Intelligent Smoke Detection System and Adaptive AI Threshold Optimization on an RA MCU Embedded Platform

Qiushi Xie^‡, Yutong Bai, Jinghuan Xiao, Yujiang Zeng^*

Apr–Aug 2025 · 8th National College Student Embedded Chip and System Design Contest — Chip Application Track · National Finals Second Prize

A lightweight smoke time-series prediction network deployed on the Renesas RA6M5, demonstrated through a dual-board comparison between AI prediction and a fixed-threshold baseline. It provides warnings 3–5 seconds earlier with a false-alarm rate below 0.5%, using a low-cost optical dust sensor and fully local MCU inference for affordable, offline operation.

ContestWebsite

Intelligent Smoke Detection · AI Prediction and Low-cost DeploymentRA6M5 · On-device AI · Dual-board A/B comparison

Dual-board comparison prototype at the national finals — National Finals Prototype · On-site A/B comparison of AI prediction and fixed-threshold boards

Smoke acquisition and prediction runtime interface — Live Data · Concentration acquisition, five-second prediction, and error display

Compact smoke detection prototype — Compact Prototype · Integrated display, alarm, and sensing

RA6M5 board inside the smoke detection system — Embedded Hardware · RA6M5, dust sensor, and audiovisual peripherals

Smoke detection system block diagram — System Loop · Acquisition, on-device prediction, display, and graded alarms

Smoke ground truth, prediction, and future trend curves — Time-series Forecasting · Ground truth, one-step prediction, and future trend

Lightweight neural network ROM, RAM, and compute requirements — Lightweight Deployment · 113,576 B ROM and 65,536 B peak RAM

Smoke prediction and graded alarm software flow — Software Flow · Calibration, prediction, state classification, and audiovisual alarms

3–5 sEarlier warning than fixed threshold≤0.5%System false-alarm rate70 msRA6M5 inference latency800+Self-collected smoke samples

‡ Team lead · * Advisor