Image-Data-Driven Deep Learning in Geosystems

2006年深度学习的突破引发了文本处理、语音识别、无人驾驶汽车、疾病诊断等领域的众多前沿创新。该项目将利用最新的计算机视觉创新背后的核心概念来解决工程中一个很少讨论但却紧迫的问题：如何分析爆炸性增长的图像数据，包括图像和视频，这些数据很难用传统方法进行分析。这些概念将被用来探索利用图像数据准确评估挡土墙安全性的可能性。这一努力旨在建立一个将工程学科与人工智能联系起来的范例，并通过利用图像数据驱动的深度学习实现地球系统的分析，从而加强地球系统作为一个重要基础设施组成部分的安全性。该项目将有助于重振岩土和其他工程领域的传统人工智能子领域，就像深度学习重新点燃了人们对人工神经网络和机器学习的兴趣，并使它们成为STEM研究和创新的领先者一样。该项目还可能改变工程师对如何以一种归因于深度学习的革命性方式创造知识的看法，即直接从数据学习，而不是间接从基于数据建立的模型学习。将通过开发移动应用程序来传播该项目的想法和产品，从而进行创新的教育和推广工作。该项目将通过在私人投资机构和专业会议的各种活动中使用项目产品(包括应用程序)向K-12学生、代表性不足的群体以及岩土工程研究人员和从业者提供服务，从而为教育做出贡献。这项研究的目的是通过对挡土墙稳定性分析的探索性调查，了解图像数据驱动的深度学习在地质系统中的作用。为了实现这一目标，将研究计算机视觉最近的突破，以评估典型地质系统(即挡土墙)的稳定性。计算机视觉后来被用作谷歌AlphaGo开发的核心技术之一。使机器在视觉分类能力上超越人类的核心概念，即卷积神经网络(CNN)，将被用于处理岩土工程中的大数据，这些大数据主要由静态和实时图像(视频)组成，使用传统的岩土工程方法很难进行分析。传统的神经网络将被用来分析挡土墙的图像，以判断墙是安全的还是失效的。对于定量分析，将使用随机方法生成挡土墙的2D和3D图像，并使用传统的极限分析和数值方法进行分析。这些标记的图像数据将被用作输入来训练卷积神经网络以进行监督学习。训练后的网络将与以与训练数据相同的方式生成的另一组独立数据进行测试。本项目将进行三个研究任务：1)理解图像数据驱动的岩土工程研究的数据科学；2)研究深度学习中图像模式与物理机制之间的联系；3)揭示深度学习方法在挡土墙稳定性分析中的稳健性和外推能力。

Breakthroughs in deep learning in 2006 triggered numerous cutting-edge innovations in text processing, speech recognition, driverless cars, disease diagnosis, and so on. This project will utilize the core concepts underlying the recent computer vision innovations to address a rarely-discussed, yet urgent issue in engineering: how to analyze the explosively increasing image data including images and videos, which are difficult to analyze with traditional methods. These concepts will be employed to explore the possibility of accurately assessing the safety of retaining walls with image data. This effort aims at setting up a paradigm for connecting engineering disciplines to artificial intelligence and enhancing the safety of geosystems as an essential infrastructure component by enabling their analysis with image-data-driven deep learning. The project will help revitalize traditional artificial intelligence sub-areas in geotechnical and other engineering areas, just as deep learning rekindled the interest in artificial neural networks and machine learning, and turned them into leading players in STEM research and innovations. The project may also change engineers' opinions regarding how to create knowledge with a revolutionary way attributed to deep learning, i.e., learning directly from data instead of indirectly from models established based on the data. Innovative education and outreach effort will be made by means of developing a mobile app to disseminate the idea and products of the project. The project will contribute to education by outreaching to K-12 students, underrepresented groups, and geotechnical engineering researchers and practitioners with the project products including the app at various events at the PIs' institution and professional conferences. The goal of this study is to understand the image-data-driven deep learning in geosystems with an exploratory investigation into the stability analysis of retaining walls. To achieve the goal, the recent breakthroughs in computer vision, which were later used as one of the core techniques in the development of Google's AlphaGo, will be studied for its capacity in assessing the stability of a typical geosystem, i.e., retaining walls. The core concept enabling machines to surpass humans in visual classification capacity, i.e., convolutional neural nets (CNN), will be used to process the big data in geotechnical engineering, which primarily consist of still and live images (videos), that cannot be readily analyzed using traditional geotechnical engineering methods. Conventional neural nets will be used to analyze images for retaining walls to tell whether a wall is safe or failed. For quantitative analysis, 2D and 3D images for retaining walls will be generated using stochastic methods and analyzed using traditional limit analysis and numerical methods for labeling. These labeled image data will be used as input to train convolutional neural nets for supervised learning. The trained nets will be tested against another independent set of data generated in the same way as the training data. Three research tasks will be conducted in this project: 1) understanding the data science for image-data-driven geotechnical engineering research, 2) investigating the connections between those image patterns in deep learning and the physical mechanisms, and 3) revealing the robustness and extrapolation capacity of the deep learning approach in the stability analysis of retaining walls.