A clear Example of a job I have done in past :
In Chemical Engineering, reactor temperature and reactant concetration is being plotted against . There is NO information known on the error/oulier distribution. task is to find similar events in time.
To do this, the python script attempts to fit the graph, via multiple types of fitting (exponential, polynomial ... ) , uses covex hull method to detect outliers - and then classifies the period of interest . Such a classification would look like :
timestep : 26 to 59
features identified : {[26,37] convex, increasing}, {[37,38] spike, up}, {[38,42] noise CANTFIT} , {[42,58] concave, decreasing}
Then, this is fed into a fuzzifier, and a weighting is generated. That is then fed into a ANN.
Then, you can feed the ANN with a longer time series or more serieses - it will then cut out regions in the time series which match your goal.
So far I don't have any recommendations. But I am interested in discussing it out more with you.
I am aalso available on skype: sean_s_con