This paper explores state-of-the-art deep learning techniques to analyse and predict structural dynamic nonlinear behaviours in civil engineering applications. In this paper, three methods, namely the piecewise linear least squares (PLLS) method, fully connected neural network (FCNN) method, and long short-term memory neural network (LSTMNN) method, are implemented and compared for structural dynamic response application under the condition of periodic, impact and seismic load. These methods are based on auto-regression model and time series estimation model, and still work when the structure is excited using immeasurable inputs. The dynamic response of a six-story steel frame analysed using the finite element method is used to validate these methods. Experimental results reveal that the PLLS and FCNN methods based on auto-regression model performs less well than the LSTMNN method based on time series estimation model, and it has a large the prediction peak mean square error. In addition, PLLS method is sensitive to noise, but FCNN and LSTMNN method based on deep learning are highly robust and anti-noise performance. These reveal the feasibility of the application of deep learning method in structural behaviours analysis in civil engineering.

本文探讨了最新的深度学习技术，以分析和预测土木工程应用中的结构动态非线性行为。本文分别对分段线性最小二乘法（PLLS），全连接神经网络（FCNN）和长短期记忆神经网络（LSTMNN）方法三种方法进行了比较，并在结构动力响应下进行了比较。周期性，冲击和地震载荷的条件。这些方法基于自回归模型和时间序列估计模型，并且当使用不可测量的输入来激发结构时仍然有效。使用有限元方法分析的六层钢框架的动力响应用于验证这些方法。实验结果表明，基于自回归模型的PLLS和FCNN方法的性能优于基于时间序列估计模型的LSTMNN方法，并且具有较大的预测峰均方误差。另外，PLLS方法对噪声敏感，但是基于深度学习的FCNN和LSTMNN方法具有很高的鲁棒性和抗噪性能。这些揭示了将深度学习方法应用于土木工程结构行为分析的可行性。