Sensor networks deployed for scientific data acquisition must inspect measurements for faults and events of interest. Doing so is crucial to ensure the relevance and correctness of the collected data. In this work we unify fault and event detection under a general
framework. We use machine learning techniques to classify measurements that resemble a training set as
and measurements that significantly deviate from that set as
. Furthermore, we aim at an anomaly detection framework that can be implemented on motes, thereby allowing them to continue collecting scientifically-relevant data even in the absence of network connectivity. The general consensus thus far has been that learning-based techniques are too resource intensive to be implemented on mote-class devices. In this paper, we challenge this belief. We implement an anomaly detection algorithm using Echo State Networks (ESN), a family of sparse neural networks, on a mote-class device and show that its accuracy is comparable to a PC-based implementation. Furthermore, we show that ESNs detect more faults and have fewer false positives than rule-based fault detection mechanisms. More importantly, while rule-based fault detection algorithms generate false negatives and misclassify events as faults, ESNs are
, correctly identifying a wide variety of anomalies.