Diagnosing a priori unknown faultsby modified supervised-unsupervised learning algorithm

Experts should analyse systems in order to define would-be faults in systems. As a result of this analysis, there will be a set of priori known faults supporting off-line teaching of neural networks. Unfortunately, it is impossible to define all faults in the design phase. As a result, a priori unknown faults may appear in systems. 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. It may either rearrange existing clusters moving their borders or add a new cluster to the existing clusters and re-arrange clusters around the new cluster. The change in homogeneity of assignment may affect the class-cluster assignment. It may either re-assign clusters to the existing classes or may add new class and reassigns clusters to the existing classes and the new class. To identify changes in distribution of input patterns and in assignment of features and classes the average proximity and the class homogeneity is used, respectively. After appearing a priori unknown faults the RBF neural network was upgraded using a modified supervised-unsupervised learning algorithm taking into account the distribution and homogeneity values.