Anomaly detection is a crucial task in time series analysis, aiming to identify unusual patterns or behaviors that deviate from expected norms. These anomalies can indicate significant events, such as system faults, fraudulent activities, or unexpected changes in real-world processes. As the volume and complexity of time series data continue to grow across various domains, developing efficient and robust anomaly detection methods remains an active area of research. Anomaly detection approaches can be broadly categorized into statistical, machine learning, and deep learning methods. Statistical techniques rely on assumptions about the data distribution and flag observations that fall outside expected ranges, such as using moving averages, autoregressive models, or hypothesis testing. Machine learning-based methods use patterns learned from the data to distinguish normal behavior from anomalies, often leveraging clustering, distance measures, or ensemble techniques. More recently, deep learning models such as autoencoders, recurrent neural networks, and transformers have been employed to capture complex temporal dependencies and detect subtle irregularities. Each category offers unique strengths and trade-offs in terms of interpretability, scalability, and sensitivity to different types of anomalies. In this thesis, I contribute to the field by integrating and adapting several unsupervised anomaly detection algorithms into a unified benchmarking framework called TSB-AD. This benchmark is designed to evaluate and compare time series anomaly detection methods under a consistent experimental setup. I ported multiple algorithms—some originally published between 2020 and 2024—into the framework, ensuring they followed a standardized API for training and inference. Each implementation was tested on labeled datasets to assess detection accuracy, computational efficiency, and robustness. My work enhances the usability and extensibility of the TSB-AD benchmark, enabling more reliable and fair comparisons between algorithms and supporting future research in this area.
Παναγιώτης Τοπαλίδης (Wed,) studied this question.