Multitier Dynamic Dual Kullback–Leibler Divergence Based on Manifold and Sensitive Features for Incipient Fault Detection

To solve the problem of incipient fault detection in an effective and timely manner, a multitier dynamic dual Kullback–Leibler divergence (MDD-KLD) based on manifold and sensitive features is proposed. First, the subspace highlighting the dynamic trend of the data is constructed using an exponentially weighted moving average (EWMA) window on the basis of manifold learning. Then, two sensitive detection indicators are constructed by leveraging the sensitivity of the KLD to minor changes in the probability distribution of the data, which is called dynamic dual (DD) KLD model. In addition, by calculating the KLD between normal data and priori fault data to determine the optimal sensitive variable subblock for a specific fault, the fault sensitive features are extracted to construct a complementary detection model, which supports the primary detection model, resulting in the MDD-KLD model. Finally, testing on an industrial case showed that the primary model attains 97% average accuracy for routine faults, while the complementary detection model further enhances both accuracy and fault detection time (FDT) for incipient faults relative to the primary model.