It is impossible to define all faults in the design phase. As a result, a priori unknown faults may appear in systems that must be managed by fault diagnosis. A priori unknown faults modify the distribution of the input patterns and the homogeneity of assignments of input patterns to the output space. The change in distribution of input patterns may modify clusters and cluster class assignment. To identify changes in distribution of input patterns a proximity measure and in cluster class assignment a homogeneity measure is used, respectively. After appearing a priori unknown faults the RBF neural network is trained on-line using a modified supervised-unsupervised learning algorithm taking into account the proximity and homogeneity values.