一种基于深度置信网络的车撞桥墩？に鹕说燃杜斜鸱椒？
A method for discriminating damage level in vehicle？？bridge pier？？collision based on deep belief network

为了解决车撞桥墩过程中桥墩受损等级不易被实时监测和提前预警的问题，在深度学习理论框架下，对车撞桥墩损伤等级判别方法进行研究。在探讨车撞桥墩过程及其影响因素的基础上，引入场论，分析碰撞过程动力学和能量转换关系，并构建桥墩结构损伤场势计算公式；借鉴风险评估概念，对不同场势值所属的损伤等级进行定义，建立车撞桥墩损伤等级判别模型。利用深度置信网络能够准确提取数据本质特征和Softmax分类器可有效判别多等级概率的优势，构建基于深度置信网络的车撞桥墩损伤等级判别模型。将已有研究中相关试验和仿真数据输入构建的网络模型中，对网络结构中参数进行调试，并将该模型判别结果与传统机器学习判别方法结果进行对比验证。研究结果表明：针对该模型设计的3种网络结构，网络深度为7层时，模型判别结果的精度最高，平均绝对误差和平均相对误差分别为0.37和14.8%；相比人工神经网络模型，提出的模型采用深层神经网络结构，判别结果更为稳定，平均绝对误差和平均相对误差分别降低了1.34和？？11.5%？？；相比随机森林模型，提出的模型可自动提取数据的特征属性，所得判别等级结果的偏差更低，不超过0.5级，且其平均绝对误差和平均相对误差分别降低了0.46和4.7%，在多损伤等级判别问题中效率和准确度更高。
In order to solve the problem that the damage level of bridge pier was not easy to be monitored in real time and early warning, under the deep learning theory framework, the method of discrimination the damage level in the bridge pier was studied. The field theory was used to analyze the dynamics and energy conversion relationship during the collision of a car with a bridge pier, and a formula to calculate the damage field potential of the bridge pier was derived. The concept of risk assessment was used to define the damage levels based on different damage field potential values, and the damage level discrimination model for the vehicle？？bridge pier collision was established. The DBN can accurately extract the essential characteristics of the data and the Softmax classifier can effectively discriminate the multi？？level probability. Hence, these two were combined to establish the damage level discrimination model for vehicle？？bridge pier collision. The relevant experimental and simulation data from existing research were input to the network model formulated, the parameters of the network structure were debugged, and then the results were compared with those obtained from the traditional discrimination methods based on machine learning. The results show that among the three network structures designed in this study, the deep belief network (DBN) model has the highest accuracy when the network depth is seven layers. The mean absolute error and the mean relative error are 0.37 and 14.8%, respectively. Compared to the artificial neural network model, the DBN model adopts the deep neural network structure, the discrimination level result is more stable, and the mean absolute error and the mean relative error are reduced by 1.34 and 11.5%, respectively. Compared to the random forest model, the DBN model can automatically extract the characteristic attributes of the data, the deviation of the obtained discrimination level result is lower, not exceeding 0.5 level, and the mean absolute error and the mean relative error are