為分析斜坡流光彈顆粒受力與傾倒所發展之深度學習模型

Abstract

本研究旨在發展深度學習模型以分析斜坡光彈顆粒流中單顆光彈顆粒在斜坡流中的影像來評估其受力和傾倒狀況。由於準確的分析需要高品質的影像，我們首先進行多顆粒影像裁切和分類等前處理工作，以獲得單顆顆粒影像。隨後，我們發展了三大類模型：第一類包括二分類和多分類模型，其中二分類模型用於識別顆粒有無光條紋及判斷是否帶有因傾倒所產生的月牙形殘影，而多分類模型則用來評估顆粒的配位數。第二類模型用於評估單顆顆粒的受力大小和受力角度，以獲得等價於離散元素法計算所得之細部資訊，為光彈顆粒應用之一大突破。第三類模型分析顆粒的傾倒資訊，包括傾倒幅度（月牙寬度）及其傾倒方位（月牙發生位置）。這些模型以自定義或遷移學習方式建置，在自定義模型中，我們通過經驗法則調整超參數及模型架構，以應對簡單的任務；在遷移學習模型中，我們選擇合適的預訓練模型，使用自定義層，並根據需要進行人工調整超參數以進行遷移學習。結果顯示，這些模型在訓練和驗證過程中表現良好，對未來光彈顆粒的影像分析效率有具體的助益。本研究更進一步統整所得之單顆粒資訊以獲得多顆粒斜坡流之流場資訊，包含顆粒中光條紋以及月牙殘影的出現與否、顆粒受力、月牙殘影厚度與發生位置於流場的分布狀況。

This study develops deep learning models to analyze single photoelastic disk images in inclined chute flows, evaluating their force and tilting conditions. To ensure accurate analysis, we first preprocess multi-disk images by cropping and classifying them to obtain single-disk images. We then developed three types of models: the first type includes binary and multiclass classification models. Binary classification model identifies the presence of fringe patterns and the occurrence of the crescent rim from tilting, while multiclass classification model evaluates the disk coordination number. The second model type assesses force magnitude and force loading angles on individual disks, providing detailed information equivalent to discrete element method calculations, marking a significant breakthrough in photoelastic disk applications. The third model type analyzes disk tilting, including tilt magnitude (crescent rim thickness) and orientation (crescent rim location). These models use either custom designs or transfer learning. In custom-built models, we adjust hyperparameters and architectures based on empirical rules for simple tasks. For transfer learning models, we select suitable pre-trained models, use custom layers, and manually adjust hyperparameters. The results show these models perform well in training and validation, enhancing the efficiency of future image analysis of photoelastic disks. This study further integrates individual disk information to derive flow field details in multi-disk slope flows, including the presence of fringes and crescent rims, disk loading forces, and the distribution of crescent rim thickness and occurrence within the flow field.