陸正昌;施政廷Cheng-Chang Lu;Cheng-Ting Shih廖靖旻Ching-Min Liao醫學影像暨放射科學系2026-03-262026-07-1220212021-06-30https://handle.ncl.edu.tw/11296/34dyz2https://scholars.csmu.edu.tw/handle/123456789/7543由於科技蓬勃發展，工程製造技術提升，現今已能夠大量生產出高度複雜之模型。然而，在工程領域中常用於作為實際測試之有限元分析模擬軟體已無法對大量且高度複雜之模型即時地進行力學分析。因此，本研究提出利用深度學習技術以即時地對在拉伸及壓縮測試下任意形態金屬結構與任意材料之力學分析。本研究利用有限元分析模擬軟體產生2998筆具隨機孔隙結構、任意金屬材料以及任意拉伸與壓縮距離之資料，且將其整合成矩陣之形式，以利力學神經網路之讀取與訓練。本研究提出四種多輸入單輸出之三維力學神經網路，分別為基於卷積神經網路的編碼解碼網路（CNN-based Encoder-Decoder, ED）、三維U型網路（three-dimentional Unet, 3D-Unet）、基於編碼解碼的力學嵌合網路（CNN-based Encoder-Decoder with mechanical embedding, ED-Mech）以及基於編碼解碼的力學嵌合殘差學習三維網路（CNN-based Encoder-Decoder with mechanical embedding and residual learning, 3D-StressNet），以進行快速馮米塞斯應力場之預測，並以均方根誤差、平均絕對誤差以及平均絕對百分誤差評估力學神經網路之效能。本文提出之四種力學神經網路在訓練過程中，隨著迭代次數的增加，訓練曲線圖呈現穩定收斂，在訓練完成時，均方根誤差也有持續下降之趨勢。其中，以具有力學嵌合及殘差學習模塊之3D-StressNet效能最佳，誤差結果為最低。其均方根誤差與平均絕對誤差分別為5853以及16.9。而在測試案例中，也以0.12%之平均絕對百分誤差顯示其預測馮米塞斯應力場之優秀性能。在大量且高度複雜模型之力學分析上，深度學習技術具有強大之潛力以進行準確且快速地馮米塞斯應力場預測。本研究在提出了國際間首個多輸入單輸出之三維力學神經網路的同時，也建置了國際間首個三維力學資料庫。未來期望能使深度學習技術更加廣泛地運用於工程領域中，在保持精準度的同時也提升計算效率，以對社會做出貢獻，增進人類福祉。Due to the vigorous development of science and technology and the improvement of engineering manufacturing technology, it is possible to mass produce highly complex models. However, the finite element analysis simulation software that is often used as actual testing in the engineering field can no longer perform mechanical analysis on a large number of highly complex models in real time. Therefore, this research proposes using deep learning technology perform mechanical analysis of arbitrary shape of metal object and material under tension and compression tests instantly.This study uses finite element analysis simulation software to generate 2998 data made up of random porous models, arbitrary metal materials, and arbitrary tension and compression distance. This data will be integrated into the form of matrix to facilitate the reading and training of the convolutional neural network. This research proposes four multi-input single-output three-dimensional mechanical neural networks, namely CNN-based Encoder-Decoder (ED), three-dimentional Unet (3D-Unet), CNN-based Encoder-Decoder with mechanical embedding (ED-Mech), CNN-based Encoder-Decoder with mechanical embedding and residual learning (3D-StressNet), to predict the von Mises stress rapidly, and use the root mean square error, average absolute error and average absolute percentage error to evaluate the performance of the mechanical neural network.In the training process of the four mechanical neural networks proposed in this research, the training curve shows a stable convergence with the increase of iterations. When the training is completed, the root mean square error also has a continuous downward trend. Among them, the 3D-StressNet with mechanical embedding and residual learning module has the best performance and the lowest error result. The root mean square error and average absolute error of 3D-StressNet are 5853 and 16.9, respectively. In the testing case, the average absolute percentage error of 0.12% also shows the excellent performance of 3D-StressNet in predicting the von Mises stress field.In the mechanical analysis of a large number of highly complex models, deep learning technology has a strong potential to predict the von Mises stress field accurately and quickly. This research not only proposes the first multi-input single-output three-dimensional mechanical neural network, but also builds the first three-dimensional mechanics database in the world. In the future, it is expected that deep learning technology can be more widely used in the engineering field, while maintaining accuracy, but also improving computing efficiency, so as to contribute to society and enhance human well-being.積層製造工程應力分析有限元分析深度神經網路additive manufacturing (AM)manufacturing stress analysisfinite element analysisdeep neural networkthesis