| Summary: | Everyday objects are becoming smart enough to directly connect to other nearby and remote objects and systems. These objects increasingly interact with machine learning applications that perform feature extraction and model inference in the cloud. However, this approach poses several challenges due to latency, privacy, and dependency on network connectivity between data producers and consumers. To alleviate these limitations, computation should be moved as much as possible towards the IoT edge, that is on gateways, if not directly on data producers. In this paper, we propose a design framework for smart audio sensors able to record and pre-process raw audio streams, before wirelessly transmitting the computed audio features to a modular IoT gateway. In this paper, an anomaly detection algorithm executed as a micro-service is capable of analyzing the received features, hence detecting audio anomalies in real-time. First, to assess the effectiveness of the proposed solution, we deployed a real smart environment showcase. More in detail, we adopted two different anomaly detection algorithms, namely <italic>Elliptic Envelope</italic> and <italic>Isolation Forest</italic>, that were purposely trained and deployed on an affordable IoT gateway to detect anomalous sound events happening in an office environment. Then, we numerically compared both the deployments, in terms of end-to-end latency and gateway CPU load, also deriving some ideal capacity bounds.
|
|---|
