ABSTRACT:

Learning methods use historic data points about a process past behavior 
to build a predictor (classifier), which can be employed to predict the 
process future behavior. However, in most real-world applications the 
data are non-stationary; they change continuously over time. This 
requires not only a continuous learning as new patterns are available, 
but the predictor needs to adjust itself (self-correct or adapt) to 
drifts in data characteristic over time or when new events happen or 
new conditions occur. The goal is to ensure an accurate prediction of 
process behavior according to the changes in the new incoming data 
characteristics. This requires a continuous learning over long period 
of time with the ability to forget data becoming obsolete and useless. 

In this tutorial, the problem of learning from data from time-based and 
complex non-stationary processes will be studied. Some efficient and 
adapted techniques to handle this problem will be presented and 
explained using several illustrative examples. Their application on 
some real-world problems will be discussed. In particular, two 
application examples will be detailed. The first one is the monitoring 
of the quality of folded pieces based on the analysis of acoustic 
signals characterizing the noises issued of the folding operation. The 
second application is the monitoring of the functioning of the steam 
generator in a nuclear reactor. This monitoring is based on the 
analysis of steam generator noises measured by two acoustic sensors.