Open Collections

Abstract

In the era of big data, driven by the advent of the Internet of Things, process industries face the challenge of analyzing massive and complex data to extract relevant information for effective process monitoring. Leveraging these data is critical for ensuring safety, optimizing performance, and maintaining competitiveness. As a result, data-driven process monitoring has received significant research attention and achieved remarkable progress. However, limitations such as limited interpretability, reliance on extensive labelled datasets, and challenges in translating research findings to practical settings persist. This dissertation introduces a novel paradigm called visual analytics. Visual analytics transforms temporal process data into visual formats to uncover hidden patterns. It bridges the gap between traditional process monitoring methods and modern data-driven approaches by reframing process monitoring problems as computer vision tasks, such as image classification. This paradigm emphasizes interpretability, allowing process experts to relate visual patterns to operational conditions and, consequently, supporting more informed decision-making. This dissertation explores three pathways within the visual analytics paradigm: (i) feature engineering, where predefined mappings are used to convert time series data into visual representations; (ii) architecture engineering, which develops neural network architectures to directly learn visual representations; and (iii) data engineering, which employs contrastive learning to highlight differences and similarities in the data without relying on annotations. The proposed methodologies are evaluated using two benchmark datasets. The first is the simulated continuous stirred tank heater dataset, which provides a controlled setting for testing the proposed methods. The second is the industrial Arc Loss dataset, which includes one year of operating data from a 60,000 ton/year pyrometallurgical plant, used to demonstrate the robustness and scalability of the proposed methods in real-world scenarios. We conduct comprehensive experiments to compare the proposed methods with state-of-the-art techniques from both traditional and advanced process monitoring approaches. The evaluation focuses on fault detection performance, qualitative interpretability, and challenging scenarios, such as limited labelled data and transfer learning. Experimental results highlight the effectiveness of the proposed visual analytics frameworks, demonstrating competitive performance across all criteria. In addition, the proposed approaches provide informative visual representations, which enhance interpretability and facilitate improved process monitoring and decision-making.