<![CDATA[Indoor air quality monitoring increasingly requires local intelligence because indoor exposure conditions can change faster than centralized systems can respond. However, many machine learning based indoor air quality predictors remain difficult to deploy on microcontrollers due to memory limits, computation constraints, network dependence, and the lack of a reproducible edge deployment workflow. This study develops an end-to-end TinyML framework for deploying a lightweight indoor air quality predictor on a NodeMCU ESP32 using EloquentTinyML and TensorFlow Lite Micro. A synthetic IAQ dataset was generated from eight environmental variables, namely temperature, humidity, PM2.5, PM10, NO2, SO2, CO, and room type, and the indoor air quality score was derived from pollutant weighted features before balancing comfort labels using SMOTE. The proposed compact MLP contains eight inputs, one hidden layer with twelve neurons, ReLU activation, dropout regularization, and a single regression output. Five-fold validation produced an average root mean square error of 5.794, mean absolute error of 4.394, and R2 of 0.877, while the converted model required only 121 trainable parameters. These results indicate that compact TinyML deployment can provide a feasible proof-of-concept for local indoor air quality estimation, although physical sensor validation remains necessary.]]>
Copyrights © 2026
