Analysis of Neural-Network-Based Anomaly Detection Methods for Time Series

Bishoy KAMEL

Diplomová práce

Abstract:

The reliability of deep learning models for time-series classification is paramount in safety-critical applications such as industrial process monitoring and healthcare. A primary threat to this reliability is out-of-distribution (OOD) data, which can cause silent model failures. While many OOD detection methods have been developed for computer vision, their comparative performance in the time-series domain remains under-explored, particularly concerning the influence of the underlying model architecture. This thesis presents a rigorous and reproducible benchmark of six post-hoc OOD detection methods, including confidence, density, and gradient-based approaches. The detectors are evaluated on four distinct time-series datasets using two different classifier backbones: a 1D Convolutional Neural Network (CNN) and a Long Short-Term Memory (LSTM) network. Performance is quantified using the Area Under the Receiver Operating Characteristic (AUROC) curve, with multi-seed trials and bootstrap confidence intervals employed to ensure statistical robustness. The empirical results demonstrate that the feature-space density detector, Deep Deterministic Uncertainty (DDU), consistently achieves the most robust and stable performance across nearly all configurations. In contrast, gradient-based methods exhibit significant performance variability,showing high sensitivity to both the dataset and the choice of model architecture. These findings indicate that the effectiveness of an OOD detector is not independent of the base model and that DDU is a strong candidate for reliable OOD detection in time-series applications.

Abstract:

The reliability of deep learning models for time-series classification is paramount in safety-critical applications such as industrial process monitoring and healthcare. A primary threat to this reliability is out-of-distribution (OOD) data, which can cause silent model failures. While many OOD detection methods have been developed for computer vision, their comparative performance in the time-series domain remains under-explored, particularly concerning the influence of the underlying model architecture. This thesis presents a rigorous and reproducible benchmark of six post-hoc OOD detection methods, including confidence, density, and gradient-based approaches. The detectors are evaluated on four distinct time-series datasets using two different classifier backbones: a 1D Convolutional Neural Network (CNN) and a Long Short-Term Memory (LSTM) network. Performance is quantified using the Area Under the Receiver Operating Characteristic (AUROC) curve, with multi-seed trials and bootstrap confidence intervals employed to ensure statistical robustness. The empirical results demonstrate that the feature-space density detector, Deep Deterministic Uncertainty (DDU), consistently achieves the most robust and stable performance across nearly all configurations. In contrast, gradient-based methods exhibit significant performance variability,showing high sensitivity to both the dataset and the choice of model architecture. These findings indicate that the effectiveness of an OOD detector is not independent of the base model and that DDU is a strong candidate for reliable OOD detection in time-series applications.

Keywords

ime series anomaly detection, out-of-distribution detection, uncertainty estimation, gradient-based methods, neural networks, training dynamics, representation learning

Jazyk práce: angličtina

Datum vytvoření / odevzdání či podání práce: 20. 8. 2025

Obhajoba závěrečné práce

• Vedoucí: prof. Dr. Ing. Udo Garmann

JIHOČESKÁ UNIVERZITA V ČESKÝCH BUDĚJOVICÍCH

Přírodovědecká fakulta

Magisterský studijní program / obor:
Artificial Intelligence and Data Science / Artificial Intelligence and Data Science