A research prototype for process model analysis and reference model mining. This webseite is currently under construction. Thank you for your patience.
Run RefMod-Miner as a Service now!

Deep-Learning-Based Process Prediction

Basic concept:

Process management at runtime entails enormous application potentials for business processes. For example, customer service representatives can quickly respond to inquiries as to when a particular case will be completed or resolved. In this way, processors can identify process instances with a high delay or failure probability and react at an early stage to avoid risks. Process prediction is an important aspect of process management at runtime. A prediction can refer to either the output of the process or the events following the process. Examples of relevant process outputs are the final state (e.g. “customer request is accepted” or “customer request is rejected”), case data (e.g. the “cost” attribute is less than a certain value), or LTL conformity formulas (e.g. “requests approved” occurs before “write a check” and the activities were carried out by different employees). Predictions can be based on the sequence of activities of the current process instance, the collected case data, the resources involved in the activities, the execution times of these activities, or any other case or workflow data available at runtime [1,2].

In the RefMod-Miner, an approach for predicting the next process steps is implemented using methods of deep learning. A recurrent neural network (RNN), which is trained on past process data, is used. For any running process instance, this network is able to predict the subsequent process steps as well as their expected execution times.


How the RefMod-Miner works:

The described approach to process prediction is implemented with the Showcase of the Smart LEGO® Factory. This presents the use of business process management methods for the industry 4.0 in an innovative scenario. A flexible production process for the production of tractors in different variants is presented. The model offers realistic design possibilities for the product and also integrates people into the workflows, through manual work steps. The production process is initially instanced in the form of a process model based on a descriptive production and product model. This is embedded in a complete model for value creation within the Smart LEGO® Factory, based on the specific product. From this, the fully automated control and monitoring of the real value added in a cyber-physical system starts, from the parts supply to the actual production up to the quality control at the end of line and the warehouse logistics. start fragment




Through the above-described approach for the prediction of process data by means of deep learning, the next process steps, the running time of the current process step and the running time of the entire process instance are predicted during the ongoing production process of the Smart LEGO® Factory for the current process instance. Here, a recurrent neural network, which has been trained on past process executions, is used. In this way, errors can be avoided and proactive responses to possible errors as well as the prognosis of subsequent work steps can be enabled. The prognosis of the associated execution times allows for an improved planning of production processes. in doing so, new concepts of artificial intelligence are combined with methods of business process management. You can see the showcase at the CeBIT from the 20th to the 24th. March in Hanover at the booth of Software AG (Hall 4. booth C11).


[1] Vgl. Joerg Evermann; Jana-Rebecca Rehse; Peter Fettke: A Deep Learning Approach for Predicting Process Behaviour at Runtime. In: Marlon Dumas; Marcelo Fantinato (Hrsg.). Proceedings of the 1st International Workshop on Runtime Analysis of Process-Aware Information Systems. International Workshop on Runtime Analysis of Process-Aware Information Systems (PRAISE-2016), located at International Conference on Business Process Management, September 18-22, Rio de Janeiro, Brazil, Springer (2016) 

[2] Vgl. Joerg Evermann; Jana-Rebecca Rehse; Peter Fettke:Predicting process behaviour using deep learning. CoRR abs/1612.04600 (2016)