PROCESS MINING FOR REENGINEERING CIRCULAR AND RESILIENT PRODUCTION PROCESSES

This thesis explores the integration of process mining techniques with deep learning models to enhance the understanding, analysis, and optimization of complex business processes. Process mining bridges the gap between model-driven and data-driven approaches through a set of techniques that extract insights from event logs. Meanwhile, deep learning has the potential to model and learn intrinsic patterns and dependencies from data. By combining these fields, the research carried out throughout this thesis addresses key challenges in a specific subfield of process mining called predictive process monitoring. In this subfield, the goal is to leverage all the historical data to predict future behaviors within a process. Relevant predictive tasks include, for example, predicting the next activity, the remaining time of a current instance, or the outcome of a current instance. In the past few years, several papers have been published on predictive process monitoring regarding different tasks. Despite the great success and convincing results in terms of predictive accuracy, researchers overlook some principles that are crucial for the advancement of the field. For instance, transforming business and semantic rules from process data into a latent space is a key step for training machine learning models. Usually, most proposals in the literature rely on naive techniques to encode categorical attributes from events, such as the well-known one-hot encoding. Furthermore, a second level of encoding is required to aggregate a trace of events into a single vector array, which is commonly done by simply averaging the attribute values of events. These approaches naturally result in loss of information and fail to properly capture all the intrinsic complexity of process data. This often-overlooked step introduces significant limitations to modern predictive process monitoring applications. For example, while deep learning models excel at tasks such as predicting the next activity or suffixes in a process, these tasks often fall short of addressing more meaningful, practical needs. In real-world scenarios, the usefulness of next-activity predictors is often questionable. A more impactful application lies in process simulation, where the goal is to model potential new behaviors within a process before actual changes are implemented, ensuring safety and minimizing risk. Unfortunately, current techniques for representing process data are inadequate, making deep learning models inflexible to effectively support this type of simulation. For example, process data is often encoded as sequences of discrete events without capturing the contextual relationships or dependencies between them. This limitation means that a deep learning model trained on such data may accurately predict the next event in a process but cannot simulate complex scenarios, such as how changes in resource allocation or timing constraints might affect the overall workflow. Therefore, in this thesis, we aim to better understand how to encode process data in latent spaces to improve the performance of learned models. We begin this work by reviewing several encoding techniques from other research areas to assess their applicability in process mining. We also investigate and motivate the development of feature engineering techniques tailored to this particular type of data. To this end, we systematically design and empirically evaluate extensive experiments for different tasks. In addition, we propose the use of declarative languages that capture linear temporal logic over finite traces to improve the data representation of process data for training deep learning models, thus overcoming the inflexibility of these applications for process simulation. Finally, we use all the knowledge and contributions developed in this thesis to propose meaningful solutions in the industrial sector. More specifically, we develop three relevant use cases at an Italian company called Avio Aero: (i) a process mining pipeline to extract insights from their data, (ii) an application of large language models to enrich this data, and (iii) a combination of both research areas to propose improvements and design a more resilient and robust process model.