Abstract

The reliability and security of today’s smart and autonomous systems increasingly rely on effective anomaly detection capabilities. Logs generated by intelligent devices during runtime offer valuable insights for system monitoring and troubleshooting. Nonetheless, the enormous quantity and complexity of these logs render manual anomaly inspection impractical and error-prone. To address this, various automated log-based anomaly detection methods have been developed. However, many of these approaches are evaluated in controlled environments with publicly available datasets, which differ significantly from the noisy, unstructured, and unlabeled logs encountered in industrial settings. This thesis explores and adapts existing machine learning and deep learning techniques for anomaly detection in real-world system logs produced by an intelligent autonomous display device. Initially, we conduct a comparative analysis of machine learning and deep learning methods using a small manually labeled dataset to evaluate the detection accuracy and computational efficiency. Our results highlight the most suitable approaches for enabling proactive maintenance and enhancing system reliability. Expanding on this, we evaluate advanced deep learning methods across weakly supervised, semi-supervised, and unsupervised learning paradigms, using heuristically labeled logs and benchmark them against fully supervised baselines to examine the trade-offs between label dependency, detection performance, and industrial applicability. Finally, we propose a systematic approach for managing unlabeled and noisy log data, providing practical guidelines for selecting suitable learning strategies based on label availability, data quality, and real-world constraints. The findings of this work provide valuable insights for the implementation of scalable, accurate, and robust log-based anomaly detection in industrial IoT environments.