MicroFed-AD: A Microcontroller-Based Federated Edge Intelligence Framework for Low-Latency Environmental Anomaly Detection

Environmental IoT monitoring requires timely anomaly detection while reducing bandwidth, energy use, and raw-data transmission. This paper presents MicroFed-AD, a microcontroller-oriented edge intelligence framework for PM2.5 and acoustic anomaly detection. The framework combines reconstruction-based TinyML inference with a federated learning evaluation layer for collaborative model refinement. Four unsupervised autoencoder architectures, namely Deep-AE, ConvAE, LSTM-AE, and Conv-VAE, are evaluated using multi-modal environmental features. Conv-AE and LSTM-AE are deployed on an ESP32-S3 development board using TensorFlow Lite Micro, with model files stored in the SPIFFS partition and runtime memory allocated through the Tensor Arena. The Conv-AE achieves the strongest practical trade-off, with an inference time of 0.549s per 24-step input window and a validation MAE of 0.0423, while fitting within the internal SRAM budget. The LSTM-AE is executable only when external PSRAM is available due to its larger Tensor Arena requirement. Federated learning is evaluated through simulation-based client partitioning using FedAvg, FedProx, and FedNAG across different client populations and data distributions. The results show that FedProx and FedNAG generally outperform FedAvg, with Conv-AE remaining the most efficient backbone for edge deployment. These findings provide a hardware-validated local inference foundation and a simulation-supported basis for future multi-node microcontroller-side federated environmental monitoring. Source code is available at: https://github.com/OswaldoEscOrn/MicroFed-AD.