Year: 2024
Type: dissertation
Abstract: Telecommunication networks are essential in modern communication, yet their complexity presents significant challenges in anomaly detection and classification. Traditional methods, focusing on single data sources like text logs or time series, often overlook complex anomaly patterns. This thesis explores the applications of multimodal approaches that synergize text logs and time series data through techniques ranging from concatenation to advanced cross-modality attention mechanisms, aiming to enhance anomaly classification. We evaluated our approach in a real-world scenario within Ericsson’s Radio Access Network (RAN) domain, using the class-average F1 score for performance benchmarking across different anomaly classes. Our results show that, although unimodal models with transformer encoders achieved an impressive class-average F1 score of 80.3%, our multimodal models were competitive at 78.64%. Importantly, in the absence of transformer encoders in the model architecture, multimodal models outperformed unimodal baselines by a huge margin, underscoring their efficacy and merit, particularly in environments with computational limitations. The results indicate that self-attention mechanisms play a crucial role in understanding complex multimodal anomaly signatures, and warrant further exploration of multimodal deep learning techniques in the telecom domain. This research sets the stage for future advancements in telecom network diagnostics, contributing to the development of more effective and adaptable anomaly classification and detection solutions.
Author: Devm Upadhyay, Shekhar
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7037969608