Introduction
Discrete-event simulation (DES) is widely used in various industries for planning production and logistics systems, allowing users to explore “what-if” scenarios. While its use during commissioning or operation is less common, it offers significant benefits. This paper distinguishes between technical and control systems, using emulation to test real control systems within simulated technical systems. DES can also be combined with process mining to streamline model creation and generate synthetic data, enhancing its value in production and logistics.
Process mining, situated at the intersection of data and process science, analyzes processes using past operational data. It includes types like process discovery, conformance checking, and action-oriented process mining. Combined with discrete-event simulation (DES), it provides insights into past and future processes.
Action-oriented process mining generates actions based on diagnostic insights, bridging the gap between insight and action. Its implementation in production and logistics is under-researched and risky, as failures can impact operational performance. Discrete-event simulation (DES) can mitigate these risks by generating event streams for different scenarios, helping to analyze and predict necessary process control actions.
Simulation model
During the execution of a simulation model, a simulation-generated event stream is created whose events are recorded by an event monitor and evaluated with regard to their temporal and behavioral conformance.
Temporal constraints can be formulated, for example, by target dates that are set for the completion of an order. The behavioral conformance is assessed by analyzing behavioral patterns (i.e., a set of activities and possible control flow relations) and a reference behavioral model. This is shown in the model below.
This detailed view of the control flow enables real-time analysis of specific behavioral patterns during process execution. It allows for the calculation of three key indicators: conformance (correctness of behavior), completeness (case progression), and confidence (stability of conformance).
Checked events are added to the event stream and sent to the simulation action controller. Based on the event’s status, an appropriate event routine is chosen and communicated to the simulation model. If an event deviates from expected behavior, a specific routine is automatically selected to handle it; otherwise, the default routine is executed. The described workflow is shown below.
The simulation action controller’s decisions are observed throughout the simulation runs. Event streams with non-conforming events can be generated to trigger automatic actions by the controller, and their effects are evaluated within the simulation.
This approach allows thorough testing of the action controller in a simulated environment before real-world implementation, mitigating risks to process sequences. Simulation results can be statistically validated and capture rare events. The number of simulation runs depends on the use case, ensuring edge cases are considered. This method enables action-oriented process mining without altering the real system, using simulation as an emulation model to reduce risks.
Use case
The application system is a conveyor system on a university laboratory scale. The goal of the use case is to illustrate how the approach can handle incorrect processing sequences (behavioral non-conformance) and processing delays (temporal non-conformance) and how automated actions can address them.
The architecture is implemented using AnyLogic 8 and open-source process mining software. A simulation model and experiment template are built, leveraging AnyLogic’s Reinforcement Learning (RL) template to define observations, actions, and configurations, adapting it for the simulation context.
The RL pipeline is repurposed for simulation-enhanced, action-oriented process mining. The simulation model is exported to a standalone Java model and linked to a Python Jupyter Notebook using the Alpyne library. Key events trigger actions, pausing the simulation to pass events to the Python controller, which logs and checks them for conformance using a Pandas DataFrame.
Streaming conformance checking systems are limited, but frameworks like PM4Py and Beamline offer solutions. This work uses pyBeamline for behavioral conformance checking, comparing new events against a normative model. Events are logged and checked for conformance, with results guiding the simulation action controller. Non-conforming events trigger specific actions, while conforming events follow default logic. This method allows thorough testing of process mining components in simulations before real-world implementation, reducing risks.
Results
This article introduces a method combining discrete-event simulation and process mining, using a simulation model as an emulation tool to implement action-oriented process mining.
This approach helps mitigate risks in automated process-regulating actions within production and logistics systems. While the current results are promising, they serve as a starting point for further research, particularly in enhancing dynamic action spaces and improving data quality.
Future studies are needed to demonstrate the quantitative benefits in more complex settings.