Speaker Name: Dr. Ranjith Chiplunkar (Imperial College London)

Date: 28-01-2026 (Wednesday)

Time: 2:30 PM

Venue: LC102

Abstract: Latent variable models offer a compact way to represent complex systems through unobserved underlying factors. In many practical settings, these latent variables exhibit intrinsic asymmetry due to physical constraints or directional dynamics, which are not well captured by Gaussian assumptions. In this talk, I present a probabilistic framework for extracting and interpreting latent process dynamics by explicitly accounting for asymmetry in both system evolution and observation models. The central theme is the use of non-Gaussian, skewed probabilistic representations to
model constrained behavior, enabling physically consistent and interpretable representations under uncertainty.

First, I introduce a state-space modeling approach for non-stationary processes in which monotonic degradation dynamics are represented using a closed skew-normal random walk, while stationary process relationships are modeled separately using Gaussian state dynamics. The resulting formulation leads to a simultaneous state-and-parameter estimation problem involving skew-normal filtering and smoothing problems. The methodology is demonstrated through simulation studies and a predictive monitoring application involving fouling in a hot lime softener system.

Building on this perspective, I then present a physics-informed probabilistic slow feature analysis framework for extracting slowly varying latent  structures from high-dimensional process data. Here, skewed output distributions are utilized to model the inequality constraints on the observed data, ensuring that the extracted latent variables remain physically consistent, interpretable, and generalizable. An industrial case study illustrates how incorporating physics into probabilistic feature extraction improves robustness and interpretability while reducing effective data dimensionality.

Together, these works highlight how skewed probabilistic modeling and physics-aware latent variable representations provide a principled foundation for monitoring and understanding complex process dynamics, particularly in systems where non-stationarity, constraints, and uncertainty play
a central role.