Minimal-invasive methods for monitoring and supervision of industrial processesSTRŽINAR, ŽIGA (Avtor) Škrjanc, Igor (Mentor) Pregelj, Boštjan (Komentor) Time series analysisClusteringClassificationEvolving methodsIndustrial processesMonitoringThis thesis deals with the development of minimally invasive methods for the supervision and monitoring of industrial processes, focussing on time series analysis techniques. In the context of Industry 4.0, where industrial machines are increasingly equipped with networking and data acquisition capabilities, efficient machine supervision is an important aspect of efforts to improve the efficiency and reliability of production lines. This dissertation addresses the need to equip existing machines with low-cost, non-invasive monitoring hardware that -- supported by the state-of-the-art algorithms presented here -- enables effective monitoring of such machines. The dissertation proposes a processing pipeline consisting of data acquisition, segmentation, clustering, classification and sequence analysis. The proposed methods take into account the ever-changing nature of industrial processes and the limited availability of labelled time series from the past. To overcome these challenges, evolving mechanisms are incorporated into several of the algorithms. The thesis consists of five publications, all focussing on time series analysis. Several algorithms are presented in these publications -- two for time series classification and three for clustering. Four of the publications deal with pneumatic signals, which are common in industrial environments. A dataset of such signals has been published and made available for wider use by the research community. One publication diverges into the field of biomedical engineering and applies time series classification methods to measurements of electrodermal activity in order to recognise stress in humans. Several data sets are used to evaluate the proposed methods. Given the application focus, particular importance is placed on an industrial data set for validation. Nonetheless, the methods are also evaluated against a broad collection of publicly available time series datasets, including the University of California, Riverside (UCR) Time Series Archive and the Wearable Stress and Affect Detection (WESAD) dataset. The evaluation includes accuracy metrics, time complexity analyses, visual comparisons of results, and comparisons of computation times across multiple scenarios. The methods developed and presented in this thesis clearly show that non-invasive measurements can be used effectively for the supervision of industrial equipment. Furthermore, methods that support unsupervised learning are particularly well suited for industrial applications where large amounts of unlabelled data are common. The work presented is an important step towards realising the vision of data-driven manufacturing.20252025-12-05 08:30:02Doktorsko delo/naloga176603sl