Search
Menu
Je me connecte

Institute of Information and Communication Technologies, Electronics and Applied Mathematics

Search

Institute of Information and Communication Technologies, Electronics and Applied Mathematics

Seminars

icteam | Louvain-la-Neuve

Scientific seminars are regularly organised by and for ICTEAM researchers in Louvain-la-Neuve. Participation to these seminars is open for anyone interested.

 

ICTEAM seminars of the week

Current Week Seminars

[INMA] 2026-11-10 (14:00) : Advances in data-driven fault diagnosis of linear dynamical systems using subspace and geometric methods

At Euler building (room A.002)

Speaker: Gabriel de Albuquerque Gleizer (TU Delft)
Abstract: Fault diagnosis, in a broad sense, is the problem of detecting, classifying, and quantifying undesired behavior in dynamical systems. These three problems, with increasing levels of difficulty – but also information power – are called respectively fault detection, isolation, and estimation. Faults can appear as changes in dynamics due to wear and tear, unexpected external disturbances, or losses of components such as sensors and actuators. For MIMO LTI systems, there exist a plethora of model-based theoretical results and tools for fault diagnosis. When only data is available, however, the traditional two-step approach – identify a model, then design a model-based diagnostic system – can be deceivingly problematic. In this talk I will present two classes of data-driven problems in MIMO LTI systems associated with fault diagnosis. The first problem is the system identification from faulty data. This is a rather extreme system-identification problem, where the goal is to identify an LTI system subject to controlled inputs and unobserved faults. It is motivated by situations where an unseen fault can impact the system dynamics while simultaneously acting as an external disturbance; or by situations where the an identification campaign cannot run in a controlled environment. Assuming faults are described by bounded additive external disturbances, I will present a behavioral approach to the problem formulation, and a solution approach that enables the determination of the minimum dimension of the fault space and the minimal subspace of fault-related matrices that explains the data. Building on tools from compressed sensing, I will then illustrate how additional assumptions on the fault signals enable a more precise identification. The second problem is direct data-driven detection of input redundancy and the estimation of invariant zeros. I will start by recalling how these two properties are related to fault diagnosis and cyber-security, and then illustrate how standard identification can completely miss the presence of them. I will then revisit the classic output-zeroing problem – determining the set of input solutions that lead to a flat zero output – from a data-driven perspective, and will show results that enable the computation of the degree of input redundancy for general LTI systems, as well as the invariant zeros of left-invertible systems. The results rely on geometric control tools, but our constructions require dodging the necessity of a state-space representation, working essentially on the input space of the system via properly constructed block-Hankel matrices. With simulation results, I will illustrate how to apply these techniques in realistic cases with noisy data, and showcase the benefits of our approach when compared to an indirect system-identification-based one.

[INGI] 2026-11-12 (13:00) : NVMe-over-QUIC: Secure and Efficient NVMe Disaggregation over IP Networks

At Shannon Room - Maxwell (A.105)

Speaker: Soyong Kim
Abstract: NVMe-over-Fabrics enables disaggregated storage, but high-performance deployments often rely on specialised lossless RDMA fabrics. This talk presents NVMe-over-QUIC, a secure IP-based transport that does not require such a fabric. Our design maps each NVMe command to an independent QUIC stream, leveraging QUIC’s built-in encryption, multiplexing, and loss recovery while avoiding cross-command head-of-line blocking. I will also describe a UDP-based core-steering architecture that preserves CPU affinity for multi-core scaling. Our evaluation shows up to 4× higher throughput than NVMe/TLS and substantially improved tail latency under packet loss and additional delay. These results show that QUIC can provide an efficient and robust alternative for secure NVMe disaggregation over conventional IP networks.