- 學位論文
利用機器學習模型應用於池沸騰中氣泡之辨識分析
Machining learning enables analysis of bubble detection in pool boiling
摘要
本研究使用機器學習技術，對HFE-7100介電流體的池沸騰所產生的氣泡進行了深入分析。通過應用StarDist[1]模型對影像中的氣泡進行識別和形狀描繪，獲得了畫面中氣泡的大小分布和數量，並進一步進行氣泡追蹤，以分析氣泡的移除率。此外，我們在訓練資料中加入了區域陰影，以比較最終模型對氣泡識別能力的影響。研究中，我們利用提出的演算法對實驗所得的氣泡影像進行分析，同時我們使用本研究的模型對其他實驗所拍攝的氣泡影像進行分析，以驗證模型的泛用性。 實驗結果顯示，神經網路對於低熱通量條件下的單獨氣泡具有良好的識別效果，能夠準確分辨氣泡的邊緣和數量。在氣泡相鄰或互相遮擋時，演算法仍能正常辨識氣泡範圍並進行分割。在此階段，大部分氣泡為小型氣泡，且背景無大量氣泡遮擋產生的陰影，因此使用無陰影模型進行分析即可。隨著熱通量的提高，背景開始出現大量氣團和陰影，影像品質下降，且氣泡的大小增加，辨識難度加大。此時，使用包含陰影資料訓練的模型進行分析，可以觀察到氣泡辨識的正確性有顯著的提升。當池沸騰進入連續氣泡區域時，氣泡容易大量合併，導致氣泡面積急劇增加。儘管如此，模型仍能正常辨識大部分氣泡，但部分大型氣泡可能出現切割現象，導致體積計算產生誤差，進而低估氣泡體積。同時我們發現本模型亦可以應用於其研究之氣泡影像，在大部分情況下也表現良好，驗證了模型的泛用性。
並列摘要
In this study, machine learning techniques were used to analyze bubbles generated during pool boiling of the dielectric fluid HFE-7100. The StarDist[1] model was applied to identify and outline bubble shapes in images, allowing us to determine the bubble size distribution. Bubble tracking was also performed to analyze the removal rates of vapor. Shadow was added to the background of the training data to assess robustness of our model for bubble recognition under different conditions. In addition, both bubble images from experiments and other studies were tested to validate its generalizability. The results showed that the neural network could effectively identify individual bubbles under low heat flux conditions, accurately distinguish bubble edges and counts. Even for adjacent or overlapping bubbles, recognition and segmentation were done successfully. At low heat flux, most bubbles and the effect of shadows were small, so the model trained by original data was good enough. As the heat flux increased, background shadows emerge and bubble grew, reducing the image quality. To address this issue, the model was re-trained by data manipulated with shadow-gradient background. This significantly improved the accuracy of bubble recognition under these conditions. When continuous bubble generation, led to coalescence, most bubbles could still be identified. However, some large bubbles were partially segmented, causing the underestimation of vapor volume. The model also performed well on bubble images from other studies, confirming its generalizability.