Sequential State Estimation for Regime and Transition Detection in Annulus Flow Videos

Abstract

Flow involving gas and liquid phases inside a vertical annular channel develops into several large-scale flow patterns. These patterns are better understood by observing how they change over time, since their defining visual characteristics emerge through continuous evolution rather than from single, static snapshots. In experimental video footage, factors such as how long certain structures persist, how irregularly they appear or disappear, the way interfaces between phases rearrange, and short-lived mixing events are often just as significant as the momentary visual state captured in any single frame. For this reason, regime recognition based only on single-image classification can miss the temporal structure that makes transitions interpretable and operationally useful. This paper develops a sequential machine-vision framework for regime identification and regime-transition detection in vertical annulus videos. The formulation treats the observed image stream as a noisy projection of an evolving latent flow state and models regime assignment as a temporally coupled inference problem with uncertain boundaries. The proposed treatment integrates frame encoding, sequence representation, transition scoring, soft boundary supervision, and causal filtering so that stable intervals, mixed intervals, and onset events can be handled within one probabilistic pipeline. Particular attention is given to the mismatch between high frame-rate redundancy and comparatively slow regime evolution, the rarity of boundary events relative to stable segments, and the fact that clip-level human interpretation often provides more faithful supervision than isolated frame labels. The paper also specifies evaluation procedures suitable for streaming annulus data, including experiment-disjoint inference, interval-aware transition metrics, calibration near boundaries, and robustness under optical degradation. The resulting methodology is intended to support flow-regime recognition systems that can do more than assign frame labels, namely estimate evolving macrostate trajectories and localize the intervals in which one regime gives way to another.