‘Symbiotic’ data-driven modelling for the accurate prediction of mechanical properties of alloy steels

Abstract

A new optimal strategy based on symbiotic modelling is proposed. The system combines Linear Regression Model (LR), Non-Linear Iterative Partial Adaptive Least Square Model (NIPALS), Neural Network Model with double loop procedures (NNDLP), Adaptive Numeric Modelling (Neural-Fuzzy modeling NF) and metallurgical knowledge in order to provide effective modelling solutions and achieve an optimal prediction performance. As a final step a fusion procedure is used to perform a routine decision making based on aggregation algorithm and clustering method that allow to systematically select the final best prediction outcome from a set of competing solutions. The proposed methodology is then applied to the challenging environment of a multi-dimensional, non-linear and sparse data space consisting of mechanical properties of 'Mild' Steel in particular Tensile Strength (TS) and Yield Strength (YS) in hot-rolling industrial processes. Using a data set containing critical information on the mechanical properties obtained from a hot strip mill, it is concluded that the developed new systematic modelling approach is capable of providing better prediction than each individual model even in data distribution areas which are reckoned to be sparse.

Enhanced particle swarm optimisation algorithms for multiple-input multiple-output system modelling using convolved Gaussian process models

“…For example, a robot arm system with multiple degrees of freedom has multiple outputs that are strongly correlated. Another example is the prediction of steel mechanical properties in (Gaffour, Mahfouf, and Yang 2010), where the yield and tensile strength are predicted from the chemical compositions and grain size. Note that these two outputs are highly correlated.…”

Predictive modelling of the granulation process using a systems-engineering approach

“…Developing such a structure does not only consist of mapping the inputs to the outputs, but also discovering knowledge that may not be easy to extract by the already available approaches. The idea of the network relies on having a number of models with different structures, thus, (i) complex input/output relationships could be captured because of the number of functions and weights included [24], (ii) models with different structures could play a complementary role in modelling the possible patterns of the process, and (iii) training the data through two stages could help to extract the associated knowledge required for accurate property predictions [25]. Generally, an RBF network consists of three layers, namely; an input layer, basis functions acting as a hidden layer, and an output layer.…”

Particle swarm optimization for convolved Gaussian process models

“…For example, a robot arm system with multiple degrees of freedom has multiple outputs that are strongly correlated. Another example is the prediction of steel mechanical properties in [11], where the yield and tensile strength are predicted from the chemical compositions and grain size. These two "outputs" are highly correlated.…”