Being SMART About Failures: Assessing Repairs in Smart Homes

Inexpensive wireless sensing products are dramatically reducing the cost of in-home sensing. However, these sensors have been found to fail often and prohibitive maintenance costs may negate the cost benefits of inexpensive hardware and do-it-yourself installation. In this paper, we describe a new technique called SMART that uses applicationlevel semantics to detect, assess, and adapt to sensor failures. SMART detects node failures at run-time by analyzing the relative behavior of multiple classifier instances trained to recognize the same set of activities based on different subsets of sensors. Once a failure is detected, SMART assesses its importance and adapts the classifier ensemble in attempt to avoid maintenance dispatch. Evaluation on two public datasets shows that SMART decreases the number of maintenance dispatches by 55% on average, identifies non-failstop failures at run-time with more than 85% accuracy, and improves the activity recognition accuracy under node failures by 15% on average.