Flow features are usually identified using instantaneous snapshots of quantities such as velocity, pressure, and vorticity. Meanwhile, the flow time history (FTH) can easily be measured by probes and used to find local time-varying flow characteristics. However, as the temporal features vary greatly at different locations, it is challenging to comprehensively analyze the entire flow domain. In this work, the flow time history autoencoder (FTH-AE) is employed to encapsulate the massive flow variable data at different locations simultaneously into a set of compacted ultra-low-dimensional latent codes. As the flow dynamic system can be accurately reconstructed using these codes, they contain the key temporal features that are then used by the Formula: see text-means clustering algorithm to distinguish temporal features at different locations. The streamwise and crosswise velocity data of unsteady flow around a circular cylinder at Formula: see text and Formula: see text were trained, and the whole field was successfully divided into multiple meaningful subdomains. It was proved that each subdomain has similar temporal flow features, such as average velocity, phase, fluctuation, and mode, which are consistent with existing knowledge of fluid mechanics. The proposed method provides an alternative approach for distinguishing flow features using deep learning of probe-based time history data.
Zhan et al. (2026) studied this question.