應用基於均勻採樣資料集之生成模型於四連桿機構之路徑合成

Abstract

四連桿機構之路徑合成使用最佳化演算法會遇到缺陷問題與初始值設定問題。雖然近年來可見到結合深度學習的研究，成功解決使用最佳化演算法的問題，但在精準度上仍還有提升的空間。因此本研究提出一個模型架構，包含目標曲線之前處理、資料集生成以及模型訓練之過程，該模型架構可以自動判斷適合目標曲線的生成模型。本研究將目標曲線擬合後重新生成，解決目標曲線點數量不同與疏密不均勻的問題。針對資料集的生成，使用非監督式學習的方式調整不同形狀之曲線所佔的比例，讓生成資料集的過程可以均勻採樣到各種形狀的曲線，提升模型預測的準確性。本研究針對不同的四連桿機構種類，分別訓練對應的模型，並使用分類器自動判斷目標曲線適用的生成模型，簡化訓練的難度。本文中的範例顯示該模型架構可以成功合成各種封閉式曲線與開放式曲線。

Solving the path synthesis of four-bar linkages via optimization algorithm may encounter issues as defects and guess of initial values. Recently, the method for resolving such problems associated with deep-learning schemes shows that it can avoid these issues. However, there is still room for improving the accuracy of the scheme. In this work, a framework including preprocessing, data generation, and model training for the path synthesis of four-bar linkage are presented. The preprocess starts by regenerating the target path with evenly distributed points along the path, followed by the normalization of the shape and feature extraction. For data generation, an unsupervised learning is employed to adjust the distribution of paths of different shapes in the dataset so that the robustness of the model can be achieved. As for model training, models based on datasets of different types of four-bar linkages as well as a classifier to determine the suitable generative model for the target path is constructed. Finally, several examples including closed and open paths are illustrated to verify the effectiveness of the framework.