Continual Learning-Enabled On-Device Audio Event Detection for Low-Latency Edge-IoT Applications

Abstract

Booming Internet of Things (IoT) applications in smart spaces have led to the need of real-time, accurate, and adaptive audio event detection (AED) engine. Traditional cloud-based AED methods are not always fast during the communication process, they raise privacy issues, and need regular retraining of the model to be accurate in the changing acoustic conditions. In this case study, the design, deployment, and assessment of a continuously learning-empowered on-device AED system has been provided that will be configured to low-latency Edge-IoT workloads. The offered architecture combines lightweight convolutional-recurrent neural network with the continual learning module based on elastic weight consolidation, allowing the model to adapt gradually to new audio events with avoiding catastrophic forgetting. The implementation is on an ARM based edge computing machine and real-life smart manufacturing floor environment with multiple sources of sound, varying noise intensity, and low bandwidth network. The continuous learning strategy on field tests shows increases in event recognition precision by nearly 18 percent over fixed designs when concurrently tested with previously unseen audio classes, with an end-to-end detection latency of <50 ms and a decrease of 92 percent bandwidth usage in contrast to cloud-based instruments. Moreover, on-device processing grants the data privacy requirements uphold and, hence, greatly saves the energy utilized making the system applicable in carrying out IoT nodes that run with batteries. The results indicate the practicality and efficiency of incorporating lifelong learning in an edge-deployed AED system in dynamic resource constrained settings and insight into scaling such architectures to larger smart city, smart healthcare, and smart industrial applications.