Klasyfikacja stanów przedkrytycznych

Praca zawiera przykład klasyfikacji danych rzeczywistych opisujących sygnały niekrytyczne, przedkrytyczne i krytyczne. Celem jest rozpoznanie stanów niebezpiecznych tak wcześnie jak to możliwe. Ze względu na brak separowalności liniowej danych w celu separacji klas użyto klasyfikacji hierarchicznej z cięciami za pomocą klasyfikatorów liniowych oraz z podejściem one-versus-rest z wyróżnioną klasą sygnałów bezpiecznych. W wyniku ośmiu cięć uzyskano ostateczny podział przestrzeni skutkujący odseparowaniem klasy sygnałów bezpiecznych od podejrzanych, tj. przedkrytycznych i krytycznych oraz dający najmniejszą liczbę błędnie sklasyfikowanych obiektów z klasy sygnałów niekrytycznych.

The paper presents an application of classification methods to time-continuous signals (1). Signals with values that exceed a certain critical maximum are called dangerous or critical, otherwise we speak about normal or routine operation of the system under consideration, Fig. 1. The problem is to recognize pre-critical states, i.e. states preceding the actual dangerous ones, and that as early as possible. False negative classifications may have very serious consequences, while false positive verdicts cause expensive but unnecessary counter-measures. As pre-processing, the input signals are characterized by a number of features, which form sequences of vector data, indexed by the cycle number (2). In a first stage, suspicious feature vectors are selected, from which in a second sweep unlikely candidates are removed. The focus of the present paper is this second stage, i.e. the distinction between actual pre-critical and the harmless routine states among the suspicious states, indicated in the first stage by a certain preliminary test. The choice of features and the logic behind the preliminary test are beyond our present scope. Let it suffice to say that the first step is a combination of Principal Component Analysis and some statistical test, and that it is very effective but unspecific in the application at hand.For the real-world data we used to develop the method, it turned out that the obtained feature vectors were linearly non-separable. For that reason a hierarchical approach was applied, where in several steps linear cuts (4,5) of the one-versus-rest type were performed in order to single out the true pre-critical states. For the example under consideration, in eight iterations separation between pre-critical and non-pre-critical ones was achieved. We succeeded to keep the number of wrong negatives at zero while reducing the number of wrong positives to a fraction of the starting value, established by the preliminary test, Fig. 3, 4, 5. The final sensitivity, for the given data set, is 100%, and the achieved specificity is at 93.15%. Numerical experiments, using nonlinear classifiers on much larger data sets, are under way. The present aim is to find an optimal set of features and a one-step criterion which further improves the quality of the classification.