A survey of datamining methods for sensor network bug diagnosis

Abstract

This chapter surveys recent debugging tools for sensor networks that are inspired by data mining algorithms. These tools are motivated by the increased complexity and scale of sensor network applications, making it harder to identify root causes of system problems. At a high level, debugging solutions in the domain of sensor networks can be classified according to their goal into two distinct categories; (i) solutions that attempt to localize errors to a single node, component, or code snippet, and (ii) solutions that attempt to identify a global pattern that causes misbehavior to occur. The first category inherits the usual wisdom that problems are often localized. It is unlikely for independent failures to coinside. Hence, while many different trouble symptoms may occur simultaneously, they typically arise from a single misbehaving component such as a failed radio or a crashed node that may, in turn, trigger a cascade of other problems. In contrast, the second category of solutions is motivated by interactive complexity problems. They seek to uncover bugs in networked sensing systems that arise due to unexpected interactions between components. The underlying assumption is that individual components are easier to test, which ensures that they work well in isolation. Therefore, practical software systems seldom fail due to a single poorly-coded component. Rather, they fail due to an unexpected interaction pattern between individually well-behaved components. The challenge is to uncover the global interaction patterns that leads to the problem, as opposed to chasing a local root cause. The chapter describes the above different techniques and concludes with a brief review of other troubleshooting work, not inspired by data mining literature.