No document available.
Abstract :
[en] Anomaly detection is an essential task in modern monitoring systems because the timely identification of abnormal behavior can prevent operational issues such as process deterioration and system failure, as well as broader impacts like economic loss and serious health consequences. Its importance spans many fields, including manufacturing, healthcare, finance, and industrial monitoring. However, anomaly detection becomes particularly difficult when the data exhibit structural constraints, temporal dependence, high dimensionality, or practical requirements related to false alarm control, interpretability, and privacy.
Traditional statistical monitoring methods provide explicit alarm rules and false alarm control, but may struggle with highly complex data structures. In contrast, machine learning (ML) methods have demonstrated strong predictive ability and flexibility, but they often behave as black boxes and do not explicitly address monitoring-oriented requirements such as alarm calibration, interpretability, and deployment without centralizing raw data. This thesis addresses these challenges by developing anomaly detection frameworks that combine statistical and ML techniques for two difficult data settings: compositional data (CoDa) and electrocardiogram (ECG) time series data.
The first part of the thesis focuses on anomaly detection for CoDa, whose components are constrained to represent relative parts of a whole. Because of this simplex structure, standard multivariate monitoring tools cannot be appropriately applied without accounting for the compositional constraints. To address this issue, the thesis proposes a variable sampling interval multivariate exponentially weighted moving average (VSI-MEWMA) control chart for CoDa based on the isometric log-ratio (ilr) transformation. Furthermore, it introduces a novel Dirichlet density transformation to effectively reduce dimensionality while removing the inherent compositional constraints. The research then advances CoDa monitoring by developing more flexible one-class approaches that integrate the MEWMA technique with support vector data description (SVDD), using both ilr and Dirichlet density transformations. The thesis also reviews the existing control chart literature for monitoring CoDa, synthesizes its main developments, and identifies limitations and directions for future research.
Simulation results show that the VSI-MEWMA CoDa chart detects process shifts more rapidly than its fixed sampling interval counterpart, while the proposed ilr-based MEWMA-SVDD chart generally provides more effective monitoring than the traditional MEWMA-CoDa chart, and the Dirichlet-based chart is competitive in several of the examined scenarios.
Building on the monitoring philosophy established in the first part, the second part of the thesis adapts these methodologies for ECG anomaly detection in healthcare. While the first part utilizes statistical transformations for CoDa, this part employs a Transformer-based variational autoencoder (VAE) to learn deep representations of ECG signals. The reconstruction error vectors obtained by comparing the original ECG signals with the VAE reconstructions are then used as the input to the MEWMA-SVDD control chart framework, thereby effectively combining deep learning (DL) with statistical decision rules designed to control false alarms. To improve transparency and interpretability, the centralized framework is further extended with SHapley Additive exPlanations (SHAP). Finally, the thesis extends this ECG monitoring framework to a federated learning setting, where multiple clients collaboratively train a shared Transformer-based VAE without exchanging raw ECG data, while the MEWMA-SVDD detectors and their control limits are constructed locally for each client. Experimental results show that the centralized Transformer-based VAE combined with MEWMA-SVDD framework achieves high anomaly detection performance while enabling false alarm control through a control limit calibrated according to a pre-specified in-control ARL. The SHAP extension provides local explanations of the ECG temporal windows contributing most strongly to the VAE reconstruction errors. In the federated setting, the framework maintains strong local classification performance and improves local false alarm control overall through locally calibrated MEWMA-SVDD control limits without sharing raw ECG data.
Taken together, the results of this thesis show that combining statistical process control (SPC) and ML can lead to anomaly detection methods that are structurally appropriate, monitoring-aware, explainable, and privacy-friendly. The proposed frameworks not only enhance anomaly detection in CoDa and ECG monitoring but also provide a foundation for future developments in interpretable and decentralized anomaly detection systems.
Institution :
ULiège - Université de Liège [HEC Liège], Liège, Belgium
ULille - Université de Lille [ENSAIT (École Nationale Supérieure des Arts et Industries Textiles)], Roubaix, France