All the trend detection algorithms(for social networks) that I have seen use some kind of an average based on the previous data and for every new recording they check how far it is from the average. The problem I see with this approach is:
how can we remove noises from the past data?
We would need something that works on a sequence of data and not on a particular point, it is not about the peaks but whether a particular sequence of data will lead to a one.
Another problem with average based methods are lack of past data