基于LM-BD-ESN模型的路面抗滑性能预测

Prediction of pavement skid resistance based on LM-BD-ESN model

针对路面抗滑性能预测任务中存在的指标单一和预测精度差等问题,在传统回声状态网络(echo state network,ESN)模型的基础上,提出了逻辑映射(logistic mapping,LM)和偏差丢失(bias dropout,BD)优化的改进回声状态网络模型(LM-BD-ESN)。其中,LM模块能够优化输入权重矩阵,从而与多变量非平稳序列数据产生更高的契合度;BD模块能够自主删除多余的存储单元,从而降低模型复杂度。针对路面材料与抗滑性能之间存在的非线性关系描述,基于三维测量仪采集路面的多组三维形貌数据,分别利用支持向量机(support vector machine,SVM)、相关向量机(relevance vector machine,RVM)、极限学习机(extreme learning machine,ELM)、ESN及LM-BD-ESN对路面抗滑数据进行分析验证。结果表明,所提LM-BD-ESN算法在预测任务中的均方根误差和平均绝对百分比误差分别为0.0858和0.0664,相较于其他算法具有更高的效率和精度。

In order to solve the problem of single index and poor prediction accuracy in the prediction of pavement skid resistance,an improved echo state network(LM-BD-ESN)model optimized by logical mapping(LM)and bias dropout(BD)was proposed on the basis of the traditional echo state network(ESN)model,which was further applied to the prediction task of pavement anti-sliding performance.Among them,the LM module can optimize the input weight matrix,thus producing a higher degree of fitting with the multivariate time series data.The deviation loss module of BD can delete redundant storage cells independently,thus reducing the complexity of the model.In view of the nonlinear relationship between pavement materials and anti-sliding performance,multiple sets of three-dimensional topography data of pavement was collected based on three-dimensional measuring instrument.Based on this,the anti-sliding data of pavement were analyzed and verified by support vector machine(SVM),relevance vector machine(RVM),extreme learning machine(ELM),ESN,and LM-BD-ESN,respectively.The results show that the root mean square error and average absolute percentage error of the LM-BD-ESN algorithm proposed in this paper are 0.0858 and 0.0664,respectively in the prediction task,which has higher efficiency and accuracy compared with other algorithms.