A Federated Approach for Adaptive Urban Sound Classification on TinyML Edge Devices.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Edge Computing

Background:

• Urban environments generate complex soundscapes with spatial and temporal variations.
• Transmitting raw audio data raises significant communication bandwidth and privacy concerns.
• Resource-constrained edge devices require efficient machine learning models for real-time processing.

Purpose of the Study:

• To develop a federated TinyML architecture for real-time urban sound classification on edge devices.
• To enable communication-efficient collaborative learning without transmitting raw audio.
• To address privacy and bandwidth limitations in urban sound monitoring.

Main Methods:

• A compact audio embedding network served as a frozen feature extractor.
• A lightweight classifier head was trained on-device and shared using MQTT.
• The system was evaluated on ESP32 hardware using cross-dataset transfer (UrbanSound8K to SONYC).
• Federated aggregation techniques were employed for collaborative learning.

Main Results:

• Federated learning recovered performance loss due to domain shift, improving accuracy from 78.27% to approximately 85%.
• Repeated aggregation enhanced macro-F1 score and class balance across heterogeneous datasets.
• Real-time inference was achieved with minimal overhead (~250 ms per window).
• Compact classifier head updates (~1.2 kB) ensured communication efficiency.

Conclusions:

• Federated TinyML enables adaptive, real-time urban sound classification on resource-constrained edge devices.
• The proposed architecture effectively balances performance, communication efficiency, and privacy in smart city networks.
• This approach facilitates distributed, collaborative learning for urban sound analysis.