The regions with uncertain decisions are typical in fault diagnosis. These regions appear as a result of model mismatches, noises and unknown inputs. To increase the efficiency of fault diagnosis it is required to improve classification accuracy in these regions. The homogeneity distribution of the codebook vectors is a key element in the accuracy of the classification process. The paper defines an appropriate homogeneity measure that is strongly correlated with the optimal misclassification error. The classification process of the Counter Propagation neural network (CPN) is investigated. Based on this homogeneity value, the paper proposes two modification algorithms for the original CPN classification algorithm to reduce the misclassification error in the regions of uncertain decisions. The accuracy of the proposed algorithm is tested with a case study. The first method alters the learning rate using the homogeneity value of the region. The second method generates an R-tree decomposition of the input space and performs redistribution of the codebook vectors. Both methods provide significant improvement in classification.