Novel data analytics to automatically detect patterns in diverse customer data

Abstract

Time series anomaly detection is a critical task across various industries, aiming to identify unusual behaviour that does not conform to expected behaviour. Anomalies in the data typically represent an indication of an underlying issue in the system or unexpected change in a pattern in one or more time series metrics. Telecom service providers are ensuring a higher level of customer satisfaction by understanding customer behaviours and providing reliable network services. However, the increasing availability of complex, multivariate time series data from large scale computer networks presents both opportunities and challenges for network monitoring and anomaly detection. This thesis addresses the critical domain of time series anomaly detection, a key area across numerous sectors due to its importance in identifying irregular patterns within data that deviate from expected behaviours.

Detecting anomalies in complex time series data is crucial in various fields, as it enables the identification of irregular trends, the prediction of system behaviour, and the prevention of negative outcomes. Chaotic systems are deterministic dynamical systems that exhibit irregular, seemingly random behaviour, and are sensitive to initial conditions. In this thesis, the efficacy of machine learning models such as the extreme learning machine and the multilayer perceptron is explored for chaotic time series prediction with a particular focus on an Online Sequential version of ELM aimed at enhancing anomaly detection capabilities. The adaptive nature of dynamic systems, demonstrating how these models learn and adjust to new data, allowing them to remain effective over time. In most practical scenarios, time series data usually arrive in different variants with distinct sequential and temporal characteristics depending on the data generating processes. A novel decomposition-based anomaly detection framework is designed to be robust and scalable, capable of accommodating the diverse and often challenging nature of time series data across various domains, thus setting a benchmark for subsequent analyses. The effectiveness of the proposed approach for anomaly detection is evaluated across diverse domains using a weighted segment-based mechanism.

The decomposition-based anomaly detection framework is applied to BT real-world telecommunication network time series metrics. Their predictive empirical significance has been evaluated through forecasting evaluation metrics. By employing dynamic forecasting in an online cross-validation framework, we have demonstrated the capability for real-time prediction and adaptation to new data. This approach is essential for telecommunications networks, where data traffic patterns evolve rapidly, necessitating models that can adjust dynamically to changing conditions. To tackle network time series metrics at large scale a profile pattern-based anomaly detection approach is developed for multiple network throughput metrics using a clustering algorithm based on their similar patterns. The features computed after creating the profile pattern are modelled with the Isolation Forest algorithm. This approach offers a scalable and efficient solution for anomaly detection in large-scale network environments, ultimately contributing to improved network performance and increased reliability.


Thesis is embargoed until 31st January 2028

Date of Award	Jan 2026
Original language	English
Sponsors	BT Ireland Innovation Centre (BTIIC)
Supervisor	Bryan Scotney (Supervisor), David Glass (Supervisor) & Shuai Zhang (Supervisor)