Information Technology Reference
In-Depth Information
network having as inputs the contents of some chemical elements and some of the
previously predicted hardenability values, according to the schematic description provided
in Fig. 1.
This application proved to be successful because: i) results were accurate and accuracy could easily be
measured; ii) the intelligent systems (a neural network in both cases) was very simple, with few
weights, and these could easily be interpreted by the technician; iii) the neural predictor has been
coupled with a user-friendly software interface allowing not only to run the model, but also to collect
data and re-train all the neural networks with new data provided by the user, so that each steel
company can progressively “specialise” the predictor on its own steel grades; iv) training time for
using the software tool which was developed was very short. It is worth noting that, as pointed out at
the beginning, the neural network itself is just a small element of the whole system (software tool, pre-
processing, data collection, result presentation, etc.
4.2 Prediction of malfunctioning during steel casting
In the standard steelmaking practice, during continuous casting, the liquid material
produced in the blast furnace is cast, after some manufacture, into the ladle and,
subsequently, into the tundish (see Fig. 2). On the bottom of the tundish, some nozzles are
located, through which the liquid steel passes into the mould or strip casters. The section of
such nozzles is far smaller with respect to the tundish dimensions. When particular steel
grades are produced, some alumina precipitation on the entry and on the lateral surface of
these nozzles can partially or even totally block the flow of the liquid steel. This
phenomenon is commonly known as clogging and is highly detrimental to casting reliability
and quality of the cast products.
a)
b)
Fig. 2. a) Location of the nozzles that can be occluded; b) Schematic description of the
labelled SOM-based classifier.
The clogging phenomenon is still not deeply understood (Heesom 1988), due to the very
high number of chemical and process factors affecting the occurrence of the precipitation of
the materials on the nozzle internal surface as well as to the impossibility of installing
complex systems of probes and sensors in order to closely observe the phenomenon itself.
For this reason, some attempts have been performed to apply intelligent systems for the
prediction of clogging occurrence on the basis of the steel chemical composition and on the
process parameters. In particular, such as it can be frequently found in fault diagnosis
applications, the prediction of clogging has been faced as a binary classification problem
Search WWH ::




Custom Search