The basic problem of classification priori unknown faults is related to rearrangement of existing classes and/or introduction of new classes that requires management of uncertain regions where input pattern vectors may belong to several classes. The Counter Propagation neural network (CPN) was selected to investigate the classification problems because it integrates both supervised and unsupervised learning to support diagnosis of both priori known and unknown faults. The CPN network is taught to have clusters that are described by codebook vectors in the training phase. In the basic CPN network the distribution of the codebook vectors is independent from the class homogeneity as a result of the Kohonen SOM unsupervised learning algorithm. Having unknown faults there are regions where the distribution of codebook vectors is in-homogenous like the class boundaries, i.e. the pattern vectors have a larger attraction power than in the homogeneous regions. To diagnose unknown faults the codebook vector distribution density should be increased in the inhomogeneous regions, i.e. in class boundary regions and decreased in homogeneous regions. The basic CPN algorithm was modified incorporating the class homogeneity to provide the rearrangement of codebook vector to manage uncertain regions and to diagnose priori unknown faults.