From alarm floods to prescriptive insights: Leveraging smart data with a domain-agnostic Bayesian approach

In interconnected Cyber–Physical Systems, the big data paradigm is often at the basis of the overwhelming alarm floods phenomenon, which compromises effective alarm management by hindering root cause identification. Although a timely response to potential failures is essential, traditional approaches typically address alarm prediction and root cause analysis as separate tasks, thereby constraining the development of integrated diagnostic frameworks that support more effective decision-making. This work proposes and develops a domain-agnostic approach that simultaneously addresses both challenges, demonstrating the ability to transform alarm floods into a transparent “chain of evidence”. The approach leverages machine alarm signals as inherently interpretable data, transforming raw Programmable Logic Controller (PLC) alarms into temporally ordered sequences. Through Mutual Information and Chi-squared statistics for feature selection, the model isolates informative signals while reducing noise, then applies Bayesian Network structure learning to uncover causal relationships and enable probabilistic prediction of future target alarm occurrences. A case study on real-world data from an Italian brass manufacturing plant demonstrates strong predictive performance, with both numerical and experts’ validation confirming the inferred causal chains. This work contributes to a scalable, domain-agnostic approach for interpretable industrial AI that leverages smart data utilization, enhancing alarm management and operational resilience in smart manufacturing systems.