以卷積神經網路預測邊坡穩定性

Abstract

土壤隨機場理論為一種得以描述土壤空間變異性之方法，因此逐漸獲得關注且發展日益趨於成熟，於邊坡可靠度穩定性分析中，隨機場有限元素分析方法亦得以有效地分析邊坡穩定性；然而，作為一個常搭配蒙地卡羅分析方法使用之工具，其費時且可能需消耗大量運算資源。使用隨機有限元素分析方法決定邊坡破壞機率時，人工智慧是個強大且有潛力於減少大量分析次數之工具；隨著科技與電腦運算效能的發展，機器學習與深度學習方法已愈具效率且熱門。卷積神經網路(convolutional neural network, CNN)為人工智慧領域下的一個類別並得以應用於大地工程問題之預測上，其優良之影像辨識能力是其一大優點，並已被發現其可以一系列之隨機場邊坡作為輸入資料，以強度折減法進行之隨機有限元素分析結果如安全係數或邊坡破壞面作為輸出標籤作模型訓練，並預測邊坡穩定性；然而，不同卷積神經網路模型於預測未知資料集之表現差異與泛化能力仍尚未被測試和檢驗。 因此，本研究使用不同卷積神經網路模型作坡角30◦與坡角50◦隨機場邊坡圖片之訓練和學習，以評估含安全係數與邊坡破壞面兩部分之邊坡穩定性。第一部分為安全係數預測任務，傳統的淺層卷積神經網路會被使用於預測坡角40◦隨機場邊坡之邊坡穩定性；與殘差神經網路(residual neural network, ResNet)結合之卷積神經網路亦被訓練且作預測結果改善能力、預測表現、以及泛化能力之比較；此外，卷積神經網路模型之泛化能力將透過多種變異係數與關聯性長度之組合進行評估。第二部分則使用另一個卷積神經網路對邊坡破壞面進行預測。 研究結果顯示，當訓練資料與測試資料兩資料集具有相同之隨機場來源時，卷積神經網路模型對安全係數與邊坡破壞面兩結果皆具備優秀的預測能力。然而當隨機場來源不同時，在預測安全係數任務中殘差神經網路相較傳統淺層卷積神經網路表現為佳；在預測邊坡破壞面任務中，卷積神經網路則只能對出現於訓練資料中之已知坡角之同坡角邊坡作預測，其尚無法預測未知坡角之邊坡破壞面。因此，卷積神經網路模型應用於未知坡角邊坡破壞面預測之能力仍有待更多研究佐證。

The random field theory of soil has gained attention and become a well-developed approach for characterizing soil spatial variability. The random finite element method, employed in slope reliability analysis, effectively evaluates slope failure probability. However, this method commonly incorporates Monte Carlo analysis, which may have drawbacks due to its time-consuming process and the computational resources required. Artificial intelligence is a powerful tool that can potentially eliminate the need for performing numerous random finite element analyses to determine slope failure probabilities. Machine learning and deep learning are becoming increasingly popular and more efficient as technology advances and computer calculations. Convolutional neural networks (CNNs) are a class of artificial intelligence that can be used to assess geotechnical engineering predictions. Among the advantages of CNN models is its ability to analyze visual imagery. It has been found that a CNN model can be trained to predict slope stability using a series of random fields of slopes as input data and random finite element results, including the factor of safety or slip surface predictions, from strength reduction analyses as output labels. However, the performance differences and generalization abilities of the CNNs in predicting unknown datasets have not yet been examined. Therefore, this study uses CNN models to evaluate slope stability predictions by learning from images of 30◦ and 50◦ random field slopes in two parts: the safety factor and failure slip surface. Firstly, for predicting the safety factors, a conventional shallow CNN model is employed to predict the slope stability of a 40◦ random field slope. The conventional CNN model and this model incorporating residual neural network (ResNet) are compared to see the improvement, performance, and generalization ability. Additionally, the generalization ability of the CNN models is evaluated under various combinations of the coefficient of variation and correlation length. Secondly, another CNN model is trained to predict failure slip surfaces. The results demonstrate that the CNN models exhibit excellent predictive capabilities for the safety factor and slope failure slip predictions when the training and testing data share the same random field source. However, when the random field sources differ, the ResNet model outperforms the conventional shallow CNN for safety factor prediction. Regarding predicting slope failure surfaces, the CNN model used in this study is currently limited to making predictions only for known slope angles included in the training data; it is not yet capable of predicting failure surfaces for slopes with unknown angles. As a result, the CNN model's ability to predict slope failure surfaces with unknown angles remains to be determined.