Alarm Root-Cause Analysis Using an Alarm Logic Directed Graph Extracted From Control Software

The growing complexity of modern automated production systems demands solutions for managing alarm floods potentially stemming from multi-root causes, while improving situational awareness of operators to ensure timely fault responses and minimize downtime. For this purpose, knowledge-driven approaches that utilize domain specific knowledge for alarm handling have shown promising results. However, knowledge extraction is expensive, leaving data-driven approaches as an alternative, which in turn requires large amounts of data. To profit from both approaches and alleviate their weaknesses, hybrid approaches have been a focus of recent development. Therefore, this work proposes the Alarm-Logic-Directed-Graph for a novel hybrid approach. This graph models the alarm logic inherent in the control code, enabling knowledge-based clustering and sorting of the alarm log, which then serves as input data for a subsequent Bayesian network learning. Moreover, with the integration of physical and heuristic knowledge, the root causes can be further refined to alarm-related component status variables. The evaluation of this work is conducted on the Tennessee-Eastman Process, a well-known continuous process simulation. The results show that the proposed approach remains robust to multi-root cause identification even in the absence of historical alarm log data or process knowledge besides programmable logic controller code.