Optimal input selection for neural fuzzy modelling with application to Charpy energy prediction

Input variables selection plays a critical role in data-driven modelling, especially for complex systems with high dimensionality between the input/output space. In this paper, a new artificial neural network based forward input selection scheme is proposed. The objective of the proposed scheme is to select the smallest number of important variables as model inputs, which will then be used for neural-fuzzy data modelling. The proposed input selection scheme is applied to a case study of Charpy impact energy prediction, with data extracted from an industrial database. Model performance has been compared with previous results where a much larger input set was used. Simulation results show that the number of inputs for the Charpy data model can be significantly reduced with little performance degradation. Also, the performance of the proposed scheme outperforms both the standard correlation analysis and fuzzy clustering based input selection schemes.