Stimulation Initiative for European Neural Applications
Esprit Project 9811
Alarm Identification with SENECA
Company background
No further information available
The problem
In a continuous casting facility molten steel is poured from a melting
pot into a tundish and from there into a watercooled mold where it
solidifies. The solid steel is then cut into pieces of a fixed lenght
for further treatment in a rolling mill. Due to the cooling of within
the mold the steel develops a solid shell which surrounds a still molten
core when it is withdrawn from the mold. From time to time the formation
of the solid shell is disturbed so that a rupture of the shell occures,
the rupture leaves the mold and liquid stell leaks into the facilities.
This causes long shutdown times an enormous costs.
Neural network application
To detect the temperature signal pattern which is characteristic for a
rupture a neural network based system has been developed. It consists of
multi-layer-perceptron networks for each sensor which share their
weights among each other and generate a response whether the network
detected a rupture or not. The system was trained with data collected
from 262 casting processes which have occured during the operation of
the casting facility. Within this data were 86 real and 171 false alarms
detected by a conventional algorithm. A selected training set of real
and false alarms has been used to train the neural network.
Benefits
Over a period of several months the system has been tested parallel to
the conventional algorithm. All real alarms have been detected by both
systems. However, the alarms from the neural detection system came up to
14 seconds earlier than those of the conventional system. In online
operation, the neural detection system's false alarm rate is
considerably lower in comparison to the conventional system.
Generalization
The ability of neural networks to detect complex patterns in noisy
signals makes them a valuable tool in almost any fault detection system
and helps increasing the production's quality standards. The described
application is one example for a neural network based system which has
been trained on historical data and which can improve continuously by
learning from experiences gathered during online operation.
Contact person
Stefan Fuhrmann -
Zentrum für Neuroinformatik GmbH.
Universitätsstraße 160,
D-44801 Bochum.
Phone: +49 234 9787-51,
Fax: +49 234 9787-77.