基于概率神经网络的地铁车站易损性分析

Fragility Analysis of a Subway Station Based on Probability Artificial Neural Network

摘要:

基于深度学习方法提出了一种地震响应概率模型，并基于此模型推导了地铁车站结构极限状态超越概率的计算公式，以评价结构的地震易损性。首先采用主成分分析对地震强度指标进行正交化和降维；为了克服传统地震响应概率模型中地震动强度指标与结构地震响应指标服从对数空间线性分布假设的局限性，基于BP神经网络建立趋势模型以预测结构的地震响应；基于概率神经网络建立误差模型以描述基于统计的趋势模型与基于物理机制的数值模型之间的误差，以拓展残差的齐次方差正态分布假设的严格限制。最后，以上海典型2层三跨地铁车站为算例，计算得到其易损性曲线。结果表明，基于深度学习建立的趋势模型较好地模拟了地铁车站结构地震响应随地震强度指标第一主成分的非线性变化特征；建立的误差模型准确地描述了趋势模型预测值的残差随地震强度指标第一主成分的方差非齐次特征。

Abstract:

The exceedance probability of a limit state of subway station is deduced based on the novel probabilistic seismic demand model （PSDM） proposed in the present paper using the deep learning method. Principal component analysis （PCA） was used to orthogonalize IMs and reduce the dimension of IMs. The trend model to predict the seismic responses of structure was established based on the back propagation （BP） neural network， which avoids the limitation of the assumption of the traditional PSDM that the demand measure （DM） of structure has a linear relationship with the intensity measure （IM） of ground motion in the log-transformed space. The error model to describe the error between the statistics-based trend model and the physical mechanism-based numerical model was established using the probabilistic neural network， which can expand the limitation of the assumption that the residuals is normally distributed with homogeneous variance. Taking a two-story and three-span subway station in Shanghai as a case study， the fragility curves of the subway station were developed based on the proposed method. The results show that the trend model established based on deep learning well simulates the nonlinear change of the seismic response with the first principal component of IMs. The established error model accurately describes the nonhomogeneous variance of residuals of the seismic responses predicted by using the trend model.