We present a new method for on-fine induction motor failure detection using
stator current monitoring. The concept of geometric signal separation in f
eature space is introduced. A set of observations from a single phase of th
e stator current is transformed into a feature vector. After establishing a
local metric in the feature space, close neighbors of the present vector a
re searched for in a database representing allowed states of the motor, whi
ch were recorded during a training period. In their absence the present sta
te is a novelty, which is considered as a failure if if, persists for a cer
tain time. This monitoring scheme successfully deals with varying load cond
itions of the motor, where also oscillating load torques can be tackled. Ad
ditionally, observations from a healthy motor, which differ considerably fr
om any known allowed state in the database due to changes of environmental
conditions can be distinguished from true failures at a significant level.
We do not use expert rules or detailed knowledge about the monitored motor.