Process-aware pattern recognition and deviation detection under uncertainty

Abstract

Humans naturally follow specific patterns in their daily activities, whether driven by best practices, physical layouts, social norms or established processes or routines. Monitoring and understanding these human activities is a challenging problem, with the ultimate goal of identifying anomalies from desired behaviour, such as those related to healthy living or efficient working. Recognising process patterns and routines, and detecting anomalies can provide crucial insights. However, this presents significant challenges across different domains. Varying structures of processes in different domains complicate the task of capturing and analysing these patterns, as well as understanding the multifaceted nature of anomalies. Some domains are structured, like manufacturing, therefore it is easier to detect workflow patterns and deviations from them, but others are semi-structured or unstructured, like daily living. In the latter, even defining what an anomaly means is a challenge. Moreover, data are often filled with noise and uncertainties, which may lead to misinterpretation of normal occurrences as anomalies.

To address these challenges, we first discuss the characteristics of different types of domains, i.e., structured, semi-structured, and unstructured domains. We delve into the diverse sources of anomalies across different domains, which present unique challenges in the identification and interpretation of anomalies. Then we propose four approaches to integrating process information into pattern recognition and anomaly detection in different domains.

Firstly, in structured domains, we propose a workflow recognition approach that can automatically correlate the generated data with its corresponding processes, which lays the foundations for leveraging process information for further analysis. Secondly, we propose a process-aware deviation detection approach, specifically designed to operate effectively with uncertain data. Thirdly, we propose a process-driven approach to recognising human activities, aimed towards leveraging process information from semi-structured domains to enhance the accuracy of activity recognition. Fourthly, in unstructured domains, we propose a temporal pattern recognition and anomaly detection approach with a focus on the domain of Activities of Daily Living (ADLs). The proposed approach involves identifying both short-term and long-term deviations in daily activities, as well as detecting changes of complex behaviours over time. In summary, this thesis underscores the effectiveness of incorporating process information to interpret patterns and detect anomalies in different domains.