IRESE: An intelligent rare-event detection system using unsupervised learning on the IoT edge

In the recent years, a rapid growth of IoT devices has been observed, which in turn results in a huge amount of data produced from multiple sources towards the most disparate cloud platforms or the Internet in general. In a typical cloud-centric approach, the data produced by these devices is simply transmitted over the Internet, for consumption and/or storage. However, with the exponential growth in data production rates, the available network resources are becoming the actual bottleneck of this huge data flowing. Therefore, several challenges are appearing in the coming years, which are mainly related to data transmission, processing, and storage along the so-called cloud-to-thing continuum. In fact, one of the most critical requirements of several IoT applications is low latency, which often hinders raw data consumption to happen at the opposite endpoint with respect to its production. In the context of IoT data stream analytics, for instance, the detection of anomalies or rare-events is one of the most demanding tasks, as it needs prompt detection to increase its significance. In this respect, Fog and Edge Computing seem to be the correct paradigms to alleviate these stringent demands in terms of latency and bandwidth as, by leveraging on re-configurable IoT gateways and smart devices able to support the distribution of the overall computational task, they envisage to liquefy data processing along the way from the sensing device to a cloud endpoint. In this paper, we will present IRESE, that is a rare-event detection system able to apply unsupervised machine learning techniques on the incoming data, directly on affordable gateways located in the IoT edge. Notwithstanding the proposed approach enjoys the benefits of a fully unsupervised learning approach, such as the ability to learn from unlabeled data, it has been tested against various audio rare-event categories, such as gunshot, glass break, scream, and siren, achieving precision and recall measures above 90% in detecting such events