Aggregate system analysis for prediction of tardiness and mixed zones of continuous casting with fuzzy methodology

This thesis presents an aggregate system analysis with fuzzy methodology for interpretation, diagnosis and prediction of the behavior of the complex systems. The proposed systematic fuzzy modeling has three significant characteristics: (a) an improved fuzzy clustering approach with covariance-norm matrix, (b) an improved strategy for input variable selection and assignment of input-output membership functions, and (c) an appropriate parametrized reasoning mechanism. Initially, we surveyed the literature on fuzzy system modeling and discussed different approaches to fuzzy cluster analysis. Some of these procedures revealed shortcomings in with real-world data. Having developed the proposed model and its related algorithms, we tested it on four sets of data from real-world case studies. We found our approach better suited the real-world problems, including the interactions and correlations among complex sets of data and variables. It also presented a suitable strategy for determining the number of clusters and the degree of fuzziness of the system. We then introduced the index and methodology for significant input selection and assignment of input membership functions and considered possible correlations between input variables, using a Mahalanobis distance measure. The parametrized inference mechanism determined the actual parameters of the system based on the data. We tuned the input-output membership functions through a supervised-learning procedure to reduce the system's error. The proposed fuzzy methodology then was applied for system analysis, diagnosis and prediction of three complex problems in continuous casting: tardiness, mixed-zone effects, and total costs of tardiness and mixed zones. In each case, we compared the results with those of previous fuzzy models with identity-norm matrices and Euclidean distance measures and with a classical multiple-regression model. The results show that the proposed fuzzy methodology is superior with respect to identifying the critical rules, critical variables, and error minimization.