Hierarchical diagnostics of analog systems based on the ambiguity groups detection

The paper presents the hierarchical approach to detect and identify faults in the analog system using combined Artificial Intelligence (AI) methods. The automated diagnostic system has two levels of fault identification, based on the unsupervised and supervised learning. The former is used in the initial stage to separate easily identifiable states of the analyzed system from the difficult ones. The latter are identified with the more sophisticated classifier. Because the difficulty of the fault identification is related with the existence of Ambiguity Groups, the Unsupervised Learning scheme is employed to detect them and decompose training data set into subsets, on which two stages of classifiers are trained. The first set (considered “simple”) is processed by the simpler machine learning algorithm. The second set is used to train the more complex classifier (operating in the uncertainty conditions). The proposed scheme is generic, therefore various algorithms can be implemented. In the presented case, the Self Organizing Map (SOM) is used in the first stage, while Random Forest (RF) - in the second one. To verify the approach, the 3rd order Bessel highpass filter was analyzed. The architecture was confronted against the traditional approach (where the standalone classifiers are employed). Results confirm usefulness of the proposed solution, regarding the higher classification accuracy and smaller computational effort than its alternatives.