Digital twins are based on physical and functional information about components, products and systems which are necessary for production processes in each phase of the life cycle. One key aspect of a digital twin are the simulation models that extend data already available in different life cycle phases, such as design, engineering, operation, and service . This article describes the application of a digital twin in one use case from the operation phase of an automated Production System (aPS): industrial alarm management.
The concept presented in this article was developed in the project “Innovative Modeling Approaches for Production Systems to Increase Validatable Efficiency“ (IMPROVE). IMPROVE is sponsored by the European Union and focuses on virtual Factories of the Future (vFoF). Data-driven and model-based digital twins are part of a holistic solution for Self-X technologies, including diagnosis and optimization of industrial components, machines and plants . Self-X technologies help to increase efficiency, reduce the frequency of failures and lower running costs. Fraunhofer IOSB-INA participates in IMPROVE with Ostwestfalen-Lippe University of Applied Sciences and 11 other partners from academia, industry, and software development.
One of the main use cases in IMPROVE is intelligent alarm management which focuses on analyzing alarms and warnings generated during plant operation. Unfortunately, many alarms that are displayed to the operator are either redundant alarms or nuisance alarms, such as chattering or lingering alarms. Eventually, the number of alarms becomes so high that it overwhelms the machine operator.
Alarm flooding is a persistent problem in industrial plant operation . It occurs when the frequency of alarm announcements becomes so high that the operator is overwhelmed and loses sight of how to solve the situation. In the worst case, that leads to critical alarms being overlooked and time-consuming searches for the root cause of the problem. This, in turn, may result in dangerous situations, significant downtime and even irreversible damage, such as the infamous explosion at a Texaco refinery which was found to be caused by a flood of alarms .
The goal of an intelligent alarm management system is to avoid alarm flooding and support the operator if it nevertheless occurs. The main reason for alarm flooding is a flawed alarm system design. As revamping the alarm design of a running system is impossible and stopping production impractical, other solutions are required. Traditionally, alarm management has employed simple methods, such as basic signal and alarm filtering to remove alarms before they are displayed to the operator, or giving operators the option of shelving alarms they consider irrelevant or redundant. For anything more advanced than traditional methods, a deep understanding of the system – expert knowledge – is needed. This is either very difficult or too time consuming to obtain. Now, the rise of machine learning and data-driven computational intelligence allows us to consider more complex and intelligent approaches to alarm flooding – by utilizing a digital twin of the system. Intelligent approaches allow us to either reduce the number of alarms or assist the operator in identifying the root cause.
In IMPROVE, the basis for the implementation of a digital twin for industrial alarm management is a simulation environment that uses the PhysX engine for discrete event simulation. The PhysX engine controls the behavior of basic physical elements, for example tracks and conveyor belts or discrete loads representing manufactured goods. The design of the simulation model is enriched with the custom behavior of production units and sensors, definitions for raising alarms based on sensor values as well as a flood detection system. Discrete event simulation allows the user to greatly speed up the flow of time and observe long-term behavior of the plant model – and record the data. Inducing failures in specific modules allows us to collect a case base of alarm flood samples with semantic annotations. Fig. 1 shows an example of such a model, simulating a scenario where a fault induced in one module triggers an alarm flood. A simulated case base is a valuable tool which can be analyzed and the results applied to the real plant. The raw case base is itself a model of faulty behavior in the plant and can be analyzed using data mining approaches in order to gain insights into recurring problems in the plant. Moreover, it can be used to directly support the operator during an alarm flood in a running plant. Machine learning approaches can find similar, previously seen and annotated cases to suggest possible solutions to the operator . Furthermore, a case base can also be used to learn a variety of further models – such as causality models based on Bayesian networks or Markov chains. Such models shed light on dependencies between alarms, which can be used to reduce the number of alarms displayed.
Fig. 1: An example of a simulation model of the SmartFactoryOWL Versatile Production System demonstrator behavior. A fault induced in Module 2 results in a production stop and an alarm flood.
The combination of the simulation model with the embedded behavior and the model-learning and analysis approaches constitutes the digital twin (illustrated in Fig. 2). In the context of industrial alarm management, the digital twin is used to construct a database of alarm floods. This is useful for analyzing the overall behavior of the alarm system as well as supporting the operator during plant operation.
Fig. 2: The concept of a digital twin for industrial alarm management: a simulation model generates data used for case base construction, machine learning and analysis to support the operator of the production system.
 Boschert S.; Rosen R.: (2016) Digital Twin – The Simulation Aspect. In: Hehenberger P.; Bradley D.: (eds) Mechatronic Futures. Springer, Cham
 IMPROVE project website: http://improve-vfof.eu/
 Health and Safety Executive (HSE). The Explosion and Fires at the Texaco Refinery, Milford Haven, 24 July 1994 (Incident Report). HSE Books, 1997.
 J. Wang; F. Yang; T. Chen and S. L. Shah: "An overview of industrial alarm systems: Main causes for alarm overloading, research status, and open problems," IEEE Transactions on Automation Science and Engineering, vol. 13, pp. 1045–1061, April 2016.
 M. Fullen; P. Schüller and O. Niggemann: Semi-supervised case-based reasoning approach to alarm flood analysis. In: Machine Learning for Cyber-Physical Systems and Industry 4.0 (ML4CPS). Lemgo, 2017.