Operational Patterns Recognition Using Multi-class Cost-sensitive Learning for Alumina Evaporation Process

The characteristics of Alumina evaporating process (AEP) are analyzed firstly. Operational pattern is defined to describe the AEP. A new framework is proposed, which formulates the operational pattern recognition problem as a multi-class class-imbalanced problem of unequal misclassification costs. Aim to the multiclass class-imbalanced problem of unequal misclassification costs in the operational pattern set of AEP, a multiclass cost-sensitive Probabilistic neural network (mc-PNN) method is proposed. A spectral clustering based on Balanced iterative reducing and clustering using hierarchies (BIRCH) is used to optimize the number of pattern layer nodes of mc-PNN. Experimental results show that the proposed method reduces effectively average misclassification costs and increases recognition rate of excellent and faulty classes.