This is very interesting but does raise the question of how to adequately define patterns. One would have to have a working understanding of the type of events that need alerts, and how they look on a sensor, as well as what type of false alarm could look the same. But it is a definite step up from threshold-based alerting, which is hugely susceptible to false alarms.
Using the patterns on their own is a bit of an iterative process, but I'm sure you can imagine how it's a substrate on which all kinds of interesting automation of pattern learning could be built.
