Identification of Industrial Alarm Floods Using Time Series Classification and Novelty Detection

Alarm flood classification (AFC) methods are used to support human operators to identify and assess recurring alarm floods in industrial process plants. State-of-the-art AFC methods, however, show shortcomings in handling an ambiguity of the activations and order of alarms and the detection of previously unobserved alarm floods. To solve these limitations, we present a novel three-tier AFC method that uses alarm series as input. In the classification stage, a linear ridge regression classifier with a convolutional kernel-based transformation (MultiRocket) is used to classify alarm floods according to their dynamic properties. In the detection stage, a novelty detection method based on the "local outlier probability" (LoOP) is used to decide whether an unknown alarm flood belongs to a known class or a novel one. Finally, we improve the classification results using an ensemble approach. Our proposed method is compared to two naive baselines and three relevant methods from the literature using a publicly available dataset based on the "Tennessee-Eastman" process. It is evident that our method shows the highest overall classification performance and robustness of all of the considered methods and effectively overcomes existing challenges in AFC.