Online alarm flood classification via interpretable template extraction and structured convolutional matching

Abstract

Alarm flood classification in industrial alarm systems is a challenging task due to variability in fault durations, process noise, and the volume of overlapping alarms. However, alarm floods triggered by similar faults often exhibit recurring structural patterns, which, if identified effectively, can support the root cause diagnosis and informed decision-making by operators. Existing classification methods often rely on opaque models, extensive retraining, or lack integration with operator-facing tools. Motivated by this practical problem, a unified visual analytics-based methodology for the real-time classification of alarm floods is proposed in this paper. The contributions are threefold: (1) The existing High-Density Alarm Plot (HDAP) is extended into a structured matrix representation to encode alarm activity over time; (2) a 2D convolution-based alignment technique is developed to extract representative templates from historical alarm floods, enabling category-specific pattern generation; and (3) a dynamic matrix representation is introduced to support real-time alarm monitoring, where similarity matching against pre-learned templates facilitates online classification with minimal delay. The proposed method is interpretable, operator-friendly, and seamlessly integrates with existing visual tools. The effectiveness of the proposed method is validated on the Tennessee Eastman Process benchmark, demonstrating robust and accurate early-stage classification.