液滴沸騰自動化觀測系統之建置

Abstract

為了分析液滴沸騰的運作機制，應進行沸騰實驗來繪製出完整的蒸發曲線。然而由於物理現象的限制，進行沸騰實驗會有耗時且不夠精準的問題。為了能夠更有效率且精準地進行沸騰實驗，因此本研究加入了webcam進行即時影像處理，利用Arduino、步進馬達及3D列印物件組成注射幫浦，搭配麥克風及風扇形成自動化系統。利用SolidWorks Flow Simulation對環形風場進行模擬，針對管長及管徑的建置進行探討，最後選用了突擴管及確認最適當的壓克力管與加熱面距離長為15 mm。懸浮液滴的即時影像處理方面，透過圓形偵測法辨識懸浮液滴的半徑，並透過連續上百幀液滴半徑的曲線擬合得出懸浮液滴半徑的時變率在半徑0.6 mm以上時為定值，此外也能推出蒸發時間；在沸騰液滴方面，透過移動平均演算法來辨識沸騰液滴與加熱平面的接觸面積及蒸發時間，並透過接觸面積與蒸發時間計算出即時熱通量。此外針對沸騰液滴的聲響，本研究亦利用深度神經網路模型來探討沸騰液滴，以及加熱平面所造成的沸騰聲響與其熱通量的關係；將影像處理所計算的熱通量作為訓練標籤，利用自動化系統擷取的音訊，並透過特徵工程將原始音訊優化成聲響特徵，且經過機器學習模型的訓練，所得模型能夠完全學習訓練集的特徵，但會有過擬合的問題；在去除掉高熱通量及過渡沸騰區低熱通量的資料後，學習曲線的loss值為49.58 kW/m2，此時其測試集有最佳的表現。最後本研究以自動化系統進行不同研究變因的萊氏實驗，也比較了人工及自動化系統進行實驗的差異，確認自動化系統位於薄膜沸騰區及核沸騰區都有很高的精準度，也發現韋伯數從7.01上升至23.18，萊氏溫度點從154oC上升至192oC，且蒸發時間從85.2秒下降至78.9秒。利用研磨及酸洗製程形成孔蝕表面，會使萊氏溫度點稍微上升、對蒸發時間的影響有限。

To analyze the progression of droplet boiling, the boiling experiments should be performed to plot a complete evaporation curve. However, due to physical constraints, droplet boiling experiments are time-consuming and uncertain. To carry out the droplet boiling experiment more efficiently and accurately, the purpose of this study was to construct an automatic system that was formed with a microphone, a fan, webcams for real-time image processing, and a syringe pump consisting of an Arduino micro-board, a stepper motor, and 3D printed objects. SolidWorks Flow Simulation was used to simulate the annular velocity field to discuss the combination of the pipe length and pipe diameter. The most appropriate result is the sudden expansion pipe, and the distance between the pipe and the hot surface is 15 mm. In the aspect of real-time image processing, the radius of the suspended droplets was identified by circular detection. Through the curve fitting of the droplet radius for hundreds of consecutive frames, it can be concluded that the rate of the suspended droplet radius is constant when the radius is above 0.6 mm, and the evaporation time also is derived. The moving average algorithm is used to identify the contact area and evaporation time between the boiling droplet and the hot surface, and the heat flux can be calculated in real-time through the contact area and evaporation time. In addition, for the audio of boiling droplets, a deep neural network model also was used to explore the relationship between the boiling audio and its heat flux in this study. The heat flux calculated by image processing was used as dataset labels, and the audio signals recorded by the automatic system were optimized by feature engineering as dataset features. Furthermore, after the training of the DNN model, the model can completely learn the features of the training datasets, but an overfitting problem occurs. After removing the data of high heat flux and lower heat flux in the transition boiling region, the loss value of the learning curve is 49.58 kW/m2, and the prediction for testing datasets shows the best performance. Finally, the automated system was used to carry out the Leidenfrost experiments with different experimental parameters, and it also was compared with the manual one to prove the high precision in the film boiling zone and the nucleate boiling zone. Also, it was found that when the Weber number increased from 7.01 to 23.18, the Leidenfrost temperature increased from 154oC to 192oC, and the evaporation time decreased from 85.2 seconds to 78.9 seconds. A pitted surface from the grinding and pickling process slightly increased the Leidenfrost temperature and had limited influence on the evaporation time.