Being SMART About Failures: Assessing Repairs in Activity Detection

One of the main challenges for activity recognition systems is that there are numerous sensors which are likely to move or exhibit failures due to hardware degradation, inaccurate readings, and environmental changes. In this paper, we propose a Simultaneous Multi-classifier Activity Recognition Technique (SMART) that uses application-level semantics to detect sensor node failures and improve the detection accuracy under those failures. Once a node failure is detected, instead of immediately dispatching maintenance, SMART evaluates the severity of the failure by using data replay analysis. Maintenance is dispatched only if the severity analysis indicates that the node failure would have caused an application-level failure in the past and the system could not have recovered from it by updating the classifier ensemble it is using. Evaluation of SMART on a set of activities from two public datasets shows that SMART decreases the number of maintenance dispatches by 45% on average and almost triples the mean time to failure of the application. SMART identifies all applicationlevel failures at run time and improves the activity detection accuracy under node failures by more than 70%. Keywords-wireless sensor networks; activity detection; machine learning; failure analysis