Smart Office Environment Monitoring System Based on Arduino Nano 33 BLE Sense

The Smart Office Environment Monitoring System based on Arduino Nano 33 BLE Sense is an embedded AI project. It uses Edge Impulse to train models, combines multi-sensor fusion and INT8 quantized neural networks to achieve real-time local evaluation of office environment quality. This system provides a low-cost solution for remote work and smart space management, with end-to-end completeness (hardware selection, sensor fusion, model training to user interface), suitable for developers learning embedded AI and users who want to quickly build monitoring systems.

With the popularity of remote work, the physical conditions of the office environment (temperature, noise, air quality, etc.) have a significant impact on efficiency, cognitive performance and health. However, most people lack objective assessment methods. Traditional devices only display raw data and rely on cloud processing (with privacy risks and latency). The rise of embedded AI provides new ideas: edge devices run lightweight neural networks to complete data collection, analysis and decision-making locally, and can provide real-time feedback without networking.

Hardware Architecture: The Arduino Nano 33 BLE Sense main control board (Nordic nRF52840 SoC) is selected, with onboard HS300x (temperature and humidity), LPS22HB (pressure), and PDM microphone (noise) sensors. Multi-sensor fusion fully reflects the environment quality. Model Design: A fully connected neural network is trained through the Edge Impulse platform, and INT8 quantization is used to compress the model size and accelerate inference (using Cortex-M4 SIMD instructions). Software Architecture: The Arduino firmware coordinates the collection-inference-output loop; the Python serial bridge server connects the device to the web dashboard; the web interface intuitively displays the environment level (color-coded) and raw data.

Model performance indicators: training set accuracy 97.2%, validation set 89.1%, test set 87.4%. Although there is some overfitting, it is acceptable in embedded applications. Classification strategy: Three states are defined: normal, overheated, and noisy. The decision logic combines hard-coded thresholds and model inference (e.g., if the noise exceeds the threshold, it is directly judged as unsuitable; if the model prediction confidence is ≥70%, it corresponds to the level; if it is below, it is general).

As an embedded AI course project, this project demonstrates the complete development process (sensor fusion, model optimization, system integration, engineering practice) and provides a reproducible reference for beginners. It represents a typical application mode of edge AI: low-cost microcontrollers deploy lightweight models to realize intelligent sensing of traditionally complex systems. With the maturity of the TinyML ecosystem, such applications will be widely used in smart home, industrial monitoring and other fields.

Current Limitations: Limited dataset size leads to insufficient generalization; dependency on USB serial port limits deployment flexibility; fixed sampling period is difficult to adapt to fast-changing environments; three-classification simplifies the loss of fine-grained evaluation. Improvement Directions: Expand the dataset to enhance generalization; use BLE for wireless transmission; develop a mobile app to support push notifications; introduce a temperature-humidity comprehensive comfort index.