液滴沸騰影像處理與聲響深度學習分析

Abstract

沸騰曲線的研究與繪製，對沸騰熱傳的研究與相關的應用至關重要，沸騰熱傳因為其高潛熱與高熱對流係數，常常被應用在加熱或散熱上，然而沸騰曲線的研究往往有準確度不高、不確定性高與研究時間人力成本高的現象，因此本研究以傳統型的表面沸騰實驗設計為出發，使用水平的熱金屬表面搭配抽風系統，建構並優化其系統模擬、自動化實驗流程並加入影像與聲響分析，以減少實驗的時間人力成本，並降低實驗的不確定性。本研究以SOLIDWORKS Flow Simulation對實驗設備的溫度與風場進行模擬，並依實驗量測與模擬結果優化實驗設計，使用電腦、LabVIEW使用者介面、注射幫浦與PID控制系統自動化實驗流程，同時以影像與音訊分析判斷液滴沸騰的沸騰狀態、移動、大小變化等等，並引入卷積神經網路深度學習以提升其效果。完成表面沸騰實驗系統-以抽風維持懸浮液滴之建置與優化，確立了熱金屬板與壓克力圓管之間距15 mm為最佳，鼓風機之工作電壓15 V為最佳。以圓形偵測演算法，自動追蹤上百次萊氏現象沸騰液滴的移動與大小變化，透過液滴半徑與時間變化之斜率，計算熱通量，繪製去離子水與95%酒精在6061鋁合金上之沸騰曲線，並得到去離子水的體積參數為2.75，95%酒精的體積參數為3.19。最後嘗試以聲響分析沸騰液滴，以ResNet CNN模型，分辨出沸騰與環境音，其準確率達9成以上。未來期望繪製更多條不同液體與加熱表面之沸騰曲線，並使用此方法輔助凹式表面沸騰實驗系統或是池沸騰實驗。

The research of the boiling curve is an important component in thermal science and plays a key role in heat transfer. However, the traditional methods of the boiling curve plotting were uncertain and time-consuming. Therefore, the purpose of this research was to establish the automatic surface boiling system by using image analysis, audio signal analysis, and CFD simulation. In this study, the SOLIDWORKS Flow Simulation was used to estimate the temperature and the wind velocity distribution in the system, and the results were used to optimize the system. For the purpose of system automation, the PID temperature control system, syringe pump, webcam, microphone, and computer were integrated by LabVIEW. Besides, the image and the audio signal analysis were used to determine the evaporation time, boiling situation, location, and size of the boiling liquid. Furthermore, CNN deep learning models were used to enhance the performance of the audio signal analysis. The CFD models of the system were established. After the optimization, the distance between the heated test surface and the acrylic pipe was set to be 15 mm. Moreover, the working voltage of the blowers was set to be 15 V. In this study, hough circle detection was used and successfully detected the boiling water on the heat plate. By tracking the boiling water, the heat flux of a certain heated surface temperature was calculated by the boiling water radius changing rate, and the boiling curves of deionized water and ethanol were plotted. Furthermore, the research found out that the volume index of deionized water was 2.75 and the volume index of ethanol was 3.19. The ResNet CNN model was used to classify the audio signal of noise and boiling water, and the accuracy of the results was more than 90%. There is abundant space for further progress in analyzing the boiling curve of different liquids and different heated test surfaces.