A pattern recognition technique a
ssociated with a new
state e
stimator i
s developed in order to
supervi
se electrical proce
ss. For thi
s purpo
se, diagno
stic feature
s are extracted from current and voltage mea
surement
s for monitoring different operating mode
s. Then, a feature
selection method i
s applied in order to
select the mo
st relevant feature
s which define the feature
space. In thi
s frame, the cla
ssification i
s realized by a non-parametric method (“k-neare
st neighbor
s” rule) with reject option
s. However, thi
s method doe
s not take into account the evolution of the operating mode
s. Thu
s, it i
s nece
ssary to enhance the initial knowledge databa
se. For that, a polynomial approach allow
s characterizing the intermediate
state
s of each operating mode
s and an original u
se of Kalman algorithm allow
s predicting the evolution of the partially known mode
s. A
simple behavioral model i
s u
sed to de
scribe the evolution of the pattern vector. An e
stimation
step provide
s the parameter of
such model and a prediction
step determine
s the future evolution of the pattern vector.
This approach is illustrated on an asynchronous motor of 5.5 kW, in order to detect broken bars under any load level. The experimental results prove the efficiency of pattern recognition methods in condition monitoring of electrical machines.