Principal Investigator: Kim Duc Tran
Project Members: Thi Thuy Van Nguyen, Guillaume Tartare, Cédric Heuchenne
PhD Student involved in the project: Thi Thuy Van Nguyen

Project Description

The project “Anomaly Detection and Applications” aims to develop advanced statistical and artificial intelligence-based methods for detecting abnormal patterns in complex, dynamic, and high-dimensional data systems. In modern industrial, cyber-physical, IoT, healthcare, and smart manufacturing environments, data are increasingly multivariate, noisy, non-stationary, and affected by evolving operating conditions. Early and reliable anomaly detection is therefore essential for improving system safety, preventing failures, supporting predictive maintenance, and enabling trustworthy decision-making.

A major scientific focus of the project is multivariate time-series anomaly detection, particularly in systems where normal behavior changes over time due to drift, degradation, regime transitions, or external disturbances. To address these challenges, the project investigates non-stationary system modeling and develops methods to distinguish true anomalies from normal temporal variability, sensor noise, distributional shifts, or gradual process drift.

One of the central methodological directions is the development of a Drift-Aware Switching State-Space Model (DSSM). This model is designed to generate realistic multivariate time-series data with controlled latent regimes, temporal dependencies, drift mechanisms, degradation patterns, and anomaly labels. By explicitly modeling switching dynamics and drift-aware behavior, DSSM can support both algorithm development and evaluation under realistic non-stationary conditions. This is especially important because many existing anomaly detection benchmarks do not adequately reflect the complexity of real-world systems, where anomalies may emerge gradually, interact with process drift, or occur under changing operating regimes.

The project also contributes to trustworthy benchmarking for anomaly detection. Instead of relying solely on static or overly simplified datasets, the project aims to construct benchmark environments in which the anomaly type, drift onset, affected variables, degradation state, and temporal context can be explicitly controlled and documented. This allows researchers to evaluate anomaly detection models more rigorously with respect to detection accuracy, robustness to drift, false alarm control, interpretability, and generalization across regimes.

From a methodological perspective, the project combines statistical process monitoring, machine learning, deep learning, explainable AI, and state-space modeling. Statistical process monitoring techniques, such as control charts, multivariate monitoring, support vector data description, and exponentially weighted moving average methods, provide rigorous foundations for process surveillance and abnormality detection. Machine learning and AI methods extend these foundations by learning complex nonlinear relationships from high-dimensional data. Explainable AI is incorporated to make anomaly detection results more transparent, helping users understand which variables, time periods, or latent mechanisms contribute to detected abnormal behaviors.

The project team brings together complementary expertise in anomaly detection, statistical learning, industrial statistics, machine learning, and applied data science. Kim Duc Tran serves as the Principal Investigator and coordinates the scientific direction of the project. Thi Thuy Van Nguyen, as a PhD student, contributes to the development of anomaly detection methods for IoT and multivariate time-series data using statistical and machine learning techniques. Guillaume Tartare and Cédric Heuchenne contribute expertise in statistical modeling, methodological validation, process monitoring, and data-driven decision support.

The expected outcomes of the project include new anomaly detection models, a drift-aware benchmark generator based on DSSM, validated experimental frameworks, scientific publications, and practical tools for intelligent monitoring systems. By integrating rigorous statistical foundations, drift-aware state-space modeling, explainable AI, and trustworthy benchmarking, the project aims to advance reliable anomaly detection solutions for real-world non-stationary systems.

Publication List

1. Title: SVDD control charts based on MEWMA technique for monitoring Compositional Data Authors: Thi Thuy Van Nguyen, Cédric Heuchenne, Kim Duc Tran, Guillaume Tartare, Kim Phuc Tran Year & Venue: 2025, Computers & Industrial Engineering 201, 110865 DOI: https://doi.org/10.1016/j.cie.2025.110865
2. Title: Monitoring the Ratio of Two Normal Variables and Compositional Data: A Literature Review and Perspective Authors: Thi Thuy Van Nguyen, Thi Hien Nguyen, Kim Duc Tran, Cédric Heuchenne, Kim Phuc Tran Year & Venue: 2025 (published 2024), Chapter in Computational Techniques for Smart Manufacturing in Industry 5.0 (Taylor & Francis) Link: https://www.taylorfrancis.com/chapters/edit/10.1201/9781003399322-8
3. Title: Explainable Transformer-Based Approach for ECG Anomaly Detection. Authors: Thi Thuy Van Nguyen, Cédric Heuchenne, Kim Duc Tran, Guillaume Tartare, Kim Phuc Tran Year & Venue: 2026 (presented 2024), Lecture Notes of the Institute for Computer Sciences... (EAI RAIDS), pp. 85-108 DOI: https://doi.org/10.1007/978-3-032-14055-5_7
4. Title: A novel Transformer-based Anomaly Detection approach for ECG Monitoring Healthcare System Authors: Thi Thuy Van Nguyen, Cédric Heuchenne, Kim Duc Tran, Kim Phuc Tran Year & Venue: 2023/2024, The Seventh International Conference on Safety and Security with IoT, pp. 111-129 (Springer) DOI: https://doi.org/10.1007/978-3-031-53028-9_7
5. Title: Anomaly Detection for Compositional Data using VSI MEWMA control chart. Authors: Thi Thuy Van Nguyen, Cédric Heuchenne, Kim Phuc Tran Year & Venue: 2022, IFAC-PapersOnLine 55 (10), 1533-1538 DOI/Link: ScienceDirect article + arXiv:2203.15438
6. Title: Machine learning for compositional data analysis in Support of the decision-making process. Authors: Thi Thuy Van Nguyen, Cédric Heuchenne, Kim Phuc Tran Year & Venue: 2022, Chapter in Machine Learning and Probabilistic Graphical Models for Decision Support Systems (Taylor & Francis)  DOI: https://doi.org/10.1201/9781003189886-8
7. Title:  Designing a Trustworthy Benchmark Generator for Multivariate Time-Series Anomaly Detection: Drift-Aware Switching State-Space Model with Distributional Calibration Authors: Kim Duc Tran, Guillaume Tartare, Hugues Annoye, Cédric Heuchenne (2022, Ongoing)