當今流體密度、黏度與表面張力量測方法大多只能一次量測單一物性且對於 待測液體有嚴格的溫度範圍限制，還有複雜、精密設備使量測成本難以降低。排液容 器法為基礎以容器盛裝一靜止流體並由容器底部的孔口排出，量測出口隨時間或隨高 度頭的質量流率數據作為流動特徵，其優勢為不論在冷、熱流場都能有高度的適應 性，在實驗操作簡單能同時量測三種流體物性，且設備結構簡單和低維護成本。 本研究捨棄了過往排液容器法中的修正排放係數或是初始猜測值，避免了排放係 數涵蓋物性範圍受限於實驗或是決定初始猜測的問題，改以深度學習作為回歸的演算 法。本文使用 COMSOL Multiphysics 商用軟體建立一個接近現實排液容器法實驗模型 的計算流體力學模型，產生大量不同流體的流動特徵數據作為回歸資料的基礎，再運 用 TensorFlow 深度學習對繁多流動特徵進行整理並建立回歸模型，省去了實驗量測資 料轉換為目標物性反覆迭代的計算過程，快速地從質量流率資料經由深度學習模型直 接算出目標物性預測值。 以甘油水溶液與丙二醇水溶液對最終預測模型進行驗證，密度由實驗資料得到的 預測值平均絕對誤差為 7.21kg/m3，平均相對誤差為 0.66%，黏度由實驗資料得到的預 測值平均絕對誤差為 0.215cP，平均相對誤差為 6.03%，在表面張力係數由實驗資料得 到的預測結果，其平均絕對誤差為 24.27mN/m，平均相對誤差為 42.98%。對比使用模 擬正算的質量流率作為輸入資料得到的預測值之平均絕對誤差為 1.52mN/m，平均相對 誤差為 3.16%，由此可知此方法在量測密度與黏度的可靠度，然而在表面張力係數的 預測上僅在理論上可使用質量流率回歸，尚無法在現實的實驗中執行。未來想要改良 此量測方法，必須要盡可能縮小模擬與實驗資料的誤差才能將此量測方法推向更穩定 更精確度的表現，並實現量測表面張力係數在實務上的應用。;Most of the measuring methods for fluid density, viscosity and surface tension coefficient nowadays can only measure one single physical property at one time and have strict restriction of fluid temperature range. In addition, the cost of measurement is difficultly reduced resulting in the complex and sophisticated equipment. Based on Draining Vessel Method, a stationary fluid is contained in a vessel and discharged from the orifice of the bottom of vessel. Make the mass flow rate over time or head measured from outlet as flowing feature. The advantage of this method is high adaptability in cold or hot flow field. Furthermore, it allows measuring three kinds of fluid physical property at one operation of experiment with simple equipment and low maintaining cost. In this study, we abandon the discharging coefficient and initial guess value in previous research of Draining Vessel Method to avoid the problem owing to lack of physical property range covered by experiment or the decision of initial guess value. We apply the deep learning as the algorithm of regression. With the CFD simulation software (COMSOL Multiphysics), we establish a model of draining vessel that is close to reality and generate a large number of flow feature data of different fluids as the basis for the regression data. Through deep learning, various flow features are sorted out and the regression model are established. It saves the iterative calculation process during converting the experiment measurement data to target physical property. The mass flow rates are rapidly figured out predicted target value through the deep learning model directly. We verify the final prediction model with the aqueous glycerol and aqueous propylene glycol. The mean absolute error of prediction of density obtained from experiment data is 7.21kg/m3 , and the mean relative error is 0.66%. The mean absolute error of prediction of viscosity obtained from experiment data is 0.215cP, and the mean relative error is 6.03%. In the prediction result of surface tension coefficient obtained from experiment data, its mean iii absolute error 24.27mN/m and its mean relative error is 42.98%, while the mean absolute error is 1.52mN/m, and the mean relative error is 3.16% with the prediction obtained from using the simulated mass flow rate the as input data. This method is reliable in measuring density and viscosity, but it hasn’t been able to measure surface tension coefficient in real experiment. If we want to improve this measurement method in the future, it is necessary to reduce error between the simulation and experiment as much as possible so as to make this method more stable and accurate performance and realize the practical application of the measurement of surface tension.
