Real-time classification and early warning of industrial alarm floods using modified TF-IDF methods

In complex industrial processes, the activation of a single alarm can trigger a cascade of consequences, affecting multiple interconnected components and rapidly increasing the number of active alarms. This phenomenon, known as an alarm flood, imposes a significant operational burden on operators by overwhelming them with high volumes of alarm notifications. This paper addresses these challenges by advancing real-time fault detection and monitoring systems. The contributions of this paper are twofold. First, it introduces a similarity analysis using a modified term frequency–inverse document frequency (TF-IDF) method to form clusters of alarm floods. Parameter optimization techniques are then applied to achieve the optimal n-gram size for improved clustering performance. Second, a novel method is proposed for real-time fault detection and a machine learning predictive model is further applied to assign real-time n-gram sequences to historical alarm floods, ensuring continuity in fault identification. Finally, an optimal alert threshold is designed to generate early warnings, enhancing operators’ responses to potential alarm floods and thus improving industrial process safety and efficiency.