A Shannon entropy-based conflict measure for enhancing granular computing-based information processing

One of the aims of the human-like computational paradigm of granular computing (GrC) is to discover - and capture - high-level knowledge from raw data in the form of information granules. In real-world applications, information is often associated with uncertainty due to factors such as measurement imprecision, low process repeatability, as well as sparse data measurements. Hence, data uncertainty should be taken into consideration while carrying out information granulation. In this paper, a framework of iterative information granulation is presented that for the first time in the literature is enhanced with measures of uncertainty during the granulation process. A special case study is investigated, one of a complex manufacturing processes and associated information consisting of sparse, often conflicting data (measurements). An algorithmic procedure is proposed for quantifying the uncertainty caused by conflict during the iterative information granulation process; this is achieved via the use of the Shannon entropy theory to capture uncertainty during the iterative merging of the information granules. The resulting conflict measure is used to `guide' the granulation process into solutions of low-conflict granules. The result is an enhanced granular information set, in terms of distinguishability and interpretability. The granular data set is then `mapped' into the linguistic terms of a Neural-Fuzzy (NF) rule-base to form a model that represents the process under investigation. The proposed GrC-NF methodology is applied to the complex manufacturing process of Friction Stir Welding (FSW) of steel. It is shown how the proposed framework is able to capture good quality information granules from raw process data, which are then mapped into a Neural-Fuzzy model that predicts the resulting torque on the FSW tool with more than 90% accuracy.