Event-triggered natural hazard monitoring with convolutional neural networks on the edge

Abstract

In natural hazard warning systems fast decision making is vital to avoid catastrophes. Decision making at the edge of a wireless sensor network promises fast response times but is limited by the availability of energy, data transfer speed, processing and memory constraints. In this work we present a realization of a wireless sensor network for hazard monitoring based on an array of eventtriggered single-channel micro-seismic sensors with advanced signal processing and characterization capabilities based on a novel co-detection technique. On the one hand we leverage an ultra-low power, threshold-triggering circuit paired with on-demand digital signal acquisition capable of extracting relevant information exactly and efficiently at times when it matters most and consequentially not wasting precious resources when nothing can be observed. On the other hand we utilize machine-learning-based classification implemented on low-power, off-the-shelf microcontrollers to avoid false positive warnings and to actively identify humans in hazard zones. The sensors' response time and memory requirement is substantially improved by quantizing and pipelining the inference of a convolutional neural network. In this way, convolutional neural networks that would not run unmodified on a memory constrained device can be executed in real-time and at scale on low-power embedded devices. A field study with our system is running on the rockfall scarp of the Matterhorn Hörnligrat at 3500 m a.s.l. since 08/2018.

A decade of detailed observations (2008–2018) in steep bedrock permafrost at the Matterhorn Hörnligrat (Zermatt, CH)

“…Along-side a number of technology-oriented publications have emerged that discuss sensor and wireless network design (Talzi et al, 2007;Hasler et al, 2008;Beutel et al, 2009;Keller et al, 2009b;Buchli et al, 2012;Sutton et al, 2015bSutton et al, , a, 2017a, performance analysis (Keller et al, 2012a(Keller et al, , 2011 and smart sensors (Sutton et al, 2017b;. More recently focus has shifted on even more complex sensing modalities including machine learning methods to the portfolio of application-specific data analysis (Meyer et al, 2017(Meyer et al, , 2019b.…”

The dual processor platform architecture

“…In order to support the long-range capabilities of the SX1262 radio, an extension of the FlockLab testbed to other rooftop locations spread throughout the city is currently ongoing 1 . The CC430 ComBoard in conjunction with eLWB [6] has been used in a wireless event-triggered geophone sensor system [4] where it is currently used in two field deployments that are publicly accessible at http://data.permasense.ch. The DPP2 platform has also been used as an educational facility in a recent lecture/lab at ETH Zurich.…”

“…In this demo we will demonstrate an implementation of the Eventbased Low-power Wireless Bus (eLWB) [6] based on an FSK version of Glossy for the CC430. The application is an event-triggered geophone platform [4] that uses the CC430 ComBoard and eLWB. This application generates periodic as well as event-based traffic that is transmitted using eLWB to a central sink (base station) running on a laptop.…”

How many climb the matterhorn?

“…The first strategy is performed using incremental network quantization [7] and reduces the size of the network's parameters by a factor of 4. The latter strategy is performed by using a novel method we coined time distributed processing [3]. Time distributed processing pipelines the inference by using a depth-first computation of the convolutional neural network and by exploiting the temporal characteristics of microseismic data.…”