Addressing the challenges posed by high complexity, ambiguity in model structure determination, and overfitting during parameter learning in the soft sensor modeling of intricate industrial processes, this study intro...
详细信息
Addressing the challenges posed by high complexity, ambiguity in model structure determination, and overfitting during parameter learning in the soft sensor modeling of intricate industrial processes, this study introduces a recursive interval type-2 fuzzy neural network utilizing a logarithmic (log) t-norm (RIT2FNN-log). Unlike the conventional interval type-2 fuzzy neural network that employs the product t-norm, which often results in insufficient excitation intensity and inaccurate numerical calculations, the proposed logarithmic t-norm enhances excitation intensity and magnifies subtle differences, thereby enhancing network accuracy. The optimization of RIT2FNN-log involves two main stages: structural learning and parameter learning. Structural learning employs an arithmetic optimization algorithm with mean square error as the fitness function to determine the optimal number of rules and initial parameter values for RIT2FNN-log. Parameter learning utilizes a hybrid learning algorithm, with the antecedent parameters trained using a gradient descent algorithm based on AdaBound to prevent overfitting, and the consequent parameters trained using the leastsquaresmethod based on recursivesingularvaluedecomposition for rapid model convergence. The efficacy of RIT2FNN-log was demonstrated through applications in modeling tasks combined cycle power plant full-load output power prediction and Box-Jenkins time series analysis. Experimental comparisons with existing models revealed superior performance of RIT2FNN-log in terms of simplicity in network structure and lower mean square error.
暂无评论