Towards NLP-driven Online-classification of Industrial Alarm Messages

Industrial alarm systems facilitate the reliable and effective management of industrial process operations. However, the manufacturer’s fear of missing a critical fault or lacking process understanding may lead to the implementation of alarm systems that overburden operators by an enormous amount of alarms. Inadequate alarm settings and maintenance, coupled with a high volume of disturbance alarms, necessitate operators making crucial decisions within a limited time. Thus, an alarm management strategy that accurately anticipates the types of incoming alarms is needed, especially in brownfield systems where it is infeasible to simplify the underlying alarming functions. This alleviates the need for human interpretation of alarms upon arrival, allowing operators to address anomalous behaviors early. This study proposes an alarm classification approach, based on active learning, latent dirichlet allocation (LDA) topic modeling and transformer-based deep learning. It focuses on addressing these challenges through a systematic natural language processing (NLP) approach to classify and map fault messages. To demonstrate the validity of this approach, real-world alarm data from an industrial hybrid non-woven manufacturing process is used.