Fault Diagnosis using Deep Neural Networks for Industrial Alarm Sequence Clustering

Significant progress has been made in the field of industrial alarm management
systems (AMS) in terms of diagnostic and prognostic accuracy. However, persistent
challenges, such as poorly configured alarm setups and floods, contribute
to an increased number of false alarms, consequently reducing the efficiency of
the monitoring system. In addition, more sophisticated models and interactive
visualization tools are needed to support supervisors and maintenance operators.
This paper proposes a novel approach based on deep learning that combines
autoencoder and self-organizing maps to extract valuable features and a clustering
algorithm to identify related alarm groups. This bi-level methodology is
applied to real manufacturing system datasets, demonstrating its effectiveness in
identifying false alarms, reducing alarm sequence interpretation time, enhancing
understanding of alarm interrelationships, and providing a basis for causal analysis
and root cause identification. The approach also compares favorably with
the classical methods in the literature, laying the foundation for improved industrial
safety management. The system also offers maintenance recommendations
to decision makers, further validating alarm sequences.