気液二相流の流量計測に向けた特徴量の選定と深層学習モデルへの影響度の調査

抄録

Cryogenic fluids, such as liquid hydrogen and liquid oxygen are used as propellants. However, these liquids have quite low boiling temperatures and easily become a gas-liquid two-phase flow in pipes leading to difficulties in controlling the flow rate and causing problems. Therefore, we have tried to develop a two-phase flow meter using a capacitive void fraction sensor and deep learning techniques (BLSTM). To realize an accurate regression model, we have selected 9 features (from 19 features) including 3-scale wavelet package transform (8 features) and moving range of capacitance with the sliding window method In line with previous research, and we have achieved a more accurate model than raw signal input model by 0.15 (RMSE). In addition, to realize a reliable regression model, we have applied XAI techniques (PFI, PD, ICE). We have found that one approximation signal from wavelet transform and moving range are important for the regression model, and the liquid flow rate increase exponentially to the signal and linearly to the moving range.