This paper proposes a prediction model for three-dimensional spatio-temporal soil deformation in bucket excavation. The prediction model for soil deformation (PMSD) consists of two machine learning processes: the long short-term memory (LSTM) and convolutional autoencoder (Conv-AE). These processes use datasets obtained from an experimental apparatus for bucket excavation developed in this work. The apparatus equips multiple depth cameras that precisely capture time-series data of soil deformation in bucket excavation. The LSTM, an extension of a recurrent neural network, successively predicts three-dimensional soil deformation. The Conv-AE is incorporated to both ends of the LSTM in order to quasi-reversibly compress and reconstruct the datasets so that the computational burden of the LSTM is relaxed. Qualitative and quantitative evaluations of the PMSD confirm the feasibility of time-series prediction of three-dimensional soil deformation. The Conv-AE shows sufficient accuracy equivalent to the measurement accuracy of the depth camera. The prediction accuracy of the PMSD is about 10 mm in most of the cases.

提出了一种斗式开挖土体三维时空变形预测模型。土壤变形预测模型 (PMSD) 由两个机器学习过程组成：长短期记忆 (LSTM) 和卷积自动编码器 (Conv-AE)。这些过程使用从这项工作中开发的铲斗挖掘实验装置获得的数据集。该装置配备多台深度摄像头，可精确捕捉铲斗开挖过程中土壤变形的时序数据。LSTM 是循环神经网络的扩展，连续预测三维土壤变形。Conv-AE 被合并到 LSTM 的两端，以准可逆地压缩和重建数据集，从而减轻 LSTM 的计算负担。PMSD 的定性和定量评估证实了三维土壤变形时间序列预测的可行性。Conv-AE 显示出足够的精度，相当于深度相机的测量精度。在大多数情况下，PMSD 的预测精度约为 10 mm。