Graph Embedding-Based Multilinear Subspace Analysis of State Ensembles in Gas–Liquid Flow
Z. Y. Li, Chao Tan, Shumei Zhang, Feng Dong · IEEE Transactions on Instrumentation and Measurement · 2024
The process states in gas–liquid flow are the synthetical consequence of multiple apparent factors related to time, space, flow conditions, and so on. Higher order tensor decomposition in multilinear algebra provides a powerful framework mathematically to analyze the multifactor structure of flow state ensembles. Focusing on entangling the constitute factors or modes embedded in various states, a novel strategy about graph-embedding-based multilinear subspace analysis (MSA) of state ensembles in gas–liquid flow is proposed. Pulse-wave ultrasonic Doppler (PWUD) sensor is adopted to acquire spatiotemporal information of flow states under varying flow conditions. Under the architecture of graph embedding, tensor dynamic local preservation projection (TDLPP) is constructed via PWUD signal to further describe the latent characteristics of various flow states in time–space dimensions. In the reconstructed graph embedding latent space, higher order singular value decomposition (HOSVD) is used for tensor analysis of flow state ensembles that combine several modes, including different state categories, time, space, and flow conditions. The purpose of the strategy is to further achieve state monitoring of typical and transitional processes in gas–liquid flow under varying flow conditions with a simple sensor, on the foundation of refining apparent factors underlying flow state formation. Dynamic experiments are designed to verify the effectiveness of the proposed scheme.