Data-driven process mining for production line optimization using IIOT and big data technologies

Abstract

In the face of increasing complexity in modern manufacturing, which is driven by rapid technological innovation, globalization, environmental sustainability requirements, and the growing demand for customized products, industrial organizations are transitioning toward smarter, more flexible, and data-driven production systems. Despite this shift, the vast volume of unstructured and high-frequency data generated at the shop floor level, originating from programmable logic controllers, sensor networks, manufacturing execution systems, and enterprise resource planning platforms, continues to present considerable challenges for process-level analysis and operational decision-making. This thesis addresses these challenges by proposing a hybrid methodology that integrates Industrial Internet of Things technologies and large-scale data processing infrastructures with process mining techniques. The aim is to convert raw machine-level signals into structured event logs, which can then be analyzed to model, evaluate, and optimize production processes more effectively. The proposed methodology is composed of four principal components: industrial data acquisition, data preprocessing, structured event log creation, and the application of multiple process discovery algorithms. It is designed to be scalable and adaptable across both discrete and continuous production environments. By applying this approach, the study aims to identify inefficiencies, detect behavioral anomalies, and uncover weaknesses in process control logic. To assess the practical applicability of the methodology, two contrasting industrial scenarios were selected as case studies: one representing a batch-based, semi-automated sauce production system and the other a continuous rubber compounding line. These cases were chosen specifically for their different production characteristics, offering a comprehensive view of the flexibility and limitations of process mining in diverse industrial settings. In the discrete sauce production scenario, digital process signals were captured from a simulated semi-automated line using the Message Queuing Telemetry Transport protocol. These signals were transformed into structured event logs using Python-based processing tools. The analysis focused on detecting structural inconsistencies such as skipped steps and improper event transitions. These issues were successfully identified through the use of the Log Skeleton and Alpha Miner algorithms. In addition, delay patterns were quantitatively evaluated using standard z-score calculations, which revealed significant bottlenecks particularly during the early phases of the production flow. These analytical insights were visually reinforced using ProsB, a custom-developed software tool that enables anomaly detection and performance profiling through graph-based process visualizations. In contrast, the rubber compounding process exhibited a highly consistent structural flow but demonstrated considerable variation in task durations. In this case, the data was collected using a Unified Namespace architecture, an Industrial Internet of Things framework that facilitated the hierarchical integration of operator inputs, machine state information, and enterprise planning data. Time-based performance analysis revealed that certain production sequences took significantly longer to complete, with delays observed in approximately ten to fifteen percent of the production cycles. These delays were not related to specific product types but were instead caused by irregular operator interventions and temporary reductions in machine efficiency. To better understand this behavioral variability, Fuzzy Miner was used to cluster similar execution patterns and identify outlier cases. This analysis demonstrated how process mining could go beyond structural modeling to provide insight into real-time operational behaviors and performance deviations. To further evaluate the effectiveness of the process discovery tools, four algorithms were compared across both case studies: Alpha Miner, Heuristic Miner, Fuzzy Miner, and All Operators Miner. The comparative analysis revealed that each algorithm had distinct advantages depending on the production scenario and data characteristics. Heuristic Miner produced clear and interpretable models for common process paths but lacked sensitivity to rare or noisy behaviors. Fuzzy Miner was more adept at abstracting complex and variable process flows, although it occasionally reduced model transparency. Alpha Miner, which builds formal Petri net representations, was useful for strict conformance checking but less suitable for capturing behavioral flexibility. All Operators Miner provided a deeper modeling capability for branching and probabilistic behavior but resulted in more complex visual models that may challenge interpretability. These findings suggest that a hybrid approach, tailored to the characteristics of each process and data type, is necessary for achieving optimal results in industrial process discovery. In summary, this research demonstrates that the integration of process mining techniques with Industrial Internet of Things-based data collection and big data analytics constitutes a powerful framework for increasing operational transparency, efficiency, and responsiveness in manufacturing environments. The hybrid methodology developed and validated in this thesis proved effective across both discrete and continuous production scenarios. It facilitated the detection of structural deficiencies, identification of performance bottlenecks, and classification of behavioral variability in real-world and simulated process settings. By enabling the transformation of raw production data into actionable insights, the study supports key goals of Industry 4.0, such as real-time monitoring, continuous improvement, and data-driven decision-making. The contributions of this work are both practical, in terms of industrial applicability, and theoretical, offering a methodological foundation for future research in intelligent process optimization and digital transformation.