Neuro-fuzzy Approach - Computational Intelligence in Time Series Forecasting - page 259

x monitoring of tool wear.

6.8.1 Application of Neuro-fuzzy Modelling to Material Property Prediction

Chen and Linkens (2001) have proposed a systematic neuro-fuzzy modelling
framework with application to mechanical property prediction in hot-rolled steel.
Their methodology includes three main phases:

x the initial fuzzy model, which consists of generation of fuzzy rules by a

self-organizing network

x the second phase, which includes the selection of important input variables

on the basis of the initial fuzzy model and also the assessment of the
optimum number of fuzzy rules (hidden neurons in the RBF network) and
the corresponding receptive fields determination via the fuzzy c -means
clustering algorithm

x third phase, dedicated to the model optimization, including parameter

learning and structure simplification on the basis of backpropagation
learning and the similarity analysis of fuzzy membership functions.

Thereafter, the neuro-fuzzy model developed is used to predict the tensile stress,
yield stress, and the like in materials engineering.
In materials engineering, property prediction models for materials are important
for design and development. This has for many years been an important subject of
research for steel. Much of this work has concentrated on the generation of
structure - property relationships based on linear regression models (Pickering,
1978), (Hodgson, 1996), developed only for some specific class of steels and
specific processing routes. Recently, some improved, neural-networks-based
models have been developed for prediction of mechanical properties of hot-rolled
steels (Hodgson, 1996), (Chen et al. , 1998), and (Bakshi and Chatterjee, 1998).
Using complex nonlinear mapping, the models provide more accurate prediction
than traditional linear regression models. But the drawback here is that the
development of these kinds of model is usually highly problem specific and time
consuming, so that the development of a fast, efficient, and systematic data-driven
modelling framework for material property prediction is still needed.
The problem of modelling of hot-rolled metal materials can be broadly stated as
follows. Given a certain material which undergoes a specified set of manufacturing
processes, what are the final properties of this material? Typical final properties, in
which metallurgical engineers are interested, are the mechanical properties such as,
tensile strength ( TS ), yield stress, elongation, etc. Chen et al . (2001) have
developed a neuro-fuzzy model for the prediction of the composition-
microstructure-property relations of a wide range of hot-rolled steels. More than
600 experimental data from carbon-manganese (C-Mn) steels and niobium micro-
alloyed steels have been used to train and test the neuro-fuzzy model, which relates
the chemical compositions and microstructure with the mechanical properties.
In the experimental data set, they have considered 13 chemical compositions,
two microstructure variables, and measured tensile stress values, which
corresponds to a system with 15 possible input variables and with one output