooks + legacy tutorial notebook
• experiments/: trained artifacts and reported result tables
• assets/: project images used in docs
• archive/: original chronological project layout for traceability

What’s In / What’s Not

In this repo

• Raw/processed data files used during development (small-to-midsize local datasets)
• Model artifacts (.tflite, .cc) and firmware source
• Reproducible scripts for preprocessing/training/export

Not in this repo

• Physical hardware setup automation (manual installation on-site)
• Full experimental environment replication (building/floor setup constraints)
• Guaranteed bit-identical retraining outcomes across all TensorFlow versions/hardware

Citation

If you use this repository, please cite:

J.-W. Kwak and I.-Y. Choi, “Addressing Inter-floor Noise Issues in Apartment Buildings using On-Sensor AI Embedded with TinyML on Ultra-Low-Power Systems,” Journal of The Korea Society of Computer and Information, vol. 29, no. 3, pp. 75–81, Mar. 2024.