PrediSense

Objective:

Nowadays, the vast majority of the data acquired by Internet-of-Things (IoT) sensors is useless from the point of view of the target applications. This is particularly evident on IoT systems for continuous monitoring, which are set up to collect data at a high rate despite interesting events occurring very scarcely, thus consuming much more energy than needed. In PrediSense, we address the challenges to design the next generation of smarter ultra-low power IoT systems by drawing inspiration from human perception. The human brain actually ignores most of the data acquired by the sensory nervous system, as our perception is driven by attention and expectations. Similarly, PrediSense advocates a sensing strategy that exploits knowledge about the acquired signals to predict their future behavior, and then dynamically tunes the sensor device to focus on the expected time and value ranges of interest. This approach supports a paradigm shift to enable radical energy savings for IoT systems and to avoid the requirements of the Nyquist-Shannon theorem. To this end, PrediSense will be built on four pillars: 1. Abductive reasoning, for selecting the most likely hypothesis leading to the prediction of the upcoming events of interest. 2. Self-supervised machine learning, for building explanatory hypotheses of time intervals with comparable event dynamics. 3. Event-based sampling, for changing the focus from a classical, high-rate and data-blind sampling to an approach based on the behavior of the signal. 4. Bayesian inference, for quantifying and narrowing the time and value ranges for data acquisition. The project will tackle three different use cases in the key domain of remote healthcare and biosignal monitoring, including cardiovascular, musculoskeletal and neurological monitoring. Hence, PrediSense will push the limits of the current technology to open new avenues into the design of smart ultra-low power devices that will shape the future of energy-efficient IoT.