Incipient Fault Detection Method Based on Probability-Related Dynamic Slow Feature Analysis

To accurately identify true faults and distinguish them from normal deviations in operating, this paper proposes a novel probability-related-based dynamic slow feature analysis fault detection method. Given that process data exhibit both static and dynamic characteristics, the process data are first dynamically expanded to construct an augmented dataset reflecting the dynamic properties of the process. Subsequently, slow feature analysis is applied to extract slow features that represent the static characteristics of the process, while their rate of change is used to measure dynamic characteristics. Additionally, to further evaluate both static and dynamic characteristics of process data, the Kullback-Leibler divergence is employed to quantify the probabilistic distribution differences of slow features and their rates of change, thereby constructing a fault detection statistic capable of distinguishing between process faults and operational deviations. Simulation experiments based on the Tennessee Eastman process verify the effectiveness and superiority of the method.